The Oxford Handbook of Reference 9780199687305, 0199687307

Citation preview

OUP CORRECTED PROOF – FINAL, 17/1/2019, SPi

   

REFERENCE

OUP CORRECTED PROOF – FINAL, 17/1/2019, SPi

OXFORD HANDBOOKS IN LINGUISTICS Recently Published

THE OXFORD HANDBOOK OF NAMES AND NAMING Edited by Carole Hough

THE OXFORD HANDBOOK OF DEVELOPMENTAL LINGUISTICS Edited by Jeffrey Lidz, William Snyder, and Joe Pater

THE OXFORD HANDBOOK OF INFORMATION STRUCTURE Edited by Caroline Féry and Shinichiro Ishihara

THE OXFORD HANDBOOK OF MODALITY AND MOOD Edited by Jan Nuyts and Johan van der Auwera

THE OXFORD HANDBOOK OF PRAGMATICS Edited by Yan Huang

THE OXFORD HANDBOOK OF UNIVERSAL GRAMMAR Edited by Ian Roberts

THE OXFORD HANDBOOK OF LANGUAGE AND SOCIETY Edited by Ofelia García, Nelson Flores, and Massimiliano Spotti

THE OXFORD HANDBOOK OF ERGATIVITY Edited by Jessica Coon, Diane Massam, and Lisa deMena Travis

THE OXFORD HANDBOOK OF WORLD ENGLISHES Edited by Markku Filppula, Juhani Klemola, and Devyani Sharma

THE OXFORD HANDBOOK OF POLYSYNTHESIS Edited by Michael Fortescue, Marianne Mithun, and Nicholas Evans

THE OXFORD HANDBOOK OF EVIDENTIALITY Edited by Alexandra Y. Aikhenvald

THE OXFORD HANDBOOK OF LANGUAGE POLICY AND PLANNING Edited by James W. Tollefson and Miguel Pérez-Milans

THE OXFORD HANDBOOK OF PERSIAN LINGUISTICS Edited by Anousha Sedighi and Pouneh Shabani-Jadidi

THE OXFORD HANDBOOK OF ELLIPSIS Edited by Jeroen van Craenenbroeck and Tanja Temmerman

THE OXFORD HANDBOOK OF LYING Edited by Jörg Meibauer

THE OXFORD HANDBOOK OF TABOO WORDS AND LANGUAGE Edited by Keith Allan

THE OXFORD HANDBOOK OF MORPHOLOGICAL THEORY Edited by Jenny Audring and Francesca Masini

THE OXFORD HANDBOOK OF REFERENCE Edited by Jeanette Gundel and Barbara Abbott For a complete list of Oxford Handbooks in Linguistics please see pp –

OUP CORRECTED PROOF – FINAL, 17/1/2019, SPi

   

......................................................................................................................

REFERENCE ...................................................................................................................... Edited by

JEANETTE GUNDEL and

BARBARA ABBOTT

1

OUP CORRECTED PROOF – FINAL, 17/1/2019, SPi

3

Great Clarendon Street, Oxford,  , United Kingdom Oxford University Press is a department of the University of Oxford. It furthers the University’s objective of excellence in research, scholarship, and education by publishing worldwide. Oxford is a registered trade mark of Oxford University Press in the UK and in certain other countries © editorial matter and organization Jeanette Gundel and Barbara Abbott  © the chapters their several authors  The moral rights of the authors have been asserted First Edition published in  Impression:  All rights reserved. No part of this publication may be reproduced, stored in a retrieval system, or transmitted, in any form or by any means, without the prior permission in writing of Oxford University Press, or as expressly permitted by law, by licence or under terms agreed with the appropriate reprographics rights organization. Enquiries concerning reproduction outside the scope of the above should be sent to the Rights Department, Oxford University Press, at the address above You must not circulate this work in any other form and you must impose this same condition on any acquirer Published in the United States of America by Oxford University Press  Madison Avenue, New York, NY , United States of America British Library Cataloguing in Publication Data Data available Library of Congress Control Number:  ISBN –––– Printed and bound by CPI Group (UK) Ltd, Croydon,   Links to third party websites are provided by Oxford in good faith and for information only. Oxford disclaims any responsibility for the materials contained in any third party website referenced in this work.

OUP CORRECTED PROOF – FINAL, 17/1/2019, SPi

C .............................................

List of Abbreviations About the Contributors

. Introduction

vii xi



J G  B A

PART I FOUNDATIONS: REFERENTIAL FORMS AND THEIR INTERPRETATION . Reference as a speech act



P H

. Referential intentions



M O’R

. Joint reference



A B

. Cognitive status and the form of referring expressions in discourse



J K. G, N H,  R Z

. Different senses of ‘referential’



N H, J G,  K B

. Definiteness and familiarity



B A

. The indefiniteness of definiteness



B A

. Indefiniteness and specificity



K  H

. De re/de dicto E K  F S



OUP CORRECTED PROOF – FINAL, 17/1/2019, SPi

vi



. Negative existentials



L C, M R,  A S

. A taxonomy of uses of demonstratives



R B. D  G W

. Contextual influences on reference



C R

PART II IMPLICATIONS AND APPLICATIONS: PROCESSING AND ACQUISITION OF REFERENCE . Reference and referring expressions in first language acquisition



A S O

. Reference resolution: A psycholinguistic perspective



E K  E F

. Accessibility and reference production: The interplay between linguistic and non-linguistic factors



J V, E K,  A M

. What can neuroscience tell us about reference?



B B

. Processing anaphoric relations: An electrophysiological perspective



C B  R K

. Computational generation of referring expressions: An updated survey



E K  K  D

. Reference in robotics: A Givenness Hierarchy theoretic approach



T W  M S

. Computational models of referring: Complications of information sharing



K  D

References Index

 

OUP CORRECTED PROOF – FINAL, 17/1/2019, SPi

L  A .........................................................................................................

PP PP

first person pronoun second person pronoun

PP

third person pronoun

SG SG

first person singular third person singular

γ a

googled (from the internet) action

A

audience, agent

 ACT

accusative Set of Activated Entities

 ART

adverb article



aspect

c/

causal-historical classifier

CNF CNP

Conjunctive Normal Form phrase of the category of common nouns



complementizer

CP D

complementizer phrase domain

 DEF

dead end definite

Det/D

determiner

DNF DP

Disjunctive Normal Form determiner phrase

DRS DRT

Discourse Representation Structure Discourse Representation Theory

DUR

durative aspect

E EEG

expression electroencephalogram

ELAN

early left anterior negativity

OUP CORRECTED PROOF – FINAL, 17/1/2019, SPi

viii

  

ERP

Event Related Potential

FAM FID

Set of Familiar Entities free indirect discourse

FOA 

Focus of Attention genitive

GH

Givenness Hierarchy

GREC H

Generating Referring Expressions in Context helper, hearer

HCRC HYP

Human Communication Research Centre hypothesization

I

intention

IA IC

Incremental Algorithm implicit causality

IFA IM

Intensional Functional Application indefinite marker

IMPERF

imperfective

IND INF

indicative infinitive

IPG 

Intersective Predicate Generalization Judicious Overspecification

KB

Knowledge Base

KR L

Knowledge Representation listener

L LAN

second language left anterior negativity

LF 

logical form lack of orientation



locative

LPC LTM

late positive complex Long Term Memory

mM

morpheme(s) member (of a dyad)

MASI

Measuring Agreement on Set-valued Items

MLU MOD

Mean Length of Utterance modifier

N NC

noun non-control

NEG

negative

OUP CORRECTED PROOF – FINAL, 17/1/2019, SPi

   NFS

Nearest-First Search

NLG NMN

Natural Language Generation nominalizer

NNS 

non-native speaker nominative, nominalizer

NP

noun phrase

Nref NS

referentially induced frontal negativity native speaker

O OBL

object, outcome oblique



object marker

 OVS

open proposition Object–Verb–Subject

P P-

property, principal pragmatic

p

proposition

PAH PART

Position of Antecedent Hypothesis partitive

PC PERF

partitive constraint perfective

PMI

Pointwise Mutual Information

PMT POSS

pure metalinguistic theory possessor

PP pro-

prepositional phrase pronoun

PROG PRP

progressive Perfect Recall Percentage

PRT

particle

Q Q

question first maxim of Quantity

Q QR

second maxim of Quantity Quantifier Raised/Raising

R

referent

RDF 

resource description framework referring expression

REF REFL

referential reflexive

REG

Referring Expression Generation

ix

OUP CORRECTED PROOF – FINAL, 17/1/2019, SPi

x

  

REL

relativizer

RM S

referentiality marker speaker, sentence, subject

SM SPS

subject marker syntactic positive shift

SUBJ

subject

SUP SVO

superlative Subject–Verb–Object

 t

salience weight time

T

map feature token, tense

ToM 

Theory of Mind topic

TP U

tense phrase utterance

VP

verb phrase

VPE

verb phrase ellipsis

OUP CORRECTED PROOF – FINAL, 17/1/2019, SPi

A  C .......................................................................................................................

Barbara Abbott is an Emeritus Professor of Linguistics and Philosophy at Michigan State University, where she taught from  to . Her main research interests are in semantics, pragmatics, and philosophy of language. She has published one book and a number of articles on topics ranging from reference and noun phrase interpretation to conditional sentences. Christopher Barkley has been a researcher in the Department of Experimental and Clinical Pharmacology at the University of Minnesota, Twin Cities, since . His research focuses on using imaging methods and pharmacological probes to better understand the relationship between language and other cognitive systems, especially working memory. Anne Bezuidenhout has taught at the University of South Carolina since , where she is Professor of Philosophy and Linguistics and affiliate member of the Mind and Brain Institute. Her research focuses on verbal communication in natural conversational settings which she studies using both formal and experimental methods. Kaja Borthen is Professor of Linguistics at Norwegian University of Science and Technology in Trondheim, Norway, where she is working in the section for Scandinavian Languages. She has done research on the semantics and pragmatics of various kinds of nominal expressions, and is currently leading a research project on pragmatic markers. Berit Brogaard is Professor of Philosophy at University of Miami and Professor II at University of Oslo. Her areas of research include philosophy of mind, philosophy of language, and cognitive science. She is the author of Transient Truths (Oxford University Press, ), On Romantic Love (Oxford University Press, ), The Superhuman Mind (Penguin, ) and Seeing & Saying (Oxford University Press, ). Leonard Clapp is a Professor of Philosophy at Northern Illinois University, where he serves as Director of NIU’s graduate program in philosophy. Most of his research concerns the interface between semantics and pragmatics, and he is particularly interested in dynamic approaches to discourse and conversation. Ryan B. Doran is Assistant Professor in the Department of Philosophy and Classics at the University of Regina (Saskatchewan, Canada). He completed his PhD in philosophy at Northwestern University in . His research interests are in the philosophy of language on topics such as reference and pragmatics.

OUP CORRECTED PROOF – FINAL, 17/1/2019, SPi

xii

  

Emily Fedele received her PhD from the University of Southern California in , focusing on reference resolution in Italian and English. In her dissertation, she investigated the interpretation of null and overt pronouns in both anaphoric and cataphoric configurations. She is currently a research associate with the Institute for Defense Analyses. Jeanette Gundel is Professor of Linguistics at the University of Minnesota, Twin Cities, where she has been teaching since . She is also Associate Director of the Center for Cognitive Science and an affiliate member of the Department of Philosophy. Her research focuses primarily on the interface between linguistic theory and pragmatics, especially reference and information structure. Peter Hanks is the John M. Dolan Professor in the Department of Philosophy at the University of Minnesota. His research is in philosophy of language and the history of analytic philosophy, with a focus on propositions and the early Wittgenstein. His book, Propositional Content, Oxford University Press, appeared in . Nancy Hedberg is Professor of Linguistics and Cognitive Science at Simon Fraser University in British Columbia, Canada. She has worked on cleft sentences and other copula sentences, on reference, and on the meaning of prosody. She is currently doing research on reference and performance aspects of oral discourse in the Salish language Hul‘q’umi‘num’. Elsi Kaiser is an Associate Professor in Linguistics at the University of Southern California. She earned a PhD in Linguistics from the University of Pennsylvania in . Her primary research focus is in language processing and psycholinguistics. She investigates the processes and representations involved in domains that involve multiple aspects of linguistic representation, such as reference resolution. She has conducted cross-linguistic work in typologically diverse languages in multiple language families (including Finnish, Estonian, French, German, Dutch, Bangla/Bengali, Hindi, Italian, Korean, Mandarin Chinese, and Vietnamese). Ezra Keshet is Associate Professor of Linguistics and Philosophy at the University of Michigan. He chiefly works on theoretical semantics, especially intensionality, focus, and anaphora. Recent research topics include conditional conjunctions (e.g., ‘Take another step and I’ll shoot!’) and a new dynamic semantics for plural anaphora. Robert Kluender is Professor of Linguistics at the University of California, San Diego, where he has been teaching since . He also serves on the Executive Committee of the Center for Research in Language. His main research interest is language and brain, focusing on syntactic processing, island constraints, and language evolution. Emiel Krahmer is a full professor of Communication, Cognition, and Computation in the Tilburg School of Humanities and Digital Sciences. In his work he studies how people communicate with each other, both in text and in speech, with the aim of subsequently improving the way computers communicate with human users. He was

OUP CORRECTED PROOF – FINAL, 17/1/2019, SPi

  

xiii

PI on various externally funded research projects, including an NWO VICI project on the production of referring expressions, which ran from  to . Alfons Maes is a full Professor and Head of the Department of Communication and Digital Sciences, Tilburg School of Humanities and Digital Sciences. His work includes research on multimodal aspects of human communication with studies on visual metaphor, visual health communication, multimodal navigation communication, and information visualization. He received his PhD in . Michael O’Rourke is Professor of Philosophy and faculty in AgBio Research at Michigan State University. His research interests include epistemology, communication and epistemic integration in collaborative, cross-disciplinary research, and linguistic communication between intelligent agents. He became interested in referential intentions in graduate school and has continued to work on related topics. He is Director of the NSF sponsored Toolbox Dialogue Initiative that investigates approaches to facilitating interdisciplinary research. Marga Reimer is a Professor of Philosophy at the University of Arizona in Tucson, where she has taught since . Her main research interests are the philosophy of language, the philosophy of psychiatry, and (more recently) the intersection between these two areas. Craige Roberts is an Emeritus Professor of Linguistics at The Ohio State University. She studies formal semantics and pragmatics in the tradition of generative linguistics. She has worked on anaphora and its relationship to presupposition and the context of interpretation since her  dissertation. Anne Salazar Orvig is professor of Linguistics at the University Sorbonne Nouvelle— Paris , where she has been teaching since . She has conducted various studies in the field of pragmatics, both on dialogue and dialogism and on language acquisition. Currently, she is working on a dialogical approach to the acquisition and usage of referring expressions by young children. Matthias Scheutz is a Professor of Cognitive and Computer Science in the Department of Computer Science at Tufts University. He earned a PhD in Philosophy from the University of Vienna in  and a Joint PhD in Cognitive Science and Computer Science from Indiana University, in . His current research focuses on complex cognitive robots with natural language capabilities. Florian Schwarz is Associate Professor of Linguistics and a member of the graduate group in Psychology at the University of Pennsylvania. His research integrates work in formal semantics and pragmatics with psycholinguistic methods, with a recent focus on experimental investigations of presuppositions and implicatures. Two of his main areas of theoretical work in semantics are definite descriptions and intensional semantics, specifically situation semantics.

OUP CORRECTED PROOF – FINAL, 17/1/2019, SPi

xiv

  

Anne Spire holds a BA and MA from the University of Arizona. Her research interests include formal semantics, pragmatics, and the use of semantics in web technologies. Kees van Deemter is Chair in Computing Science at the University of Aberdeen. He works in Computational Linguistics, and his main area of expertise is Natural Language Generation. He is interested in logical and philosophical issues arising from this work, and has collaborated extensively with psycholinguists interested in algorithmic models of human language production. His book “Computational Models of Referring: a Study in Cognitive Science” (MIT Press, ) examines the production of referring expressions, combining computational and experimental methods, and placing Referring Expressions Generation in its historical context. Jorrig Vogels is a postdoctoral researcher at the Center for Language and Cognition, University of Groningen (CLCG), where he works on a personal NWO VENI project on the role of perspective taking in the production of referring expressions. He is interested in both experimental and theoretical approaches to reference production and comprehension, especially in the influence of factors that exceed the local linguistic context. He received his PhD from Tilburg University in . Klaus von Heusinger is Professor and Chair of German Linguistics and director of the Collaborative Research Center “Prominence in Language” at the University of Cologne. His research interests include the semantics and pragmatics of referring expressions and semantics-pragmatics in general. His current research focuses on differential object marking, discourse structure, and the encoding of definiteness and specificity in different languages, including Germanic, Romance, Altaic, and Semitic. He has edited various volumes and published in various journals, including Journal of Semantics, Lingua, Research on Language and Computation, and Theoretical Linguistics. Gregory Ward received his PhD in Linguistics from the University of Pennsylvania in . He is currently Professor of Linguistics, Gender, and Sexuality Studies and, by courtesy, Philosophy, at Northwestern University, where he has taught since . His research is in the general area of discourse pragmatics, with specific interests in pragmatic theory, information structure, and reference/anaphora. Tom Williams is an Assistant Professor of Computer Science at Colorado School of Mines. He earned a joint PhD in Cognitive Science and Computer Science from Tufts University in . His current research focuses on enabling natural language capabilities for intelligent agents operating in uncertain and open worlds. Ron Zacharski is Associate Professor of Computer Science at the University of Mary Washington. He has worked on machine translation and multilingual information extraction. He is currently working on deep learning approaches to natural language processing.

OUP CORRECTED PROOF – FINAL, 17/1/2019, SPi

CHAPTER



......................................................................................................................

INTRODUCTION ...................................................................................................................... JEANETTE GUNDEL AND BARBARA ABBOTT

. S C O P E

AND INTENDED AUDIENCE

.................................................................................................................................. T HE intended audience for this volume includes linguists, psycholinguists, computational linguists, philosophers of language, and others interested in the relation between referring expressions in natural language and the things they apply to/pick out— roughly speaking, their meaning or intended interpretation. We will not attempt to pin down a more precise definition of the term ‘reference’ as it is intended in the title of this volume, because usage varies widely both across disciplines, and even within a single discipline, depending on the perspective of the user and the kinds of questions and broader concerns that motivate interest in the topic. Philosophers of language distinguish ‘singular reference’, the relation between singular terms and what they apply to, as a special case of ‘reference’. In this usage, definite descriptions (e.g., the woman in black, the most elegant solutions), proper names (e.g., Barack Obama), demonstrative expressions (e.g., this apple, that, these) and personal pronouns (e.g., she, them) are generally considered to be singular terms, regardless of whether they are grammatically singular or plural. Singular terms are distinguished from quantificational phrases (e.g., few wickets, most pianos), which cannot be assigned a unitary denotation. A wider conception of the term ‘reference’ comes about due to Gottlob Frege’s classic  paper ‘Über Sinn und Bedeutung’ [‘On sense and reference’]. In Frege’s theory, the reference/meaning (‘Bedeutung’) of a sentence is its truth value, and his principle of compositionality dictates that the reference of a complex expression is determined by the references of its parts (plus the syntactic structure that holds these together). Absence of reference of a part of a sentence results in absence of truth value for the whole. Thus, for any sentence with a truth value, each of its nonfunctional constituents—including the individual words (not just singular terms)—must have a reference. A different conception of reference, sometimes termed ‘pragmatic reference’, was sparked most notably by the philosopher P. F. Strawson’s  classic ‘On Referring’

OUP CORRECTED PROOF – FINAL, 17/1/2019, SPi



JEANETTE GUNDEL AND BARBARA ABBOTT

(a reply to Russell’s  classic ‘On Denoting’). Strawson argued that it was wrong to speak of expressions as referring—since a single expression (e.g., the present king of France) could be used by different speakers at different times to speak of different individuals (even though there may not be any existing individual that fits that expression at a given point in history). Instead, Strawson proposed that referring is something people use expressions to do, a three-place relation between speakers, linguistic expressions, and the world/their (intended) interpretations. Linguists have taken this pragmatic view of reference one step further by being concerned not just with what a speaker’s use of a referring expression contributes to the truth value of an utterance, but also with how speakers—by using a particular expression (e.g., she, that woman)—succeed in picking out their intended referent for the addressee (given that a single form rarely if ever uniquely determines a single referent), and how the use of different types of expressions in different contexts affects addressee interpretations. The distinction between semantic reference (something expressions do) and pragmatic reference (something speakers do) also has implications for the kinds of expressions that are involved. On the semantic conception, most kinds of linguistic expression, from individual words/morphemes to whole sentences, might be said to have reference. On the pragmatic conception, it would only be NP/DPs¹ that are involved. In this volume, we attempt to provide a state-of-the-art overview of reference used in the different senses above, with primary emphasis on pragmatic reference. We will consider different expression types and their contribution to the interpretations of sentences/utterances, as well as the knowledge/activities of speakers/writers who use such expressions to refer and of hearers/readers in (their efforts at) interpreting acts of reference. We will also look at implications and applications of reference from the perspective of its development in the individual and from psycholinguistic, neurolinguistic, and computational perspectives. The book is divided into two parts: Part I (Chapters –) covers foundational issues including different types of referring expression and their interpretation. Part II (Chapters –) covers applications and implications, specifically the acquisition and processing of reference from the perspectives of psycholinguistics, neurolinguistics, and computational linguistics, including robotics. An overview of the chapters is presented below.

¹ In the initial decades of Chomskyan generative grammar, phrases such as ‘the woman’, ‘Barack Obama’, ‘a horse’, ‘they’ were considered to be projections of the category of noun; hence the label ‘Noun Phrase (NP)’. While this usage continues, many linguistics have since adopted the view that such phrases should be considered projections of the category of determiner (Det/D), hence the label ‘Determiner Phrase (DP)’. Under the latter approach, an NP is a constituent within the DP that serves as the complement of D and contains the head N of the NP and any restrictive modifiers. DPs without an overt determiner (e.g., dogs in dogs are barking) are argued to have an empty D node. Some linguists consider forms traditionally referred to as ‘pronouns’ (e.g., she, this) to be determiners with an empty NP complement.

OUP CORRECTED PROOF – FINAL, 17/1/2019, SPi

INTRODUCTION

. O V E R V I E W



OF CHAPTERS

..................................................................................................................................

.. Part I. Foundations: referential forms and their interpretation In Chapter , ‘Reference as a speech act’, Peter Hanks discusses the view that reference is an act, a relation between people and objects, or perhaps more basically, between people, thoughts, and perceptual experiences, rather than just between linguistic forms and objects. He challenges the primacy of this kind of speech act reference over linguistic reference, arguing for a more nuanced view of the relation. Hanks shows that it is possible to use a referring expression without thereby performing an act of referring and conversely to refer to something in speech without uttering a linguistic expression, arguing that the circularity is broken if one distinguishes—within speech acts of reference—between semantic reference and speaker reference, which are related to each other in a complex, interdependent system, where speaker reference is not prior to semantic reference as there is no simple linear relation between acts of speaker reference and the relation of linguistic reference. Thus, speaker reference is an act the speaker performs using a linguistic expression where the speaker has some particular entity in mind and intends to single out/identify that entity, and linguistic expressions are meaningful only insofar as they make contributions to the sentences—and thereby illocutionary acts—in which they occur. To the extent that acts of speaker reference involve semantic reference, speaker reference implies/includes linguistic reference and semantic reference, and the latter is meaningful only insofar as it makes a contribution to the sentence—and thereby illocutionary act—in which it occurs. Michael O’Rourke in Chapter , ‘Referential intentions’, draws together work on referential intentions from philosophy, linguistics, and psychology. He provides a critical review of referential intentions themselves, pointing out that there is no referring device that hasn’t received theoretical treatment in terms of intentions of language as viewed at least in part in the practical context of speech activity, addressing their role in securing reference or their status as intentions. He discusses what he believes to be the primary motivation for introducing referential intentions into semantic and pragmatic theory, namely speaker control over one’s meaning, and supplies two taxonomies of referential intentions: levels of intention in discourse, and intentions that attach to specific lexical items. In Chapter , ‘Joint reference’, Anne Bezuidenhout, adopting the view that reference is primarily something speakers do and only secondarily something words do, argues that in everyday conversational settings, reference is a purposive joint action performed by interlocutors. She proposes that reference is an act performed in the course of some collaborative exchange between two or more communicators and referential communication succeeds when the hearer/addressee has a cognitive fix on the same object that the speaker does. Thus it is the dyad that succeeds in referring, not only the individuals

OUP CORRECTED PROOF – FINAL, 17/1/2019, SPi



JEANETTE GUNDEL AND BARBARA ABBOTT

taken singly. Bezuidenhout argues further that what is involved is not the code or even the speaker’s attempt to use the code to refer to the object she has a cognitive fix on, but rather thought, which does not involve a particular language (only precepts). It is thus the joint actions involved in conversational interactions that are the sites for achieving actions and allowing the interlocutors to arrive at internal representations of the same object. In this way meaning is jointly constructed converging on a joint interpretation, which could take several turns. Thus, the processes underlying an action of referring are distributed processes, which are not confined to the heads of single individuals, but are distributed across the heads of two or more individuals as well as the external environment in which they interact. While Chapters , , and  address basic questions about the nature of reference in human language—such as what it takes to refer and who refers—Gundel, Hedberg, and Zacharski in Chapter , ‘Cognitive status and the form of referring expressions in discourse’ (reprinted from a paper that appeared in Language in ), address the question of what makes it possible to understand referring expressions, given that a speaker’s intended referent is rarely, if ever, uniquely determined by the descriptive/ conceptual content encoded in the form alone, and thus a given object can be referred to in many different ways and the same form can be used to refer to many different objects. To answer this question they propose that in addition to conceptual descriptive content, referring expressions in human language also contain determiners and pronouns that encode information about how and where the intended referent can be accessed in the mind of the addressee. Specifically, these forms encode one of six cognitive statuses on what the authors call a ‘Givenness Hierarchy’ where each status entails all lower statuses, but not vice versa. The proposal is supported and illustrated by results of corpus studies of correlations between referring forms and cognitive status in five different languages: English, Japanese, Mandarin Chinese, Russian, and Spanish. Hedberg, Gundel, and Borthen in Chapter , ‘Different senses of “referential”’, show how the term ‘referential’ is used within the Givenness Hierarchy framework presented in the previous chapter, namely as an instruction to the hearer to construct a unique representation by the time the sentence has been processed, and they discuss how this both differs—and is similar to—other senses of the term ‘referential’ that have been used in the linguistic and philosophical literature. Barbara Abbott, in Chapter , ‘Definiteness and familiarity’, discusses the notion of definiteness, which has received various treatments in the literature, with some authors defining it in terms of uniqueness and others arguing that the conventional import of definite descriptions is familiarity. Abbott provides a detailed history as well as comparison and evaluation of the two theories, providing a summary of problems with the familiarity theory, as well as contexts which give rise to familiarity implicatures and explanations for them. In Chapter , ‘Indefiniteness of definiteness’, a reprinted paper from a volume that appeared in , Barbara Abbott is concerned with the difficulties involved with different attempts to characterize the property of definiteness and different types of referring expressions that can be definite, and asks whether existing

OUP CORRECTED PROOF – FINAL, 17/1/2019, SPi

INTRODUCTION



characterizations of the notion of definiteness capture its essence. She reviews a number of these attempts in detail, concluding that referential uniqueness appears to be the strongest contender. Chapter , ‘Indefiniteness and specificity’ by Klaus von Heusinger, is concerned with different ways in which specificity, ‘having a particular referent in mind’, is modeled, focusing primarily on indefinite expressions. Using the classic example ‘Mary wants to marry a Swede’ on the reading where Mary has a particular referent in mind (the specific reading), as distinct from the interpretation where she simply wants to marry one or another Swede, von Heusinger notes that this distinction, which has been treated as similar to Quine’s  distinction of indefinite phrases in opaque contexts and the de dicto/de re distinction of definites, yields ambiguity, even though the readings are truth conditionally equivalent. He provides a classification and detailed discussion of seven kinds of contexts for specificity-related phenomena and illustrates the variety of theoretical approaches to specificity with four families of theories of specificity, providing a critical evaluation of each. He argues that there is a core notion of specificity underlying the intuitive concept, namely referential anchoring, where the referent of a specific indefinite is dependent on a discourse participant, and where the content of the anchoring function must be unfamiliar to the hearer to distinguish specific indefinites from definites. Ezra Keshet and Florian Schwarz in Chapter , ‘De re/de dicto’ discuss the ambiguity that arises with respect to the distinction between de re and de dicto readings, which overlaps with the specific/non-specific distinction discussed in the previous chapter. They provide a brief history of this distinction in the philosophical literature which goes back to Aristotle, and review the basics of a scope theory (which goes back to Quine) which attempts to provide a formal account of this distinction and lay out an extensive overview of empirical phenomena that have been described in terms of the de dicto/de re distinction. Keshet and Schwarz also discuss approaches that have been proposed as alternatives to the traditional scope theory, provide a detailed comparison of the two theories they consider to be the most promising, and discuss de dicto/de re ambiguities that go beyond the noun phase. Finally, they note the broader practical relevance of the distinction and briefly discuss specific examples of its application in the practice of law. In Chapter , ‘Negative existentials’, Leonard Clapp, Marga Reimer, and Anne Spire address the problem raised by sentences of the form ‘Pegasus does not exist’, namely the fact that we can use such sentences to make true and informative statements despite the fact that they assert that the subject does not exist, while at the same time presupposing that it does. They review a number of proposals made by analytic philosophers to avoid or explain away this supposed contradiction, and evaluate these, noting that no consensus as to the solution has been arrived at yet, concluding that the problem cannot be solved independently of providing an adequate theory of verbal communication as a whole. Chapter , ‘A taxonomy of uses of demonstratives’, by Ryan Doran and Gregory Ward, provides a classification of thirteen broadly defined uses of demonstratives such as English ‘this’, ‘that’, ‘these’, and ‘those’, including some which have not previously

OUP CORRECTED PROOF – FINAL, 17/1/2019, SPi



JEANETTE GUNDEL AND BARBARA ABBOTT

been distinguished in the literature. The classification is based on the semantic value of the demonstratives, specifically whether they are used to refer to an entity, a kind, a predicate, or a variable in a quantification domain, as well as on pragmatic factors, and is made in the hope of providing a basis for identifying common features of these referring expressions. In the last chapter in Part I, Chapter , ‘Contextual influences on reference’, Craige Roberts addresses two basic questions that arise in a compositional account of reference interpretation: (i) how do speakers use linguistic expressions to refer in contexts of utterance and (ii) how does conventional content constrain such use. Roberts argues that all types of referring expressions involve dependence on context, and provides a wide range of examples and a detailed discussion of each type of context dependence, noting that for most referring expressions there is no fixed reference outside of context, and rather than noun phrases themselves referring, it is speakers that use these phrases to refer in context. She also discusses some trends in the literature which suggest the state of the art in explaining how context interacts with conventional content in the course of interpretation.

.. Part II. Applications and implications Part II begins with Chapter , ‘Reference and referring expressions in first language acquisition’ by Anne Salazar Orvig. Salazar Orvig reviews the growing literature on first language acquisition of referring expression in a wide variety of languages, drawing mainly on specific data from French, and summarizes the cross-linguistic milestones in order of acquisition of different types of referring expressions, including different determiners. She also points out that while studies of children’s narratives consider acquisition of anaphoric reference to be relatively late (as late as nine years in some cases) with earlier uses of referring expressions and determiners being egocentric, studies of naturally occurring dialogues yield very different results, with children having acquired a competence in use of referring expression by the age of three. She considers the factors that may account for this paradox including the child’s involvement in the co-construction of reference in a dialogic situation. In Chapter , ‘Reference resolution: a psycholinguistic perspective’, Elsi Kaiser and Emily Fedele provide a review of the laboratory-based experimental psycholinguistic research on reference, focusing mainly on comprehension of anaphoric reference by native speakers in a number of different languages. After providing an overview of the cross-linguistic variety of referring expressions, the authors review claims about how reference resolution is influenced by properties of the antecedent and also discuss a view according to which pronoun interpretation is a by-product of general inferencing about coherence relations between clauses. They suggest a way of reconciling the coherencebased and salience-based approaches by distinguishing ‘likelihood of mention’ from ‘likelihood of pronominalization’, and also consider the role of processing limitations and related considerations in the production and interpretation of pronouns.

OUP CORRECTED PROOF – FINAL, 17/1/2019, SPi

INTRODUCTION



Chapter , ‘Accessibility and reference production: the interplay between linguistic and non-linguistic factors’ by Jorrig Vogels, Emiel Krahmer, and Alfons Maes, reviews the literature concerned with the factors that influence a speaker’s referential choices in language production, exploring the data that exist from a variety of different languages on how linguistic factors interact with nonlinguistic factors in the choice of a referring expression in discourse. They also discuss the question of how speakers choose what they mention first in an utterance and how they refer to it, as well as the role the accessibility of mental representations plays therein. Brit Brogaard in Chapter , ‘What can neuroscience tell us about reference?’, argues that Discourse Representation Theory (DRT), or a related dynamic theory that (unlike traditional formal semantics) equates meaning with mental representations and thus can model discourse updates, can provide an adequate semantics of natural language for which findings from neuroscience can provide significant evidence. She reviews how neuroscientific findings lend support to reference-related phenomena such as anaphoric reference and proper names, and also considers methodological concerns that have been raised about existing neuroscientific approaches to natural language interpretation. In Chapter , ‘Processing anaphoric relations: an electrophysiological perspective‘, Christopher Barkley and Robert Kluender also address the implications of neuroscientific findings on reference, focusing specifically on anaphoric relations. They discuss studies that take advantage of the high temporal resolution of the Event Related Potential (ERP) technique to determine what the brain is sensitive to and when that sensitivity occurs when anaphoric relations are being processed. They provide a review of ERP studies on anaphoric relations, and also evaluate claims in the literature about the identity of brain response with different types of anaphoric relations and the consequences of these claims for the architecture of the language system and the functional specificity of brain responses to language input. Chapter , ‘Computational generation of referring expressions: an updated survey’, by Emiel Krahmer and Kees van Deemter, provides a survey of the computational literature on generation of referring expression, an update of a review that was published in  in Computational Linguistics. The survey describes a range of computational approaches to the problem of generating referring expressions and discusses how algorithms addressing this problem have been experimentally tested. Among the new developments discussed in this volume are the Bayesian methods which are reviving the idea that speakers’ choice of referring expressions may be driven primarily by rationality, increasing interaction between researchers in vision and language, and modeling ways in which speakers and hearers collaborate to ensure together that they are attending to the same referent. In Chapter , ‘Reference in robotics: a Givenness Hierarchy theoretic approach’, Tom Williams and Matthias Scheutz propose a domain-dependent open-world reference resolution system/algorithm (GH-Power) for robotics, making use of the Givenness Hierarchy system discussed in Chapters  and  of the present volume. They also review theoretical concerns which provide motivation for future work and discuss their

OUP CORRECTED PROOF – FINAL, 17/1/2019, SPi



JEANETTE GUNDEL AND BARBARA ABBOTT

approach in relation to other approaches to reference resolution in robotics as well as shortcomings being addressed in ongoing work. Chapter , ‘Computational models of referring: complications of information sharing‘, by Kees van Deemter, examines the limitations of algorithms in the area of referring expression generation in models of human language production, focusing on issues that arise not from the logical or linguistic structure of the expressions, but from the situation in which the referring expression is uttered, and also provides a discussion of how these issues can be addressed. Building on the concept of information sharing (information that is shared, as distinct from the rest of the sentence that presents new information) van Deemter discusses challenges that arise when sharing is not a simple interplay between given and new information, for example when a speaker is unsure about what the hearer knows or when there are situations in which identification of the referent is not the goal of the description.

ACKNOWLEDGMENTS We would like to thank Mai Al-Khatib, Link Swanson, and Paul Tilleson, graduate students in the Cognitive Science and Linguistics programs at the University of Minnesota, for their invaluable help at various stages of this project.

OUP CORRECTED PROOF – FINAL, 17/1/2019, SPi

P A R T I .............................................................................................................

FOUNDATIONS: REFERENTIAL FORMS AND THEIR INTERPRETATION .............................................................................................................

OUP CORRECTED PROOF – FINAL, 17/1/2019, SPi

OUP CORRECTED PROOF – FINAL, 17/1/2019, SPi

CHAPTER



......................................................................................................................

REFERENCE AS A SPEECH ACT ...................................................................................................................... PETER HANKS

. I N T R O D U C T I O N

.................................................................................................................................. D ISCUSSIONS of reference in philosophy and linguistics are often framed as investigations into a semantic relation between linguistic forms and things. Proper names, pronouns, indexicals (e.g., ‘I’ or ‘today’), demonstratives (e.g., ‘this’ and ‘that’), and (perhaps) definite descriptions all refer—in some of their instances and often only in context—to objects. Let’s call this kind of reference linguistic reference. There is, however, a more fundamental relation of reference that holds between people and objects. People refer to things, and their doing so makes it possible for expressions to refer. Linguistic expressions, after all, are just collections of sounds or marks, with no intrinsic ability to stand for anything else. It is because of the way people use these sounds and marks that they come to refer to things outside themselves. To refer in this more basic sense is to do something—it is to perform an action. Reference in this sense is a kind of action. Familiar examples of this kind of action occur in speech, when someone utters a word or phrase and thereby refers to something. There are also non-verbal examples involving pointing, grasping, touching, and other gestures. Acts of reference also occur (arguably) when someone has a thought or perceptual experience that is directed at a particular individual. It may be that such reference is more basic than reference in speech and other behavior, in which case a full account of reference would lead ultimately to the philosophy of mind and psychology.¹ In this chapter, however, I will only be concerned with reference as a spoken act, that is, acts of reference in which a speaker refers to something through the use of a linguistic expression. Let’s call this speech act reference. I will attempt to do two things. The first will involve a closer examination of the priority of speech act reference over linguistic reference. As we will see, this claim of priority can be challenged, and meeting this challenge leads to a more nuanced view of the relationship between spoken acts of ¹ See e.g., Searle () and Campbell ().

OUP CORRECTED PROOF – FINAL, 17/1/2019, SPi



PETER HANKS

reference and linguistic reference. The second will be to argue that spoken acts of reference are possible only as components of illocutionary acts with satisfaction conditions, such as assertions, questions, orders, and promises. This rules out freestanding acts of reference. It also draws a conceptual connection between spoken acts of reference and satisfaction conditions. The speech act of reference is always an element in a larger speech act with truth conditions, answerhood conditions, or fulfillment conditions. To refer to something in speech is to make a certain sort of contribution to these sorts of satisfaction conditions. This entails that it is possible to use a referring expression without thereby performing a speech act of reference, and, conversely, that it is possible to refer to something in speech without uttering a linguistic expression that refers to that thing. Examples of both will be provided below.

. A C T S

OF REFERENCE AND

LINGUISTIC REFERENCE

.................................................................................................................................. Linguistic reference, as I have been using the term, is a relation that holds between a linguistic expression, the object to which it refers, and, in many cases, a context of utterance. The name ‘Barack Obama’ refers to Barack Obama. The occurrence of the word ‘I’, in my utterance of ‘I am cold’, refers to me. I claimed above that this relation of linguistic reference is based on the explanatorily prior fact that people use expressions to refer to things. In a simple and naïve formulation, an expression E refers to x because people use E to refer to x. But it could be argued that the order of priority is in fact the reverse—that linguistic reference should be used to explain speech act reference. In the introduction to a recent volume on reference Andrea Bianchi writes: But it is almost indisputable . . . that a crucial role is played by the fact that (some of ) the linguistic expressions we use have semantic properties that connect them to extralinguistic entities. Thanks to these properties, they may be used by us to refer to things. Bianchi (: )

In other words, it is because of the fact that a linguistic expression refers to a certain entity that people refer to that entity by using that expression. A natural way of putting this is to say that referring expressions exert a kind of control over the acts people perform by using them. They control not only the kind of linguistic act a speaker performs but also the content of that act. A speaker performs an act of reference using the name ‘Barack Obama’, as opposed to an act of predication, quantification, or something else, because this name is a referring expression. Furthermore, she refers to Barack Obama, and not, say, Hillary Clinton, or the city of Chicago, because the name ‘Barack Obama’ refers to Barack Obama. Or consider John Perry’s example of Heimson, who is confused and thinks he is David Hume (Perry ). Heimson says

OUP CORRECTED PROOF – FINAL, 17/1/2019, SPi

REFERENCE AS A SPEECH ACT



‘I am David Hume’, believing his use of ‘I’ to refer to David Hume. But he is wrong. The meaning of ‘I’ controls who Heimson refers to.² We are already in a bit of a quandary here. On the one hand, there is a strong intuition behind the idea that the semantic relation of reference between linguistic expressions and things must be based on prior facts about spoken acts of reference. Something has to explain how an expression comes to refer to what it does. It can’t be magic. The only plausible place to look for such an explanation is in the uses that people make of words. On the other hand, the semantic properties of the referring expression a speaker uses seem to control both the fact that the speaker performed an act of reference and, in many cases, the identity of the object to which the speaker referred. That makes the semantic properties of referring expressions explanatorily prior to the speech act of reference. This has us going around in a circle, with speech act reference explaining linguistic reference and vice versa. We can break the circle by invoking a distinction—within speech acts of reference— between semantic reference and speaker reference. The distinction is best illustrated with Kripke’s Jones–Smith example: Two people see Smith in the distance and mistake him for Jones. They have a brief colloquy: ‘What is Jones doing?’ ‘Raking the leaves’. ‘Jones’, in the common language of both, is a name of Jones; it never names Smith. Yet, in some sense, on this occasion, clearly both participants in the dialogue have referred to Smith, and the second participant has said something true about the man he referred to if and only if Smith was raking the leaves (whether or not Jones was). Kripke (: )

When the second participant says ‘Jones is raking leaves’, the semantic referent of his use of ‘Jones’ is Jones and the speaker referent is Smith. Put another way, he semantically refers to Jones but speaker refers to Smith. Now, it should be clear that linguistic reference controls the speech act of reference only for semantic reference. When the man says ‘Jones is raking leaves’ he semantically refers to Jones, and does so because the name ‘Jones’ refers to Jones and not Smith. And the speaker refers to Smith, even though Smith bears no semantic relation to the name ‘Jones’. With this distinction in hand we can now say that facts about speaker reference ground and explain linguistic reference; and facts about linguistic reference ground and explain semantic reference. Linguistic expressions refer to things, perhaps relative to contexts, because people have used those expressions to speaker refer to things in contexts. And once expressions refer to things it is possible for speakers to perform acts of semantic reference. The (grossly oversimplified) picture that emerges is one in ² This example works because ‘I’ is a ‘pure’ or ‘automatic’ indexical, i.e., a context-sensitive expression whose meaning is sufficient for determining its semantic value in a context. Other kinds of nonautomatic context sensitive referring devices (e.g., ‘that’ and ‘this’) require supplementation in order to secure a referent, such as a demonstration, or a referential intention on the part of the speaker (see King a, b for discussion). In these cases the meaning of the referring expression does not have full control over the referent of an act of reference, although it still controls the fact that it is an act of reference and not some other kind of linguistic act.

OUP CORRECTED PROOF – FINAL, 17/1/2019, SPi



PETER HANKS

which, prior to the existence of semantic relations between linguistic expressions and objects (in linguistic prehistory, as it were), there were acts of speaker reference in which speakers used sounds and marks to refer to things. These acts of speaker reference became conventionalized, eventually solidifying into the semantic properties of referring expressions. With those semantic properties in place, linguistic expressions took on the ability to control the acts of semantic reference performed by speakers. Summing up, first came acts of speaker reference, then linguistic reference, and finally acts of semantic reference. The picture is pleasingly simple and intuitive, but in fact it is a distortion. It obscures the fact that the speech act of reference—whether semantic reference or speaker reference—depends on a set of background practices and norms involving both referring expressions and acts of reference. Linguistic reference, semantic reference, and speaker reference are all tied up with one another in a complex, interdependent system. Kripke’s own account of speaker reference helps make the point. He says that ‘we may tentatively define the speaker’s referent of a designator to be that object which the speaker wishes to talk about, on a given occasion, and believes fulfills the conditions for being the semantic referent of the designator’ (Kripke : , my emphasis). This defines the notion of speaker reference partly in terms of the notion of semantic reference. If Kripke is right, and the concept of semantic reference figures crucially in the definition of speaker reference, then it cannot be that speaker reference is somehow prior to semantic reference. Here is another way to draw out the difficulty for the simple, intuitive picture. The picture asks us to imagine a time when linguistic reference did not exist, yet people referred to things with sounds or marks that were not (yet) referring expressions. Now, it is not hard to imagine someone pointing at, or touching, something while making a sound, but in that case the referential work is done behaviorally, not via the utterance of the sound. Nor is it hard to imagine someone referring to something with a previously meaningless noise or symbol, as long as this is set against a background of established practice of using names and other linguistic expressions to refer to things. What is hard to imagine, I submit, is a case in which someone refers to something just by uttering a meaningless sound, in the absence of any general practice of using sounds to refer to things. An analogy might be helpful here. Consider the rook in chess.³ There are pieces in chess that are rooks, and there are moves in chess games that are moves of the rook. Armed with this distinction we can ask: which came first, rooks or rook-moves? That is similar to asking: which came first, referring expressions or acts of speaker reference? The analogy with chess is useful because the answer is obvious. The game of chess came first, and it brought both rooks and rook-moves with it. Similarly, in the case of reference, the answer is that the set of practices and norms that constitute the general

³ The analogy with chess is, of course, an allusion to Wittgenstein’s use of language-games to investigate meaning and understanding (Wittgenstein ).

OUP CORRECTED PROOF – FINAL, 17/1/2019, SPi

REFERENCE AS A SPEECH ACT



phenomenon of reference in language came first, and it brought linguistic reference, semantic reference, and speaker reference along with it. The lesson is that there is no simple linear explanatory relationship between acts of speaker reference and relations of linguistic reference. Acts of reference and referring expressions are elements in a complex and holistic phenomenon, with each element allowing for and reinforcing the others. It may be that this phenomenon as a whole is founded on more basic acts of behavioral, cognitive, or experiential reference, but it is a mistake to think that one strand in this web can be pulled out and made the basis for the others.

. A C T S

OF REFERENCE AND

ILLOCUTIONARY ACTS

.................................................................................................................................. To this point we have been examining the relationship between the speech act of reference and linguistic reference. Let us now focus on the nature of the speech act itself, and in particular on the act of speaker reference.⁴ What is it for someone to speaker refer to something? As a start, we might say that it is an action the speaker performs, using a linguistic expression, where the speaker has some particular individual in mind, intends to single out or identify that individual, and, following Kripke, believes that the individual is the semantic referent of the referring expression she uses. This seems right as far as it goes—the problem is that it is not very informative. What is it to have a particular individual in mind? And what goes into intending to single out or identify an individual? These concepts are no less philosophically problematic and mysterious than the concept of reference. Progress is possible once we realize that acts of reference are possible only as elements of complete illocutionary acts with satisfaction conditions, such as assertions, questions, commands, or promises.⁵ This claim is the analog—for speech acts—of Frege’s Context Principle: “never ask for the meaning of a word in isolation, but only in the context of a proposition,” (Frege : x). Another way to put this is that linguistic expressions are meaningful only insofar as they make contributions to the meanings of sentences in which they occur.⁶ This is a claim about linguistic expressions, but a close ⁴ The act of semantic reference, based as it is on linguistic reference, is relatively unproblematic: for someone to semantically refer to an object x is for her to competently use an expression that refers to x. That said, the main conclusion of this section also applies to acts of semantic reference. Acts of semantic reference only occur as components of more comprehensive illocutionary acts with satisfaction conditions. ⁵ The concept of an illocutionary act was introduced by Austin () in his pioneering work on speech act theory. The concept was refined and systematized by Searle (, a). ⁶ Wittgenstein’s version of the Context Principle in the Tractatus Logico-Philosophicus captures this clearly: ‘Only the proposition has sense; only in the context of a proposition has a name meaning, (.)’ (Wittgenstein : ).

OUP CORRECTED PROOF – FINAL, 17/1/2019, SPi



PETER HANKS

analog applies to speech acts. Acts of reference, predication, quantification, etc. are the analogs of words, and whole illocutionary acts are the analogs of sentences. For acts of reference, then, the principle says: ‘never ask about an act of reference in isolation, but only in the context of an illocutionary act.’ Equivalently: acts of reference are possible only as they make contributions to complete illocutionary acts. Let us put reference aside for a moment and consider the act of predication. In the simplest case, an act of predication is an act of applying or attributing a property to an object. If I hold up a lime and say ‘This is green’, I predicate the property of being green of the lime. Now, it is fairly easy to see that an act of predication cannot occur in isolation. It makes no sense to just predicate, without an accompanying act of reference, or something else that serves to identify an object. You cannot perform an act of predication by just saying ‘green’, or ‘is green’. There has to be something else in the conversation that provides an object to be the target of your use of ‘green’. Maybe you have pointed to something, or you are answering a question about a previously identified object, or there is some salient object in the conversation. Note that in all of these cases you are asserting that something is green, and to assert something is to perform an illocutionary act. The speech act of predication cannot occur except as part of an assertion, or some other complete illocutionary act.⁷ The same goes for acts of reference. Acts of reference occur as components in acts of making assertions, asking questions, giving orders, making promises, and other illocutionary acts with satisfaction conditions. I can imagine a philosopher trying to refute this by staring at something and saying ‘that’, or by just uttering someone’s name.⁸ But that is something only a philosopher would do. In ordinary conversation we never just refer to something. We always do so as part of a broader speech act. The restriction to illocutionary acts with satisfaction conditions is important, and potentially controversial, because it rules out certain speech acts that may be thought to involve acts of reference. Suppose Alice is on the other side of a crowded room and I am trying to get her attention. I call out ‘Alice!’. This is a speech act, in particular, an act of trying to get someone’s attention, but it does not have satisfaction conditions. It has success conditions, but not satisfaction conditions. I may or may not succeed in getting Alice’s attention, but it doesn’t make sense to say that this act is true or false, or answered or unanswered, or fulfilled or unfulfilled in the way that an order or promise can be fulfilled. All speech acts have success conditions, that is, conditions for their successful performance, but only some have satisfaction conditions. Here is another example: suppose Andy has just entered the room, and I say ‘Hello Andy’. This is a greeting. My act is successful if Andy hears me and recognizes that I am greeting him.

⁷ See Hanks () for more on the act of predication. ⁸ ‘Naming appears as a queer connection of a word with an object—and you really get such a queer connection when a philosopher tries to bring out the relation between name and thing by staring at an object in front of him and repeating a name or even the word “this” innumerable times’ (Wittgenstein : §: ).

OUP CORRECTED PROOF – FINAL, 17/1/2019, SPi

REFERENCE AS A SPEECH ACT



But it does not make sense to evaluate my act as true or false, or answered or unanswered, or fulfilled or unfulfilled. One more example, from the TV show Seinfeld. A recurring trope on the show would have Jerry caught up in some mishap due to his nemesis Newman. Whenever that happened Jerry would scowl and say, with consternation and annoyance, ‘Newman’. The speech acts Jerry was performing in these cases were expressives, that is, acts of expressing emotional or psychological states (see Searle a: Ch. ). Other examples of expressives are utterances of ‘Thank you’, or ‘Sorry’—acts in which someone gives voice to a feeling of gratitude, or remorse, or some other mental state. In Jerry’s case, he was expressing feelings of anger or frustration about something Newman did. Now, an expressive can be successful or unsuccessful, depending on whether the speaker really has the underlying psychological state and succeeds in making it manifest. But, again, it doesn’t make sense to say that Jerry’s acts of saying ‘Newman’ are satisfied or not. The claim that acts of reference are always components of illocutionary acts with satisfaction conditions implies that saying ‘Alice!’ to get her attention, or ‘Hello Andy’, or ‘Newman’, are not acts of reference to Alice, Andy, or Newman. I am not referring to Alice when I call out ‘Alice!’—I am trying to get her attention. When I say to Andy ‘Hello Andy’, I am not referring to him, I’m greeting him. When Jerry says ‘Newman’, he is expressing his anger at Newman, and he could have done that by saying ‘Arrrgh’, or shaking his fist, or scowling, or any number of other things, none of which involve any kind of reference to Newman. These are all cases in which someone uses a referring expression that refers to some individual without performing a speech act of reference to that individual. If this is correct it means that we can analyze acts of reference in terms of the satisfaction conditions of the illocutionary acts of which they are components. To refer to an individual x is to perform an illocutionary act whose satisfaction conditions essentially and ineliminably involve x. For example, when I say ‘Hillary Clinton lives in New York’, I perform an assertion that is true if and only if Hillary Clinton lives in New York. Because these truth conditions essentially involve Hillary Clinton, I referred to Hillary Clinton in my utterance. Or suppose I ask ‘Does Hillary Clinton live in New York?’. In this case my question is answered by assertions whose truth conditions essentially involve Hillary Clinton, that is, the assertion that she lives in New York or the assertion that she does not. Suppose someone replies ‘Yes’ to my question. On the present account, the speaker referred to Hillary Clinton by saying ‘Yes’, since she said something whose truth conditions essentially involve Hillary Clinton, even though she hasn’t uttered any linguistic expression that refers to Clinton. It follows that it is possible to refer to an individual in a speech act without using any linguistic device that refers to that individual in that context. Here is another example: suppose I say to Hillary Clinton: ‘Please move to Minnesota’. In this case my request is satisfied if and only if Hillary Clinton moves to Minnesota. These fulfillment conditions essentially involve Hillary Clinton, and so I referred to her when I made the request, even though I didn’t use a name or any other referring device that refers to Hillary Clinton.

OUP CORRECTED PROOF – FINAL, 17/1/2019, SPi



PETER HANKS

Obviously, much more needs to be said by way of refining and clarifying this account of the speech act of reference. I cannot do that here, but let me close with a few final remarks about what it means for satisfaction conditions to essentially and ineliminably involve some individual x. Suppose I assert ‘The  Democratic presidential nominee lives in New York’. The truth conditions for this assertion involve Hillary Clinton, but they don’t essentially involve Hillary Clinton. It could have been that Bernie Sanders won the  Democratic nomination. In that case, my assertion would be true if and only if Bernie Sanders lives in New York. By contrast, when I say ‘Hillary Clinton lives in New York’, the truth of my assertion invariably depends on whether Hillary Clinton lives in New York, even when we consider different possible ways for the world to be. To adapt Kripke’s terminology, these truth conditions rigidly depend on the facts about Hillary Clinton (Kripke ). No matter which possible world we consider, it is always Hillary Clinton’s place of residence that matters for the truth of an assertion that Hillary Clinton lives in New York. Generalizing this to other kinds of speech acts, the claim is that to refer to an individual is to perform an illocutionary act whose satisfaction conditions rigidly depend on or involve that individual. More needs to be said to clarify what this means, but that would take us into the intricacies of speech act theory and beyond the scope of this chapter.

. C O N C L U S I O N

.................................................................................................................................. Reference can be understood as a relation between linguistic expressions and things, or as an action that people perform. This chapter has focused on the latter, in particular the acts of reference that people perform using linguistic expressions. I started with the natural thought that these actions are more fundamental than the relation of linguistic reference, but that gave way to the idea that both acts of reference and linguistic reference are mutually dependent aspects of the complex set of practices and norms that constitute the more encompassing phenomenon of reference. In the second half of the chapter I argued that acts of reference are always elements of more complex illocutionary acts with satisfaction conditions, such as assertions, questions, orders, and promises. This allowed for an account of acts of reference in terms of the satisfaction conditions of these acts: to refer to an individual is to perform an illocutionary act whose satisfaction conditions essentially or rigidly depend on that individual. This locates the analysis of the speech act of reference squarely within the theory of illocutionary acts, and the theory of intentional action more generally.⁹

⁹ Thanks to Ben Caplan and Jeanette Gundel for helpful comments and suggestions on an earlier draft of this chapter.

OUP CORRECTED PROOF – FINAL, 17/1/2019, SPi

CHAPTER



......................................................................................................................

REFERENTIAL INTENTIONS ...................................................................................................................... MICHAEL O’ROURKE

. I N T R O D U C T I O N

.................................................................................................................................. T HE practice of introducing intentions into semantic and pragmatic accounts of reference is widespread. It comes as no surprise that accounts of reference avail themselves of intentions when they are built to explain context sensitive lexical items such as certain indexicals (e.g., ‘now’, ‘she’, ‘you’) and demonstratives (e.g., ‘that’, ‘this’).¹ If I say to you as you reach into the refrigerator, ‘That is my beer,’ it is natural to interpret my use of ‘that’ as reflecting my intention to refer to the bottle you are about to grab. Positing referential intentions reflects our (defeasible) assumption that, in uttering the words intentionally, the speaker knows what she is talking about. This practice is not restricted to demonstratives and indexicals. Famously, Donnellan () distinguished referential and attributive uses of definite descriptions, with speaker intentions being one of the mechanisms adduced to explain the difference. Kripke () and Evans (/) both introduce intentions in making sense out of how we can refer with proper names. Indeed, it would appear that there is no referring device that has not received theoretical treatment in terms of intentions. This is not inevitable, of course— those who adopt formal approaches and prefer to treat language as a Platonic object will have little use for intentions; however, if one is committed to the idea that language should be understood at least in part in the practical context of speech activity, then intentions are ready to hand for use in both semantic and pragmatic explanations.

¹ I follow Kaplan (a) in dividing the context sensitive referring expressions into indexicals and demonstratives. Perry () offers a useful, related account involving automatic and intentional indexicals, with the intentional category comprising those terms whose reference is not determined automatically by the nature of the context. Indexicals in the Kaplanian usage comprise both Perry’s automatic category and some of his intentional category. It is in accounting for the semantics and pragmatics of those terms in the intentional category that theorists often introduce intentions.

OUP CORRECTED PROOF – FINAL, 17/1/2019, SPi



MICHAEL O’ROURKE

But while referential intentions are common, they have not been the subject of a critical review. This is that review. In surveying the landscape of referential intentions, I aim to take referential intentions seriously. It is not uncommon for referential intentions to function as placeholders in a theory of language, marking a site where speaker control is deemed relevant to interpretation but receiving little attention either in terms of their specific role in securing reference or their status as intentions. In what follows, I draw together work on referential intentions from philosophy, linguistics, and psychology, focusing both on what makes them referential and how they function as intentions. This project is distinctively philosophical—the concept referential intention combines elements of language with those of action, and a full account of it should blend theoretical work on reference in linguistics and the philosophy of language with theoretical work on intention in psychology and the philosophy of action; while such an account is beyond the scope of this chapter, I aim to make progress toward it by outlining ways in which referential intentions are conceptually constrained by reference on the one side and intention on the other. My specific goal is to supply an overview of these states that does justice to their variety while introducing constraints on their implementation in semantic and pragmatic theories of natural language. To this end, I begin by discussing what I take to be the primary motivation for introducing referential intentions into one’s semantic or pragmatic theory, viz., capturing the assumption of speaker control over one’s meanings. I then supply two taxonomies of referential intentions, focusing first on levels of intention in discourse and second on intentions that attach to specific lexical items, such as demonstratives and proper names. These taxonomies frame the subsequent analysis, which aims to detail important aspects of the form, content, and function of referential intentions. I conclude by briefly considering two critical concerns, one having to do with the threat of hyperintentionalism and the other with the explanatory status of referential intentions.

. M O T I V A T I O N

.................................................................................................................................. Referential intentions are a type of speaker intention, formed by speakers in the process of using language to refer to things in the world. They appeal as an explanatory element to theorists who take language to be a lived phenomenon, typically used by people to create, mediate, and extend complex networks of interpersonal relationships. So deployed, referential intentions are asked to illuminate language production and reception, revealing language to be a key feature of goal-directed activity. As such, referential intentions operate in the theoretical space opened up by the later Wittgenstein, Austin, and Grice. They are featured in semantic theories (e.g., Kaplan b; King ), pragmatic theories (e.g., Kripke ; Bach ), and theories of communication and discourse (e.g., Grosz and Sidner ; Roberts ; Stone ). In this section, I examine in more detail what motivates theorists of reference to

OUP CORRECTED PROOF – FINAL, 17/1/2019, SPi

REFERENTIAL INTENTIONS



employ referential intentions, arguing that these intentions reflect key assumptions about discourse generally and referential communication in particular. Before getting down to work, I simplify my task in this chapter in two ways. First, I won’t address whether the concept should play a semantic role or a pragmatic role, keeping the discussion free from the inevitable entanglements that arise when you engage in theoretical negotiation at that interface. My interest is in surveying the role that referential intentions play in theories of linguistic significance, which include both semantic and pragmatic theories. Second, I will not be concerned to specify the number of argument positions in the reference relation; that is, what I will say about referential intentions should be applicable whether or not we are talking about reference as a twoplace relation between a word and an object, a three-place relation between speaker, word, and object, or a four-place relation involving a listener as well. These are important topics, and they will arise occasionally as I move through the chapter, but they are largely orthogonal to my interest in referential intentions. A third simplification is applicable only to this section: what I have to say in here applies with equal force to speaker intentions, which is the more general category that includes referential intentions; given this, most of what I will have to say is directed at the broader category, unless otherwise indicated. While referential intentions—and speaker intentions more generally—appear in many accounts of language, their theoretical employment is not inevitable. This is perhaps obvious if you consider theories other than those that focus on utterances and language use, such as those that consider language to be a formal object and the project of language theory to be a province of logic, going back to much of the work of Russell and Frege. But even if you locate language in the lives of language users, it is not necessary to introduce these states into semantic or even pragmatic explanations. Some contributions to discourse are closely considered productions containing elements that issue from deliberate wordsmithing; for example, consider a prepared speech or a written document such as this chapter—the language these contain has been carefully crafted, sometimes down to the last word. In light of this, the suggestion that these words are intended by their authors seems reasonable; of course, whether these intentions are promoted into an explanatory role by a theorist is another matter—the words will also be written out in some graphical system, but that fact doesn’t seem relevant to their significance—but at least it seems clear that speaker intentions are available for promotion. By contrast, much discourse is more casual and free-flowing, comprising contributions produced without a lot of advance work or conscious monitoring. When we talk with friends, discuss the events of our day, or just ‘shoot the shit’, there are things we wish to say, but how we say them, especially at the lexical level, will often seem to be at least as much a thing revealed to us in the saying of them as it is to our audience. This is true, in particular, about our selection of referring expressions during referential discourse. Here the idea that it is intentions all the way down, or even down very far, is not as striking. Lack of conscious awareness is no proof that intentions are not

OUP CORRECTED PROOF – FINAL, 17/1/2019, SPi



MICHAEL O’ROURKE

involved, of course, but here their promotion into an explanatory role is more hypothetical, relying on analogies with other cases that are arguably similar.² If adoption of speaker intentions for theoretical work is not forced on the theorist, what explains their appeal? Why regard them as relevant to the project of explaining the significance of language? The primary reason has to do with the assumption that when it comes to language production, the speaker is the one who will (typically) have the last word on what she means. This is beyond challenge as an assumption underlying pragmatics, but it is also assumed by many doing semantics. Of course, unless we want to go over the wall with Humpty Dumpty, it would be a mistake to ignore the role played by linguistic convention in determining or at least constraining linguistic significance; however, the realities of contextually sensitive language and semantic underdetermination (see Bach ) forces even the biggest fan of convention to acknowledge that other factors are relevant to the determination of utterance significance. Discourse is rife with examples of speech situations involving linguistic items that are context sensitive and must be understood in light of what the speaker is talking about. Examples include many indexical pronouns (e.g., ‘you’, ‘she’), demonstratives (e.g., ‘that’), restricted quantifiers (e.g., ‘some of the students’), indefinite definite descriptions (e.g., ‘the table’), certain adjectives (e.g., ‘local’, ‘actual’), and various adverbs (e.g., ‘here’, ‘presently’).³ In addition, semantic underdetermination, exhibited by sentential forms that are in some way incomplete, is a common feature of standard discourse (e.g., ‘I have had breakfast’; When? Today? At some point in your life? (see Taylor ) and ‘The park is some distance from here’ (see Carston ; Sperber and Wilson )). When a speaker uses context sensitive expressions or sentences that must be completed to be understood, we presume that they have a specific, recoverable meaning because there is a fact of the matter about what the speaker means by them. If there is confusion or lack of clarity about what the speaker meant and they are available, we might consult them and ask what they meant, taking their response to be compelling evidence in favor of one interpretation. One way to model this type of speaker-centrism about utterance significance is by attributing intentions, including referential intentions, to the speaker as a way of indicating that the speaker typically has the last word on what their utterances mean. We can think of this as a speaker control assumption, and it is an assumption that underpins theories in which speaker intentions are given privilege of place. Linguistic

² “Explicitly formulated linguistic . . . intentions are no doubt comparatively rare,” Grice () tells us. “In their absence we would seem to rely on very much the same kinds of criteria as we do in the case of nonlinguistic intentions where there is a general usage” (: ). ³ The indexicals and demonstratives that figure into this list are what Perry  calls “discretionary”. For good, general discussions of these, along with the contextually sensitive adjectives and adverbs, see Perry  and Braun . For quantifier domain restriction, see Stanley and Szabo . For indefinite definite descriptions, see Donnellan  and Wettstein . See King  for discussion of all of these understood as supplementives.

OUP CORRECTED PROOF – FINAL, 17/1/2019, SPi

REFERENTIAL INTENTIONS



convention constrains meaning, perhaps even determining it in certain cases,⁴ but theorists interested in accounting for the significance of natural language, given its contextual flexibility, will acknowledge the need to attend to a speaker’s interests in speaking. Speech is not digestion—it is initiated, monitored, and concluded intentionally and its contours reflect conscious, intentional influence if not outright control. One important means for exerting control over a complex task is to be deliberate in acting, selecting moves carefully so that they combine to support the pursuit of one’s objectives. This sort of deliberate action is modeled in terms of intentions that arise out of practical reasoning about the task (e.g., Bratman ). The similarity of speech to other forms of complex, intentional behavior inclines language theorists to employ similar theoretical machinery, including intentions. In the case of referential discourse, the selection and coordinated use of referring devices to introduce referents for consideration can be seen as deliberate actions that issue from intentions to refer to those items.⁵ In developing a theoretical account of linguistic significance that makes room for intentions, speaker intentions are specifically motivated by characteristics of utterance behavior that prominently evince speaker control. These include the global decision, shaped by one’s discourse goals and knowledge of the language, to participate in the discourse by producing an utterance, and local decisions to shape the specific syntactic, lexical, and phonological profile of an utterance.⁶ In some cases, the relevance of intention to the significance of local aspects of the utterance is patent, such as when one says, ‘Don’t move—you have a hornet on your shirt’, to one individual in a group of people; here the audience will be very keen to determine to whom the speaker intended to refer. In other cases, however, recognizing the relevance of intention takes counterfactual reflection on the ways in which the utterance can go awry. Pathological utterances include out and out utterance failures, but also utterances that function in unexpected ways. It is not always the speaker’s fault when utterances fail to meet expectations,

⁴ The degree to which conventional meaning, rooted in the logical form of a sentence, determines the semantic content of the sentence is an issue at the heart of the debate about the semantic/pragmatic interface. Here I am interested in meaning understood broadly as what Grice called “total significance” (Grice : ; see also O’Rourke ), comprising both semantic and pragmatic inputs. ⁵ This discussion is not an endorsement of Humpty Dumpty’s position on linguistic meaning. Unless speakers are just talking because they like the sound of their own voice, they are typically out to achieve some end for which language production is a means; pursuit of that end requires them to respect the conventional forms and constraints that underwrite the utility of words in the first place. Even granting this point, this discussion might evoke Kaplan’s (b) discussion of subjectivist semantics, where “everyone runs their own language” (b: ). What I have said about the combination of linguistic convention and speaker control, though, is as applicable to what Kaplan calls “consumerist semantics” as it is to subjectivist semantics—even if we “adopt the typical consumer’s attitude of compliance” and “defer” to the meaning of the words we use, we still use them intentionally, and often with pragmatic intent that goes well beyond their “prepackaged” semantic value (b: ). ⁶ For an example of the theoretical use of intentions in modeling the phonology of language, see Bromberger and Halle ().

OUP CORRECTED PROOF – FINAL, 17/1/2019, SPi



MICHAEL O’ROURKE

of course—road noise can make it difficult to hear, listeners stop listening, and even conscientious auditors make mistakes. But some cases are pathological for reasons that have to do with the speaker. In these cases, utterance pathology reveals degrees of linguistic freedom available to the speaker, i.e., different ways in which the speaker might act that could either go well or go poorly. When the utterance is healthy, the speaker gets credit in the form of attributed intentions, marking their control over the relevant variables. For example, consider the use of a definite description to talk about a person in view, as in the classic example, “Who is the man drinking the martini?” (Donnellan : ). Donnellan asks us to consider the case in which the man is drinking water out of his martini glass; here, the definite description would fail to denote him. This failure, though, does not undermine the referential connection between the description and the teetotaler, according to Donnellan, and this is due to the fact that the speaker who asks the question intends to refer to the man with the definite description, with that intention securing the referential connection (: ). In asking this question, the speaker is free to select the description, and in this case he makes an honest mistake; however, he is also free to identify the target of his remark and use the description to refer to him, and Donnellan believes that if the speaker decides to use the description in this way he can be successful even if he fails to select an adequate description. This degree of control over the linguistic situation is grounded in an intention to refer with the description to the teetotaler. Reflection on pathological cases is a common part of theoretical modeling—you model the paradigm case and then complicate the model to accommodate relevant phenomena that depart from the paradigm. A locus classicus of this type of modeling is Grice (), in which a series of modifications are made to the initial analysis of utterer’s occasion meaning in response to challenges to the sufficiency and the necessity of this analysis. This type of theoretical modeling is also evident in theories of referential intentions, such as Perry (), Simchen (b), and King (). Consider King’s introduction of referential plans into his referential intention-based Coordination Account. These are introduced to deal with cases of “conflicting intentions,” where a speaker uses a term to refer to an object that he thinks possesses a certain attribute which it does not in fact possess. As King notes, the “classic case” of this in the literature is the Carnap/Agnew case from Kaplan (), where Kaplan says of a picture of Agnew behind him that it is a picture “of one of the greatest philosophers of the twentieth century”, thinking that it is a picture of Carnap. The conflicting intentions are the intention to refer to the picture behind him and to refer to the picture of Carnap. On one rendering of his Coordination Account—the “Best Laid Plans” version—King attempts to do justice to the intuition that Kaplan refers with his demonstrative ‘that’ to Agnew by introducing an assembly of three intentions that figure into referential plans, the initiating, intermediate, and controlling intentions (King : ). Of these intentions, the initiating intention to refer to a picture of Carnap and the controlling intention to refer to the picture behind him conflict, but it is the controlling intention that secures the semantic value of the demonstrative

OUP CORRECTED PROOF – FINAL, 17/1/2019, SPi

REFERENTIAL INTENTIONS



according to the Coordination Account. King is motivated here by a pathological case to complicate his story about referential intentions, increasing the number involved in making sense out of the semantic landscape and then embracing one of these as part of his explanatory account.⁷ Here again, consideration of the counterfactual profile of utterance behavior motivates introduction of referential intentions as a way of capturing aspects of significance over which the speaker is taken to exert some control. We can summarize the speaker control motivation for introducing speaker intentions into theories of linguistic meaning with the help of two guiding principles:

· Speaker Control Principle: If there is an aspect of an utterance that could fail to ·

come off and thereby undermine the speaker’s purpose in producing the utterance (e.g., lexical composition, phonological profile) because of something the speaker does but did not have to do, or doesn’t do but could have done, then that aspect falls under intentional control. Intention Generation Principle: If something falls under intentional control, then it could be modeled as the object of an intention.

The Speaker Control Principle expresses the idea that the speaker can decide (or not) to modify or manipulate an utterance in a number of ways so that it advances her toward her communicative goals, and a number of those decisions (or omissions) in any particular case could cause the utterance to go awry. The Intention Generation Principle is not an uncommon assumption in the history of the philosophy of action, although there are many who would dispute it (e.g., Bratman ). I will return to these principles below, but for now it is worth noting that they seem to be responsible for the explosion of intentions that operate in at least some intention-friendly accounts of reference.

. F I E L D

GUIDE TO REFERENTIAL INTENTIONS

.................................................................................................................................. In this section, I provide a field guide for identifying referential intentions. Referential intentions are instances of the broader class of speaker intentions, and so I will begin by providing a speaker intention framework and then locating referential intentions within it. Utterance production by human speakers is paradigmatically intentional, and this is often taken to be the effect of executed intentions as represented by the two speaker control principles in §.. To say that utterance production is intentional is to say that it was something the speaker did but could have done differently. Intentionality so understood applies at various levels, corresponding to the different contributions the ⁷ This move is similar to the one that Perry () makes: both are motivated by the Carnap/Agnew case to think of referential intentions as plans, complicating their theoretical accounts.

OUP CORRECTED PROOF – FINAL, 17/1/2019, SPi



MICHAEL O’ROURKE

utterance makes to the goals that motivated the speaker to produce it.⁸ Consider an utterance U made in a discourse of the sentence: ()

She was on a plane at that time.

We can view U from ‘above’, seeing it from a comprehensive discourse perspective or from the perspective of the specific speaking turn to which it contributes, taking it to be part of the speaker’s pursuit of her global or more local discourse goals. One can also view U reflexively, asking whether it expresses what the speaker wants expressed at that moment. One can also view it from below, as the aggregate sum of lexical decisions motivated by very local concerns about how to frame one’s specific point.⁹ If you understand utterance intentionality as operating on multiple levels, the speaker control principles will entail the legitimacy of positing intentions on each of those levels to ground the intentionality. Taking the category of utterances to include the production of the smallest meaningful units (i.e., morphemes), one could understand intentions to concern constituents of discourse ranging from morphemes at the lowest level to the discourse as a whole at the most global level.¹⁰ Within the nested set of discourserelevant constituents, there are multiple perspectives that can be taken to generate different intentions. Consider this nesting for the utterance U mentioned above: ()

morpheme ! word ! sentence ! assertion ! speaking turn ! discourse

U can be understood as comprising morphemes, which form the constituent words within the sentence. Following Austin (), these could be understood as pieces of the locutionary act, suggesting that one can also look at U as an illocutionary act, such as an assertion. Moving out, U is part of a speaking turn that could include more illocutionary acts, which itself contributes to a more general discourse. Since the speaker could have changed things up at any of these levels, our principles imply that intentions could be introduced at any of these levels to ground utterance intentionality. We can say more about these speaker intentions. Most broadly, at the level of discourse or interpersonal communication, we find discourse intentions (cf. Litman and Allen ). These could be intentions to communicate something to an audience, such as a complex position on a philosophical issue or a set of directions to the local brewery (e.g., Grosz and Sidner ; Roberts /). Less broadly, at the level of

⁸ Cf. Grosz and Sidner’s () discussion of discourse purposes and discourse segment purposes (: ). ⁹ It is worth pointing out that U is itself the aggregate sum of smaller utterances (e.g., utterances of the words ‘she’ and ‘plane’). In what follows I will want to understand utterance in this broader way, to include the production of smaller, sub-sentential linguistic elements. ¹⁰ What an intention concerns can be described in terms of contact points at each level. Linguistic intentions, when executed, result in actions, and the actions in question are productions of linguistic items. These items are the contact points. At the discourse level, the whole discourse is the primary contact point, whereas with illocutionary intentions speech acts are the primary contact points. At the lexical level, the contact points are words, and perhaps morphemes (e.g., when a speaker emphasizes a prefix or a suffix, as in ‘I’m UNconvinced’).

OUP CORRECTED PROOF – FINAL, 17/1/2019, SPi

REFERENTIAL INTENTIONS



the speaking turn or the produced sentence, we find what I’ll call illocutionary intentions. These include the intention to make a specific move toward one’s discourse objectives, make a specific claim, or perhaps have a specific effect on an audience (e.g., Grice ; Searle ; Kripke ). Most narrowly, at the level of the units of meaning, including morphemic, lexical, and sentential items, we locate locutionary intentions.¹¹ These include the intentions to utter certain words and phrases, produce certain prosody contours, and produce a particular sentence as a way of making a point (i.e., the intention to produce one sentence in particular, such as ‘I do not think it means what you think it means’).¹² In a typical conversation, there could be the following nested intentions behind the production of a specific contribution, moving from the most global at the top down to the most local: I¹₁: [participate in the conversation I²₁: [take my turn when I have something to say and it is appropriate I³₁: [make a cooperative contribution to the conversation I⁴₁: [make an assertion (or ask a question or . . . ) I⁵₁: [produce an utterance with a certain locutionary structure I⁶n: [utter and then . . . and then ]]]]]] This nested list is intended to be illustrative, not exhaustive. Intentions I¹ to I³ have the ring of conversational policies—i.e., general intentions that frame one’s engagement with conversational discourse whenever one chooses to engage in it—although circumstances can change that lead to suspension or withdrawal of those intentions (cf. Bratman ). Intentions I⁴ to I⁶ are more utterance specific, changing moment to moment during the course of the conversation. We can locate referential intentions within this framework. A paradigm case of referential intentions—the case of demonstrative reference—resides at the locutionary level, but one could argue that referential intentions are operative at higher levels as well. For example, if the discourse is about a particular thing—say you and your partner are discussing one of your children—then the intention to participate in that conversation is an intention to participate in referential discourse. This becomes perhaps more compelling if the content of the intention at that level were to mention the person under discussion, for example, intending to talk about Junior. Similarly, if you intend to make a complex referential point in your current speaking turn, uttering several sentences as a way of getting at a particular item, then the intention to take that turn

¹¹ For examples of theories that employ locutionary intentions, see Donnellan , Kaplan b, Perry , Simchen a, and King . ¹² All three types of intentions—discourse, illocutionary, and locutionary—can be understood to be communication intentions in the Gricean sense (Grice ; see Bach and Harnish ). Whether at the discourse level or more locally, the utterances produced are intended to have their effect via their recognition, or as Bach and Harnish () put it, “their fulfillment consists in their recognition” (: ).

OUP CORRECTED PROOF – FINAL, 17/1/2019, SPi



MICHAEL O’ROURKE

would appear to inherit a referential character. And this holds as well for illocutionary intentions to produce sentential utterances that are about particular items. Thus, intentions to perform illocutionary acts or combinations of illocutionary acts can be characterized as referential because they are about specific items, which could be represented in their content. They are referential because they make certain contributions to the discourse, they depend for their significance on how things stand with a particular item (or items, in the case of plural reference), and they do this in part by involving that item in their content. The referential character of these intentions, though, is in general derivative, arising out of the fact that in performing illocutionary acts the speaker will produce specific referring terms as a way of connecting their acts with the items that interest them.¹³ In light of this, close attention paid to reference at the locutionary level should provide us with what we need to make general sense out of referential intentions. Not all locutionary-level intentions are relevant to our purposes; for example, an intention to produce a sentence might be referential, but only in the derivative way just canvassed. The primary contact points for referential intentions at the locutionary level are noun phrases, such as definite descriptions and demonstratives. Let’s call locutionary intentions that fix aspects of the production of particular referential items (e.g., demonstrative pronouns or definite descriptions) sub-sentential referential intentions. These intentions will concern the production of devices of reference, such as indexicals and demonstratives, proper names, descriptions, and possibly other linguistic items that depend on a referential connection for their significance, such as restricted quantifiers and modifiers. How referential intentions are rendered varies across theories that make use of them. They are typically introduced either in general, ranging over a number of different types of referring expressions, or as part of an analysis of the semantics or pragmatics of a specific expression type, such as indexicals or descriptions. The standard form of a referential intention is this: ()

Intend to refer (L) to R with E

This schematic form highlights reference as an action—as something the speaker does in executing the intention to refer (cf. Searle ). Even so, the relation of reference involved can be interpreted either as a relation between an expression E and a referent R, or as something that involves a listener L, as when you refer someone to something.¹⁴ Examples of intentions introduced to do general work across types of ¹³ Roundabout ways of speaking about an object could be an exception to this, depending on how roundabout they are. In this case, we might regard an illocutionary act or perhaps a speaking turn as referential even though there is no term in the mix that stands for the item under discussion and the intentions that lead to the utterance do not mention it. This is not typical, though, and so while it is an interesting case, I will not consider it in what follows. ¹⁴ There is a process/product ambiguity at work here between reference understood as an intentional process grounded in a relation between speakers, listeners, words, and the world, and reference understood as a stable relation between words and the world. While traditionally within philosophy

OUP CORRECTED PROOF – FINAL, 17/1/2019, SPi

REFERENTIAL INTENTIONS



terms include the specific and general intentions Kripke () introduces into his account of speaker reference and semantic reference, respectively, where the specific intention is “on a given occasion, to refer to a certain object” and the general intention is “to refer to a certain object whenever the designator is used” (: ). Kronfeld () introduces conversational and functional referring intentions, with the former focused on getting L to think of R “via a particular mode of presentation” and the latter getting L to think of R without privileging any particular way of thinking about it (: ; see also : –). In both cases, these distinctions between types of referential intentions apply across the categories of referring devices;¹⁵ these accounts assume that if reference is to be treated as an intentional phenomenon, what matters is the speaker’s orientation to the referent and the listener, rather than the specific type of term that is employed for the purpose. As for specific expression types, referential intentions receive the most attention in the discussion of indexicals and demonstratives, which is natural given the semantically incomplete and contextually dependent character of these terms. The composition of these two categories remains in dispute, although our primary interest is in what Kaplan calls the true demonstratives, which typically need to be augmented by something (e.g., a demonstration) to make their referent salient. These include pronouns such as ‘she’, ‘it’, and demonstrative pronouns such as ‘this’ and ‘that’, among others (Kaplan a: ). Perry calls true demonstratives the “intentional” or “discretionary” indexicals (Perry : –, : ), marking the fact that the speaker has control over which item(s) in the world is introduced into the discourse by the term. Indexicals and demonstratives are the subject of an extensive literature, and much of it makes room for intentions. Kaplan’s (a, b) work on how true demonstratives secure their reference gave rise to a debate about the semantic relevance of intentions versus demonstrations that continues to this day (see Reimer ; Bach ; L. Roberts ; Mount ; Perry ; King b). In Kaplan (b), directing intentions are proposed as the mechanism for securing the referent at least in the case of perceptual demonstratives. Directing intentions are “aimed” at perceived objects as their referential targets (b: ), but beyond this, Kaplan offers little analysis; others, though, have come forward with their own analyses, such as Reimer (), who takes directing intentions to be intentions “to refer to, and say something of, a perceived object or individual on which one has ‘focused’” (: ).¹⁶ These analyses emphasize the role of the speaker in reference. Other speaker-centric of language the focus has been on reference understood as a word/world relation, some take the process of referring to be the more basic explanandum (e.g., Grice ; Clark and Wilkes-Gibbs ; Kronfeld ). On these views, referring is a speech act performed by a speaker with a linguistic item (e.g., a name or a description), while reference is a relation between word and world that is obtained by abstracting away from the act, and in particular from the speaker and listener. (See Ludlow  for discussion.) ¹⁵ Other accounts that are similarly designed with multiple expression types in mind are Bach (/ ) and King (, b). ¹⁶ Other developments of this concept are supplied by Mount  and Perry . For discussion of the ways in which these intentions can be understood as directing, see §...

OUP CORRECTED PROOF – FINAL, 17/1/2019, SPi



MICHAEL O’ROURKE

referential intentions discussed in connection with demonstrative reference include the intention to refer to what one is “thinking of ” (Perry : ), the intention to demonstrate what one has in mind (Reimer : ), and the intention to refer to what one is demonstrating (Bozickovic ). An importantly different type of referential intention is audience involving, and is represented by Bach’s () Gricean suggestion that when we use demonstratives, we intend that “one’s audience identify, and take themselves to be intended to identify, a certain item as the referent by means of thinking of it in a certain identifiable way” (: ). And there are also proposals that combine the speaker and audience, for example, Mount’s () “second-order” referential intention to “use demonstrations in a way that correctly identifies for others the object the speaker has in mind” (: ), as well as King’s “Coordination Account” (King , b). The influence of Donnellan’s () referential/attributive distinction has ensured that referential intentions also receive attention in theoretical treatments of descriptions. As he argued, we can use definite descriptions to talk about what is denoted by the property expressed (i.e., the attributive use) or we can use them to talk about a specific item in particular whether or not it is denoted by the property expressed (i.e., the referential use). When we use a description referentially, we “intend our audience to realize whom we have in mind” (: ), or alternatively, we intend that the audience identify what we have in mind on its basis (Donnellan : ). In Donnellan (), this audience-directed referential intention is paired with a truthconditional intention: the speaker who uses a description referentially “intends that truth or falsity shall be a function, in part, of the properties of the person or thing he has in mind” (: ). Others have followed him in agreeing that intentions play a significant role in grounding the referential deployment of definite descriptions (e.g., Kripke ; Wettstein ; Bach /; Kronfeld ). A recent development of this idea understands referential intentions for descriptions in this way: An intention to use a description to refer*, is an intention to induce the hearer to have descriptive cognitive fix on its designation, and thereby, in the light of knowledge the hearer has or can acquire, come to have a cognitive fix on it of the sort standardly imparted by names, demonstratives, and indexicals. Korta and Perry : 

Like Donnellan () and Wettstein (), Korta and Perry () expand the role of referential intentions beyond complete definite descriptions to those that are incomplete (e.g., ‘the book’), and others expand it to include indefinite descriptions (e.g., Donnellan ; Dekker ).¹⁷ Bach (/) goes even further, generalizing the referential/attributive distinction into a distinction between “referring objectually”

¹⁷ Hawthorne and Manley () also rely on referential intentions in accounting for the semantics of indefinite descriptions. Van Rooij () deploys referential intentions as part of accounts of the semantics of pronominal anaphora, particularly pronouns that are anaphoric on indefinite descriptions. For critical discussion of these attempts and others, see Lewis ().

OUP CORRECTED PROOF – FINAL, 17/1/2019, SPi

REFERENTIAL INTENTIONS



(i.e., the referential use) and “referring descriptively” (i.e., the attributive use), where the latter is rooted in an intention “to be talking about whatever uniquely satisfies a certain individual concept” (/: ).¹⁸ The third principal type of referring expression is the proper name, and it has also been analyzed in terms of referential intentions. In describing what he calls the “undifferentiated Description Theory” of proper names, grounded in the work of Searle (), Evans (/) argues that it depends on a speaker’s intention to refer “by his use of a name [to] that which satisfies or fits the majority of descriptions which make up the cluster of information which the speaker has associated with the name” (/: ). A causal theorist, Evans sets this aside in favor of a view that gives privilege of place to this intention: the intention “to refer to the item that is the dominant source of his associated body of information” (/: ), making reference track the etiology of what we know about the referent. Two other referential intentions that have received significant attention in connection with proper names are what we can call the deferential intention and the dubbing intention, both introduced in Kripke (). The deferential intention is the intention to use a name “with the same reference as the man from whom he heard it” (: ), and the dubbing intention is the intention to create a new practice in which a name is used to refer to some item.¹⁹ A more recent development of an intention-based theory is found in Simchen (a), who makes proper name reference depend on an intention to use a morpheme in a way that contributes to its “overall trajectory” as a name (a: ). Evans, Kripke, and Simchen each treat their preferred referential intentions as relativized to social name-using practices, which create the context within which the intention can secure a referential connection in a particular case.²⁰

. A N A L Y Z I N G

REFERENTIAL INTENTIONS

.................................................................................................................................. In this section, I consider in more detail the nature of referential intentions, focusing in particular on their form, content, and function in discourse. Working primarily with the sub-sentential intentions introduced at the end of the last section, my goal is to evaluate these states both from the perspective of reference and the perspective of intention, thereby discharging my promise to take referential intentions seriously.

¹⁸ For more discussion of intentions associated with the attributive use, see O’Rourke (). ¹⁹ For more discussion of these, see Evans (/: –) and Kaplan (a: ). ²⁰ Additional devices that have received treatment in terms of intentions include expressions such as relational expressions, quantifiers, and gradable adjectives, all of which are classified as supplementives by King (a) (see also Glanzberg ). These are not standardly regarded as devices of reference, but the intentions involved in supplying the implicit arguments necessary to parameterize these devices are referential, selecting specific items or ranges of items relative to which these terms are to be interpreted.

OUP CORRECTED PROOF – FINAL, 17/1/2019, SPi



MICHAEL O’ROURKE

.. The form of referential intentions I will begin by discussing the form, or structure, of referential intentions. We can think about the form of referential intentions extrinsically, in terms of their relations with other intentions and cognitive attitudes that give rise to and shape practical activity— speech, after all, is rarely disconnected from other goal-directed activities. We can also think about them intrinsically, in terms of their own structure qua cognitive states and, depending on the theoretical orientation, in terms of their involvement in assemblies of intentions that structure reference and referring. I will consider referential intentions from each of these points of view in turn. Extrinsically, it is helpful to think about referential intentions in relation to the levels of speech activity mentioned above, specifically the discourse and illocutionary levels. Roberts () suggests that “relations over intentions are the central organizing features of discourse” (: ) and provides a model of discourse context that depends on domain goals, or “separate goals in the real world” (: ), and discourse goals represented as questions under discussion. Grosz and Sidner () also take intentions, in the form of an “intentional structure,” to be a key part of what constitutes discourse (: ). On both of these accounts, the intentions involved will include referential intentions. Both accounts take the intentions that organize discourse to be Gricean, effecting the transfer of meaning between interlocutors via their mutual recognition. This is a common feature of accounts that locate reference in the context of communication, such as those found in Kronfeld () and Bach (). For example, Bach () takes referential intentions to be “part of a communicative intention,” that is, a reflexive intention of the Gricean sort.²¹ Reference clearly contributes to communication intentions on the illocutionary and discourse levels, since achievement of communication goals will often turn on the individual involvement of specific aspects of the world. One way to model this contribution is by taking communication intentions to involve referential intentions in some way, as we see with Bach. And further, given that speech is often instrumental in its value, this mereological relationship between communication and referential intentions makes broader practical goals an appropriate extrinsic context for referential intentions as well. This raises the question, what is it to be part of a communication intention? One way to conceive of this relationship is supplied by the theory of planning. Plans can be understood as “intentions writ large,” exhibiting both partiality and a hierarchical, means-end structure (Bratman : ).²² So understood, communication plans provide extrinsic structures within which to locate referential intentions, which could be seen as means toward ends that might be linguistic (e.g., identify something in order to make a claim about it) or nonlinguistic (e.g., identify the proper flour for use in the rye bread). Communication plans can be understood as communication intentions ²¹ See also Kronfeld (: –) for a discussion of the “discourse purpose of referring”. ²² Cf. Pollack (), who argues that a plan can be “usefully seen as a complex mental attitude, one comprising a structured collection of beliefs and intentions” (: ).

OUP CORRECTED PROOF – FINAL, 17/1/2019, SPi

REFERENTIAL INTENTIONS



writ large, inheriting their reflexive character and their requirement of mutual recognition—consider that in performing illocutionary acts, taking speaking turns, or making larger contributions to discourse, you want the recognition of your intentions to be part of the reason why your audience generates their interpretation of your remarks. One way to be part of a communication intention, then, is to be part of a communication plan, specifying a particular means that moves a speaker toward her discourse goals. A number of noteworthy attempts at developing plan-based theories of discourse can be found in the computational linguistics literature (e.g., Grosz and Sidner ; Litman and Allen ). In addition to being used in the philosophical literature as a way of making sense out of discourse (e.g., Grice ; Thomason ), plans have also been used in accounts of reference (e.g., Perry ; King ), as we will see below. Intrinsically, we can think of referential intentions as states with content, and if we take them seriously as intentions, we will want to understand these as cognitive states (Bratman , ; Pollack ). Bratman () takes intentions to be conductcontrolling pro-attitudes (: ), which is to say that they motivate action (unlike representational attitudes such as belief ) and then control action instead of merely influencing it (unlike other pro-attitudes such as desire). If we consider them attitudes, then it is natural to ask toward what are they attitudes? In the case of a typical intention, for example, the intention to get another cup of coffee, it seems natural to think that the attitude is toward a (possibly future) action that you intend to bring about. With referential intentions of the form expressed in (), the attitude would be toward the action of referring (L) to R with E. One can also represent these intentions as propositional attitudes, taking the proposition in question to be specified by a sentence describing the state of the world in which the action is accomplished (Pollack ); with a referential intention, an example would be: ()

I intend that I refer to Liela with ‘she’.

The terms to in () and that in () are correlated with the object of the attitude, that is, with the action or the proposition, respectively. Although any intention to could be rewritten as an intention that in which the action expressed in the former is replaced by a proposition expressing the performance of that action, these are not simply linguistic variants. In general, intend to presents an action to which one is committed and which one will plan around; intend that, by contrast, announces a commitment to make something the case.²³ Further, they do different work in intention-based theories of communication from Grice onward (e.g., Donnellan : ; Grosz and Sidner : ). Intend that can be used, for example, to call attention to attitudes that constitute mutual agency (Bratman ); an account of referential intentions that ²³ Intend that is often a way of speaking that indicates certain elements of the target state of affairs are outside of one’s control. For example, I intend that coffee be served at the workshop I’m running next week, where this reflects my commitment to making sure that coffee be available but also expresses the fact that I will be relying on someone else to perform the relevant actions associated with serving coffee.

OUP CORRECTED PROOF – FINAL, 17/1/2019, SPi



MICHAEL O’ROURKE

takes them to be joint intentions, then, might use this form to cash out the interlocking nature of collaborative reference. Alternatively, intend that might announce a goal to be accomplished by an intention to. And a third difference: intend that can indicate certain extrinsic structural elements of a referential intention attributable to its inclusion in a complex communication act, while intend to would be used to express a more specific contribution to that act (Bach ). That referential intentions can stand in complex relationships with propositions, much like other propositional attitudes, is highlighted by the fact that the de re/de dicto distinction can be applied to them. For example, I can intend de re of the man in line in front of me at the bank to refer to him with ‘you’ in saying ‘You can go over to that teller’; in other circumstances, I might unknowingly refer to the same man when I say, ‘I have a three o’clock meeting with Stanley Philips’, intending de dicto to use the name ‘Stanley Philips’ to refer to the man so named.²⁴ Among those theorists whose analyses of referential intentions involve use of the de re/de dicto distinction are Reimer (), Simchen (b), and King (). We saw above that referential intentions are often taken to be part of communication intentions due to their contribution to the pursuit of communicative goals; this extrinsic point implies for some a correlative intrinsic point, namely, that referential intentions are themselves communication intentions, exhibiting reflexivity and an essential dependence on mutual recognition. Indeed, we can spell referential intentions out in Gricean terms: in using a lexical item to refer to something, you intend that the audience think of that item in part because they realize that they are intended to do so. For example, Bach () understands the distinctively referential intention to be “the intention that one’s audience identify, and take themselves to be intended to identify, a certain item as the referent by means of thinking of it in a certain identifiable way” (: ). This is a reflexive intention of the sort that figures into assemblies of intentions designed to structure interpretation. Of course, this intention is not disconnected from the other communication intentions at work here—they are mutually supporting, providing different cues and clues that can be used to recognize the discourse point of the utterance. Up to this point, I have treated referential intentions as single intentions that can be expressed either in the intends to or the intends that form. But many theorists argue that this is too simplistic by drawing on a variety of problem cases, such as Kaplan’s Carnap/Agnew described in §., or Kripke’s () case of Smith, raking leaves, who is mistaken for Jones (: ). In these cases, reference appears to go awry, and one common diagnosis is that there are multiple referential intentions that come into conflict in these cases. Thus, we can think of reference as depending on the joint

²⁴ It might seem mistaken to classify this consideration under form and not content, given the role of the content specification in making out the de re/de dicto distinction. Here form and content grade off into one another, although the fact that this distinction turns on the nature of the relationship between the state and its content (i.e., whether the referential intention is related to the referent directly or to the proposition) suggests a structural feature of referential intentions.

OUP CORRECTED PROOF – FINAL, 17/1/2019, SPi

REFERENTIAL INTENTIONS



execution of multiple referential intentions that constitute some sort of assembly; when all goes well, these intentions do not conflict—they are executed together in the performance of a referential act. I will look more closely at two different types of intention assemblies that serve as intrinsic analyses of referential intentionality in particular cases. The first has to do with different referential intentions that are related as means to end; the referential act amounts to an attempt to execute these at once, where you intend to refer to X by intending to refer to Y. These assemblies are best conceived as plans. Once again, then, plans crop up as a relevant structure for referential intentions, but whereas before they supplied context for referential intentions, here they supply the structure of the referential intentions themselves. For example, in developing her “modified quasiintentional account” of demonstrative reference, Reimer () distinguishes between primary and secondary intentions. The primary intention concerns the main topic, the object the speaker “has in mind” (: ), while the secondary intention is generated through practical reasoning from the primary intention and relevant beliefs (de re or de dicto) about the object of the primary intention. The secondary intention “will do the work of disambiguating the demonstration” that accompanies the referential act (: ). For example, in the Carnap/Agnew case, Kaplan (the speaker) intended to refer to the picture of Carnap (the primary intention) by referring to the picture in the direction of his ostensive gesture (the secondary intention), given that he (falsely) believed that he could refer to the picture of Carnap by pointing in that direction. This constitutes a referential plan to ostensively identify the picture, with the end given by the primary intention and the means by the derived, secondary intention(s).²⁵ The second type of intention assembly has to do with different referential intentions that are best conceived as hierarchical structures. These intentions are related, to borrow from Dretske (), as structuring elements to triggering elements. Typically, these assemblies involve standing intentions that function like referential policies (Bratman : –), supplying infrastructure for acts of reference, and intentions that trigger specific acts of reference. Kripke’s () distinction between the general intention to use a term according to its conventional meaning and a specific intention to refer to a specific object is an example (: ).²⁶ Both of these intentions are ²⁵ For other plan-based accounts of referential intentions, see Kronfeld (), Perry (), and King (). Perry (), for example, supplies an analysis of Kaplan’s notion of directing intentions in terms of referential plans, which involve intending to say of some object O that it is P by referring to O and predicating P of it (: ). Here we have a schema for building illocutionary-level communication plans of a certain type out of, in part, a referential intention. King () provides a schema for how one might generate the referential intention in the form of a controlling intention as the result of means-end reasoning from an initiating intention that is more general and expresses the communicative goal of the speaker (: ). For a discussion of the role of reasoning in reference and its implications for the complex of referential intentions formed, see Bach (: –). ²⁶ Compare this view to the one developed by Simchen (a), involving primary and secondary intentions. The primary intention secures the success or failure of the act, rather like the specific intention, while the secondary intention could be a descriptive, policy intention “to refer to whatever the person from whom I picked up the name referred”, rather like the general intention (a: ).

OUP CORRECTED PROOF – FINAL, 17/1/2019, SPi



MICHAEL O’ROURKE

operative at the same time, with the former structuring execution of the latter; however, they can come apart, as when you intend to use the name ‘Jones’ to refer to Jones (i.e., the general intention) but you also intend to use it to refer to the person you see, who happens to be Smith (i.e., the specific intention).²⁷ In both cases, a single act of reference can be understood as involving the execution of all the intentions in the assembly, but the nature of the intentions and the logical relationships among them are different. In the case of the plan-like assemblies, the different intentions are both specific, in that they relate to the referential episode in question, and they are sequential in their origin, with the means intention deriving as output from practical reasoning that involves the end intention as input. In the case of hierarchical assemblies, the intentions involved operate at different levels of generality, typically with one functioning like a standing referential policy and another focusing on the referential episode at hand; further, they are hierarchically related, with the more general intention(s) structuring the formation of the more specific intention(s) but not as ends to means.²⁸ I will close this subsection by commenting on another way in which assemblies of intentions have been used in theories of reference. The approaches that I have examined privilege the speaker, even though some insist that the content of the speaker’s intentions essentially involve the listener. Rather than interpreting the relationship between reference and referential intentions from the perspective of the speaker, one might treat it as the mutual product of the coordination of speaker and listener. This approach is adopted by Clark and Wilkes-Gibbs (), who develop a “conversational model” of speaker reference as “a collaborative process requiring actions by both speakers and interlocutors” (: ).²⁹ On this approach, then, referential intentions could be joint intentions or ‘We’-intentions (Searle ; Tuomela ). One can also embed referential intentions so understood in the planning model, treating the plans themselves as joint intentions (Stone ).

.. The content of referential intentions The first point to make about the content of referential intentions concerns the implications of the structural distinction between intend to and intend that. Referential intentions understood as intentions to have non-propositional content—the content specifier is a verb phrase describing a referential act; however, we can understand them ²⁷ Another example worth noting is Donnellan’s () distinction between audience-directed and truth-conditional referential intentions, mentioned above. The latter can be specified in light of the particular referential episode, but nevertheless has the feel of a referential policy, that is, a standing intention that structures referential discourse. For discussion, see O’Rourke (). ²⁸ In speaking of generality here, I mean that a single structuring intention applies across a range of referential acts, whereas a single triggering intention concerns the performance of a specific referential act. ²⁹ See also Tomasello .

OUP CORRECTED PROOF – FINAL, 17/1/2019, SPi

REFERENTIAL INTENTIONS

S

E



A

R : produces : connected to : refers to

: received by : identifies

FIGURE . A model of relationships of a referential episode involving the speaker S, the expression E, the audience A, and the referent, R.

as propositional if we understand them as intentions that. A full account of the content of referential intentions on offer in the literature should address its specification in these two modes. For our purposes in this section, I consider a higher-level classification scheme based on a simple model representing the typical situation in referential discourse, introduced in Figure ., that allows us to concentrate on aspects of content that do not vary appreciably between the two modes. The model features the various relations promoted in different theories of referential intentions, including the fix speaker S has (or might have) on the referent R, the production of an expression E that refers to R, the reception of E by the audience A, and the identification of R by A as a result of receiving E. First, consider the relationship between the speaker and the referent. In all of the proposals, the referent R is involved, since these are referential intentions and so the connection with the referent is a defining feature of this type of speaker intention.³⁰ But how S is connected to R can vary. Most accounts take referential intentions to be singular in content, involving the object in a directly relational, de re way (e.g., Donnellan ; Kripke ; Kaplan a; Reimer ; Simchen a; King ). Simchen a, for example, makes being “c-related” to the intended referent a “cognitive prerequisite” of having a referential intention (a: –). The c-relation is the “causal-historical peg . . . upon which the subject may hang subsequent cognitions” (a: ), and it underwrites a singular relation between speaker and referent. But not all theories of reference that rely on referential intentions understand the content of these intentions in singular or relational terms. There are those, like Kronfeld (), that are explicitly general, building on the descriptive theory of reference. For Kronfeld, “reference is entirely a matter of associating a mental state with descriptive content” (: ), and

³⁰ Of course, we use referring terms in situations where there is no R. Empty cases such as these come in two types. The first type is when the speaker is mistaken about the existence of a referent; in such a case, what matters for the content of the referential intention is what S takes to be the referent. The second type concerns terms like ‘Santa Claus’, which a speaker will use without confusion about its ontic implications. From the point of view of referential intentions, these can be understood as ersatz cases in which the term is used as if it were referential.

OUP CORRECTED PROOF – FINAL, 17/1/2019, SPi



MICHAEL O’ROURKE

referential intentions understood as embedding modes of presentation of the intended referent are a principal vehicle of this association (: ).³¹ One connection between S and R that has figured centrally into the discussion of the content of referential intentions concerns the cognitively mediated relationship between speaker and referent that Wettstein () has called the “discriminating cognitive fix” (: ). This is represented in the literature in a number of ways. Two are prominent: having in mind and aiming. These metaphors are suggestive, and are used to call attention to a connection between a thinker/speaker and an item about which they are thinking/speaking. Having in mind is featured in Donnellan’s () discussion of the referential/attributive distinction, and also figures into various other treatments of reference.³² While there is no canonical analysis of this relation, it is clear that it is not (merely) belief, even occurrent belief. To have something in mind when uttering a referring term indicates that the speaker is differentially related to the item in question and can track it for the purposes of discourse. This does not entail a robust know who/what, but instead indicates that independent resources exist that can allow the speaker to isolate and think or speak about the item. Tracking the item by having it in mind is a way of speaking that resonates with aiming. One can think about referring to an item as being analogous in certain ways to aiming at the item with a term, or perhaps aiming the audience at the item. Kaplan (b), for example, speaks of the directing intention being “aimed at a perceived object” (b: )—one way to interpret directing here is in terms of directing one’s audience through the use of a referring term.³³ If you have an audience, the idea is to call their attention to the item in part by making them aware of your intention to refer to it, which is done via use of the term in question. The referent is the target of your referential utterance, and you select your referring term as a mechanism for aiming your audience’s attention at that target. The relevance of aiming as a metaphor for reference highlights other contentrelevant elements in the model of reference above, in particular, the expression E and the audience A. Those theorists who see referential intentions as part of communication intentions will take their content to involve A essentially. This is a common way of thinking about referential intentions—these intentions aim to use E to introduce an item into the discourse contribution for A to identify and about which S wishes to say something. King (), for example, builds a “responsibility to be understandable” by the “ideal” audience into his Coordination Account of demonstrative reference, ³¹ For another account that discusses referential intentions understood in this more general, descriptive way, see Bach (/). For one that combines causal-historical (i.e., singular) aspects with general, see Evans (/). ³² Here is a short list of works discussing this concept in connection with reference, either favorably or critically: Kripke (); Wettstein (); Kaplan (a, b); Reimer (); Bach (); O’Rourke (); Bozickovic (); Mount (); and Capuano (). See also the contents of Almog and Leonardi (), especially the contributions from Wettstein and Kaplan. ³³ For other discussions of this metaphor as applied to reference, see Donnellan (, ), Kronfeld (), and O’Rourke ().

OUP CORRECTED PROOF – FINAL, 17/1/2019, SPi

REFERENTIAL INTENTIONS



making the audience an essential part of the analysis (: ). But while it is a common way of thinking about them, it is not necessary. For example, Perry () rejects the idea that an account of demonstrative reference should be built on some “forensic” sense of speaker responsibility; rather, it should be tied to “the actual structure of intentions” involved in referring (: ). For him, “reference is easy”—as he sees it, “if you can think of it you can form an intention to refer to it, and can do so” (: ). The difference between these approaches corresponds to whether you take referential intentions to underwrite the act of referring A to R, or rather merely the act of referring to R, respectively. The former requires content that includes all four elements in the model while the latter requires content that includes just S, E, and R.

.. The function of referential intentions Form and content combine to supply a way of thinking about what referential intentions are and how they relate to other states and elements, but they don’t address the ways in which these intentions function. If we take them seriously as intentions, referential intentions are formed, maintained, and executed in the course of producing referential utterances. I will examine their function from three perspectives: constraints on their formation, their role in reference determination and recognition, and their influence on speech behavior. First, it is important to recognize that, in general, referential intentions cannot be formed willy nilly—this is what allows Donnellan () to answer MacKay’s () charge that his is a ‘Humpty Dumpty’ conception of the referential use of definite descriptions. There are constraints and conditions that govern the formation of intentions, and this applies just as well to referential intentions. Among the constraints and conditions discussed in the literature are these:

· Expectation Constraint: One can form an intention only if one can expect to · ·

successfully execute it; in the case of reference, “[w]hether he can form that intention . . . may depend upon what expectations he has about his audience and their ability to grasp his intention” (Donnellan : ).³⁴ Epistemic Constraint: An intention should be consistent with your beliefs and what you know, or at least you should regard it as consistent (Neale : ). Incentive Condition: If the only way to accomplish a goal is to A, then you may under certain very unusual conditions form the intention to A even if it is very

³⁴ Compare this to what we might call the ‘Possibility Constraint’ in Neale (: ), expressed as follows: “A cannot (intend to) refer to some particular person α by uttering some expression X on a given occasion if he believes it is impossible for his audience B (or at least any rational, reasonably well-informed interpreter in B’s shoes) to construe him as referring to α” (: ). This version of the Expectation Constraint does not admit of exceptions, unlike the guarded one expressed in Donnellan ().

OUP CORRECTED PROOF – FINAL, 17/1/2019, SPi



MICHAEL O’ROURKE

unlikely; for example, “[a] man in the water from a sinking ship might move his arms with the intention of swimming a hundred miles to shore, if that is the only hope, even though he has no rational expectation of doing it” (Donnellan : , n. ).³⁵ The audience is introduced as a part of the Expectation Constraint, but this constraint is operative even in the absence of an external audience. This is due to the fact that one is always one’s own audience, and the ‘talking to yourself ’ option makes it possible to form referential intentions and perform referential acts without an external audience present.³⁶ Second, referential intentions, when formed and executed, underwrite a relationship between the uttered expression and the referent that can be viewed from different perspectives. From the perspective of the speaker, the intention can be taken to be partially determinative of reference. Whether this is part of the semantic story (see Donnellan ) or not (see Bach ), the intention can be defended as establishing a relationship between speaker and referent that is involved in the total significance of the episode (e.g., Kripke ; Bach ). The flip side of this is recognition of the referential intention, which is necessary if the referential act is to be audience-involving and successful. The determination role can be seen as a part of the metaphysics of referential intentions, and the recognition role a part of their epistemology (Neale : , ). Finally, consider the influence referential intentions have on speech behavior. Referential intentions, qua intentions, can structure reasoning and deliberation during speech. As we have seen, planning accounts of referential intentions take these intentions to be inputs into practical reasoning, issuing in further intentions that provide means to accomplishing referential ends (e.g., Reimer ; King ). They can also operate at a higher level, contributing to broader discourse purposes by guiding the selection of terms and phrases needed to achieve one’s communicative goals (e.g., Grosz and Sidner ). It is instructive to think here of writing, and in particular wordsmithing, which highlights this role even at the lexical level.

. T W O

CONCERNS

.................................................................................................................................. While reliance on intentions in theories of reference is fairly common, there are of course criticisms that can be leveled at their theoretical employment. Rather than ³⁵ This is introduced here not as a constraint on intention formation but rather as a condition that, if satisfied, would allow intention formation in violation of the Expectation Constraint. Allowing for this might explain the guarded way in which the Expectation Constraint is expressed in Donnellan (). In the event that this is deemed unallowable, the kind of intention that one might form in the condition described by Donnellan might better be identified as an intention to try to swim  miles. ³⁶ This could be taken to shore up Perry’s () suggestion that reference is ‘easy’, even in the absence of an audience. For discussion of the audience-less case, see Grice (: –) and King (: –).

OUP CORRECTED PROOF – FINAL, 17/1/2019, SPi

REFERENTIAL INTENTIONS



consider criticisms of specific implementations of referential intentions (e.g., Reimer’s  criticism of Kaplan b), I will close by briefly considering two general concerns: explanatory failure related to the content of the intentions, and what I will call hyperintentionalism. I regard these as challenges that must be addressed by any adequate account of reference that is grounded in referential intentions.

.. Referential intentions and explanatory failure First, if you take referential intentions to determine reference, there is the threat of explanatory failure due to explanatory deferral in the form of regress or circularity. On the referential intentions approach, if you refer to A, you execute an intention to refer to A; you execute an intention to refer to A if and only if you execute an intention the content of which is about A; the content of your intention to refer is about A just in case _______. How are we to fill in the blank? There are a number of alternatives. One involves focusing on intentions as products of practical reasoning involving beliefs and desires, and arguing that they inherit their intentionality from the states that gave rise to them. Pursuing this line threatens an unacceptable type of representational regress, where we track the referent from one cognitive state to another. Another would be to stop the slide with the content of the intention itself, arguing that it contains an element that stands in an appropriate referential connection to A. Filling this out in a way that avoids circularity is a key success condition for theories of reference based in referential intentions. The metaphors of having in mind and aiming do not help here, since they fail to provide any candidate for the content of the intention that would be independently associated with the referent in question.³⁷ The two leading candidates are the usual semantic suspects: a relational condition, which selects A as standing in a non-denotational relation with the content element (e.g., a causal-historical condition, such as in Donnellan  or Kripke , or perhaps a perceptual relation as in Kaplan b),³⁸ or a descriptive condition, which selects A by virtue of satisfaction or denotation (e.g., Kronfeld ).³⁹ ³⁷ Bach () puts it this way: “Referential intentions determine reference, inasmuch as they are the component of communicative intentions that concerns what an utterance is about. Now to say that referential intentions determine reference may seem to suggest that they succeed by magic or are somehow self-fulfilling. However, such a suggestion misconstrues their role. You cannot utter any old thing and gesture in any old way and expect to be taken to be referring to whatever you have in mind. You do not say something and then, as though by an inner decree (an intention), determine what you are using the expression to refer to. You do not just have something ‘in mind’ and hope your audience is a good mindreader” (: ). ³⁸ See Simchen (a), who develops an account of name reference in terms of referential intentions that involve c-relations and an embedded notion of m-reference, which is tied to morphemes used in the production of the name. Here the c-relations support m-reference, which is used to analyze reference. This is done to avoid giving rise to, in his words, a “peculiar form of self-generation” (a: ). ³⁹ As I have noted, Evans (/) combines descriptive conditions and a causal relation into a relational theory.

OUP CORRECTED PROOF – FINAL, 17/1/2019, SPi



MICHAEL O’ROURKE

The suggestion, then, is that spelling out the content of a referential intention in terms of one of these conditions provides an independent way of fixing the referent of the term under consideration. But it is natural at this point to wonder why we don’t just use this condition as an analysis of reference and dispense with referential intentions altogether. Why doesn’t reliance on referential intentions amount to pushing the explanation from the language back into the head, without really providing any sort of independent analysis? For instance, in the case of an indexical expression, why not just treat the referential intention as scaffolding that serves up the determining condition without itself being a part of the explanation?

.. Hyperintentionalism I use ‘hyperintentionalism’ to refer to the unreasonable proliferation of intentions in the service of referential discourse. Grice () saw this threat, noting that “some questions may be raised about my use, fairly free, of such words as ‘intention’ . . . I must disclaim any intention of peopling all our talking life with armies of complicated psychological occurrences” (: ). But not all intention-based theories of linguistic significance come with such a disclaimer, and that includes various theories of referential intentions. The temptation to explain speaker control over speech with intentions, motivated by the two principles supplied in §., could result in their profligate employment. Numerous intentions are posited to underwrite various aspects of reference in order to explain pathological cases (e.g., cases of conflicting intentions— see King ), even at the level of the morpheme (e.g., Simchen a). We could introduce intentions to control other aspects of speech situations, like predicates (Neale ) and phonology (Bromberger and Halle ). If we adopt a planning approach to reference, that adds another dimension of complexity just for the referential parts of an utterance—if there are plans for the other parts, then the intentions executed to produce the utterance themselves constitute only the tip of the intentional iceberg. It is natural to take these philosophical theories to be contributions to theoretical psychology, formulating empirical hypotheses about the use of referential language.⁴⁰ If we go down this road, the result is a large and potentially unwieldy structure of intentions governing many if not all aspects of any given utterance, and at several levels at once. For example, the rather mundane sentence, ‘You and you need to see the principal right now,’ would be analyzed against a highly complex background of ⁴⁰ One could also take these to be instances of a more philosophical project that aims to provide formal constraints on the concept of reference, thereby locating it in a network of concepts. Here the intentions could represent modal constraints identified counterfactually by the speaker (as with the Speaker Control Principle), or epistemic constraints recoverable as part of abductive reasoning. Of course, a conceptual project of this sort should be understood to put pressure on an empirical project, but the goal of philosophical work need not be the generation of testable hypotheses. See O’Rourke (: –). Thanks to Bruce Glymour for discussion of this point.

OUP CORRECTED PROOF – FINAL, 17/1/2019, SPi

REFERENTIAL INTENTIONS



intentions. Whether or not this is true is to some extent an empirical hypothesis, and it is not exactly the safe play to bet against the brain’s ability to support complex activity. It would be a mistake to think that taking referential intentions seriously implies that they must be consciously available when we employ them—they could function tacitly and still determine reference. Still, it seems unlikely on its face that we form and execute so many intentions whenever we speak—the ability exhibited by young speakers of the language would seem to defy the suggestion that language use requires cognitive overhead this substantial. Further, there are examples that challenge the presence of a complex fabric of intentions underlying every referential utterance. For example, consider the prevalence of formulaic language in both writing and conversation, that is, larger, multiword conventionalized linguistic items, such as compound words, phrases, and clauses. Formulaic language is likely not produced from the ground up (Wray ), yet it can contain referring expressions (e.g., ‘you can say that again’, ‘as we saw earlier’). In these cases, it is plausible to think that intentions may be operative at the level of the formulae but not below that level. Fortunately, one can be congenial toward referential intentions and resist hyperintentionalism. In particular, we can follow Bratman (), who introduces the notion of the motivational potential of an intention as follows: A is in the motivational potential of my intention to B, given my desires and beliefs, just in case it is possible for me intentionally to A in the course of executing my intention to B . . . [W]e need not suppose that if A is in the motivational potential of an intention of mine, then I intend to A. Bratman : –

Applying this concept to our situation, the idea would be that one might have a higherlevel communication intention to convey some point to an audience that has within its motivational potential various lower-level lexical performances, for example, the production of referring terms. If so, then one can intentionally engage in referring without the need to posit intentions to represent speaker control over the various moving parts of reference. This suggestion has the dual virtues of retaining a commitment to speaker control while keeping the ‘armies’ of referential intentions small. It is also consistent with the idea that in certain cases—for example, wordsmithing a question you hope to ask of a speaker during the discussion session after a talk—you really do intend to utter the referring terms you utter.

. C O N C L U S I O N

.................................................................................................................................. In this chapter I have sought to take referential intentions seriously by conducting a wide-ranging survey of their use in semantic and pragmatic accounts of reference and referring. Referential intentions are a common part of explanations of reference—their appeal is rooted in a commitment to control by the speaker over her meaning, with

OUP CORRECTED PROOF – FINAL, 17/1/2019, SPi



MICHAEL O’ROURKE

concomitant deference to the speaker in cases where it is unclear what she meant. Referential intentions do not operate in isolation from other practical concerns—they help structure utterance behavior and contribute to the pursuit of a speaker’s broader practical goals. Their paradigmatic contributions concern the production of subsentential phrases, especially for intentional or discretionary indexicals but also for proper names, definite descriptions, and a range of other linguistic items that depend for their significance on how things stand with a particular item or items in the world. As intentions, they can be analyzed in terms of their form, content, and function; while there are many theoretical options considered in accounts of referential intentions, planning approaches stand out as especially popular because they underwrite important intrinsic and extrinsic characteristics. While there are utterances (e.g., delivering a carefully crafted press release) where referential intentions are explicitly formed and executed, most are less engineered and more spontaneous, accommodating referential intentions only as tacit semantic or pragmatic mechanisms. Typically, then, referential intentions are creatures of theory, posited as parts of semantic or pragmatic explanations that must deliver on their explanatory promise without exploding in number beyond what can be supported by psychological reality.

ACKNOWLEDGMENTS I owe a debt of gratitude to many people for their willingness to provide me with guidance in thinking about the issues discussed in this chapter. Included in this group are Michael Bratman, Eros Corazza, Robin Jeshion, John Perry, Liela Rotschy, and Howard Wettstein. I am especially grateful to Barbara Abbott, Bruce Glymour, Jeanette Gundel, and Graham Hubbs for assistance with this chapter.

OUP CORRECTED PROOF – FINAL, 17/1/2019, SPi

CHAPTER



......................................................................................................................

JOINT REFERENCE ...................................................................................................................... ANNE BEZUIDENHOUT

. I N T R O D U C T I O N

.................................................................................................................................. I will argue that in everyday face-to-face conversations, referring is a joint activity. I begin by adopting two crucial assumptions. Firstly, I adopt the view that it is not words that refer but rather people who refer by using words in appropriate contexts. This is a view that was articulated by Strawson () in his famous critique of Russell’s theory of denoting. On this view, referring is a kind of linguistic action. This does not mean that we cannot talk about words as referring (and indeed, I will frequently talk of referring expressions in this chapter). It simply means that referring is primarily an action performed by people and only secondarily something words do. Words inherit this property from the fact that people use (or could use) them in the performance of certain types of actions. This is similar to the sort of transference that happens in our talk about tools or instruments. We say the jackhammer hammers, although strictly it is the person using the jackhammer in a certain kind of way that is doing the hammering. Similarly, we say our smart phone has alerted us to bad weather conditions when in fact it was people working for the National Weather Service who programmed these alerts and who are monitoring the weather (a process which in its turn has perhaps been automated) in order to determine when and where to send out alerts. Secondly, it is possible to perform a purposive action jointly. That is to say, it is possible for two or more agents in coordination with each other to do their part to bring about a joint accomplishment that would not be brought about by any of the individuals acting alone. Since linguistic actions are just one specialized kind of purposive action, anything that is true of purposive actions in general should be true of linguistic actions too. Putting these two assumptions together, I will argue that—in everyday conversational settings—actions of referring are joint actions in which two or more interlocutors do their part to produce an outcome that would not have resulted from the efforts of any of the individuals acting alone. I will say no more here in defense of the view that referring is primarily an action performed by an agent (or agents). However, in §.,

OUP CORRECTED PROOF – FINAL, 17/1/2019, SPi



ANNE BEZUIDENHOUT

I say a little more in defense of the view that there are joint actions and what they entail. I turn, in §., to defending the notion of joint referring. There are weaker and stronger views that could be defended. On the stronger view, all referring is joint referring. On the weaker view, jointly referring to things is a species of referring actions that includes cases in which individuals refer to things. I will be arguing for the stronger view and will show how cases that we might describe as individuals referring to things are parasitic on cases of two or more individuals jointly referring to things. In §., I extend my view to cases of category reference.

. J O I N T

ACTION

.................................................................................................................................. Suppose you want to move a heavy upright piano up some stairs. Unless you have superpowers, you will need to recruit some people to help you achieve your aim. You determine, based on the weight and size of the piano, that you will need three helpers. You find three such willing and able people. They may be people who share your desire to move the piano up the stairs. But perhaps they don’t care about the piano per se but only about the money you’ve promised to pay them to help. Or maybe you are a slave owner and the people don’t have much choice in the matter, but they prefer to help rather than face the consequences of not helping. You may at this point strategize amongst yourselves as to how to proceed. Or perhaps you all just take up a position around the piano and then each of you begins to lift at your spot and starts to move in the direction of the stairs. If you are an experienced work crew who has done this sort of job hundreds of times before, you may just get right to it, figuring that the others on the crew will know what their roles are. However, whether or not you formulated a plan in advance as to who would do what when, it is likely that you will have to communicate to each other at various stages in the process so that appropriate adjustments can be made. Perhaps something unanticipated happens. One of the crew members failed to put his gloves on at the start and now needs everyone to stop and brace the piano so he can release his hold and put his gloves on. Or perhaps the turn in the stairwell is narrower than anticipated and new angles will have to be plotted. But eventually the task will be done and the piano will be up the stairs and in its new resting place. This moving of the piano—a temporally extended event that began at some time with the piano and the movers at a point below the stairs and ended at some later time with the piano in its location at the top of the stairs—is a joint action, because none of the series of individual actions carried out by any one of the moving crew would by itself constitute the moving of a piano up the stairs. Of course it is conceivable that some subset of the individual actions that constitute this particular moving of the piano up the stairs would have been sufficient for the action. Perhaps one of the crew members was not really pulling his weight and he simply moved about as though he were carrying a crucial part of the load. Even if this individual had been dropped from

OUP CORRECTED PROOF – FINAL, 17/1/2019, SPi

JOINT REFERENCE



the crew, the piano would have been moved in much the same way, with the others basically acting just as they did. However, as the assumption was that you could not move the piano by yourself, at least some of the crew members’ individual actions had to be added to yours to achieve the desired result. So, what is needed for a joint action is as follows: (i) There is a principal, P, who desires outcome O. The principal could be an individual or a group of individuals. If it is a group, they must have a group/joint desire for O. Clearly, we would need to step back and explain this concept of a group desire, or else our explanation of joint action threatens to be circular. So, for the time being, I will assume the principal is an individual. (ii) P is unable to achieve O by solo means. Of course, sometimes there are things that you could do by yourself but still you ask for help. You could move the TV set up the stairs by yourself, but it would be much easier if someone lent you a hand. Here I will focus on cases of joint action where P needs help or O cannot be achieved. (iii) P is able to recruit the requisite number of helpers, H₁, . . . , Hk who are willing and able to do their parts to bring about O. The helpers may not desire O for its own sake, but they must have the desire to help P achieve O. As already mentioned in the example of the piano moving, their desire to help may be purely instrumental. They may have been promised pay for the job or they are in a subordinate power relation to P, or whatever. And of course the helpers must be able to do their part. Your frail grandmother may desire to help you move the piano up the stairs, but she isn’t able to do her part. (iv) P and H₁, . . . , Hk must each perform a series of individual actions, , , . . . , , which are coordinated with each other at the respective times t₁, . . . , tn such that at each time ti, the set of coordinated individual actions {Pai, H₁ai, . . . , Hkai} is sufficient to constitute a complete segment of the joint action (e.g., it constitutes the piano’s moving up one step). See Figure .. This condition is a simplification, because it is possible that there are stages where one or more of the agents who are collaborating needs to perform an individual action or series of actions while the rest maintain a holding pattern. I will construe an agent Aj’s remaining in a constant state during a certain temporal stage ti as a kind of action—what Aj is doing during ti is holding steady. So, we can still say that for every ti, there is an action Ajai performed by Aj at that time and that this action must be coordinated with those of all the other collaborators’ actions at that time.¹ Another ¹ Vendler () made a distinction between states, actions, accomplishments, and achievements. In some contexts it is crucial to make these distinctions. Strictly speaking, moving a piano to a place

OUP CORRECTED PROOF – FINAL, 17/1/2019, SPi



ANNE BEZUIDENHOUT










→ → Flow of time → → FIGURE

. Schematic representation of a joint action.

kind of limiting case is one where the principal initiates the action but then delegates the execution to the helpers (e.g., you hire a set of movers to move your piano upstairs but you just stand on the sidelines while they do all the heavy lifting, your role being only to specify where you want the piano to end up). I do not offer the conditions (i)–(iv) articulated above as a set of necessary and sufficient conditions for joint action, as I am sure that some clever philosopher can imagine a scenario in which (i)–(iv) are satisfied and yet intuitively we would deny that a joint action has been performed. However, I do think that these conditions are sufficient to explain some central cases of joint action and they will enable me to make some headway in explaining how joint reference works. Note also that we need to say some more about what is involved in the coordination by the collaborators of their individual actions at a time. We need to be careful here not to smuggle in concepts that beg the question. We could just say that the collaborators have a group intention to perform the set of coordinated individual actions {Pai, H₁ai, . . . , Hkai} at each time ti. However, we would then need to step back and unpack the notion of a group intention. For my purposes in this chapter, I will say that a set of agents is collaborating at a time if each has an individual intention to perform an individual action at that time on the expectation that all the others will do their part and would not intend to perform that individual action unless he/she believed that the others had relevantly similar intentions. For example, you would not intend to lift up your corner of the piano unless you believed that the other crew members intended to do their bit because if you acted alone either you would not succeed in budging the piano at all or you would tip it over. In addition to each person’s having an individual intention to do their part, individuals must intend that their first-order intentions be mutually manifest to the others. For upstairs is an accomplishment, since it has a specified endpoint (the point at which the piano will be at its resting place upstairs), whereas moving a piano is simply an action, because no beginning or endpoints are specified. States are different from either of these. I do not think these distinctions affect my main line of argument in this chapter. However, since I have built the notion of an objective into my account of a joint action (see condition (i) above), perhaps the title of this section should be changed from ‘Joint action’ to ‘Joint accomplishment’.

OUP CORRECTED PROOF – FINAL, 17/1/2019, SPi

JOINT REFERENCE



these intentions to be mutually manifest it is not necessary for anyone to consciously entertain the thought that others have these intentions but only that they are able to do so should it prove necessary to check these intentions. Each of the sets of coordinated individual actions at some time ti (represented by the gray bar in Figure .) could itself be treated as a joint action—a temporal stage of a more complex joint action. Clark (: –) argues that each joint action consists of a basic joint activity and a coordinating joint activity. The coordinating joint activity consists of turns-at-talk or other acts of communication involving material signals that help align the basic actions of the collaborators. For example, suppose Ann and Bob are assembling a TV stand. A basic joint action might consist of Ann’s holding the top piece steady while Bob inserts and tightens a screw. A coordinating joint activity might consist of a pair of turns-at-talk where Ann suggests that she can hold the top piece while Bob fastens the screws and Bob verbally signals agreement (e.g., saying ‘Yeah’ or ‘Good idea’). This is not very explanatory, as Clark simply helps himself to the notion of a joint activity (either basic or coordinating) in his account of what constitutes a joint activity. Note that what Clark calls a basic joint action is really just a set of individual actions. So in order to avoid the appearance of circularity in the account, we can simply talk of a set of basic individual actions and then say on Clark’s behalf that what turns them into a joint activity is the coordinating joint activity. This is still problematic, as we are appealing to the notion of a coordinating joint activity to explain the notion of joint activity. We need now to unpack Clark’s notion of a coordinating joint activity. Clark does in fact seem to think it can be unpacked by appeal to the notion of communicative acts of a certain kind. He writes: “It takes coordination for people to do things together, no matter how simple, and it takes communication to achieve coordination.” (Clark, : ) The kind of communication Clark has in mind is the already mentioned kind that consists of basic turns-at-talk, such as questionanswer or suggestion-response pairs, or cases of non-verbal communication where the communicator uses material signals (gestures, eye gaze, objects-at-hand, etc.) to direct addressees to objects, locations, people, etc., or where the communicator places objects or people in what I will call ‘action spaces’—places that are conventionally associated with, or which conventionally call for, special kinds of routines, such as placing groceries on the checkout counter for the cashier to ring up. This account that appeals to certain types of (verbal or non-verbal) communicative acts may help Clark to escape the circularity charge. I will remain neutral on this matter. However, this is not an account that I will be able to help myself to when it comes to giving an account of joint reference. This is because the activity of jointly referring to something is an example of the sort of communicative act that Clark helps himself to in his account of joint coordination. For this reason, I will stick to the fairly anodyne account of collaboration/coordination mentioned above, which appeals to the individual intentions of the principal and the helpers to do certain things on the expectation that the relevant others have similar intentions and to their individual intentions that these first-order intentions be mutually manifest. In the joint reference

OUP CORRECTED PROOF – FINAL, 17/1/2019, SPi



ANNE BEZUIDENHOUT

case, this will involve the interlocutors in a particular conversational context having such (mutually manifest) individual intentions. I have been assuming that joint actions are a species of purposive action. However, I am not assuming that all purposive actions involve conscious deliberation and planning on the part of the agents involved. All the beliefs, desires, intentions, and plans involved could be carried out in a fairly automated way, especially if the actions are ones that have been performed numerous times before. I already mentioned the case of the moving crew who have moved countless pianos upstairs before and so don’t need to strategize much at all at the start and simply get down to the task, making adjustments along the way as called for by the particular situation. These adjustments too need not be consciously planned. The crew members may have learned to read each other’s behavior so well that no verbal communication is needed to make adjustments to their own intentions and actions to accommodate one another. Communication between the group members may even be a matter of adjustments that are part of a feedback loop between the individuals and the object(s) deployed and/or acted on that do not run through any sort of mental representation of the situation by these people. For example, you slightly release the tension in your muscles and the piano dips slightly on your end. In response to this slight movement of the piano, my muscles tighten and my grip on my end of the piano becomes slightly firmer and the angle of my grip is slightly adjusted. This sort of motor response doesn’t require thinking on either of our parts. There is no need for a sense-think-act cycle; the cycle can be a direct sense-act one. See Clark (: ff ). On the other hand, when the collaborators are all relative novices at the task, a fair amount of preplanning may be needed and there may have to be a lot of talk at each stage to make sure that everyone is still on task. For example, you and your spouse may cooperate to assemble a chest of drawers that you bought at IKEA based on the set of assembly instructions included in the kit. Since neither of you is a carpenter or very familiar with furniture assembly, this may take quite a bit of negotiation. Clark () reports on conversational data from pairs of people who were given tasks such as assembling a TV stand or building a Lego model. The pairs typically follow an overall action plan that matches the ones followed by people acting alone. What is interesting about the joint action case is the negotiation that must go on at each stage along the way to ensure that the agents are appropriately aligned with each other. A related discussion of situations in which people work together can be found in Clark and Henetz (). Other relevant research involves the recording and analysis of the conversations of pairs of individuals who are collaborating on tasks that collectively are known as referential communication tasks, such as the map task studied by Brown () or the game board task used by Schober (). I will say more about this research on referential communication tasks in §.. So far I have given examples of joint actions that are not too complicated and where it is relatively easy to think that we could, in principle, lay out each of the individual actions required by each of the agents involved at each of the stages of the process. However, many joint actions are ones where it is highly unlikely that we could

OUP CORRECTED PROOF – FINAL, 17/1/2019, SPi

JOINT REFERENCE



explicitly lay things out in this way. (Frankly, I think it is an illusion that we could do so in the piano moving case either, partly for the reason that many aspects of the task could involve the sorts of feedback loops described above that bypass conscious and unconscious decision-making mechanisms.) For example, Hutchins () has a fascinating ethnographic and social psychological account of what it takes for a navigation team (in collaboration with other onboard personnel) to perform the joint action of bringing a , plus ton US Navy amphibious helicopter transport into port. Even with the aid of sophisticated technology, such as satellite navigation systems, radar equipment, and other electronic navigation aids, this is clearly not a task that can be carried out by one person. Hutchins gives a detailed description of the distributed cognitive processes involved in pilotage and shows that these processes depend on an elaborate sociocultural and institutional background. Aspects of his account of navigation team performances will be woven into the account of joint reference laid out in §.. It should be noted that, like Clark (), Hutchins assumes that communication plays a crucial role in coordinating behavior. Moreover, the theory of communication he relies on makes use of the idea of joint interpretation (: ) and of the idea that meanings “are negotiated by the participants in the context of their understandings of the activities underway” (: ). Hutchins argues that the “illusion of meaning in the message is a hard-won social and cultural accomplishment” (: –) and that the illusion can only really be sustained in cases of communication where there are very strong and stable sets of constraints on the interlocutors’ expectations of one another’s behavior. In my account of joint referring in the following section, I cannot simply help myself to the idea of communication as joint interpretation. This must fall out as a conclusion of my arguments.

. J O I N T

REFERRING

..................................................................................................................................

.. Referential communication tasks If referring is a kind of (purposive) activity, and if at least some (purposive) actions can be performed jointly, we can ask whether referring is one of the kinds of activities that is (or could be) performed jointly. That is, is referring something that would not occur unless two or more individuals co-operated with each other in some way? It seems clear that if there are cases of two or more people jointly referring to something, such joint referring will never be as complex an event as the joint action of bringing an amphibious helicopter transport into port described by Hutchins (). Nevertheless, linguistic practices are sophisticated sociocultural practices and so referring takes place against a sociocultural and institutional background, just as the joint action of pilotage does. Thus aspects of Hutchins’s account of the joint activity of pilotage can potentially be used in an account of joint referring.

OUP CORRECTED PROOF – FINAL, 17/1/2019, SPi



ANNE BEZUIDENHOUT

Just as Hutchins chose a navigation team engaged in pilotage to make his case for the idea of a distributed cognitive system that has cognitive properties different from those of any individual and capable of a joint activity that is larger than the sum of the individual actions performed by the team members, I will begin with an examination of the behavior of dyads engaged in a map task, which is one species of what have been called referential communication tasks. See Yule () for a summary of the types of tasks that have been used in this referential communication research.² The behavior of such dyads is not ‘cognition in the wild’, unlike the activity of pilotage analyzed by Hutchins. The map task conversations that have been recorded and analyzed were carried out in an experimental setting with materials designed for the purpose by an experimenter. However, the situations are naturalistic enough to think that map task conversations in which interlocutors try to establish mutual understandings about map locations and map symbols are very similar to naturally occurring conversations in which interlocutors are trying to talk about places or things in the world (which could of course also include talk about maps, map features, and map locations). In classic versions of the map task, each member of a pair of participants has a schematic map in front of them, not visible to the other participant. Each map is comprised of an outline and some number of labeled features. Most features are common to the two maps, but some features are not shared, a fact known to the participants. One of the maps has a route drawn in, while the other does not. The task is for the participant without the route (the follower) to draw in a route on the basis of a discussion with the other participant (the guide). The pair will be successful if at the end the two maps have identical routes drawn in. For example, in the HCRC Map Task Corpus, which contains a total of about eighteen hours of spontaneous speech that was recorded from  two-person conversations, involving sixty-four different speakers, the maps that were used to elicit data contained a dozen labeled features (e.g., a whitewashed cottage, a chestnut tree, Green Bay, etc.). See Anderson et al. (). These map tasks are harder than they might seem. One reason is that participants cannot see each other’s maps and so pointing gestures, eye gaze, and so on cannot be

² I should note that I do not share Yule’s understanding of the term ‘referential communication’. He distinguishes two types of language that he calls transactional and interactional (Yule : –). The former, he says, is used in situations, such as in school or work or other institutional settings, where information needs to be exchanged and formality and precision are called for in one’s messages. The latter is used in informal settings with familiars where one chats for no purpose other than to reinforce social bonds and where inexact and vague language use is rarely challenged. Yule says that referential communication is a kind of transactional use of language. I do not doubt that there is a continuum of communicative situations of the sort Yule posits (from informal chat sessions at the one extreme and formal information exchanges at the other). However, what I am calling referential communication occurs in all these situations, because even in a situation where you are simply chatting with familiars, it is still necessary to talk about objects and for your interlocutors to be able to figure out what you are talking about. I do agree that the standards of precision required in object identification may differ from context to context. But referring is not absent in chitchat.

OUP CORRECTED PROOF – FINAL, 17/1/2019, SPi

JOINT REFERENCE



used as cues to help establish a referential fix.³ This is one way in which map task conversations deviate from face-to-face conversational exchanges. Secondly, the map task is made difficult by the fact that the director’s and the follower’s maps do not share all their features, so some feature descriptions by either the director or the follower may potentially be hard to interpret at first. When one hears a strange description, one has to wonder whether this is intended as a description of a feature that is in fact missing from one’s own map or is simply a deviant description of a shared map feature.⁴ A third difficulty is that no orienting information, such as a compass rose, is printed on the maps. Thus the dyads have to either explicitly or implicitly work towards a shared means of talking about relative locations on the map. Three possibilities are to use an absolute, or an intrinsic, or a relative frame of reference. Depending on which of these is chosen, one and the same map feature would be described in different ways. For example, a map feature could be described as ‘the blacksmithy to the west northwest of the carpenter’s hut’ in the absolute frame (assuming that the up direction on the map is north), or as ‘the blacksmithy to the right of and slightly behind the carpenter’s hut’ in the intrinsic frame, or as ‘the blacksmithy to the left and slightly above the carpenter’s house’ in the relative frame.⁵ In fact, dyads in the HCRC Map Corpus also resorted to other ways of establishing map locations, such as “so you should be about oh two inches off the left of the paper” (HCRC Map Corpus, observation qec) and “just look at that the d—in between the ‘d’ and ‘e’ just put a straight line straight above that” (HCRC Map Corpus, observation qec). The ‘d’ and ‘e’ mentioned here are letters in one of the map labels, ‘alpine garden’. There are also cases in which frames are mixed together. For example, one guide mixed directions using a relative frame with ones using an absolute frame, such as in the instruction “and then go to the right just a wee shade about w—one inch an in—inch and a half stop an inch and a half and then sort of go northeast a wee bit” (HCRC Map Corpus, observation qec, lines –). One way in which the map task resembles Hutchins’ pilotage case is that followers in the map task need to trace out routes on their maps on the basis of information they get from directors, just as the plotters on a navigation team need to plot a fix on their charts of their ships’ current location on the basis of information passed to them from bearing ³ In some variants of the map task, an opaque barrier separates participants from each other. In other variants, they are allowed eye contact (although of course they still cannot see each other’s maps). Even in the eye contact condition, pointing and eye gaze will not help to establish a referential fix. However, being able to observe facial expressions and bodily postures might help participants to make judgments about how their messages are being received and whether or not adjustments are needed. So, potentially, the task is easier in this condition. ⁴ For example, in one case reported in Brown (: ), a participant described a map symbol for an electric pylon as a ‘colon’. The pylon was a shared map feature and the listener in this case appeared to accept the description and to understand it as pertaining to the pylon symbol. ⁵ Some referential communication tasks have been designed to test people’s abilities to adopt alternative referential frames and to make adjustments when the other participant has an orientation different from their own. Schober () designed a task where the follower’s display was rotated either  or  with respect to the director’s display, a fact known to the participants.

OUP CORRECTED PROOF – FINAL, 17/1/2019, SPi



ANNE BEZUIDENHOUT

recorders, who in turn get their information from bearing readers. That is, there is a need for some sort of action in the world in order for the map task to be successfully completed. Success does not end inside the skull of the follower, with the follower in some sort of mental state (e.g., a state of understanding or belief). A second way in which the map task is like the pilotage case is that both situations involve working with things, namely schematic maps or navigation charts, that can only be used and acted upon by people who have learned and share certain sociocultural representational conventions. The referential communication literature is vast and it is not my intention here to summarize the findings from the research reported in this literature. Referential communication tasks have been used at least since the s, when Glucksberg and Krauss, along with their colleagues, devised a task which became the model for countless others. Their intention was to study the development of referential communication abilities in children. See, for example, Glucksberg et al. (). One of the principal drivers of the early work on referential communication was to address the issue of whether children are egocentric communicators; that is, communicators who are incapable of providing helpful informational cues to an audience who lacks the information available from their own perspectives and/or who are apt to use whatever information is available to them for language understanding, regardless of their interlocutors’ potentially different informational states. After decades of such research, it is hard to say that any overall consensus has been reached. Although a good deal of evidence has accumulated that children are not egocentric and that their poor performance on referential communication tasks can be attributed to other causes (see Dickson  and Lloyd et al.  for summaries), there are still studies being produced that assert the opposite (e.g., Girbau ).⁶ More recently, work on referential communication has enjoyed a renaissance, although the focus has shifted to adult communication, with many studies claiming to show that adults are egocentric communicators or at least have an egocentric bias. Keysar and his colleagues have adapted the old referential communication paradigm from the s and married it with eye tracking using the Visual World paradigm in ⁶ As the debate about children’s referential communication abilities was losing some steam in the late s and early s, the debate about children’s theory of mind (ToM) was gaining momentum. The concern about children’s egocentrism was preserved in this new literature, although the methods used to investigate this issue were slightly different, with variants on the so-called ‘false belief task’ being the principal way in which children’s knowledge of other minds was tested. Again, this literature is vast and I cannot summarize the findings here. I mention this simply to note that the issue about children’s egocentric communication did not go away but simply became subsumed under a more general concern about children’s social cognition skills. After all, knowledge of one’s interlocutor’s beliefs, preferences, and intentions is a crucial part of verbal communication in general and referential communication in particular. Pertinent to the theme of this volume, Gundel and Johnson () investigated children’s use of referring expressions in the context of the debate about children’s ToM, concluding that there may be two stages in the development of ToM abilities, since children are able to use referring expressions appropriately before they have developed fully-fledged metarepresentational skills. See also Chapter  of this volume, by Anne Salazar Orvig, on children’s use of referring expressions.

OUP CORRECTED PROOF – FINAL, 17/1/2019, SPi

JOINT REFERENCE



order to assess adult referential communication abilities. Participants in their studies reveal an egocentric bias, in the sense that they use their interlocutors’ perspectives only when a problem is detected, in which case they make (consciously controlled and effortful) adjustments to account for alternative perspectives on potential referents in order to figure out what the speaker might have intended to refer to. See Keysar () and Keysar et al. (, ). In addition to findings in line with Keysar’s claim, there are competing findings that show that adults use common ground information (which includes information about what information one’s interlocutors have access to) from the very first moments of the comprehension process, thereby casting doubt on the claim that perspective taking only occurs during a second, effortful phase of communication. See Brown-Schmidt et al. (); Hanna et al. (); and Heller et al. (). See also the meta-analysis by Rubio-Fernández (). There is even some recent work using this new updated eye-movement-monitoring paradigm to examine the perspective-taking abilities of both adults and children. See Epley et al. (). There has also been an interest in using referential communication tasks in second language acquisition research. Much of this research is applied and is focused on the issue of whether referential communication tasks can be used in L classrooms as tools for encouraging second language learners to engage in conversational exchanges that are task oriented and more like the types of situations they might encounter in the real world. There is also more theoretical work on L referential communication. The focus of this work tends to be on communication strategies used by L learners. For example, what sorts of referring expressions do they choose, how do they sequence questions and answers, how do they signal communication problems, and how do they accommodate feedback? See Yule () for a summary of the tasks, procedures, and analytic frameworks used in this L research.

.. Models of referential communication In the research referenced in §.., referential communication tasks were used essentially as a lens on perspective-taking abilities rather than on referential communication for its own sake. These researchers assume that when pairs of people communicate about objects in a display of some sort (where the displays available to the two individuals are different in controlled ways), this will give us a good window on whether and how communicators will make adjustments in their messages to take account of their interlocutor’s information states regarding objects in the displays. The purpose of these studies is not to analyze what is involved in referring to an object. Moreover, the notion of communication that results in a mutually shared referential fix on an object is not problematized in any way. On the contrary, some bland assumptions about ideal speaker–hearer exchanges seem to be simply taken for granted. For example, it is assumed that an ideal speaker gives the ideal listener just the right amount of information (no more and no less) to enable the listener to get a cognitive fix on the speaker’s

OUP CORRECTED PROOF – FINAL, 17/1/2019, SPi



ANNE BEZUIDENHOUT

intended referent. This of course assumes that the ideal speaker knows how much information is the right amount for the listener, which explains why researchers think referential communication tasks are a good tool for examining people’s perspectivetaking abilities (and their social cognition and theory of mind skills more generally). My intention, on the other hand, is precisely to problematize the idea of referential communication and to ask what is involved in such communication and whether such communication is a potential site for the activity of jointly referring to an object. A good place to start is with a model of referential communication that I reject. This model is a version of the code model of communication that Sperber and Wilson (, ) argue against. According to the code model, a referential communication cycle begins with the speaker forming a clear idea of an object that she wants to communicate something about. Call this idea a ‘singular concept’ (i.e., a concept of an individual thing). It represents the speaker’s cognitive fix on the object in question. The speaker then forms the intention to refer to that object and searches for an appropriate linguistic vehicle to express that singular concept (e.g., a pronoun, an indexical, a definite description, or some other kind of singular referring expression).⁷ I will ignore the complicated intervening steps that require this expression to be composed with others to formulate a complete message (which may not be a complete sentence).⁸ On the assumption that we are dealing with face-to-face verbal communication between two hearing individuals who belong to the same linguistic community, this encoding of the speaker’s ideas is then spoken out loud (after appropriate information has been sent to the speaker’s articulatory system). The hearer (or the hearer’s auditory system) detects the sounds produced by the speaker and after various processing stages in which these sounds are given appropriate phonological representations and in which a process of linguistic decoding is completed, the hearer will now (on the assumption that there was no noise in the communication channel that produced distortions) have access to the same singular concept that the speaker had. (This is a simplification, as certain changes may be called for to take account of the fact that speaker and hearer don’t occupy exactly the same perspective. For example, what the speaker thinks of as this F may need to be thought of as that F, if what is proximal to the speaker is distal to the listener.) The end result is that the hearer will have a cognitive fix on the very same object that the speaker did, and referential communication will have succeeded. On this view of referential communication, the cycle begins and ends with a mental state (a cognitive fix) and the only reason for any kind of verbalization is that this is the only way to get an idea from one mind to another. If somehow we could hook the two minds up together in some more direct way, we could bypass the need for any verbal

⁷ In her map task studies, Brown (: –) found that people did not in fact confine their choices of expressions to definite expressions but sometimes chose indefinites and even bare NPs. ⁸ For example, it could be just a prepositional phrase, as in the example from Stainton (), in which someone utters the phrase ‘On the stoop’ to direct his companion to the missing table leg that he knows she is searching for.

OUP CORRECTED PROOF – FINAL, 17/1/2019, SPi

JOINT REFERENCE



coding and decoding.⁹ This is also a very individualist picture of referring. The speaker, without any aid from the hearer, comes to have a cognitive fix on an object and thus comes to formulate an intention to refer to that object for the purposes of communicating with the hearer. If all goes well along the communication channel, the hearer will also come to have a cognitive fix on that same object, thereby satisfying the speaker’s referential intention. The hearer’s mental state is as individualistic as the speaker’s initial state, even though of course the speaker played a causal role in bringing that mental state about. There is no joint action of referring in this picture. There are just two individual mental acts of cognitively fixating on an object—one and the same object if there is no noise in the communication channel. On the alternative view I am suggesting, referring is an act performed in the course of some sort of collaborative exchange between two or more communicators and occurs in the external world shared by the communicators. In the situation I am imagining of two people collaborating on a map task, it is the dyad that succeeds in referring to a map feature or map location (or to the map itself, or to one of the tools being used for the task, such as the writing tool being used to draw in a route on the map, etc.), not either of the individuals taken singly. Of course, just as in the case of the joint action of moving a piano up some stairs, the individuals in the dyad must be doing their own parts in this joint action. We can accept at least some aspects of the code model, namely that the individuals will need to harness their own cognitive and attentional resources in order to establish a cognitive fix on an object. However, an individual’s obtaining a cognitive fix on an object is not an act of referring, just as an individual’s lifting up his end of the piano is not an act of moving the piano up the stairs. In both cases these are merely parts of a joint action. These individual actions must be performed in coordination with the individual actions of the other agents involved for the joint accomplishment to be realized. Is it a consequence of my view that a single person engaged in (silent) thought cannot be referring to anything? Not necessarily. Conscious thinking is always ‘thinking for speaking’. That is, it is thinking that, if verbalized, would be expressed in a shared public language or code (e.g., if the thinker is a mathematician, the code could be a mathematical or logical symbolism).¹⁰ Thus we can allow that a silent thinker is referring to an object in a conditional sense. The silent thinker is referring to an object O with her thought T just in case, if T were to be verbalized to a suitable audience, she and her interlocutors would succeed in jointly referring to O. ⁹ There are in fact people actively trying to realize such a scheme of mind-to-mind connection, in the hope that if people could directly exchange ideas, they could communicate with each other despite the fact that they speak mutually unintelligible languages. See Hamzelou (), in particular the quotation from one Andrew Jackson from Newcastle University. Needless to say, I believe this sort of speculation is based on a profoundly mistaken view of mind and language. ¹⁰ A fuller treatment of this issue would require an argument against the possibility of a private language, presumably along the lines given by Wittgenstein (: paras. –). The ‘secret’ languages developed by children (e.g., twins) that are sometimes given as examples of private languages are in fact public languages, because they rely on a shared and potentially publically available set of conventions. For a discussion of private language, see Candlish and Wrisley ().

OUP CORRECTED PROOF – FINAL, 17/1/2019, SPi



ANNE BEZUIDENHOUT

It is also compatible with my view that there is such a thing as unconscious thinking that consists of the manipulation of mental representations that are written in a mental code distinct from any human public language (natural or artificial). For example, perhaps the percepts involved in sensory processing are like this: if percepts are indeed such mental representations, we must give some account of their representational properties—we must explain how they are hooked up to properties in the world. I have no favored view of my own in this regard. We could argue for a causalcounterfactual account of the sort defended by Fodor () or the teleosemantic account defended by Millikan (). However one resolves this issue, these mental representations are not consciously available ones and thus the sort of thinking that involves the manipulation of such representations does not involve acts of referring, even in the conditional sense articulated above. I agree that these sorts of unconscious (or subconscious) processes are necessary causal precursors to conscious thinking. For example, in order to consciously think about an object, unconscious perceptual and attentional mechanisms may be at work in the background, but as soon as one becomes consciously aware of an object, one is thinking of it using publically available categories and so we are once again dealing with the thinking-for-speaking case, which is parasitic on overt acts of verbalization in the way already described above. One might also wonder whether my account of joint reference is compatible with the idea of strategic conversation defended by Asher and Lascarides (). They argue that strategic conversations, such as courtroom cross-examinations or political debates, are non-cooperative, because these are situations in which the interlocutors have misaligned motives. Nevertheless, people surely succeed in referring to things during such strategic conversations. How can I explain this, if my account requires all reference to be joint reference and all joint reference to be defined in terms of coordination between the interlocutors? The apparent tension is resolved once we get clear about the level at which the cooperation or non-cooperation occurs. In a strategic conversation, one may want to trick one’s adversary into saying something disadvantageous to them or play on their cognitive biases or ask them leading questions or get them to believe falsehoods by saying something strictly and literally true that nevertheless strongly implicates something false. Here you may be doing things that are non-cooperative in the Gricean sense. However, in all these cases, one must be relying on the sort of coordination needed to ensure successful joint reference, otherwise these higher order goals of trickery and deceit would not even get off the ground.

.. Some examples of successful and unsuccessful joint reference My discussion thus far has been very abstract. It is worth looking in detail at some examples of dyads performing a version of the map task. The following is an excerpt of a map task conversation cited in Brown (: –). The participants here are

OUP CORRECTED PROOF – FINAL, 17/1/2019, SPi

JOINT REFERENCE



Scottish and ‘++’ signals a long pause of over one second, ‘+’ a pause of half a second, ‘–’ a very brief but audible discontinuity, and ‘?’ a rising intonation: (B&P) A: B: A: B: A:

you’ve got the start hm have you got palm trees – palm beach on your map aye well go – about an inch down from them and go left ++ have you got the swamp on your map? B: crocodiles A: yeah well ++ you avoid that + you go round it – round in nearly a circle – round it ++ so that have you got the waterfall B: mhm A: you go round the waterfall B: left or right or what A: eh + left round the waterfall + and up This seems to be a relatively successful dyad. The director (A) has assumed a relative spatial frame where directions (up, down, left, right) are determined from the map viewer’s perspective. The follower (B) simply accepts this frame, as evidenced by B’s question in line . A’s descriptions ‘the start’ and ‘palm trees – palm beach’ appear to be effective in enabling B to get a cognitive fix on the relevant map features and B signals successful coordination by responding with ‘hm’ and ‘aye’ on lines  and  respectively. What is interesting about this conversation is the next set of turns. A asks whether B’s map has ‘the swamp’ and B replies simply ‘crocodiles’. A then signals, by uttering ‘yeah well’, that he/she takes B’s reply as confirmation that there is a relevantly similar map feature on B’s map and so the task can proceed. The long pause after ‘yeah well’ may indicate that A is making needed updates to the mutual cognitive environment (i.e., the information that is common to A’s and B’s cognitive environments). In other words, after the update it will be mutually manifest to A and B that swamps and crocodiles are associated with each other, so that an indicator of the presence of one is in effect an indicator of the presence of the other. (For discussion of the notion of mutual manifestness and cognitive environments, see Sperber and Wilson .) This shared assumption allows for a case of successful joint reference. Note also that successful joint reference does not require that A and B share a precise nor an exactly similar cognitive fix. Their respective cognitive fixes are similar enough to ensure that their act of jointly referring is good enough for the map task purposes. In fact, participants in the HCRC Map Task corpus do at times note that this is not a task that calls for precision. One participant expresses this by saying: “it doesnae need to be that accurate we’re just trying to avoid all these other things” (HCRC Map Task corpus, observation qec, line ). One might argue that what this dyad jointly referred to was a map feature type, whereas each member of the dyad made a successful individual reference to a map

OUP CORRECTED PROOF – FINAL, 17/1/2019, SPi



ANNE BEZUIDENHOUT

feature token. Firstly, I agree that in this case the dyad had to coordinate on a feature type rather than a feature token. However, this is a peculiarity of this particular conversational setting, because there were two maps that needed to be brought into relation to one another. It is always possible to imagine cases where two people are consulting a single map, where the goal is joint reference to a map feature token, and yet in which a similar set of coordinating moves would be needed to achieve joint reference. Imagine for instance that the map is displayed on a giant screen on a wall, above the heads of the two people, so that they cannot directly put a finger on any map feature. Moreover, they don’t have a laser pointer, etc. So joint reference is not always to types rather than tokens. Secondly, to the implicit objection that in the map task case there are two cases of successful individual reference, I would argue that the acts of individual referring posited here are parasitic on cases of joint reference (in the same way that I argued in §.. that reference in thinking is parasitic on cases of public joint reference). In other words, the members of the dyad, M₁ and M₂, are referring individually in a counterfactual conditional sense; M₁ is individually referring to a map feature token T on M₁’s map just in case M₁ and M₂ would succeed in jointly referring to T if the two were working together on M₁’s map. Similar remarks apply to the case of M₂ referring in an individual way to a map feature token on his/her map. Alternatively, I could simply deny that there is individual reference here. There are individual cognitive fixes, but these are individual basic actions that must be put into coordination in order for reference to be achieved. This easy coordination between dyad (B&P) is interesting because apparently not all dyads in Brown’s map task study were as successful. For an example of a dyad that fails to establish joint reference, see the conversation between R&H from the same corpus (Brown : ). However, I will not analyze this case here. A clearer example of lack of coordination that leads to failure of joint reference is given in the following conversational snippet between a male native speaker (NS) of English and a female non-native speaker (NNS), cited in Pica et al. (: ), in which the NNS was asked to draw and describe a picture:¹¹ NS: does the TV have antennas? NNS: terrace? NS: eas,12 like two things coming up in the back antennas? ah . . . NNS: eh . . . NS: ok, we’ll pass

¹¹ Pica et al. () are interested in the role that gender plays in such conversations and this example is given as an illustration of the greater tendency of male NSs compared to female NSs to discontinue negotiation when a NNS seeks clarification. ¹² The transcript reads ‘eas’ and Yule (: ), who cites this same snippet, corrects it to ‘yes’. However, it is possible that the native speaker intended to says ‘ears’, as TV antennae are sometimes called ‘rabbit ears’ in colloquial American English.

OUP CORRECTED PROOF – FINAL, 17/1/2019, SPi

JOINT REFERENCE



It is clear in this case that the interlocutors have not been able to coordinate on the feature of a TV antenna. This is presumably partly because of a gap in lexification on the part of the NNS but also because the NS’s attempts at refining the NNS’s cognitive fix were unsuccessful. The descriptions of TV antennas as ‘ears’ or as ‘two things coming up in the back’ were insufficiently precise to help the NNS. Pica et al. () do not say whether the NS’s speech was accompanied by hand gestures. It is possible that the NS put his hands up on either side of his head to represent rabbit ears, but this culture-specific correlation of antenna with rabbit ears would undoubtedly have been lost on the NNS (not to mention that hands at the side of one’s head representing rabbit ears would itself have been hard to process). At any rate, the NNS responds on line  with ‘eh . . . ’, signaling puzzlement. And the NS ends this sequence of turns with an admission that coordination has failed. Thus joint reference is not achieved in this case. Another example of negotiated reference where it is clear that work is done by a dyad to accomplish the goal of referring is given in Hutchins (: –). Hutchins does not cite this as an example of joint referring, but rather as an example of learning on the job. In his book, Hutchins generally takes a notion of referential communication for granted. His primary aim is to elucidate the joint action of pilotage. So my account extends his analysis, establishing that there are joint actions all the way down to the level of referring to objects. The conversation cited by Hutchins takes place between two members of the navigation team aboard a Navy helicopter transport ship as it is leaving port. One of the team members, SW, is the bearing taker who is responsible for operating the special telescopic sighting device on the starboard wing of the navigation bridge. The other is the navigation team’s plotter, C, who happens to be the highest ranking and most experienced member of this team. SW is unable to locate the north end of the th Avenue terminal, which he is supposed to use as a landmark to take his next bearing reading to help the plotter fix the ship’s current location. C goes out on the wing to point out the landmark to SW. C puts his arm across SW’s shoulders and aims SW’s body in the right direction. The conversation then proceeds as follows: C: The north one, all the way up SW: OK C: If you can’t see the light, just shoot the tangent right on the tit of the, the last end of the pier there. SW: OK, that pier, where those two . . . C: Yeah, all the way at the end. SW: All right. C: There should be a light out there, but if you can’t see the light out there at the end of the pier just shoot the end of the pier. In this example a lot of the communicative work is being done non-verbally. In order to coordinate their gaze directions, C physically manipulates SW’s body in such a way that he and SW are looking in the same direction. This does not yet guarantee that SW

OUP CORRECTED PROOF – FINAL, 17/1/2019, SPi



ANNE BEZUIDENHOUT

has a suitable cognitive fix on the object C has in mind, so he uses a definite description to narrow the focus. However, this still requires SW to do a fair amount of processing/ inferencing. C’s description ‘the north one’ leaves it implicit as to what type of thing C has in mind. Of course, SW has already been told to mark the bearing of the north end of the th Avenue terminal, so SW can use his general knowledge about ports to narrow his search to things that can plausibly be called ‘terminals’. In his next turn, C uses two additional definite descriptions, ‘the light’ and ‘the last end of the pier there’. These are presumably intended to further refine SW’s focus. It is not clear that C is using these descriptions to characterize things he can actually clearly see or whether these are things he knows from memory.¹³ It is very possible that these fine details can only be seen through the telescopic sighting device that SW is standing in front of. SW’s next turn is one where he apparently tries to confirm that he has a cognitive fix on the right object by asking whether it is the pier ‘where those two . . . ’, but C cuts SW off before his turn is completed and in C’s next two turns, he essentially reiterates the information he has already given (about the light and the last end of the pier). It is unclear what SW is attempting to get confirmation about by his use of the phrase ‘those two’. It is possible that C immediately understands what SW has in mind (perhaps because C has himself viewed this landmark through the telescopic sights many times before) and that C cuts SW off because it is already clear that SW has the right cognitive fix. However, it is more likely that C cuts SW off because there is no time for elaborate negotiations of joint reference. When the ship is leaving the port, the navigation team takes a fix on the ship’s location every three minutes. With such a short fix cycle, the plotter cannot take too much time out on the wing instructing the bearing taker, as he needs to be on the navigation bridge at the chart table. That it is lack of time that explains why C cuts SW off is also suggested by the fact that C’s next two turns essentially reiterate the information already given. C is trusting that he and SW are already sufficiently coordinated so that once SW looks through the telescopic sights, he will be able to locate the intended landmark precisely enough to be able to ‘shoot’ the target and get the desired bearing reading. C and SW have succeeded in jointly referring to the landmark. Note that for successful joint reference there is no need to converge on an object with sharp boundaries. Language in everyday interactions is not used to carve nature at its precise joints. For one thing, one might doubt that nature has such precise joints. Secondly, even if nature had such precise joints, the standards of precision operative in a particular context are likely to be less than perfect. Rather, a joint reference is good

¹³ Hutchins makes the point that the various positions in the navigation team also represent the typical career trajectory of a navigation team member. One starts out as a pelorus operator (the position held by SW) and moves up to bearing recorder and then plotter (the position held by C). Thus it is likely that, earlier in his career, C would have had SW’s job and had many opportunities to mark the bearing of the north end of the th Avenue terminal himself.

OUP CORRECTED PROOF – FINAL, 17/1/2019, SPi

JOINT REFERENCE



enough if it allows the interlocutors to proceed with whatever larger task is at hand, which is what happens in the case of C and SW discussed above. I mentioned above that the account of joint reference being defended here could be thought of as extending Hutchins’s analysis of the joint action of pilotage and as establishing that there are joint actions all the way down to the level of referring to objects. Hutchins argues that the joint action of pilotage/navigation relies on distributed cognitive processes—that is, cognitive processes that are not confined to the heads of single individuals but are distributed across the heads of many individuals, as well as the external environment in which these individuals are interacting. He claims that distributed cognitive processes involve the transfer of representations across a series of representational media. So, for example, each ‘fix cycle’ that is one of the components of the temporally extended and complex task of navigation of a ship into or out of a port involves the transmission of information about the current bearings of three predetermined landmarks with respect to the ship. Each landmark bearing is initially represented in analog format based on a compass reading on a telescopic sighting device, and then is transformed into a digital representation (a number) that can be entered into a written log. That digital representation is then converted back into an analog representation of the bearing, in the form of a line of sight drawn on a chart. The point of convergence of three such lines of sight gives a fix on the ship’s current location and thus concludes one fix cycle. If we apply this idea that joint actions depend on distributed cognitive processes that involve ‘transfer of representations across representational media’ to the case of joint reference, it may seem that we are forced back into the code model of referential communication that I rejected in §.., which posits a transfer of a representation of an object (a ‘singular concept’) from one person’s mind to another. However, the example of C and SW negotiating joint reference to a landmark allows one to see that the code model is a faulty model. To see the interaction between C and SW as one in which an idea is transmitted from one mind to another and to focus primarily on internal cognitive factors seems inadequate at best. For one thing, it downplays the crucial role of physical gestures, eye gaze, etc.¹⁴ Moreover, the highly underspecified and incomplete nature of the explicit verbal cues used in these sorts of exchanges is an indicator of the degree to which interlocutors are relying on information that is available in the immediate, local shared physical, and social context. These are factors that extend beyond the minds of any one individual and even beyond the minds of the members of the dyad. The code model privileges the internal cognitive states of individuals and sees conversational interactions as necessary only because minds cannot be directly

¹⁴ The role of gestures and bodily movements in talk-in-interaction has become a focus of recent research in interactional linguistics. See, for example, Arnold (); Bolden (); Enfield (); Goodwin (); and Kita (). See also work in social psychology, such as Krauss et al. () and Goldin-Meadow ().

OUP CORRECTED PROOF – FINAL, 17/1/2019, SPi



ANNE BEZUIDENHOUT

connected. The alternative view sees conversational interactions as indispensable. Such interactions are the sites for achieving reference and the internal cognitive states of participants are merely one part of a larger, distributed process. These cognitive states by themselves do not refer.¹⁵ Referring is an action and a cognitive state cannot act. A person can act by recruiting a cognitive state for current conversational purposes, but this cognitive act by itself is not an act of referring. It is merely part of an individual’s contribution to a joint activity of referring. Just as one individual’s contributions to moving a piano to a location upstairs do not by themselves constitute moving the piano upstairs, so too a person’s recruiting a cognitive state (or a series of cognitive states followed by the utterance of certain words and accompanied by certain gestures) does not by itself constitute an act of referring. Some recent work on the use of complex demonstratives could be construed as generally supportive of my claims about reference to objects being a joint achievement. Acton and Potts (: ) claim that the use of a complex demonstrative presupposes “perspectival alignment” and helps to create a shared perspective on the referent of the demonstrative. For example, they argue that there is a distinction between the use of ‘the’ as opposed to ‘that’ in an utterance such as: ‘Could you turn around and get me the/that box of cereal’. They state: “demonstratives don’t presuppose that there is a unique individual in the context satisfying the descriptive content of the NP. Instead, they presuppose that there is a particular individual that satisfies the descriptive content of the NP and that the addressee is equipped to identify it by consulting the speaker’s perspective” (Acton and Potts : ). Furthermore, they argue that the precise pragmatic and social effects that arise from the use of demonstratives are “ultimately determined jointly by the discourse participants during interaction” (Acton and Potts : ). Sometimes the use of demonstratives can be deemed inappropriate, because people feel they do not share common ground with the speaker. This can arouse negative affective states in the audience. Acton and Potts have an extended case study of Sarah Palin’s (over-)use of demonstratives, showing that this resonated with her supporters but alienated others. I am generally sympathetic to these arguments and think Acton and Potts have identified some important ways in which meaning is jointly constructed. However, as regards my specific claim that reference is a joint achievement, it is not clear that the arguments by Acton and Potts () help establish my claim. What they are saying is jointly constructed is the social/emotional effect of demonstrative reference (whether this be positively or negatively valenced). They never explicitly say that demonstrative reference itself is a joint achievement.

¹⁵ Such internal cognitive states may of course carry information, as I acknowledged in §... An account along the lines of Fodor () or Millikan () might explain how.

OUP CORRECTED PROOF – FINAL, 17/1/2019, SPi

JOINT REFERENCE

. N E G O T I A T I N G



JOINT REFERENCE

TO A CATEGORY

.................................................................................................................................. We saw in the previous section that it may take several conversational turns for joint reference to be negotiated. In this section I want to briefly note how this sort of negotiation is also at work in helping people to converge on an appropriate descriptive category for the things they are talking about. Convergence on appropriate descriptive categories is part of what is needed for achieving joint reference, since the application of such categories is what helps fine-tune the interlocutors’ cognitive fixes on an object or a property. So coordination is needed at multiple levels in a conversational interaction, both at the level of meaning and at the level of reference.¹⁶ One sort of case of negotiation over categories involves what Overstreet and Yule () call locally contingent categorization. They have in mind categorizations that depend on the use of what they call general extenders, which are expressions such as ‘and stuff (like that)’, ‘and what have you’, ‘and such’, ‘or something (like that)’, ‘or whatever’. Their point is that sometimes there may be no word or ready-made phrase that picks out just the category of things a speaker has in mind. Thus the speaker uses a term for a central exemplar of the intended category plus one of these general extenders and then relies on the listener to use shared information from the conversational context (including both perceptually and conceptually available shared information) to narrow down the category being referred to. Converging on a joint interpretation may take several conversational turns. Overstreet and Yule (: –) use examples to illustrate cases of both successful and unsuccessful category negotiations.

. C O N C L U S I O N S

.................................................................................................................................. I have argued for the view that referring is a joint accomplishment between two (or more) people in a conversational context that may need to be negotiated over several conversational turns. This joint action of course requires the individuals ¹⁶ The idea that settling on an interpretation is something that happens in interaction and is negotiated over several conversational turns is a view that is by now well entrenched in the pragmatics literature. For example, Arundale (), Haugh (), and others have argued that speech act interpretations are negotiated over several conversational turns. Was someone’s utterance intended as a request or a teasing challenge? Haugh (: ) writes: “we cannot simply talk about inferences made about putative speaker intentions in interpreting utterances in isolation. Utterance interpretation occurs in sequentially organized discourse, and thus [people’s] inferential work about [one another’s utterances] is necessarily dependent on their inferences about each other’s inferences. In other words, inferences about (speaker) meanings are always reflexive and contingent on the understandings by self and other displayed in prior and subsequent turns”.

OUP CORRECTED PROOF – FINAL, 17/1/2019, SPi



ANNE BEZUIDENHOUT

concerned to play their own individual roles in this interaction and to do this in a coordinated way. These individual roles include recruiting certain mental resources in order to establish a cognitive fix on an object (which will include the activation of many subconscious and hence unreportable processes of attention, memory recall, etc.). However, these cognitive states by themselves do not constitute actions of referring. Referring is an action that happens in the world outside the heads of individuals and includes such things as physical gestures and other bodily movements, the coordination of eye gaze, and the manipulation of things in the ‘action space’ defined for a particular conversational setting. All of these things are dependent on socially learned routines and conventions, which is another reason that the processes underlying an action of referring are distributed processes, which are not confined to the heads of single individuals but are distributed across the heads of many individuals, as well as the external environment in which these individuals are interacting. This is the stronger view of joint reference that I said I would be defending. On the stronger view, all referring is joint referring. A weaker view would argue that jointly referring to things is a species of referring actions that includes cases in which individuals refer to things. I have argued that individual cognitive fixes are not instances of referring (since they are not actions at all) and that even if we choose to say that an individual’s recruiting of such a cognitive state is an act of referring, such acts of referring are parasitic on cases of two or more individuals jointly referring to things. The primary sense in which referring happens is the case of two (or more) people jointly referring to something, and when we talk of individuals as referring, or of silent thoughts as referring, or of words as referring, these are all secondary instances of referring, defined in terms of the primary case of the joint action of referring. These are convenient shorthand ways of talking, just as when we talk of jackhammers as hammering or cell phones as alerting us to bad weather conditions.

OUP CORRECTED PROOF – FINAL, 21/1/2019, SPi

CHAPTER



......................................................................................................................

COGNITIVE STATUS AND THE FORM OF REFERRING EXPRESSIONS IN DISCOURSE ...................................................................................................................... JEANETTE K. GUNDEL, NANCY HEDBERG, AND RON ZACHARSKI

. I N T R O D U C T I O N

.................................................................................................................................. O NE of the more interesting facts about human language is that we can use different forms to refer to the same thing, and the same form can be used to refer to many different things. Yet people somehow manage to understand one another. A particular issue of Language for example, can be referred to as an issue of Language, the issue, that issue, this issue, that, this, or it, and any one of these forms could be used to refer to other objects on different occasions. The question then is: what do speakers/writers know that enables them to choose an appropriate form to refer to a particular object and what do hearers/readers know that enables them to identify correctly the intended referent of a particular form? The study of reference has a long tradition in the philosophical literature, and has been investigated from various perspectives within linguistics and psychology (see, e.g., Karttunen ; Nunberg ; Hawkins , , ; Clark and Marshall ; Grosz ; Heim ; Maclaran ; Givón ; Ariel ; Kronfeld ; and numerous works cited therein). Although many important insights and observations have come out of this work, basic facts concerning the distribution and understanding of different forms of referring expression in natural language discourse still remain unexplained. In this chapter we outline a theory

OUP CORRECTED PROOF – FINAL, 21/1/2019, SPi



JEANETTE K. GUNDEL, NANCY HEDBERG, AND RON ZACHARSKI

whose main premise is that different determiners and pronominal forms conventionally signal different cognitive statuses (information about location in memory and attention state), thereby enabling the addressee to restrict the set of possible referents. In §. we introduce the Givenness Hierarchy, a set of implicationally related statuses which we propose are necessary for explaining the relation between referring forms and conditions for their appropriate use and interpretation across languages. In §. we show how the Givenness Hierarchy accounts for restrictions on the distribution of forms for a particular type of reference which Prince (b) calls “inferable”. In §. we propose correlations between statuses on the hierarchy and different forms in Mandarin Chinese, English, Japanese, Russian, and Spanish, and in §. we present results of an empirical study of the use of referring forms in these five languages. Finally, we show in §. how the Givenness Hierarchy interacts with the Gricean Maxim of Quantity to predict the actual distribution and interpretation of forms when necessary conditions for the use of more than one form are met.

. T H E G I V E N N E S S H I E R A R C H Y

.................................................................................................................................. It is widely recognized that the form of referring expressions, like such other aspects of language as word order and sentence intonation, depends on the assumed cognitive status of the referent, that is, on assumptions that a co-operative speaker can reasonably make regarding the addressee’s knowledge and attention state in the particular context in which the expression is used (cf. e.g., Chafe , ; Gundel , ; Prince b; Grosz and Sidner ). But the nature of such cognitive statuses and the logical and empirical relations among them is still a matter of some debate. Moreover, researchers have not always distinguished the statuses themselves (e.g., whether or not an addressee already has a mental representation of a referent and whether attention is focused on the referent) from the means by which a referent acquires a particular status (e.g., whether it has been linguistically introduced, whether it is part of general cultural knowledge, etc.). We propose that there are six cognitive statuses relevant to the form of referring expressions in natural language discourse, and that these are related in the Givenness Hierarchy shown in (). ()

The Givenness Hierarchy in uniquely type > activated > familiar > > referential > identifiable identifiable focus 8 9 < that = {it} this {that N} {the N} {indefinite this N} {a N} : ; this N

OUP CORRECTED PROOF – FINAL, 21/1/2019, SPi

COGNITIVE STATUS



Each status on the hierarchy is a necessary and sufficient condition for the appropriate use of a different form or forms.¹ The relevant English forms are given in (). In using a particular form, a speaker thus signals that she assumes the associated cognitive status is met and, since each status entails all lower statuses, she also signals that all lower statuses (statuses to the right) have been met. For example, the definite article the signals ‘you can identify this’, the demonstrative determiner that signals ‘you are familiar with this, and therefore can identify it, and so on’. We thus agree with Garrod and Sanford  and Ariel  that the different forms serve as processing signals to the addressee. However, while these authors (as well as others who have proposed degrees or types of ‘givenness’) view the statuses signaled by different forms as mutually exclusive, in the model we propose here the statuses are implicationally related (by definition), such that each status entails (and is therefore included by) all lower statuses, but not vice versa. The statuses are thus ordered from most restrictive (in focus) to least restrictive (type identifiable), with respect to the set of possible referents they include. For example, an entity which is in focus is necessarily also activated, familiar, uniquely identifiable, referential, and type identifiable. However, not all uniquely identifiable entities are familiar and not all familiar entities are either activated or in focus. In presenting this framework, we make only minimal assumptions about reference processing and about the representation of referents in long- and short-term memory. None of these are particularly controversial. The individual statuses are characterized below. TYPE IDENTIFIABLE: The addressee is able to access a representation of the type of object described by the expression. This status is necessary for appropriate use of any nominal expression, and it is sufficient for use of the indefinite article a in English. Thus, a dog in () is appropriate only if the addressee can be assumed to know the meaning of the word dog and can therefore understand what type of thing the phrase a dog describes. ()

I couldn’t sleep last night. A dog (next door) kept me awake.

REFERENTIAL: The speaker intends to refer to a particular object or objects. To understand such an expression, the addressee not only needs to access an appropriate type representation, he must either retrieve an existing representation of the speaker’s intended referent or construct a new representation by the time the sentence has been processed. The status ‘referential’ is necessary for appropriate use of all definite expressions, and it is both necessary and sufficient for indefinite this in colloquial English.² Thus, while () can have an interpretation where the speaker intends to say ¹ We assume here that demonstrative this and indefinite this are two distinct forms rather than just different instances of the same form. This assumption predicts correctly that a (proximal) demonstrative whose referent must be activated will not necessarily have an indefinite use in other languages. ² There has been considerable debate in the philosophical and linguistic literatures concerning the referential status of both definite and indefinite expressions (cf. e.g., Russell ; Strawson ; Donnellan ; Partee ; Chastain ; Fodor and Sag ; Ludlow and Neale ). To the extent that we are concerned here with USES of referential expressions, i.e., with speaker reference, rather than referential interpretation in a purely semantic sense (cf. Kripke ) our work is independent of

OUP CORRECTED PROOF – FINAL, 21/1/2019, SPi



JEANETTE K. GUNDEL, NANCY HEDBERG, AND RON ZACHARSKI

something about a particular dog or one where she is simply asserting that there is at least one dog (next door) that kept her awake, () is appropriate only if the speaker intends to say something about a particular dog.³ ()

I couldn’t sleep last night. This dog (next door) kept me awake.

UNIQUELY IDENTIFIABLE: The addressee can identify the speaker’s intended referent on the basis of the nominal alone. This status is a necessary condition for all definite reference, and it is both necessary and sufficient for appropriate use of the definite article the.⁴ Identifiability may be based on an already existing representation in the addressee’s memory, as would probably be the case in () without the material in parentheses, but as Hawkins () and others have pointed out, identifiability does not have to be based on previous familiarity if enough descriptive content is encoded in the nominal itself. For example, the phrase the dog next door in () would be perfectly felicitous even if the addressee had no previous knowledge that the speaker’s neighbor has a dog. ()

I couldn’t sleep last night. The dog next door kept me awake.

Thus, expressions which are referential, but not uniquely identifiable, require the addressee to construct a new representation as determined by the content of the much of this debate. Thus, like Chastain () and Fodor and Sag () we believe that indefinites may be used either referentially or non-referentially; but we agree with Chastain (), Searle (b), Bach (b) and Birner () that definite expressions are always used referentially in the sense that speakers intend to refer to a particular entity in using them—either one they are acquainted with and intend to refer to irrespective of whether the description actually fits (Donnellan’s “referential use”) or one which the description actually fits, irrespective of whether the speaker is directly acquainted with it (Donnellan’s “attributive use”). The sense of ‘referential’ that we define here is not to be confused with the sense of ‘specific’ whereby the phrase a student in the syntax class in (i) is necessarily specific since it can only have a wide-scope existential reading (i.e., There is a student in the syntax class who . . . ). (i) A student in the syntax class cheated on the final exam. This phrase could be used either referentially or non-referentially here, since a person who utters (i) “might be intending to assert merely that the set of students who cheated on the syntax exam is not empty [the non-referential reading]; or he might be intending to assert of some particular student, whom he does not identify, that this student cheated [the referential reading]” (Fodor and Sag : ). See Enç () for a discussion of some other senses in which the term ‘specific’ has been used. ³ A number of researchers have found that referents of indefinite noun phrases introduced with this are more likely to be continued in subsequent sentences than referents of phrases introduced with a (Prince a; Wright and Givón ; Gernsbacher and Shroyer ). Such findings would be expected if speakers always intend to refer to a particular entity when using phrases introduced by this, whereas phrases introduced by indefinite a, which requires only type identifiability, are ambiguous between a referential and a non-referential interpretation. Indefinite this is very likely an extension of the cataphoric use of the proximal demonstrative, that is, its use in referring to an object which will not be activated for the addressee until the next sentence is processed, as in What I wanted to tell you is this. Last night, . . . (see Perlman  and Maclaran  for discussion). ⁴ The Givenness Hierarchy thus allows us to define explicitly the notion of definiteness. A noun phrase is definite if its referent is necessarily at least uniquely identifiable. Since ‘type identifiable’ and ‘referential’ are the only statuses that don’t entail ‘uniquely identifiable’ it follows that all forms listed under statuses to the left of ‘referential’ (i.e., all but a and indefinite this) are associated with noun phrases that are definite.

OUP CORRECTED PROOF – FINAL, 21/1/2019, SPi

COGNITIVE STATUS



referential expression along with the rest of the sentence. For expressions which are both referential and uniquely identifiable, on the other hand, the addressee is expected to retrieve a representation on the basis of the referring expression alone (see Webber  and Millikan  for further discussion).⁵ FAMILIAR: The addressee is able to uniquely identify the intended referent because he already has a representation of it in memory (in long-term memory if it has not been recently mentioned or perceived, or in short-term memory if it has). This status is necessary for all personal pronouns and definite demonstratives, and it is sufficient for appropriate use of the demonstrative determiner that.⁶ Thus (), unlike (), is appropriate only if the addressee already knows that the speaker’s neighbor has a dog. ()

I couldn’t sleep last night. That dog (next door) kept me awake.⁷

ACTIVATED: The referent is represented in current short-term memory. Activated representations may have been retrieved from long-term memory, or they may arise from the immediate linguistic or extralinguistic context.⁸ They therefore always include the speech participants themselves. Activation is necessary for appropriate use of all pronominal forms, and it is sufficient for the demonstrative pronoun that as well as for stressed personal pronouns. The pronoun that in () can thus be used appropriately to refer to the barking of a dog only if a dog has actually been barking during the speech event or if barking had been introduced in the immediate linguistic context.

⁵ This does not mean, of course, that the content of the remainder of the sentence cannot assist the addressee in correctly identifying the speaker’s intended referent. ⁶ There is a relatively infrequent use of the demonstrative that which does not require familiarity. For example, the referent of the demonstrative phrase in (i) is uniquely identifiable, but not necessarily familiar. (i) It has great potential for those who must read technical documents. [message from electronic newsgroup] We believe this to be a special ‘precision’ use of the demonstrative that, which emphasizes the exhaustiveness of the referent. For example, the presence of those in (i) forces and emphasizes an interpretation where the referent is ALL the people that meet the description. There were no instances of ‘precision’ that in the present study, and only % of the tokens of ‘demonstrative’ that in our previous studies were of this type. ⁷ Our characterization of the distinction between demonstrative determiners and the definite article appears to disagree with that of Hawkins (), who proposes that ‘identifiability’ is a necessary property of demonstratives, but not of the definite article. However, this disagreement may be more terminological than substantive, since Hawkins’s definition of identifiability is closer to our ‘familiar’ than to our ‘uniquely identifiable’. Thus, like Hawkins, we maintain that familiarity is not part of the conventional meaning of the definite article. This point is discussed further in §.. ⁸ Cf. Sgall et al. ; Chafe , ; and Gundel  for earlier uses of the term ‘activated’. A number of important questions arise concerning the nature of representations in memory and their relation to (representations of ) linguistic forms. For example, does the process of retrieving a representation from long-term memory involve first constructing a representation from the discourse and then somehow matching this one to one in memory, or are the previous representations accessed directly? While such questions need to be addressed in a complete theory of reference, they are beyond the scope of the present study.

OUP CORRECTED PROOF – FINAL, 21/1/2019, SPi

 ()

JEANETTE K. GUNDEL, NANCY HEDBERG, AND RON ZACHARSKI

I couldn’t sleep last night. That kept me awake.

Activation is also necessary for appropriate use of the definite demonstrative determiner this. Both determiner and pronominal this require the referent to be not only activated, but speaker-activated, by virtue of having been introduced by the speaker or otherwise included in the speaker’s context space (cf. Lakoff ; Fillmore , ; Halliday and Hasan ). The phrase this dog in () is therefore inappropriate in the context of A’s question. ()

A: Have you seen the neighbor’s dog?   ??this dog B: Yes, and kept me awake all night. that dog

But in (), where the dog has been introduced by the speaker, either this or that is appropriate.   This dog kept me awake last night.⁹ () My neighbor has a dog. That IN FOCUS: The referent is not only in short-term memory, but is also at the current center of attention. This status is necessary for appropriate use of zero and unstressed pronominals. The entities in focus at a given point in the discourse will be that partially ordered subset of activated entities that are likely to be continued as topics of subsequent utterances. Thus, entities in focus generally include at least the topic of the preceding utterance, as well as any still-relevant higher-order topics.¹⁰ To the extent that syntactic structure and prosodic form encode topic-comment structure and serve to highlight constituents whose referents the speaker wants to bring ⁹ When this is used to refer to an entity not activated by the speaker, the speaker-activation condition is being exploited to convey special effects, such as solidarity. One such use is in polite interruptions, generally clarification questions like (i) (see Hedberg ). (i) This is Chris you’re talking about, right? [Frederickson tapes] Another example is in expressions like this is true, where. as Georgia Green has suggested to us, the speaker, by using this rather than that appropriates an idea introduced by the addressee. The speaker-activation condition may also be extendable to uses of this for extralinguistic objects relatively close to the speaker and for intervals including speech time, and uses of that for objects relatively far away from the speaker and for times prior to speech time. ¹⁰ By ‘topic’ we mean what the speaker intends a sentence to be primarily about. While the topic is often in subject position, it does not have to be. In fact it need not be overtly represented in the sentence at all (see Gundel , ). The term ‘focus’ has been used in two distinct ways in the literature (see Hajičovà (). We use ‘in focus’ here to refer to the psychological notion of focus of attention (Hajičovà’s focusAI cf. Linde ; Grosz and Sidner ). This is to be distinguished from the notion of focus as the position of linguistic prominence in the part of the sentence that expresses the comment (Hajičovà’s focusL cf. Halliday ; Chomsky ; Jackendoff ). These two senses of focus are related, however, in that elements tend to be linguistically focused because the speaker wants to bring them into the focus of attention. In addition, like the topic of a sentence, the referent of a linguistically focused element is likely to be in focus in subsequent utterances in the discourse.

OUP CORRECTED PROOF – FINAL, 21/1/2019, SPi

COGNITIVE STATUS



into focus, membership in the in-focus set is partially determined by linguistic form (cf. the centering and focusing algorithms of Grosz et al. ; Sidner ; and Dahl ). For example, subjects and direct objects of matrix sentences are highly likely to bring a referent into focus, whereas this is not the case for elements in subordinate clauses and prepositional phrases. Thus, in (a) the bull mastiff is not currently in focus because it has not been mentioned in the immediately preceding discourse. But since it is introduced in matrix subject position (and is most likely also the topic) in (a), it is brought into focus, and can therefore be appropriately referred to with either that or it in (b). But in (), where the bull mastiff has been introduced in a prepositional phrase that functions primarily to restrict the referent of the indirect object, reference with it is inappropriate. ()

a. My neighbor’s bull mastiff bit a girl on a bike.   It’s b. the same dog that bit Mary Ben last summer. That’s

() a. Sears delivered a new bike to my neighbors with the bull mastiff.   #It’s the same dog that bit Mary Ben last summer.¹¹ b. That’s c. Anyway, this new bike is really hideous and . . . While linguistic form plays an important role in determining what will be brought into focus, actual inclusion in the ‘in-focus’ set depends ultimately on pragmatic factors, and is not uniquely determinable from the syntax. For example, a large wind energy project in (a) is in a syntactic position similar to that of the bull mastiff in (a), but its referent, unlike that of the bull mastiff, is brought into focus because of its importance in this context. Subsequent reference with the unstressed pronoun it, as in (b), is therefore appropriate here. ()

a. However, the government of Barbados is looking for a project manager for a large wind energy project. b. I’m going to see the man in charge of it next week. [personal letter]

. I N F E R A B L E S

.................................................................................................................................. In her seminal work on givenness, Prince (b) proposes different types of givenness/newness and suggests that these are related in the following preference hierarchy, though she does not explicitly link the statuses with particular forms.

¹¹ We use # here to indicate unacceptability in the given context.

OUP CORRECTED PROOF – FINAL, 21/1/2019, SPi



JEANETTE K. GUNDEL, NANCY HEDBERG, AND RON ZACHARSKI

()

FAMILIARITY SCALE (Prince b): 8 9 Evoked > > > > < = > Unused > Inferable Situationally > > > > : ; Evoked

Containing > Inferable

Brand Brand New > New Anchored

An important distinction between the statuses in () and those in the Givenness Hierarchy is that the Familiarity Scale does not distinguish between ‘activated’ and ‘in focus’; the status ‘evoked’ covers both. Furthermore, while statuses in both scales are ranked according to degree of givenness (from most familiar to least familiar), the relation between statuses in the Givenness Hierarchy is one of entailment, while statuses in the Familiarity Scale are mutually exclusive. Some of Prince’s statuses correspond to set differences between ours. For example, ‘unused’ corresponds roughly to ‘familiar’ but not ‘activated’; ‘containing inferable’ corresponds to ‘identifiable’ but not ‘familiar’; and ‘brand new’ corresponds to ‘type identifiable’ but not ‘uniquely identifiable’. As we show in §., entailment relations among the statuses allow for a straightforward explanation of the interaction of the Givenness Hierarchy with Grice’s Maxim of Quantity in predicting the actual distribution of forms in discourse. Of particular concern to us here is the category Prince calls ‘inferable’ (see also the ‘bridging inferences’ of Clark and Haviland , the ‘associated anaphors’ of Hawkins , and the ‘indirect anaphors’ of Erkü and Gundel ). In such cases, the speaker assumes that the hearer can infer an entity “via logical—or more commonly, plausible—reasoning from discourse entities already Evoked or from other Inferables” (Prince b: ). For example, the referents of a whole paragraph in () and the pulse in () are inferables. ()

[Boss to secretary who just typed an affidavit he is reading] ‘Miss Murchison,’ said Mr Urquhart, with an expression of considerable annoyance, ‘do you know that you have left out A WHOLE PARAGRAPH?’ [Dorothy Sayers, Strong Poison ()]

()

Members of the jury—there is no need, I think, for me to recall the course of Philip Boyes’ illness in great detail. The nurse was called in on June st, and during that day the doctors visited the patient three times. His condition grew steadily worse . . . On the day after, the nd, he was worse still—in great pain, THE PULSE growing weaker, and the skin about the mouth getting dry and peeling off. [Dorothy Sayers, Strong Poison (: )]

To account for such examples, Garrod and Sanford () distinguish between explicit focus, which contains representations of entities directly mentioned in a discourse, and implicit focus, which contains information from situational scenarios that is not specifically mentioned but is directly relevant to something that is mentioned. Similarly, Chafe () suggests that an entity can be “semi-active” (in an individual’s consciousness but not in focus) by being a member of a “set of expectations associated with a schema” evoked by the discourse. Thus, for these authors, as for

OUP CORRECTED PROOF – FINAL, 21/1/2019, SPi

COGNITIVE STATUS



Prince, inferable entities have a separate cognitive status on a par with different types or degrees of givenness. An interesting property of inferables is that they typically do not allow reference with a pronominal (see Garrod and Sanford : –) or with a demonstrative determiner (see Webber : ). The pulse in () cannot be replaced with that pulse or with it, for example. Such facts can be accounted for naturally if ‘inferable’ is viewed not as a separate cognitive status but rather as a way that something can achieve a particular status by association with an entity that has been activated. We would thus expect inferables to have different statuses, and to be coded by different forms, depending on the nature and strength of the link between the inferable and its associated discourse entity. If, as suggested in Garrod and Sanford , hearers/ readers do not automatically construct token representations of entities that can be inferred but are not directly mentioned in a discourse, then most inferables would have a status lower than familiar, as is the case in () and () above. Thus, in () the addressee is expected to identify the subclass of paragraphs that the referent of a whole paragraph belongs to, namely paragraphs in the affidavit that the speaker is holding. But she is not expected to uniquely identify the particular paragraph in question. That is, a whole paragraph is referential (and therefore type identifiable), but not uniquely identifiable (and therefore also not familiar). In () the referent of the pulse is uniquely identifiable, but not familiar. What must be familiar in order for () to be felicitous is the knowledge that patients have pulses, but there is no reason to expect that the mention of a patient will automatically evoke a representation of that particular patient’s pulse. Since most instances of inferables are of the type in () and (), at most uniquely identifiable, but not familiar, our analysis predicts correctly that they generally cannot be referenced by pronouns (which require their referents to be at least activated) or by demonstrative determiners (which require their referents to be at least familiar). But when the link between an inferable and its associated discourse entity is strong enough to create (or activate) an actual representation of the inferable, a demonstrative determiner as in (), or even a pronoun as in (), is possible. () We went to hear the Minnesota Orchestra last night. That conductor was very good. () There was not a man, woman, or child within sight; only a small fishing-boat, standing out to sea some distance away. Harriet waved wildly in its direction, but THEY either didn’t see her or supposed that she was merely doing some kind of reducing exercises. [Adapted from Dorothy Sayers. Have His Carcase (: )] In () the link between orchestras and their conductors is strong enough to make a conductor familiar in this case. The conductor can therefore be referenced with a demonstrative determiner. Similarly, in () the mention of someone waving towards a boat is enough to create and bring into focus a representation of people in the boat. Reference with a pronoun is therefore possible here.

OUP CORRECTED PROOF – FINAL, 21/1/2019, SPi



JEANETTE K. GUNDEL, NANCY HEDBERG, AND RON ZACHARSKI

A similar explanation can be given for the contrast between well-known examples like those in ().¹² ()

a. I dropped ten marbles and found all of them except for one. It’s probably under the sofa. b. #I dropped ten marbles and found only nine of them. It’s probably under the sofa.

In (a), reference with one creates and brings into focus a representation of the missing marble, thus licensing reference with the unstressed pronoun it in the second sentence. In (b), by contrast, the fact that there is a missing marble is inferable from the first sentence, but the possibility of inference is not sufficient to create a representation of that marble. Reference with it in the second sentence is thus inappropriate. Since demonstrative pronouns and determiners also require familiarity (i.e., they require an already existing mental representation of the referent), we would predict that a demonstrative determiner or pronoun, like a personal pronoun, could appropriately refer to the missing marble in (a) but not in (b). This prediction is correct, as illustrated in (). () a. I dropped ten marbles and found all of them, except for one.   That ’s probably under the sofa. That missing marble b. I dropped ten marbles and found only nine of them.   That # ’s probably under the sofa. That missing marble However, a form with a definite article can appropriately refer to the missing marble in both the (a) and (b) examples, given a description which is sufficient to uniquely identify the marble. ()

a. I dropped ten marbles and found all them, except for one. The missing marble’s probably under the sofa. b. I dropped ten marbles and found only nine of them. The missing marble’s probably under the sofa.

These facts are also predicted by our analysis, since a definite article, unlike a personal pronoun or demonstrative, requires only that the referent be uniquely identifiable, but not necessarily that it be familiar.¹³

¹² Quoted in Heim (: ) and originally due to Barbara Partee. ¹³ As far as we can see, the contrast between () and () remains unexplained under Heim’s analysis. Note that it would be too strong to require that referents of pronouns and demonstrative phrases must be identical to referents introduced in the linguistic or spatiotemporal context. While correctly ruling out examples like (b) and (b), such a restriction would also rule out acceptable reference with pronouns and demonstratives, as in () and ().

OUP CORRECTED PROOF – FINAL, 21/1/2019, SPi

COGNITIVE STATUS



. U N I V E R S A L I T Y O F T H E G I V E N N E S S HIERARCHY

.................................................................................................................................. We have proposed six implicationally related cognitive statuses, and have shown that each of these is necessary and sufficient for the appropriate use of a different form or set of forms in English. In this section, we discuss correlations between cognitive status and different forms of reference in four additional languages: Mandarin Chinese, Japanese, Russian, and Spanish. These correlations are presented in Table ..¹⁴

Table . Correlation between linguistic form and highest required status. IN FOCUS

ACTIVATED

FAMILIAR UNIQUELY REFERENTIAL TYPE IDENTIFIABLE IDENTIFIABLE

CHINESE Ø TA tā ‘s/he, zhè ‘this’ it’ nèi ‘that’ zhè N ENGLISH it

HE, this, that, this N

JAPANESE Ø

kare ‘he’ kore ‘this’ sore ‘that’ medial are ‘that’ distal kono N ‘this N’ sono N ‘that N’ medial

RUSSIAN Ø ON on ‘he’ èto ‘this’ to ‘that’ SPANISH Ø él ‘he’

yi N ‘a N’ ØN

nèi N

that N

the N

indefinite this N

aN

ØN

èto N to N

ÉL ese N ‘that N’ el N ‘the N’ éste ‘this’ medial ése ‘that’ medial aquel N ‘that aquél ‘that’ distal N’ distal este N

ØN

ØN un N ‘a N’

¹⁴ For simplicity, we have excluded proper names, generics, and indefinite plurals from our study. We have also excluded zero NPs in conjoined and nonfinite clauses, in relativized position, and in special uses of language such as English casual speech (e.g., smells good) and in recipes (e.g., bake for five minutes). Thus, we did not include English among the languages which allow zero (Ø) NPs. We have included only one form on the chart to represent members of a whole class. (In English, for example, it in the ‘in focus’ column, represents all unstressed personal pronouns, and HE in the ‘activated’ column represents all stressed personal pronouns. (As Bolinger  and others have noted, that is the typical stressed form of it). Abbreviations in the glosses are ACC ‘accusative’, ADV ‘adverb’, ASP ‘aspect’, CLF ‘classifier’, COMP ‘complementizer’, GEN ‘genitive’, NOM ‘nominative’, LOC ‘locative’, OM ‘object marker’. TOP ‘topic’, and Q ‘question’.

OUP CORRECTED PROOF – FINAL, 21/1/2019, SPi



JEANETTE K. GUNDEL, NANCY HEDBERG, AND RON ZACHARSKI

Forms are listed below the highest cognitive status necessary for their appropriate use, For example, the proximal demonstrative determiner or pronoun zhè in Chinese requires that the referent be at least activated: a zero (Ø) pronoun in Spanish requires that the referent be in focus, and so on.

.. The universality of cognitive statuses The six statuses in the Givenness Hierarchy appear to be adequate for describing appropriate use of demonstratives, articles, and pronouns in the five languages we examined. However, not all six statuses are required for all the languages. Thus, as seen in Table ., only English has a form (indefinite this N) for which the status ‘referential’ is both necessary and sufficient.¹⁵ The remaining five statuses are not all required for Chinese, Japanese, and Russian, which lack distinct forms for articles. As illustrated in ()–(), a noun with no preceding determiner in these languages can be interpreted as either uniquely identifiable (definite) or merely referential or type identifiable (indefinite). ()

Chinese: Tā zài bıˇsài zhōng huò jiăng he in game during win prize ‘He won a prize in a/the game.’

()

Japanese: Kare wa akai kingyo o hoshii He TOP red goldfish ACC want ‘He wants a/the red goldfish.’

()

Russian: V ruke deržali bilety In hand held tickets ‘In (their) hand(s), (they) held tickets.’

These languages differ, however, as to which status is unnecessary. As Table . shows, Japanese and Russian have no forms for which the status ‘uniquely identifiable’ is both necessary and sufficient. In Chinese, by contrast, this status appears to be sufficient for appropriate use of the distal demonstrative determiner nèi; but Chinese apparently has no determiner which requires the referent to be familiar, but not necessarily activated. Thus, with respect to cognitive status requirements, nèi behaves more like the definite

¹⁵ But see Enç  for a discussion of how her notion of specificity (close to the notion of referentiality described here) is needed to account for the omission of morphological case markers in Turkish, as well as universal constraints on the use of existential constructions.

OUP CORRECTED PROOF – FINAL, 21/1/2019, SPi

COGNITIVE STATUS



article in English and Spanish than like the distal demonstratives in these languages. According to the Chinese speakers we consulted, (), unlike its counterpart in the other languages, is appropriate even if the addressee has no previous knowledge that the speaker’s neighbor has a dog. Gébì-de nèi tiáo () Zuótiān wănshàng wŏ shuì-bù-zháo. yesterday evening I sleep-not-achieve next.door that CLF gŏu jiáo de lìhai dog bark ADV extremely. ‘I couldn’t sleep last night. The dog next door was barking.’ This supports the observation, made for example by Li and Thompson (: –), that the unstressed distal demonstrative in Chinese is beginning to function like a definite article.

.. Correlations between cognitive status and linguistic form In the languages we investigated, the statuses necessary for appropriate use of corresponding forms are the same for all forms except demonstrative determiners and the indefinite article.¹⁶ It is also of interest that when the required status for a corresponding form is the same across languages, the correlation appears not to be arbitrary. Thus, forms which signal the most restrictive cognitive status (in focus) are always those with the least phonetic content, namely unstressed pronouns, clitics, and zero pronominals (cf. Givón ; Kameyama ; Levinson ; and Ariel  for similar observations). In addition, all pronouns (including demonstrative pronouns) require the referent to be at least activated, which is no doubt related to the fact that the minimal descriptive content of a pronoun provides little if any basis for identifying the referent. Finally, in English and Spanish—the languages that have a definite article—the referent of a phrase introduced by this article must be at least uniquely identifiable, but not necessarily familiar. Thus, the phrases headed by the conclusion in (a) and la sombra in (a) do not require previous familiarity. However, the definite article would be inappropriate without the modifiers in these contexts, as in (b) and (b), because the referent would no longer be uniquely identifiable from the description.

¹⁶ We are only talking about cognitive status here: there are, of course, other conditions which differ across languages, such as restrictions on the use of definite determiners with generics and proper names. Languages also exploit morphological devices such as noun incorporation and syntactic devices such as preverbal versus post-verbal position to signal cognitive status.

OUP CORRECTED PROOF – FINAL, 21/1/2019, SPi



JEANETTE K. GUNDEL, NANCY HEDBERG, AND RON ZACHARSKI

() a. I’ve come to THE CONCLUSION THAT I WILL [personal letter] b. #I’ve come to THE CONCLUSION.

NOT EXTEND MY CONTRACT AT THE BANK.

negra de los hombres siguió. () a. La sombra larga y the shadow large and black of the men followed ‘The long and black shadow of the men followed’ [Juan Rulfo, No oyes ladrar los perros] b. #LA SOMBRA siguió. the shadow followed ‘The shadow followed’ While pronominals and the definite article appear to require the same status across languages, the situation is more variable for demonstratives and indefinite articles.

... Demonstratives As noted in §., the referent of a noun phrase introduced by a demonstrative has to be at least familiar in English. A demonstrative determiner is inappropriate when the referent is uniquely identifiable but not familiar to the addressee. In (), for example, the conclusion cannot be replaced by that conclusion or this conclusion. This is also the case for demonstrative determiners in Japanese, Russian, and Spanish. Thus, la sombra in () could not be replaced by esta sombra, esa sombra, or aquella sombra in the particular context in which () occurs. And the Spanish, Japanese, and Russian examples in ()–() are appropriate only if the addressee can be assumed to know that the speaker’s neighbor has a dog. ()

Spanish:   perro de al lado No pudo dormir anoche Ese Aquel not could sleep last.night dog of to.the next.door no me dejó dormir not me let sleep ‘I couldn’t sleep last night. That dog next door kept me awake.’

()

Japanese: Kinoo wa hitobanjuu nemurenakatta. Tonari no ie couldn’t.sleep neighbor gen house yesterday TOP all.night no ano inu no sei da. gen that dog gen reason is ‘I couldn’t sleep last night. That neighbor’s dog is the reason.’

()  Russian:   Eta sobaka u soseda mne vsju noč’ ne davala spat’. Ta dog at neighbor me all night not allow to.sleep ‘That dog next door kept me awake all night.’

OUP CORRECTED PROOF – FINAL, 21/1/2019, SPi

COGNITIVE STATUS



However, as we noted in §.., the Chinese distal demonstrative determiner nei in a sentence like () only requires the referent to be uniquely identifiable, but not necessarily familiar. The languages also differ as to which demonstrative determiners, if any, require the referent to be activated. In Spanish, which has a three-way distinction in demonstratives, only the proximal determiner este requires activation. Thus, both the medial ese and the distal aquel are possible in () above, even if the dog has not been activated by the immediate linguistic or extralinguistic context.¹⁷ But in Japanese, which also has three demonstratives, both proximal and medial determiners, like pronouns, require activation. Although the facts are not entirely clear, it appears that the medial demonstrative sono would be appropriate in () only if the addressee were currently aware of the dog, that is, if it could be assumed that a representation of the dog is already activated.¹⁸ While Japanese and Spanish differ as to whether or not the medial demonstrative determiner requires activation, the proximal demonstrative determiner requires activation in both languages, as it does in Chinese and English. Russian, however, differs from the other languages in lacking the activation condition on the proximal determiner. Thus the proximal èta is possible (and in fact preferred) in () above, even if the dog has not been recently mentioned and is not present in the immediate discourse context, that is, if the dog is familiar, but not activated. Finally, the languages appear to differ in whether one or more demonstrative forms require not only that the referent be activated, but that it be speaker-activated. As we noted in §., the referent of English that can be used to comment on the remarks of another speaker (activated, but not speaker-activated), as in () above, or it can be used to comment on the speaker’s own remarks (speaker-activated), as in (). () John, this speech was a magnificent triumph for the President. He showed he could stay awake for twelve whole minutes. He showed that he could speak every word off his teleprompter, even the long ones. But the speech doesn’t have any chance of putting the scandal behind him, because the scandal is not about mistakes, as he said, and it’s not about mismanagement, as the Tower Commission said. It is about a betrayal of principles, it’s about lying, and it’s about breaking the law. And THOSE ISSUES remain. [PBS. The McLaughlin Group, //] However, pronominal and determiner this REQUIRE speaker activation. Thus, while those issues can be replaced by these issues in (), this is unnatural across speaker boundaries in (). ¹⁷ Although both forms require only familiarity, they are not equally appropriate in all contexts. In languages which have two demonstratives that require only familiarity, the more distal form is generally restricted to contexts where distance (either spatial or cognitive) is being emphasized. Thus, Spanish speakers find the medial ese to be more natural than the distal aquel in an example such as (). ¹⁸ There has been considerable debate regarding conditions on the appropriate use of sono and ano (cf. Kuroda ; Hinds ; Kuno ; and Kitagawa ). Our claim that sono requires activation while ano requires only familiarity appears to be at least consistent with all positions. But it is not clear whether this distinction can account for all differences in the distribution of these two forms.

OUP CORRECTED PROOF – FINAL, 21/1/2019, SPi



JEANETTE K. GUNDEL, NANCY HEDBERG, AND RON ZACHARSKI

()

A: I think that my novels are better than his. B: I agree with that (statement)/??this (statement).

Speaker activation also appears to be required for proximal demonstrative pronouns and determiners in Japanese (kore, kono) and Spanish (ésta, esta). Thus kore in () is inappropriate in B’s utterance, because the referent has been activated by A. ()

Japanese: A: Watashi no hon wa Mishima no yori ii. GEN book TOP Mishima GEN than good I ‘My book is better than Mishima’s.’   B: Sore wa honto da. #Kore TOP true be ‘That/this is true.’

And, while the facts are less clear than in Japanese, most Spanish speakers that we consulted find the proximal esto to be less natural than the medial eso in (). ()

Spanish: A: Mis cuentos son mejores que los de Ortega my stories are better than those of Ortega ‘My stories are better than Ortega’s.’   B. Estoy de acuerdo con eso ??esto am in agreement with ‘I agree with that/this’

Finally, as illustrated in () and (), proximal demonstratives in Chinese and Russian do not require speaker activation. (As noted above, the proximal determiner in Russian does not require activation at all.) In fact, speakers prefer the proximal (zhè, èto) to the distal (nèi, to) forms in these examples. ()

Chinese: A: Wŏ juéde wŏde xiăoshuō bi lŭxūn xiĕ-de hăo. I think my novel than Luxun write-NOM good ‘I think my novels are better than Luxun’s.’   wŏ chéngrèn. B: Zheige ?Neige I admit ‘I agree with this/that.’

()

Russian: A. Moi rasskazy lučše čem ego rasskazy. my stories better than his stories ‘My stories are better than his stories.’

OUP CORRECTED PROOF – FINAL, 21/1/2019, SPi

COGNITIVE STATUS

 B. Ja s

etim ?tem



 soglasen.

I with agree ‘I agree with this/that.’

... The indefinite article Of the languages we examined, only Chinese, English, and Spanish have an indefinite article, and these languages are alike in that the indefinite article appears to require only type identifiability. But distribution of the indefinite article differs somewhat in the three languages. As noted in §., a noun phrase introduced by English a may be referential or merely type identifiable. Thus, a sentence such as (a) can have an interpretation where the speaker intends to refer to a particular car and wants the addressee to construct a representation of that car (the referential reading), or it can have an interpretation where the speaker intends only to assert that Ellen bought at least one car, that is, she is now a car owner (the non-referential, merely type-identifiable reading). The two interpretations are distinguished in (b) and (c), respectively. () a. Ellen bought a car. b. We went to Southtown Toyota yesterday. Ellen bought a car. It’s a Regatta Blue Corolla with a sunroof. c. A: You’ll never guess what happened today. B: Don’t tell me. Ellen bought a car. In contrast to a, which requires only type identifiability, the referent of indefinite this must be referential. Thus, (a) is appropriate in the context of (b), but not in (c). () a. Ellen bought this car. b. We went to Southtown Toyota yesterday. Ellen bought this car. It’s a Regatta Blue Corolla with a sunroof. c. A: You’ll never guess what happened today. B: #Don’t tell me. Ellen bought this car. The Spanish indefinite article un also requires only type identifiability, as illustrated by examples like () and (), where the referents of the phrases una gaseosa and una celosa fanatica are type identifiable, but not referential. pero no encuentre. () El quiere (una) gaseosa he wants a soft.drink but not find ‘He wants a soft drink, but can’t find (one).’ () No sos (una) celosa fanática. not be a jealous fanatic ‘Don’t be a fanatic jealous (woman).’

OUP CORRECTED PROOF – FINAL, 21/1/2019, SPi



JEANETTE K. GUNDEL, NANCY HEDBERG, AND RON ZACHARSKI

However, unlike the English indefinite article, the Spanish indefinite article is optional in clearly non-referential contexts like () and (), and in some non-referential contexts it is not used at all. Chinese does not have an obligatory indefinite article. However, as we will see in the next section, the distribution of yi ‘one’ in actual discourse is similar to that of the indefinite article in Spanish. Most singular referential indefinites and some nonreferential ones are introduced by yi.¹⁹ Bolinger  (Abstract) notes that “in Old English the indefinite article, now required in the John is a lawyer type of sentence, was optional.” Bolinger also points out that the indefinite article is still optional in Modern English “in certain peripheral constructions” such as He’ll never make (a) captain and He’s running for (an) office. Chinese, English, and Spanish thus appear to be at different stages in the development of the indefinite article from the numeral ‘one’. Chinese represents a very early stage in which the indefinite article is optional and is generally restricted to referential contexts. In Spanish the indefinite article is optional in some non-referential uses; but it has become grammaticalized for referential indefinites. Finally, the English indefinite article has become grammaticalized not only for referential indefinites, but also for most non-referential ones.

. T H E G I V E N N E S S H I E R A R C H Y

AND

NATURAL LANGUAGE DISCOURSE

.................................................................................................................................. The Givenness Hierarchy and the associated forms in Table . predict that a particular form will be inappropriate if the required cognitive status is not met. Since the statuses are implicationally related, this analysis also predicts that a form can appropriately encode the necessary and sufficient status (the status immediately above the form in the table) as well as all higher statuses (statuses to the left). For example, the referent of an NP with the definite article the in English may be just uniquely identifiable, or it may also be familiar, activated, or in focus; and the referent of a demonstrative pronoun may be simply activated or it may also be in focus. We would thus expect forms to be distributed across more than one status in actual discourse. We tested these predictions by investigating the distribution of different forms of reference in naturally occurring discourse for each of the five languages in our study.²⁰ The results of this investigation ¹⁹ Li and Thompson (: ) also note that “the numeral yi ‘one’, if it is not stressed, is beginning to function as a.” ²⁰ Our data comes from a variety of spoken and written sources which differ in formality and degree of planning. These include novels, short stories, magazine articles, news broadcasts, interviews, and casual conversations. In addition, for all languages except Russian, we also analyzed narrative film descriptions which were collected for another study. The methodology here was similar to that of the Pear Stories (Chafe ). Speakers viewed a silent film called The Golden Fish and, immediately after viewing the film, described it to another native speaker of their language.

OUP CORRECTED PROOF – FINAL, 21/1/2019, SPi



COGNITIVE STATUS

Table . Distribution of Chinese forms according to highest status IN FOCUS ACTIVATED FAMILIAR UNIQUE REFERENTIAL TYPE TOTALS Ø Tā Zhè Nèi zhè N nèi N yi N N

 



  

 

 

 









 

 

   

TOTALS

















Table . Distribution of English forms according to highest status IN FOCUS ACTIVATED FAMILIAR UNIQUE REFERENTIAL TYPE TOTALS it HE this that this N that N the N indefinite this N aN



TOTALS



  

      



 



      





 



 







are given in Tables .–., which show the distribution of the different forms according to the highest cognitive status met by that form in the particular context in which it was used.²¹

²¹ Cognitive status involves assumptions that a cooperative speaker can reasonably make regarding the addressee’s knowledge and attention state in the context in which an expression is used. As discussed in §. (particularly note the discussion of examples () and ()), the cognitive status of a referent is not uniquely determined by syntactic structure. This is reflected in the methods we used for analysis. Two trained coders analyzed each transcript. While there were coding guidelines based on syntax and recency of mention, decisions on cognitive status were not completely mechanical, but also involved judgments based on relevance and the shared knowledge and beliefs of the speaker and hearer. The two coders agreed on approximately % of the tokens examined. Most disagreements were between familiar versus activated or activated versus in-focus. We believe this is because the boundaries between statuses involving attention state are not discrete, even though they map onto discrete forms. Disagreements were resolved by discussion among the coders.

OUP CORRECTED PROOF – FINAL, 21/1/2019, SPi



JEANETTE K. GUNDEL, NANCY HEDBERG, AND RON ZACHARSKI

Table . Distribution of Japanese forms according to highest status IN FOCUS ACTIVATED FAMILIAR UNIQUE REFERENTIAL TYPE TOTALS   

Ø kare kore sore are kono N sono N ano N N

  

 

TOTALS

   



   

   







   















Table . Distribution of Russian forms according to highest status. IN FOCUS ACTIVATED FAMILIAR UNIQUE REFERENTIAL TYPE TOTALS Ø ono ONO èto tot èto N to N N TOTALS

 

   



  





































Table . Distribution of Spanish forms according to highest status. IN FOCUS ACTIVATED FAMILIAR UNIQUE REFERENTIAL TYPE TOTALS Ø él (subj) él (nonsubj) éste ése aquél este N ese N aquel N el N un N TOTALS

  

     

    





























 







OUP CORRECTED PROOF – FINAL, 21/1/2019, SPi

COGNITIVE STATUS



As predicted, all but a few occurrences of referring expressions examined met the necessary conditions posited for that expression for the language in question (see Table .) For example, in Spanish, all referents of zero pronominals were in focus; all referents of overt pronouns (both personal pronouns and demonstratives) were at least activated;²² all phrases with the demonstrative determiner ese were at least familiar; all phrases with the definite article el were at least uniquely identifiable; and all phrases with the indefinite article un were at least type identifiable. When the referent of a particular form does not have the required cognitive status, the result may be infelicitous, as in (), where the addressee was not able to identify the referent of the pronoun these because it wasn’t activated. () M: These. Do these go in here or there? K: These? M: The ones I just got done writing.23 [Fredrickson tapes] A more typical result of using a form whose referent does not meet the required status is that the addressee accommodates and is able to associate the correct referent with the form in spite of the fact that it was used inappropriately. This is illustrated in (), the one instance of an unstressed personal pronoun in our English data where the referent is not in focus. () K: N: K: N: K: N: K:

Barb got it. Catmopolitan? Yeah. Catmopolitan. She got it. Yup. I suspicion she was a cat in her other life. Oh did I tell you that THEY have a cat, they have two cats; one is Maynard and one’s Dudley. [Frederickson tapes]

The referent of they in K is Barb and her husband. Although the addressee knows that Barb is married (i.e., the couple is familiar), it is probably not the case that mentioning Barb automatically brought her husband into focus (or that the couple had even been activated in this context). But the reference succeeded nevertheless. The results in Tables .–. thus clearly support the necessary conditions hypothesized for the forms in Table .. Moreover, we found that forms were not only used in coding noun phrases whose referents met the minimal required status; they were also ²² Because we had tapes for only a small portion of our data, we were not able to distinguish stressed from unstressed pronouns in most cases. Note, however, that the referents of all but a very few personal pronouns in any of the languages were in focus. Either a stressed or an unstressed personal pronoun would therefore have been appropriate. ²³ This example and some others in the paper are drawn from data outside the study reported on in Tables .–.. The ‘Frederickson tapes’ are transcribed recordings of conversations during family gatherings, collected by Karen Frederickson of the University of Minnesota Linguistics Department in –.

OUP CORRECTED PROOF – FINAL, 21/1/2019, SPi



JEANETTE K. GUNDEL, NANCY HEDBERG, AND RON ZACHARSKI

used in coding higher statuses. For example, in English there are some tokens of the N for all statuses to the left of, and including, uniquely identifiable, and there are some tokens of the demonstrative pronoun that for both activated and in focus. Similarly, in Russian, bare nouns with no preceding determiner are found for all statuses; and there are tokens of the demonstrative determiner èto for referents that are activated, as well as for those that are familiar but not activated. This is as expected since each of the cognitive statuses in the Givenness Hierarchy entails all lower statuses, and a particular status thus implies the possibility of reference with forms associated with lower (entailed) statuses.²⁴ However, the distribution across statuses varies considerably for different forms. For example, we found a relatively high occurrence of phrases with the definite article in English and Spanish and of a bare noun in the other languages for all statuses that imply uniquely identifiable. By contrast, demonstrative pronouns are rarely used for referents that are in focus, and demonstrative determiners are rarely used for referents that are familiar. Similarly, there were no occurrences of indefinite articles for statuses higher than referential, even though all statuses meet the necessary conditions for the use of this form. What then explains the choice among forms when necessary conditions for the use of more than one form are met? This question will be addressed in the next and final main section of the chapter.

. T H E G I V E N N E S S H I E R A R C H Y A N D GRICE’S MAXIM OF QUANTITY

.................................................................................................................................. Since each of the cognitive statuses in the Givenness Hierarchy entails all lower statuses, a particular form can often be replaced by forms which require a lower status. For example, the proximal demonstrative determiner these, which requires that its referent be activated, could be replaced by those, which requires only familiarity, by the, which requires only unique identifiability, and sometimes even by an indefinite article (or zero if it is plural), which requires only type identifiability. This is illustrated in (), where these systems was the form used in the cited magazine article, but any of the other alternatives would also have been acceptable in this context.

²⁴ The fact that all occurrences of the Chinese distal demonstrative determiner nèi were at least familiar may be just an accident of the data we examined; or it may be that the use of nèi for referents that are not familiar is relatively rare in Chinese, even though necessary conditions for this form are the same as for the definite article in English and Spanish (see §.). In any case, these findings are not inconsistent with the claim that the determiner nèi requires the referent to be uniquely identifiable, but not necessarily familiar.

OUP CORRECTED PROOF – FINAL, 21/1/2019, SPi

COGNITIVE STATUS



() These incredibly small magnetic bubbles are the vanguard of a new generation of ultradense memory-storage systems. 8 9 These systems > > > > < = Those systems > > > The systems > : ; New generation ultradense memory storage systems are extremely rugged: they are resistant to radiation and are nonvolatile. (Graff ) As noted in §., however, the distribution of forms across statuses which meet necessary conditions for their appropriate use is not random. And some forms rarely occur, even when necessary conditions for their use have been met. Thus, as seen in Tables ., ., and ., all noun phrases in our data coded with the indefinite article in Chinese, English, and Spanish are at most referential or type identifiable, even though the status ‘type identifiable’ is entailed by all other statuses on the hierarchy and therefore all statuses meet necessary conditions for use of the indefinite article. Similarly, entities in focus are generally coded by unstressed pronouns or by zero, even though a demonstrative pronoun in all five languages requires only activation, and any entity that is in focus is necessarily also activated. Moreover, a given form is often inappropriate, or conveys some special effect, even when necessary conditions for its use have been met. Thus, it in (bi) is naturally interpreted as referring to the topic, Simplified English, which is in focus at this point in the discourse. But if it is replaced by this, as in (bii), the most natural interpretation is one where this refers to the whole statement about Simplified English. () a. Simplified English disallows the use of passive, progressive, and perfective auxiliary verbs, among other things. bi. It requires engineers to break up long compound nouns and technical expressions into chunks of three or less elements. [message from electronic news group] bii. This requires engineers to break up long compound nouns and technical expressions . . . In the remainder of this section, we will argue that such facts involve conversational implicatures which result from interaction of the Givenness Hierarchy with Grice’s maxim of quantity, stated in ().²⁵

²⁵ Following Grice (), we use the term ‘conversational implicature’ to refer to non-deductive inferences that arise when the maxims of conversation are being observed. These are, of course, not restricted to conversation, but are assumed in all cooperative uses of language, spoken and written. There has been important work which aims to reduce one or both parts of the quantity maxim to more general principles (cf. e.g., Atlas and Levinson ; Horn ; and Sperber and Wilson ). We have returned to Grice’s original formulation because we find it adequate for explicating the facts at issue here (see also Levinson ). However, we highly recommend the more recent work for essential insights into ways in which the two parts of the quantity maxim interact (e.g., Horn’s Division of Pragmatic Labor).

OUP CORRECTED PROOF – FINAL, 21/1/2019, SPi



JEANETTE K. GUNDEL, NANCY HEDBERG, AND RON ZACHARSKI

()

Maxim of Quantity (Grice ): Q: Make your contribution as informative as required (for the current purposes of the exchange). Q: Do not make your contribution more informative than is required.

The classic examples that give rise to quantity implicatures are ones which form an implicational scale (see Horn , ; Hirschberg ). In Q-based implicatures, use of a weaker (entailed) form conversationally implicates that a stronger (entailing) form does not obtain. For example, while () entails (and is therefore consistent with) (), a speaker who utters () normally implicates the negation of (), that is, she implicates that she does not agree with everything the addressee said. ()

I agree with all of what you said.

()

I agree with some of what you said.

In Q-based implicature, by contrast, use of a weaker (entailed) form implicates a stronger (entailing) form. For example, () entails (), but not vice versa. However, a speaker who utters () normally intends to convey (), that is, she implicates that she will go to the meeting only if the addressee goes. ()

I’ll go to the meeting if and only if you go.

()

I’ll go to the meeting if you go.

Atlas and Levinson () propose that forms which give rise to such implicatures are usually ones where the meaning associated with the stronger form is stereotypical. Thus, since conditional relations discussed in everyday discourse are stereotypically bi-conditional, the bi-conditional form (if and only if ) would be “more informative than is required”. Since the statuses in the Givenness Hierarchy form an implicational scale, we find, as expected, that use of forms which overtly signal different cognitive statuses also gives rise to quantity implicatures.

.. Q As noted above, we found no examples of the indefinite article in Chinese, English, or Spanish for statuses above referential, even though all statuses meet necessary conditions for this form. Moreover, use of an indefinite article typically implicates by Q that the referent is not uniquely identifiable and hence also not familiar, activated, or in focus.²⁶ Thus, while we claim that the conventional meaning of the indefinite article ²⁶ Grice himself observed that “Anyone who uses a sentence of the form ‘X is meeting a woman this evening’ would normally implicate that the person to be met was someone other than X’s wife, mother, sister, or perhaps even a close platonic friend” (: ). In other words, even though Grice does not

OUP CORRECTED PROOF – FINAL, 21/1/2019, SPi

COGNITIVE STATUS



(what it explicitly signals) is simply that the referent is at least type identifiable, use of this form conversationally implicates by Q that the addressee cannot uniquely identify the referent. Since conversational implicature (unlike entailment) is not a necessary inference, our analysis correctly predicts that the referent of an indefinite noun phrase can be uniquely identifiable, or even familiar, in some contexts. For example, in () the definite article (the form actually used in the original citation) is appropriate in referring to the $. million Host Senate office building because the latter is uniquely identifiable. However, since the descriptive information necessary to identify the referent is fully encoded in the nominal itself, an explicit marker of identifiability is unnecessary and an indefinite article would have been equally appropriate here. () The senate on August  voted  to  to spend $, for a third Senate Gymnasium due to be built in THE/A . MILLION-DOLLAR HOST SENATE OFFICE BUILDING OPENING IN JANUARY. (U.S. News and World Report,  August , ) Similarly, as discussed in Dahl , the phrase a doctor in () does not introduce a new entity into the discourse. Since it is the property of being a doctor, and not the identity of this particular doctor, that is relevant here, an indefinite article can be used without implicating that the referent is not familiar. If non-identifiability (and therefore non-familiarity) were part of the conventional meaning associated with the indefinite article, () would necessarily have the interpretation that the speaker believes exercise helps because she heard it from someone other than Dr. Smith. () Dr. Smith told me that exercise helps. Since I heard it from inclined to believe it.

A DOCTOR,

I’m

Our analysis also predicts correctly that the interpretation of non-identifiability (or non-familiarity) associated with the indefinite article can be cancelled without contradiction, as in (), and reinforced without redundancy, as in (). () I met A STUDENT before class. A STUDENT came to see me after class as well—in fact it was the same student I had seen before. () But forged provenance papers still did not mean that the kouros was fake . . . The Getty decided that the fake documents were not reason enough to ask Mr. Becchina, the Basel dealer who had sold the kouros, to take back the sculpture. (Attempts by The Times to reach Mr. Becchina were unsuccessful.) Then last April, an independent scholar in London, Jeffrey Spier, was shown a

explicitly state it in these terms, use of an N implicates that the referent is not someone uniquely identifiable to the addressee. Note, however, that the indefinite article does not implicate nonreferentiality. This is so because with the exception of indefinite this in colloquial English, the languages we examined have no separate form that signals referentiality. The indefinite article would thus be the strongest possible form for coding something that is referential, but not uniquely identifiable.

OUP CORRECTED PROOF – FINAL, 21/1/2019, SPi



JEANETTE K. GUNDEL, NANCY HEDBERG, AND RON ZACHARSKI

photograph of a fake torso of a kouros, belonging to A BASEL DEALER (NOT MR. BECCHINA), that looked similar to the Getty’s sculpture. [The New York Times,  August , :)²⁷ We also find Q operating in the coding of noun phrases whose referents are in focus. As Tables .–. show, the number of in-focus referents coded by the strongest, most restrictive possible form (zero or unstressed pronoun) ranges from % (Russian) to % (English); and a demonstrative pronoun was rarely used in such cases.²⁸ In English, for example,  of the  noun phrases that were in focus were unstressed personal pronouns, and only one was a demonstrative pronoun. Similarly, in Japanese eighty-seven of the  noun phrases whose referents were in focus were coded by a zero pronominal, and only five by a demonstrative pronoun. This is so even though all forms meet the necessary conditions for coding an in-focus referent since the status ‘in focus’ entails all other statuses. Moreover, use of a demonstrative pronoun, which requires only activation, often implicates by Q that the referent is not currently in focus; that is, it signals a focus shift (cf. Isard ; Linde ; Bosch ; Sidner ).²⁹ Some illustrations of the focus-shift function of pronominal demonstratives are given in ()–().

²⁷ Example () is adapted from Hawkins (: ), who argues, as we do here, that use of the indefinite article conventionally implicates the negation of what is conventionally signaled by the definite article. Hawkins explicitly defines the conventional implicature of the as follows: “the conventionally implicates that there is some subset of entities, [P], in the universe of discourse which is mutually manifest to S and H online and within which definite referents exist and are unique” (: ). We believe that the extension of this definition is essentially equivalent to what we mean by ‘uniquely identifiable’. Example () has another interpretation where the phrase a Basel dealer is used to express the inability of the London scholar to identify the dealer as Mr. Becchina. As pointed out to us by an anonymous reviewer, the expression not Mr. Becchina does not reinforce an implicature for the reader on this interpretation. According to the native speakers we consulted, the facts illustrated in ()–() are similar in Chinese and Spanish. It is tempting to propose that non-identifiability and non-familiarity is conversationally (not conventionally) implicated by the indefinite article in all languages which have such an article. This would be the case if articles, personal pronouns, and demonstratives conventionally signal a cognitive status on the Givenness Hierarchy in all languages, as they do for the languages we have investigated here. In any case, this is an interesting empirical claim which is worthy of further investigation. ²⁸ Although all the languages but English are ‘pro-drop’, we find some interesting differences among them in the distribution of pronoun versus zero. While % of the referents in focus were coded with zero in Japanese, Chinese had only % zeroes, and Russian had only %. The preferred form for referents in focus in Chinese and Russian appears to be unstressed personal pronouns (% in Chinese and % in Russian.). In Spanish, which allows zero only in subject position, % of in-focus pronominal subjects were zeroes, and % were overt subject pronouns. The relatively high number of zero pronominals in Japanese is predicted by our analysis, since only a zero argument requires the referent to be in focus in this language (i.e., Japanese has no unstressed personal pronouns). The facts in the other languages are consistent with our analysis, but we have no explanation for the differences in frequency of zero pronominals. ²⁹ Since the topic of the immediately preceding utterance is always in focus at the beginning of the next utterance (see §.), a shift in topic is generally also a focus shift.

OUP CORRECTED PROOF – FINAL, 21/1/2019, SPi

COGNITIVE STATUS

() Chinese: a. Xiăo háizi hěn gāoxing. Suŏĭ tā small child very happy so he jīnyú de yīzhī huā fàng zài goldfish NOM one flower put in ‘(The) small child was very happy. So (the) goldfish into (the) bowl.’



bă tā dài gĕi hóng he bring give red yúgāng lĭmiàng. bowl LOC he put a flower he had brought for OM

shì zhèi-ge gùshi. b. ZHÈI jiù this then be this-CLF story ‘This then is the (lit. ‘this’) story.’ () Japanese: a. Toori e dete shibaraku hashitteku. street to go.out for.some.time run ‘He goes out onto the street and runs for some time.’ b. To nanka yatai mitatna omise ga atte. and something stall seem shop NOM be ‘There is a shop like a stall.’ c. KARE wa sono omise no toko e itte. TOP that shop GEN place to goes he ‘He goes to that shop.’ () K: : : A: K:

And . . . So what he did was . . . came in, set up the tree . . . And then he made wassail, with rum in it? And made it in coffee cans and heated it on the stove in the graduate lounge. Oh, gee. And THIS was the solstice tree. [Frederickson tapes]

() When Snepp makes a speech he has to submit a text to CIA censors first. When he wrote a book review for the Los Angeles Times, he had to show it to the agency before he sent it to the newspaper, and when the editor asked for a change, he had to show THAT to the censors too. [Anthony Lewis, Secrecy policy has no sense. Minneapolis Star and Tribune //] () Anyway going on back from the kitchen then is a little hallway leading to a window, and across from the kitchen is a big walk-through closet. On the other side of THAT is another little hallway leading to a window . . . [personal letter] In the Chinese example in (), the proximal demonstrative pronoun zhèi in (b) implicates that focus has shifted from the boy—who is the center of attention at the end of (a)—to the story. In the Japanese example in (), focus shifts from the boy in (a) to the shop in (b). Thus, when the boy is reintroduced in (c), the pronoun kare (which, like demonstrative pronouns, requires only activation) shifts focus back to the boy again. In (), the tree has been activated

OUP CORRECTED PROOF – FINAL, 21/1/2019, SPi



JEANETTE K. GUNDEL, NANCY HEDBERG, AND RON ZACHARSKI

by its mention in Kl, thus licensing the use of a pronominal in K. However, since the tree is not in focus in K, K, or A, reference to it in K constitutes a focus shift and thus requires a stressed demonstrative form. The fact that demonstrative pronouns, unlike unstressed personal pronouns and zero, can be used to refer to something that is activated, but not currently in focus, follows from the necessary conditions posited for appropriate use of these forms. What is of interest here is that demonstratives not only don’t require the referent to be in focus, but often implicate that the referent is not currently in focus. This is particularly striking in cases where there is potential ambiguity of reference. For example, this in () would not normally be interpreted as the currently in-focus wassail; the referent of that in () is interpreted as something other than the first version of the book review; and the referent of that in () is not understood as the kitchen. If that in () were replaced by it, however, the kitchen would be the only possible referent since it is in focus here. We believe that use of pronominal this and that in referring to previous statements (cf. Webber ) is just a special case of focus shift, since the focus of attention at the point after a statement is made is typically not the event or state of affairs described by that statement, but rather the entity which is the topic of the statement. Thus, in (bi) above, use of it continues the topic and refers to Simplified English; but use of this in (bii) implicates that the referent is not in focus, and is interpreted as referring to the whole statement about what Simplified English disallows.³⁰

.. Q We have seen in §.. that the choice among forms when the necessary cognitive status for more than one form is met can be partly explained in terms of the first part of the quantity maxim—be as informative as required. Interaction of this maxim with the Givenness Hierarchy correctly predicts that (a) an indefinite article will normally not be used for referents that are uniquely identifiable, since this form explicitly signals

³⁰ While we don’t necessarily want to claim that interaction of cognitive status and implicature account for all functions of particular forms of referring expression, we believe that at least some frequently discussed functions of particular forms may insightfully be explained in terms of focus shift. The tendency for overt (stressed) pronouns in pro-drop languages to signal emphasis contrast, or the opening of a new thematic unit, the tendency for use of a distal demonstrative to signal closure of a thematic unit, and the tendency of demonstratives to signal contrast may all be subcases of focus shift, In (i), for example, the speaker could have referred to the currently in focus travel journal with pronominal it. Use of this implicates by Q that the referent should be viewed in a new way, thus implicitly bringing into focus other travel journals with which it might be contrasted. (i)

I’ve been working some more on my book, on my travel journal from ’. That sounds funny, but I have all my notes. However, I want to write a little better than I am, put a little more effort, make THIS a little more quality effort than the last one was. [Frederickson tapes]

OUP CORRECTED PROOF – FINAL, 21/1/2019, SPi

COGNITIVE STATUS



only type identifiability, and (b) for referents that are in focus, an unstressed personal pronoun or zero, which explicitly delimits the set of possible referents to those that are in focus, will normally be chosen over a demonstrative pronoun, which gives less information about cognitive status because it only requires that the referent be activated. However, as is evident from Tables .–., use of a particular form doesn’t always conversationally implicate that necessary conditions for a form requiring a higher status don’t obtain. Over % of full noun phrases whose referents were at least familiar were introduced by a definite article in English and Spanish, and by a bare noun in Russian, even though necessary conditions for a demonstrative determiner (the stronger, more restrictive, form) were met. While demonstrative determiners occurred somewhat more frequently in Chinese and Japanese, a bare noun (the weaker form) was also common for full familiar NPs in these languages (% in Chinese and % in Japanese).³¹ Thus, use of a definite article or a bare noun clearly doesn’t implicate by Q that the referent is not familiar; rather, it is the second part of the Quantity Maxim—don’t be more informative than required—that is relevant here. The question that naturally arises is: why should choice among definite determiners be primarily dictated by Q when, as just argued, choice among definite pronominals is primarily dictated by Q? We believe the answer to this question is related to the fact that pronominal forms have little if any descriptive content, so that information about cognitive status is crucial in delimiting the set of possible referents. For full noun phrases, however, signaling identifiability is often sufficient for identifying the referent, given the descriptive content of the noun and its modifiers, and an explicit signal of a more restrictive cognitive status is therefore unnecessary. Moreover, since familiarity is the most common basis for identifiability, the Q implicature here would follow from the observation that the second part of the quantity maxim induces stereotypical interpretations (cf. the observation in Atlas and Levinson  noted above.) Since most references which are uniquely identifiable in a discourse are also at least familiar, explicitly signaling a status higher than uniquely identifiable would be more informative than required. Thus, while some researchers consider familiarity to be part of the conventional meaning of the definite article and treat the non-familiar cases as exceptional (cf. Heim ), we propose that the definite article conventionally signals only that the referent is uniquely identifiable and that familiarity is conversationally implicated by Q. Since the definite article explicitly signals that the speaker expects the addressee to uniquely identify the referent, the more restrictive cognitive statuses

³¹ The higher frequency of demonstrative determiners in Chinese and Japanese may be due partly to the fact that these languages lack a separate form for the definite article. It is of interest, however, that Russian—which also lacks a definite article—patterns more like English and Spanish here, with demonstratives accounting for less than % of the full definite NPs. The fact that the highest number of demonstrative determiners was found in the Chinese data would appear to support observations that the demonstrative determiner is beginning to function like a definite article in Chinese (see §..). But notice that it is the proximal rather than the distal demonstrative that occurs most frequently. Our data thus suggest that both demonstratives are beginning to function like a definite article in Chinese, with the proximal form predominating for referents that are at least activated.

OUP CORRECTED PROOF – FINAL, 21/1/2019, SPi



JEANETTE K. GUNDEL, NANCY HEDBERG, AND RON ZACHARSKI

associated with demonstrative determiners often have little information value and do not need to be signaled explicitly. This is especially true when the referent has just been introduced by a phrase which is at least partially identical in form, as in the following examples from English and Spanish.³² ()

‘How in the world,’ demanded Harriet, ‘did you get here?’ ‘Car,’ said Lord Peter, briefly. ‘Have they produced the body?’ ‘Who told you about THE BODY?’ [Dorothy Sayers. Have His Carcase. : ]

()

Spanish: Y él atrajo mucho al pez, y él miraba AL PEZ, and he attracted a.lot to.the fish and he looked at.the fish EL PEZ miraba a él. y and the fish looked at him ‘And he was attracted to the fish, and he looked at the fish, and the fish looked at him.’

Examples ()–() illustrate the use of a bare noun for an activated referent in Chinese, Japanese, and Russian. ()

Chinese: Zhèi shíhou qiàqiăo jīnyú yĭjīng; hé niăor zài yóuxì this time so.happen goldfish already with bird at play yĭjīng tiào-dào le Zài yóuxi-de shíhou JĪNYÚ goldfish already jump ASP at play-NOM time z.huōmiàn-shàng. table-top ‘At this time, (it) so happened that (the) goldfish was already playing with (the) bird. At (the) time (they) were playing, (the) goldfish had already jumped onto the table top.’

()

Japanese: De sono kake de otokonoko ga katte. NOM win and that gambling in boy Akai kingyo go sono otokonoko no toki wa kingyo ga red goldfish NOM that boy GEN time TOP goldfish NOM kantan.ni patto toreta to iu kanjji de easily rapidly caught Q say seem be de OTOKONOKO wa yorokonde uchi e modotte TOP happily home to go.back and boy

³² Notice that Q can be invoked here only to explain why the definite article is used instead of a demonstrative determiner. It cannot explain why a full NP rather than a pronominal is used in the activated cases. Some possible reasons for this including ambiguity resolution and global focus shift were discussed by Marslen-Wilson et al. ; Guindon ; and Fox .

OUP CORRECTED PROOF – FINAL, 21/1/2019, SPi

COGNITIVE STATUS



iku-tte iu bamen ga atte NOM is go-COMP say scene ‘And the boy wins the bet. It seems that the red goldfish is easily caught in that boy’s turn: and there’s a scene where the boy goes home happily.’ () Russian: Ètot kot poterjal vsjakuju sovest’. KOT—brodjaga i bandit. this cat lost every conscience cat vagabond and bandit ‘This cat has lost all conscience. The cat is a vagabond and a bandit.’ Since Q predicts use of the weakest possible form for full definite NPs, when demonstrative determiners do occur there is often a good reason for conveying the stronger cognitive-status information. For example, a demonstrative determiner that requires familiarity but not activation (the proximal form in Russian and the distal form in the other languages) often facilitates long-term memory for a familiar referent. In such cases, which we refer to as ‘reminder that’, the determiner that explicitly signals that the referent is familiar and conversationally implicates by Q that the referent is not activated. Examples of ‘reminder that’ in English and Russian are given in () and ().³³ () Exxon Oil claims it will take several million dollars to clean up THAT OIL SPILL OFF THE COAST OF ALASKA. [beginning of radio newscast] () Russian: A sošli my na ostanovke kotoraja nazyvalas’ sorok-vtoroj and descended we at stop which was.called forty-second kilometr, a spustilis’ v ÈTOT KAN’ON. kilometer and went.down in this canyon ‘And we got off at a stop which was called forty-second kilometer, and (we) went down into that (lit. ‘this’) canyon.’ Redescriptions of an activated referent provide a compelling example of a stronger than strictly necessary form being required. Determiner this which requires activation, becomes crucial in identifying the referent in such cases by serving as an explicit signal that the referent has already been activated. Examples of such uses of this N in English, which appear to be restricted to more formal, written genres, are given in ()–(). () Nearly lost in the polemic was Judge Kennedy himself. That was ironic, because in many ways THIS FORMER SMALL-CITY LAWYER with the stable marriage and three attractive children and the fine reputation appears to personify just those values that made the image of Ronald Reagan so attractive after the convulsions of the ’s and ’s. [New York Times //, : ]

³³ Other special effects associated with demonstratives, such as emotional uses discussed, for example, in Lakoff , may also be attributed to quantity implicatures.

OUP CORRECTED PROOF – FINAL, 21/1/2019, SPi



JEANETTE K. GUNDEL, NANCY HEDBERG, AND RON ZACHARSKI

()

Poll Return The attachment feature sends THIS INBOUND TAG to the series/I channel controls to indicate a poll capture for interrupt servicing or nonburst cycle steal servicing. It is not used to signal a burst transfer. [technical document]

()

One valuable outcome of these organizational studies was the refinement of our notions of three different approaches that could be incorporated in an automated message filtering system. We refer to THESE TECHNIQUES as the cognitive, social, and economic approaches to information filtering. Malone et al. ()

As seen in Tables .–., demonstrative forms (both pronouns and determiners) are relatively infrequent in the five languages we investigated. Similar observations have been made by other investigators (cf. e.g., Ariel ). The analysis we have proposed here provides an explanation for why this should be the case. Since demonstrative pronouns require only activation, they signal a weaker, less restrictive cognitive status than unstressed personal pronouns or zero, which require the referent to be in focus. Demonstrative pronouns are thus less informative than are unstressed personal pronouns, because anything which is in focus is also activated, but not vice versa. And since demonstrative determiners (other than the distal demonstrative in Chinese) require the referent to be at least familiar, they signal a stronger, more restrictive cognitive status than the definite article or zero determiner, which require only identifiability. Demonstrative determiners are thus more informative than the definite article or zero determiner, because anything which is familiar is also (uniquely) identifiable, but not vice versa. The application of Q (give as much information as necessary) for definite pronouns and Q (don’t give more information than necessary) for full definite NPs thus conspires to result in relatively low frequency of demonstratives, both pronoun and determiner, in natural language discourse.

. C O N C L U S I O N

.................................................................................................................................. We have proposed that six implicationally related cognitive statuses are relevant for describing speakers’ ability to appropriately use and interpret different forms of reference in natural language discourse. We have shown that each of the statuses is a necessary and sufficient condition for the use of one or more different forms, and that interaction of these form–status correlations with the Gricean maxim of quantity allows us to account for facts regarding the actual distribution of different forms of reference (both within and across languages) which remain unexplained in previous analyses. The form of referring expressions is only one of a number of linguistic phenomena which have been shown to depend on factors relating to speakers’ assumptions about

OUP CORRECTED PROOF – FINAL, 21/1/2019, SPi

COGNITIVE STATUS



the addressee’s knowledge and attention state. Others include intonation, topic/focus marking particles, and a wide range of syntactic structures. (See, e.g., Prince ; Lambrecht ; Sgall et al. ; Gundel ; Kuno ; Givón ; Hedberg ; Rochemont and Culicover ; Steedman ; Ward et al. ; and other references cited in Green : –.) We hope that the theory of cognitive statuses outlined in this chapter will contribute to more adequate and insightful analyses of these phenomena as well.

ACKNOWLEDGMENTS This chapter first appeared as a paper in Language, vol. , No. , , and is being reprinted with permission of the Linguistic Society of America. We gratefully acknowledge the University of Minnesota Graduate School for a Grant-inAid and Faculty Summer Research Grant to Jeanette Gundel and Cornell University for a Mellon Postdoctoral Fellowship to Nancy Hedberg. Ron Zacharski was partly supported by UK Economic and Social Research Council grant #R   to Edinburgh University. This chapter represents a revised and expanded account of research introduced in Gundel, Hedberg, and Zacharski , , . We are indebted to Thorstein Fretheim, Alison Marchant, Gerald Sanders, Candace Sidner, and two anonymous referees for their critical comments and suggestions. Special thanks also to Ann Mulkern, Suellen Rundquist, and Karen Schaffer for their work on various stages of this project and to Xinghui Chen, Karen Frederickson, Hui-chuan Huang, Svetlana Krylova, Sahya Luck, Yoshiko Matsumoto, Silas Oliveira, Satoko Suzuki, and Zhu Yao for help with data collection, translation, and analysis. The authors’ names appear in alphabetical order.

OUP CORRECTED PROOF – FINAL, 17/1/2019, SPi

CHAPTER



......................................................................................................................

DIFFERENT SENSES OF ‘REFERENTIAL’

...................................................................................................................... NANCY HEDBERG, JEANETTE GUNDEL, AND KAJA BORTHEN

. I N T R O D U C T I O N

.................................................................................................................................. T HERE exist a range of different notions of referentiality in the literature. Although meaningful and clearly defined in and of themselves, it is not always clear how the various notions differ and how they might be seen as intersecting. The goal of the present chapter is to review these different notions, focusing on application of the Givenness Hierarchy theory of Gundel, Hedberg, and Zacharski () to various linguistic phenomena sensitive to referentiality. In particular, we will be concerned with the interaction between the cognitive status ‘referential’, one of six cognitive statuses postulated by the Givenness Hierarchy, and related categories proposed in the literature. In the next section, we introduce the Givenness Hierarchy. In the following sections, we discuss types of linguistic phenomena that are sensitive to some sense of referentiality.

. T H E G I V E N N E S S H I E R A R C H Y

.................................................................................................................................. According to the Givenness Hierarchy Theory (Gundel et al.  and subsequent work), nominal expressions encode two kinds of information: descriptive/conceptual information about the speaker’s intended referent/interpretation; and procedural information about where in memory a representation of the associated referent can be found in the mind of the addressee, or whether and how to represent it if it is not already in memory—its ‘cognitive status’. The latter is encoded by various determiners or pronouns. The six cognitive statuses assumed to be relevant in this sense across

OUP CORRECTED PROOF – FINAL, 17/1/2019, SPi

DIFFERENT SENSES OF ‘REFERENTIAL’



languages are represented in (), along with the respective English forms which encode these statuses: ()

The Givenness Hierarchy In focus > Activated > Familiar > it, she1

that, this, SHE, this N

that N

Uniquely > Referential > Identifiable the N Indefinite this N

Type Identifiable aN

The way to read this table is that each determiner or pronominal form encodes the cognitive status under which it is placed, as part of its conventional meaning. For instance, whereas in English an unstressed personal pronoun such as she can in principle apply to an infinite number of feminine singular entities, the procedural information that the intended referent is in the addressee’s current focus of attention narrows down the set of candidates to a very small number, often just one. The indefinite article a, on the other hand, encodes only the status ‘type identifiable’, indicating that the addressee is to associate an interpretation of the type described by the conceptual content of the nominal expression. One fundamental property of the Givenness Hierarchy is the fact that the statuses are in a unidirectional entailment relation, by definition. Thus, if an entity is in focus of attention, it is necessarily also activated (in short-term/working memory); if it is activated, it is necessarily also familiar (in memory, possibly long-term memory); if it is familiar it is also uniquely identifiable (the addressee is expected to be able to associate a unique representation), and so on. This means that all nominal forms except the ones that explicitly encode the status ‘in focus’ are predicted to be appropriate for more than one cognitive status. The definite article the, for instance, signals explicitly that the addressee can associate a unique representation with the referent of the definite article phrase, and the form can thus be used with any one of the four more restrictive cognitive statuses, including ‘in focus’. This is so because if the referent has the cognitive status ‘in focus’, it is necessarily also activated, familiar, uniquely identifiable, referential, and type identifiable—which means that the necessary condition for the definite article, ‘uniquely identifiable’, is met. This is why, for example, a phrase headed by the definite article, which is not overtly specified for the status ‘in focus’, can be used in the second sentence of (), even though the man can be expected to be in focus and an unstressed pronoun (he) would have been acceptable as well. ()

A tall, handsome man came into the restaurant. Suddenly the man shouted.

The unidirectional entailment relation between statuses on the Givenness Hierarchy predicts that there will be a one-to-many correlation between forms and cognitive statuses; but there are still constraints on individual forms imposed by the minimal ¹ The forms it and she stand for all unstressed personal pronouns here. Although we assume that pronouns are determiners, following the DP hypothesis (Abney ), differing only in that pronouns lack an NP complement, we will continue to use the more traditional term ‘pronoun’ for convenience.

OUP CORRECTED PROOF – FINAL, 17/1/2019, SPi



NANCY HEDBERG, JEANETTE GUNDEL, AND KAJA BORTHEN

status signaled by that form. For instance, if the intended referent of a nominal expression is at most referential, that is, not uniquely identifiable or higher, the necessary condition on the use of the definite article the is not met, and the prediction is that the cannot be used felicitously in the given context. Consider the examples in (): ()

a. Yesterday, I went to a pub in Minneapolis, and after a while this strangelooking guy entered the room. b. Yesterday, I went to a pub in Minneapolis, and after a while the strangelooking guy entered the room.

Whereas the indefinite determiner this² can be used in (a) in a case where no strangelooking guy has previously been mentioned or can otherwise be uniquely identified by the addressee, the definite article the cannot. These facts follow from the association between forms and statuses in () above and the definitions of the cognitive statuses ‘referential’ and ‘uniquely identifiable’ below: ()

UNIQUELY IDENTIFIABLE: The addressee can identify the speaker’s intended referent on the basis of the nominal alone. (Gundel et al. : )

()

REFERENTIAL: The speaker intends to refer to a particular object or objects. To understand such an expression, the addressee not only needs to access an appropriate type representation, he must either retrieve an existing representation of the speaker’s intended referent or construct a new representation by the time the sentence has been processed. (Gundel et al. : )

The main difference between the statuses ‘uniquely identifiable’ and ‘referential’ is the following: The instruction for a nominal form encoding the status ‘uniquely identifiable’ is for the addressee to associate (retrieve or establish) a unique representation of the referent based on the nominal alone.³ For a form encoding the status ‘referential’, on the other hand, the addressee is only instructed to retrieve or construct a representation of the referent by the time the whole sentence has been processed.

² The Givenness Hierarchy treats indefinite this, which explicitly signals only referentiality in English, and proximal demonstrative determiner this, a form which explicitly signals activation, as separate lexical items; but the function of indefinite this perhaps reflects its historical origins as a proximal demonstrative determiner. ³ Barker () correctly notes that the referent of a definite pronoun cannot be uniquely identified on the basis of the descriptive content of the pronoun, and views this as an objection to the claim that anything in focus is also uniquely identifiable. But note that the definition of ‘uniquely identifiable’ does not require that the referent be identified on the basis of the descriptive content of the pronominal form. Although the distinction between conceptual and procedural information encoded by nominal forms was not made as explicit in Gundel et al. (), as it was in later work (e.g., Gundel , and Gundel, Hedberg, and Zacharski ), the Givenness Hierarchy claim has always been that the unidirectional entailment holds between the cognitive statuses of the mental representations of referents, not between the forms that encode these statuses. Thus, since unstressed definite personal pronouns in English conventionally encode the information that the referent is in focus, this allows the referent to be uniquely identified, as there is frequently no more than one entity in the addressee’s focus of attention.

OUP CORRECTED PROOF – FINAL, 17/1/2019, SPi

DIFFERENT SENSES OF ‘REFERENTIAL’



Since the description ‘strange-looking guy’ is not sufficient to establish a unique representation in the given context, the prediction is that the definite article is not appropriate here. The indefinite determiner this, on the other hand, is predicted to be acceptable, since the addressee will be able to establish a representation of the referent by the time the sentence has been processed, in this case a mental representation of the strange-looking man that entered the pub the day before when the speaker was there. Another cognitive status distinction that will be important in this chapter is ‘type identifiable’. The cognitive status ‘type identifiable’ is defined as follows: ()

TYPE IDENTIFIABLE: The addressee is able to access a representation of the type of object described by the expression. (Gundel et al. : )

All that is signaled by a form that encodes the status ‘type identifiable’ is that the addressee is expected to access or construct a representation of the type of entity described. Due to the unidirectional entailment of statuses on the Givenness Hierarchy, the indefinite article in English, which encodes the status ‘type identifiable’, is predicted to be able to be used for referential entities as well as entities that are at most type identifiable, since anything that is referential is type identifiable.⁴ Whether or not an indefinite expression is interpreted as referential or merely type identifiable will depend on the content of the rest of the clause and other contextual factors that may reveal the likelihood of the speaker intending to refer to a particular entity. The italicized phrases in () will most likely be interpreted as merely type identifiable, not referential. ()

a. I don’t have a VCR and neither does my neighbor. b. Peter is a doctor.

In the examples in (), the speaker most likely intends the addressee to only associate an appropriate type representation of a VCR and a doctor. Thus, as expected, if one substitutes the indefinite determiner this for the indefinite article a in these cases, the interpretation changes significantly. In () below, on the other hand, the indefinite phrases are most likely interpreted as not only type identifiable, but also referential. ()

a. When my youngest child was three or so, we were visiting at the house of a friend of mine and my friend was babysitting her infant nephew. b. I want to tell you about a strange guy I saw today.

In (a), it is evident from the fact that the friend is mentioned subsequently that the speaker intended the addressee to establish a representation of the friend. Similarly, in (b) it follows from the rich description of the referent and the content of the clause that the speaker has a particular individual in mind and intends to continue to talk about him. As expected, these phrases may be substituted by indefinite this, which explicitly encodes the status ‘referential’. ⁴ This explains why an a phrase can be interpreted as either referential or non-referential (or what has sometimes been called ‘specific’ versus ‘non-specific’). See §. and von Heusinger (Chapter  of this volume) for discussion of ‘specificity’.

OUP CORRECTED PROOF – FINAL, 17/1/2019, SPi



NANCY HEDBERG, JEANETTE GUNDEL, AND KAJA BORTHEN

The correlation between forms and cognitive statuses in () makes predictions about acceptable and unacceptable contexts for the various forms. As noted above, a further prediction is that each form is in principle also compatible with all statuses higher than the one it explicitly encodes. However, most forms are used more typically with the status they explicitly encode than with higher ones that entail that status, and the likelihood that a given form will be used for such a higher status depends on more general, pragmatic factors. For instance, the definite article is much more likely than the indefinite article to be used for reference to entities that are uniquely identifiable, and possibly also familiar, activated or in focus (see Gundel et al. ). These facts do not follow from the Givenness Hierarchy itself, but can be given a pragmatic explanation. A form can be used with a higher status than the one it minimally requires if doing so is in accordance with general pragmatic principles. From the point of view of Relevance Theory (Sperber and Wilson ; Wilson and Sperber ), speakers are expected to use the form that leads to adequate cognitive effects for minimal processing effort. To illustrate how this affects choice of nominal form, consider the examples in () below, based on similar examples in Scott (: –). ()

a. b. c. d. e.

John was late for work again. He never learns. John was late for work again. That rascal never learns. John was late for work again. The rascal never learns. John was late for work again. A rascal never learns. John was late for work again. He/that rascal/the rascal/?a rascal doesn’t like to get up early.

In these examples, John is in the focus of attention by the time the first sentence has been processed since he has been referred to in subject position (cf. the coding guidelines of Gundel et al. ), which means that the speaker can use an unaccented personal pronoun to refer to him the second time, as in (a). Doing so requires no extra processing effort since a pronoun explicitly signals that its referent is in focus, and John is likely to be the only male entity that is in focus at the point when the pronoun is encountered. In (b), ‘that rascal’ is also acceptable in this context in reference to John, as anything that is in focus is also familiar (in memory). The reason why the determiner that, which signals the weaker (entailed) cognitive states ‘familiar’ is equally acceptable here, according to a relevance-theoretic view, is that the description ‘that rascal’ gives rise to important information about the speaker’s attitude towards John’s repeated tardiness, and thus has additional cognitive effects compared to the pronoun (cf. Scott : –). Thus, the extra processing effort expended in having to interpret a less informative form with respect to cognitive status, is paid back in terms of extra cognitive effects arising from the descriptive content of the phrase. Similar reasoning applies in (c) and (d), which only instruct the addressee to explicitly assign a unique or appropriate type representation respectively, since the addressee could easily infer, by the Principle of Relevance (Sperber and Wilson : ), that the speaker intends to convey the information that she thinks John is the/a rascal. This is a highly accessible assumption, which also adds additional cognitive effects without undue processing

OUP CORRECTED PROOF – FINAL, 17/1/2019, SPi

DIFFERENT SENSES OF ‘REFERENTIAL’



effort since someone who is repeatedly late for work may reasonably be called a rascal. In (e), by contrast, while it is possible to infer that John is a rascal because he doesn’t like to get up early, it takes more effort to do so and it is also more difficult to achieve positive cognitive effects by relating this to the fact that John was late for work again. In the remainder of this chapter, we discuss how the sense of ‘referential’ on the Givenness Hierarchy subsumes and goes beyond various senses of the notion of ‘referential’ that have been discussed in the semantics and pragmatics literature. For example, since all definite phrases are predicted to be used referentially on the Givenness Hierarchy, the Givenness Hierarchy sense of referentiality subsumes uses that have been termed ‘attributive’ as well as those that have been termed ‘referential’ by Donnellan ().

. R E F E R E N T I A L I T Y

AND DEFINITENESS

.................................................................................................................................. The Givenness Hierarchy takes a uniqueness rather than a familiarity view of the semantics/pragmatics of definite phrases (see also Birner and Ward ; Abbott ; and this volume, inter alia). More specifically, definite nominals are defined in terms of the hierarchy as those whose referents have the cognitive status ‘uniquely identifiable’ or higher, in the sense that the addressee is expected to associate a unique representation of the referent on the basis of the phrase itself, either because he already has a unique representation of it in memory or because he is able to construct a new unique representation, possibly by way of a bridging inference to a recently activated referent (Clark and Haviland ).⁵ The fact that the referent of a definite article phrase in English and many other languages is often also familiar follows from the fact that familiarity implies uniqueness, but it is not a defining property of the definite article. Most non-pronominal definite phrases are marked by the definite article in English, but definite pronouns, possessor-marked phrases⁶, proper names, demonstrative phrases, and generics are also definite in this sense, and obtain a similar semantics/pragmatics.⁷ The

⁵ A nominal phrase resulting from a bridging inference is also known as an ‘associative anaphor’ (Hawkins ), an ‘inferrable’ (Prince b), or an ‘indirect anaphor’ (Erkü and Gundel ). The process of constructing a new representation by way of a bridging inference is also possible for referential indefinites, as in (i): (i) Jane has a new car. She already broke a window. ⁶ By ‘possessor-marked phrases’ we include only those where the possessor is definite. See Abbott (: ). ⁷ The reason why proper names, possessives, and generics are often not marked with a definite article is no doubt related to the fact that they are typically definite in the sense that we have defined here, and a definite article is thus redundant. But many languages, e.g., German for proper names, Norwegian for possessives, and French and Arabic for generics, do mark these forms with a definite article or suffix, either obligatorily or optionally.

OUP CORRECTED PROOF – FINAL, 17/1/2019, SPi



NANCY HEDBERG, JEANETTE GUNDEL, AND KAJA BORTHEN

referents of the italicized phrases in (), which are all definite, thus all have at least the status ‘uniquely identifiable’, and therefore also ‘referential’.⁸ ()

a. b. c. d. e.

I have to go walk the dog. He hasn’t been out all day. My mother is coming to visit this weekend. John won’t be coming to the party tonight. Dogs are mammals. That car is old, but the engine is new.

. O V E R T

MARKERS OF REFERENTIALITY

.................................................................................................................................. As noted above, on the Givenness Hierarchy, the status ‘uniquely identifiable’ entails the status ‘referential’; so all definite references are predicted to be referential. This follows from the definitions of the two statuses. If the addressee can associate a unique representation with the referent of the nominal (uniquely identifiable), at the point after it is processed, then she can associate a unique representation with the referent by the time the whole sentence has been processed (referential). On the other hand, nominal phrases that are not definite can be either referential or nonreferential. In the case of referential indefinites, the speaker has a particular entity ‘in mind’, and expects the addressee to associate a unique representation with that entity by the time she is finished processing the sentence. Data from a variety of languages show that referentiality can be grammatically marked even when the language lacks a means of marking definiteness, that is, unique identifiability. Thus, there are a number of languages that don’t have a definite article, but do explicitly encode referentiality, often called ‘specificity’ in the literature (see Karttunen , ; and von Heusinger, Chapter  of this volume).⁹

.. Differential object marking In some languages, a system of ‘differential object marking’ (Aissen ; von Heusinger ) includes a way of marking a referential object. For example, in ⁸ As noted in footnote , associating a unique representation does not mean that the descriptive/ conceptual content of the phrase must uniquely apply to the referent or that the nominal has to contain any descriptive/conceptual information at all. All that is necessary for the intended referent of a form to be uniquely identifiable is that the addressee is able to associate a unique mental representation of the referent. In the case of definite pronouns, for example, the addressee is able to do this by associating a unique representation that is already in focus. ⁹ It should be noted that not all authors agree that definiteness entails specificity, as they consider attributive noun phrases in Donnellan’s () sense as definite, but not specific (See Partee ; von Heusinger , and Chapter  of this volume; Hedberg et al. ; and §. below).

OUP CORRECTED PROOF – FINAL, 17/1/2019, SPi

DIFFERENT SENSES OF ‘REFERENTIAL’



Turkish, referential objects—whether uniquely identifiable or not—are marked with the accusative case marker; and in Persian, referential objects are marked with the particle RA. Hedberg, Görgülü, and Mameni () present the examples in () and () that illustrate this. The (a) sentences show that definite phrases are marked with these forms. The (b) and (c) sentences show that indefinite phrases that are so marked are interpreted referentially. () Turkish: a. Bugün avukat-ı gör-üyor-um. today lawyer-ACC see-PROG-SG ‘I am seeing the lawyer today.’ b. Bugün bir avukat-ı gör-üyor-um. today one lawyer-ACC see-PROG-SG ‘I am seeing a (particular) lawyer today.’ c. Bugün bir avukat gör-üyor-um. today one lawyer see-PROG-SG ‘I am seeing a lawyer today (some lawyer or other).’ () Persian: a. Emruz vakil-o mi-bin-am. today lawyer-RA DUR-see-SG ‘I am seeing the lawyer today.’ b. Emruz ye vakil-(i)-o10 mi-bin-am. today a/one lawyer-I-RA DUR-see-SG ‘I am seeing a (particular) lawyer today.’ c. Emruz ye vakil mi-bin-am. today a/one lawyer DUR-see-SG ‘I am seeing a lawyer today (some lawyer or other).’

.. Referential determiners In other languages, referentiality is marked on any noun phrase, whether subject, object, or even some other grammatical relation. Salish languages are a case in point, where apparent referentiality marking is distributed across different syntactic argument types. The following examples are taken from a text published in Beaumont () of a story in Sechelt, a Salish language of British Columbia:¹¹

¹⁰ Note that ye ‘a/one’ and -i are both optional, but one of the two morphemes is necessary if the -RA marked NP is to be interpreted as indefinite. Some analysts (e.g., Sadrai ) thus view -RA as a definiteness marker because -RA used without ye or -i indicates that the referent is uniquely identifiable. ¹¹ This story was reglossed by Kaoru Kiyosawa, Simon Fraser University.

OUP CORRECTED PROOF – FINAL, 17/1/2019, SPi



NANCY HEDBERG, JEANETTE GUNDEL, AND KAJA BORTHEN

λ’álsam-nú-mut ?e te sét’álín sísiya. () t’i aux catch.on-NC-REFL OBL ART stick.out small.tree ‘She managed to hang on to a small tree protruding from the bank.’ λ’eq te míman stúmish. () a. t’í λ’um s-cú-s man aux then NOM-go-SG.POSS out ART child ‘Then the young man went out.’ b. t’í cú ?émíwet ?e te ?élewim-s. aux go get.home OBL ART home-SG.POSS ‘He went home to his place.’ Example () shows the introduction, for the first time in the story, of a tree (at most referential, i.e., not yet uniquely identifiable), using the masculine article te. Example (a) shows the same article used for the tenth reference to the main character of the story, a young man (familiar), and (b) shows the same article used for the first reference to the man’s home (uniquely identifiable). Analysis of the text reveals that the relevant set of articles can be used for referents that are in focus, activated, familiar, uniquely identifiable, or at most referential, as would be predicted by the Givennesss Hierarchy, since all these statuses imply the status ‘referential’. There is no definite article in the language. Matthewson () and Gillon () offer detailed discussion of the semantics of determiners in the Salish languages Lillooet and Squamish, respectively. Their discussion suggests that these languages contain determiners that mark referentiality in the Givenness Hierarchy sense. Thus, for example, ‘assertion-of-existence’ determiner phrases in Lillooet and ‘deictic’ determiners in Squamish are reported to introduce new referents but obligatorily take widest scope with respect to negation and modals.¹²

.. Indefinite proximal demonstrative forms As noted in §., in some languages, a form identical to the proximal demonstrative is used for referential indefinites. For example, the form this in casual, spoken English encodes the status ‘referential’ according to Gundel et al. (). According to authors such as MacLaran (: ), use of indefinite this “draws attention to the fact that the speaker has a particular referent in mind, about which further information may be given.” This is shown in MacLaran’s examples below:

¹² However, Lyon (, ) shows that while Okanagan Salish determiner i? phrases can introduce new referents as well as co-refer to established referents, they also can occur in non-specific contexts and within the scope of negation and modals.

OUP CORRECTED PROOF – FINAL, 17/1/2019, SPi

DIFFERENT SENSES OF ‘REFERENTIAL’



() a. He put a/this -cent stamp on the envelope, and only realized later that it was worth a fortune. b. He put a/this -cent stamp on the envelope, so he must want it to go airmail. Use of indefinite this is more natural in (a) where the speaker continues to talk about the -cent stamp, than in (b), where there is no further mention of the stamp. However, noteworthiness and likelihood of subsequent mention cannot be part of the definition of the status ‘referential’; if it were, we would make false predictions about perfectly felicitous, if less frequent, uses of both indefinite this and uses of other forms (such as the definite article) that entail the status ‘referential’. To sum up, indefinite this in English is often used when it is relevant to call attention to the referent of this form for some reason, for example, the referent is noteworthy and likely to be talked about further. When those conditions hold, the attention of the addressee is drawn to the representation of the referent and referentiality is highlighted. Other languages may also have referential this.¹³ For example, Spanish has an indefinite use of the proximal demonstrative determiner in syntactic contexts that are obligatorily presentational. () Spanish: a. Ya este tio a mi lado, y . . . go this guy to my side and ‘There’s this guy next door to me, and . . . ’ b. Hay el tio este, y . . . there.is the guy this, and . . . ‘There’s this guy, and . . . ’ Consultants suggest that this usage is a stereotypical way to start off a joke. It is clear from the syntactic context that these phrases are introducing a referent that is brand new to the discourse, and the proximal marking functions as an affective demonstrative, connoting nuances of vividness and emotional engagement on the part of the speaker, just like the features identified by Lakoff () as characteristic of indefinite this in English. To conclude this section, referentiality in the Givenness Hierarchy sense can be overtly marked grammatically. We gave three types of cases from a variety of languages that mark this sense of referentiality directly. All definites are referential in this sense, but not everything that is referential is also definite (i.e., uniquely identifiable by the time the nominal form alone is processed). In the next section, we turn to examples that have been described and labeled as non-referential in some of the philosophical and linguistic literature, but which would be considered referential within the Givenness Hierarchy framework. ¹³ von Heusinger (Chapter  of this volume) provides discussion of indefinite this expressions, including discussion of indefinite this in other languages, e.g., German.

OUP CORRECTED PROOF – FINAL, 17/1/2019, SPi



NANCY HEDBERG, JEANETTE GUNDEL, AND KAJA BORTHEN

. T H E

REFERENTIAL–ATTRIBUTIVE DISTINCTION

.................................................................................................................................. As noted above, the status ‘uniquely identifiable’ unilaterally entails the status ‘referential’ on the Givenness Hierarchy. This means that all definite phrases are referential by definition, so there can be no ‘non-referential’ definites. This sense of ‘referentiality’ is distinct from the sense of ‘referentiality’ discussed by Donnellan (). Following that tradition, Partee () and von Heusinger () posit a dichotomy between ‘specificity’/‘referentiality’ and definiteness. For them, phrases that would be treated as type-identifiable but not referential in the Givenness Hierarchy theory are cases of non-specific (or ‘non-referential’) indefinites, and phrases whose referents are uniquely identifiable in the Givenness Hierarchy theory but which Donnellan would call ‘attributive’ are cases of non-specific (or nonreferential) definites. The entailment relation that the Givenness Hierarchy theory posits between unique identifiability/definiteness and referentiality/specificity is thus challenged on the Partee/von Heusinger account. Let us look more closely at Donnellan’s original distinction. According to Donnellan (: ), the definite description subject of ‘Smith’s murderer is insane,’ has two distinct uses.¹⁴ On the ‘attributive’ use, the speaker states that Smith’s murderer—whoever turns out to fit that description—is insane. On the ‘referential’ use, the speaker states about a particular person (e.g., Jones) that he is insane and uses the description ‘Smith’s murderer’ because he assumes that this will enable the hearer to ‘pick out’ Jones as the person that he is talking about. Gundel et al. () point out that the cognitive status ‘uniquely identifiable’ covers both these cases, since the addressee can associate a unique representation in both cases, namely the person who the speaker describes as Smith’s murderer, regardless of whether the speaker, or anyone else, knows who it is.¹⁵

¹⁴ The same would hold for ‘the murderer of Smith is insane.’ ¹⁵ We would give a similar analysis to at least some ‘de dicto’ utterances containing definite phrases, such as the example in (i), modeled on an example by Quine (), where the speaker is not trying to say that Ralph believes a contradiction. (i)

Ralph believes that the man in the brown hat is a spy, but does not believe that the man at the beach is a spy.

The speaker presents both descriptions as uniquely identifiable in that each description determines a unique representation for the addressee, and each description is referential in the sense that the speaker intends the hearer to access a representation of the referent by the time the sentence has been processed. The two representations are individuated by different properties. Perhaps the speaker uses two different descriptions even if she knows (and maybe wants to convey) that they have the same referent because she wants to indicate that Ralph doesn’t realize that the descriptions pick out the same man. For more thorough discussion of the de dicto/de re distinction, see Keshet and Schwarz (Chapter  of this volume).

OUP CORRECTED PROOF – FINAL, 17/1/2019, SPi

DIFFERENT SENSES OF ‘REFERENTIAL’



Birner (, ) argues that Donnellan’s distinction is not linguistically relevant because both referentially and attributively used definite descriptions in Donnellan’s sense denote discourse referents, either in the sense of Webber (, ), where discourse referents are mental ‘coat hooks’ on which descriptions are hung, or in Kronfeld’s (, ) sense where a discourse referent is a set of descriptions believed to denote a unique object. Birner argues that the closer the individuating set of properties expressed in a definite description comes to containing all the information that enables the entity to be distinguished from all others, the more referential as opposed to attributive the description intuitively seems to be. Both types of description are referential when viewed in a mentalist way, and both can be anaphorically co-referenced with pronouns. As mentioned above, Partee () and von Heusinger (), suggest that the specificity distinction, originally defined for indefinite phrases, should be amalgamated with Donnellan’s referential–attributive distinction for definite phrases. On this view, both non-specific indefinites and Donnellan’s definite attributive phrases would be classified as ‘non-specific’/‘non-referential’. However, in opposition to this move, we wish to point out that languages such as Turkish and Persian that mark ‘specificity’/ ‘referentiality’ on direct objects (as discussed in §.) obligatorily mark Donnellanattributive phrases as specific/referential. Thus, Hedberg et al. () give the examples in () and () of superlative and ordinal object phrases in Turkish and Persian: () a. Turkish (Görgülü ): Okutman bul-abil-diği en zor alιştιrma-yι ver-di. instructor find-MOD-NMN most hard exercise-ACC give-PAST ‘The instructor assigned the most difficult exercise he could find.’ b. Persian: Ostad saxt-tarin tamrin-(i)-(ro) ke tun-est REL can-PAST-SG instructor difficult-SUP exercise-I-RA peyda kon-e daad. find do-SG give-PAST-SG ‘The instructor assigned the most difficult exercise he could find.’ () a. Turkish (Görgülü ): Cem ilk gör-düğü araba-yι beğen-di. John first see-NMN car-ACC like-PAST ‘John liked the first car that he saw.’ b. Persian: Jân avval-in mâshin-(i)-ro ke did kosh-esh umad. that saw good-SG came John first-DEF car-I-RA ‘John liked the first car that he saw.’ The object phrases in these sentences are obligatorily marked with -ACC and -RA, respectively, but can easily be used attributively (as well as referentially) in Donnellan’s sense. The Givenness Hierarchy sense of ‘referential’ correctly accounts for this data.

OUP CORRECTED PROOF – FINAL, 17/1/2019, SPi



NANCY HEDBERG, JEANETTE GUNDEL, AND KAJA BORTHEN

These phrases would have to be inherently interpreted as uniquely identifiable because there can be only one most difficult exercise or first car seen. Since they are uniquely identifiable, they are referential in the Givenness Hierarchy sense—and crucially they are morphologically marked as such. Indeed, we suggest that any referring form, including indefinite this, can be used in ways that seem just as ‘non-referential’ as Donnellan-attributive definite phrases. See, for example, the discourse in (), where all of the italicized phrases would be classified as ‘referential’ in the Givenness Hierarchy sense. ()

There’s this rapist running around. No one knows who he is. The rapist has struck twice within the last week and everyone is terrified. The police need to be doing everything they can to catch that rapist soon.

Finally, it should be noted that the conclusion that Donnellan’s original distinction is never marked linguistically in any language is challenged in Borthen (, ). Borthen reports that Norwegian, which does have a definite article, allows for superlatives and ordinals to be expressed with or without the definite article, as shown in () below. The claim that such phrases are inherently uniquely identifiable is supported by the fact that the adjective is morphologically marked as definite, regardless of whether or not the definite article is present. What is relevant here is that Borthen reports that phrases of this type without the definite article, as in (b), can only be used attributively—not referentially in Donnellan’s original sense. ()

Norwegian: a. Den beste skiløperen får en premie. the best-DEF skier-DEF gets an award ‘The best skier will get an award.’ b. Beste skiløper får en premie. best-DEF skier gets an award ‘Whoever turns out to be the best skier will get an award.’

Whereas the subject phrase in (a) (with a definite article as well as a definite suffix on the adjective and noun) can have both an attributive use (to talk about the best skier, whoever it turns out to be) and a referential use in Donnellan’s sense (to talk about a particular best skier, e.g., Kjell), the subject phrase in (b) (with no definite article, and a definite suffix on the adjective but not the noun) can only be used attributively. Thus, if the speaker knows who the specific referent is, and intends to refer to that referent, the claim is that she may utter only (a), not (b).¹⁶ The facts in ()–() show that superlatives and ordinals are not always specified for unique identifiability by a definite article, but they are still interpreted as uniquely ¹⁶ The use of bare superlatives in Norwegian, though frequent, is not fully productive. These phrases can only occur as arguments of certain predicates (see Borthen ). The same holds for the bare nominals presented in () below.

OUP CORRECTED PROOF – FINAL, 17/1/2019, SPi

DIFFERENT SENSES OF ‘REFERENTIAL’



identifiable (and therefore also referential and type identifiable, in the Givenness Hierarchy sense) due to the uniquely-identifying conceptual content inherently expressed by the description. Borthen () posits a [-REF] feature for examples like (b) to account for the intuition that they can only be used attributively in Donnellan’s sense. However, we wish to emphasize here that such phrases are referential in the Givenness Hierarchy sense. A possible pragmatic explanation for the attributive intuition may emerge from the fact that in Norwegian, there is an alternative, (a), that is explicitly and redundantly marked for definiteness with a definite article (and definite suffix on the noun). This situation could give rise to a contrasting interpretation for the shorter (also marked definite) (b) that focuses on the descriptive/ conceptual content itself (the type identifiable aspect of the phrase). This does not require that the speaker know anything about the individual described other than the fact that he or she satisfies the description. More data would be needed to decide between these two accounts—specifically, data indicating whether the attributive reading of sentences like (b) can be suspended in certain contexts, and also whether the facts are similar for phrases that are not superlatives.

. B A R E

NOMINAL PHRASES

.................................................................................................................................. English allows bare nominals (without a determiner) for mass nouns, for generic and indefinite plurals, and for some singular predicative uses, as illustrated in (). The bare nominal in (a–c) could be referential in the Givenness-Hierarchy sense since the speaker would be expecting the hearer to construct a representation of the particular portion of wine or group of bees, or to access an existing representation of the cat. However, the bare nominal in (d) would not be referential in the GivennessHierarchy sense since the speaker would most likely be using the nominal only to evoke the role of chair rather than a specific referent. () a. b. c. d.

There is wine in the refrigerator. Cats make good pets. The garden was swarming with bees. Mary was elected chair.

In many languages, bare nominals are even more widespread and can be used for a range of different cognitive statuses. For example, the interpretation of bare nominal phrases in Mandarin and Japanese can have any cognitive status, whether nonreferential or referential. Examples are shown in () and (),¹⁷ where the relevant bare nominals are in boldface. ¹⁷ The data here was collected for a study by Fuller and Gundel (). Speakers viewed a silent film called ‘The Golden Fish’ and, immediately after viewing the film, described it to another native speaker of their language.

OUP CORRECTED PROOF – FINAL, 17/1/2019, SPi

 ()

NANCY HEDBERG, JEANETTE GUNDEL, AND KAJA BORTHEN

Mandarin: a. At most Type-identifiable (not Referential) Shang jie qu mai jiu. mount street go buy wine ‘[He] went out to buy some wine.’ b. Referential Ta feichang xiang dedao yi-zhi hong-de he very want get one-CLS red-PRT ‘He very much wanted to get a red goldfish.’ danshi hong-de jinyu cang zai shitou but red-PRT goldfish hide in rock ‘But the red goldfish was hiding behind a rock.’

()

jinyu, goldfish houmian behind

Japanese: a. Referential ichiban hajime ni detekita bamen ga first beginning at appeared scene SM ano ee mannaka hen ni kuroi neko ga ugoiteite SM be.moving well eh center about at black cat ‘The scene that first appeared has a black cat moving in the center’. b. Familiar and In Focus (and therefore Referential) kingyo to kotori ga uta o utattari SM song OM sing goldfish and bird sorekara kingyo ga kurukuru mawattari and.then goldfish SM round.and.round turn ‘The goldfish and the bird sing songs. And the goldfish turns round and round.’

Since Mandarin and Japanese don’t have definite or fully productive indefinite articles, it is perhaps to be expected that bare nominals will have a wide distribution across statuses. Indeed, Gundel et al. () showed in their corpus study that interpretations of bare nominals in these languages can have any status on the Givenness Hierarchy. However, in some languages that have both definite and indefinite articles, singular bare nominals occur in contexts where they are reportedly not used with higher cognitive statuses than ‘type identifiable.’ For example, Borthen (: ) considers the Norwegian data in (). ()

Norwegian: a. Jeg ønsker meg en svart sykkel. I want REFL a black bike ‘I want a black bike.’ b. Jeg ønsker meg svart sykkel. I want REFL black bike ‘I want a black bike.’

OUP CORRECTED PROOF – FINAL, 17/1/2019, SPi

DIFFERENT SENSES OF ‘REFERENTIAL’



Borthen reports that whereas the speaker can use the indefinite phrase in (a) referentially to express the desire for a particular black bike, she cannot do the same with (b)—this utterance can only mean that she wants something that fits the description ‘black bike’. Spanish also contrasts singular bare nominals with indefinite article phrases. The examples in () show that an indefinite article is required if the phrase is used referentially. Whereas the expression with an indefinite article in (a) can naturally achieve a referential interpretation, the speakers we consulted rejected a referential interpretation of (b) without the indefinite article. However, when the interpretation is non-referential, an indefinite article is optional and the singular nominal can appear either with an article (c) or without (d). () Spanish: a. María quiere comprar un coche. Lo encontró en Craigslist. a car it found.SG on Craigslist. Maria want.SG to.buy ‘Maria wants to buy a car. She found it on Craigslist.’ b. #María quiere comprar coche. Lo encontró en Craigslist. car it found.SG on Craigslist Maria want.SG to.buy ‘Maria wants to buy a car. She found it on Craigslist.’ c. María quiere comprar-se un coche. Pero todavía no ha encontrado Maria want.SG to.buy-REFL a car but still not has found uno que le guste. one that her pleases ‘Maria wants to buy a car. However, so far, she hasn’t found one that she likes.’ d. María quiere comprar coche. Pero todavía no ha encontrado car but still not has found Maria wants.SG to.buy uno que le guste. one that her pleases ‘Maria wants to buy a car. However, so far, she hasn’t found one that she likes’. A similar situation arises in Turkish and Persian. As we saw earlier in ()–(), Turkish and Persian don’t have definite articles but do have indefinite articles, or at least use numeral ‘one’ determiners in most contexts where an indefinite article occurs in English. Hedberg et al. (: fn. ) report that direct objects can occur bare in both Turkish and Persian. In such cases, the nominal receives an interpretation that is reportedly not only obligatorily non-referential but is also number neutral. Thus, with the Turkish and Persian sentences in () and (), it is possible (but not necessary) that the speaker is seeing more than one lawyer. Öztürk () analyzes such noun phrases in Turkish as quasi-incorporated into the verb.

OUP CORRECTED PROOF – FINAL, 17/1/2019, SPi



NANCY HEDBERG, JEANETTE GUNDEL, AND KAJA BORTHEN

()

Turkish: Bugün avukat gör-üyor-um. today lawyer see-PROG-SG ‘I am lawyer-seeing today.’

()

Persian: Emruz vakil mi-bin-am. today lawyer DUR-see-SG ‘I am lawyer-seeing today.’

Gundel et al. () proposed that bare nominals in Mandarin and Japanese contain a zero determiner that encodes the status ‘type identifiable.’ However, because they contain no overt determiners and because we wish to remain neutral on the issue of the syntactic structure of bare nominals in different languages, we now suggest that bare nominals are unspecified for cognitive status. Thus, in languages that permit them syntactically, they can in principle be used for referents/interpretations with any cognitive status. We suggest that the cognitive statuses that the use of bare nominals can signal depends upon the inventory of determiner forms in the language. If there is neither a definite article nor a very fully productive indefinite article, as in Mandarin and Japanese, bare nominals have complete freedom of occurrence. However, if the language contains a fully or quite strongly productive indefinite article, as is the case with Norwegian, Spanish, Turkish, and Persian, bare singulars are restricted to non-referential interpretations. Finally, if the language has a definite but not an indefinite article, as is the case for Moroccan Arabic, bare singulars can have a referential or non-referential indefinite interpretation, as shown in (), but cannot be interpreted as definite.¹⁸ ()

Moroccan Arabic: a. shrit tumubil lyoum. today buy.PERF.S car ‘I bought a car today.’ b. bghit neshri tumubil, walakin ma qdit want.PERF.S buy.IMPERF.S car but NEG able/can sh nlqa whda lli tajabni. NEG find.IMPERF.S one that like.IMPERF.S ‘I want to buy a car, but I can’t find one that I like.’

Whether such non-referential or non-uniquely-identifiable interpretations are absolute or pragmatically derived through implicature is a topic that we leave for future research.

¹⁸ Dechaine and Tremblay () posit a similar distinction between ‘pragmatically conditioned covert D’, exemplified in the Bantu language Shona, and ‘paradigmatically conditioned covert D,’ exemplified in English, Italian, and Hebrew.

OUP CORRECTED PROOF – FINAL, 21/1/2019, SPi

CHAPTER



......................................................................................................................

DEFINITENESS AND FAMILIARITY ...................................................................................................................... BARBARA ABBOTT

“What you have is a Supergun. No, not ‘a’ Supergun, it’s ‘the’ Supergun. The only one of its kind.” – Louise Penny, The Nature of the Beast (), Chapter 

. I

.................................................................................................................................. THE main topics of this chapter (as suggested by the title) are definiteness—more specifically definite descriptions—and familiarity, and the relation between them. Of course on Russell’s () famous theory, the primary difference between ‘the’ and ‘a/an’ is that the former conveys unique satisfaction of the descriptive content associated with it, a theory supported by the quotation from The Nature of the Beast at the start of the chapter. As many have pointed out since then (e.g., Hawkins ), this unique satisfaction must be relativized to a constrained universe, relevant to the conversational context—what Hawkins termed a ‘pragmatic set’ or ‘P-set’. In the happy phrase of Birner and Ward () the denotation must be assumed to be “uniquely identifiable” by the addressee, and Hawkins’s P-set is what enables this in many cases. For Russell—and following him Hawkins ()—the unique satisfaction proposition is an entailment of the utterance, while for Strawson () it is presupposed. Horn and Abbott () endorsed a neo-Russellian approach as well, though one on which uniqueness is conveyed as a conventional implicature (see also, e.g., Gundel et al. , ; Abbott a, b; Gundel ; Gundel and Hedberg ). Familiarity theories, instead, view the conventional import of the definite article as one of familiarity of some type—we’ll look at three possibilities. After a brief look at the theories, we will review a very familiar problem for these theories, and then look at

OUP CORRECTED PROOF – FINAL, 21/1/2019, SPi



 

some more problems (not new either). Following that we will look at that subset of cases in which occurrences of definite descriptions do indeed convey an assumption that the addressee is familiar with the referent. Abbott and Horn () addressed that issue, but did not really give a fully satisfactory explanation of the familiarity implicature, so I’d like to give that a try here. But so as not to leave my reader in suspense, I’ll confess right now that this attempt will not be completely successful.

. F   

.................................................................................................................................. Paul Christophersen is often cited as originating familiarity theories of definite descriptions. As he put it: Now the speaker must always be supposed to know which individual he is thinking of; the interesting thing is that the the-form supposes that the hearer knows it too . . . A condition of the use of the is that there is a basis of understanding between speaker and hearer. This basis comprises the subjects and things known by both parties . . . Christophersen (: )

On this view, definite descriptions denote entities which the speaker assumes are Hearer-old, to use a term from Prince (). More recently Irene Heim () revived this idea in somewhat more restricted terms. As is well known, initially in this work she proposed that definite descriptions are used to refer to entities which have already been introduced into the discourse. On this view they must be Discourse-old (Prince ), a subset of Hearer-old entities. Later this was modified to include cases in which the denoted entity is in the extralinguistic context, or related (‘bridged’) to something that has been introduced: Prince’s Inferables. Finally on Stalnaker’s (, ) account, presuppositions such as the existence of a referent for a definite description belong to the common ground of a conversation. With this account we go back to something like Christophersen’s view—that the denotation of a definite description need only be Hearer-old.

. A      

..................................................................................................................................

.. The problem The main problem associated with the familiarity theory of definite descriptions is the large number of counterexamples (cf. e.g., those in Abbott , , , a, b,

OUP CORRECTED PROOF – FINAL, 21/1/2019, SPi

  



, ; Horn ; Abbott and Horn ; Horn and Abbott ; and the works cited in all of these works). My  paper cited the existence of counterexamples in Hawkins ; Gundel et al. ; Birner and Ward ; and Poesio and Vieira . Here are some of those examples (underlining added when not in the originals). ()

a. What’s wrong with Mary? b. Oh, the guy she went out with last night gave her a hard time.

()

Joe’s fed up with the book he just got for his birthday. (These are from Hawkins : –, cited in Hawkins : )

()

I couldn’t sleep last night. The dog next door kept me awake. (From Gundel et al. : )

() a. In her talk, Baldwin introduced the notion that syntactic structure is derivable from pragmatic principles. b. If you’re going into the bedroom, would you mind bringing back the big bag of potato chips that I left on the bed? (From Birner and Ward : exx. (a, b)). () a. Mr. Ramirez, who arrived late at the Sharpshooter with his crew because he had started early in the morning setting up tanks at another site, just got the first raise he can remember in eight years, to $. an hour from $. b. They wonder whether he has the economic know-how to steer the city through a possible fiscal crisis, and they wonder who will be advising him. (From Poesio and Vieira : exx. (a, c)). Note that none of these definite descriptions conveys an assumption that the addressee has prior acquaintance with the entities the descriptions denote. There are others whom I should have cited in my  paper, who have made the same point—most notably, perhaps, Sebastian Löbner, whose  (!) paper ‘Definites’ distinguished two types of definite descriptions. The denotation of a SEMANTIC definite is determined independently of the context of utterance, while PRAGMATIC definites require contextual information. (Rothschild  made a similar distinction, terming Löbner’s semantic definites ROLE-TYPE, and his pragmatic definites PARTICULARIZED.). Thus semantic definites are those like ‘the meaning of the definite article’ and ‘the second wife of the former husband of my father’s lover’ (from Löbner : ; note that the second of these examples actually does require contextual information, but I guess not in such a way as to determine the uniqueness of the referent). As Löbner pointed out (: ), semantic definite descriptions are particularly problematic for familiarity theories of definiteness. However they are not the only problematic cases—examples (b), (), (), (b), and (a), above, are all pragmatic definites but still problematic for familiarity theories—including the least-restrictive kind that require only that definite descriptions be Hearer-old. And I’ve added more, naturally-occurring, examples, such as those in () below (from Abbott b: exx. (b, c)).

OUP CORRECTED PROOF – FINAL, 21/1/2019, SPi



 

() a. The New York City Board of Health voted yesterday to adopt the nation’s first major municipal ban on the use of all but tiny amounts of artificial trans fats in restaurant cooking . . . [The New York Times on line, //.] b. Police said Thursday that they have caught the Baseline Killer, the gunman responsible for nine slayings that spread terror across the Phoenix area for nearly a year and a half. [Traverse City Record Eagle, //, p. A.] In that paper, I suggested trying the newspaper to find lots more examples—here are a couple from today’s edition (in both cases the cited text falls at the very beginning of the article). () a. Property owners who live along Duck Lake’s shores could soon be on the hook for a new special property tax. The proposed tax would fund the maintenance and operation of a dam used to control the lake’s level . . . [Traverse City Record Eagle, //, p. A.] b. Jason Anthony Ryan was caught in a lie when he initially told investigators he didn’t know Geraldine Montgomery. That’s part of the argument Kalkaska County Prosecutor Mike Perreault made Wednesday in his opening statements in Ryan’s trial for the  rape and murder of a Kalkaska resident, Geraldine Montgomery. [Traverse City Record Eagle, //, p. A.] Again, none of the underlined definite descriptions above has been used with an assumption that the addressees are already familiar with the entity in question. And that also goes for the epigraph of this chapter, where the uniqueness criterion is even spelled out: “What you have is a Supergun. No, not ‘a’ Supergun, it’s ‘the’ Supergun. The only one of its kind.”

.. The proposed solution: accommodation The most usual response to the existence of counterexamples for familiarity theories is to invoke David Lewis’s (a) concept of accommodation. Briefly, the idea is that when an utterance requires background knowledge on the part of the addressee, and the addressee does not have the knowledge in question, they simply add it. Although I have no difficulty with accepting the fact that accommodation does occur in conversation (Lewis gave many other examples in addition to presuppositions), adding accommodation to familiarity gives us an extremely uninsightful theory: it says basically that definite descriptions either do denote entities that are familiar to the addressee or they do not. I have another theory, one that’s equally informative about the nature of definite descriptions: definite descriptions either denote red barns or they do not.

OUP CORRECTED PROOF – FINAL, 21/1/2019, SPi

  



. M    

.................................................................................................................................. Larry Horn and I have pointed out additional problems for familiarity theories.

.. Written versus spoken texts In my  paper I described the results of some empirical research as follows: The basic finding was that written texts included a greater number of definite descriptions that introduce new discourse entities than did the spoken texts. Such descriptions are often notable for their length [cf. e.g., the newspaper examples in () and () above], which makes sense given that a novel definite must uniquely identify a referent from the world at large, whereas one that denotes a discourse entity must only distinguish it from other entities in the discourse. Abbott (: )

As I pointed out in that paper, a familiarity theorist armed with accommodation would apparently have to argue that readers are more accommodating than hearers, which does not seem reasonable. Instead, what seems to be going on in these cases is a consequence of the need to confine one’s assertion to a single proposition (as argued by, e.g., Du Bois  and Lambrecht ). In ordinary speech one can go on and on with assertions (and many people do), but newspaper editors want to convey as much information in as small a space as possible, so anything that is not the main point of the utterance is packaged in non-asserted content—frequently a definite description. And of course readers can take their time processing these complex utterances, while listeners do not have that luxury.

.. ‘The’ versus ‘a/an’ Another group of problems for familiarity theories involves the relation between the definite article and its natural paradigmatic alternative: the indefinite article. As I noted in my  paper, stressing the, as in example () below (ex. () in my paper), automatically stresses the uniqueness of the denotation. ()

The person who could have convinced me to join that club just quit himself.

Stressing ‘the’ also naturally heightens its contrast with ‘a/an’. As was argued by Hawkins , the two form a ‘Horn Scale’—‘the’ being the stronger, of course, and ‘a/an’ the weaker, thus conversationally implicating that the stronger is not appropriate.

OUP CORRECTED PROOF – FINAL, 21/1/2019, SPi



 

Horn and my  and  papers noted a number of examples in which the implicature in question is one of non-uniqueness, not non-familiarity. Some of these are repeated below (from Horn and Abbott , exx. (a), (a), and (b), respectively; as you may know, the gamma is Horn’s symbol for examples googled from the internet). () a. Deforestation “was a or the major factor” in all the collapsed societies he describes. (Jared Diamond, quoted in New York Times book review by Gregg Easterbook,  Jan ). b. γYet time and again, North Korea is cited as not only “a” but “the” major threat to US security. c. γI do not aspire, like some others, to creating “a” or even “the” philosophy of mathematics education. And Horn () noted a more subtle kind of case; compare the following examples (his ex. ()). ()

a. A churlish John Bolton ( . . . reacted angrily to the committee). b. The churlish John Bolton ( . . . was a poor choice to be UN ambassador).

Concerning these examples Horn remarked: While the indefinite in [(a)] suggests a temporary state, the definite in [(b)] indicates a permanent condition. Further, the use of the indefinite suggests a multiplicity of (potential) Bolton-stages or guises, whereas the definite presents a property of the individual without individuating among stages. Horn (: )

Thus the typical contrasted feature in the choice between ‘the’ and ‘a/an’ is the presence or absence of uniqueness. As we have acknowledged, and will consider in more detail below, use of ‘the’ can indeed convey familiarity and thus use of ‘a/an’ instead can convey non-familiarity. Given an explanation for those cases in which definite descriptions implicate familiarity, this is not a problem for the uniqueness theory. However the implicatures above, which have nothing to do with familiarity, are indeed a problem for familiarity theories.

.. Implications or assertions of non-familiarity What might be the biggest problem for theories according to which familiarity is conventionally associated with definiteness is the existence of examples which implicitly or explicitly deny any such assumption, but without causing anomaly. My example (a) below (from Abbott a: ex. ()) has achieved a certain level of notice, probably because it mentions curling; but it’s quite easy to come up with others, as shown in (b–e).

OUP CORRECTED PROOF – FINAL, 21/1/2019, SPi

  



() a. The new curling center at MSU, which you probably haven’t heard of, is the first of its kind. b. I’m sure you don’t know about the newly-proposed solution to global warming (so let me tell you about it). c. Let me also tell you about the interesting art show taking place in Grand Rapids this weekend. d. I’m betting you are totally unaware of the guy sitting behind you in the scary zombie costume. e. γWe all know about the day of the rapture but do you know about the day of the warning which Jesus Christ is going to send during which the stars will collide and the flames of his mercy will shoot across the world producing a red sky. (Did you know about this day of warning? I sure didn’t!) Each of these examples either explicitly or implicitly conveys an assumption of non-familiarity. The fact that these examples show no anomaly at all argues strongly against any claim that familiarity is semantically encoded in definite descriptions.

.. Summary and section conclusion In this section we have seen three additional problems facing the familiarity theory of definite descriptions—problems that have been well known for some time. One might wonder why I am going on and on about this now—isn’t everybody convinced? Well, apparently not. I have been reviewing papers for a volume on reference, and time after time the author in question simply assumes that the familiarity theory is correct! It’s driving me bananas. At this point I would like to use some very strong language conveying my feelings about familiarity theories, but such comments don’t have a place in scholarly literature. So instead we will turn to the issue of the implicature of familiarity on occasions when it does arise.

. T  

.................................................................................................................................. We turn now to cases in which use of a definite description does indeed convey an assumption that the addressee is acquainted with its denotation. We first look at some examples, then we turn to a proposal about which contexts such examples occur in, and finally to the issue of explanation.

.. Descriptions which convey familiarity In looking at examples of familiarity implicating descriptions, I want to separate out two different kinds of cases. To illustrate the first kind of case we will see two examples,

OUP CORRECTED PROOF – FINAL, 21/1/2019, SPi



 

both from novels. The first of these (from John Sandford’s Silken Prey, Ch. ) occurs in a conversation between Lucas Davenport, of Minnesota’s Bureau of Criminal Apprehension, and Turk Cochrane, a Minneapolis police officer. (I’m grateful to Larry Hauser for finding this example.) The situation is that a body has been discovered in a small house—apparently killed by a single shot, with no evidence of other shots having been fired—and the police are wondering about the time of the murder. Cochrane has found that a neighbor up the street from the murder site heard a loud sound earlier. ()

Davenport: “So you haven’t figured out the boom?” Cochrane: “No, but who knows. Maybe that was the shot.”

Cochrane could not have made this assertion felicitously had he not known that Davenport knew about the shot he was talking about, the single shot that killed the victim. The second example (ex. () from Abbott and Horn ) is from Sue Grafton’s book U is for Undertow (p. ), and the context is a phone conversation between a recently introduced character named Walker McNally and someone named Jon whom Walker has called and whom the reader had not heard of prior to this phone conversation. Walker is describing a police search in his neighborhood: ()

“I saw them just now, on my way home from work. I pulled over and chatted with a gal I knew. She said they thought a child was buried on the hill. They dug up the dog.”

Here again, use of the definite description ‘the dog’ automatically conveys to the reader that Walker knows that Jon knows about the dog in question. The second group of examples is less dramatic and, to my intuitions, they do not convey an implicature of addressee familiarity quite as strongly as the two above. The first one was pointed out to me by Andrew Kehler, some years ago. His example, in () below, seems to convey that the addressee has already heard about the book in question. ()

John is reading the book he found on the beach today.

Similarly Ken Taylor suggested the example in () below to Larry Horn (also cited in Abbott and Horn , ex. ()): ()

The man in the kitchen is eating pizza.

Ken suggested that this utterance assumes either that the addressee has already heard about the man, or can see him there in the kitchen. And yet another example from our  paper (ex. ()) is () below: ()

The student I met with three days ago came to see me after class.

Here again, the speaker seems to implicate that the addressee knows about this student.

OUP CORRECTED PROOF – FINAL, 21/1/2019, SPi

  



Note that, unlike conventional aspects of meaning, conversational implicatures can be suspended or neutralized, either contextually or verbally. This is illustrated in the following modifications of these examples. () a. But maybe that was the shot—a single shot that killed a victim I discovered this morning. b. . . . They dug up the dog, the one that I’m about to tell you about. c. John is reading the book he found on the beach today—I’ll tell you about it later. d. The man in the kitchen—the only one you’ll see if you go in there—is eating pizza. e. I only meet with one student a day; the student I met with three days ago came to see me after class. In these examples, the assumption that the addressee has prior knowledge of the entity in question is neutralized. We now have two tasks. One is to try to clarify when definite descriptions have this implicature, and the other is to finally explain how it comes about. First, though, I would like to make clear that these examples do not support familiarity theories. That is because, as we have seen, such theories must make use of accommodation to deal with the many counterexamples like the ones cited in §. and §.. But once accommodation has been invoked, it then becomes necessary to explain why it is not in play in connection with examples ()–(). That is, in contexts in which the addressee is not familiar with the denoted item, these utterances as they stand would be anomalous.

.. What contexts give rise to the familiarity implicature? In our  paper, we noted that Hawkins () had distinguished cases in which singular indefinite descriptions do not suggest the existence of more than one entity satisfying their descriptive content, ones which are ‘neutral’ with respect to uniqueness. He gave the example in () (his ex. (), plus underlining): () A movie that Mary was watching last night was really interesting. Hawkins noted that () is quite consistent with Mary having watched more than one movie last night, or only one. Now observe what happens if we replace the indefinite description with a definite one, as in (). () The movie that Mary was watching last night was really interesting. Now we seem to have one of those familiarity-implicating definite descriptions, which suggests that the addressee already knew about this movie. We suggested, in our  paper, that this is in fact the characteristic situation in which definite descriptions convey familiarity—when the corresponding indefinite

OUP CORRECTED PROOF – FINAL, 21/1/2019, SPi



 

description would be uniqueness-neutral and would not convey anything about the possibility of other entities satisfying its descriptive content. Note that this holds true of our examples ()–() above. Below they are reproduced with indefinite descriptions replacing the definite ones. ()

a. b. c. d. e.

But maybe that was a shot. They dug up a dog. John is reading a book he found on the beach today. A man in the kitchen is eating pizza. A student I met with three days ago came to see me after class.

For (a), the police officer could have heard more than one shot, and similarly for (b–e). These indefinite descriptions are all uniqueness-neutral. We also provided some additional evidence for this claim about the descriptions which suggest addressee familiarity. We noted that if we constructed a context in which uniqueness is relevant, the description in () no longer conveys an assumption of familiarity. We gave the example in () below (our ): ()

Please get me more information about the student I met with three days ago.

Clearly the uniqueness of the student is important in this case. As we noted: “We need not assume, from the utterance of [()], that the addressee has prior acquaintance with the student in question, but only that they can determine in some way (say from the speaker’s appointment book) who that individual is” (Abbott and Horn : ). And of course use of the corresponding indefinite description would not be appropriate here—() was our example (). ()

#Please get me more information about a student I met with three days ago.

In this case it is relevant whether the speaker met with more than one student, since it is unlikely that the speaker would not care about which student the addressee gets the information about. Of course with a little change in context, we can make such a request reasonable, as we pointed out—() is an example we gave in n. : ()

Please get me more information about a student—ANY student—I met with three days ago. (I just have to prove I was on campus then, and not on the site of the terrorist act I’ve been charged with.)

These data confirm our generalization about when the familiarity implicature arises. We now turn to the question of explaining the implicature.

.. Explaining the familiarity implicature Abbott and Horn () really had only one sentence attempting to explain the relation between uniqueness-neutral indefinites and familiarity-implicating definites:

OUP CORRECTED PROOF – FINAL, 21/1/2019, SPi

  



“In such a context [i.e., one where the corresponding indefinite description would be uniqueness-neutral], to flag an entity as uniquely identifiable instructs the addressee to search for a relevant shared domain in which this entity is unique” (: ). But that one sentence has never seemed sufficiently explanatory to me. So, the situation is that the speaker is speaking about an entity, and it doesn’t matter whether or not there are other entities meeting the description used. The speaker uses a definite description instead of an indefinite description. Use of ‘the’ conveys uniqueness, within the relevant domain, of the entity meeting the description used. Why should this only be appropriate if the addressee already knows about that entity? Here’s where Hawkins’s  analysis might come in handy. Hawkins proposed that definite descriptions conventionally implicate “P-membership”. This is spelled out below (Hawkins’s ex. ()). (H) “The conventionally implicates that there is some subset of entities, {P}, in the universe of discourse which is mutually manifest to S & H on-line and within which definite referents exist and are unique.” Hawkins ()

Now consider our first two examples: ‘the shot’ and ‘the dog’. The descriptive content here is so minimal that the requirement of a P-set, “mutually manifest to S & H online” and small enough to contain only a single gunshot (remember these are police having the conversation) or a single dog, is going to guarantee that the addressee knows about the entity in question. Given the satisfaction of the conventional implicature described in (H), it seems clear that the addressee must be acquainted with the entity in question. Let us now turn to the other group of examples that convey familiarity, repeated below (from ()–() above): () a. John is reading the book he found on the beach today. b. The man in the kitchen is eating pizza. c. The student I met with three days ago came to see me after class. Let us consider (a). Presumably the addressee knows which John is being referred to. (People usually don’t use proper names, especially first names, when they don’t know whether their addressees know who they’re talking about.). The rest of the sentence provides a P-set—books John found on the beach today. However it is apparently not sufficient that the addressee has been told about this set of books; it must be that they also have to know that it contains only one book—they can’t just take the speaker’s word for it. And similarly for (b) and (c). For (b), the addressee presumably knows about the kitchen and the typical things a kitchen contains—it might even be the addressee’s own kitchen. So we must construe (b) as requiring that they also know about the man there. And for (c), again, the addressee can’t simply take the speaker’s word for there being only one student they met with three days ago. Now, how are these different from, for example, the examples below (repeated from examples (), (), (), and (b) above)?

OUP CORRECTED PROOF – FINAL, 21/1/2019, SPi



 

()

a. What’s wrong with Mary? b. Oh, the guy she went out with last night gave her a hard time.

()

Joe’s fed up with the book he just got for his birthday.

()

I couldn’t sleep last night. The dog next door kept me awake.

()

If you’re going into the bedroom, would you mind bringing back the big bag of potato chips that I left on the bed?

In these cases the addressee doesn’t need to have prior knowledge of the entity in question—they can just take the speaker’s word for it. So how are these examples different from the ones in ()? That’s what I can’t figure out. One thing in the back of my mind is that I believe there are dialects of German (and maybe this is even true of standard German) which do have two definite articles with morphological differences between them, where one of these articles is associated with uniqueness and the other with familiarity. I don’t know of anything like that for English—no morphological differences with ‘the’ that correlate with when its use suggests addressee familiarity and when it doesn’t. And if ‘the’ were ambiguous in this way, shouldn’t all the examples in ()–() be ambiguous—with one reading suggesting familiarity and the other one not? So I don’t think we can conclude that English is like those German dialects in this respect. Examples (a) and () above are close to being a minimal pair. (a)

John is reading the book he found on the beach today.

()

Joe is fed up with the book he just got for his birthday.

Is the difference in familiarity implicature here a consequence of a difference in roles played by the books in question? Perhaps () aims to tell us why Joe is fed up, so which book it is that he is fed up with doesn’t really matter. While just to say that John is reading a book doesn’t really say enough—so that we have to know which book it is? No, because if we switch the verbs, as in (), we still have the same difference in familiarity. ()

a. John is fed up with the book he found on the beach today. b. Joe is reading the book he just got for his birthday.

Is it because it’s unusual to find a book on the beach, but not to get one for your birthday? But that can’t be it—it’s not unusual for there to be a man in the kitchen, or for a teacher to have met with a student three days ago. Speaking for myself, I’m fed up with this issue, and happy to give up on it.

. C 

.................................................................................................................................. In this chapter I have reviewed the many serious problems for familiarity theories of the difference between definite and indefinite descriptions. This was followed by a

OUP CORRECTED PROOF – FINAL, 21/1/2019, SPi

  



review of some cases in which definite descriptions do indeed convey familiarity, subdivided into two sorts. We were able to tentatively give an explanation for the implicature of familiarity in cases of the first sort—very brief descriptions. But not for the second group of cases.

A This is a revised version of a paper which was presented in November  at Yale University’s Hornucopia. I’m grateful to that audience for their comments, especially David Beaver and Larry Horn. Larry Horn also gave me the epigraph at the beginning of this chapter. Larry Hauser read an early version of this chapter, and I’m grateful for his helpful comments, as well as those of my co-editor, Jeanette Gundel.

OUP CORRECTED PROOF – FINAL, 17/1/2019, SPi

  ......................................................................................................................

THE INDEFINITENESS OF DEFINITENESS ......................................................................................................................

 

. I

.................................................................................................................................. T chapter is about definiteness, and more specifically about the difficulties involved in getting clear on which noun phrases (NPs) should be classified as definite, or more properly, which NPs have uses which can be so classified. (I use ‘NP’ here the way many linguists now use ‘DP’. I also use ‘CNP’, following Montague , to mean ‘phrase of the category of common nouns’, i.e., for the head N plus any restrictive modifiers.) Intuitively, as a rough first approximation, an NP should be considered definite only if it can be used to talk about some particular entity, where an entity may be either concrete or abstract, and may be a group of entities, or a mass of stuff. Many people agree that there are at least four categories which have such uses: proper names, definite descriptions, demonstrative descriptions, and (personal and demonstrative) pronouns. However, the question arises whether these are the only kinds of NPs that deserve the label ‘definite’, and if so why. As we will see, universally quantified NPs, partitives, possessive NPs, and specific indefinites all raise issues concerning definiteness. In the following sections we will look at a number of different attempts to characterize the property of definiteness. We start in §. with three ‘traditional’ proposals: the notion of strength, which arose in connection with the so-called ‘definiteness effect’ in existential sentences; uniqueness/exhaustiveness, a legacy of Bertrand Russell’s  analysis of definite descriptions; and familiarity, which has been a major competitor to uniqueness following Irene Heim’s  dissertation. In §. we turn to three proposals which can be seen as descendants of Richard Montague’s classic  work analyzing NPs as generalized quantifiers (sets of sets): those of Jon Barwise and Robin Cooper (), Barbara Partee (), and Sebastian Löbner (). The final section contains a few brief concluding remarks. Throughout the chapter I will be

OUP CORRECTED PROOF – FINAL, 17/1/2019, SPi

   



ignoring NP uses described as ‘generic’ or ‘bound variable’ unless otherwise mentioned. Even so, space prevents anything like a thorough examination of the topic at hand; the discussion will be necessarily condensed and we will be forced to skip over many important issues.

. C 

..................................................................................................................................

.. Strength The possible role of definiteness within early Chomskyan approaches to English grammar arose in connection with  (Abbott ) existential (‘there-be’) sentences. Such existentials, which may occur discourse initially, do not allow all NP types, as illustrated in () and (). () a. There was a/some (student’s) dog in the yard. b. There were some/several/many/too few/no dogs in the yard. () a. *There was Bill/it in the yard. b. *There was the/that/every/each/neither/Mary’s dog in the yard. c. *There were all/most/both (of the) dogs in the yard. As indicated, the NPs following ‘be’ in () are welcome in this type of existential while those in () are not. Initially the distinction was thought to be one of definiteness and the term ‘definiteness effect’ is often used to describe these differences in felicity. Gary Milsark’s classic work on this topic (, ) revealed many of the complications surrounding this criterion of definiteness, and it is to his credit that he created the new terms  and  for those NPs which can—and cannot—occur felicitously in an existential. Based on examples like those above in () and (), we may sort NPs (and determiners) into two categories as shown in (). ()

Weak: Strong:

a/some (student’s) dog, some/several/many/too few/no dogs Bill, it, the/that/every/each/neither/Mary’s dog, all/most/both dogs

As can be seen, our basic four kinds of definites (proper names, definite and demonstrative descriptions, and pronouns) do not occur felicitously in non-contextualized existentials and are correspondingly classified as strong. In the case of possessive NPs (‘a/some student’s dog’, ‘Mary’s dog’), it appears that the weakness or strength of the genitive NP determiner is transferred to the NP as a whole. (This has been noted by McNally ; Barker ; and Peters and Westerståhl , among others. Cf. also the property of ‘transparency’ noted by Löbner .) Compare too the related example in () (from Woisetschlaeger : ). ()

There was the wedding photo of a young black couple among his papers.

OUP CORRECTED PROOF – FINAL, 17/1/2019, SPi



 

The underlined focus NP in this example is intuitively in the same class with the possessives, but with a postposed ‘possessor’ phrase (‘a young black couple’). There are at least a couple of potential difficulties here. One concerns the universally quantified NPs—those with ‘all’, ‘every’, or ‘each’ (hereinafter the ‘universals’). They are intuitively definite in many of their uses, so their exclusion from existentials seems natural. However they are often considered not to be definite, especially if definiteness is associated with referentiality, which is traditionally opposed to quantification. But then, part of my purpose is to question these traditional oppositions. On the other hand NPs with ‘most’ as determiner are more problematic. They do seem intuitively to be indefinite. Consider a sentence like the following: ()

When the power went off, most students headed for the dorm.

This sentence does not specify which actual students are involved—the speaker clearly does not intend to be talking about any particular students. It is true that morphologically, ‘most’ is a superlative—thus requiring the definite article in its adjectival use. However the definiteness in this case seems to be associated with the quantity involved rather than the denotation of the NP as a whole. (That is, assuming ‘most students’ amounts to more than half of them, the complement of this group does not allow another subset as big.) So the exclusion of NPs with ‘most’ presents a genuine problem for viewing non-occurrence in a non-contextualized existential as an adequate criterion for definiteness. (For more discussion of this issue vis-à-vis this construction, see Abbott : Ch. .)

.. Uniqueness We’ll begin this subsection by reviewing Russell’s classic analysis of definite descriptions, as well as some additions and modifications that have been proposed for it. Following that we turn to more recent variations on the uniqueness theme, and see how well it applies to other sorts of NPs which are usually considered to be definite.

... Russell’s analysis of definite descriptions As is well known, Russell () analyzed denoting expressions quantificationally. () and () below show the difference between indefinite and definite descriptions, in his view. ()

a. A representative arrived. b. ∃x[representative(x) & arrived(x)]

()

a. The representative arrived. b. ∃x[representative(x) & 8y[representative(y) ! y=x] & arrived(x)]

On this analysis, definite descriptions share with indefinites an implication of existence of an entity meeting the descriptive content of the CNP. (Following Frege () and

OUP CORRECTED PROOF – FINAL, 17/1/2019, SPi

   



Strawson (), this element of content may be viewed as  in the case of definite descriptions. We return briefly to this issue later.) For Russell, the crucially differentiating element was the implication that this descriptive content apply uniquely—spelled out in the underlined portion of (b). The formal analysis shown above in (b) can be extended to definite descriptions with mass or plural heads, as shown by Sharvy (; see also Hawkins ). In such cases it is the totality of stuff or entities that is in question. Definite descriptions like ‘the representative’ in (a) are called ‘incomplete’ or ‘indefinite’, since there is an abundance of representatives in the world. In order to maintain Russell’s analysis we must assume that the uniqueness element in (b) is relativized to context in some way. The issue of incomplete definite descriptions is a complex one which we will skip over for the most part here; see Abbott ; Horn and Abbott  for discussion, as well as Löbner’s () ‘pragmatic’ definites, which we will come to very shortly. It is important to note that the uniqueness aspect of Russell’s analysis is separable from the quantificational aspect. That is, definite descriptions could be seen as simple referring expressions (as in the views of both Frege and Strawson) which nevertheless require unique applicability of their descriptive content. This is true of the approach of Löbner (, ), according to which the definite article is a marker of ‘functionality’, in the sense that the CNP with which it is combined is taken to denote a function from contexts to individuals. Some CNPs (e.g., ‘king of France’, ‘first person to swim the English Channel’, ‘claim that pigs can fly’) do this automatically; Löbner () termed these  . The others (e.g., ‘representative’, ‘red car’, ‘person who called last night’) he called  . (Rothschild , apparently unfamiliar with Löbner’s work, introduced the terms   and  for the two subcategories, respectively.) Incomplete definite descriptions, noted above, fall into Löbner’s category of pragmatic definites. Löbner () argued specifically against any interpretation of definite descriptions as quantificational. We return to that issue in §..

... Semantic versus referential uniqueness It will be useful to distinguish two distinct but closely related ways in which an NP could be described as ‘uniquely referring’. If Russell’s theory of definite descriptions, as amended by a suitable approach to incomplete descriptions, correctly captures their contribution to the truth conditions of utterances in which they appear, then the essence of definite descriptions is that there is at most one thing (which may be an atomic entity or a group or mass individual) in the relevant context or situation which matches that descriptive content. Let us call this   (cf. also Roberts  for a slightly different concept). There is another way of viewing uniqueness, which takes into account the goals a speaker has with respect to their addressee. On this view, the essence of definiteness in a definite description is that the speaker intends to use it to refer to some particular entity, and (crucially) expects the addressee to be able to identify that very intended

OUP CORRECTED PROOF – FINAL, 17/1/2019, SPi



 

referent (cf. the concepts of ‘unique identifiability’ and ‘individuation’ discussed by Birner and Ward : f; and see also Gundel, Hedberg, and Zacharski ). This is a pragmatic property which I have called   (Abbott ) (cf. Löbner’s () functional analysis, and also the remarks of Bach : ).

... Extending uniqueness to other NPs We must now check to see how well the idea of uniqueness fits the other categories of NP which are commonly considered to be definite. We’ll start with proper names and then move on to pronouns. For the purposes of this discussion, it will help to separate demonstrative pronouns from the personal pronouns, and group them instead with demonstrative descriptions. ....   It seems clear that proper names are similar to definite descriptions in possessing both semantic and referential uniqueness. First, proper names present themselves as being associated with a single referent; the term ‘proper’ indicates this property, which is also reflected in the fact that proper names in English, used as such, constitute a complete NP and do not accept determiners or restrictive modifiers. On the pragmatic side, as with definite descriptions, speakers can expect their addressees to be able to determine, from the use of a proper name, who or what is being spoken about as long as those addressees are already familiar with the name and its referent (see Prince : ). ....  When we consider personal pronouns it quickly becomes clear that most of them are not semantically unique. Third person pronouns in English incorporate only minimal descriptive content. Although this minimal content may occasionally apply uniquely in a constrained or shrunken universe of discourse, it need not, as shown most clearly by examples like the following (from Winograd : ): ()

The city councilmen refused the demonstrators a permit . . . a. . . . because they feared violence. b. . . . because they advocated revolution.

The city councilmen and the demonstrators are both plural objects suitable for the pronoun ‘they’. However, importantly, in the pair of sentences in (), the content of the predication makes it clear who is being referred to. Use of a pronoun in a context in which a typical addressee would not be able to determine a referent uniquely results in infelicity, as in (). ()

# I told Sue and Betty about the problem, and she said she would work on it.

So it seems that use of a personal pronoun shares with uses of definite descriptions and proper names an assumption that the addressee is expected to be able to determine a referent uniquely—they are referentially unique.

OUP CORRECTED PROOF – FINAL, 17/1/2019, SPi

   



....  Demonstratives are different from the kinds of definite NP we have been considering in requiring (in their demonstrative uses) some kind of ‘demonstration’ (pointing, nod, etc.) from the speaker. Such indicators may of course be used with other definites, but the other kinds of definite NP do not incorporate this requirement as a part of their semantics. As a result, as pointed out by King (: ), a single demonstrative phrase may be used repeatedly in an utterance for different intended referents, unlike definite descriptions or personal pronouns: () a. I want that cookie, and that cookie, and that cookie. b. #I want the cookie, and the cookie, and the cookie. () a. I want that, and that, and that. b. #I want it, and it, and it. The requirement of a demonstration helps demonstratives achieve referential uniqueness without semantic uniqueness.

... Subsection conclusion: the universals As we have seen, definite descriptions, proper names, pronouns, and demonstratives all seem to share referential uniqueness—an intention on the part of the speaker using them to speak about a particular entity which they assume that the addressee should be able to identify. Thus this property has a strong claim to be the essence of definiteness. Furthermore, if that claim holds up then it would seem that the universals (those NPs with ‘all’, ‘every’, or ‘each’ as determiner) should also be included in the category of definite NPs, since in at least some of their non-generic uses their denotation should similarly be identifiable to an addressee.

.. Familiarity We turn now to a competitor to Russell’s uniqueness theory. On Heim’s (, ) approach to semantics, definite and indefinite descriptions both introduce a variable with information concerning some entity (the information contained in the CNP). The difference, in Heim’s view at that time (following Christophersen ), was that indefinite descriptions were required to introduce novel entities while definite descriptions were required to denote familiar ones. When we introduced Russell’s uniqueness theory, we noted in passing that both Frege and Strawson had proposed that definite descriptions presuppose the existence of a referent, rather than asserting it as Russell’s theory seems to imply. Furthermore, on the common ground view of presuppositions, they are propositions which the speaker assumes are shared beliefs between speaker and addressee (see e.g., Stalnaker , ; but also Abbott b). Thus the familiarity theory of definiteness comes close to being just a special case of this view of presuppositions.

OUP CORRECTED PROOF – FINAL, 17/1/2019, SPi



 

The main problem with this approach to definiteness (and in general with the common ground theory of presuppositions) is that there are many counterexamples. As noted above, Löbner () (and Rothschild ) have distinguished semantic or role-type definite descriptions, where the CNP content itself determines a unique referent, from pragmatic or particularized definite descriptions, where the uniqueness in context is signaled by the definite article itself. As Löbner pointed out, the familiarity theory of definiteness neglects the first kind, which can naturally be used to introduce new entities into the discourse (: ). However pragmatic, particularized definite descriptions may also be used to introduce new entities into the discourse, as shown by () below. ()

The case of a Nazi sympathizer who entered a famed Swedish medical school in , seven years after being convicted of a hate murder, throws a rarely discussed question into sharp focus . . . [The New York Times online, //; underlining added.]

In this example, the entity in question was being mentioned for the first time. Newspapers yield many such examples. In fact empirical research by Fraurud () and Poesio and Vieira () has shown that more than % of definite descriptions in naturally occurring discourse may introduce new entities. Supporters of the familiarity theory typically respond to such examples by citing  in the sense of David Lewis (a). However, as has been observed by, for example, Gazdar (: ); Soames (: , n. ); Abbott (: ), among others, appeals to accommodation in this case make the familiarity theory virtually vacuous—definites denote familiar entities unless they don’t. More importantly, such appeals do not explain the fact that it is possible to explicitly or implicitly deny any assumption that the referent of a definite description is familiar to the addressee, as shown in (). () a. The new curling center at MSU, which you probably haven’t heard of, is the first of its kind. [= Abbott a: ex. ()] b. I’d like to introduce you to the idea that Scientology is a gigantic moneylaundering scheme. If it were correct that familiarity was conventionally encoded in definite descriptions, then examples like those in () should be anomalous, but they are not. By contrast, uniqueness apparently is conventionally encoded in definite descriptions. For one thing, when ‘the’ and ‘a/an’ are explicitly contrasted it is always uniqueness that is at issue, not familiarity. (See Horn and Abbott  for many examples.) For another, denying uniqueness for the content of a definite description results in anomaly, as shown in (). ()

#Russell was the author of Principia Mathematica: in fact, there were two. [= Abbott a: ex. ()]

OUP CORRECTED PROOF – FINAL, 17/1/2019, SPi

   



The natural conclusion is that uniqueness is part of the meaning of the definite article while familiarity is not. Instead, familiarity may be derived as a conversational implicature—something that may be cancelled or otherwise neutralized in context. (See Abbott and Horn  for further discussion of this interesting issue.)

.. Section conclusion Of the criteria considered here—strength, uniqueness, and familiarity—it is clear that uniqueness, especially viewed as referential uniqueness, comes closest to characterizing definite NPs. We turn now to some other proposals.

. P 

.................................................................................................................................. In this section, we consider three proposals which arose in the wake of Montague’s () treatment of NPs as expressing generalized quantifiers, or sets of sets (ignoring intensionality, which we will continue to do for the duration of this chapter). They have in common focusing on those generalized quantifiers with non-empty generator sets. There are some differences among the proposals, with possibly different conclusions about which NPs would be considered to be definite. We will also look at some syntactic evidence.

.. The proposals ... Barwise and Cooper  We consider first the definition of definiteness in Barwise and Cooper (: f; italics in original). () DEFINITION. A determiner D is definite if for every model M = and for every A for which ||D||(A) is defined, there is a non-empty set B, so that ||D||(A) is the sieve {X  E|BX}. (Hence, ||D||(A) is what is usually called the principal filter generated by B.) In more or less ordinary language, this definition requires definite determiners, when combined with a set term, to yield a set of sets with a nonempty intersection—the generator set for the filter. NPs with a definite determiner are then definite. This definition raises a couple of issues. The first is that it does not include NPs without determiners, such as pronouns and proper names. However it should be relatively easy to revise this kind of definition of definiteness to include them, since the generalized quantifiers interpreting them would also be principal filters. The second

OUP CORRECTED PROOF – FINAL, 17/1/2019, SPi



 

issue is more complicated. Barwise and Cooper intended to include definite descriptions (of course) while excluding the universals. One reason for this is that they assumed (following Jackendoff ) that a crucial property of definite NPs is the ability to serve as the embedded NP in a partitive, and that the universals cannot appear there. However excluding the universals while including definite descriptions required a couple of stipulations. One was that definite descriptions for which the CNP set is empty (like ‘the present king of France’) are undefined. Thus Barwise and Cooper follow Frege () and Strawson () in their view that definite descriptions semantically presuppose the existence of a referent. The other stipulation is that the universals do not share this presupposition of existence. (This latter stipulation does not follow Strawson; cf. Strawson : .). We will return to the issue of partitives in §...

... Partee  Our second characterization was not actually proposed as a definition of definiteness, but is nevertheless highly congruent with the Barwise and Cooper idea. Broadly within the Montagovian framework, there are three possible (extensional) types for NPs: e (the type of NPs which denote entities), (the type of NPs which denote sets of entities), and (the type denoting generalized quantifiers—sets of sets). Partee () noted that many NPs can appear in more than one type, depending on the context, and she proposed a number of ‘type shifting principles’, to provide appropriate interpretations. The principle of interest here is one called ‘lower’, which applies to NPs of type and yields NPs of type e. Lower only applies to generalized quantifiers which are generated by single entities (where, as we have noted, plural sums and masses are also considered to be entities) and maps them on to those entities. Overtly quantificational NPs like ‘most chickens’ and ‘no good ideas’ are not subject to this principle (although it is not clear that the universals are excluded, given that totalities can be considered to be plural sums). Partee pointed out that the traditional division between referential and quantificational expressions seems to correlate well with the division between NPs which may be of type e and those which may not be (cf. Partee : ). Partee’s type e NPs are the same as Barwise and Cooper’s definites with one exception: indefinite descriptions. Partee’s criterion for being of type e was the ability to serve as the antecedent of a singular discourse pronoun. As shown in () (from Partee , exx. () and ()), this criterion groups indefinite descriptions with definite NPs. ()

a. John/the man/a man walked in. He looked tired. b. Every man/no man/more than one man walked in. *He looked tired.

As noted above, Heim () treated indefinite descriptions similarly to definites. Chastain ; Kamp ; and Fodor and Sag  have also argued that indefinite descriptions can be referential. In order to achieve a type e interpretation for indefinite descriptions, Partee (following Zeevat ) suggested that they might receive a generalized quantifier interpretation based on a particular variable—roughly,

OUP CORRECTED PROOF – FINAL, 17/1/2019, SPi

   



the set of sets containing x, where x would be assigned a value in context. This would allow lower to apply, yielding x. If this proposal is right, then we would have to conclude at this point either that indefinite descriptions can be definite (this would be when they introduce a discourse referent), or, more plausibly, that being of type e does not correspond to being definite.

... Löbner  Löbner () explored the interaction between negation and NP interpretation. He argued that definite descriptions are associated with a (semantic)   , so that predicates apply to them as a whole. This is most easily illustrated with a plural definite description, as shown in (): () a. The cows are in the field. b. The cows are not in the field. Löbner argued that (a) is true only if all of the cows are in the field, and (b) is true only if all of the cows are not in the field. If some of the cows are in the field and some are not, then neither (a) nor (b) has a truth value. This property then plays an essential role in Löbner’s characterization of definiteness, and also serves to distinguish plural definite descriptions from the universals (as well as other quantified NPs). Note that the negation of (a) is not (b), but rather (c). () a. All the cows are in the field. b. All the cows are not in the field. c. Not all the cows are in the field. It is this difference in behavior which separates definite NPs from quantificational NPs definitively, in Löbner’s view. It is worth noting that this characterization of definiteness, like Partee’s criterion for type e NPs, could be held to include specific indefinite descriptions. As the examples below in () suggest, the specific indefinite descriptions, like definites, take scope outside of negation. (This in the examples below should be read as the specific indefinite this (cf. Prince a), and not the demonstrative ‘this’). () a. This/A certain strange cow is in the field. b. This/A certain strange cow is not in the field. c. No (#certain) strange cow is in the field. The natural negation of (a) is (b), not (c). Löbner supported this classification of NPs with some syntactic characteristics. One concerned general scope taking ability. He asserted that “definite NPs do not have scope at all”, while quantificational NPs, of course, do take different scopes. Another property concerned behavior in partitives; like Barwise and Cooper, Löbner assumed that only definite NPs may appear embedded in a partitive. In the next two subsections we take up these assumptions.

OUP CORRECTED PROOF – FINAL, 17/1/2019, SPi



 

.. Scope taking In this subsection we will take a look at the scope taking abilities of NPs. It will help to break these down into two possibilities—the ability to take narrow scope, and the ability to take wide scope—since these might differ. Of particular interest will be similarities and differences between definite descriptions and quantified NPs.

... Narrow scope Proper names, and pronouns and demonstrative NPs when they are used demonstratively, generally speaking do not take narrow scope with respect to other operators. This is true whether the operator in question is a quantifier, a propositional attitude predicate, or a modal. This behavior is unlike that of quantificational NPs. Each of the examples in () is ambiguous, and has an interpretation in which the underlined NP is interpreted with narrow scope relative to the boldface operator. ()

a. Everybody loves somebody. b. Rush Limbaugh hopes that many liberals will fail. c. Several philosophers might have gone into plumbing.

In contrast to the ambiguity of the examples in (), the univocality of those in () illustrates the fact that proper names, pronouns, and demonstratives do not take narrow scope. ()

a. Everybody loves Madonna/her/that singer over there. b. Rush Limbaugh hopes that Obama/he/this person sitting here will fail. c. Aristotle/he/those philosophers might have gone into plumbing.

Exceptions to these generalizations have been argued for, but by and large the pattern holds. In these contexts (i.e., in sentences with other operators), definite descriptions can pattern with the quantificational NPs rather than the referential ones, as shown in (): ()

a. Each of those people loves the color they look best in. b. Rush Limbaugh hopes that the current president will fail. c. The number of US states might have been odd.

So as far as classifying an NP type as definite or indefinite, the ability to take narrow scope does not seem to give us good results. The kind of narrow scope Löbner was particularly concerned with was narrow scope with respect to negation. And it is true that (b) above, repeated here as (), seems strongly to suggest that the cows as a group fail to be in the field. ()

The cows are not in the field.

However ‘each’ also refuses to take narrow scope with respect to negation; () is likewise unambiguous.

OUP CORRECTED PROOF – FINAL, 17/1/2019, SPi

   



() Each of the cows is not in the field. But Löbner does not seem to allow that quantification involving each involves definite reference.

... Wide scope As sentence operators, quantifiers can take wide scope with respect to other sentence operators: propositional attitude predicates, modals, or other quantifiers. Thus the examples in () above also have readings where the underlined NP has wide scope. Wide scope would seem also to be unexceptionable for definite NPs, if we take the examples above in () to be ones in which the underlined NPs have wide scope. An alternative, however, is to conclude that those NPs are simply scopeless. On the other hand it has been argued that we need to recognize actual scope taking on the part of proper names (and other definites) in order to account for cases of sloppy identity. Consider, for example, (): () Mary likes her boss but Jane doesn’t. The relevant reading here is the sloppy one where Jane doesn’t like her own boss. On the standard analysis, the pronoun (‘her’) in the first VP is construed as a variable bound by ‘Mary’. (This allows for the second, pronominal, VP to be interpreted as identical with the first.). In order to bind this pronoun, it is typically assumed that the NPs ‘Mary’ and ‘Jane’ must have sentential scope taking capabilities (cf. Heim and Kratzer ; following Keenan ; Partee ). Peters and Westerståhl () gave example () as evidence that definite descriptions are not quantifiers. () The novices chose a mentor. They remarked, concerning this example, that it “unambiguously entails that all novices have the same mentor” (: , n. ). However that is not so clear. Consider example (): () In their sophomore year the students chose a major. This sentence does not seem to me necessarily to imply that each of the students chose the same major. And likewise () seems to have a non-anomalous reading. () For lunch, the children ate an apple. On the other hand it is not so clear that these facts require a quantificational analysis of the definite descriptions in question. Why isn’t it simply an issue of distributive versus collective predication? In which case why did Peters and Westerståhl take this kind of fact to be significant concerning the quantificational status of definite descriptions? I must confess to being at a loss as to how to interpret these kinds of scopal facts.

OUP CORRECTED PROOF – FINAL, 17/1/2019, SPi



 

... Conclusion concerning scope It seems to me we must conclude from this exploration that scope facts are quite complex and require further investigation. The statement that definite NPs do not have scope at all seems definitely too strong. If we want to include definite descriptions in the super category of definite NPs, then we cannot take inability to take scope as a characteristic of the category. On the other hand Löbner’s claims about negation have held up well, with the exception that each does not behave like the other quantificational determiners in this regard. We turn now to partitive NPs, which have been taken by Löbner as well as many others to provide a criterial context for definites.

.. Partitives Superficially, partitive NPs seem to have the form Det of NP. Some examples are given in () below¹. ()

some of the apples, few of those options, all of Mary’s dogs

If that is indeed their structure, they present a problem: ordinarily determiners combine with CNPs, which denote sets of entities. But with the exception of predicate nominals, NPs denote either generalized quantifiers (sets of sets), or entities, and so they do not provide a suitable interpretation for another determiner to combine with. Jackendoff () argued that the NP embedded in a partitive had to be definite; he termed this the ‘Partitive Constraint’. As was mentioned in §..., Barwise and Cooper () assumed that Jackendoff was correct, and proposed an analysis of ‘definite NP’ which was tailored to that assumption. As we saw, this definition requires a definite NP to always have a non-empty intersection (the generator set)—thus including definite descriptions (on the assumption that they semantically presuppose existence), but excluding the universals (on the assumption that they do not semantically presuppose existence) as well as all other overtly quantified NPs. The beauty of the Barwise and Cooper analysis was that it suggested an explanation for the Partitive Constraint: only a generalized quantifier with a non-empty intersection would yield a set for the initial determiner to combine with. Barwise and Cooper suggested that ‘of’, in this construction, acts as an instruction to take the generator set of the NP with which it combines (: f ). Given that the universals have a non-empty intersection (unless the CNP fails to denote, as in, e.g., the case of ‘every unicorn’), Barwise and Cooper’s reasoning would

¹ Unfortunately, my discussion in this paper will focus exclusively on English. For a consideration of definiteness in a variety of languages, see Lyons . Much of the material in this section was first presented in Abbott . That paper also presents an analysis of partitives, which is not attempted here.

OUP CORRECTED PROOF – FINAL, 17/1/2019, SPi

   



lead us to expect them to be able to occur embedded in a partitive. Consider an example like (): () Most of every apple was in the bowl. It is true that () cannot mean most of the apples are in the bowl. However that is because ‘every’ is a necessarily distributive quantifier—hence its denotation cannot be taken as a group. Below we will see that the universal quantifier ‘all’ does allow a group interpretation, and does appear embedded in partitives. We should note () does have an interpretation as a  partitive, where the initial determiner (‘most’ in this case) applies to the individual apples rather than the apples as a group. On this interpretation the individual apples have been cut up or mashed, and a majority portion of each put in the bowl. The necessary distributivity of ‘every’ forces the initial determiner to apply to the individuals in the denotation of its NP. The failure of NPs with ‘each’, ‘most’, or ‘both’ to occur embedded in a group partitive is a result of the same property. They may readily occur in mass partitives however, as shown in (): () a. The Smithsonian donated most of both rare book exhibits. b. One third of each book Chomsky writes is footnotes. c. At least some of most fruits consists of rind and seeds. In mass partitives like these the initial quantifier is applying to individuals within the denotation of the embedded NP, and not that NP as a whole. Returning to the main theme, a number of researchers (most of whom concentrate on group partitives) have concluded that partitives do not, in fact, have the structure Det of NP. Keenan and Stavi (: ) and Peters and Westerståhl (: ) argue that partitives have the structure [[Det of Det] CNP]; while Barker (); Löbner (); and Ionin et al. () (among others), support a two NP structure. If one of these different structures is correct, the rationale for the Barwise and Cooper analysis disappears. Despite this fact, there is still widespread confidence in Jackendoff ’s partitive constraint. Keenan and Stavi assumed that the second Det in their Det of Det partitive structure must be definite (: ). Peters and Westerståhl conclude that it must be either definite or possessive (cf. Peters and Westerståhl : ). Similarly, as we have seen, Löbner holds that only definite NPs, and not quantificational NPs, occur embedded in a partitive (: ). Leaving aside the case of the necessarily distributive NPs, there remain many counterexamples to the Partitive Constraint, viewed as a requirement of definiteness. Thus consider the examples below in ()–(). (The original sources are given following the examples; ()–() also appeared in Abbott .) () He ate three of some apples he found on the ground. (Stockwell, Schachter, and Partee : ) () This is one of a number of counterexamples to the PC. (Ladusaw : )

OUP CORRECTED PROOF – FINAL, 17/1/2019, SPi



 

()

They called the police because seven of some professor’s manuscripts were missing. (Keenan and Stavi : )

()

I would hate for my boyfriend and me to be two of seventeen housemates—we would never be able to kiss in private. (Ionin et al. : )

()

a. Ants had gotten into most of some jars of jam Bill had stored in the basement. b. Three quarters of half the population will be mothers at some point in their lives. c. Any of several options are open to us at this point. d. Each student only answered a few of many questions that they could have. e. Half of all dentists who chew gum prefer Trident. (Abbott : passim)

With the exception of the last example ((e)), each of the underlined NPs above is intuitively indefinite, and would not be classified as definite by any of the analyses of that concept which we have looked at so far. The last example has an embedded universal (‘all dentists who chew gum’) with a group interpretation. Data like those above (and more examples can be easily found) suggest that any NP that can have a group interpretation can appear embedded in a group partitive. The only exception is bare plural and mass NPs, which are not welcome there, as illustrated by (). ()

a. *Most of books by Chomsky are on politics. b. *Some of green slime is created by bacteria.

As is also indicated by these two examples, it does not matter whether the bare NP is interpreted as indefinite (as in (a)) or generically (as in (b)). Neither is possible embedded in a partitive. (Interestingly, in this characteristic bare plural and mass NPs are different from proper names; viz., e.g., ‘Most of Australia is desert’). The upshot of this investigation into partitive NPs is that they do not provide a good diagnostic for definiteness.

.. Section conclusions In this section we have looked at three further attempts to characterize a distinction in NPs. In two cases the authors were explicitly attempting to get at the essence of definiteness (Barwise and Cooper, and Löbner), while Partee suggested that being of type e might turn out to coincide with the closely related concept of referentiality. However, as we have seen, there are substantial problems with each of these attempts— viewed as definitions of definiteness. In one (and possibly two) cases the universals seem to be excluded only arbitrarily. Furthermore in the case of the Barwise and Cooper analysis their exclusion was motivated by an assumption which we have seen ample reason to question. On the other hand specific indefinite descriptions would be

OUP CORRECTED PROOF – FINAL, 17/1/2019, SPi

   



regarded as definite according to two of the characterizations, which raises further questions about their adequacy—again, as definitions of definiteness. I would certainly not want to claim that the semantic properties brought to light in the three works considered here are not interesting, or that they do not correspond to significant linguistic properties of NPs. My only claim is that none of them appear to give a satisfactory definition of definiteness, or one that is superior to referential uniqueness.

. C 

.................................................................................................................................. This chapter has explored the concept of definiteness, and in particular whether existing characterizations of this notion seem to capture its essence. In the course of this exploration we have seen a number of issues and problem areas which make the selection of a single characterization as the correct one difficult. My own feeling is that referential uniqueness is the strongest contender. However, I want to reiterate a comment from the beginning of the chapter: that space prevents anything like a full examination of the issues under investigation here. Interested readers are urged to consult both the references listed below and the works they cite.

A This chapter originally appeared as a paper in Thomas Gamerschlag, Doris Gerland, Rainer Osswald, and Wiebke Petersen, eds., Frames and Concept Types: Applications in Language and Philosophy, , New York: Springer, –. It is used here with permission. Much of the material was drawn from Abbott : Ch. . An earlier version was presented at the Second Conference on Concept Types and Frames in Language, Cognition, and Science (Düsseldorf, August ), and I would like to thank that audience for their comments. I’m especially grateful to Sebastian Löbner for his comments, and for inviting me to the conference. And finally, I’m grateful to my co-editor, Jeanette Gundel, for her excellent comments.

OUP CORRECTED PROOF – FINAL, 21/1/2019, SPi

CHAPTER



......................................................................................................................

INDEFINITENESS AND SPECIFICITY ...................................................................................................................... KLAUS VON HEUSINGER

. I

.................................................................................................................................. S PECIFICITY is a semantic–pragmatic notion for certain ‘strong’ readings or interpretations of indefinite noun phrases that can affect truth conditions or felicity conditions as well as morpho-syntactic phenomena such as movement or Differential Object Marking. The notion is broadly employed in recent work and it is associated with an open set of properties and various concepts, but there is no common definition. The generally shared intuition behind this category is that a speaker uses an indefinite noun phrase specifically if he or she intends to refer to a particular referent, the referent ‘the speaker has in mind’. In the course of this chapter, we will discuss different approaches to model this rather informal characterization and I will conclude that the theory of ‘referentially anchored indefinites’ provides an appropriate account of this characterization and linguistic phenomena associated with specificity. At the same time this theory also delimits the concept of specificity from other familiar concepts such as partitivity, noteworthiness, or discourse prominence. The range of specificity related phenomena can be illustrated by examples ()–() below: ()

Mary may want to marry a Swede. (Karttunen /, ex. ())

Example () has two readings, one in which there is a particular Swede and Mary wants to marry him, and one in which Mary wants to marry one or another Swede. The first reading expresses an existential entailment that there is a Swede, but the second reading does not; thus the readings have different truth conditions. Therefore, Karttunen (/) and others model this contrast by the scope of the existential quantifier associated with the indefinite noun phrase a Swede with respect to the intensional verb want. The contrast was linked to Quine’s (: ) remark that indefinite noun

OUP CORRECTED PROOF – FINAL, 21/1/2019, SPi

  



phrases in opaque contexts have two readings, a ‘relational’ and a ‘notional’ one, which makes example () ambiguous, similarly to the de re versus de dicto readings of definite noun phrases.¹ Example () has one reading according to which the speaker has a particular student in mind and one where the speaker just asserts that at least one student cheated without implying any particular one. ()

A student in the syntax class cheated on the final exam. (Fodor and Sag : , ex. ())

There is no truth conditional difference between the two readings of (), supporting the assumption that the contrast is pragmatic and not semantic. However, Fodor and Sag (: ) argue that () [their (), KvH] “is nevertheless ambiguous”. Its indefinite noun phrase may be semantically interpreted in two distinct ways. One semantic interpretation is that of a quantified expression such as each student or few students; the other interpretation is that of a referring expression such as a proper name or demonstrative phrase. They further argue that this lexical ambiguity shows truth conditional effects. Enç () claims that specificity is closely related to partitivity, if not identical. She argues that the second sentence in () has two interpretations in English, which can be disambiguated in Turkish by accusative case marking on the indefinite a girl. A case marked indefinite direct object must be partitive, that is, part of the children introduced in the first sentence. An unmarked direct object signals that the referent of the girl is not included in the set of children introduced before (English translation of Turkish examples (see also () in §..). ()

Some children entered the room. I knew a girl.

While Enç defines specificity in terms of d-linking or inferability, Ionin () undertook another extension of the notion of specificity towards forward discourse properties. She proposed that the use of indefinite this in English, as in (), not only behaves like a specific indefinite in the sense of Fodor and Sag (), but also (strongly) suggests that the speaker talks further about the discourse referent introduced by the this-indefinite. The example would be less coherent if we would continue after this man with an unrelated topic. ()

There is this man who lives upstairs from me who is driving me mad because he jumps rope at  a.m. every night. (Ionin : , ex. (), quoted from MacLaran : )

Ionin (: ) connects the referential meaning of the indefinite this with its discourse function by the concept of noteworthiness; the use of indefinite this is only felicitous if there is a noteworthy property to be assigned to the indefinite.

¹ Baker () termed these two readings of indefinites ‘specific’ versus ‘non-specific’ in his master’s thesis. See also Fillmore () for one of the first uses of ‘specific’.

OUP CORRECTED PROOF – FINAL, 21/1/2019, SPi



  

As we can see from the examples, a specific interpretation of an indefinite has been described as, for instance, expressing a direct referential or rigid reading (cf. () and ()), a wide scope reading (cf. ()), an epistemic reading, i.e., the speaker can identify the referent (cf. (), ()), the indefinite being discourse linked and presuppositional (cf. ()), or the indefinite signaling discourse prominence and noteworthiness of the introduced discourse referent (cf. ()). This list is not comprehensive and could be extended, and there is no agreed set of characteristics of a specific reading. The semantic–pragmatic category ‘specificity’ is notoriously difficult to define or to demarcate from other semantic–pragmatic categories since we have three closely interacting parameters, which sometimes lead to circular argumentation: different linguistic means to mark specificity contrasts, types of specificity contrasts, and the theoretical models that describe and analyze specificity. These parameters are the topic of the following sections: §. discusses some linguistic means to express specificity such as article systems or specific adjectives. In §., I present different types or contexts of specificity contrasts. §. presents a short overview of four families of models of specificity and §. provides a short summary and outlook.

. L    

.................................................................................................................................. The English examples ()–() do not show any formal or overt marking of the contrast between a specific and a non-specific reading. Still there is a broad variety of linguistic means to mark or signal specificity across languages, which range from universal means to language specific means. Fodor and Sag (: –; based on Karttunen ; Fodor  and others) provide a helpful list of (universal) linguistic characteristics that favor either specific or non-specific readings of indefinites, which I illustrate as modifications of () in (b–f ) below: (i) A main indicator is the content of the noun phrase: the more descriptive content a noun phrase has, the more likely it is to have a specific reading (cf. (b)). However, even long descriptions can be read non-specifically (see below). (ii) Non-restrictive relative clauses obligatorily trigger specific interpretations (cf. (c)). (iii) Topicalization and left dislocation favor a specific interpretation (cf. (d)).² (iv) Indefinite or presentative this strongly favors a specific reading (cf. Prince a) (cf. (e)). (v) Imperatives only allow nonspecific readings, such that even the long indefinite in (f ) cannot be understood specifically.

² Jeanette Gundel kindly informed me that Fodor and Sag’s () claim about topicalization favoring a specific interpretation is too strong (see Gundel  for an extensive discussion).

OUP CORRECTED PROOF – FINAL, 21/1/2019, SPi

   ()

a. b. c. d. e. f.



Mary wants to marry a Swede. Mary wants to marry a very talkative Swede from the small village of Lärbro. Mary wants to marry a Swede, who lives in the small village of Lärbro. A Swede from Lärbro, Mary wants to marry him. Mary wants to marry this Swede. Mary, marry a very talkative Swede from the small village of Lärbro!

Besides this set of universal grammatical indicators, there is a long list of language particular linguistic indicators of specificity such as mood in relative clauses (Rivero ), articles (Chung and Ladusaw ) or the complex system of indefinite pronouns (see Haspelmath  for an overview). Spanish, like other Romance languages, marks the contrast between a specific versus a non-specific reading of an indefinite (and the referential versus attributive reading of a definite) noun phrase by the mood in the relative clause that modifies that noun. Subjunctive marks or determines a nonspecific (or an attributive) reading, and indicative marks a specific (or a referential) reading (Rivero :  ex. ()): ()

a. Quiere casarse con {la, una} muchacha que sea (SUBJ) rubia y con pecas. ‘He wants to marry {the, a} girl who may be blonde and with freckles.’ (NON-SPECIFIC only) b. Quiere casarse con {la, una} muchacha que es (IND) rubia y con pecas. ‘He wants to marry {the, a} girl who is blonde and with freckles.’ (SPECIFIC only)

Ionin () argues that if a language has an article system with two semantically different articles, the contrast expressed is either definiteness or specificity. While most European languages that have articles mark definiteness, various other languages mark specificity, like Maori, which distinguishes between the specific têtahi and the non-specific he indefinite article (Chung and Ladusaw : –, ex. (a); (a), T for tense): () a. Kâore têtahi tangata i waiata mai. T.not a person T sing to.here ‘A particular person didn’t sing (= There was a person who didn’t sing).’ b. Kâore he tangata i waiata mai. T.not a person T sing to.here ‘No one at all sang.’ But: ‘*A particular person didn’t sing.’ English (like many other languages) can express specificity contrasts by means of certain determiners, adjectives, quantifiers, or bare nouns (see Farkas ): ()

a. b. c. d. e. f. g.

Every student recited a poem of Pindar. Every student recited thisindef poem of Pindar. Every student recited SOME poem of Pindar. Every student recited a certain poem of Pindar. Every student recited at least one poem of Pindar. Every student recited poems of Pindar. Every student recited sm poem of Pindar.

OUP CORRECTED PROOF – FINAL, 21/1/2019, SPi



  

(a) is the unmarked form with the indefinite article, which is ambiguous between a specific and non-specific interpretation; the indefinite or presentative this in (b) forces a specific (or referential) reading, the focused some in (c) allows a wide scope reading more easily than the form with the indefinite article in (a); the specificity marker a certain in (d) forces a specific (i.e., wide scope) reading. The expression at least one in (e) has a quantificational (i.e., non-specific (quantificational)) reading (still it shows a scopal behavior similar to (a)). The bare noun in (f) only allows for a narrow scope (i.e., non-specific) reading, and likewise the form in (g) with the phonologically reduced sm. There are many more ways of marking (non-)specificity by lexical items, functional markers, or other constructions. Most of these means are restricted to certain contexts or to certain specificity contrasts. We therefore assume as reliable tests for specificity in the list of Fodor and Sag (): (i) non-restrictive relative clauses, (ii) this-indefinites (if a language allows for it), (iii) the use of a certain or a particular (in languages with similar expressions). Tests for non-specificity are (iv) imperatives, and (v) the replacement by bare nouns, if a language allows for it.

. T   

.................................................................................................................................. The broad variety of data associated with specificity, the different types of specificity, and the many theoretical approaches to specificity make a comprehensive overview and a straightforward classification very difficult (but see Karttunen , /; Fodor ; Abbott ; Fodor and Sag ; Abusch ; Farkas , ; Yeom ; Ruys ; Kamp and Bende-Farkas ; Ionin ; Endriss ; Ebert and Hinterwimmer ). In the following I present a classification of seven kinds of contexts for specificity related phenomena, extending Farkas’s () division into scopal, epistemic, and partitive specificity by the following four additional categories of indefinites (see von Heusinger ): referential specific, topical, noteworthy, and discourse prominent indefinites.

.. Referential contrasts Indefinites show an ambiguous interpretation in opaque contexts, which is comparable to the de re versus de dicto readings of definites (see Keshet and Schwarz, Chapter  of this volume). ()

a. Mary wants to marry a Swede. b. Mary wants to marry the richest man in Sweden.

Thus, both sentences in () allow a reading where the speaker intends to refer to the particular person Mary wants to marry and where the indefinite noun phrase in (a)

OUP CORRECTED PROOF – FINAL, 21/1/2019, SPi

  



and the definite noun phrase in (b) serve to identify this person to the hearer or addressee. But there is also a reading where the speaker simply intends the hearer to attribute to Mary the wishing attitude to marry a Swede or the richest man in Sweden (whoever this might be). As a test to distinguish the two readings, we use existential inferences: the specific reading of (a) allows for the inference that there is a Swede, while the non-specific does not. Analogously, the specific reading allows for anaphoric linkage (as in a continuation He lives in Göteborg), and the non-specific reading is preferred with the continuation But she hasn’t found one yet.

.. Scopal contrasts Scopal specificity (often also including referential specificity) concerns the interpretation of the indefinite if there are other semantic operators in the sentence. I present three types of contexts with growing complexity: simple scope interaction (‘local configuration’), scope island escaping properties (‘global configuration’, ‘long distance indefinites’), and intermediate readings. First, Karttunen (/) discussed examples such as () where the indefinite can take wide scope over the universal quantifier every day, as in (a). The indefinite can also take narrow scope with respect to the universal quantifier, as in (b). We can model these two readings if we represent the indefinite as an existential quantifier that scopally interacts with the universal quantifier representing every day. () Peter visited a museum every day. a. There is a particular museum (say the Louvre), such that Peter visited it every day. b. For each day, there was a different museum that Peter visited. Second, Fodor and Sag () present an observation that shows that indefinites have readings that cannot be modeled by regular quantifier scope interaction. They note that specific indefinites, or what they call referential indefinites, are able to escape ‘scope islands’, while other quantifiers are not. Scope islands are structural configurations that do not allow quantifiers to take scope over them. Scope islands are created for instance by that-complements (with lexical heads) as in () and () or by conditionals as in () and (). Example () can have a reading such that there is a rumor that a student of mine had been called before the dean, as in (a). This reading does not entail that there is a student of mine. The example has a second reading according to which there is a student such that there is a rumor that the student had been called before the dean, as in (b). This reading entails that there is a student. Interestingly, the universal quantifier each cannot get the wide scope reading (b). This is explained by the assumption that the that-clause is a scope island for quantifiers. () and () show that conditionals are also scope islands, but allow indefinites to take wide scope.

OUP CORRECTED PROOF – FINAL, 21/1/2019, SPi



  

()

John overheard the rumor that a student of mine had been called before the dean. a. the rumor that there is one student (but there might be no student of mine) b. there is a student of mine such that there is a rumor . . .

()

John overheard the rumor that each of my students had been called before the dean. a. the rumor that each of my students . . . b. *for each of my students there is a rumor that the student . . .

()

If a friend of mine from Texas had died in the fire, I would have inherited a fortune. a. if any friend of mine from Texas . . . b. there is a friend of mine and if he . . .

()

If each friend of mine from Texas had died in the fire, I would have inherited a fortune. a. if all of my friends from Texas had died in the fire, . . . b. *for each of my friends, if one of them . . .

Fodor and Sag () assume that the indefinite is (lexically) ambiguous between a referential expression (similar to proper names or this-indefinites), and an existential expression (similar to existential quantifiers). They assume that the referential indefinite is scopeless and therefore gives the appearance of escaping the island, while the existential behaves like other quantifiers and therefore cannot escape the island. Thus they can explain the behavior in scope islands by assuming a lexical ambiguity. Third, based on their ambiguity theory they make a prediction for the reading of indefinites in scope islands with two quantifiers, namely that there is no intermediate reading. This prediction was refuted and different accounts proposed (see for discussion §. below; Farkas ; Abusch ; Schwarz ; von Heusinger : §)

.. Epistemic contrasts Epistemic specificity comes closest to the notion of the ‘referential intentions’ of the speaker, paraphrased as “the speaker has a particular individual in mind” (Karttunen : ). Farkas () uses the term ‘epistemic specificity’ to describe the contrasts related to referential intentions and that we can best discuss in sentences without other operators, as illustrated in () from Karttunen (: ). ()

a. b. c. d. e. f.

I talked with a logician. I talked with Rudolf. I talked with a famous philosopher. I talked with the author of Meaning and Necessity. . . . , and not with a linguist. . . . , therefore I now understand the first and second syllogism.

OUP CORRECTED PROOF – FINAL, 21/1/2019, SPi

  



The speaker could use (b–d) instead of (a) if the speaker has talked with Rudolf Carnap, a famous philosopher and the author of Meaning and Necessity, and the speaker has this referent in mind. Thus (a) in its epistemic specific reading is an answer to the question ‘Who did you talk with this morning?’. The non-specific reading of the indefinite is an answer to ‘What kind of person did you talk with this morning?’ This reading is favored by the continuations in (e–f ) and contrastive accent on logician. Fodor and Sag (: , ex. ()) argue based on example (), repeated in (), that epistemic specificity behaves very similarly to referential and scopal specificity and is therefore one instance of the same category. () a. A student in the syntax class cheated on the final exam. It was the guy who sits in the very back. b. A student in the syntax class cheated on the final exam. I wonder which student it was. In the specific interpretation (a) the speaker ‘has a referent in mind’ and makes an assertion about this referent. In the non-specific reading (b), the speaker just makes an assertion that the set of students in the syntax class who cheated on the final exam is not empty. The contrast between epistemic specific and non-specific indefinites seems intuitively clear and for many quite obvious, but it is often difficult to operationalize the contrast. First of all, there is no contrast in truth conditions. Second, in the absence of semantic operators we can use the indefinite in either interpretation as an antecedent for an anaphoric expression. Third, additional factors may play a role and overwrite epistemic specificity. For example, the indefinite in () is in subject (and thus preferred topic) position and therefore more likely to be interpreted as specific. Fourth, past tense promotes an epistemic specific reading, in particular in sentences with a first person singular subject, as in (). Still we can use a kind of ignorance test, which generally works much better with subjects in the third person, as in (). And we can also use imperatives to clearly select non-specific interpretations of indefinites: () a. I talked to a student of mathematics. (And I do know her). b. I talked to a student of mathematics. (#And I do not know her). () a. Peter talked to a student of mathematics. (And I do know her). b. Peter talked to a student of mathematics. (And I do not know her). () a. #Talk to a (specific) student of mathematics. b. Talk to a student of mathematics. The contrast between epistemic specific readings and epistemic non-specific readings is often aligned to Donnellan’s () contrast between a referential reading and an attributive reading of definites (see Partee  for discussion). Neale () and Heim () provide overviews of the controversial discussion on the semantic or pragmatic status of this distinction.

OUP CORRECTED PROOF – FINAL, 21/1/2019, SPi



  

.. Partitive contrasts Partitive specificity has been related to other types of specificity since Enç (), who discusses direct object marking in Turkish. Indefinites generally introduce new discourse referents together with a description. Partitive indefinites pick out one (unmentioned) referent from a discourse-familiar group. Obviously, such indefinites presuppose existence and behave like strong quantifiers. Enç () relates partitive indefinites to Pesetsky’s (: ) notion of d(iscourse)-linking that accounts for the different presuppositions of ‘which’ versus ‘who’. ()

a. Oda-m-a birkaç çocuk gir-di. room-.sg.-Dat. several child enter-Past ‘Several children entered my room.’ (Enç : ex. ()) b. Iki kız-ı tanı-yor-du-m. two girl-Acc. know-Prog.-Past-.sg. ‘I knew two [of the] girls.’ (Enç : ex. ()) c. Iki kız tanı-yor-du-m. two girl know-Prog.-Past-.sg. ‘I knew two girls.’ (Enç : ex. ())

(a) introduces a set of children, and the accusative case in (b) indicates that the two girls are part of that set. The expression iki kızı ‘two girls-acc’ presupposes existence, while the unmarked iki kız ‘two girl’ refers to girls not included in the set of children. Enç takes this observation as a strong indicator that the accusative expression is specific and proposes that specificity can be derived from partitivity, or more exactly from familiarity of the superset involved. Diesing () and de Hoop () take partitivity as an instance of Milsark’s () contrast between a weak (cardinal, non-specific) and a strong (presuppositional, specific) interpretation. However, partitive specificity is orthogonal to referential (or scopal) specificity, as in (a) and to epistemic specificity, as in (b) (see Abbott ; Farkas ; van Geenhoven  for discussion). ()

a. John wants to marry one of Steve’s sisters. (He doesn’t care which one.) b. One of Steve’s sisters cheated on the exam. (We have to find out which one.)

Partitives and specific indefinites show similar behavior: they restrict the set of possible referents, they show wide scope behavior and can introduce discourse referents even in the scope of semantic operators. Therefore, Farkas () takes them as one kind of specific indefinites. However, I assume that partitives are specificity related phenomena, rather than specific.

.. Topical contrasts Topicality has also been closely related to specificity. Portner and Yabushita () assume that the restrictor set of the indefinite is topical, either explicitly as in the case of

OUP CORRECTED PROOF – FINAL, 21/1/2019, SPi

  



partitives, or implicitly via other information. Portner and Yabushita () argue on the basis of Japanese and Portner () of Chinese data that a topical and very narrow restrictor set triggers specificity effects. This perspective on specificity is very similar to Schwarzschild’s () extreme domain restriction approach, where the restrictor set is a singleton (‘singleton indefinites’) and the indefinite shows specificity effects. A different approach assumes that the whole indefinite is topical in the sense of a sentence or ‘aboutness’ topic (see Cresti ; Endriss ). The intuitive idea is that the speaker introduces the topic by a separate speech act. Thus the topic is identified independently of the assertion in the sentence, giving rise to the typical specificity contrasts.

.. Contrasts in noteworthiness English has an indefinite use of the proximal demonstrative this that introduces an indefinite that does not interact with other operators, much like a deictically used demonstrative. The use of indefinite this is licensed if it introduces a discourse referent that becomes the theme of the subsequent discourse (Prince a) or that is ‘noteworthy’, that is, it has an unexpected and interesting property (McLaran ; Ionin ), as illustrated by the contrast below (MacLaran : ). ()

a. He put √a/#this  cent stamp on the envelope, so he must want it to go airmail. b. He put √a/√this  cent stamp on the envelope, and only realized later that it was worth a fortune because it was unperforated.

The indefinites in both sentences introduce a discourse referent, and there is no other operator and no referential versus attributive contrast. Nothing prevents either indefinite from introducing a discourse referent. Still there is an important difference: the marked indefinite in (b) introduces a significant theme for the subsequent discourse. Indefinite this signals particular, interesting, new information, while unmarked indefinites just signal that they introduce a discourse referent with more or less important properties. Noteworthiness does not necessarily force frequent anaphoric links, as Ionin (: , based on an example of Prince a: ) illustrated with the following example, where indefinite this refers to a noteworthy property, while the regular indefinite article does not. ()

A: Why do you like him? B: Oh, he has this/#a nose . . .

.. Contrasts in discourse prominence Gernsbacher and Shroyer (: ) reported a minor contrast between this-indefinites and indefinite noun phrases with the indefinite article with respect to anaphoric expressions that continue referents introduced by the two forms: expressions anaphoric to thisindefinites are more often realized as zero or pronouns than anaphors to indefinite noun

OUP CORRECTED PROOF – FINAL, 21/1/2019, SPi



  

phrases with the indefinite article, while definite descriptions are more often used to link to regular indefinites than to this-indefinites. Chiriacescu and von Heusinger () reported similar effects between case-marked indefinite direct objects and their non-case-marked versions in Romanian. A second discourse property is characterized by the parameters ‘referential persistence’ and ‘topic shift potential’. Chiriacescu () showed that informants use a higher frequency of anaphoric expressions or a higher referential persistence as well as a higher rate of topic shifts if the indefinite has special marking, such as indefinitethis in English, indefinite so’n in German or differential object marking in Romanian. The forward referential properties in the last two subsections were illustrated by the contrast between this-indefinites (or other specially marked indefinites) and indefinite noun phrases headed by the indefinite article, which may be ambiguous between a specific and a non-specific reading. Alternatively, the indefinite article could also be characterized as underspecified for specificity or referentiality, while indefinite this is specified for referentiality (see Gundel et al. ). In a paragraph continuation experiment, Deichsel () gave short contexts to participants and asked them to continue the fragment with five naturally sounding sentences. One kind of context is (a), where the indefinite dies noun phrase diesen Kommillitonen (‘this fellow student’) is picked up by the pronoun ihn in the next sentence. In the second context, the indefinite noun phrase einen Kommillitonen (‘a fellow student’) is also picked up by a pronoun thus enforcing a specific reading. ()

a. Laura will diesen Kommilitonen mit auf die Party bringen. Sie hat ihn im Englischkurs kennen gelernt. ‘Laura wants to bring this fellow student to the party. She met him in the English class.’ b. Laura will einen Kommilitonen mit auf die Party bringen. Sie hat ihn im Englischkurs kennen gelernt. ‘Laura wants to bring a fellow student to the party. She met him in the English class.’

If the specific interpretation of the indefinite noun phrase with the indefinite article has the same discourse properties as the indefinite dies noun phrase, we predict that the referential structure of the continuation stories would be the same. However, diesindefinites were significantly more discourse prominent than noun phrases with the indefinite article; they showed four times more anaphoric expressions and they were used ten times more as a topic (Deichsel : ). This result on the discourse properties suggests that dies-indefinites have different discourse properties than specific indefinites with an indefinite article, even though their referential properties in sentences seems to be very similar if not identical.

.. A family resemblance notion of specificity This short survey shows that specificity is a linguistic notion covering different contrasts that are related to each other in some way or other. But not all notions can closely be linked

OUP CORRECTED PROOF – FINAL, 21/1/2019, SPi

  



to the original intuition that the speaker has the referent of a specific indefinite ‘in mind’. We therefore distinguish between specificity in a narrow sense and specificity related notions, which show similar features, but can be orthogonal to specificity in the narrow sense. Family tree of specificity specificity related phenomena

specificity (narrow sense)

referential-semantic notion

(i) referential specificity (opaque context)

(ii) scopal specificity

semantic-pragmatic notion

(iii) epistemic specificity

familiarity based notion

(iv) partitivity, presuppositionality, backgrounding

(v) topicality

discourse prominence (forward reference)

(vi) noteworthiness

(vii) discourse prominence

There are different ways to group these types of specificity together: Fodor and Sag () assume that type (i) to (iii) are one and the same phenomenon. Farkas () argues that epistemic specificity, scopal specificity—which includes referential specificity in her view—, and partitive specificity are independent of each other and can cross-classify. Still they show the common effect of reducing the restrictor set of the indefinite, that is, the set of potential referents is restricted to few or possibly to only one element. This concept of ‘referential stability’ (Farkas and von Heusinger ) can generalize over various types of specificity and motivate why languages use the same encodings for these types. Kamp and Bende-Farkas () argue that the core notion is epistemic specificity and an appropriate analysis can also explain referential and scopal effects. Von Heusinger () modifies this picture into an analysis of specific indefinites as ‘referentially anchored’ indefinites covering (i)–(iii) and relating it to (vi) and (vii). Enç () derives specificity contrasts from her core notion of partitive specificity, while Prince (a) and Ionin () have focused on the discourse properties of the ‘referential intention’. Wright and Givón () argue that the semantic contrasts of (i)–(iii) can be derived from a discourse pragmatic notion of specificity. Many other groupings and categorizations of these contrasts are found in the literature.

. T  

.................................................................................................................................. As illustrated in the last sections, specificity is a semantic–pragmatic notion with various instances that are related by familiarity resemblance. There is no agreement in the literature on what counts as core data for specificity and what only as related phenomena (this has partly to do with the fact that most European languages have no clear overt marking for specificity). Theories also differ with respect to the semantic–pragmatic level they take as the essential contribution of linguistic forms.

OUP CORRECTED PROOF – FINAL, 21/1/2019, SPi



  

I illustrate the great variety of theoretical approaches with four families of theories on specificity: (i) Exceptional scope theories, (ii) Referential theories, (iii) Familiarity theories, and (iv) Discourse prominence theories. This division into four families of approaches follows the different semantic–pragmatic perspectives on indefinite reference, discussed in §§..–.., and it is partly orthogonal to one classic distinction into structural ambiguity theories (scopal theories) and lexical ambiguity theories. The lexical ambiguity approach assumes two indefinite articles in the lexicon, a referential indefinite article and an existential indefinite article, which happen to be homophonous in English (Karttunen ; Fodor and Sag ; Kratzer , among others), but may be overtly expressed in other languages, such as in Maori or many other languages (see §.). The scopal theories are discussed in the next section.

.. Exceptional scope theories In early approaches, (referential) specificity was associated with the different readings of indefinite noun phrases in opaque contexts (Quine ; Karttunen , / ; Fodor ; Abbott ). Sentence (), repeated as (a), from Karttunen (/, ex. ()), is ambiguous in three ways (b–c): ()

a. b. c. d.

Mary may want to marry a Swede. There is some Swede whom Mary may want to marry. It may be the case that there is some Swede whom Mary wants to marry. It may be the case that Mary wants her future husband to be a Swede.

The readings of indefinites in opaque contexts, as in (), show a contrast similar to the contrast of the de re versus de dicto readings of definite noun phrases (Quine , ; Partee ). Karttunen (/) argues on examples like () that there could not be a categorical distinction between a specific reading of an indefinite and a non-specific reading, since that would not account for the three readings. He rather suggests a scopal analysis of specificity: the indefinite is represented as an existential quantifier that can be Q(uantifier) R(aised) and therefore freely take scope over the other two operators yielding the correct truth conditions for the three readings. An alternative to the QR-theory is the type-shifting approach (Zimmermann ; van Geenhoven and McNally ) which is based on the observation that indefinites, like definites, can either behave as regular arguments of type e, as properties of type , or as quantifiers of a higher type (see Partee ). Intensional verbs like to want, to seek, to hunt, to owe, etc. can take the indefinites in any of their forms. In the property type the indefinite is semantically incorporated into the predicate and does not introduce a discourse referent or allow for existential entailments. A more recent variant of the type-shifting approach is Chung and Ladusaw’s () distinction between two compositional operations, Saturation and Restriction: the indefinite can be of type e and then saturate an argument of the verb, or it can be of type and modify or restrict the event expressed by the verb.

OUP CORRECTED PROOF – FINAL, 21/1/2019, SPi

  



Fodor and Sag () raise a critical argument against the QR analysis. They show on indefinites in scope islands, as in ()–() in §.. that the indefinite does not behave like a quantifier, since it can escape scope islands, which is not possible for regular quantifiers. There are various analyses of this ‘exceptional scope behavior’, or ‘long distance indefinites’. I briefly discuss three prominent approaches: (i) The longdistance scope shift approach ascribes fewer restrictions on movement to the existential quantifier (Abusch ; von Stechow ; Schwarz ). The other approaches all assume that the indefinite is not moved but stays in situ. (ii) In the existentially closed choice function approach, scope is derived by assuming that the indefinite article introduces a choice function variable that can be bound freely at different scope sites (Reinhart ; Winter ); (iii) Under the singleton indefinite or implicit domain restriction approach the indefinite is enriched by descriptive material until it expresses a singleton and therefore gives the illusion of wide scope, similarly to other domain restriction approaches (Portner ; Schwarzschild ). First, the long-distance scope shift approaches assume that indefinites along with other weak quantifiers can escape scope islands, they scope freely (upwards), as in (c) (Abusch ). This approach has the advantage of accounting for the data without introducing new concepts into the analysis, just by loosening some restriction. However, the disadvantages are that there is no uniform behavior of quantifiers and that the theory overgenerates in predicting intermediate scope readings, even if they are not acceptable. However, von Stechow (), Schwarz (), and Heim () opt for the long-distance scope shift or flexible scope theory of indefinites. Second, existentially closed choice function approaches assume that the indefinite article introduces a choice function f. A choice function f is a function that selects one element out of a non-empty set or, more generally, assigns to a set one of its elements (Reinhart ; Winter ). This choice function can be bound by existential quantifiers at different levels, that is, inside the relative clause yielding the narrow scope reading, or outside the relative clause but still in the scope of the universal quantifier, as in (d). The advantage of the choice function approach is that we can derive all available readings without QR and therefore without scope island violations. (As we will see in the next section, choice functions can be used to substitute an existential quantifier elaborating a scope theory, or they can be used for substituting a referential expression, modeling a referential approach.) The choice function approach has not only created an original formal tool for indefinites but has also instigated much controversy on the adequacy of this tool. There are problems with empty sets and with readings in downward entailing contexts (see Schwarz  and Chierchia  for an exhaustive discussion). ()

a. Every professor rewarded every student who read a book on the semantics– pragmatics interface. b. intermediate scope: every professor > a book on the s-p-i > every student For every professor there is a certain (possibly different) book on the s-p-i, such that the professor rewarded every student who read that book. c. ∀x [professor(x) ! ∃y [book-on-s-p-i(y) & x rewarded every student who read y]]

OUP CORRECTED PROOF – FINAL, 21/1/2019, SPi



   d. ∀x [professor(x) ! ∃f [ch(f ) & x rewarded every student who read f(bookon-s-p-i)]] e. ∀x [professor(x) ! [x rewarded every student who read a book on the s-p-i x that x had liked most.]]

Third, Schwarzschild () proposes an alternative view on the exceptional scope behavior of indefinites. He applies the domain restriction approach adopted with other quantifiers and shows that enriching the descriptive material of the indefinite leads to truth-conditional effects that are equivalent to Kratzer’s approach. The wide-scope reading is entailed by an indefinite that is restricted to a singleton set (‘singleton indefinite’), while the intermediate-scope reading is derived by a restriction resulting in a function that depends on the highest quantifier, as in (e), expressing a functional reading. Even though domain restriction is necessary for other quantifiers, it is not clear whether the restriction to a singleton set is always justified, as examples with partitives show (see Endriss : ).

.. Referentiality or indexicality theories The central intuition underlying a specific reading is the ‘referential intention’ of the speaker, who introduces a discourse referent that he or she can identify, but not the hearer (Dekker ).³ Focusing on this aspect, several analyses have developed. We discuss in the following (i) Fodor and Sag’s () indexicality theory, (ii) Kratzer’s () contextually bound choice function approach, (iii) Kamp’s ‘entity representation’ approach, (iv) von Heusinger’s () referential anchoring approach, and (v) the application of referential anchoring in Sæbø’s () reported speech contexts. (i) Indexicality theory Fodor and Sag (: ) give a purely indexical interpretation of specific (their ‘referential’) indefinites, which refer to the intended referent. They propose a Kaplan style semantics of specific indefinites, but do not give an explicit definition. Heim (: ex ()) formulates the original idea in a two dimensional semantics with a context set c and an evaluation point i. The indexical or direct referential meaning of an indefinite only depends on the utterance context, as it is the case for regular indexical expressions. ()

a. [[aquant α]]c,i = λQ. ∃x[[[α]]c,i(x) & Q(x)] b. [[aref α]]c,i is defined only if there is a unique individual that the speaker of c has in mind in c, and this individual is in [[α]]c,c; where defined, [[aref ]]c,i = this individual.

³ Barbara Abbott kindly informs me that there are specific uses of indefinites, as in (i) where the referent is also hearer known, but still discourse new. (i)

Someone got up on the wrong side of the bed this morning.

OUP CORRECTED PROOF – FINAL, 21/1/2019, SPi

  



(a) is the representation for the existential meaning and allows for the scopal mobility, as with other quantifiers. (b) is the representation for the specific indefinite, which introduces a referent that only depends on the context of utterance like other indexical or demonstrative expressions (see also the modification by Ionin  below). The indexical semantics (b) seems to be adequate for the direct referential meaning of indefinite this in English (and other languages), but not for the specific reading of an indefinite with the indefinite article, as shown by the existence of the intermediate scope reading, which is neither a direct referential nor a quantificational reading that obeys scope islands. (ii) Kratzer’s () contextually bound choice function approach Kratzer () combines the indexical approach of Fodor and Sag () with the choice function approach and assumes contextually bound choice functions, that is, the choice function variable is contextually determined, entailing a wide-scope reading (similar to the original Fodor and Sag approach). The intermediate reading, however, can be forced by a bound variable in the descriptive content, for example, book on the s(emantics)–p(ragmatics) i(nterface) she has recommended. Thus the set of books depends on the professor and the selected element co-varies with the values for professors, as in (a), yielding an ‘apparent intermediate’ or ‘pseudoscope’ reading (Kratzer ). The representation (a) leads to a new problem: if two professors have recommended the same books, the choice function f would select the same book for both professors, since the sets are extensionally identical. This is too strong a restriction for the intermediate reading, which intuitively allows for different choices of books depending on professors, even if they recommend the same set of books. Therefore, Kratzer () introduces a ‘Skolemized’ choice function in (b), that is, a contextually given Skolem function g that takes one individual argument (or parameter) and a set argument and yields one element of the set. () a. ∀x [professor(x) ! [x rewarded every student who read f(book-on-s-p-i x recommended)]] b. ∀x [professor(x) ! [x rewarded every student who read g(x)(book on the s-p-i x has recommended)]] with g assigning choice functions to professors such that the choice function selects a book on the s-p-i that the professor has recommended. Using choice functions allows dissociating the scope of the indefinite from its descriptive content. While the descriptive content stays in situ, the choice function variable can be bound at different places in the sentence resulting in different scopal properties of the indefinite. Choice functions also capture the intuitive idea that a specific indefinite can be understood as selecting an element out of a set according to a certain method. In a very general sense, choice functions are term-creating operations corresponding to type shifting from a set to an individual, which seems necessary for independent reasons. On the other hand, choice function approaches are controversial, as the representation of indefinites with choice functions seems to be too flexible: Choice functions do not allow for existential entailments, since they are not defined for

OUP CORRECTED PROOF – FINAL, 21/1/2019, SPi



  

the empty set (see above). It is an open issue whether this is a welcome result for fictional objects (see Ruys ) or whether this has to be repaired (see Winter ). Existentially bound choice functions predict wrong readings in downward entailing contexts (see Schwarz ; Chierchia  for discussion and additional restrictions on choice function construals). This problem, however, does not arise with contextually bound choice functions (see Kratzer , but Chierchia ). A final criticism is that once we are forced to use Skolemized choice functions (i.e., functions with one individual argument and a set argument) we may as well take Skolem functions with n-individual arguments and do without the problematic choice functions altogether (see Hintikka ; Steedman ; Kamp and Bende-Farkas ; Onea and Geist , among others). (iii) Kamp’s ‘entity representation’ approach Kamp () and Kamp and Bende-Farkas (, based on a manuscript from ) develop a version of Discourse Representation Theory that allows for the representation of direct referential expressions like proper names and demonstratives. They distinguish between external anchors, that is, functions that relate a discourse referent to an object in the world (like proper names to their bearers) and internal anchors, that is, functions that relate the representation to other discourse referents. Kamp (, ) elaborates this approach and models external anchors as ‘Entity Representations’, that is, a representation that directly links the discourse representation to the intended referent. While the Entity Representation is part of the speaker representation, the recipient has to establish the same Entity Representation for proper names and demonstratives via principles of communication, that is, principles that allow building the common ground from speaker and recipients’ representations. Epistemic specific indefinites are similar to proper names and demonstratives in that they both require Entity Representations, that is, direct reference to their objects, just as in the analysis of Fodor and Sag (). However, they are different from proper names and demonstratives in that the communicative principles do not force the recipient to establish such a stable discourse referent—therefore, the properties of specific indefinites vary between direct referential expressions and existential expressions. (iv) von Heusinger’s () referential anchoring approach Von Heusinger () extends the Kaplan style semantics of Fodor and Sag () in a different direction. While other approaches assume that (epistemic) specific indefinites have referents that can or must be identified by the speaker, von Heusinger also allows for other potential agents in the linguistic context, which are able to establish reference. He starts from the semantics in () and allows referential anchoring not only to the speaker, but also to ‘a context agent’. The example () from Higginbotham () illustrates such a case: the indefinite ‘a (certain) student from Austin’ is in both (a–b) specific, but in (b) not speaker known or identifiable. The specificity can rather be modeled by ‘referential anchoring’ to the subject George.

OUP CORRECTED PROOF – FINAL, 21/1/2019, SPi

  



() [[aref α]]c,i is defined only if there is a unique individual that the speaker in c has in mind in c, and this individual is in [[α]]c,c; where defined, [[aref α]]c,i = this individual. () a. George (to Lisa): I met a certain student from Austin today. b. Lisa: George said that he met a (certain) student from Austin today. Von Heusinger ( based on Egli and von Heusinger ; von Heusinger ) cashes out the idea of referential anchoring in terms of ‘indexed epsilon terms’, which are equivalent to parameterized choice functions discussed above. The idea is that the indefinite article can translate into the complex pronominal element fx (originally εi) with x being a parameter that must be bound by a context agent. The function f applied to the anchor yields a choice function that is applied to the set denoted by the descriptive content of the indefinite yielding the referent. Additional evidence for variation in what can act as the anchor comes from the analysis of German specificity adjectives ein gewisser and ein bestimmter (‘a particular’, ‘a certain’). Ebert, Ebert, and Hinterwimmer (: ) observe that the specificity adjective ein bestimmter in (a) can establish a referential link to the speaker or the subject Peter, as indicated by explicit information about the ignorance of the speaker. The adjective ein gewisser, however, must be linked to the speaker. ()

a. Peter Peter CD – CD –

sucht schon seit Stunden nach searches already since hours after keine Ahnung, welche genau er no idea which.one exactly he

einer bestimmten a bestimmt sucht. searches

b. Peter Sucht schon seit Stunden nach einer gewissen Peter searches already since hours after a gewiss CD – # keine Ahnung, welche Genau er sucht. CD – no idea which.one exactly he searches ‘Peter has been looking for a certain CD for hours now—I have no idea which one exactly he is looking for.’ We can summarize the characterization of referential anchoring as follows: in the prototypical case the anchoring function takes the speaker as its argument, and its value is the referent of the specific indefinite. However, other arguments are possible (see von Heusinger  for an extensive discussion). (v) Sæbø’s () on specific indefinites in speech reports I present a final observation that shows that specific indefinites not only carry a referential intention but also show referential effects in a discourse: Sæbø () analyses specific indefinites that serve as antecedents for direct referential terms in speech reports. Sæbø (: ) provides the following example (a) and (b): Suppose that Sæbø says (a) to his wife and that his wife has no idea about the identity of the ‘someone else’. Still, the ‘someone else’ can later utter (b).

OUP CORRECTED PROOF – FINAL, 21/1/2019, SPi

 ()

   a. I have met someone else. b. He has told his wife he has met me.

Sæbø observes that the use of the direct referential pronoun me in the speech report (b) is conspicuous, as it is more specific than its antecedent someone else in (a). This is surprising since the content of a reported speech is equal or less informative than its original utterance. However, the pronoun me in (b) is more informative than the indefinite someone in (a). Sæbø argues that this is only possible if the indefinite in (a) has a referential intention such that in the reported speech that referential intention can be spelled out by a pronoun (see for details Sæbø ).

.. Familiarity theories The main distinction between definite and indefinite expressions in a discourse model is that definites introduce familiar discourse referents and indefinites novel ones (Kamp ; Heim ). In section .. we discussed instances of partitive indefinites as in (), where in the partitive reading the indefinite introduces a discourse referent that is part or a subset of an already introduced discourse referent (here: children), see Enç (). ()

Several children entered my room. I knew a girl.

Partitives, or more correctly implicit partitives, are an instance of inferrable expressions, expressions that have a more complex referential structure than simple definites or indefinites. Inferrable definites as well as inferrable indefinites introduce (i) a novel discourse referent, by (ii) linking it via a hearer given relation to (iii) a discourse given anchor (or antecedent). The difference between definite inferrables and indefinite inferrables is the uniqueness (or exhaustiveness) condition (Prince b, see §..). Enç (:  ex. ()) models this complex structure by introducing a second index for the existence of an anchor for the expression (she assumes for partitivity a subset relation for plural and an element relation for singular). Each noun phrase comes with two indices: the index i for the familiarity status of the introduced referent and an index j for the familiarity status of the anchor. A specific indefinite has a novel i, but a familiar j. ()

Every [NP α] is interpreted as α(xi) and xi ⊆ xj if NP is plural {xi} ⊆ xj if NP is singular

There are various ways to modify this model: Roberts (: ) introduces the term ‘weak familiarity’ by focusing on the linking relation, rather than the availability of the anchor. All such familiarity approaches have in common the fact that they account for the presuppositionality of partitive indefinites as a way to link the indefinite to the previous discourse. Familiarity theories are extensions of theories that distinguish between definite and indefinite noun phrases based on their discourse properties. So they are hearer oriented and cannot model the speaker oriented perspective of ‘having a

OUP CORRECTED PROOF – FINAL, 21/1/2019, SPi

  



referent in mind’. Therefore, partitives easily cross-classify with referential, epistemic, and scopal specificity (see §.., ex. ()).

.. Discourse prominence theories There are various extensions of models of specificity trying to account for the function of indefinites in the subsequent discourse. Givón’s () work on ‘topic continuity’ and Wright and Givón’s () analysis of the pragmatics of indefinite reference assume that special indefinites such as this-indefinites or article forms in their early grammaticalization stages (such as xad ‘one’ in Hebrew) signal discourse prominence of the referent. This discourse prominence is expressed by early topic shifts, high referential persistence, etc. Givón assumes that the pragmatic function triggers the semantic properties of specific indefinites. But we can also take the reverse perspective: it is the sentence semantic properties of indefinites that govern their discourse pragmatic ones. The two approaches we discussed in §.. and §.. differed in the assumption of whether or not epistemic specificity is part of the semantic representation. But even in an approach that does not assume an immediate sentence semantic effect, some authors assume a discourse effect, as illustrated by the position of Stalnaker (: ), who holds that the difference between specific and non-specific indefinites is crucial for discourse structure: “The account I am sketching suggests that this difference matters, not to the interpretation of the indefinite expression itself, but only to the evaluation of subsequent statements made with pronouns anaphoric to the indefinite expression.” A similar position is taken by Kamp (), who assumes that the specific use of an indefinite strongly signals that the speaker intends to say more about the referent and thus that the indefinite serves as the antecedent for a referential chain. There is a small, but growing corpus of evidence that this link exists (see Givón ; Gernsbacher and Shroyer ; Chiriacescu and von Heusinger ; Deichsel and von Heusinger , among others), but there is no fleshed out theory that bridges the semantic properties with the discourse pragmatic properties of indefinites. A pragmatic account may go like this: the use of a specific indefinite forces the hearer to establish a permanent discourse referent. By coherence principles and Gricean maxims, the speaker would only force the hearer to do that if she intends to say more about that referent. On the other hand, a semantic account may assume the following relation between referential properties of specific indefinites and their discourse properties: if the speaker has a referent in ‘mind’, then he or she generally plans to provide more information about this specific referent—thus there is an implicature from a specific indefinite to a discourse prominent indefinite. Jeanette Gundel (pc) informs me about another variant of the accounting for the relation between semantics and discourse pragmatics: a specific indefinite is referential in the sense that it introduces a referent while processing the sentence. This referent is then available for further predication. A non-specific indefinite just introduces an attribute

OUP CORRECTED PROOF – FINAL, 21/1/2019, SPi



  

or concept, in terms of Gundel et al. () ‘type identifiable’, which does not suggest itself for continuation, but does not block referential continuation either. Now, we have to explain why a referent once introduced in the discourse model ‘creates a pressure’ to re-use it.

.. Evaluation of the theories I illustrate the differences between the four families of approaches on the following selected issues: (i) are the referential theories too strong? (ii) who can act as holder of the referential intention? (iii) what is the status of the familiarity theories? and (iv) what is the relation between the sentence semantics of specific indefinites and their discourse pragmatics? The main controversy between existential theories of specificity and referential theories (cf. §.. and §..) is that referential theories correctly represent the referential intention, but are too strong with respect to the truth conditions of a sentence, as illustrated in (): even if the speaker has a particular poem of Pindar in mind, the utterance of (a) only contributes an existential statement, as in (d). Since the hearer cannot know which poem the speaker has in mind, the speaker commits himself or herself only to the existential statement, which would be true even if Ann would read a different poem from what the speaker has in mind (see for discussion King ; Ludlow and Neale ; but Kratzer ). ()

a. b. c. d.

Ann read a poem of Pindar. Ann read thisindef poem of Pindar. Ann read a certain poem of Pindar. Ann read sm poem of Pindar.

The lexical expressions thisindef and a certain in (b–c) contribute the referential reading by their semantic content and render the utterance false if Ann is reading a different poem, as predicted by the referential theory. If we assume that the indefinite article is ambiguous between a specific and non-specific reading, the speaker is only committed to the weaker reading, but this is independent of the assumption that the specific reading has a referential representation. (ii) There is a second controversy: Fodor and Sag’s () referential semantics for specific indefinites seems to fit quite well the behavior of this-indefinites, but not of a certain-indefinites or indefinites with a (specific) indefinite article. So we might want to distinguish between three classes of indefinites: referential indefinites, specific indefinites, and non-specific indefinites. But then we need a different semantics for specific indefinites, which was sketched in §.. (iv) under the notion of ‘referential anchoring’. The holder of the referential intention can shift from the speaker to other discourse referents that are able to hold such an intention. (iii) The referential anchoring must not be taken as just an instance of discourse linking as assumed in the familiarity theory for partitive indefinites. Such theories model partitive indefinites by presupposing (rather than asserting) the

OUP CORRECTED PROOF – FINAL, 21/1/2019, SPi

  



restrictor set of the indefinite and they also further restrict that set out of which an indefinite selects its referent. Still they do not deal with the identification of the referent. This means partitivity is orthogonal to specificity, and familiarity theories may not be able to account for specificity if understood as the linguistic expression of referential intention. (iv) A very open issue is the interaction between the sentence semantic contribution of specific indefinites and their discourse pragmatic ones as discussed in the last subsection. Here we have to learn much more about the interaction, but we also have to model the communicative interaction of speaker and hearer (see Kamp ). I have to leave this perspective open for further research.

. S

.................................................................................................................................. The semantic–pragmatic category ‘specificity’, which is informally described by the speaker ‘having a referent in mind’ is used to describe various semantic and pragmatic contrasts. I have argued that there is a core notion of specificity underlying the intuitive concept, namely referential anchoring. The referent of a specific indefinite is dependent on some discourse participant. The anchor must be familiar to speaker and hearer, while the content of the anchoring function must be unfamiliar to the hearer (to distinguish specific indefinites from definites). Still the hearer has to accommodate the fact that there is a function and must establish a permanent representation for the specific indefinite. I have shown that this approach is quite flexible and can account for various particular constraints associated with special specificity markers. However, it cannot explain all phenomena associated with different types of specificity, which might get different kinds of explanation (such as genuine intermediate scope indefinites via embedded topics, see Ebert, Endriss, and Hinterwimmer ). I discussed the similarities between specific indefinites and partitive indefinites as well as ‘topical’ indefinites and showed that they are independent notions, but with similar effects. Finally, I compared the semantic properties of specific indefinites with their discourse pragmatic functions, which open up a new domain of research, namely the interaction of semantic and pragmatic properties with discourse properties of nominal expressions.

A I would like to thank the two editors of this volume, Barbara Abbott and Jeanette Gundel for their encouragement, patience, and very helpful and constructive comments. I would also like to thank the following persons for comments and long discussions: Chiara Gianollo, Cornelia Ebert, Stefan Hinterwimmer, Hans Kamp, Edgar Onea, and Umut Özge. I gratefully acknowledge that the research for this chapter has been funded by the German Research Foundation (DFG) as part of the SFB  “Prominence in Language” in the project C “Conceptual and referential activation in discourse” at the University of Cologne.

OUP CORRECTED PROOF – FINAL, 21/1/2019, SPi

  ......................................................................................................................

D E R E /D E D I C T O ......................................................................................................................

    

. I

.................................................................................................................................. T de re/de dicto distinction is based on the ways in which (broadly speaking) intensional operators can affect the interpretation of other elements—especially noun phrases—in a sentence. While it is most commonly discussed in connection with modal expressions, which are standardly construed as quantifiers over possible worlds, entirely parallel phenomena arise relative to tense, construed as involving quantification over times, and other temporal expressions. In some lines of work, both of these dimensions are tied together by talking about situations, which can be seen as parts of worlds extending over a given time span. As we believe that these different dimensions pattern together in terms of the de re/de dicto distinction, we will draw on examples from all of these realms. We begin with a brief historical background of the distinction, and then lay out the basics of a traditional analysis in terms of scope. Next, we review the challenges to such a theory, and then sketch possible revisions to capture the problematic data. We close with some loose ends and possible further areas of related phenomena.

. H 

..................................................................................................................................

.. Early observations and origins of terms The earliest known observation that a modal may affect some elements in a sentence without affecting others is due to Aristotle,¹ who discusses sentences such as the following, each of which is interpretable in the two ways shown:

¹ Dutilh Novaes () traces it to Sophisitics Elenchis, a–a.

OUP CORRECTED PROOF – FINAL, 21/1/2019, SPi

  /   ()

It’s possible for a sitting man to walk. a. A sitting man still possesses the ability to walk. b. It’s possible for a man to simultaneously sit and walk.

()

It’s possible for a man who is not writing to write. a. A man who is not writing still possesses the ability to write. b. It’s possible for a man to simultaneously write and not write.



The (a) interpretations of the sentences above make perfect sense: your ability to walk or write remains even at times when you are not performing these activities. The (b) interpretations are paradoxes, though: no one has the ability to walk and sit or write and not write simultaneously. The distinction hinges on which portions of the sentence are part of the ability (or possibility) being described. If the entirety of the ability being described is the ability to write (or walk), then the sentence makes sense. However, if the ability described includes both writing and not writing (or both walking and sitting), the paradox arises. We can view this difference as a difference of whether or not the modal operator is affecting the interpretation of the phrase sitting in () or not writing in (). According to Kneale (), the next serious discussion of this effect did not appear until  years later, when Peter Abelard distinguished two types of interpretations of modal sentences: de re, where a modal sentence is about a thing (re), and de sensu, where a modal sentence is about a linguistic statement. For instance, the (a) interpretations above would be de re, since they are predicated of a person: the man who is sitting or not writing. The (b) interpretations would be de sensu, though, since the modal expression it’s possible is not predicated of a thing, instead taking only a clausal complement: a man to simultaneously sit and walk and a man to simultaneously write and not write. This seems to be the origin of dividing modal sentences into ones about things (de re) and ones not about things. The next evolution of the terms, again according to Kneale, is a distinction between modal sentences about dicta—roughly the de sensu interpretation of Abelard—and sentences not about dicta. The idea is roughly that certain modal expressions modify what a particular clause says (id quod dicit propositio in Latin). For instance, in (b) above, the possibility modal modifies a linguistic statement akin to ‘a man simultaneously writes and does not write,’ claiming that it is possible for this statement to be true. The first full use of the terms de re and de dicto is due to Thomas Aquinas, who was also the first to define the terms syntactically: ()

A modal proposition is either de dicto or de re. A modal proposition de dicto is one in which the whole dictum is the subject and the modal is the predicate, as when it is said ‘For Socrates to run is possible.’ A modal proposition de re is one where the modal is interpolated in the dictum, as when it is said ‘For Socrates it is possible to run.’ [translated by Dutilh Novaes ]

Aquinas divides the sentence syntactically into the subject and the predicate. The subject may be a full clause (in the de dicto case) or a thing (in the de re case), as illustrated below:

OUP CORRECTED PROOF – FINAL, 21/1/2019, SPi

 ()

     de dicto:

S

Subject

Predicate

For Socrates to run

()

de re:

possible

S

Subject For

is

Predicate

Socrates

it is possible to run

We see a similar (actually clearer) syntactic distinction in English, where certain modals can take noun phrases as subjects and clauses as objects, while others only take clauses: ()

de dicto: S Subject

Predicate

For a man to simultaneously write and not write

()

is

possible

de re: S Subject

Predicate

A man who is not writing

(still) has the ability to write

The terms used by Aquinas were adopted by von Wright () in his book on modal logic. Prior () and Kneale () took up this thread and soon the terms were standard in modal logic. However, they still referred to sentences, and not to noun phrases: a modal sentence was de re if it comprised a modal predicate with a nounphrase subject.

.. The modern distinction Following up on Frege’s () observations about proper names in intensional contexts, Russell () discusses the following pair of examples: ()

a. George IV wished to know whether Scott was the author of Waverley. b. George IV wished to know whether Scott was Scott.

OUP CORRECTED PROOF – FINAL, 21/1/2019, SPi

  /  



While serving as Prince Regent of England in , George IV famously remarked that he wanted to meet “the author of Waverley,” presumably being impressed by the novel. Russell’s example in (a) is meant perhaps to describe the state of affairs after George IV has enough of an idea of who wrote the novel to ask whether Scott did so, but before he knows for sure that Scott wrote the book. Russell points out that although Scott actually is the author of Waverley, (a) means something quite different from (b); as Russell puts it, upon hearing (a), “an interest in the law of identity can hardly be attributed to the first gentleman of Europe” as it could be upon hearing (b). Although Russell does not use the term, his example represents the first clear exposition of a de dicto reading as the phrase is used today. Consider the following quote perhaps attributable to George IV before he has any idea who wrote Waverley: ()

I want to meet the author of Waverley.

All that the Prince means to say is that in situations where his desires are met, he meets the author of Waverley—whoever it is. Since the Prince does not know that Scott wrote this book, he would probably not agree that the expressed desire is to meet Scott. We now call this the de dicto reading of the noun phrase the author of Waverley. Quine () was the first to discuss sentences as exhibiting an ambiguity hinging on the de re/de dicto distinction, although his terminology characterizes the contrast as relational versus notional readings. The examples involve indefinite noun phrases (a sloop, a spy), and both have two readings as shown below: () I want a sloop. a. There is a certain sloop that I want. b. I seek relief from slooplessness.

(de re) (de dicto)

() Ralph believes that someone is a spy. a. There is someone whom Ralph believes to be a spy. b. Ralph believes there are spies.

(de re) (de dicto)

Note that () does not involve an overt clausal complement, though Quine’s proposal (as well as later authors’) analyzes them as equivalent to overt cases. Leaving this issue aside, we take overt cases such as () to be prototypical instances of the modern notion of a de re/de dicto distinction.

.. Basics of a scope theory The first formal account of the de re/de dicto distinction, going back at least to Quine () (with a precursor in Russell ), ties it to the notion of scope. Couched in a possible world theory of modality, the scope theory of de re/de dicto rests on the assumption that predicates in natural language may differ in extension from world to world and from time to time. For instance, the predicate President of the United States always has a singleton set as its extension, but the member of this set may vary

OUP CORRECTED PROOF – FINAL, 21/1/2019, SPi



    

depending on the year. If we represent the year of evaluation as a superscript parameter on the interpretation function, we get the following values: ()

a. ⟦President of the United States⟧²⁰⁰⁷ = {George W. Bush} b. ⟦President of the United States⟧²⁰¹⁴ = {Barack Obama}

Similarly, taking w₀ to be the actual world, and w₁ to be an alternative possible world where Mitt Romney won the  election, we have: ()

a. ⟦President of the United States⟧w⁰ = {Barack Obama} b. ⟦President of the United States⟧w¹ = {Mitt Romney}

The term scope may be defined semantically or syntactically. The semantic scope of an intensional operator comprises the material whose intensional status is directly affected by the meaning of the operator. For instance, in (), the material George W. Bush was President forms the semantic scope of the intensional phrase in . ()

In , George W. Bush was President.

The syntactic scope of an operator, on the other hand, is its c-command domain, defined as the operator’s sister and nodes dominated by its sister. The standard scope theory of de re/de dicto proposes that an operator’s syntactic scope at the level of logical form (LF) maps directly onto its semantic scope, as illustrated below, where the scope of the operator In  is indicated by a dotted box:² ()

a.

2014

S PP In 2007 Bush

S VP

= 1 iff

was President b. ⟦Bush was President⟧2007 = 1 iff c. Bush was President in 2007.

In other words, under the scope theory, an intensional operator sets the world/time parameter (a semantic notion) of material in its c-command domain (a syntactic notion). In this system, any material remaining in the domain of an intensional operator at LF will be interpreted de dicto, as illustrated in the (actually false) LF for the sentence in ().

² Note that for purpose of illustration, we’re ignoring the contribution of the past tense here and pretending that in  is the only temporal operator at play.

OUP CORRECTED PROOF – FINAL, 21/1/2019, SPi

  /   () a.

2014

S PP In 2007



S NP

VP was a senator the president from Illinois

= 1 iff

b. ⟦was a senator from Illinois⟧2007 (⟦the President⟧2007)= 1 iff c. George W. Bush was a senator from Illinois in 2007.

In order to receive a de re reading, a noun phrase must raise to a syntactic position at LF above the operator, as shown in (). Note that we follow Heim and Kratzer () in assuming that adjunction of a raised noun phrase is combined with λ-abstraction over its trace right below the adjunction site. () a.

S NP

2014

S S

the president λ1 PP

In 2007 t1

S

= 1 iff

VP was a senator from Illinois

b. [λx. ⟦t1 was a senator from Illinois⟧2007, g

] (⟦the President⟧2014) = 1 iff

1→x

c. Barack Obama was a senator from Illinois in 2007.

. E 

..................................................................................................................................

.. Affected expressions and linguistic environments We now move on to lay out a more extensive (though by no means fully exhaustive) overview of the empirical landscape with regards to the de re/de dicto distinction, starting with the types of noun phrases that exhibit the relevant readings. In addition to the cases already illustrated above, (strong) quantifiers—such as most and a (even in its strong reading)—can also be interpreted relative to an operator that shifts the time of evaluation. For instance, the following sentences are

OUP CORRECTED PROOF – FINAL, 21/1/2019, SPi



    

most easily understood as conveying de dicto interpretations of the relevant noun phrases: ()

a. In , most senators belonged to the Democratic Party. b. In , a senator from Ohio belonged to the Democratic Party.

Delegations from particular states and political majorities in the U.S. Senate of course change over time, and if we replaced  by another year, the truth of both of these statements may change based on the range of individuals described by the nominal predicates senators/senator from Ohio. The same effect can, of course, be found with modal expressions that shift the world of evaluation: ()

a. It’s possible that most elected senators belong to the Republican Party. b. It’s possible that a senator elected by Michigan belongs to the Republican Party.

Imagine waking up the day after an election before checking the results. In such a scenario, these sentences can be understood to convey that different individuals (whose identity need not be determined at the time of utterance) may have been elected members of the U.S. Senate, resulting in a change of majority (without requiring any changes in party affiliation for any given individual). Crucially, however, these expressions do not have to be interpreted relative to the shifted world (or time), but can also be interpreted relative to the actual world (or time of utterance): ()

a. It’s possible that, by , most senators will be enjoying their retirement. b. It’s possible that, by , a senator from Michigan will be enjoying his retirement.

In these examples, the relevant individuals are understood to be ones that actually are senators, in the real world at the time of utterance (in ), not the ones that are senators in  or any other set of possible utterance-time senators. Thus, the relevant noun phrases are said to have a de re interpretation: the noun phrase is anchored in the utterance context with respect to the time and world of evaluation, rather than being evaluated relative to the time and world at which the main proposition expressed by the sentence is evaluated. The same is also available for definites and relative to time shifters, as was shown in () above. While both de re and de dicto variations along the temporal dimension are very easy to come up with, modal examples typically require more care in constructing an appropriate context, so it is worth reflecting for a moment on some of the general properties of types of contexts where the difference between de re readings and de dicto readings becomes relevant. With regards to epistemic or doxastic modal operators, one typical such situation is one where the extension of a given predicate in the actual world does not agree with all of the epistemically (or doxastically) accessible worlds. In other

OUP CORRECTED PROOF – FINAL, 21/1/2019, SPi

  /  



words, in at least some of the worlds that a speaker considers as viable epistemic alternatives, the individuals falling under the predicate senator from Michigan are different from the individuals that actually are senators from Michigan. It is in such contexts that one might wish to talk about either the individuals that happen to fall under the predicate in the epistemically accessible worlds—whoever they may be, varying from world-to-world—or, alternatively, about those very individuals that actually happen to fall under the description, regardless of whether they fall under it in other epistemically accessible worlds. For example, in evaluating (b), we are most plausibly talking about properties of individuals that might be elected senators from Michigan, in particular as we try to determine whether there are accessible worlds where at least one of the individuals that happen to be elected senator from Michigan in the respective world also is Republican (in the same respective world). In contrast, in (b), the predicate enjoying one’s retirement cannot plausibly be applied to individuals who are senators in the respective accessible worlds (in ), since being an active senator is incompatible with being in retirement. Therefore, a senator from Michigan is here most plausibly understood to be about the individuals who actually (and currently) happen to be senators from Michigan, and the evaluation of the sentence requires determining whether these very individuals are enjoying their retirement in any of the accessible worlds (in ). Having established the basic distinction in light of the various types of noun phrases that it applies to, we now can turn to considering the inventory of embedding expressions that exhibit the two types of interpretations. These include all modal and temporal operators that operate at the sentential level (we will not consider modal expressions operating at the level of, say, noun phrases, such as the adjectives former or fake). In order to maintain a comparable inventory of examples throughout, we will mostly rely on constructed examples building on the patterns already introduced, but see the references for original examples involving the relevant expressions. Attitude verbs. As noted above, one of the basic cases that was discussed early on in connection with the de re/de dicto distinction is that of attitude verbs (Quine ). The types of contexts where the difference in interpretations becomes relevant with these is quite parallel to those for epistemic and doxastic modals. Take the following variation of our sentences above, uttered the night after an election before John checks the news: () John thinks that most elected senators are Republican. While we could imagine that John is said to have a specific set of individuals in mind for every seat in the senate, the more likely interpretation here is that however any given individual race for open seats in  will go, it will result in an outcome where the majority of individuals elected to the senate are Republican. As before, the sets of individuals that happen to be elected in the worlds compatible with John’s beliefs are not the same throughout.

OUP CORRECTED PROOF – FINAL, 21/1/2019, SPi



    

In contrast, the same type of sentence can also be used to convey that John has a belief about those individuals who actually happen to be senators at the time and world of utterance: ()

John thinks that in , most senators will be retired.

The most plausible interpretation of this sentence is that John has a belief about the actual current senators—the de re interpretation. The same contrast in readings is, of course, available with other attitude verbs (e.g., think, need, want, etc.). Modals. Another type of expression that displays the de re/de dicto ambiguity is that of modal verbs, such as must, can, might, have to, should, etc. To illustrate these, we’ll introduce another type of context that will be useful for some of the phenomena to be considered below as well. Imagine a game show, where the contestant has to answer questions about different trades in competition with professionals in those trades. First, let’s consider a scenario compatible with a de dicto reading below a doxastic modal. Imagine that the order in which the various professionals enter for different rounds is determined by randomly lining them up behind stage. We’re watching that line, but can’t see the first two people. But we think that the first two people are plumbers, namely John and Sue, though actually Sue is a carpenter. In such a situation, we might utter the following sentence: ()

(According to our beliefs,) the contestant has to face a plumber next.

Given a doxastic interpretation of the universal modal have to, this is true on a de dicto interpretation, since in all worlds compatible with what we believe, the contestant faces either John or Sue, who are plumbers in those worlds. But it’s not true on the de re interpretation, since there is no one individual that is actually a plumber whom the contestant will face in all worlds compatible with what we believe. Now let’s consider a scenario compatible with a de re reading above a deontic reading of the same modal (in the same sentence). Imagine a variant where, based on the game rules, the order of professsionals is by alphabetical order of their first names. The next in the alphabet is John. Now using the modal deontically, we could describe this situation as follows: ()

(Based on the rules of the game,) the contestant has to face a plumber next.

The sentence is now true on a de re interpretation of a plumber, since there is an actual plumber—John—that the contestant has to face next based on the rules of the game and given where we are in the alphabet. But the rules of the game say nothing about the occupation of the next professional, so there are worlds compatible with the requirements of the game where the next professional is not a plumber. Thus, the sentence would be false on a de dicto interpretation. Of course, scenarios favoring de re readings with doxastic modals and de dicto readings with deontic modals can also be constructed. (We leave these as an exercise for the reader!). Adverbial quantifiers. Parallel observations can be made for modal adverbs, such as possibly, necessarily, probably, etc. Rather than repeating what would be very much the

OUP CORRECTED PROOF – FINAL, 21/1/2019, SPi

  /  



same type of example, we instead turn back to the temporal domain. In addition to the effects with temporal prepositional phrases that we used above, we can find parallel phenomena with various temporal operators, including adverbial quantifiers such as always and usually, which can be seen as quantifying over points (or intervals) of time. Starting with a de re example, consider the following: () Michelle Obama usually spends her vacation with the president in Hawaii. This is most plausibly interpreted as conveying that Michelle Obama is typically spending her vacations with the actual current president, during, before, and after his presidency, rather than with whoever is president at a given point in time. In contrast, going back to the game show scenario above, we could describe the variant where the professionals go up in alphabetical order as follows: () The contestant always starts out by facing the professional whose name comes first in alphabetical order. Here we are not talking about some specific professional, since these may vary across different instantiations of the game show. Rather, what is being said here is that no matter who the professionals are on a given day, it will always be the case that the one whose name comes first in the alphabet amongst the professionals on that day will be the first one the contestant has to face (a de dicto reading). Tense. Given the effects with temporal adverbials and prepositional phrases, it is not surprising that we can find parallel phenomena relative to tense, that is, temporal information expressed as part of the verb phrase (either via tense morphology on the main verb or through a temporal auxiliary). While there are different well-established analyses of tense in the literature which differ in whether they see tense itself as an operator over times or rather as introducing reference to a contextually salient time, any account sees the overall temporal information expressed by tense as something that other expressions in the clause can interact with scopally. Let’s start with a de re example: () Most senators went to college in their home state. Since senators don’t typically attend college while being members of the senate, the most natural interpretation of this sentence is that for most of the individuals who are senators now there is a time in the past (plausibly prior to their membership in the senate) at which they attended college in their home state. An illustration of a simple de dicto example can be construed most easily when we provide an explicit temporal point of reference, either by a temporal prepositional phrase, as in the initial examples, or by having an explicit context, for example, in the form of a question: () a. What was the political situation in Congress in ? b. Most senators were Republican. As in the initial examples, the most plausible interpretation of the reply is that most individuals who were senators in  were members of the Republican party.

OUP CORRECTED PROOF – FINAL, 21/1/2019, SPi



    

On this interpretation, the sentence is indeed historically true, whereas on a potential de re interpretation, it is false. Some issues and questions Up to this point, we have considered the various examples exhibiting the de re/de dicto distinction, which included a range of different noun phrases and embedding operators, as an entirely uniform phenomenon, and that is indeed the obvious starting point as we strive for theoretical parsimony. But this simple theoretical picture may come under fire if we find substantial differences in the behavior of different expressions involved. Some such differences will indeed come up below. In light of these, the main question will be whether these differences undermine a unified picture, or whether we can maintain the unified picture and attribute any potential differences to independent factors that interact with the core mechanisms giving rise to the different interpretations. As far as the data considered so far is concerned, the traditional scope theory sketched above can handle all the basic cases, and its simplicity makes it useful for illustration of the basic phenomena. However, a number of more complicated examples are by now well-established as showing that a simple traditional theory based on scope is empirically inadequate. We turn to a discussion of the most important phenomena that are beyond its reach next.

. F 

..................................................................................................................................

.. Multiple embeddings and intermediate readings To this point, the examples we have considered involve only one operator relative to which a noun phrase can be interpreted de re or de dicto. If this was all there was, a very simple alternative account of the data would be that the world or time of utterance relative to which the entire sentence is interpreted always remains accessible in embedded environments. Technically, this could be implemented by having a special parameter on the interpretation function that does not get shifted (something similar is necessary for indexicals anyway). This would make it possible to account for at least some of the phenomena considered above without any scope-taking or movement. For example, our sentence () could simply be interpreted as in (b), assuming we have some mechanism in place to ensure that the definite gets evaluated relative to the utterance-world parameter. Here, the time used for evaluation is indicated via underlining: ()

a. In , the president was a senator from Illinois. b. ⟦was a senator from Illinois⟧h2007, 2014i (⟦the President⟧h2007, 2014i) =  iff c. Barack Obama was a senator from Illinois in .

However, once we consider more complex examples with more than one operator, we see that mere access to the world of utterance is not enough. A noun phrase in the most

OUP CORRECTED PROOF – FINAL, 21/1/2019, SPi

  /  



embedded clause can be interpreted de re relative to the lower operator but de dicto relative to the higher one. Take the example in (a), a variant of the game show context considered above. Specifically, assume that the next two professionals in line for facing the contestant are painters. Sam, who’s observing with us from backstage, incorrectly assumes that they are plumbers. Furthermore, Sam does not know which one of the two professionals is first in line, but he is fully informed about the actual rules of the game. We could appropriately describe this situation with the following sentence, as made clear by the paraphrase. () a. Sam thinks that the contestant has to face a plumber in the next round. b. ‘In all worlds w 0 compatible with what Sam believes in w₀ there is some x who is a plumber in w 0 and in all worlds where the rules of the game in w 0 are followed, the contestant has to face x in the next round.’ The noun phrase a plumber here is interpreted relative to Sam’s belief-worlds. Note that there is no particular person of whom Sam thinks that they are next—it could be either one of the two. But he takes both of them to be plumbers, thus he believes a plumber is next. But he is under no false impression about the rules of the game, and he does not relate their status as (presumed) plumbers to the rules at all. So in this case, it will not suffice for the noun phrase in question to either have access to the world quantified over by its immediate embedding operator or the world of utterance. Instead, it is interpreted relative to the higher embedding operator. Thus, the notions of de re/de dicto are relative notions: the noun phrase in question is de re relative to the modal but de dicto relative to the attitude verb in (a). The scope theory can account for this by moving a plumber to a landing site between the two operators. Of course, one could also consider adding yet another parameter slot on the interpretation function. We will return to this possibility in §... For now, the main points are that a) mere access to the world (or time) of utterance does not suffice to capture the full range of phenomena, and b), de re/de dicto are relative notions to be understood with respect to a specific operator.

.. Scope islands, paradoxes, and Fodor’s third reading The second set of issues we would like to turn to concerns the relationship of the noun phrases in question to two components of modal intensional operators.³ On the one hand, these operators affect the intensional status of noun phrases in their semantic scope. But since both they and—at least certain ones of—the noun phrases also involve quantification (over worlds or times and individuals, respectively), they also end up in a scopal relationship. On the traditional scope theory, these two properties are tied together. Consider the structure in (a), showing a noun phrase δ c-commanded by ³ This section is closely based on one from Keshet ().

OUP CORRECTED PROOF – FINAL, 21/1/2019, SPi



    

an intensional operator ω. Under the traditional scope theory of de re and de dicto, this configuration is only compatible with a de dicto reading for δ relative to ω. In order to receive a de re reading relative to ω, δ must move to a position above ω, as schematized in (b). ()

a. De dicto:

b. De re: …

ω …

δ1 δ

… ω

… …

t1

This simple feature of the analysis makes two predictions about a noun phrase δ and an intensional operator ω in such a configuration: ()

a. If δ is trapped below ω (due to a syntactic island or another barrier to movement), δ may not be de re relative to ω. b. The quantificational force of δ will scope above the quantificational force of ω if δ is de re relative to ω and below the quantificational force of ω if δ is de dicto relative to ω.

Counterexamples have been raised for both of these predictions. May: finite clauses. One counterexample to the prediction in (a) is due to May (). To set up this counterexample, May first points out that finite clauses act as islands, blocking quantificational noun phrases inside them from scoping out.⁴ For instance, in (a), the noun phrase every rally in John’s district can scope above some politician, yielding a reading where (potentially) different politicians will speak at each rally. However, this reading—and hence, presumably, this scoping—is unavailable in (b): ()

(= von Fintel and Heim (: )) a. Some politician will address every rally in John’s district. b. Some politician thinks that he will address every rally in John’s district.

Let us now turn our attention to (a), an illustration of May’s counterexample. Its de dicto reading attributes a contradictory belief to Mary, as she would consider the same set of people to be outside and inside. However, this sentence clearly also has the more sensible de re reading which asserts that everyone actually in this room is such that Mary thinks that he or she is outside. According to the scope theory, then, the phrase ⁴ Wilder () later refuted the strongest form of this claim. However, it still seems that the subject of a finite clause cannot scope out of that clause.

OUP CORRECTED PROOF – FINAL, 21/1/2019, SPi

  /  



everyone in this room must move to the position shown in (b) in order to receive its de re reading. Based on the data in (), however, May calls into question whether such a phrase could move to the position it holds in (b).⁵ This contradiction poses a problem for the scope theory. () a. Mary thinks that everyone in this room is outside. b. [everyone in this room]x [Mary thinks that x is outside] Now, one conceivable way to resolve this problem would be to make an exception to allow quantificational noun phrases to scope out of islands under certain circumstances. For instance, perhaps such a noun phrase is allowed to move to become de re, but not allowed to move for (other) scope reasons. As seen next, though, such a relaxation of the rules is not enough to solve the problem. De re noun phrases in if-clauses. Another island for syntactic movement is an if-clause: () Some politician will be happy if everyone votes for him. Similar to the example above, () lacks the reading where the quantifier everyone scopes above the quantifier some politician, that is, where each person x is such that there is a particular politician who will be happy if x votes for him. And yet, despite this restriction on quantifier movement from within if-clauses, such noun phrases may be de re, as seen in the following variation of (a): () If everyone in this room were outside, it would be empty. Considerations parallel to (a) apply: since no one can be in this room and outside in the same world, the noun phrase everyone in this room must be de re relative to the modal governing the conditional in order for () to make sense. Therefore, in order to maintain the scope theory, one would have to add if-clauses to the list of islands that quantifiers may sometimes escape.⁶ Relaxing the rules on when quantifiers may move out of syntactic islands does not solve all of the problems with the scope theory, though. As described in (b), the scope ⁵ This also ties in with facts concerning Antecedent Contained Deletion (Sag ) that involve the subjects of finite clauses: (i) Mary wants to report everyone that Bill does. a. . . . Bill reports. b. . . . Bill wants to report. (ii) Mary thinks that she reported everyone that Jill did/*does. a. . . . Jill reported. b. # . . . Jill thinks she reported. In (i), where the noun phrase with an elided phrase is not inside a finite clause, the ellipsis can refer to the entire clause, as shown. However, in (ii), where the noun phrase is inside a finite clause, the ellipsis can only refer to the inner clause, presumably because the noun phrase may not raise to the top of the sentence. ⁶ In fact, since an if-clause is clearly a finite clause and arguably a complement, this may be another instance of the type of island discovered by May.

OUP CORRECTED PROOF – FINAL, 21/1/2019, SPi



    

theory predicts that the quantificational force of a de re noun phrase will scope in a position above the relevant intensional operator. Consider the following structure for (), though (see Lewis ; Kratzer ): () NP1 CP

everyone in this room would

CP if

it be empty S

t1

VP were outside

The scope theory predicts that the quantificational force of everyone in this room should take scope above that of the modal would. A paraphrase of the meaning of this structure is everyone in this room x is such that if x were outside, this room would be empty. This suggests that the sentence means that the absence of any one person in the room would render the room empty. However, as several researchers have pointed out for parallel examples (von Stechow ; Abusch ; Percus ), the sentence can convey that it is the absence of the totality of the people actually in the room which renders the room empty, not the absence of just one person. Thus, we seem to have a combination of a de re reading with narrow quantificational scope. A similar puzzle can be constructed for sentences involving temporal intensionality, as shown in (): ()

When everyone in this room was outside, it was empty.

In (), the items being quantified over are presumably times, not possible worlds, but the problem remains. () does not mean that for everyone in this room x, when x was outside, the room was empty. These two examples reveal that merely relaxing the rules on movement out of syntactic islands is not enough to solve the problems of the scope theory. Fodor: non-specific de re. Another case where the quantificational force of a de re noun phrase takes scope below the relevant intensional operator was pointed out by Fodor (). Consider yet another variation of our game show context, where the contestant is shown headshots of various professionals and gets to choose who to face next. She’s in the process of deciding and has narrowed it down to two, John and Sue, whom we know to be plumbers (while the contestant is ignorant in this regard). We can describe this new situation with the sentence in (). But this is unexpected on either the de re or the de dicto reading, which both are false in this scenario. Thus the point of Fodor-style examples like this one is that they have more than two readings. ()

Mary wants to face a plumber next.

OUP CORRECTED PROOF – FINAL, 21/1/2019, SPi

  /  



a. Non-specific de dicto: Mary has a preference for what type of professional to face next: she wants it to be a plumber. b. Specific de re: There’s a specific individual, say John, that Mary wants to face next. He actually happens to be a plumber, but Mary may or may not know this. c. Non-specific de re: Mary wants to face either John or Sue next, for instance, although she had not decided which yet. Both John and Sue actually happen to be plumbers, although Mary may or may not know this. (a) and (b) describe what we have been calling the de dicto and de re readings, neither one of which makes the right prediction for the above context. First, consider the non-specific de dicto reading: it is false in that context, since Mary has no idea that John and Sue are plumbers, so their property of being a plumber is not at all reflected in her beliefs. Secondly, the specific de re reading is also false: there is no specific individual that Mary wants to face next—she wants to face either John or Sue. Therefore, Fodor argues that there is another reading, given in (c), and it is this reading that accounts for the context laid out above: in all of Mary’s desire worlds, the person she faces next is a plumber in the actual world (but not necessarily in the respective desire worlds). Fodor’s account is based on the insight that the quantificational force of an indefinite like a plumber can scope separately from its intensional status. As shown above, she calls readings where the quantificational force scopes above the intensional operator specific and those where it scopes below non-specific.⁷ Fodor’s three readings also carry over to the domain of times, as shown in the following example: () Between  and , John always took a buddy (of ) his same weight to the world series. a. Non-specific de dicto: John took a different buddy to each world series and each time he weighed the same as John at that time. b. Specific de re: There is a particular buddy who is now John’s weight that John took to each world series. c. Non-specific de re: John took a different buddy to each world series and each one weighed the same (at that time) as John does now. If you take always to be a universal quantifier over times, () sets up a similar threeway split to (). The specific reading is one where the quantificational force of a buddy (of) his same weight scopes above always, and the non-specific readings are those where this noun phrase scopes below always. The de re readings are those where the weight is

⁷ Fodor actually calls readings where the intensional status scopes above the intensional operator transparent and those where it scopes below opaque. In this chapter, however, we will continue to refer to items as de re when their intensional status scopes above an intensional operator (i.e., they are transparent in Fodor’s terminology). Similarly, we will continue to refer to items as de dicto when their intensional status scopes below an intensional operator.

OUP CORRECTED PROOF – FINAL, 21/1/2019, SPi



    

the same at the speech time, and the de dicto reading is one where it is the same at the time being quantified over (in this case each world series).⁸ In summary, the data reviewed in this section establishes two important facts for theories of de re and de dicto: first, de re and de dicto are relative notions, such that a given noun phrase’s property of being de re or de dicto is always relative to a specific intensional operator. Secondly, quantificational scope and de re/de dicto are not entirely tied together, contrary to what the traditional scope theory predicts.

. R 

.................................................................................................................................. In light of the problems for the traditional scope theory just reviewed, several alternative approaches for capturing the various readings have been considered in the literature. The crucial feature they all share is that they loosen the connection between quantificational scope properties of a noun phrase on the one hand and its intensional status on the other. But they do so through different strategies, and to varying extents. We will review three such approaches here.⁹ The first abandons the idea that the world of interpretation for an utterance is supplied as a parameter on the interpretation function, and instead assumes natural language to have the power of explicit quantification over worlds (and times) in the object language. This removes the need to raise a noun phrase out of the scope of an operator to yield a de re interpretation. The second and third continue to rely on movement to derive de re readings, but they provide more freedom with regards to the consequences of movement for quantificational scope.

.. Intensional variables in the object language Early work in temporal and modal logic (e.g., by Kripke () and Prior ()), as well as in formal semantics for natural language (Montague ), treated times and worlds differently from individual variables.¹⁰ In particular, modal and temporal operators were seen as merely shifting the appropriate evaluation index on the interpretation function, while individuals could be quantified over in the object language. However, as was first argued for tense (Kamp ; Vlach ; Benthem ), and later generalized to worlds (Cresswell ) and situations (Kratzer ), there are examples which show that natural language has the expressive power of quantification over worlds and times (or situations) in the object language:

⁸ One other problem for the scope theory is raised by Bäuerle (), but we refer readers to Keshet () for more information. ⁹ Additional and more recent proposals include Sternefeld (); Romoli and Sudo (). ¹⁰ This section is largely based on parts of §. of Schwarz ().

OUP CORRECTED PROOF – FINAL, 21/1/2019, SPi

  /  



() There will be times such that all persons now alive will be happy at the first or miserable at the second. (Cresswell : ) () If it might have been that everyone actually rich was poor then the economy would have been in bad shape. (Cresswell : ) () If, whenever it snowed, it had snowed much more than it actually did, the town plow would have removed the snow for us. (Kratzer : ex. ()) Roughly speaking, what these examples show is that even in the context of an intensional operator, we are able to make reference to times, worlds, or situations introduced at the level of a higher clause. For example, in (), “we have to be able to consider for each actual snowfall s a set of counterfactual alternatives and compare the amount of snow in each of them to the actual amount of snow in s. This means that we have to be able to ‘go back’ to the actual snowfall situations after considering corresponding counterfactual situations” (Kratzer ). Since the effect can be iterated at will, its analysis requires the expressive power equivalent to that of quantifying over the relevant entities in the object language. Technically, this can be implemented either by representing variables of the right kind in the syntax and allowing intensional operators to bind them quantificationally, or by allowing infinite sequences of evaluation indices (Cresswell ). For purposes of presentation, we will adopt the former perspective. The notion that there are syntactically represented variables for worlds (and times and situations)—commonly referred to as possible world pronouns—has also been put to use to account for de re readings with narrow quantificational scope (Percus ; Schwarz ; Elbourne ). On such an approach, all noun phrases contain an unpronounced, but syntactically represented, possible world (or time) pronoun, which saturates the world (or time) argument of the predicate (Percus ; von Fintel and Heim ).¹¹ Possible world pronouns can be bound by different λ-abstractors associated with intensional operators (as well as the topmost node of the sentence, on the present implementation), which has the desired effect of (partially) disentangling quantificational scope of a noun phrase from the intensional status of its nominal predicate, as can be seen in the sketch of the relevant LFs for () and the corresponding truth conditions, adapted from Percus (): () If everyone in this room were outside, it would be empty. () de re interpretation of everyone in this room: λw0 ½S if ½S λw1 ½S everyone in this room in w were outside in w  [it would be empty in w₁]]] (adapted from Percus ) ¹¹ An important question that we’ll abstract away from for the moment is where in the noun phrase these pronouns appear: as arguments of the noun or of the determiner? Schwarz (), building on Büring (), argues for the latter. See §... for brief discussion.

OUP CORRECTED PROOF – FINAL, 21/1/2019, SPi

 ()

     For any world w, (0 ) is true in w iff for every accessible world w0 such that everyone in this room in w is outside in w0 , the room is empty in w0 .

Given the indexing on the pronoun associated with the noun phrase, everyone in this room will be interpreted relative to the world of evaluation for the entire sentence, which renders a sensible reading of the sentence. Alternatively, the possible world pronoun in the noun phrase could also be indexed as w₁, which yields a de dicto reading (though this renders the antecedent contradictory in the present example). The same strategy also deals with the other problematic cases. (a) gets the desired reading (which does not attribute a contradictory belief to Mary) without having to scope the quantifier out of the finite clause: (a)

Mary thinks that everyone in this room is outside.

() λw [Mary thinks in w [λw every one in this room in w is outside in w]] Similarly, Fodor-style examples such as () with a quantificational narrow scope de re reading (Fodor’s ‘non-specific de re’) are also captured: ()

Mary wants to face a plumber next.

()

λw₀ [Mary wants in w₀ λw₁[to face (in w₁) a plumber in w₀ next]]

Thus, the option of binding a possible world pronoun in a noun phrase that relates it to an operator higher up in the structure provides the necessary disentangling of quantificational scope and intensional status to deal with the data that is problematic for a traditional scope theory.

.. A scope theory with higher type traces The next two approaches to be considered are variants of the scope theory, which turns out to be able to provide some leeway for distentangling scope and intensional status as well. The first, based on discussion by von Fintel and Heim (), modifies the assumptions about the semantic type of traces. While these are most commonly taken to be of type e, it is also possible to consider traces of other types (see also discussion in Heim and Kratzer : –). In particular, we can consider leaving traces of type het, ti. When applying this to the Fodor-example in (), as in the LF below, the result turns out to be a quantificational narrow scope de re reading.¹²

¹² The additional step of quantifier raising within the embedded clause is needed to deal with the quantifier in object position; alternatively, one could assume higher-typed entries for the verb.

OUP CORRECTED PROOF – FINAL, 21/1/2019, SPi

  /  



() VPt NP1⟨et,t⟩

VP1⟨⟨et,t⟩,t⟩

a plumber λ1,⟨et,t⟩

VPt NPe

VPet

Mary V⟨st,et⟩

VPst

wants λw1

VPt NP⟨et,t⟩ t1

VPet λ2,e

VPt NP

VP⟨e,t⟩

pro V⟨e,et⟩ NPe to face

t2

To see how this comes about, consider the composition of the crucial pieces at the topmost level: () ⟦ VP1 ⟧ w ¼ λQhet;ti : ⟦ wants ⟧ w ðλw1 :Qðλxe : ⟦ face ⟧ w1 ðxÞð ⟦  ⟧ w1 ÞÞÞð ⟦ Mary ⟧ w Þ ()

⟦ NP1 ⟧ w ¼λPet :∃y½y is a plumber in w and PðyÞ ¼ 1

()

⟦VP1 ⟧ w ð⟦NP1 ⟧ w Þ ¼ ⟦wants⟧ w ðλw1 :½λPet :∃y½y is a plumber in w & PðyÞ ¼ 1 ðλxe : ⟦ face ⟧w1 ðxÞð ⟦  ⟧w1 ÞÞÞð ⟦ Mary ⟧w Þ

As can be seen in (), even though the quantificational force of a plumber winds up in the scope of want, the nominal predicate nonetheless gets interpreted relative to the world of evaluation for the entire sentence. Thus, if we permit traces to have type het, ti, we can account for Fodor-style sentences within the traditional scope theory. However, the other two problems above still remain: finite clauses and if-clauses will still require the noun phrase to be moved out of the relevant clause to get the reading in question. While it is in principle possible to relax constraints on movement for special cases (e.g., to derive de re readings), it is unclear whether that can be motivated on more general grounds. Be this as it may, the current extension shows that the scope theory is open to ammendments that allow a partial disentanglement of scope and intensional status. The following section explores another proposal for doing so, which does not require relaxing constraints on movement for de re readings.

OUP CORRECTED PROOF – FINAL, 21/1/2019, SPi



    

.. Split intensionality: a modern scope theory Keshet (a, ) attacks the problem of quantificational narrow-scope de re readings by taking a closer look at the interpretation of intensionality in general.¹³ If we start out from an extensional system and try to extend it minimally to incorporate intensions where needed, one straightforward approach is to use a rule such as Intensional Functional Application (IFA: Heim and Kratzer ), which is basically a type repair strategy for intensionality: when a function (e.g., that introduced by a modal or attitude verb) requires an intensional argument but the syntax only supplies an extensional one, IFA shifts the type of the argument from an extension to an intension and then applies Functional Application as per usual. The main innovation of the system described in Keshet (a, ) is that it introduces the type-shift in a separate step, namely through the insertion of the ∧ operator and a new rule: ()

Intensional Abstraction If α is a branching node and {β, γ} is the set of its daughters, where β dominates only an ∧ operator, then, for any situation s and variable assignment g, ⟦α⟧s, g = λs0 0 ∈Ds. ⟦γ⟧s , g. (Modeled after Heim and Kratzer’s Predicate Abstraction : )

The operator ∧ may be inserted freely—if it yields a type mismatch, the derivation will simply fail. There is no longer a repair strategy when a function requires an intension but is supplied with an extension. Instead, the idea is that a derivation will only succeed if ∧ has already been inserted by the time the function takes its argument. Giving the type-shifting operation a separate place in the structure, perhaps even several nodes away from the intensional operator that required it in the first place, makes it possible to raise a noun phrase to a position between the two. This has the effect that the intensional status of the noun phrase is not affected by the embedding, but its quantificational scope still is below the quantifier over worlds introduced by the operator. This deals with all three problematic cases, as illustrated below.¹⁴ Finite clauses The first problem for a traditional scope theory along the lines sketched in §.., due to May (), was that quantificational noun phrases could never scope out of finite clauses as required for de re readings of sentences like (a) above. Based on the new rule above, a de dicto reading of everyone in this room is derived when everyone in this room is below the ∧ and a de re reading is derived when this noun phrase is above the ∧. But being above ∧ no longer requires outscoping the relevant embedding expression, as seen in (). This option immediately solves May’s objection to the scope theory, since the syntactic movement involved is no longer illegal; as illustrated in (), the de re noun phrase scopes above the new operator ∧, but still within the finite complement: ¹³ This section is largely a condensed version of the key sections in Keshet (). ¹⁴ Wide-scope (specific) de re readings will continue to require movement out of the scope of the relevant embedding expression.

OUP CORRECTED PROOF – FINAL, 21/1/2019, SPi

  /  



(a) Mary thinks that everyone in this room is outside. ()

Narrow Scope de re for everyone in this room: VPt NPe

VPet

Mary V⟨st,et⟩ thinks

VPst NP⟨et,t⟩

VPest

everyone λ in this room 1

VPst ^

VPt t1

VPet is outside

A complete derivation for a sentence like () is shown in Keshet (), but here is a sketch of how it proceeds to illustrate the effect of ∧: first, the ∧ applies to the VP, of type t, to form a node of type st. The subject everyone in this room moves above this node, first abstracting over a type-e argument to form a node of type est. This node and the quantifier of type het, ti can compose using the semantic function Combine independently proposed in Büring (),¹⁵ resulting in another node of type st. Last, the verb thinks takes this type-st node as its argument. If-statements The fact that under Split Intensionality a noun phrase de re relative to an intensional operator ω can remain below ω explains why its quantificational force can scope below that of ω. Recall our counterexample () to the scope theory from §.., repeated here: () If everyone in this room were outside, it would be empty. The original scope theory incorrectly predicts () to mean that for each person x, if x were outside, the room would be empty. Split intensionality, on the other hand, captures this case correctly. As shown in (), inside the if-clause, the noun phrase everyone in this room has raised to a position above the ∧. Only the items below ∧ (those in the box shown in ()) are interpreted in the supposition worlds of the conditional. Since everyone in this room has moved out of this box, it is evaluated in the actual world, although it still scopes below the modal in terms of quantificational force. ¹⁵ Büring () proposes a syntactic operator κ that allows the combination of a quantifier with a node representing any function requiring a type-e argument. In short, the quantifier fills this open argument slot and the remaining arguments are passed up the tree.

OUP CORRECTED PROOF – FINAL, 21/1/2019, SPi



     VPt

() VP⟨st,t⟩

TPst

V⟨st,⟨st,t⟩⟩ would

^ this room be empty

CPst C

TPst NP⟨et,t⟩

if

TPest

everyone in this room

λ1

TPst ^

TPst t1

TPt were outside

Fodor’s paradox The split intensionality theory also captures the data discovered by Fodor: ()

Mary wants to face a plumber.

As discussed above, () has a reading where the noun phrase a plumber is de re, in the sense that Mary does not know that the professionals she wants to face next are plumbers, but the noun phrase still takes scope below the verb want in the sense that there is no one single professional that Mary wants to face next. This reading is not a problem in the split intensionality system, where a noun phrase may take quantificational force below an intensional verb ω and yet still be interpreted de re relative to ω. For instance, consider the structure in () and the corresponding truth-condition paraphrase in (): VPt

() NPe

VPet

Mary V⟨st,et⟩

TPst

wants T to

VPst NP⟨et,t⟩

VPest

a plumber λ1

VPst ^

VPt NPe

VPet

Pro Veet t1 face

OUP CORRECTED PROOF – FINAL, 21/1/2019, SPi

  /  



() In all of Mary’s desire worlds w0 , there’s an x such that x is a plumber in the real world w₀ and Mary faces x in w0 . So, the split intensionality system predicts that there should be a reading of this sentence where, as Fodor describes, there is no one particular plumber in the real world which Mary wants to face next and yet the description plumber holds (only) in the real world.

.. Differentiating the theories—empirical and conceptual considerations The two most promising candidates for refined theories, based on intensional variables in the object language (§..) and the notion of Split Intensionality (§..), are equally successful in terms of the basic empirical ground they cover, in that they account for the three types of readings discussed in §.. This raises the question of whether there are any differences between them that might inform a decision as to which version to adopt. Two main issues are relevant in this regard: first, de re readings are more constrained than a basic version of a theory that assumes possible world pronouns in the object language would predict. Secondly, the range of readings predicted by the split intensionality theory is more restricted than that of possible world pronoun theories when it comes to complex embedding environments.

... Constraints on de re readings Generalization X (Percus ) While the expressive power gained by representing intensional variables in the syntactic structure allows us to capture de re interpretations, the standard implementation of this turns out to introduce a problem of overgeneration, as was first discussed in detail by Percus ().¹⁶ In particular, if we assume that all predicates—including verbs—come with a syntactically represented situation pronoun,¹⁷ we expect—barring further assumptions—de re (or ‘transparent’, in Fodor and Percus’s terminology) interpretations to be available for all predicates. This expectation is not borne out, however, as Percus shows in great detail. A case in point are the situation pronouns introduced with verbal predicates. Percus provides the example in (a) and considers an LF with the co-indexing possibility in (b).

¹⁶ This subsection is largely based on §. of Schwarz (). ¹⁷ Since the relevant literature frames these issues in terms of situation pronouns, we will adopt that terminology here as well. Situations are generally taken to be parts of worlds by the relevant authors, following (Kratzer ).

OUP CORRECTED PROOF – FINAL, 21/1/2019, SPi

 ()

     a. Mary thinks my brother is Canadian. … b. S′

λs7 NP my brother s7

VP is Canadian

s0

The noun phrase my brother here is de dicto, as its situation pronoun is co-indexed with the λ-abstractor for the embedded clause. In contrast, the situation pronoun in the verb phrase is Canadian is co-indexed with the topmost λ-abstractor. If this indexing of the pronouns would indeed represent an available LF-configuration, “we would take the sentence to be true whenever there is some actual Canadian who Mary thinks is my brother—even when this person is not my brother in actuality, and even when Mary mistakenly thinks that he is not Canadian” (Percus : ). However, in such a situation we clearly judge the sentence to be false, which shows that the indexing in the LF in (b) is not available. More generally, verbal predicates have to be evaluated relative to the same situation as the clause they appear in.¹⁸ Percus concludes that there is a general constraint on the interpretation of situation pronouns introduced with verbal predicates, which he labels ‘Generalization X’:¹⁹ ()

Generalization X: The situation pronoun that a verb selects for must be coindexed with the nearest λ above it. (Percus : )

Generalization Z and the intersective predicate generalization Building on Percus’s insights, Keshet (a, ) discusses a further constraint on the interpretation of situation pronouns, which concerns the distinction between weak and strong noun phrases. As is standard, weak noun phrases are understood to be precisely those that can appear in existential there constructions, following Milsark (). The starting point for this line of thought comes from Musan (), who showed that not all noun phrases display temporal independence (contra Enç ):

¹⁸ The noun phrase can of course be interpreted as de dicto (Mary doesn’t know who my brother is, but she assumes whoever he is to be Canadian) or de re (Mary thinks that John is Canadian, and unbeknownst to her, John actually happens to be my brother), as usual. ¹⁹ Percus also makes a parallel point for adverbs, based on parallel data, which won’t play a central role in our discussion: (i)

Generalization Y: The situation pronoun that an adverbial quantifier selects for must be coindexed with the nearest λ above it. (Percus : )

OUP CORRECTED PROOF – FINAL, 21/1/2019, SPi

  /   () a. Every fugitive is in jail (again). b. #There is a fugitive in jail (again).



(Musan ; Kusumoto )

() Some members of congress knew each other in college. In fact, . . . a. . . . three U.S. Senators were attending Harvard together in . b. # . . . there were three U.S. Senators attending Harvard together in . (Keshet a, adapted from Musan ) The contrast observed in both of these pairs of examples is that while the (a)-sentences have a perfectly reasonable interpretation, which comes about by interpreting the nominal predicate at a time different from that of the verbal predicate in its clause, the existential there variants in (b) have no sensible interpretation. (b) is contradictory, and the continuation in (b) only has the implausible interpretation that the relevant individuals were senators while attending Harvard in . Keshet (a, ) furthermore showed that this effect, too, is paralleled in the domain of possible worlds (or situations): () a. Mary thinks that someone in this room is outside. b. #Mary thinks there’s someone in this room outside. () a. Mary thinks three professors are (still) in college. b. #Mary thinks {there’s/there are} three professors still in college. (both examples from Keshet a: ) Both (b) and (b) are odd in that they can only be understood as attributing inconsistent (or implausible) beliefs to Mary, unlike their counterparts in the (a)-sentences. This shows that the predicates of weak noun phrases have to be interpreted relative to the same situation as the verbal predicate in their clause, that is, in Mary’s ‘thought-worlds’ in the present sentences. Keshet proposes to add a further generalization based on these findings: () Generalization Z: The situation pronoun selected for by a noun in a weak NP must be coindexed with the nearest λ above it. (Keshet a: ) Following Milsark (), the interpretation of the existential there-construction can be seen as involving intersection of the two predicates and existential closure over the resulting property (see Keshet , for details and a modern implementation of Milsark’s idea). Adopting the proposal by Landman () that weak noun phrases in general denote properties, Keshet () argues that Generalization Z is a special case of a more general constraint that requires any two predicates that are interpreted intersectively to be evaluated relative to the same world and time (or situation):²⁰

²⁰ Note that relative clauses constitute an important exception to this generalization (as Keshet also points out).

OUP CORRECTED PROOF – FINAL, 21/1/2019, SPi



    

() Intersective Predicate Generalization (IPG): Two predicates interpreted intersectively may not be evaluated at different times or worlds from one another. (Keshet ) Keshet (a, ) presents evidence for this with examples involving nouns and their modifiers, the have-construction, and depictives. Take the following examples of the first case as a brief illustration: () a. #In , every U.S. Senator at Harvard got straight A’s. b. #Mary thinks the married bachelor is confused.

(Keshet )

In (a), the noun U.S. Senator and the prepositional phrase at Harvard are interpreted intersectively, and the sentence only has a reading where the relevant individuals were senators and at Harvard at the same time. Similarly, the adjectival modifier married and the noun bachelor are interpreted intersectively, and (b) can only be interpreted as attributing inconsistent beliefs to Mary. The upshot of this discussion is that accounts that posit intensional variables in the object language face a problem of overgeneralization. Assuming that all predicates that have a semantic situation argument (can) take a situation pronoun as a complement predicts unattested readings, as—descriptively speaking—the situation pronouns of verbs and weak noun phrases need to be bound by the closest λ-abstractor (Generalizations X and Z: Percus ; Keshet ). Furthermore, intersectively interpreted predicates have to be interpreted relative to the same situation (IPG: Keshet a, ). But note that these problems only arise if we do indeed assume that all predicates (can) combine with a situation pronoun. One possible variation of this type of account, proposed by Keshet (c) and Schwarz (), is that situation pronouns have a more restricted distribution and only are introduced as arguments of strong determiners. This accounts for the generalizations considered above. Furthermore, Schwarz (, ) argues that situation pronouns introduced by strong determiners also are behind standard effects of quantificational domain restriction (von Fintel ). It thus seems possible to amend this type of account in an independently motivated way so as to avoid the problems of overgeneralization reviewed above. Turning to the split intensionality theory, the relevant problems of overgeneration do not arise. Given that it ties all de re readings to movement, only expressions that can move will have such readings. This straightforwardly accounts for Generalization X (and Y), since verbs (and adverbs) do not undergo movement. It also accounts for Generalization Z (and the IPG) if we assume that weak quantifiers denote properties (Landman ), and that no movement is possible out of noun phrases.

... Complex embeddings Another difference between the split intensionality theory and one based on intensional variables in the object language concerns the predictions for complex embedding environments with multiple intensional operators, at least to the extent to which we maintain the assumption that covert movement is constrained in the same (or very

OUP CORRECTED PROOF – FINAL, 21/1/2019, SPi

  /  



similar) ways as overt movement. On the split intensionality theory, a noun phrase can raise to the edge of a syntactic island but not further. This predicts that in cases where the island is itself embedded, the only de re reading available to it is one that is relative to that embedding operator. Schematically, using an if-clause island: …

()

…

λs0

…

believes

…

λs1

…

if

…

NP λ1

… λs2 … t1 …

The nominal predicate in the noun phrase here can thus be interpreted relative to the lower-clause if-situations (s₂)—if the noun phrase doesn’t raise—or the think-situations (s₁)—if it does raise—but not the situations relative to which the entire sentence is evaluated (s₀), as far as the split intensionality theory goes. In contrast, on accounts using syntactically represented situation pronouns, it can be interpreted relative to any situation variable corresponding to a c-commanding λ-binder. Keshet (a, ) claims this as a feature of the split intensionality theory, citing several cases where a de re reading is disallowed from within a syntactic island. For instance, consider the sentence in (a), which receives a plausible interpretation only when the noun phrase every paper John wrote is interpreted de re relative to the verb thinks. For instance, the sentence is true in a scenario where John turns in a series of papers but the teacher suspects that Sally had written them for John. A de dicto reading is in principle possible as far as the grammar is concerned, but it implausibly suggests that Sally and John both are the sole author of the paper. Turning to (b), the implausible interpretation is the only one available, and the sentence as a whole is perceived as odd: () a. The teacher thinks that Sally wrote every paper John wrote. b. #The teacher thinks John should be punished because Sally wrote every paper John/he wrote. Note that the same sentence except ending with . . . every paper John turned in sounds fine, which shows that the deviance is due to the specific nature of the predicate. Now, if every paper John wrote could be interpreted de re relative to the verb think in (b) as it is in (a), the sentence would have a perfectly consistent interpertation, and thus

OUP CORRECTED PROOF – FINAL, 21/1/2019, SPi



    

should not be perceived as odd. The Split Intensionality account explains the unavailability of the de re reading by appealing to the fact that the because-clause creates an island for movement, which makes it impossible for the relevant DP to escape the intensional abstraction introduced by think. In addition to these cases, though, there are also some cases that seem to contradict the prediction that islands restrict de re readings of noun phrases. Consider the following sentence in the described scenario, based on one in Bäuerle (): () a. Scenario: George meets a woman getting off a bus who has just hugged everyone on the bus. Although she is not from the South, George is bad at accents and forms the opinion that she is a Southerner. Furthermore, although George does not know this, the bus she got off is the team bus containing every Red Sox player. The woman leaves before giving George her contact information, so he is looking for her. b. George is looking for a Southerner who hugged every Red Sox player. This is a true and felicitous method of describing the scenario when the speaker knows the bus is the Red Sox team bus (even if George does not) and knows that George believes the woman is Southern (even if she is not). Considering (a) that the noun phrase a Southerner who hugged every Red Sox player must be de dicto in this scenario, and (b) that relative clauses like who hugged every Red Sox player are islands for movement, the split intensionality theory predicts that the embedded noun phrase every Red Sox player must also be de dicto. However, as seen in this scenario, it can be de re, contrary to this prediction. A possible reply for the split intensionality theory would be to claim special status for the overall world (or situation) of evaluation. This could be seen as part of the overall information about the utterance context, parallel to the way other information about the utterance context, such as who the speaker is, is generally assumed to be available via parameters on the interpretation function. (Note that such a move would also necessitate an independent explanation for (b).) But this move will not generalize to even more complex embeddings, since there we can consider cases where a noun phrase is evaluated relative to a non-global situation that is still out of reach, as it were, for scope-based accounts (again, assuming standard constraints on movement). Empirically, this gets harder to evaluate, but the following is an attempt at construing such a case: ()

a. Scenario: Same as above, except Al has observed the events in question and understood them correctly except that he has formed the (incorrect) conclusion that the bus is for the White Sox, not the Red Sox. b. Al thinks George is looking for a Southerner who hugged every White Sox player.

Once again, as long as the speaker knows the relevant details, this is a true and felicitous description of the scenario. Analogously to the sentence above, the lowest noun phrase (every White Sox player) is predicted by the split intensionality theory to be de dicto

OUP CORRECTED PROOF – FINAL, 21/1/2019, SPi

  /  



relative to the intensional context established by is looking for. However, in this scenario, this noun phrase is in fact de re relative to looking for; but it is (crucially) de dicto relative to the higher attitude verb thinks. If the sentence can serve as an adequate description of the described scenario, then that is a case where the split intensionality theory undergenerates, unless it drastically relaxes the relevant constraints on movement. Situation-pronoun accounts, on the other hand, face a problem of overgeneration if the sentence cannot adequately describe the provided scenario. Settling the empirical question will require more attention than we have space for here. We’ll content ourselves with having sharpened the difference between the two types of accounts for now. Relying on movement for de re readings while maintaining standard assumptions about constraints on movement commits the modern scope theory to various other predictions. Indeed, Keshet () discusses various island-related phenomena that arguably support these predictions (e.g., (b) above). At the same time, however, there are other examples that seem more at odds with these predictions. For example, the expressive power arguments by Cresswell () and Kratzer () would seem to involve dependencies between a noun phrase’s world (or situation) variable and embedding operators or binders that should be out of reach from the scope theory’s perspective. As a variation of an example from Cresswell () (who in turn attributes it to Angelika Kratzer (p.c.); see also Kratzer ; Elbourne  for further discussion of related examples), consider the following: () If Obama had written some novels and Fox news alleged that a ghost writer confessed that he had written every novel Obama had written, Obama would sue. Crucially, it seems possible here to refer back to the novels introduced in the if-clause from inside the most embedded clause—the noun phrase every novel Obama had written cannot plausibly be interpreted relative to the confession or the allegation. And while Keshet () argues that conjunctions do exhibit island effects and thus do not allow for de re readings, the following variant of the above example (again adapted from Elbourne , who in turn built his on Kratzer  and Cresswell ) does seem to permit a de re reading for a noun phrase appearing inside of a conjunction: () If Obama had written some novels and composed some songs and Fox news alleged that a ghostwriter confessed that he had written every novel Obama had written and composed every song Obama composed, Obama would sue. Finally, yet another difference between scope-based and pronoun-based accounts related to examples like these concerns a possible independent use of situation pronouns, particularly for domain restriction. Most recently, Schwarz (, ) as well as Elbourne () argue that situation pronouns provide all we need to account for nominal domain restriction (also see Recanati , ; Kratzer , , among others), which furthermore provides an account of donkey sentences with definites.

OUP CORRECTED PROOF – FINAL, 21/1/2019, SPi



    

While a more detailed review of these phenomena would lead us too far astray, any independent motivation for the presence of situation pronouns in the syntax would of course bolster a pronoun-based account (though if there are remaining problems of overgeneration, these of course will still have to be thrown into the mix for an overall evaluation).²¹

. F 

..................................................................................................................................

.. More on the relation between scope and intensional status One of the main lines of argument in the literature, as well as in this chapter, has been that the connection between quantificational scope and intensional status posited in the traditional scope theory is too strong, in that it predicts a perfect correlation between the two. Both of the refined theories considered above loosen the connection, but only to a certain extent. Essentially, they transform a biconditional dependence into a conditional one: de re no longer implies wide quantificational scope, but wide quantificational scope still implies de re. In other words, wide-scope de dicto readings still are predicted to not exist by these theories. While this is assumed to be empirically adequate in much of the literature (e.g., Percus ; von Fintel and Heim ), Fodor () claimed that a fourth reading (non-specific opaque in her terms) of sentences like () exists. ()

Mary wants to buy an inexpensive coat.

Under such a reading, as Fodor puts it, there is a particular coat that Mary wants to buy and that she wants to buy under the description an inexpensive coat (see Fodor : ). In this case, it is not necessarily true that the coat in question is actually inexpensive. This seems like a reasonable idea to express, and in fact this is what (a) means. However, it does not seem like () can really mean that, if we consider (b). The use of the word it in the second sentence of (b) forces a specific reading of an inexpensive coat (see Ioup ). However, once this reading is forced, it is impossible to deny that the coat is inexpensive.

²¹ One more case where Keshet () admits that the split intensionality theory breaks down involves definite descriptions, which seem to be able to refer to salient times, even those described in separate sentences. For instance, in the last sentence of (i), the term that five-year-old girl refers to someone who is clearly no longer five years old. And yet, no tense in the same sentence refers to a time when the woman was five years old—this description crucially relies on a time described in a previous sentence. (i)

In , I visited my friend Joanne. I met her two-year-old son and her daughter, who was three years older. Last year, I watched that five-year-old girl graduate from medical school.

OUP CORRECTED PROOF – FINAL, 21/1/2019, SPi

  /  



() a. There’s a coat that Mary wants to buy. She thinks it is inexpensive. But really, it is quite expensive. b. Mary wants to buy an inexpensive coat. #But really, it is quite expensive. More recently, Szabo () has provided another set of cases that he argues to support the notion of a fourth reading. First, he suggests the contrast between (a) and (b) is due to the presence versus absence of think, and points to the fact that () seems to work just fine: () Mary thinks she bought an inexpensive coat. It is actually quite expensive. The felicity of this example leads Szabo to claim that “If there is no such thing as a specific opaque [wide-scope de re, in our present terms] reading, the contrast is a bit of a mystery” (Szabo : ). However, one possible solution to this mystery would be to see it as a case of accommodation—unless Mary is known to be delusional, her belief about having bought something that she thinks is an inexpensive coat quite likely will lead a hearer to conclude that there indeed is a coat that she bought, which might suffice to license pronominal reference. Thus, the availability of it would not be due to a wide-scope LF, but due to pragmatic reasoning. While more might need to be said about this, let us briefly mention another type of example, the main one in Szabo’s paper, which he labels as summative reports. It is based on a scenario where police show a paranoid person, Alex, a large number of photographs of people from his neighborhood, his task being to identify the ones that he thinks are terrorists, and to specify where they live. Alex is not keeping count of his allegations, but one of the police officers later tallies up Alex’s input in the following way: () Alex believes that eleven terrorists live across the street from him. This, Szabo claims, should be seen as a wide-scope de dicto case, as Alex has no belief about the number of terrorists he identified to be living across the street. Parallel examples with various other quantifiers, including the strong most, are also provided by Szabo. If these are indeed instances of wide-scope de dicto readings, this has serious theoretical consequences, since the main theories reviewed here exclude the possibility of such a fourth reading. Szabo lays out a proposal in terms of split quantifiers, where the quantifier head and its nominal complement can wind up scoping separately. While we cannot go into further details here, we’d like to note that there may be reason to be skeptical about the conclusion that these cases—felicitous as they are—indeed instantiate genuine wide-scope de dicto readings. In particular, the summative examples seem to turn on the imperfections of our belief states. If there are eleven individuals of whom Alex believes that they are terrorists that live across the street, then the embedded clause of () logically follows from what he believes, regardless of whether he’s aware of the count. And if Alex is at all rational, despite his paranoia, he would have to agree with () if he reflected on his individual beliefs. So there’s a sense in which the sentence in question conveys something along the lines of ‘it logically follows from what Alex believes that there are eleven terrorists living across the street from him.’ But

OUP CORRECTED PROOF – FINAL, 21/1/2019, SPi



    

then the exceptional nature of these summative reports with respect to scope and the de re/de dicto distinction may be due to the nature of beliefs, and not the existence of widescope de dicto LFs. More cases, both involving logical properties like counting and other ones, should be looked at. Note also that even if we follow Szabo’s line of argument, the absence of a wide-scope de dicto reading for () would still need to be explained.

.. de re/de dicto beyond noun phrases Our focus here has been on the scope and intensional status of noun phrases, following the bulk of the literature.²² However, there are other types of expressions that exhibit similar—and perhaps exactly the same—phenomena. The following illustrates with some examples, without going into any serious level of analysis. Comparatives One classical case, noted early on by Russell (), involves a comparative and a than-clause: ()

Bill thinks that my yacht is longer than it is.

For the sensible interpretation of this sentence on which Bill is not holding incoherent beliefs, the predicate in the than-clause has to be interpreted relative to the actual world, whereas the predicate in the main clause has to be interpreted relative to the thought-worlds. Based on this and similar examples with counterfactuals, von Stechow () has argued that to account for the full range of data, we have to allow for de re interpretations of predicates in the than-clause (to allow for an interpretation of the predicate relative to the actual world even though it remains in the scope of the intensional operator at LF). Relative clauses Another case, somewhat intertwined with noun phrases, where we observe similar phenomena is that of relative clauses. In the temporal domain, it has been argued that the tense in a relative clause does not have to be interpreted relative to the tense in the matrix clause, even when the noun phrase containing it has to take scope below the matrix verb (which rules out a scope account, as shown by Kusumoto ). The following example is an illustration of this so-called later-than matrix interpretation:²³ () Hillary married a man that became president.

Kusumoto ()

The fact is, of course, that Bill only became president after he married Hillary, which means that the past tense on became has to be interpreted relative to the time of utterance, rather than relative to the time introduced by the past tense on married. Thus, it is possible for the relative clause to be evaluated relative to the utterance ²² This section has been adapted from §. of Schwarz (). ²³ This example alone doesn’t rule out a scopal account. Kusumoto () argues that certain variants do, but see Keshet (b) for a rebuttal.

OUP CORRECTED PROOF – FINAL, 21/1/2019, SPi

  /  



time (or situation), even though it occurs in the scope of a temporal operator (the matrix tense). One set of data that suggests that relative clauses have special properties with respect to their intensional status as well is the following (Schwarz , building on observations by Breheny ). Note that prenominal modifiers and relative clauses that—on their simplest analysis—should be equivalent to them differ in terms of their behavior in intensional contexts such as the one created by fake: () a. A fake American philosopher was at the conference. b. A fake philosopher who is American was at the conference. Only (a) is compatible with a scenario in which the person in question is a real philosopher that pretends to be American. Once again, this suggests that the intensional status of an expression in an embedded position can be independent of its immediate embedding operator. De re verbs? As discussed in §..., Percus () argued that de re readings are unavailable for verbal predicates (Generalization X). More recently, though, Cable () has argued that at least in certain cases, it seems as if verbs indeed can be interpreted de re. He imagines a situation where we are regularly practicing a juggling routine. Our friend Mary only knows that we are practicing some kind of act at the given time. One day, she thinks we’re at practice, although we actually decided to skip it last minute. He then observes that this can be characterized as follows: () Mary thinks we’re juggling right now. But this requires a de re interpretation of juggling. Interestingly, however, it need not necessarily contradict Generalization X. In short, Cable points out that the relevant reading can be paraphrased as follows: () Mary thinks we’re doing (our daily) juggling right now. He then goes on to argue that the availability of a de re reading for () in fact provides evidence for a decompositional analysis of verbs like juggle, as do + NP. Thus, what looked like a case of verbal de re may boil down to the usual de re noun phrase, at least underlyingly.

.. Broader relevance Beyond its obvious importance in semantic theory, the de re/de dicto distinction can have significant practical consequences, for example, in the interpretation of law. Anderson () details several important misunderstandings in English and American case law involving intensional contexts. One particularly egregious example is the case of Whiteley v Chappell, brought against a defendant who pretended to be a

OUP CORRECTED PROOF – FINAL, 21/1/2019, SPi



    

deceased neighbor of his in order to vote.²⁴ The relevant voter fraud statute at the time prohibited “[im]personat[ing] any person entitled to vote.”²⁵ The court found the defendant not guilty because the person he impersonated was not a “person entitled to vote”, being dead. Anderson details how this precedent became famous as an example of taking a statute (allegedly) quite literally, instead of invoking what was clearly the spirit of the legislation. As Anderson points out, though, the dichotomy is not actually between the literal meaning and the spirit of the law, but rather between two literal meanings: one de re and the other de dicto. The reading the court based its decision on (and the one later labelled the ‘literal’ meaning) was de re: the statute would only be violated if there was an actual person entitled to vote such that the defendant impersonated that person. However, the equally ‘literal’ de dicto reading would prohibit anyone from pretending that they were entitled to vote, even without necessarily impersonating any particular person so entitled. This meaning also happens to accord more closely to the apparent spirit of the statute. As a much more recent example, Anderson () recounts the case against the accounting firm Arthur Anderson, which served as auditor to the bankrupt energy company Enron while they engaged in some (fraudulently) creative bookkeeping.²⁶ Anticipating the potential—but still future—legal action against them, Arthur Anderson destroyed incriminating evidence up until the date when they received a subpoena. The relevant statute, paraphrased by Anderson (), prohibited anyone from “corruptly endeavor[ing] to influence the due administration of justice” (: ). Arthur Anderson was found not guilty because, before the investigation began and a subpoena was issued, there was no existing instance of “administration of justice” for the firm to endeavor to influence. This is obviously the de re interpretation of the statute, requiring a particular investigation to exist in order for the law to hold. Of course, though, there is a rather sensible de dicto reading, too, in which any attempt to obstruct justice—even in the absence of a particular existing investigation—would be covered by the law.

²⁴ () L.R.  Q.B. . ²⁵ Id., . Note that here we seem to be dealing with a transitive verb creating an intensional context for its object NP, without the overt presence of a clause, similar to Quine’s (). Regardless of the particular analysis of such cases, we are looking at a version of the de re/de dicto. ²⁶ Arthur Andersen LLP v United States,  U.S. , – ().

OUP CORRECTED PROOF – FINAL, 17/1/2019, SPi

        ......................................................................................................................

NEGATIVE EXISTENTIALS ......................................................................................................................

 ,  ,   

. T    

.................................................................................................................................. A singular negative existential is a sentence whose subject is a singular term and whose predicate is ‘does not exist’.¹ A paradigmatic example is: ()

Pegasus does not exist.

Such sentences have been discussed by philosophers going back at least as far as Plato’s The Sophist in  BC. Not surprisingly then, it is difficult to specify a unique “problem of negative existentials”; rather such sentences raise a family of related linguistic, logical, and metaphysical issues.² Here we will be concerned with one linguistic problem, a problem that has received considerable attention from analytic philosophers of language, and whose solution would have significant consequences for theories of reference and meaning. The problem that will concern us here is based upon the observation that we seem to be able to use a singular negative existential sentence such as () to make true and even informative assertions. This seemingly innocuous observation is problematic. For ‘Pegasus’ is a singular term (a name), and there are compelling theoretical reasons to suppose that, given that ‘Pegasus’ is a singular term, an assertion of () is true only if ¹ A singular term is more precisely a singular definite noun phrase. This category includes demonstratives and pronouns, proper names, and definite descriptions. Our discussion will focus on singular negative existentials whose subject terms are definite descriptions or names. ² Many of the relevant logical issues concern the tendentious notion of semantic presupposition, and its relation to negation and logical entailment. Many of the metaphysical issues concern the nature of existence and nonexistence. Here we are concerned with semantic and pragmatic issues, though of course these logical and metaphysical issues are relevant to our concerns.

OUP CORRECTED PROOF – FINAL, 17/1/2019, SPi



 ,  ,   

there is an entity referred to by the utterance of ‘Pegasus’ and that entity satisfies the predicate ‘does not exist’. But, given that there are no winged horses, it seems that utterances of ‘Pegasus’ do not refer to anything, and certainly do not refer to something that satisfies ‘does not exist’. Indeed, the assertion of () itself seems to imply that utterances of ‘Pegasus’ do not succeed in referring to anything. Thus assertions of singular negative existentials are in a way paradoxical: if what an assertion of () says is true, then it would seem that assertions of () would not be true (or false). So it seems that assertions of () cannot be true, but this conflicts with the observation with which we began. This is what we will take to be the problem of negative existentials. The problem can be presented as a valid argument, the conclusion of which is incompatible with the observation. The relevant premises, followed by some remarks of clarification and justification, are as follows: Referential Compositionality: An assertion of a sentence of the form ┌τ does not exist┐, where τ is to be replaced by a genuine singular term, is true only if the utterance of τ refers to an entity that does not exist, and false only if the utterance of τ refers to an entity that does exist.

This premise follows from a general principle of semantic compositionality, according to which assertions of sentences of the form τΠ (where τ is to be replaced by a genuine singular term and Π is to be replaced by a grammatically appropriate predicate) are true if and only if the utterance of τ succeeds in referring to something that satisfies Π, and false if and only if the utterance of τ succeeds in referring to something that does not satisfy Π.³ This general compositionality principle accounts for many intuitive observations. For example, it explains why assertions of ()

Obama smokes

are true only if there is an entity referred to by the utterance of ‘Obama’ and that entity smokes, and it explains why assertions of (∼) Obama does not smoke are true only if there is an entity referred by the utterance of ‘Obama’ and that entity does not smoke. It also explains why assertions of both ‘Pegasus is sleeping’ and ‘Pegasus is not sleeping’ are judged to be neither true nor false: if an utterance of ‘Pegasus’ fails to refer to an entity, then it fails to refer either to an entity that is sleeping or to an entity that is not sleeping.⁴

³ The general principle of referential compositionality originated in the work of Frege (, ). Similar principles of semantic compositionality underlie all systematic approaches to natural language semantics. (See Davidson ; Montague a, ). ⁴ Referential Compositionality thus assumes that the negation in negative existentials is presupposition preserving ‘choice’ negation as opposed to presupposition cancelling ‘exclusion’ negation. The distinction between these types of negation will be explained in §... The proposal that ‘not’ is ambiguous between choice and exclusion negation is discussed in §...

OUP CORRECTED PROOF – FINAL, 17/1/2019, SPi

 



The second premise simply states that the subject term of the particular negative existential under consideration is a genuine singular term. The relevant instance of the second premise for sentence () is Genuine Singular Term:

‘Pegasus’ is a genuine singular term.

It is not trivial to say what it is for a word to be a genuine singular term, but for now it will suffice to say that a name such as ‘Pegasus’ is a genuine singular term just in case the principle of Referential Compositionality applies to sentences in which it occurs. Thus, for example, if Russell () is correct that assertions of sentences containing denotationless definite descriptions (e.g. ‘the winged horse flies’) are false because the definite description is denotationless, then such definite descriptions are not genuine singular terms. (See §. for discussion.) From Referential Compositionality and Genuine Singular Term (by instantiation), we derive, Referential Presupposition: An assertion of ‘Pegasus does not exist’ is true only if the utterance of ‘Pegasus’ refers to an entity that does not exist, and false only if the utterance of ‘Pegasus’ refers to an entity that exists.

According to this (derived) premise, an assertion of () has a truth value only if the utterance of ‘Pegasus’ succeeds in referring to some entity; that is, referential presupposition states that an assertion of the relevant negative existential presupposes that the utterance of the subject term succeeds in referring to an entity, or, as we will sometimes put it, the assertion presupposes that the subject term is not empty. The next premise simply states that the referential presupposition described above fails for the negative existential under consideration. Thus, with regard to (), since nothing can be the referent of a relevant utterance of ‘Pegasus’ unless it is a winged horse and there are no winged horses, we have: Reference Failure: Relevant utterances of ‘Pegasus’ do not refer to anything. The restriction to relevant utterances is required because, for example, one might name one’s pet aardvark ‘Pegasus’ and then successfully refer to it with utterances of ‘Pegasus’. But of course it is not such uses of ‘Pegasus’ that are relevant to the problem of negative existentials. Thus, an utterance of ‘Pegasus’ is relevant just in case it is an utterance in a relevant assertion of a negative existential, and an assertion of a negative existential is relevant just in case it is intuitively true. Taken together then, Referential Presupposition and Reference Failure imply the conclusion that assertions of the negative existential under consideration lack a truth value. Thus the conclusion of our argument with regard to sentence () is: Presupposition Failure: Relevant assertions of ‘Pegasus does not exist’ are neither true nor false.

The conclusion, viz. Presupposition Failure, follows from Referential Compositionality, Genuine Singular Term, and Reference Failure, and, at least with regard to sentence (), these premises are intuitively correct, or at least very plausible. The problem, of course,

OUP CORRECTED PROOF – FINAL, 17/1/2019, SPi



 ,  ,   

is that the entailed conclusion contradicts the observation that ‘Pegasus does not exist’ can be used to make true and informative assertions. The present entry will review the proposals of a number of prominent analytic philosophers for avoiding, or perhaps explaining away, this contradiction. These proposals have been selected on the basis of the criteria of originality, plausibility, influence within mainstream analytic philosophy, clarity of exposition, and general accessibility. Some important work will, alas, not get the attention that it arguably deserves. (One noteworthy omission is the ‘pretense’ approach of Kendall Walton , and Gareth Evans .) After presenting each proposal and making clear how it responds to the recently rehearsed argument, criticisms will be presented, although not necessarily endorsed. Responses to these criticisms will (in some cases) be noted, but not discussed in any detail. Technicality and jargon will be eschewed so as to maximize accessibility. As a result, some degree of precision may be lost but not so much so as to distort the view in question. Finally, we should warn the reader that we do not here advocate any of the proposals that we consider, though in the concluding section we do offer a brief reactive summary.

. T   : R (, , )

.................................................................................................................................. Russell’s response to the problem of negative existentials has been very influential and has served to draw widespread attention to the problem in analytic philosophy. Russell motivates his (, , ) response by arguing for its superiority over the (then) available alternatives. Russell () attributes one of these alternatives to Meinong (, ), though in an earlier work () Russell himself advocates a version of Meinong’s response. According to Meinong’s response all meaningful utterances of singular terms refer to entities, yet some of these entities satisfy the predicate ‘does not exist’. Thus, in terms of the above argument, the response of Meinong and early () Russell is to deny Reference Failure: relevant utterances of ‘Pegasus’ do refer to an entity, an entity that has being, yet does not exist. If relevant utterances of ‘Pegasus’ do refer to an entity that does not exist, then the principle of Referential Compositionality entails that assertions of () are true. Against this proposed solution, later Russell (, ) levels two compelling criticisms. First, insofar as it appeals to a realm of non-existent beings, the response fails to preserve that “feeling for reality” which, according to the more mature Russell, ought to be preserved in “even the most abstract studies” (Russell : ). Second, the response violates the law of non-contradiction; it entails, for example, that the round square—a denizen of the disputed realm of being—is both round and (being square) not round.⁵ ⁵ Though Meinong’s view seems to lead to such contradictions, a number of theorists have attempted to develop Meinong’s general strategy and avoid the contradictory consequence. (See Parsons ; Fine ; and Zalta ).

OUP CORRECTED PROOF – FINAL, 17/1/2019, SPi

 



Russell () claims that theorists, such as his earlier self, who advocate Meinong’s response are “misled by grammar” and have “regarded grammatical form as a surer guide in analysis than, in fact, it is” (: ). On Meinong’s view, the subjectpredicate grammatical form of a negative existential is isomorphic with the logical form of the proposition expressed. Thus, for example, on Meinong’s view the subject-term of the sentence: ()

The winged horse does not exist.

refers to the winged horse (which has being) and the sentence predicates nonexistence of it. Russell’s later (, , ) proposal rejects this isomorphism; on Russell’s later view, there may be significant differences between ‘logical form’ and ‘grammatical form’. For our purposes, we may take grammatical form to be a feature of sentences; the grammatical form of a sentence is the relevant syntactic structure of the sentence. In contrast, logical form is a feature of the propositions expressed by sentences; the logical form of a proposition expressed by a sentence is the form that must be attributed to the proposition in order to explain the truth-conditions of the sentence.⁶ Russell () maintained that this sort of divergence between grammatical form and logical form is evident in sentences containing (singular) definite descriptions, phrases such as ‘the current president of the United States’, ‘the center of mass of the solar system’ and ‘the winged horse’. According to Russell’s theory of definite descriptions, the proposition expressed by ‘The king of France is bald’ has a logical form that is more accurately depicted by, ()

There is exactly one king of France and it is bald.

or in terms of familiar logical notation, ()

(∃x)[King-of-France(x) & (∀y)(King-of-France(y) → x=y) & Bald(x)]⁷

Thus, though the definite description ‘the king of France’ is the grammatical subject of the sentence ‘The king of France is bald’, it does not correspond to an entity in the proposition expressed by the sentence; the logical form of the proposition corresponds to the existential quantification displayed in (), and in () there is no genuine singular term, neither simple nor complex, which purports to refer to an entity that is king of France. As Russell () somewhat enigmatically explains, “the phrase per se has no meaning, because in any proposition in which it occurs the proposition, fully expressed, does not contain the phrase, which has been broken up” (: ). ⁶ Or so we shall understand Russell’s notion of logical form. Note that here we are attempting to follow Russell in assigning propositions to sentences, as opposed to utterances of sentences. It is difficult to discern a consistent usage of the term ‘proposition’ in Russell’s writings. ⁷ Russell (, , ) describes the relevant logical forms in terms of propositional functions; for Russell, determiners such as ‘all’ and ‘no’, as well as the indefinite and definite articles, express properties of propositional functions. For the sake of accessibility, we present Russell’s view in terms of classical first-order logic.

OUP CORRECTED PROOF – FINAL, 17/1/2019, SPi



 ,  ,   

Russell’s theory of descriptions thus analyses sentences whose grammatical subjects are definite descriptions as expressing existential quantification. This analysis implies that negated sentences such as ‘The king of France is not bald’ are structurally ambiguous; one grammatical form corresponds to two logical forms. The two logical forms that correspond to the grammatical form ┌ The Φ is not Ψ ┐ differ with respect to the relative scopes of the negation and the existential quantifier: ()

(∃x)[(Φx & (∀y)(Φy → x=y) & ∼Ψx]

()

∼(∃x)[(Φx & (∀y)(Φy → x=y) & Ψx]

Note that () entails that there exists a unique satisfier of Φ, whereas () can be true even if there is no satisfier of Φ. How does Russell’s analysis of definite descriptions provide a response to the problem of negative existentials? Assuming that a negative existential such as () is analogous to ‘The king of France is not bald’, a straightforward application of the analysis presented above yields the result that () is also ambiguous:⁸ ()

(∃x)[Winged-horse(x) & (∀y)(Winged-horse(y) → x=y)) & ∼Exist(x)]

()

∼(∃x)[Winged-horse(x) & (∀y)(Winged-horse(y) → x=y) & Exist(x)]

The proposition indicated by () is contradictory and false, as it claims that there exists a unique winged horse that does not exist. Indeed, it is precisely because () seems to express something like what () expresses that () is problematic; the problem is that () can nonetheless be used to make true and informative assertions.⁹ The interpretation provided by () is thus of no help in solving the problem. But note that, given that there are no winged horses, the reading provided by () is unproblematic and true. For () does not require for its truth that utterances of ‘the winged horse’ refer to an entity; on the contrary, since under this reading the negation takes wide-scope over the existential quantifier, () is true only if there is not a unique winged horse. Thus this reading provides a response to the problem of negative existentials. In terms of the argument presented in the introduction, Russell’s theory of definite descriptions provides a reason for rejecting the relevant instance of the premise dubbed Genuine Singular Term: according to Russell’s theory, the definite description ‘the winged horse’ is not a genuine singular term, and thus Russell has reason to deny that () is true only if the grammatical subject ‘the winged horse’ refers to an entity that does not exist. Indeed, on Russell’s later view, definite descriptions are not used to refer to entities at all; rather, as is displayed in () and (), definite descriptions are quantifier phrases

⁸ As will be explained below, Russell himself rejected the analogy. Russell maintained that though ‘the king of France is not bald’ is structurally ambiguous, negative existentials such as () are not structurally ambiguous. ⁹ The qualifier ‘something like’ is intended to allow room for a Strawson-inspired view, according to which an utterance of () does not entail that there is a winged horse, but merely presupposes that there is. The puzzle of course also arises for Strawson’s view.

OUP CORRECTED PROOF – FINAL, 17/1/2019, SPi

 



(what Russell called ‘denoting phrases’) and thus in Russell’s view definite descriptions are no more used to refer to an entity than is the quantifier phrase ‘nobody’. Does Russell’s theory of definite descriptions then solve the problem of negative existentials, at least for negative existentials whose grammatical subjects are definite descriptions? Russell’s proposal clearly has merit, and not surprisingly, amended versions of Russell’s proposal are endorsed by contemporary philosophers.¹⁰ It is not, however, without its shortcomings. For one thing, the ambiguity predicted by Russell’s proposal is problematic: it does not seem that sentences such as ‘The current king of France is not bald’ really are structurally ambiguous in the way predicted by Russell’s theory.¹¹ As Strawson (, ) observed, it is very difficult to interpret (an utterance of) a sentence with the grammatical form ┌ The Φ is not Ψ ┐ as expressing the truth-conditions indicated in (). For example, it is difficult to interpret a contemporary assertion of ‘The present king of France is not a king’ as being true because France does not presently have a king.¹² Similarly, it is very difficult to interpret an assertion of ‘The senator from California is not a senator’ as true because there are two senators from California.¹³ And negative existentials such as () also do not seem to be ambiguous in the way predicted by Russell’s theory. That is, () seems to have neither a contradictory and false reading nor an unproblematic and true reading. Rather, it seems that () has only one reading, and under this reading assertions of () are true, yet puzzling. And note that it is the reading predicted by Russell’s theory that is difficult to obtain—the reading represented by () and ()—that is required in order for Russell’s theory to provide a solution to the problem of negative existentials. The reading that is readily available—the reading provided by () and ()—simply reinforces the problem. In short, if the grammatical form of () is ambiguous in the way predicted by Russell’s theory of definite descriptions, then this theory provides a response, at least for negative existentials involving definite descriptions. But there is reason to doubt that negative existentials such as () really are structurally ambiguous in this way. A noteworthy, though often ignored, aspect of Russell’s own analysis of negative existentials is that under Russell’s analysis sentences such as () are not ambiguous, because Russell (, ) rejects the assumption that negative existentials such as () are analogous to, for example, ‘The winged horse does not fly’. The reason the analogy ¹⁰ Defenders of (versions of) Russell’s theory include Neale (); Abbott (); and Graff (). For a Strawson and Frege inspired alternative, see Elbourne (). ¹¹ Note that English does not allow negation of the definite article ‘the’ (at least not at the level of grammatical form). So, while ‘Not all dogs bark’ is perfectly grammatical, ‘Not the dog barks’ is not. If the reading indicated by () is available for (), this grammatical fact is surprising. ¹² It seems that the interpretation indicated by () is available only in special cases, such as presupposition denial, as is illustrated in the following exchange: Mary: ‘The king of France is bald.’ John: ‘No, the king of France is not bald. France has no king.’ John’s utterance of ‘the king of France is not bald’ is intuitively correct in part because he goes on to deny that France has a king. Neale () suggests that such uses of, for example, ‘the king of France is not bald’ provide evidence in support of Russell’s ambiguity thesis. It is not clear, however, that such special uses constitute evidence in support of the alleged structural ambiguity. (See Horn ; Burton-Roberts ; and Carston .) ¹³ See Horn (), especially Chapter , for an extensive discussion of the problematic ambiguity.

OUP CORRECTED PROOF – FINAL, 17/1/2019, SPi



 ,  ,   

fails for Russell is that he maintains that the grammatical predicate ‘exists’, and its negation ‘does not exist’, are to be treated in a radically different way than are the grammatical predicates ‘flies’ and ‘does not fly’; here again there is divergence between grammatical form and logical form.¹⁴ Russell (: –) maintains that though the sentence ‘Men exist’ has a grammatical subject-predicate form, the logical form of the proposition expressed by this sentence is more accurately rendered by ‘There is something that is a man’, or, utilizing some familiar logical notion ‘(∃x)(Man(x))’. Note that the predicate of this logical formula, viz. ‘Man(x)’, is not even remotely related to the inflected verb ‘exist’, which is the grammatical predicate of the sentence. Rather, according to Russell (: –), the inflected verb ‘exist’ corresponds in logical form not to a predicate, but rather to an existential quantifier.¹⁵ This analysis of the grammatical predicate ‘exists’ as corresponding in logical form to an existential quantifier is apparently what motivates Russell to maintain that () is not ambiguous between the logical forms depicted in () and (), but instead expresses a unique logical form that contains no predicate corresponding to the verb ‘exists’: ()

∼(∃x)[(Winged-horse(x) & (∀y)(Winged-horse(y) → x=y)]

This analysis of () has the advantage of avoiding the problematic ambiguity that results from the more straightforward application of Russell’s theory.¹⁶ But, the proposal that the predicate ‘does not exist’ corresponds in logical form to a negated quantifier and not to a negated predicate is problematic for other reasons. First, this analysis of the grammatical predicate ‘exists’ seems to leave us with one too many existential quantifiers in the logical forms of both, for example, ‘The flying horse exists’ and ‘The flying horse does not exist’ because Russell seems to maintain that both the grammatical subject ‘the flying horse’ and the grammatical predicate ‘exists’

¹⁴ Russell’s analysis of the grammatical predicate ‘exists’ is at least reminiscent of Kant’s view that “‘Being’ is obviously not a real predicate” (Critique of Pure Reason, A/B-A/B)). But, as Kripke (: , note ) notes, it is difficult to interpret Kant as proposing that the grammatical predicate ‘exists’ expresses a second-level property, a property of something like Russell’s propositional functions, for Kant does not seem to invoke anything analogous to Russell’s propositional functions. ¹⁵ In Russell’s system, the verb ‘exists’ corresponds in logical form to a property of proposition functions. ¹⁶ Kripke (/: –) observes that the unique analysis of negative existentials that results from Russell’s treatment of ‘exists’ as corresponding in logical form to an existential quantifier yields incorrect results when applied to negative existentials containing modal operators whose logical subjects are rigid designators. So, for example, ‘Obama might not have existed’ can be analyzed by Russell only as something like (i)

◊[∼(∃x)[th-president(x) & (∀y)(th-president(y) → x=y)]]

That is, the following alternative reading is not available to Russell: (ii)

[(∃x)[th-president(x) & (∀y)(th-president(y) → x=y) & ◊[∼Exist(x)]]

But, while (ii) represents a plausible analysis of ‘Obama might not have existed’, (i) does not.

OUP CORRECTED PROOF – FINAL, 17/1/2019, SPi

 



correspond to existential quantifiers in logical form.¹⁷ In order to derive the logical form indicated by (), Russell must maintain that grammatical definite descriptions correspond to existential quantifiers in logical form, but only if they are not somehow cancelled by grammatical existence predicates. This is implausible and ad hoc. Thus, even those who endorse Russell’s theory of definite descriptions have reason to be skeptical of his treatment of the grammatical existence predicates ‘exists’ and ‘does not exist’.¹⁸ A further, and widely noted, problem for Russell’s proposed solution concerns negative existentials that (at least in grammatical form) contain names or indexicals (pronouns or demonstratives) instead of definite descriptions. Granting for the sake of argument that Russell’s theory of definite descriptions provides a solution for negative existentials such as (), how might the theory be extended to apply to negative existentials such as ()? Russell claims that his theory of definite descriptions also applies to negative existentials such as (), because, again, the grammatical form of these sentences is not an accurate guide to the logical form of the propositions they express; what grammatically are names also correspond to what logically are existential quantifier phrases. Thus, Russell () claims that “a [sentence] about Apollo means what we get by substituting what the classical dictionary tells us is meant by [‘Apollo’], say ‘the sun-god’” (: ). And Russell () maintains that this holds for almost all singular terms: We may even go so far as to say that, in all such knowledge that can be expressed in words—with the exception of ‘this’ and ‘that’ and a few other words of which the meaning varies on different occasions—no names, in the strict sense, occur, but what seem like names are really descriptions. We may inquire significantly whether Homer existed, which we could not do if ‘Homer’ were a name. . . . If ‘a’ is a name, it must name something: what does not name anything is not a name . . . . (: , emphasis in original)

Russell’s proposal for extending his solution so that it applies not only to negative existentials whose subject terms are definite descriptions, but also those whose grammatical subjects are (apparent) names is problematic, and controversial. First, it should be noted that for any negative existential whose subject term is a genuine singular term

¹⁷ A related problem is that Russell’s analysis seems to imply that utterances of definite descriptions all by themselves, independent of any grammatical predicate, should express truth-evaluable existentially quantified propositions. That is, if the grammatical sentence ‘The king of France is bald’ corresponds to an existentially quantified conjunctive logical form that is accurately represented by (), then it would seem that the definite description alone, without the grammatical predicate ‘is bald’, ought to correspond to an existentially quantified proposition: (*) (∃x) [King-of-France(x) & (∀y)(King-of-France(y)→ x=y)] ¹⁸ For example Neale (: –) defends Russell’s insight that definite descriptions are quantifier phrases, but proposes treating quantifiers as restricted quantifiers in the style of Barwise and Cooper (). But this very plausible view of definite descriptions is incompatible with Russell’s () analysis of ‘exists’ and ‘does not exist’.

OUP CORRECTED PROOF – FINAL, 17/1/2019, SPi



 ,  ,   

(what Russell would call a name in the ‘strict sense’) Russell’s proposed solution simply does not apply. Thus, on Russell’s view, there can be no true negative existential of the form ┌τ does not exist┐ where τ is a genuine singular term. Indeed, given his logical analysis of the grammatical predicate ‘exists’ as corresponding to existential quantification, Russell is forced to maintain that such a negative existential is “a mere noise or shape, devoid of significance” (: ). This is implausible. While it may, perhaps, be plausible to question whether there are any genuine singular terms (Quine, as we will see, does perhaps question this.), it cannot plausibly be denied that if there are genuine singular terms, then they can occur in grammatical and meaningful negative existentials. But even if one rejects Russell’s (independently implausible) analysis of grammatical existence predicates, the case of negative existentials whose subject terms are genuine singular terms is problematic: for, as Russell intimates in the above passage, if we can even coherently suppose that a person might think that a genuine singular term τ refers when in fact it does not refer, then the problem of negative existentials will arise for occurrences of ┌τ does not exist┐. If we can conceive of a situation in which somebody might falsely believe what is expressed by┌τ exists┐, we can thereby imagine being able to meaningfully and truly assert ┌τ does not exist┐. Avoiding this difficulty is one of the motivations for Russell’s doctrine of acquaintance: “in every proposition that we can apprehend (i.e., not only in those whose truth or falsehood we can judge of, but in all that we can think about), all the constituents are really entities with which we have immediate acquaintance” (: ). And, of course, Russell maintains that one can be acquainted only with entities that exist. No wonder then that Russell suggests that perhaps only (uses of) ‘this’ and ‘that’ and other indexical terms qualify as uses of genuine singular terms. For it is at least plausible that if one apprehends the proposition expressed by an utterance of ‘that does not exist’, one is guaranteed to know that the proposition is false.¹⁹ By allowing that there are genuine singular terms (names ‘in the strict sense’), Russell has painted himself into a corner: if it were possible for one to apprehend the propositions expressed by sentences in which genuine singular term τ occurs even if τ had no referent, then it would be possible to truly and meaningfully assert┌τ does not exist┐. But, by the argument presented in the introduction, this leads to a contradiction, and the theory of definite descriptions, by assumption, does not provide a solution in cases involving genuine singular terms. So, if genuine singular term τ has no referent, it must be impossible to apprehend the propositions expressed by sentences in which τ occurs. The doctrine of acquaintance rules out this problematic possibility, but this doctrine is, to say the least, controversial. Of course one could avoid this controversy by simply denying that there are any genuine singular terms; perhaps all apparent singular terms are really descriptions, in Russell’s sense. This view would follow Russell in ¹⁹ That uses of ‘I’ are also plausibly viewed as uses of a genuine singular term is suggested by the cogency of Descartes’s cogito: it is plausible that my merely apprehending the propositions expressed by my utterances of ‘I exist’ and ‘I do not exist’ guarantees that I know that the former is true and the latter false.

OUP CORRECTED PROOF – FINAL, 17/1/2019, SPi

 



proposing to solve the problem of negative existentials by rejecting relevant instances of the premise dubbed Genuine Singular Term, but would differ from Russell’s view in that it would deny that there are any genuine singular terms: according to this sort of extreme descriptivism, any apparent singular term which we understand is, in the relevant logical sense, a description. This is the view advocated by Quine (, ).

. T     G S T: Q (, , )

.................................................................................................................................. In ‘On what there is’ () Quine expands on Russell’s criticisms of Meinong’s solution to the problem of negative existentials, and he also promotes and develops Russell’s (, ) solution. Quine considers a version of Meinong’s proposal, which he attributes to a fictional philosopher dubbed ‘Wyman,’ according to which a posited non-existent entity such as the alleged referent of ‘Pegasus’ is “an unactualized possible” (: , in Quine ). In addition to hinting that such entities offend against the ontological principle known as Occam’s razor, Quine presents a number of reasons which weigh against an ontology of mere ‘possibles.’ First, the view under consideration faces perplexing questions concerning the identities of the posited ‘possibles’: Let us suppose that there is in fact no actual man in the doorway. Wyman’s theory then posits that there is a non-existent being that is the referent of ‘the fat man in the doorway’ as well as a non-existent being that is the referent of ‘the tall man in the doorway’. Is the fat man in the doorway identical to, or distinct from, the tall man in the doorway? There seems to be no way of making sense of this question, and Quine maintains that we ought not posit entities for which there are no identity conditions. Second, as Russell (, ) noted, Meinong’s proposed solution seems to require positing contradictory non-existent entities such as the round square. Thus it seems that Wyman cannot stop at positing ‘possibles,’ prompting Quine to ask rhetorically, “Can we drive Wyman now to admitting a realm of unactualizable impossibles?” (: , in Quine ). Quine’s most forceful criticism of Meinong’s response, however, is that it does not really solve the problem of negative existentials, but merely makes it slightly more difficult to state the problem. Quine disparages Wyman, as an advocate of Meinong’s solution, as “one of those philosophers who have united in ruining the good old word ‘exist’” (: , in Quine ).²⁰ Quine’s point here is not simply that Wyman is distorting the meaning of the verb ‘to exist’. Rather Quine’s point is that Meinong’s proposal merely succeeds in ruining the good old word ‘exist’; the problem of negative ²⁰ Given Russell’s proposal to analyze the verb ‘exists’ as an existential quantifier, one wonders if Quine would include Russell as one of the united philosophers.

OUP CORRECTED PROOF – FINAL, 17/1/2019, SPi



 ,  ,   

existentials does not disappear once one agrees to use ‘exist’ such that there might be real entities that satisfy ‘does not exist’. For, even if one agrees to use ‘exist’ so as to allow that some real things might satisfy the predicate ‘does not exist’, Quine will nonetheless think there is no such thing as the entity posited by Meinong (and Wyman) to serve as the ‘non-existent’ referent of ‘Pegasus’. Of course, given the fact that Meinong has “ruined” the verb ‘exist’, Quine cannot express his differing ontological view by asserting the negative existential “the entity posited by Wyman to serve as the ‘non-existent’ referent of ‘Pegasus’ does not exist”, but creating this lexical inconvenience does not solve the problem. Quine explains, Wyman . . . genially grants us the nonexistence of Pegasus, and then, contrary to what we meant by nonexistence of Pegasus, insists that Pegasus is. Existence is one thing, he says, and [being] is another. The only way I know of coping with this obfuscation of issues is to give Wyman the word ‘exist’. I’ll try not to use it again; I still have ‘is’. (: , in Quine )

Quine’s point is that Wyman’s proposed solution only succeeds in making it slightly more difficult to raise the problem of negative existentials. For consider, Quine rejects Wyman’s “overpopulated universe” of mere possibles; Quine thinks there are no such beings. Thus Quine might assert his rejection of Wyman’s ontology of non-existent referents by asserting a sentence such as: ()

The entity posited by Wyman to serve as the ‘non-existent’ referent of ‘Pegasus’ is not.

This sentence, though it avoids use of the verb ‘to exist’, is equivalent to a negative existential. And since it does not use the ruined predicate ‘does not exist’, Meinong’s proposal is powerless to explain the puzzling fact that Quine thinks that assertions of () are true.²¹ Quine’s positive proposal for solving the problem of negative existentials is to extend Russell’s (, ) solution in two ways. First, Quine advances a novel proposal with regard to how grammatical proper names such as ‘Pegasus’ are to be, as Russell (: ) puts it, “broken up,” and analyzed at the level of logical form as quantifier phrases. A potentially serious difficulty with Russell’s proposal, noted by Quine, concerns the fact that names do not always admit of “pat translation into a descriptive phrase” (Quine : , in Quine ). In response, Quine proposes a convenient and reliable method for translating such names into appropriate descriptive phrases. As he explains: If the notion of Pegasus had been so obscure or so basic a one that no pat translation into a descriptive phrase had offered itself along familiar lines, we could still have

²¹ Quine does not develop this criticism of Wyman’s proposed solution to the problem of negative existentials in detail. This is because Quine () is more interested in rejecting Wyman’s ontology than he is in rejecting Wyman’s proposed solution to the problem of negative existentials. We will return to this difference in focus below.

OUP CORRECTED PROOF – FINAL, 17/1/2019, SPi

 



availed ourselves of the following artificial and trivial-seeming device; we could have appealed to the ex hypothesi unanalyzable, irreducible attribute of being Pegasus, adopting for its expression, the verb ‘is Pegasus’ or ‘pegasizes’. The noun ‘Pegasus’ itself could then be treated as derivative, and identified after all with a description: ‘the thing that is Pegasus’, ‘the thing that Pegasizes’. (: , in Quine )

Second, Quine, at least in theory, proposes applying Russell’s theory of definite descriptions to all grammatical singular terms. Quine explains the motivation for extending Russell’s proposal in this way: The theoretical advantages of so doing are overwhelming. The whole category of singular terms is thereby swept away, so far as theory is concerned; for we know how to eliminate descriptions. In dispensing with the category of singular terms we dispense with a major source of theoretical confusion, . . . In particular, we dispense altogether, in theory, with the perplexing form of notion ‘a exists’: for we know how to translate singular existence statements into more basic logical terms when the singular term involved is a description. (Quine : )

By treating all singular terms as descriptions, and thus analyzable via Russell’s theory, Quine can avoid the above described perplexities concerning Russell’s (, ) notion of a genuine name, a name to which the theory of definite descriptions does not apply. For if there are no such genuine names—not even ‘this’ and ‘that’—then every utterance of a negative existential (and positive existential) will be amenable to explanation by way of Russell’s theory of descriptions. Does Quine’s extension of Russell’s proposal succeed in solving the problem of negative existentials? There are reasons to be skeptical. First, it should be noted that, as it is an extension of Russell’s proposal, Quine’s proposal inherits many of the problems faced by Russell’s proposal. In particular, even if one accepts Quine’s proposal for how Russell’s theory of definite descriptions can be applied to grammatical singular terms, Quine’s proposal will also have the problematic consequence of counterintuitive ambiguities. So, for example, even if one accepts that ‘Obama’ can be analyzed in terms of the artificial verb ‘obamaizes’, one will still face the counterintuitive result that, for example, ‘Obama does not smoke’ is predicted to be structurally ambiguous. Second, it is not clear how to extend Quine’s Russell-inspired analysis so that it applies to indexicals and deictic uses of pronouns. Quine (, ) does not consider such cases in detail, but in Quine () he maintains that “we can assimilate the demonstrative singular terms to singular descriptions, treating ‘this (that) apple’ as ‘the apple here (there)’” (: ). But this sketch merely describes how the apparent genuine name ‘this apple’ might be analyzed in terms of the apparent genuine name ‘there’; the analysis merely trades one demonstrative expression for another. Indeed, the sorts of phenomena described by Perry () would seem to be deeply problematic for Quine’s dispensing of indexicals and demonstratives via Russell’s theory of descriptions. Perry, following Castañeda (), argues that the intuitive meanings

OUP CORRECTED PROOF – FINAL, 17/1/2019, SPi



 ,  ,   

of many assertions of sentences containing indexicals and demonstratives cannot be expressed by statements that do not contain such context-sensitive expressions. If some uses of indexicals are essential in this way, Quine’s proposal for “dispensing with” all singular terms would seem to be doomed. Third and finally, it is far from clear that Quine’s admittedly “artificial” analysis can be invoked to provide a solution to the problem of negative existentials. For, unlike Russell’s proposed analyses of grammatical names, Quine’s “artificial” proposal does not seem to be explanatorily adequate. Russell proposed that “when we ask whether Homer existed, we are using the word ‘Homer’ as an abbreviated description” (: , my emphasis). Russell is not merely claiming that if we were to use ‘Homer’ as an abbreviated description, then we would avoid the problem of negative existentials; rather Russell seems to be claiming that we do in fact use ‘Homer’ as an abbreviated definite description, and moreover that it is because ‘Homer’ abbreviates (something like) ‘the author of the Iliad and the Odyssey’ (: ) that we can coherently wonder whether or not Homer exists, and can coherently assert the negative existential ‘Homer does not exist’. In contrast, Quine’s “artificial and trivial-seeming device” seems incapable of providing such explanations. For consider, we know what the newly coined verb ‘pegasizes’ means only because Quine tells us that it means what ‘is Pegasus’ means. But we know what the predicate ‘is Pegasus’ means only if we know what the singular term ‘Pegasus’ means. But what is the meaning of the singular term ‘Pegasus’? More specifically, what could the meaning of ‘Pegasus’ be, given that we can coherently wonder whether or not Pegasus exists, and moreover use ‘Pegasus does not exist’ to say something true, and even informative?²² Being told that “the noun ‘Pegasus’ itself could . . . be treated as derivative, and identified . . . with a description” (Quine : , in Quine , my emphasis) does not seem relevant to answering the question of why utterances of ‘Pegasus does not exist’ actually are understood as saying something true. Our question is, ‘Why do utterances of () seem to be true?’; it is of no help to be told that we could express what is expressed by an utterance of () by uttering ‘there exists a unique thing that pegasizes’. Before ending our discussion of Quine’s proposal, it should be noted that Quine himself would probably not be troubled by any of the above criticisms. We have been assuming that Quine offered his proposal as a solution to a problem presented in the introduction; we have been assuming that Quine, following Russell, is attempting to solve the problem of negative existentials by motivating the rejection of all instances of the second premise, viz. Genuine Singular Term. But Quine seems not to have advanced his proposal as a solution to the problem of negative existentials. In characterizing his proposal Quine admits, in what would seem to be a direct conflict with Russell’s view, that “truth values seem to attach to singular statements only conditionally upon existence of the named object. . . . there would seem, under ordinary usage, to be no way of adjudicating the truth values of ‘Pegasus flies’ and ‘∼Pegasus flies’; the ²² Horn (: ) complains that Quine cannot “take much pride in barring Pegasus from the front door only to have the property pegasizing fly in through the window.”

OUP CORRECTED PROOF – FINAL, 17/1/2019, SPi

 



nonexistence of Pegasus seems to dispose of the question without answering it” (: ). How is this acceptance of the inadequacy of Russell’s theory with regard to, for example, ‘Pegasus flies’ to be rendered consistent with Quine’s utilization of Russell’s theory to “dispense with” all singular terms? The answer is that in proposing to “dispense with” all singular terms, Quine is not advancing an explanation of any phenomenon in natural language, rather he is proposing a revision of natural language. Quine is not interested why we judge some occurrences of negative existentials to be true; rather Quine is proposing a revision of natural language which will enable us to discuss the ontological question of what there is without having to face the problem of negative existentials. Given that Quine’s prescription to dispense with all singular terms is not an explanatory proposal, it would seem to be irrelevant to the problem of negative existentials presented in the introduction.

. D : K (/)

.................................................................................................................................. The irrelevance of Quine’s Russell-inspired revisionary proposal does not of course preclude the possibility of there being some other Russell-inspired explanatory proposal for analyzing grammatical proper names as quantifier phrases. However, in  Saul Kripke gave a series of lectures, later published as Naming and Necessity, that persuasively criticize Russell’s basic idea that grammatical proper names are “disguised” or “abbreviated” definite descriptions. Kripke’s arguments have persuaded even staunch defenders of Russell’s quantificational analysis of definite descriptions to reject Russell’s proposed analysis of grammatical proper names as quantifier phrases.²³ Kripke presents several influential criticisms, but here we will review only what is known as the semantic argument. The main idea of the semantic argument is straightforward: it is possible to successfully refer to an object or individual x by uttering its name τ even if the description one associates with τ does not denote x or even if the description denotes some other object/individual y. Thus, to use Kripke’s now famous example, one can refer to Gödel by uttering ‘Gödel’ even if it turns out that the description one happens to associate with that name (e.g., ‘the discoverer of the incompleteness of arithmetic’) fails to uniquely describe Gödel, or even if the associated description turns out to uniquely describe someone else. Kripke has us suppose that the description we associate with ‘Gödel’ is ‘the discoverer of the incompleteness of arithmetic’, and then he playfully asks us to imagine that actually it was poor Schmidt—an unknown, though talented,

²³ According to Neale, “In view of the important differences in the logical behavior of genuine referring expressions and definite descriptions . . . we must certainly follow Kripke . . . in rejecting the view that ordinary proper names are disguised definite descriptions” (: ).

OUP CORRECTED PROOF – FINAL, 17/1/2019, SPi



 ,  ,   

logician—who is the real discoverer of the incompleteness proof. If it were revealed that it was actually Schmidt who discovered the proof, we would not take ourselves to have been referring to Schmidt with our uses of ‘Gödel’. (We would not, for example, announce our surprise by uttering, ‘Oh my gosh, Gödel is Schmidt!’) In addition to advancing compelling arguments against descriptivism, in Naming and Necessity Kripke also advanced a positive “picture” of how the referent of a name is determined. According to Russell and the description theory, the referent of a name is determined indirectly, by way of the ‘identifying’ description with which that name is associated; for example, ‘Feynman’ refers to whichever unique individual (if there is one) is correctly described by the definite description that the name “abbreviates.” Kripke, as we noted above, has compelling reasons to reject this view, and in its place he proposes an alternative picture: . . . a baby is born; his parents call him by a certain name. They talk about him to their friends. Other people meet him. Through various sorts of talk the name is spread from link to link as if by a chain. A speaker who is on the far end of this chain, who has heard about, say Richard Feynman, in the market place or elsewhere, may be referring to Feynman even though he can’t remember from whom he first heard of Feynman or from whom he ever heard of Feynman. He knows that Feynman is a famous physicist. A certain passage of communicate reaching ultimately to the man himself does reach the speaker. He is then referring to Feynman even though he cannot identify him uniquely. (Kripke : )

In Kripke’s new picture the referent of a use of a name is not determined privately by a description that the speaker mentally associates with the name, but rather publicly by a chain of communication that is created by a community of speakers. This picture of how the referents of names are (or are not) determined has significant implications with regard to the meaning of names. In Kripke’s chain-of-communication picture, being a competent user and interpreter of a name does not require cognizance of a uniquely identifying description of the referent. Consequently, Kripke’s picture implies that the meanings of names cannot be adequately characterized as associated descriptions or concepts of the purported referent. Rather, the new picture supports the “Millian” view (named after early advocate J. S. Mill /) that names do not have meanings, and instead are mere “tags” for their referents. So on the new picture the semantic contribution of a name to propositions expressed by sentences in which the name appears is not a description that serves to indirectly fix the referent of the name; rather, the semantic contribution of a name is simply the referent of the name, assuming, of course, that the name is not empty. The new picture thus implies that grammatical names are directly referential terms, very similar to Russell’s genuine logical names.²⁴

²⁴ Direct reference theorists allow that some so-called “descriptive names” are correctly viewed as disguised descriptions. See Kripke’s discussion of ‘Jack the Ripper’ (: ).

OUP CORRECTED PROOF – FINAL, 17/1/2019, SPi

 



Taking their lead from Kripke (/) a wave of direct reference theorists from the mid s onward have rejected Frege and Russell’s descriptive view of names, and endorsed the Millian view described above.²⁵ But, rejection of the Frege–Russell descriptive approach to the meanings of names comes at a significant cost. If Kripke’s arguments compel us to reject the descriptive view of grammatical names, then we are precluded from endorsing Russell’s proposed solution to the problem of negative existentials, at least for cases involving grammatical names.²⁶ That is, even if Russell’s theory of definite descriptions enables one to reject Genuine Singular Term for negative existentials whose grammatical subjects are definite descriptions—and for reasons presented above the adequacy of Russell’s proposed solution even for these cases is dubious—the direct reference view of grammatical names would seem to preclude rejection of Genuine Singular Term for negative existentials whose grammatical subjects are names.²⁷ In the remainder of this section we will consider four alternative responses to the problem of negative existentials, all of which accept that the relevant occurrences of names are genuine singular terms.²⁸

.. The “Historical Block” view: Donnellan () Donnellan advocates a “historical explanation” theory of reference for names that is very similar to the chain-of-communication “picture” sketched by Kripke. Thus Donnellan, as a direct reference theorist, must provide an alternative response to the problem of negative existentials. Moreover, Donnellan, like the later Russell and Quine, seeks to avoid “the Meinongian population explosion” (: ). Donnellan’s approach is deceptively simple. As a direct reference theorist, Donnellan is compelled to maintain that an utterance of a negative existential involving a name is true if and

²⁵ For Frege (), empty names lack a reference and so statements containing them lack a truth value. Frege suggests assigning a dummy reference in such cases but then the resultant statement—even if an intuitively true negative existential—presumably will emerge as false on the grounds that the assigned referent does indeed exist. ²⁶ The direct reference theorist must also reject the solutions descriptivism provides for other problems. Thus the direct reference theorist is compelled to provide alternative solutions to “Frege’s puzzle of identity,” as well as the problem of “substitution failure in opaque contexts.” Solutions to these other problems that are compatible with the direct reference view of grammatical names are advanced in Salmon () and Soames (, ). Solutions to these other problems that are compatible with the direct reference view of demonstratives and indexicals are advanced in Kaplan () and Perry (, ). ²⁷ Kaplan () persuasively argued that indexicals and demonstratives are also directly referential; indeed, the term ‘directly referential’ can be traced to Kaplan. ²⁸ In subsequent sections we will consider the metalinguistic response of Bach (a, ) and Katz (, ). These theorists claim that though utterances of grammatical proper names are directly referential when they appear in the sorts of sentences and communicative contexts utilized in Kripke’s influential arguments, within assertions of negative existentials utterances of grammatical proper names are nonetheless appropriately analyzed as Russell-inspired descriptions.

OUP CORRECTED PROOF – FINAL, 17/1/2019, SPi



 ,  ,   

only if the utterance of the name lacks a referent. His “historical explanation” theory articulates necessary and sufficient conditions for an individual to be the referent of an utterance of a name. In this obvious way then, the historical explanation theory provides us with the conditions under which an utterance of a negative existential involving a name is true: the utterance is true just in case there is no individual who satisfies the necessary and sufficient conditions for being the referent of the utterance of the name. Let us walk through this derivation of the truth-conditions for utterances of negative existentials step by step. The first step is to articulate the necessary and sufficient conditions for an utterance of a name to have a referent. According to Donnellan, an utterance of a name refers to an individual just in case there is the right sort of historical explanatory relationship between the utterance and the individual: When a speaker uses a name intending to refer to an individual and predicate something of it, successful reference will occur when there is an individual that enters into the historically correct explanation of who it is that the speaker intended to predicate something of. That individual will then be the referent and the statement made will be true or false depending on whether it has the property designated by the predicate. (Donnellan : )

An utterance of name will thus be empty just in case there is no individual that stands in the right sort of historical explanatory relationship with the utterance. Donnellan considers the case of a child asserting ‘Santa Claus comes tonight’: “The historical explanation of this . . . does not involve any individual who could count as the referent of ‘Santa Claus’; rather it ends in a story given to him by his parents” (: ). Donnellan says that “when the historical explanation of the use of a name (with the intention to refer) ends in this way with events that preclude any referent being identified” the history ends in a block. Thus, in general, an utterance of a name is empty just in case the historical explanation of the utterance ends in a block. The second step is to deploy this account of the conditions under which an utterance of a name is empty to provide an account the truth-conditions for utterances of negative existentials involving names. Donnellan articulates these truth conditions with the following “rule”: (R) If N is a proper name that has been used in predicative statements with the intention to refer to some individual, then “N does not exist” is true if and (Donnellan : )²⁹ only if the history of those uses ends in a block.

²⁹ It is significant that rule (R) applies to only names that have “been used in predicative statements with the intention to refer to some individual.” Donnellan’s category of predicative statements excludes statements made in “discourses about fiction” and also excludes statements affirming or denying the existence of something. What is hinted at here is that a negative existential involving a name N is used to correct previous utterances that are believed to mistakenly presuppose the existence of the referent of N.

OUP CORRECTED PROOF – FINAL, 17/1/2019, SPi

 



Donnellan acknowledges that his idea of a “block,” and the historical explanation theory itself, is insufficiently precise. But Donnellan claims that, despite its imprecision, rule (R) has significant consequences for the problem of negative existentials: Even without worrying about the vagueness of the idea of a “block,” R may look unexciting, but its consequences are interesting. In the first place its form is completely antithetical to the principle of identifying descriptions, for it has nothing to do with whether an individual of a certain description existed or not. Second, it does not involve our theory of reference in any difficulties: there is the connection with the notion of historical explanation and so it ties in neatly with the positive aspects of the view, but it has no Meinongian implications, no overpopulation with entities whose existence is being denied. (Donnellan : –)

Does rule (R) provide a solution to the problem of negative existentials? If the historical explanation theory of reference for names is assumed, then endorsing rule (R) requires denying, Referential Compositionality: An assertion of a sentence of the form ┌τ does not exist┐, where τ is to be replaced by a genuine singular term, is true only if the utterance of τ refers to an entity that does not exist, and false only if the utterance of τ refers to an entity that does exist.

Indeed, for a direct reference who wants to avoid the “Meinongian population explosion” (: ), there does not seem to be any other response available. For, in rejecting descriptive theories of reference, a direct reference theorist is precluded from denying Genuine Singular Term, and in rejecting the “Meinongian population explosion,” a direct reference theorist is precluded from rejecting Reference Failure. And thus denying Referential Compositionality seems to be the only remaining option. But does endorsing Rule (R), and thereby denying the fundamental premise of Referential Compositionality, suffice as an adequate solution to the problem? Donnellan himself registers doubts concerning the explanatory adequacy of rule (R). Immediately before stating rule (R) Donnellan warns that “this rule . . . does not provide an analysis of [negative existentials involving names]; it doesn’t tell us what such statements mean or what propositions they express” (: ). What motivates this hedging? An answer is suggested by Donnellan’s admission that the interesting consequences of rule (R) (cited in the passage above) are “bought, to be sure, at the price of making a name function differently in existence statements as opposed to predicative statements” (: ). But why does rule (R) exact this price? As we have seen, rule (R) is incompatible with Referential Compositionality. And, as was explained in the introduction, this premise is an instance of a very plausible general principle of semantic compositionality: assertions of sentences of the form τΠ are true if and only if the utterance of τ succeeds in referring to something that satisfies Π, and false if and only if the utterance of τ succeeds in referring to something that does not satisfy Π. Donnellan accepts this compositionality principle for what he calls predicative

OUP CORRECTED PROOF – FINAL, 17/1/2019, SPi



 ,  ,   

statements involving names.³⁰ But, since rule (R) is incompatible with this general principle, Donnellan must pay the price of rejecting this general semantic principle in the special case of negative existentials involving names. The upshot is that because Donnellan endorses rule (R), he is precluded from invoking the general compositionality principle to explain why utterances of, for example, ‘Pegasus does not exist’ are true. In summary then, Donnellan correctly determines that direct reference theorists who reject Meinong’s “population explosion” must deny Referential Compositionality; but Donnellan provides no alternative explanation as to why negative existentials involving names are true, given that the names occurring therein lack referents.³¹

.. The “No Such Proposition” view: Kripke (/) In , three years after presenting the lectures published as Naming and Necessity, Kripke presented the John Locke Lectures, and in this later series of lectures he advances a proposal for how a direct reference theorist might solve the problem of negative existentials. (Kripke’s John Locke Lectures have recently been published as Reference and Existence, .) Kripke begins by noting that a direct reference theorist is committed to the view that assertions of (simple) sentences containing names that lack referents do not express propositions. Let us suppose, following Kripke, that ‘Vulcan’ is such an empty name. Kripke argues that the direct reference view is committed to the claim that assertions of, for example, ‘Vulcan is red’ do not express propositions; and of course, if such assertions do not express propositions, they do not express false propositions. Nonetheless, Kripke argues, we have a “strong inclination” to say of somebody who made such an assertion that they are wrong, and moreover to judge the assertion to be false. Kripke considers the case of an astronomer who mistakenly believes that ‘Vulcan’ refers to a planet that is red: . . . the astronomer believes that there is a true proposition about Vulcan, that it is red. He is wrong, not because the proposition is false but because there is no such true proposition. Nevertheless, first, it is quite plainly correct to call him wrong, in some sense, when he says that Vulcan is red, because he believes wrongly that there is a true proposition which he expressed by the words ‘Vulcan is red’. Second, one might, though this is somewhat sloppy, have a strong inclination to call this false. (Kripke : , italics in original)

³⁰ In the passage cited above, a passage in which Donnellan is invoking the intuitions that support the historical explanation theory, Donnellan claims that for a simple predicative statement involving a name and a predicate, “the statement made will be true or false depending on whether it has the property designated by the predicate” (: ). ³¹ Despite its limitations, Donnellan’s view has been developed in various ways by contemporary direct reference theorists. Amie Thomasson (), for example, invokes Donnellan’s notion of a “block” in defending her Meinong-inspired account of negative existentials, and Perry (), invokes the notion in his Russell-inspired account.

OUP CORRECTED PROOF – FINAL, 17/1/2019, SPi

 



Though the inference is not entirely explicit, Kripke infers that to the extent that we are—perhaps only because we are sloppy—inclined to judge an assertion of ‘Vulcan is red’ as false, we will be inclined to judge assertions of the negation, viz. ‘Vulcan is not red’, as true. That is, if we are inclined to judge an assertion of ‘Vulcan is red’ as false, not because it expresses a proposition that is false, but rather because it does not express a true proposition, then we will be inclined to judge assertions of the negation, viz. ‘Vulcan is not red’, as true, not because ‘Vulcan is red’ expresses a proposition that is false, but because ‘Vulcan is red’ does not express a true proposition.³² And note that this “strong inclination” to judge an assertion of ‘Vulcan is red’ as false is a “strong inclination” to reject the general compositionality principle which supports the fundamental premise of Referential Compositionality. The above explanation of our “strong inclination” to judge an assertion of ‘Vulcan is red’ as false is readily extended to an explanation of our “strong inclination” to judge an assertion of ‘Vulcan does not exist’ as true. The one difference, as Kripke notes, is that whereas one might judge that an assertion of ‘Vulcan is red’ does not express a true proposition without knowing whether or not ‘Vulcan’ has a referent and thus without knowing whether or not the assertion expresses a proposition at all, in the case of an assertion of ‘Vulcan exists’ one knows that if no true proposition is expressed, then no proposition is expressed at all. For example, if one knew that nothing was red, then one could thereby know that ‘Vulcan is red’ does not express a true proposition, even if one did not know whether or not ‘Vulcan’ has a referent. In contrast, if one knew that ‘Vulcan exists’ does not express a true proposition, then one would thereby know that ‘Vulcan’ does not have a referent. This difference, however, does not significantly affect Kripke’s explanation for our judgment that an assertion of the corresponding negative existential, viz. ‘Vulcan does not exist’, is true: We judge that an assertion of ‘Vulcan exists’ would not express a true proposition (nor any proposition at all), and, perhaps only because we are “somewhat sloppy,” we “lump together” the case of expressing a false proposition and the case of not expressing a true proposition (or any proposition at all), and are thus “strongly inclined” to judge such an assertion as false; but to the extent that we are inclined to judge an assertion of ‘Vulcan exists’ as false, we are inclined to judge assertions of its negation, viz. ‘Vulcan does not exist’, as true. Kripke explains that . . . we lump the two cases together [viz. the case of a statement being false, and the case of a statement expressing no proposition at all] and it is our ability to do so which gives the negative existential its use. The negative existential says that there is ³² Thus Kripke is assimilating our judgment that negative existentials are true to the judgments that are problematic for Strawson’s (, ) view of the presuppositions triggered by indexicals, demonstratives, names, and definite descriptions. That is, Kripke is arguing that what explains our, perhaps sloppy, judgment that utterances of ‘Vulcan does not exist’ are true also explains our judgment that utterances of, for example, ‘I had breakfast with the king of France’ are false.

OUP CORRECTED PROOF – FINAL, 17/1/2019, SPi



 ,  ,   

no such true proposition as that [‘Vulcan’] exists in fact, really no such proposition at all as that [‘Vulcan’] exists. (Kripke : )³³

According to Kripke’s proposal then, the reason we judge an utterance of ‘Vulcan does not exist’ to be true is that, as result of our sloppiness, we interpret the negation as a variety of presupposition-cancelling exclusion negation; this interpretation of negation differs from the presupposition-preserving choice negation which is assumed in Referential Compositionality. The choice negation of a sentence that is neither true nor false (due to an empty singular term) is itself neither true nor false; but the exclusion negation of such a sentence is true.³⁴ Thus, if the negation in ‘Vulcan does not exist’ is interpreted as choice negation, then, as was explained in the introduction, relevant utterances of the sentence are predicted to be neither true nor false. But if the negation is interpreted as exclusion negation, then such utterances are predicted to be true. Kripke’s proposal thus goes beyond Donnellan’s rule (R) in that Kripke provides at least the beginning of an explanation as to why in the case of negative existentials involving proper names we are strongly inclined to reject Referential Compositionality: according to Kripke, we judge utterances of ‘Vulcan does not exist’ to be true because we sloppily interpret the negation as exclusion negation. After presenting his proposal Kripke admits that he feels “very tentative about this complicated and messy view” (: ). Perhaps what motivates this hesitancy is that Kripke is proposing an error theory: on Kripke’s view, negative existentials involving empty names are not really true, and our “strong inclination” to judge them true is a consequence of our “lumping together” cases that really should be kept distinct. Other things being equal, a theory that did not ascribe such sloppiness to competent speakers and interpreters would be preferred. Braun () and Salmon () propose solutions to the problem of negative existentials involving grammatical names that are compatible with the direct reference view of names, yet do not attribute widespread error, or sloppiness, to competent speakers and interpreters.

³³ In criticizing Kripke’s account, Katz (: –), Evans (: –), and Salmon () have focused on the final sentence of this passage, and ignored the sloppy process of “lumping together” of which the final sentence is a—perhaps sloppy—summary. On Kripke’s view, assertions of ‘Vulcan does not exist’ do not express propositions at all, so of course he cannot provide an analysis of the propositions expressed by such utterances; Kripke cannot maintain that an utterance of ‘there is no . . . such proposition at all as that [Vulcan] exists’ expresses the same proposition as is expressed by ‘Vulcan does not exist’. ³⁴ The distinction between choice negation (‘∼c’)and exclusion negation (‘∼e’) is represented in the following truth-tables, where ‘*’ indicates a lack of truth-value:

φ ~cφ T F F T * *

φ ~eφ T F F T * T

OUP CORRECTED PROOF – FINAL, 17/1/2019, SPi

 



.. The Existing Fictional Entities view: Salmon (, ) Salmon’s () general strategy is to argue that “on closer examination” the problem of negative existentials “frequently vanishes” (: ). Salmon presents four varieties of negative existentials involving names: (i) negative existentials involving names that refer to entities that literally satisfy ‘does not exist’³⁵; (ii) negative existentials involving names that refer to fictional entities that do not literally satisfy ‘does not exist’, where fictional entities are abstract objects produced by the creation of works of fiction; and (iii) negative existentials involving names that refer to mythical entities that also do not literally satisfy ‘does not exist’, where mythical entities are abstract objects produced by the formulation of mistaken scientific theories; and (iv) negative existentials whose subject terms truly lack referents.³⁶ According to Salmon, the problem “vanishes” in cases (i)–(iii) because in these cases the fundamental premise Reference Failure is false. Since most of the negative existentials considered by philosophers fall into categories (ii) and (iii), our remarks will focus on these cases. Why, according to Salmon, is Reference Failure false in the category (ii) sort of negative existentials? Why, for example, does Salmon maintain that fictional names such as ‘Sherlock Holmes’ actually have referents? Salmon maintains that such fictional names have fictional characters as referents. According to this view, which is known as fictional realism, there really are such things as Hamlet, Sherlock Holmes, and (probably) Santa Claus and Pegasus. Fictional realism of course does not maintain that such entities exist as real people and animals; rather according to fictional realism such entities exist as abstract fictional characters.³⁷ Fictional realists maintain that through the artistic process of creating works of fiction the characters of such works are created, and once created, such fictional entities really exist. As a consequence of granting the full-fledged existence of fictional entities, however, Salmon’s view makes problematic predictions concerning the truth values of both negative and positive existentials. For according to Salmon (and other fictional realists), assertions of, for example, () Sherlock Holmes does not exist.

³⁵ An example of a negative existential that falls into category (i) would be a contemporary assertion of ‘Socrates does not exist’. Salmon (: –) argues that Socrates is the referent of the utterance of ‘Socrates’, though Socrates nonetheless does not (now) exist. Salmon extends this view to account for negative existentials involving reference to future specific—yet non-existing—entities, and even merely possible specific, yet non-existing, entities. ³⁶ Salmon maintains that the negative existentials in category (iv) are “rare . . . and bizarre” (: ), and that they give rise to “what is known as a Headache” (: ). Salmon’s proposed remedy for the headache is discussed in §... ³⁷ Other theorists who endorse versions of fictional realism include Saul Kripke (/); Peter van Inwagen (, ); Edward Zalta (, ); Amie Thomasson (, ); and even David Braun (). Cartwright () provides motivation for—though not an explicit endorsement of—fictional realism.

OUP CORRECTED PROOF – FINAL, 17/1/2019, SPi



 ,  ,   

are false. But these predictions are contrary to the intuitions of ordinary speakers. Salmon and other fictional realists thus face a somewhat peculiar form of the problem of negative existentials.³⁸ In response to this peculiar form of the problem, Salmon maintains that assertions of intuitively true negative existential involving fictional names involve a “nonstandard use” (Salmon : ) of those names. When a name is used in this nonstandard way, it functions “a lá Russell as a disguised improper definite description” (Salmon : ). More specifically, Salmon claims that if ‘Sherlock Holmes’ is used in this nonstandard way, an assertion of () expresses the proposition that the person who is both Holmes and Holmeseque does not exist, and this claim involving a definite description is in turn taken to be equivalent—a lá Russell—to a negated existential quantification: “there is in reality no such person – no such person, no person who is both [Holmes-the-abstract-object] and sufficiently like that (as depicted in fiction)” (Salmon : ). This negated existentially quantified sentence is true because, as Salmon explains, nothing could be both Holmeseque and Holmes, since being Holmeseque requires being a human being (for Holmes is depicted as such in the stories), and Holmes, as an abstract object, cannot be a human being. Does Salmon’s appeal to a conventional, albeit nonstandard, use of fictional names, solve the peculiar problem of negative existentials faced by fictional realists? Everett () points out that though in the case of the fictional name ‘Holmes’ most people have some idea what it would be for a thing to be Holmesesque (because they have some idea as to what is attributed to Sherlock Holmes in the stories), this condition is not met by all intuitively true negative existentials involving fictional names. Everett has us imagine that there is a story involving a fictional character Yugo, but all we know about Yugo is that he, she, or it, is some sort of fictional entity; we have no idea what is attributed to Yugo in the story. In our state of ignorance, we have no conception of what it would be for something to be Yugoesque—even if there is such a property, we have no conception of it. Nonetheless, merely based upon our knowledge that Yugo is some sort of fictional entity, we would judge an assertion of ()

Yugo does not exist.

to be true. Given our ignorance, it cannot be that this judgment is explained by our interpreting the assertion as conveying that the Yugoesque Yugo does not exist.³⁹

³⁸ This problem for fictional realism, though acknowledged in Kripke’s  lecture, was first pressed as objection by Walton (: ). ³⁹ Here is a related concern: Suppose that Yugo is a fictional fictional entity—suppose that in the story Yugo is a fictional entity. Further suppose that the story attributes to Yugo nothing other than being a fictional entity of some sort or other, and thus all we know about Yugo from the story is that Yugo, in the story, is a fictional entity. Thus, for us, to be Yugoesque is simply to be some sort of fictional entity. But then, given that Yugo is a sort of fictional character, there is something that is both Yugoeseque and Yugo. And thus understood in accordance with Salmon’s non-standard use, an assertion of ‘Yugo does not exist’ would be interpreted as false.

OUP CORRECTED PROOF – FINAL, 17/1/2019, SPi

 



Another, perhaps more significant, problem for Salmon’s (, ) approach to the general problem of negative existentials concerns his claim that the solution for negative existentials falling in category (ii) (i.e., those involving fictional names) is “extendable” (: ) to negative existentials falling in category (iii) (i.e. those involving mythical names). Braun (: ) notes that on Salmon’s view of mythical names, if Le Verrier were to assert ‘Vulcan exists’, he would express a truth.⁴⁰ (Le Verrier is the astronomer who incorrectly posited the existence of a planet between Mercury and the sun and dubbed his posit ‘Vulcan’.) This result, Braun observes, is “an indication that Salmon’s theory is incorrect” (Braun : ).⁴¹ Braun agrees with Salmon that mythical entities posited by mistaken theories exist as abstract objects, but Braun persuasively argues that such entities are not the referents of the theorists who mistakenly believe those theories. Braun observes that in using ‘Vulcan’, Le Verrier “certainly wished to speak about a planet, and not about another completely different type of object. He surely thought that if there were no intra-Mercurial planet, then Vulcan would not exist” (Braun : ). Braun concludes that given these referential intentions, “Le Verrier’s utterances of ‘Vulcan’ failed to refer to a mythical planet, or anything else” (: ). Braun’s objection is compelling, but it should be noted that it does not undermine Salmon’s general strategy for solving the problem of negative existentials. For even if Braun is correct that many uses of mythical names must be viewed as lacking referents altogether, Salmon still has a proposal for solving the problem of negative existentials in the case of mythical names, since he has a proposal for solving the problem for the category (iv) cases in which the subject term of a negative existential is truly empty. In other words, the upshot of Braun’s criticism is that Salmon was wrong to think that the category (iii) cases involving mythical names can be assimilated into the category (ii) cases involving fictional names; instead, the category (iii) mythical cases would have to be assimilated into the category (iv) cases in which the subject term truly lacks a referent. Salmon’s treatment of these category (iv) cases is very similar to Braun’s own proposal, which is discussed in the following section.

.. The Gappy Proposition view: Braun (, ) Braun’s proposal is similar to Kripke’s in that Braun also explains our judgment that utterances of, for example, ‘Vulcan does not exist’ are true even though ‘Vulcan’ is an empty genuine singular term. But Braun, unlike Kripke, maintains that this interpretation is not a consequence of our sloppiness; according to Braun, such negative ⁴⁰ In a lengthy footnote Braun explains that on Salmon’s view some of Le Verrier’s initial uses of ‘Vulcan’ failed to refer to anything, though his later uses did succeed in referring to a mythical planet. These subtleties do not affect Braun’s objection. ⁴¹ The extension of fictional realism’s reach has also been challenged by Almog (), and Richard ().

OUP CORRECTED PROOF – FINAL, 17/1/2019, SPi



 ,  ,   

existentials really are true. Braun achieves this result by rejecting Kripke’s claim that sentences such as ‘Vulcan is red’ do not express propositions at all; according to Braun, such sentences express a peculiar sort of false proposition. Braun assumes that propositions are structured entities; the proposition expressed by an (utterance of) a sentence is a structure isomorphic to the syntactic structure of the sentence, and this structure “contains” as constituents the semantic values of the words in the sentence. Since Braun is assuming the direct reference theory of names, Braun assumes that the propositions expressed by utterances of sentences containing names with referents have such referents as constituents; such referent-containing propositions are known as singular propositions. But what about the propositions expressed by sentences containing empty names? According to Braun, “a sentence containing an empty name expresses something, namely an unfilled proposition. A person who utters such a sentence says and believes an unfilled proposition” (: ). Moreover, Braun maintains that “these things people say and believe have truth values” (: ). Braun proposes that the unfilled—or “gappy”—proposition expressed by, for example, ‘Vulcan is red’ can be represented by a set-theoretic entity:⁴² ()

This represents the gappy proposition that has no subject-constituent, and has the property being red as the predicate-constituent. (Note that the brackets ‘{‘ and ’}’ are used to indicate the subject-position of the proposition. Thus the proposition depicted in () is not the proposition that the empty-set is red.) According to Braun (), all such gappy propositions are simply false. This stipulation obviously conflicts with the fundamental premise of Referential Compositionality, since this premise entails that an utterance of, for example, ‘Vulcan is red’ is false only if the utterance of ‘Vulcan’ refers to an entity that does not satisfy ‘is red’. It is by interpreting sentences containing empty names as expressing false propositions that Braun achieves the result that the negations of such sentences are really true. That is, on Braun’s view ‘Vulcan is not red’ expresses the following negated gappy proposition: ()

Since the proposition in the scope of NEG is stipulated to be false, the entire negated proposition is true. And the result that utterances of ‘Vulcan does not exist’ express a true proposition follows in a straightforward manner: the gappy atomic proposition expressed by ‘Vulcan exists’ is—like all gappy propositions—false, and thus its negation, viz., ()

is true. Braun’s gappy proposition view has several theoretical virtues. First, Braun’s view, like Kripke’s no proposition view, is compatible with the direct reference theory. But on ⁴² Braun does not maintain that propositions are such theoretic entities; he claims only that they can be represented by such entities.

OUP CORRECTED PROOF – FINAL, 17/1/2019, SPi

 



Braun’s view, unlike on Kripke’s view, our judgments that utterances of negative existentials are true is not a result of our sloppily interpreting the negation in negative existentials as exclusion negation; on Braun’s view the gappy proposition that is “strictly speaking” said by an utterance of ‘Vulcan does not exist’ is true. And second, as Braun notes, the unfilled proposition view accommodates the intuition that there is something somehow “incomplete” about intuitively true negative existentials, their patent meaningfulness notwithstanding. Such incompleteness is captured by the unfilled nature of the proposition expressed. Braun’s unfilled proposition account of negative existentials is not, however, without its challenges. One particularly counterintuitive consequence of the account, acknowledged by Braun himself, is that there is only one true negative existential proposition: that represented in (). Thus, utterances of the following sentences express the very same gappy proposition: ()

Pegasus does not exist.

() Sherlock Holmes does not exist.⁴³ A perhaps more serious challenge for Braun’s view becomes apparent when one appreciates the similarities between Braun’s gappy proposition view, and the Russellian analysis of negative existentials as negated existential quantifications. On the Russellian analysis, an occurrence of ‘Pegasus does not exist’ as analyzed as the widescope negation of a false existentially quantified proposition; the negation of the false embedded proposition—which does not contain a referent corresponding to the grammatical subject—yields a true proposition. Similarly, on Braun’s view an occurrence of ‘Pegasus does not exist’ is analyzed as the negation of a false gappy proposition; the negation of the false embedded proposition—which also does not contain a referent corresponding to the grammatical subject—yields, again, a true proposition.⁴⁴ Appreciation of similarity between Braun’s unfilled proposition view and the Russellian analysis reveals, however, that the intuitions that Strawson (, ) ⁴³ Braun seeks to explain away this counter-intuitive result by distinguishing between the propositional content of mental states and the mental states themselves: The propositional content of a belief expressed by an utterance containing an empty name is an unfilled proposition. So the beliefs that a person expresses by sincerely uttering ‘Vulcan does not exist’ and ‘Ossian does not exist’ have the same unfilled propositional content. But they may nevertheless be distinct beliefs, for distinct beliefs may have the same unfilled propositional content. (Braun : ) Another response, advocated by Adams and Stecker (), is to appeal to the standard distinction between the proposition “literally expressed” by an assertion and the proposition “meant” by the speaker. ⁴⁴ Given this similarity, it is tempting to extend the unfilled proposition view to cover sentences containing denotationless definite descriptions. On this extended version of the unfilled proposition view, utterances of ‘The winged horse does not exist’ are predicted to be true for the same reason utterances of ‘Pegasus does not exist’ are predicted to be true. And the unfilled proposition view of negative existentials containing definite descriptions would seem to be superior to the Russellian analysis of such sentences, since the extended unfilled proposition view would not predict an unwanted structural ambiguity.

OUP CORRECTED PROOF – FINAL, 17/1/2019, SPi



 ,  ,   

invokes against Russell’s analysis can be invoked with equal force against Braun’s () analysis. The problem is that for almost all predicates Π, utterances of ┌ Pegasus Π ┐ are not judged to be either true or false, but rather are judged to be neither true nor false.⁴⁵ But Braun’s unfilled proposition view, like Russell’s quantificational analysis, predicts that all such assertions will be either true or false. So, for example, the unfilled proposition view incorrectly predicts that assertions of, for example, ‘Pegasus is sleeping’ are false, and that assertions of ‘Pegasus is not sleeping’ are true.⁴⁶ In a more recent paper (), Braun presents two potential responses to this criticism. First, Braun claims the stipulation that gappy atomic propositions are false is not essential to his view; Braun claims that even “admitting that sentences that express gappy propositions are neither true nor false would not force a Gappy Proposition theorist to say that all negations of those sentences are neither true nor false” (Braun : ). Braun claims that the literal truth of such negative existentials could be achieved by claiming that negation is ambiguous between a presuppositioncancelling exclusion negation, and a presupposition-preserving choice negation. The idea is that utterances of, for example, ‘Vulcan is not red’ are interpreted as neither true nor false because the negation in such utterances is interpreted as choice negation, and utterances of, for example, ‘Vulcan does not exist’ are interpreted as true because in such utterances the negation is interpreted as exclusion negation. (Invoking this sort of ambiguity of negation is also how Salmon (: ) proposes to explain the intuitive truth of category (iv) negative existentials, i.e., negative existentials whose subject terms refer to nothing at all.) The second option Braun considers is to maintain that utterances of, for example, ‘Vulcan does not exist’ really are neither true nor false, but our (sloppy) judgment that they are true is readily explained: we judge that they are true because we believe the (gappy) proposition expressed by such utterances, viz. the gappy proposition () depicted above. That is, even if it is granted that gappy propositions are neither true nor false, gappy propositions can still be the objects of belief, and our strong inclination to assert, for example, ‘Vulcan does not exist’ can be explained by the fact that, despite its non-truth, we believe the gappy proposition that this negative existential expresses. ⁴⁵ The notable exceptions are of course the predicate ‘exists’ and its synonyms. ⁴⁶ Salmon objects to Braun’s () stipulation that gappy propositions are false on the grounds that this stipulation “illegitimately makes the problem too easy” (Salmon : ). In response to this charge, Braun () argues that gappy propositions are false: . . . some things are untrue, and yet are not false, for instance Piccadilly Circus and the Eiffel Tower. The most salient difference between untrue objects that are false and untrue objects that are not false is that the former are propositions (or items that semantically express propositions). Thus, it is reasonable to conclude that this is the crucial difference between being untrue and being false: untrue objects that are also propositions (or things that express propositions) are false. Atomic gappy propositions are propositions and are untrue. Therefore, they are false. (Braun : ) The objection we are raising, however, is not that Braun has no reason to suppose that gappy propositions are false; rather the objection we are raising is that this claim—justified or not—has very counterintuitive consequences. This problem is considered by Everett () and Caplan ().

OUP CORRECTED PROOF – FINAL, 17/1/2019, SPi

 



These are interesting proposals, and they cannot be given the consideration they deserve here; but two brief remarks are in order. The first proposal is to construe negation as semantically ambiguous between choice and exclusion negation. This proposal secures the result that under one interpretation negative existentials are literally true, but it also incorrectly predicts that there is an interpretation under which intuitively true negative existentials are neither true nor false. Similarly, the semantic ambiguity proposal predicts that under one interpretation an utterance of, for example, ‘Vulcan is not red’ expresses a (gappy) proposition that is neither true nor false, while under the other interpretation this utterance is predicted to express a proposition that is true. The biggest problem with the ambiguity proposal is that, as a matter of empirical fact, natural language negation does not seem to be semantically ambiguous in this way.⁴⁷ The second proposal represents a return to something like Kripke’s error theory. Recall that on Kripke’s no-proposition view, an assertion of, for example, ‘Vulcan does not exist’ expresses no proposition, and thus is neither true nor false, yet according to Kripke we interpret it as true as a result of a sort of sloppiness. Braun’s second proposal is also an error theory: according to Braun’s second proposal, utterances of ‘Vulcan does not exist’ express gappy propositions that are really neither true nor false, though we are strongly inclined to judge such assertions true because we hold the attitude of belief toward the gappy propositions that are expressed. The one difference between these error theories is that where Kripke has no propositions, Braun has gappy and truth-valueless propositions. One might well wonder if Kripke’s no-proposition error theory would be preferable to Braun’s gappyproposition error theory merely on the grounds of parsimony.⁴⁸

. T P M D : K (, )

.................................................................................................................................. Difficulties in arriving at a satisfactory analysis of negative existentials from the perspective of direct reference theory might prompt one to reconsider quantificational analyses in the spirit of Russell and Quine. The question is whether such analyses can be embraced in light of Kripke’s persuasive arguments against the descriptive theories of Frege and Russell. Jerrold Katz (, ) proposes a metalinguistic descriptive theory of the meanings of proper names, and he defends his theory from Kripke’s ⁴⁷ According to Gazdar (: –), no natural language contains two distinct lexical items corresponding to choice and exclusion negation. And see Horn (: –) for extensive criticisms against the semantic ambiguity of negation. ⁴⁸ The argument from parsimony gains strength when one considers that both Kripke and Braun will have to rely on something like Braun’s belief states. It would seem that our inclination to assert ‘Vulcan does not exist’ could be adequately explained by the corresponding belief state alone; there would seem to be no further explanatory work for the alleged gappy truth-valueless proposition to do.

OUP CORRECTED PROOF – FINAL, 17/1/2019, SPi



 ,  ,   

(/) arguments against the classic descriptive view of names. Katz also claims that his theory solves the problem of negative existentials for cases involving empty names.⁴⁹ According to Katz (: ) a name ‘N’ has a sense that is characterized by the definite description, “the thing which is a bearer of ‘N’”. Katz dubs this theory the “pure metalinguistic description theory” (PMT); Katz categorizes such a description as “pure” because it involves “no real property or relation in addition to the purely nominal bearer relation that makes the sense metalinguistic” (Katz : –). As Katz notes, this “purity” undermines the ability of these descriptions to fix the referent of the name; for example, the pure metalinguistic description “the thing that is a bearer of ‘Gödel’” clearly fails to determine the referent of a use of ‘Gödel’, since every member of the Gödel family is a bearer of ‘Gödel’. (Indeed, every member of every Gödel family is a bearer of ‘Gödel’.) Thus, according to PMT the descriptions associated with proper names are woefully incomplete—they obviously fail to denote a unique entity—and thus they fail to fix the reference of the name with which they are associated. Katz, however, does not think that this is a weakness of PMT. Indeed, Katz maintains traditional descriptivism conflates issues concerning the sense (or meaning) of words in a language and pragmatic issues concerning the use of words to refer to entities. According to Katz, the fundamental error of traditional descriptivism was its assumption that the descriptions associated with names must both give the meaning (or sense) of a name and also fix the referent of the name. In contrast, Katz maintains that PMT correctly “locates information about the orthographic character of proper nouns and about their tag function in their sense, and, at the same time, relocates the determination of reference . . . in the pragmatics of language use” (Katz : ). Because PMT denies that the pure metalinguistic descriptions associated with proper names fix the referent of names, PMT is immune to Kripke’s semantic argument. Kripke pointed out that because classical descriptivism makes the referent of a name ‘N’ depend upon whatever thing it is that satisfies the (probably contingent) properties associated with ‘N’, the view implies that ‘N’ could turn out to refer to something wholly unexpected. Recall Kripke’s famous Gödel-Schmidt case: if the description associated with ‘Gödel’ is “the discoverer of the incompleteness proof” and it turned out that poor Schmidt is the discoverer of the incompleteness proof, then it would turn out, implausibly, that in using ‘Gödel’ we were really referring to Schmidt. Katz’s view is immune to this sort of problem. According to Katz, a relevant occurrence of ‘Gödel’ is associated with “the thing that is a bearer of ‘Gödel’”, but this description is incomplete and does not determine a referent at all. Determining the referent of an occurrence of ‘Gödel’ “requires further extralinguistic information to know which of the bearers is the referent in the context” (: ). On Katz’s pragmatic view of reference, relevant occurrences of ‘Gödel’ refer to Gödel because he is the particular ⁴⁹ Kent Bach (a, ) defends a similar metalinguistic description view. We here focus on Katz’s theory because, though Bach () presents a more detailed response to Kripke’s (/) arguments, Katz () is more explicit with regard to the problem of negative existentials.

OUP CORRECTED PROOF – FINAL, 17/1/2019, SPi

 



bearer of ‘Gödel’ the speaker intends to speak about in the context, and thus the speaker refers to him regardless of Schmidt’s intellectual achievements. How does PMT purport to solve the problem of negative existentials? Katz (: ) states that the proposition conveyed by a typical use of a negative existential of the form ‘N does not exist’ is provided by the following schema: (C) Nothing is the thing with the property P that is a bearer of ‘N’. Katz explains that “‘P’ marks the place at which descriptive information may be introduced . . . to make the reference of tokens of ‘N’ contextually definite by distinguishing their intended bearer from the other known bearers” (: ). Thus, according to Katz, a typical use of ‘Santa Claus does not exist’ is “an assertion about the things in the world, to the effect that none of them is the such-n-such thing that is a bearer of ‘Santa Claus’” (: ). The proposal then is to follow Russell in viewing negative existentials involving names as expressing negated existential quantifications. Thus, as with Russell’s proposed solution, the metalinguistic description view of Katz proposes solving the problem by rejecting the premise we have dubbed Genuine Singular Term. Does PMT succeed in solving the problem? Again, there are reasons to be skeptical. First, the proposal seems to assume that Russell’s analysis of both names and definite descriptions as existential quantifier phrases is adequate, and in §. we noted that this claim is dubious for several reasons. Thus, for example, PMT seems to predict that an utterance of ‘Obama does not smoke’ is structurally ambiguous between several readings, one of which is the general proposition to the effect that none of the things in the world is the such-n-such that bears ‘Obama’. This proposition would be true in the circumstance wherein Obama smokes, yet neither he nor anybody else bears ‘Obama’. Similarly, PMT seems to predict that there is an interpretation of a use of ‘Santa Claus does not exist’ under which it expresses the contradictory proposition that there is something that is such-n-such, bears ‘Santa Claus’, and does not exist.⁵⁰ These predictions of structural ambiguity are as problematic for PMT as they are for Russell’s descriptive view. Finally, despite Katz’s insistence to the contrary, the proposal that proper names are semantically equivalent to pure metalinguistic descriptions is prima facie implausible. To be sure, it cannot be doubted that speakers at least sometimes use proper names to convey propositions involving metalinguistic descriptions. But what is less plausible is the claim that such metalinguistic descriptions are the literal meanings of proper names. Kripke (: ) observed that though in using names speakers do convey information about those names, and not merely information about the referents of names, the same can be said about any other word. For example, if you do not know what ‘monkey wrench’ means, I can point at one and utter ‘that’s a monkey wrench’, ⁵⁰ Note that in Katz’s schema (C), the verb ‘exists’ is nowhere represented. Katz provides no explanation for its absence in his analysis, but it is reminiscent of Russell’s (, ) similarly implausible analysis.

OUP CORRECTED PROOF – FINAL, 17/1/2019, SPi



 ,  ,   

thereby conveying that that thing is a bearer of ‘monkey wrench’.⁵¹ But this is not a good reason for thinking that the literal meaning of ‘monkey wrench’ is metalinguistic. The much more plausible view is that the metalinguistic propositions that this word can be used to convey are speakers’ meanings, and not sentence meanings. And similar remarks apply in the case of proper names.

. C:   

.................................................................................................................................. Despite over a hundred years of debate, contemporary analytic philosophy has yet to achieve anything approaching a consensus on how to solve the problem of negative existentials. Upon reflection, it should come as no surprise that philosophers of language have yet to agree upon a solution. The problem concerns how a particular sort of sentence—a singular negative existential—can be used to make a true and informative assertion, but, as of yet, there is no consensus as to how any sentence is used to make true and informative assertions. It is not as if every other phenomenon in the realm of linguistic communication has been adequately explained and the problem of negative existentials is the sole recalcitrant unsolved problem. Rather, the problem of negative existentials is a particularly puzzling phenomenon among a sea of unresolved issues and poorly understood phenomena. Seen from this perspective, it seems likely that the problem of negative existentials cannot be solved independently of providing an adequate theory of verbal communication as a whole. (And this is to say nothing of the related logical and metaphysical issues.) We will therefore refrain from concluding this entry by proposing a solution. In the way of a sort of reactive summary, however, we will offer some brief, and, at best, programmatic remarks.⁵² First, we are inclined to endorse Salmon’s observation that some negative existentials—in particular those pertaining to fictional characters—are not as puzzling as others; fictional realism provides a plausible motivation for rejecting the premise of Reference Failure for many assertions of singular negative existentials. And in cases of successful reference to fictional entities, corresponding reinterpretation of the predicate ‘does not (really) exist’ does not seem ruinous.⁵³ But, we agree with Kripke and Braun that this strategy cannot be extended to cover all puzzling assertions of negative existentials. For the remaining cases we suggest that the argument presented in the introduction is, when interpreted appropriately, sound. There is a sense in which, taken ⁵¹ In a suitable context one could also use ‘Monkey wrenches might not have been monkey wrenches’ to convey the proposition that monkey wrenches might not have borne ‘monkey wrench’. ⁵² Our reactive remarks here are broadly reflective of our own published views concerning the problem of negative existentials. (See Reimer ; Clapp ). ⁵³ As Evans (: –) notes, the modifier ‘really’ seems to affect the truth, or at least the naturalness, of the relevant negative existentials. For example, ‘Sherlock Holmes does not exist; he’s a fictional character’ sounds contradictory, or at least very odd, whereas ‘Sherlock Holmes does not really exist; he’s merely a fictional character’ seems to be unproblematic and true.

OUP CORRECTED PROOF – FINAL, 17/1/2019, SPi

 



literally, utterances of such puzzling singular negative existentials cannot be true. We suggest, moreover, that it is the soundness of the argument that explains the indisputable fact that there is something deeply puzzling about such intuitively true assertions of negative existentials; indeed, it is a mark against Russell’s and Meinong’s proposals that they do not provide an explanation of why intuitively true assertions of singular negative existentials are so puzzling. For these reasons we tend to think Kripke (/ ) is at least getting closer to the mark in claiming that intuitively true negative existentials do not literally express propositions. But, we are optimistic that this seemingly counterintuitive result can be rendered palatable by appeal to a pragmatic perspective—which incorporates Katz’s pragmatic conception of reference—with regard to our intuitive judgments concerning what is said by assertions. From this pragmatic perspective what needs to be explained is not how an utterance of, for example, ‘Vulcan does not exist’ manages to convey a proposition that is true. Rather, from this pragmatically enlightened perspective what needs to be explained is how a speaker can manage to use this sentence to make an assertion that interpreters will accept as informative and true. Finally, we suggest that providing such an explanation will probably require a detailed account of the illocutionary purpose of such assertions. Several theorists note in passing that an intuitively true assertion of a negative existential involving a singular term is typically understood as denying and correcting previous infelicitous assertions involving that very singular term.⁵⁴ We suggest that understanding the nature of such denials and corrections is probably required for adequately solving the problem.⁵⁵

⁵⁴ For example, the necessity of such previous assertions is explicitly noted in Donnellan’s rule (R) (: ). Similarly, Kripke observes that it is “legitimate” to use a negative existential involving the empty phrase ‘the bandersnatch’ in response to “a child who asked to be taken to see a bandersnatch in the zoo” (/: ). And Thomasson (: ) and Salmon (: ) build anaphora to a previous assertion into their accounts. ⁵⁵ We are grateful to David Braun, Michael Nelson, Nathan Salmon, and Barbara Abbott for helpful comments and suggestions.

OUP CORRECTED PROOF – FINAL, 21/1/2019, SPi

  ......................................................................................................................

A TAXONOMY OF USES OF DEMONSTRATIVES ......................................................................................................................

 .    

W previous semantic and pragmatic accounts of demonstrative NPs have identified a wide range of uses, there is still no comprehensive account to date that includes all of their felicitous uses. Such a unified account is not on offer here; rather our goal in this chapter is to offer a preliminary taxonomy of the various uses of demonstratives that have been discussed in the linguistic and philosophical literature and to provide a basis for classifying them. Following Kaplan (a), the philosophical literature has focused on examples in which a demonstrative expression is used by a speaker to refer to something in the immediate perceptual surroundings. This focus has led to the term ‘demonstrative’ being used almost exclusively in the philosophical literature for these referential uses of demonstrative NPs. In contrast, the linguistics literature has also used the term to include other referential uses (see Wolter  for an overview). More recent studies, however, have identified additional uses of demonstratives in which the demonstrative NP is not interpreted as referring to any particular entity at all (King ; Roberts ; Elbourne , inter alia). Our focus in this chapter will be on demonstrative uses in English, but many of these uses are also found in other languages. We take as our starting point the full range of the various uses of the English demonstratives this, that, these, and those, whether used individually as pronominal NPs or as determiners of lexical NPs.¹ In doing so, we remain theoretically neutral as to which examples of demonstratives, if any, should be considered paradigmatic for any particular use. Our goal here is to bring together the full range of uses of

¹ As our starting point is demonstrative NPs, we will not be considering the degree modifier use of this and that, as illustrated in (i): (i)

I won’t date anyone that short. [corpus]

See Maclaran  for a discussion of demonstrative degree modifiers.

OUP CORRECTED PROOF – FINAL, 21/1/2019, SPi

     



demonstratives discussed in the literature—including some uses that have not been previously distinguished—and to provide a classification that will assist in the identification of common features that emerge from an examination of the various ways in which speakers use demonstrative NPs. Our taxonomy—consisting of thirteen broadly defined uses of demonstratives—begins by distinguishing uses based on the semantic value assigned to a demonstrative NP, that is, whether it is used to refer to a discourse entity, or whether it is used generically, predicatively, or quantificationally. Subsequent classifications within these broader categories are made on the basis of various pragmatic factors that affect the interpretation of the demonstrative.

. R   

.................................................................................................................................. The first category of demonstrative uses that we will consider involve reference to a particular discourse referent.² Included in this category are deictic uses, anaphoric uses, uses that rely upon mutual knowledge, and the so-called indefinite-this use. We take up each of these in turn.

.. Deictic uses The most frequently discussed uses of demonstratives are demonstratives used deictically, whereby the referent of the demonstrative is a particular entity within the immediate spatio-temporal context of utterance. In such uses both the pronominal forms and full lexical NPs are felicitous, as are the proximal and distal forms, given an appropriate context, as shown in () with the contexts indicated: ()

a. This/that is delicious. [pointing at a bowl of ice cream]³ b. These/those students are late. [pointing at a group of students]

Often these uses are accompanied by ostensive gestures made by the speaker, thus warranting their classification as deictic. Of course, a gesture made by the speaker is not always required, as the referent of the expression may already be sufficiently prominent in the context for the hearer to be able to identify the relevant entity. The precise role of demonstrations is a matter of debate and one that we will not address here (see e.g., McGinn ; Kaplan b). What seems clear, however, is that for felicitously used deictic demonstratives, the hearer must be in a position to distinguish the intended referent with respect to the innumerable entities present in a given context. The ² These uses are commonly called ‘singular referring’ in the philosophical literature. ³ Here, and in all subsequent examples, we underline the demonstrative NP in question.

OUP CORRECTED PROOF – FINAL, 21/1/2019, SPi



 .    

speaker’s choice of demonstrative form indicating the number (singular or plural) and relative proximity of the intended referent(s), the descriptive information encoded in a lexical demonstrative NP, plus any other relevant contextual cues, all serve to provide the hearer with information sufficient to identify the referent.⁴ In the canonical examples of deictic uses, the referent of the demonstrative NP is an ordinary physical object in the perceptual surroundings. But we also include within the category of deictic uses reference to properties, sounds, events, facts, and other abstract entities, including the discourse itself, as shown in (): () a. b. c. d. e.

That’s a really nice shade of blue. Did you hear that? And that’s a fact! You are wrong—that’s exactly what she said. [= Levinson : ex. ()] It sounded like this: whoosh. [= Levinson : ex. ()]

In each of the examples in (), the pronominal forms are used (although (a) and (b) would be felicitous with lexical NPs as well) and in each of these examples the demonstrative NP receives an interpretation in which the speaker is referring to a particular entity which is identified with respect to the current context of utterance. For example, in (d), the referent is the specific claim or utterance evoked by the first sentence, and it is the utterance itself which is the referent for the demonstrative. As noted by Levinson (), the linear nature of speech precludes reference to the utterance itself as it is being spoken; thus the speaker must refer to the previous utterance, as in (d), or to the subsequent utterance, as in (e). Interestingly, it has been observed that, despite being used to refer to various different types of entities, English demonstrative pronouns generally cannot be used to refer to people (Maclaran ). When demonstrative pronouns are used to refer to people in English, they convey an ‘objectifying’ or ‘dehumanizing’ stance on the part of the speaker, as if the person referenced were being treated as an inanimate object, as illustrated in (): () [bar patrons commenting on server] a. That could serve me anytime! b. I’d definitely ask that out. The notable exception to this generalization is when demonstrative pronouns are used with identificational post-copular clauses (Higgins ), as in (): () a. That/this is my brother. b. Those/these are my students. While the semantic and syntactic properties of such sentence-types have been the subject of considerable debate (Heycock and Kroch ; Mikkelsen ; Heller and ⁴ For discussion of the interpretation of deictically used demonstratives see e.g., Diessel , ; and Levinson .

OUP CORRECTED PROOF – FINAL, 21/1/2019, SPi

     



Wolter ; Lee-Goldman , inter alia; see also the discussion in Ward ), the exceptional status of these uses remains to be fully explained. It has also been claimed that deictic uses of referring expressions are somehow fundamental or ontogenetically prior to other uses (Lyons ). However, Levinson (, ) has compellingly argued that the deictic system cannot be privileged in this way. As he notes, “The deictic system in language is embedded in a contextindependent descriptive system in such a way that the two systems produce a third that is not reducible to either” (: ). Thus, we follow Levinson in not considering deictic uses of demonstratives to be any more basic or fundamental than the other uses that we survey in this chapter.

.. Anaphoric uses Related to the deictic uses of demonstrative NPs is the anaphoric use of a demonstrative to refer to a discourse referent introduced in previous discourse, as shown in the singular (a) and in the plural (b), felicitous with both proximal and distal forms: () a. The cowboy entered. This/that man was not someone to mess with. [= Levinson : ex. ()] b. Three years ago, she took in four or five feral cats, and those cats had litters. [corpus⁵] Here, the demonstrative NPs are being used to refer to entities that had been previously evoked, with the referent for the demonstratives in (a) evoked by the antecedent expression the cowboy, and the referents for the demonstratives in (b) evoked by four or five feral cats. The prior mention of the cowboy and cats, respectively, introduces these entities into the discourse thereby rendering these entities available for the speaker to subsequently refer back to them using demonstratives. In both examples, however, the speaker or hearer may have no other means of identifying the referent for the demonstrative aside from the descriptive content supplied in the antecedent expression; for example, in (a) the speaker may not know the name of the cowboy—only that he is not someone to mess with. What distinguishes these uses from deictic ones is that the demonstrative NP is interpreted as referring to a previously evoked entity. In the anaphoric use of demonstratives, pronominal forms are also possible, as shown in the examples in ():⁶

⁵ This, and all subsequent tokens marked ‘corpus’, were drawn from our corpus of naturallyoccurring data collected between May  and February . ⁶ While demonstrative pronouns can be used anaphorically (as in ()), the use of pronominal demonstratives is generally more contextually restricted than the use of the corresponding full NP forms; see Gundel et al.  inter alia for discussion.

OUP CORRECTED PROOF – FINAL, 21/1/2019, SPi



 .    

() a. There are some really nice strawberries in the refrigerator but I was saving those for dessert. b. John insulted the ambassador. That happened at noon. [= Gundel et al. : ex. ()] c. Multiply  times , and then divide that by . [= Webber et al. : ex. ()] In (a), the referent of the plural demonstrative pronoun is explicitly evoked by the prior mention of the strawberries. In (b), the referent of the pronoun is the event of John insulting the ambassador, evoked in the previous sentence. The referent for anaphoric uses can also be implied by previous discourse, rather than being explicitly evoked, as illustrated in (c) in which the referent of the demonstrative is the inferred product of  and .

.. Uses based on private shared knowledge While deictic and anaphoric uses of demonstratives depend on the immediate spatiotemporal context or prior discourse for their interpretation, there is another use in which the interpretation of the demonstrative depends instead on a background of (presumably shared) ‘mutual knowledge’ or ‘common ground’. The relevant type of mutual knowledge on which this use depends is ‘private’ in the sense that it is assumed to be shared by speaker and hearer on the basis of their shared experience rather than more general knowledge about the world that could be assumed to be known to virtually anyone in the speech community (Joshi ). Others have termed this use ‘familiar that’ (Partee ), ‘reminder that’ (Gundel et al. ), and ‘recognitional’ (Levinson ). In () below, the demonstrative NPs are being used to refer to the relevant entities on the basis of the speaker’s assumed private shared knowledge of the referents in question. () a. I couldn’t sleep last night. That dog (next door) kept me awake. [= Gundel et al. : ex. ()] b. I couldn’t sleep last night. Those dogs kept me awake. In uses that rely upon such private shared knowledge, the speaker assumes (a) that the referent in question is among the stock of entities that are familiar to this particular speaker–hearer dyad, rather than being something that is simply generally known or inferable on the basis of broader world knowledge, and (b) that this assumption is reciprocally shared with the hearer. Thus, for the demonstrative in (a) to be felicitously used, there must be one particular dog that the speaker assumes that the hearer will uniquely identify and, moreover, that the information provided in the description of the demonstrative NP will be sufficient in the context for the hearer to identify this presumably mutually known dog.

OUP CORRECTED PROOF – FINAL, 21/1/2019, SPi

     



In the private shared knowledge uses, however, demonstrative pronouns are not felicitous, as the hearer requires some descriptive content in order to pick out the intended referent from among the entire set of entities that are mutually known, as can be seen by comparing (b) with (c). ()

c. I couldn’t sleep last night. #Those kept me awake.

Here, the demonstrative pronoun is infelicitous given that it—like pronouns in general—requires that the referent be salient but the context in (c) provides no basis for the requisite salience of the referent. However, contrary to what is suggested by the use of such terms as ‘familiar that’ found in the literature to refer to these uses, demonstratives that are interpreted on the basis of private shared knowledge are not restricted to the distal form, as shown in (): ()

. . . Look I know how upset you are about what happened to this boy. [corpus]

In the context of (), an episode of Breaking Bad, there is no prior mention of the relevant boy in the current episode, nor is he contextually present (in fact the relevant boy had not been mentioned for several episodes). What is essential to the private shared knowledge use of demonstratives (with either the distal or proximal form) is that the speaker assumes that there is a mutually-shared—and dyad-specific—basis for identifying the referent in question. Two further observations are warranted with respect to the private shared knowledge use. First, in this use the demonstrative must be deaccented. In contrast to a demonstrative used deictically, a demonstrative use that relies upon private shared knowledge cannot felicitously bear a prosodic pitch accent conveying contrast while at the same time retaining the intended interpretation (cf. #THOSE dogs kept me awake). Second, the private shared knowledge use is an exception to the generalization that definite NPs are infelicitous in existential sentences (see the discussion in Ward and Birner ). Consider (): ()

a. A: I think I’ll take a few days off. B: But there’s that project deadline looming over our heads. b. A: Too bad nobody showed up at the city council meeting to argue for the zoning change. B: There were those neighbors at the City Council meeting yesterday. [adapted from Ward and Birner : ex. ()]

In (a), the speaker is referring to a deadline that is assumed to be mutually known by the hearer, and likewise in (b) the speaker assumes that the relevant set of neighbors is mutually known to the particular hearer (by some means or other). However, in existential sentences the post-verbal NP generally represents information that is not known to the hearer (Ward and Birner ). Thus, the private shared knowledge use of demonstratives is one example of definite NPs that are felicitous in (non-list readings of) existential sentences. As Ward and Birner () note, definite NPs in an existential

OUP CORRECTED PROOF – FINAL, 21/1/2019, SPi



 .    

sentence are felicitous just in case the post-verbal NP represents a hearer-new entity (in the sense of Prince ), that is, an entity believed by the speaker to be not already known to the hearer. The private shared knowledge uses of demonstratives are felicitous with the existential, despite representing an entity that is hearer-old, because they serve a reminder function, in which the entity is being treated as hearer-new. As Ward and Birner explain: “use of the demonstrative signals that the speaker is able to recognize the entity in question, while the use of the there-sentence signals that the hearer is not expected to recall it” (Ward and Birner : ). That is, use of the demonstrative in an existential there-sentence indicates that the referent is familiar while at the same time that it is not currently being attended to by the hearer.⁷

.. Inferable uses While the private shared knowledge use depends upon a speaker’s assumed mutual knowledge about entities familiar to a particular speaker–hearer dyad, demonstratives may also be used for referents that are inferred on the basis of more general knowledge, as in (): ()

a. I’m getting acquainted with quite a few of the band guys; they’re all the best kind of people. Already know six of the other trombonists quite well, especially one guy from Springfield, Ky. You should hear that drawl. [corpus] b. I was told that when I conceive again I will need to take an extra  mg. folic acid daily along with  mg. of aspirin. Does anyone feel like they should take this prior to conceiving? Or wait till you see those two double lines? [corpus]

In (a), the identification of the referent for the singular demonstrative is based on the prior mention of the trombonist from Kentucky, which licences the plausible inference that he has a drawl. Similarly, in (b) with the plural form, the referent of the demonstrative is inferable given that the discussion is taken from a fertility website. The inferences here are ‘bridging inferences’ (Clark and Haviland ) by which the referent may be inferred via an association with an element in the previous discourse. However, not all inferable uses of the demonstrative are equally felicitous. In (a), an inference is available that licenses the use of the definite; whereas in (b) the same inference does not support the use of the demonstrative. This particular example, however, becomes felicitous when the demonstrative is used contrastively, as in (c). ()

a. A car drove by. The horn was honking. [= Wolter : ex. ()] b. A car drove by. #That horn was honking.

⁷ In addition to the private shared knowledge use, there are two other uses of demonstratives that are acceptable in existential sentences, i.e., the indefinite this use and the restrictive that use; see below.

OUP CORRECTED PROOF – FINAL, 21/1/2019, SPi

     



c. A car drove by. The horn was honking. Then another car drove by. That horn was honking even louder. [= Wolter : ex. ()] Thus, in addition to private shared knowledge about particular referents, some inferences drawn on the basis of world knowledge can provide the basis for the identification of the referent of a demonstrative expression while others cannot.⁸ We are currently investigating the precise conditions that license the felicitous use of inferable demonstratives (Doran and Ward in prep.).

.. Indefinite-this use In each of the uses of demonstratives surveyed thus far, both the proximal and distal forms can be felicitous, given an appropriate context. In contrast, the final use of demonstratives referring to particular discourse entities that we consider is felicitous only with the proximal form. This use has been variously termed ‘presentational this’ or ‘indefinite this’ (Maclaran , ; Prince a) and is illustrated in () with both the singular and plural forms: ()

a. There was this funny rattle under the hood. [= Maclaran : ex. (a)] b. I’m afraid I can’t come as I’m expecting these friends to call later. [= Maclaran : ex. (b)]

In the examples in (), on the intended interpretations, the speaker refers to an entity that is new to the discourse and is otherwise unfamiliar to the hearer. For example, it is assumed by the speaker of (a) that the hearer is unfamiliar with the referent for this funny rattle. Likewise, in (b) with the indefinite this use, the relevant friends are assumed to be unknown to the hearer. The indefinite use of this is crucially distinct from the other entity-referring uses of the demonstrative because its felicitous use does not require that the hearer is able to identify the speaker’s intended referent. Interestingly, this use is possible only with the proximal form of the demonstrative; if the proximal form is changed in (b) to those, the speaker is assuming that the relevant friends are familiar to the hearer and that they can be identified on the basis of private shared knowledge. Furthermore, only lexical NPs are felicitous with the indefinite-this interpretation (cf. #There was this under the hood), which we saw above is also the case with the private shared knowledge use. The discussion of the use of indefinite this in Prince a identifies a number of further features showing how it is distinct from the use of the indefinite article. First, in the indefinite this use, the demonstrative can be followed by numerals. () John wants to marry this one Norwegian.

⁸ See Wolter  for further discussion of demonstratives involving bridging inferences.

OUP CORRECTED PROOF – FINAL, 21/1/2019, SPi



 .    

The corresponding sentence with the indefinite article in place of the demonstrative would be not only infelicitous but downright ungrammatical. Second, () is unambiguously specific (with or without the numeral one); that is, the speaker is referring to a specific Norwegian and not asserting that John wants to marry some Norwegian or other. The use of the indefinite article, rather than this one in (), would be ambiguous between the specific and non-specific interpretations. These two observations indicate that indefinite this is not interpreted the same as the indefinite article, as the nonspecific interpretation is simply unavailable for indefinite this. Like uses of indefinite NPs, the indefinite this use also serves to introduce novel and unfamiliar entities into the discourse; however, it also functions as a topic marker (Prince a), indicating that the newly introduced discourse entity may be the subject of subsequent discourse.⁹ In this way, the indefinite use of this serves an information status function, indicating the status of the referent with respect to the inferred knowledge store of the discourse participants.

. R      

.................................................................................................................................. While the uses of demonstratives that make reference to a discourse entity have received the most attention in the literature, there are also a number of other uses of demonstratives that involve generics and reference to kinds. While generic NPs with bare plurals or the (in)definite article in English have been widely discussed, examples of generic demonstrative uses have been less well investigated. However, demonstratives can be felicitously used both in referring to abstract kinds and in making assertions about what is usual, typical, or expected—by means of so-called ‘habitual’ or ‘characterizing’ sentences (Krifka et al. ). In this section, we turn to examine four distinct generic uses of demonstratives.

.. Taxonomic and non-taxonomic kind-referring uses The use of demonstrative NPs for reference to kinds is discussed by Bowdle and Ward (), who discuss the following examples: ⁹ Further comment on the newly introduced discourse entity can appear as subsequent discourse or as elaboration on the referent. Compare: (i) A: Where did you get that beautiful scarf? B: I bought it at this store I went to yesterday. (ii) A: Where did you get that beautiful scarf? B: #I bought it at this store.

OUP CORRECTED PROOF – FINAL, 21/1/2019, SPi

      ()



a. Those Labradors are extremely loyal, you know. [= Bowdle and Ward : ex. ()] b. This IBM ThinkPad is amazing! (uttered in front of a computer) [= Bowdle and Ward : ex. (b)]

These examples show three things: first, that demonstrative NPs with a kind-referring interpretation are felicitous in both the singular and the plural; second, that such uses of demonstratives are felicitous with both the proximal and distal forms; and, third, that these demonstratives are felicitous with both object- and kind-level predication. In (a), the speaker asserts that it is generally true of individual Labrador dogs that they are loyal. In (b), on the intended interpretation, it is the kind itself, not an individual instance of the kind IBM ThinkPad, that the speaker is claiming to be amazing. In (), the demonstrative NPs are interpreted non-taxonomically. That is, they are interpreted as referring to the kinds denoted directly by the noun of the NP (Labradors in (a)). However, generic uses of demonstratives can also receive a ‘taxonomic’ interpretation in which the speaker is referring to a kind that bears a subtype–type relationship to the kind denoted by the noun of the NP, as shown in (): ()

a. These Spaniels make great pets. b. That ice cream is delicious.

On the taxonomic interpretation, the speaker of (a) is referring to a subtype of Spaniel, that is, some particular breed of spaniel (e.g., English Springer Spaniel, Cocker Spaniel, etc.).¹⁰ This constitutes a taxonomic interpretation in that the particular breed is a subtype of the more general type (SPANIEL). Similarly with (b), on the intended interpretation, the speaker is referring to a particular subtype of ice cream (e.g., Ben and Jerry’s), rather than the more general type (ICE CREAM). It is noteworthy as well that the examples in () allow a contrastive pitch accent on the demonstrative (as would be expected with the taxonomic interpretation), but such stress is not felicitous for the examples in () on the intended (non-taxonomic) interpretations (cf. #THOSE Labradors are extremely loyal, you know). One way to induce the generic interpretation of a demonstrative NP is through the presence of an instance of the relevant kind in context—known as the ‘representative object interpretation’ (Krifka et al. : ). In (), for example, the narrator of a nature documentary is describing the dangers of poisonous snakes as the featured researcher collects specimens: () Three friends and colleagues have been killed by snakebites, including one from this very snake: the multi-banded krait. [corpus]

¹⁰ Note that both (a) and (b) also allow for non-generic interpretations, i.e., ones in which the speaker is referring to a specific group of dogs and some particular portion of ice cream, respectively; these interpretations are not at issue here.

OUP CORRECTED PROOF – FINAL, 21/1/2019, SPi



 .    

The image of the researcher collecting a krait snake and the inclusion of very in the demonstrative NP suggest an interpretation in which the demonstrative is being used to refer to the particular snake that is being shown in the documentary. However, the continuation following the demonstrative NP indicates that the narrator is referring to the kind by means of the image of an instance of the kind. In this way, () parallels the deictic use of demonstratives in that the referent is evoked on the basis of entities present in the context. With the pronominal forms of the demonstratives, both the taxonomic and nontaxonomic interpretations are available, as illustrated in (): () a. These are amazing dogs. [pointing at a Cocker Spaniel] b. This is my favorite ice cream. [pointing at a carton of Ben & Jerry’s ice cream] c. A: My sister owns a Honda Civic. B: Those are great cars. B: That’s a great car. [= Bowdle and Ward : ex. ()] These examples show that demonstrative pronouns permit reference to kinds with both the proximal and distal forms, and in both the singular and the plural. In (a), the demonstrative pronoun is interpreted as referring to a particular breed of dog, which bears a taxonomic relationship to the evoked kind DOG. Likewise in (b), the speaker is referring to a particular kind of ice cream. In both cases, the presence of a representative object facilitates the identification of the relevant kind that bears a taxonomic relation to the evoked kind. In contrast, B’s responses in (c) are interpreted non-taxonomically. Here, with either the singular or plural form, the demonstrative pronouns are interpreted anaphorically, with the speaker referring to the kind previously evoked by A’s utterance. With respect to the non-taxonomic kind-referring demonstrative NPs (as in () above), Bowdle and Ward () argue that the relevant kind must be assumed to be familiar to the hearer to ensure felicity. In (), for example, this assumption is required for felicitous use of the demonstrative (although the definite article would be felicitous in these sentences to introduce an unfamiliar kind). Bowdle and Ward’s () discussion also highlights three other constraints that distinguish the nontaxonomic generic uses of demonstratives from other kind-referring NPs. First, they note that the non-taxonomic use of generic demonstratives is restricted to evaluative predicates and is in fact infelicitous with ‘factual’ predicates, as in (): ()

#Those Labradors were first bred in Newfoundland.

Here, the demonstrative is being used to make a purely factual claim about where the breed was first bred and is, as a result, infelicitous because it lacks the evaluative aspect that was present in example (a).¹¹ Second, they note that the plural generic ¹¹ Note that the demonstrative can be felicitous with a taxonomic interpretation: (i)

Those dogs were first bred in Newfoundland. [pointing to a particular Labrador]

OUP CORRECTED PROOF – FINAL, 21/1/2019, SPi

     



non-taxonomic uses of demonstratives are only felicitous with kinds found at a subordinate level with respect to a folk taxonomy (Rosch ). For example, (a) is felicitous because the kind referred to with Labradors is a relatively subordinate kind of dog. In contrast, the basic level of the taxonomy is less available for generic reference, as in (): () #Those dogs are extremely loyal, you know. This example allows for a variety of different interpretations (e.g., deictic, with the speaker referring to a particular group of dogs, as well as a taxonomic generic one), but the non-taxonomic generic interpretation, in which those dogs is used to refer to the kind DOG, is unavailable because this kind is not at a subordinate level within the folk taxonomy of dogs. Finally, Bowdle and Ward also note that the kind referred to with a felicitous nontaxonomic generic demonstrative must be a coherent and non-ad hoc kind in order for the requisite constraint on familiarity to be satisfied. This requirement regarding the information status of kinds has also been noted with respect to other examples of generic reference. In (), from Carlson  (and attributed to Barbara Partee there), the kind COKE BOTTLE is sufficiently coherent or well defined in order to allow reference to the kind with the definite article, whereas the kind GREEN BOTTLE is not: ()

a. The Coke bottle has a narrow neck. b. #The green bottle has a narrow neck. [on the generic interpretation]

Krifka et al. () use these examples to argue that the requirement for a coherent kind holds for all uses of the definite article to refer to kinds, so this constraint is not unique to the use of demonstratives. What is distinctive about using demonstratives to refer to kinds is that, on the non-taxonomic interpretation, the kind referred to must be at a subordinate level on the relevant taxonomy which the speaker assumes constitutes mutual knowledge and at which functional homogeneity obtains.

.. Uses involving exemplars of identifiable kinds In addition to uses in which the speaker is referring to a kind, there are uses of demonstratives in which the speaker is referring to an instance of a kind assumed to be part of a shared cultural narrative, as shown in (): ()

a. Just because she’s not saying no doesn’t mean she’s saying yes. Don’t be that guy. [corpus] b. Looking for that perfect gift? Father Time Antiques has a selection of antique watches for . . . [corpus] c. Take that trip you’ve always wanted. [corpus] d. That moment when you have so much shit to do, that you decide to take a nap instead. [corpus] e. We were going to have this perfect life. [corpus]

OUP CORRECTED PROOF – FINAL, 21/1/2019, SPi



 .    

With this use of the demonstrative, the speaker is not referring to any specific entity: in, for example, (a), there is no single specific guy that the speaker is directing you not to be, nor is there any particular gift that is being sought out in (b). Instead, the demonstratives in these examples are being used to pick out an instance of a kind that is assumed to be familiar. In (a), the kind of guy is inferable from the image depicted on the poster and in (b) the speaker is assuming that the hearer is familiar with the kind PERFECT GIFT (whatever the instantiation of that kind ultimately turns out to be). What these examples require for their felicitous use is that the relevant kind be sufficiently familiar to the hearer in virtue of instances of this kind being either ideal, prototypical, or stereotypical for things of that sort. As there is no one specific referent for the demonstrative NP in this use, however, it is distinct from the private shared knowledge use discussed earlier, which requires a specific referent. While the most readily identifiable examples of this use all involve the distal form, (e) shows that the proximal form can be felicitous as well. With this use, however, only a full lexical NP is felicitous, as the common noun is needed to evoke the relevant kind. We also find this use with reference to events that form part of a broader cultural narrative, in which the entity referred to with the demonstrative stands proxy for the event-type within the relevant narrative. Examples are provided in (): ()

a. Megan graciously shared her story of the realities of a military wife, and how she felt the day she got that knock on the door. [corpus] b. It was not that I disliked the man, but like many people, I’m not really fit to talk in the morning until I’ve had that first cup of coffee. [corpus]

In these examples, the entity referred to by the demonstrative stands proxy for an event that is an instance of an event-type within a familiar narrative. In (a), the speaker presumes familiarity with the culturally-shared narrative about military families in which the event of a knock on the door stands proxy for the event of being notified by the military that a family member has died. Likewise, in (b), the speaker presumes that the hearer shares the cultural narrative that drinking coffee is required in order to function normally in the morning. The cup of coffee referenced by the speaker is emblematic for the event of drinking coffee, which is part of that narrative. As we discuss below, the features of this use also license many of the variable interpretations discussed in §..

.. Stereotypical use with proper names Genericity is also present in uses of demonstrative with proper names. It was first reported by Lakoff () that demonstratives can be used with proper names. There are two distinct interpretations, each with a characteristic intonational contour. On the first interpretation, the demonstrative is being used contrastively, in order to single out a specific individual from among those that share the name. The second interpretation

OUP CORRECTED PROOF – FINAL, 21/1/2019, SPi

     



is one in which the speaker is making a claim about what is typical or expected of the referent. Under the contrastive interpretation, the demonstrative receives a pitch accent in order to indicate that the referent is singled out from a set of alternatives, whereas under the stereotypical interpretation the demonstrative remains deaccented while the proper name receives a pitch accent. Both interpretations are available for (): () Oh, that John! On the contrastive interpretation of (), the name John functions like a common noun, denoting a set of persons that share the name John. The speaker uses the demonstrative NP to refer to one person among the (contextually-determined) set of individuals named John. Under this interpretation, the referent for the demonstrative NP may be determined either by means of deixis, anaphora, or assumed shared knowledge (as discussed above in §§..–..). On the stereotypical interpretation of the demonstrative NP in (), on the other hand, the speaker is referring to the relevant individual named John—presumably known to both speaker and hearer in virtue of the use of the proper name—and conveys that John has done something typical for him, for example, John is up to his old tricks again.¹² While not a kind-referring use, the stereotypical use of demonstratives with proper names nonetheless resembles the kind of genericity that Krifka et al. () call ‘characterizing sentences’. In a characterizing sentence, the speaker asserts that which is typically or usually true of the referent. Consider (): ()

a. Oh, that Justin Bieber, getting into trouble again. [corpus] b. #That Justin Bieber is from Canada.

Since the property of ‘getting into trouble’ is something that is characteristic of the referent, the demonstrative in (a) is felicitous. The stereotypical use of demonstratives with proper names requires mutual knowledge between the speaker and hearer about what is usual for, or typical of, the referent and, as a result, exemplifies features of genericity while also making reference to a specific individual. With a predicate that does not describe a stereotypical behavior of the referent, such as his provenance, the result is infelicitous, as shown in (b).

. P 

.................................................................................................................................. In §.. above, we discussed one use of demonstratives that is only felicitous with the proximal form: the so-called indefinite this use. Here, we discuss an additional use in ¹² Lakoff () characterizes this use as a distinct type of ‘emotional deixis’. Our taxonomy, however, does not include a separate category of emotional deixis. That said, each of the examples Lakoff provides can be classified in our system based on the usage conditions that underlie each of the categories we propose.

OUP CORRECTED PROOF – FINAL, 21/1/2019, SPi



 .    

which only the proximal form is felicitous. With this use, the demonstrative NP receives a property-denoting interpretation (rather than an entity-denoting one). These two features are illustrated in the examples in (): ()

a. Megan bought me the scariest thing alive!! It’s this/#that hamster and it makes noise and is voice activated and its cheeks light up and it moves around and follows commands!!!! [corpus] b. Look, the dealers are the popular kids, but they’re not normal popular. They’re these/#those crunchy granola dudes that have convinced everyone that they’re cool. [corpus]

In both (a) and (b), the demonstrative NP is not used by the speaker to refer to any particular entity; rather it is used to predicate a property of the referent, thus constituting a distinct use from those previously surveyed.¹³ Doran and Ward () argue that with the predicative use of proximal demonstrative NPs, the speaker classifies the referent with respect to a discourse-new property (in the sense of Prince ). The felicitous use of such demonstratives requires that there be a contextually salient ‘open proposition’ (OP) of the form ‘r is an x’, where r is the referent and x is the category that classifies the referent. Compare (a) and (b): () a. A: Who’s Joan?/What does Joan do?/What’s Joan like? B: She’s this lawyer. b. A: Who’s a lawyer? B: #Joan’s this lawyer. [cp. Joan’s a lawyer.] In (a), B’s response instantiates the variable of the OP by classifying the referent, Joan, as belonging to the category LAWYER. A’s question evokes the relevant OP given her request that B provide her with a contextually-appropriate category that classifies the referent. In (b), on the other hand, A’s question does not evoke the appropriate OP. Rather than asking B to provide a category that classifies Joan, A’s question asks instead for a member of the relevant category. Thus, the use of the proximal demonstrative is infelicitous in (b) because there is no contextually salient OP of the appropriate form.¹⁴ In addition to instantiating an appropriate OP, the property that classifies the referent must also be discourse-new (in the sense of Prince ). In fact, when the relevant property has been previously evoked, infelicity results, as shown in (): () a. A: My cousin Bob is coming over today. He’s this fashion model. B: What a coincidence! Have you met my cousin Sam? #He’s this fashion model, too. ¹³ As with indefinite this, predicative this is truth-functionally equivalent to the corresponding NP with the indefinite article. However, as predicative this is not used to refer to specific entities, we maintain that it constitutes a distinct use. ¹⁴ Note that the corresponding use of the indefinite article in (b) does not require any particular form of OP for its felicity.

OUP CORRECTED PROOF – FINAL, 21/1/2019, SPi

     



b. A: My cousin Bob is coming over today. He’s this fashion model. B: Have you met my cousin Sam? He’s a fashion model, too. c. A: My cousin Bob is coming over today. He’s this fashion model. B: Have you met my cousin Sam? He’s this interior designer. In (a), the property of being a fashion model has been evoked by A’s utterance and thus B’s use of the proximal is infelicitous because the property has become discourseold. In (b), the property evoked in B’s utterance is discourse-old, but B’s utterance is nonetheless felicitous if realized with the indefinite article. Finally, in (c), B’s use of the demonstrative is felicitous because the relevant property is discourse-new, despite the fact that A’s previous utterance also contained a predicative use of the proximal demonstrative. Thus, a discourse-new property is required for the predicative use of the proximal to be felicitous, whereas there is no such requirement for the corresponding use of the indefinite article. Since the proximal form is felicitous whenever the indefinite article would also be, it follows that the speaker’s choice of the demonstrative NP conveys additional information to the hearer. Typically, use of the proximal demonstrative conveys additional information about the referent itself beyond mere category membership based on various social and cultural stereotypes associated with the relevant properties, as illustrated in (): ()

a. Everyone thinks I’m this New Yorker. b. #Everyone thinks I’m this South Dakotan.

Here, the property of being a New Yorker is rich with stereotypical associations whereas the property of being a South Dakotan, for us at least, lacks such associations and thus renders the use of the demonstrative infelicitous.¹⁵

. Q 

.................................................................................................................................. With the exception of the predicative use of the proximal form, the uses surveyed thus far are all readily classifiable as different kinds of referential uses—distinguished on the basis of the nature of the referent and how it is interpreted. In contrast to these referential uses, there are a variety of uses noted in the literature in which the demonstrative functions semantically as a variable. Such examples pose interesting problems for the semantic analysis of demonstratives, as they provide straightforward counter-examples to the claim often appearing in the philosophical literature that demonstratives are

¹⁵ Of course, South Dakotans (and perhaps North Dakotans as well) might find this example felicitous, given their familiarity with the kind in question.

OUP CORRECTED PROOF – FINAL, 21/1/2019, SPi



 .    

essentially devices of direct reference (Kaplan a). In these quantificational uses, however, the demonstrative NP is not being used to refer to a single entity, but rather denotes entities within a domain of quantification. The implications for the semantic analysis of demonstratives are beyond the scope of our concerns here; instead, we will concern ourselves with describing three different variable interpretations of demonstrative NPs.

.. Bound variable uses with explicit links In the first quantificational use we discuss, the demonstrative NP is interpreted as a bound variable, as shown in (): ()

a. Mary talked to no senator before that senator was lobbied. [= Elbourne : ex. ()] b. Every dog in my neighborhood, even the meanest, has an owner who thinks that that dog is a sweetie. [= Roberts : ex. ()] c. Every scientist who was fired from the observatory at Sofia was consoled by someone who knew that scientist as a youth. [= Abbott : ex. ()] d. On every team there is one player who is not as strong as the rest. That weakest member is the one to play hardest against. [= Maclaran : ex. ()]

In each of these examples, there is a reading available in which the demonstrative expression is interpreted as a variable bound by a quantifier expression; no senator in (a), every dog in my neighborhood in (b), every scientist who was fired from the observatory at Sofia in (c), and every team in (d). The variable interpretation that the demonstratives receive in () is due to the explicit link to the preceding quantified expression, which need not occur in the same sentence as the demonstrative itself, as shown in (d). In each of these examples, the bound variable interpretation of the demonstrative is the preferred one in the absence of a gesture or something else to indicate a deictic interpretation of the demonstrative. While each of the examples in () uses the singular distal form, (a) would also be felicitous with the plural those senators and (c) is perhaps possible with the singular this scientist. However, note that the examples in () are each with lexical NPs and would not be felicitous with pronominal forms.¹⁶ We include within this category a range of examples involving demonstratives that parallel familiar examples from the literature on anaphora.

¹⁶ As the examples of the bound variable use that have been identified in the literature are quite limited in number, we will leave unresolved the question of whether pronouns may in fact be felicitous on such interpretations.

OUP CORRECTED PROOF – FINAL, 21/1/2019, SPi

      ()



a. Michelin is hoping to hire ten more tire inspectors. These new employees would be required to work the night shift for the first three weeks. [= Maclaran : ex. (b)] b. If Pedro owns a donkey, Pedro will ride that donkey to town tomorrow. [= Abbott : ex. (h)] c. When a bishop meets another bishop, that bishop always blesses that other bishop. [= Abbott : ex. (b)]

These examples are distinct from the earlier anaphoric examples, as in each example there is no particular discourse referent to which the demonstrative is being used to refer. For this reason we have not classified these examples with the anaphoric uses discussed above. Rather, we have included them under the bound variable interpretations, as in each of the examples in () the demonstrative NP receives a variable interpretation that depends upon the values that the antecedent expression receives. Maclaran () discusses (a) as an example of a non-specific use of the demonstrative, as there are no specific employees that are required to work the night shift for three weeks. The relevant interpretation of (a) is that if ten more inspectors are hired, then they would be required to work the night shift, without an existential commitment to there being any actual inspectors. Example (b) shows that demonstratives are felicitous in so-called ‘donkey-sentences’ and (c) is an example of an ‘indistinguishable participants’ sentence with a demonstrative NP. The precise treatment of these types of sentences, even apart from the examples with demonstratives, is a matter of much current theoretical debate. We note here only that demonstrative NPs can be felicitously used in a range of interesting constructions in which they are interpreted as variables, rather than as referring to particular discourse entities.¹⁷ The bound variable uses of demonstratives are unexpected under the assumption that demonstratives are essentially referential devices: in these uses, the demonstrative NPs receive a variable interpretation and there is no particular entity being referred to by the speaker. One puzzling feature of many of these examples is that they are, at least to our ears, much less natural with the definite article in place of the demonstrative. Consider (): () On every team there is one player who is not as strong as the rest. #The weakest member is the one to play hardest against. Here, the definite description does allow a variable interpretation parallel to the relevant interpretation of (d), although it seems considerably less felicitous than the corresponding example with a demonstrative. In light of this, a demonstrative NP with a bound variable interpretation cannot simply be replaced with a definite description preserving the bound interpretation. Thus we see that, even with the bound variable uses, there is something pragmatically significant to the speaker’s choice of the demonstrative as opposed to the definite article. ¹⁷ See Abbott  for further discussion of demonstratives in donkey sentences, and more generally for the issues surrounding the interpretation of these sentences.

OUP CORRECTED PROOF – FINAL, 21/1/2019, SPi



 .    

.. Bound variable uses with inferential links The second quantificational use is one in which the demonstrative NP is interpreted as co-varying with a quantified expression through an inferential link, as illustrated by the distal and proximal demonstratives in (): () a. Most avid skiers remember that first black diamond hill. [adapted from King : Chapter , ex. ()] b. Every bride looks forward to this day. [corpus] c. Here are three steps to help getting you to take that first plunge underwater as a certified Scuba Diver. [corpus] The demonstrative NP in (a) is most naturally interpreted as selecting, for each skier, the first time he or she attempted to ski an advanced-level hill. Likewise in (b) the speaker’s use of the demonstrative picks out, for each bride, the day of her wedding, rather than, say, the day of utterance. In (c), the relevant first plunge underwater is the first dive of each person who has newly become a certified scuba diver. This relation between avid skiers and black diamond hills in (a) is such that it is stereotypical for avid skiers qua avid skiers to have attempted black diamond runs. Thus it can be assumed that for each avid skier there is one particular black diamond run that was his or her first. Similarly for example (b), while there are many inferences that can be drawn based on stereotypical information associated with brides, that there is one unique day for each bride—her wedding day—is inferable on the basis of what is assumed to be shared stereotypical knowledge for someone in virtue of being a bride. In (c), anyone reading the guide is looking to become a certified scuba diver, and so every such individual would be associated with a unique first dive as a scuba diver. The examples in () each involve lexical NPs, but pronominal forms can be felicitous as well, given an appropriate context, as illustrated in (): () A: I just skied my first black diamond hill! B: Every avid skier remembers that. Here, the pronominal form in B’s utterance is felicitous on the intended interpretation as A’s utterance has evoked black diamond ski hills which B then stereotypically associates with avid skiers. King () has argued that such interpretations are evidence that demonstrative NPs are quantificational because the demonstrative NPs may take narrow scope with respect to a quantifier. However, there is no quantified expression in (c) to account for the variable interpretation. Instead, the pronoun you licenses the intended interpretation, as it refers to whoever is reading the guide. For the examples in (), what is required for the intended interpretation is a bridging inference; that is, in (b) there is a particular day that is salient for each bride and this inference is based on stereotypical world knowledge about brides that is presumably shared by speaker and hearer. It is possible, of course, for the variable interpretations of demonstrative NPs to be

OUP CORRECTED PROOF – FINAL, 21/1/2019, SPi

     



generated by scope interactions, as King claimed, but such interpretations do not require explicit quantification into the demonstrative NP. This kind of variable use requires that an inferential link hold between entities in the domain of quantification and the values of the variable represented by the demonstrative. For example, in (a) the speaker is assuming it to be mutual knowledge that avid skiers have attempted to ski down difficult (i.e., black diamond) hills and, likewise, for (b), the speaker is assuming it to be mutual knowledge that brides have wedding days. The basis for associating each set member with a value for the variable is heavily constrained as not every association will support the use of the demonstrative, as shown in (): ()

a. #At lunch, every girl was quietly reading that book. b. At lunch, every girl was quietly reading her book.

In (a), we see that not all entities associated with a quantified expression are felicitous with this use of the demonstrative. While every girl may have a unique book associated with her, as evidenced by the felicitous use of the possessive in (b), such a relationship does not support the demonstrative. It would appear that the use of a possessive NP indicates an explicit mapping from entities in the quantified set to values of the variable while, in the absence of such an explicit indication, the relationship must be left to inference. The relevant interpretation is available in the absence of an explicit link just in case the discourse context licenses an interpretation in which the demonstrative specifies an entity that is associated with a significant event assumed to be familiar to the hearer on the basis of shared cultural narratives. Taken together, the result of these constraints is that the kinds of entities specified by the demonstrative are emblematic of milestone events within the relevant narrative. In support of this analysis, we note first that the kinds of things specified by the demonstrative NP must be associated with events that are significant. Consider the contrast between (a) and (b): ()

a. Every professor cherishes {that/their} dissertation. b. Every professor cherishes {#that/their} office.

The demonstrative in (a) is licensed by the culturally salient professor narrative in which a dissertation stands proxy for the event of writing, completing, and defending a doctoral dissertation. In contrast, the demonstrative in (b) is infelicitous, as a professor’s office does not typically stand proxy for an event within the professor narrative. Further, the demonstrative NP must specify entities that are stereotypically related to the kind denoted by the subject NP. Consider (): ()

a. Every pilot remembers that first flight. b. #Every pilot remembers that first car.

In (a), given stereotypical assumptions about pilots, that is, that they fly planes, it can be safely assumed that each pilot has had a first flight and that this flight constitutes a

OUP CORRECTED PROOF – FINAL, 21/1/2019, SPi



 .    

significant event in the pilot narrative, hence the felicity of the demonstrative. In (b), on the other hand, the demonstrative is infelicitous given that acquiring a first car is not a significant event in the pilot narrative (even under the assumption that every pilot has owned a car). However, the demonstrative is rendered felicitous when acquiring one’s first car is a significant event in some culturally salient narrative—for example, that of a teenager, as illustrated in (): ()

a. #Every pilot remembers that first car. [= (b)] b. Every teenager remembers that first car.

In (b), the car specified by the demonstrative is standing proxy for the event of acquiring one’s first car, which is part of the culturally salient teenager narrative. The demonstrative here is felicitous because the entity specified by the demonstrative is emblematic of a significant milestone within the larger narrative. Note further that the demonstrative in (b) is licensed by the shared narrative regarding teenagers and cars, rather than a presupposition to the effect that every teenager has had a first car. Indeed, the failure of a single teenager in the relevant domain to have had a first car does not render (b) truth-valueless; rather, the truth value of (b) depends on whether, for those teenagers who have had a first car, it is the case that each of them remembers it.

.. Restrictive that uses The final quantificational use of demonstrative NPs was first noted by Maclaran (), and involves only the distal forms, that and those. This use occurs with both lexical NPs (singular or plural) and pronouns (also singular or plural), as shown in (): ()

a. Only that income above $ earned outside the country will be taxable. [= Maclaran : ex. ()] b. Those swans which live on Britain’s waterways are the Queen’s. [= Maclaran : ex. ()] c. That which doesn’t kill you makes you stronger. [corpus] d. Those who know drink Don Julio. [corpus]

Maclaran terms this use ‘restrictive that’, as the relevant interpretations of the examples in () are felicitous only with the distal forms and involve a post-nominal clause that restricts the set of entities within the relevant domain.¹⁸ The demonstratives in () are each interpreted as universal quantifiers: in (a), all and only income above $ ¹⁸ In our corpus, we have found exactly one example of the proximal form with the restrictive interpretation, from Wilfred Owen’s poem ‘Anthem for Doomed Youth’: What passing-bells for these who die as cattle? We’re not quite sure what to make of this example other than to note that poetry often takes liberties with linguistics conventions of usage.

OUP CORRECTED PROOF – FINAL, 21/1/2019, SPi

     



earned outside the country is taxable; in (b), every swan living on British waterways belongs to the Queen; in (c), all things that don’t kill you make you stronger; and in (d), each person that knows good tequila drinks Don Julio tequila. These restrictive interpretations are possible when the demonstratives are used without prior mention of the entities satisfying the descriptive condition (and are not used deictically). Two features of this use are noteworthy: first, this use of the demonstrative does not presuppose that there are any discourse referents that satisfy the descriptive component of the demonstrative NP; that is, (a) can be used to make a true assertion even if there is no income above $ earned outside the country. Second, the restrictive that use is felicitous in existential there-sentences, as in (): () There are those who say that, since we have no militia, the amendment no longer applies; they would just ignore it. [corpus] Here, the existence of the relevant individuals is being asserted—and not presupposed—by the existential.¹⁹

. C

.................................................................................................................................. In this chapter, we have classified a total of thirteen distinct uses of demonstrative NPs on the basis of two factors. First, we began by considering the semantic value of the demonstrative, that is, whether the speaker is using the demonstrative to denote or refer to an entity, a kind, a predicate, or a variable in a quantificational domain. Second, further divisions within these four broad categories were made on the basis of how the relevant value for the demonstrative is assigned in context. We have produced the preliminary taxonomy of demonstrative uses—summarized in Table .—with the goal of providing future researchers with a framework for analyzing new uses of demonstrative in English as well as other languages.

¹⁹ As mentioned above, demonstratives may felicitously occur in existential sentences under two other interpretations: indefinite this and uses based on private shared knowledge.

Lexical NP only

N (proximal only)

Y

Y

Non-taxonomic, kind-referring uses

Uses involving exemplars of identifiable kinds

Reference to kinds and other generic uses Taxonomic, kindY referring uses

Lexical NPs only

Y

Y

Lexical NPs only

Y

Inferrable uses (including bridging inferences) ‘Indefinite’ this use

Lexical NPs only

The kind mentioned bears a taxonomic relationship to the evoked kind; demonstrative must be accented Kind is restricted to subordinate kinds in a familiar taxonomy; demonstrative must be deaccented Referent is a non-specific exemplar of the relevant kind

Looking for that perfect gift?

Those dogs make great pets. [referring to a particular breed of dog] Those Labradors make great pets.

. . . but I was saving those (strawberries) for dessert. Referent assumed to be familiar to speaker There were those neighbors at and hearer; demonstrative must be the City Council meeting deaccented yesterday. Referent is inferable based on assumed John is worried about moving to familiarity with an evoked kind the south and picking up that drawl. Referent assumed to be unfamiliar to There was this funny rattle hearer and is being marked as topical under the hood.

Y

Referent evoked in previous discourse

Referent evoked in immediate situational context

This (cookie) is delicious.

Representative example

Y

Both pronominal and Distinctive features of this use lexical NP forms felicitous?

Uses based on private Y shared knowledge

Reference to discourse entities Y Deictic uses (including discourse deixis) Anaphoric uses Y

Both proximal and distal forms are possible?

Table . A preliminary taxonomy of the uses of English demonstratives

OUP CORRECTED PROOF – FINAL, 21/1/2019, SPi

Y

N (distal only)

Restrictive that uses

Y

N (proximal only)

Y

Bound variable uses with inferential links

Quantificational uses Bound variable uses with explicit links

Predicative use Predicative use

Stereotypical uses with proper names

Y (but post-nominal modifier is required for both forms)

Y

N(?)

Lexical NPs only

Lexical NPs only

Oh, that Justin Bieber getting into trouble again.

The variable interpretation of the demonstrative is explicitly linked to the set denoted by the preceding quantificational expression (e.g. via repetition or coreference) The variable interpretation of the demonstrative is inferentially linked to the set denoted by the preceding quantificational expression (e.g. via world knowledge) No existential presupposition associated with the post-nominal modifier

Those who know drink Don Julio.

Every professor cherishes that dissertation.

Mary talked to no senator before that senator was lobbied.

Speaker classifies the referent of the subject She’s this lawyer. NP with respect to a discourse-new property

Referent is an individual behaving in a typical or characteristic manner; demonstrative must be deaccented

OUP CORRECTED PROOF – FINAL, 21/1/2019, SPi

OUP CORRECTED PROOF – FINAL, 17/1/2019, SPi

  ......................................................................................................................

CONTEXTUAL INFLUENCES ON REFERENCE ......................................................................................................................

 

. I

.................................................................................................................................. I is not linguistic expressions themselves which refer, but speakers who refer in using them (Strawson ). The difference lies in what it is to make an utterance in discourse. Following Bar-Hillel (), we take an utterance to be an ordered pair of a linguistic constituent s and a context of utterance c. As is common, we take c to specify a great deal of information not only about the concrete situation in which the utterance was made—the speaker, addressee, location, time, etc.—but also about the interlocutors’ Common Ground and perhaps other information that is systematically tracked in a conversational record. The linguistic constituent s itself has a conventional content (or character, if one follows Kaplan ) which we’ll assume is truth conditional. But its meaning is another matter. A simple example: ()

[To a companion seated in a café, looking out the window:] Look at that UPS guy over on the corner. The man obviously needs help with that big package.

Assuming there’s no more than one UPS guy on the corner in question, the speaker will have succeeded in referring to him with the man, even if there are other men on the same corner. Yet in and of itself, that definite description fails to have sufficiently rich descriptive content to pick out the intended referent. It is clear that the speaker is assuming that her addressee will draw the inferences required from the context of the preceding discourse, in which she’s just pointed out the UPS guy (and no other men), to conclude that said UPS guy was who she meant to refer to. Consider the matter from a Gricean point of view: ‘U meantnn something by uttering x’ is true iff, for some audience A, U uttered x intending: () A to produce a particular response r

OUP CORRECTED PROOF – FINAL, 17/1/2019, SPi

   



() A to think (recognize) that U intends () () A to fulfill () on the basis of his fulfillment of (). (Grice ) [Moreover,] there is no inference-element E such that U uttered x intending both (') that A’s determination of r should rely on E and (') that A should think U to intend that (') be false. (Grice : ) If in uttering a sentence s in context c a speaker U meantnn to proffer that the proposition p is true, then the speaker’s intended meaning in uttering was p (or, perhaps more strictly, the speaker meant for the addressee to recognize that the speaker intended to proffer p). But the relationship between an uttered sentence s and the proposition expressed p may be quite complex, involving the resolution of presuppositions, vagueness, and ambiguity, and the recognition of intended conversational implicatures. Similarly, if the constituent uttered was a definite noun phrase np, it may be that by the utterance the speaker meantnn to refer to some singular individual e. But the relationship between the conventional content of np and the intended referent e may require the resolution of the same kinds of factors as in the sentential case, always mediated by the understood context c. Arguably, linguistic reference is typically far less direct (in the sense of ‘conventionally given’) than is often assumed. Then the leading questions that arise in a compositional account of interpretation are: (i) How do speakers use linguistic expressions to refer in contexts of utterance? And (ii) how does conventional content constrain such use? If we aim for a generative account of meaning, in which we predict all and only the meanings that will arise given conventional content, syntactic structure and context of utterance, then we take the gap between content and meaning to be a problem of context dependence. As we will see in what follows, the following types of context-dependence are found across all types of referring expressions: anaphora and familiarity presuppositions descriptive incompleteness and domain restriction dependence on a shifted perspective dependence on context-sensitive predicates Even those NP-types that most theorists agree to be referential—proper nouns— display these types of context-dependence, and hence cannot be successfully used to refer without adequate contextual clues to their intended referent. For example, Mary doesn’t uniquely pick out any individual. It’s only in context that its use conveys singular reference, permitting the addressee to pick out the intended referent. In this sense, we might say that Mary is an incomplete name, and thus its use to refer is essentially context-dependent. Similarly, the central puzzles pertaining to de re belief attribution involve proper names like Ortcutt (Quine ) and London versus Londres (Kripke ). One might say that the interpretation of these names in a given context depends on the guise under which the res denoted is known to the speaker or some other agent relevant in the context of utterance, and hence on that agent’s perspective

OUP CORRECTED PROOF – FINAL, 17/1/2019, SPi



 

on the res. This argues that these types of context dependence are central to how we retrieve intended reference generally, even in NP types often taken to be relatively context-insensitive. Following common usage, I’ll refer collectively to those NP types which are canonically used to refer as definite NPs (Lyons ; Abbott ). Depending on the inventory of NPs in a given language, these will certainly include proper names; indexicals and demonstratives; and pronouns. In some languages, like English, it also includes definite descriptions, or their closely related kin. For example, in Bulu (Bantu) (Barlew ), there is a non-demonstrative definite determiner te- which differs from English the in important respects, but forms NPs which are used to pick out entities that the speaker has reason to believe the addressee is paying attention to. And in languages without definite determiners, like Japanese or Serbo-Croatian, bare NPs—those without article or determiner—may sometimes also be used to refer. But here we will focus on the English definites, though I take them to illustrate problems of more general interest.

. R    

..................................................................................................................................

.. Anaphora and familiarity presuppositions: Even with proper names there’s reason to think that their felicitous use by competent speakers presupposes familiarity on the part of the addressee. This familiarity doesn’t require some kind of personal acquaintance with the individual which a name picks out. Rather, in order for a proper name to be felicitously used, the name must be properly introduced to the addressee. For example, the use of Ernest in () is a bit odd: ()

There is a gentleman in Hertfordshire. Ernest is engaged to two women.

Either (a) Ernest is an individual already known to the addressee (at least ‘by name’), in which case the second sentence seems like a non sequitur when uttered after the first, or else (b) the speaker intends by use of the name to refer to the aforementioned gentleman in Hertfordshire. However in the second case, if the gentleman were familiar to the addressee then the speaker wouldn’t have used the first sentence; but if he wasn’t familiar prior to this, the speaker hasn’t really properly introduced the name, so that the addressee is being required to accommodate the answer to Who’s Ernest? and what’s he got to do with Hertfordshire? There are a number of ways that a name can be properly introduced. Here are just a few: Cumming’s () “naming construction”: ()

There is a gentleman in Hertfordshire by the name of ‘Ernest’. Ernest is engaged to two women.

OUP CORRECTED PROOF – FINAL, 17/1/2019, SPi

   



Appositives, also a construction for naming: ()

There is a gentleman in Hertfordshire, Ernest. Ernest is engaged to two women.

()

I’d like you to meet my friend Ernest. Ernest is engaged to two women.

Deictic introduction ritual: ()

A is speaking to two companions: Charles (nodding and gesturing to Charles), this is Ernest (nodding and gesturing to Ernest). Earnest (nodding and gesturing to Ernest), Charles (nodding and gesturing to Charles).

But the original way that a name is introduced into the language is via a dubbing event: Dubbing: () A: Who’s that? B: I dunno. Let’s call him ‘Ernest’. I think Ernest is awfully cute, don’t you? Compare () with I hereby dub thee Sir Ernest, or a naming by parents. In these cases, there is a socially granted authority vested in the dubber. But nicknames and nonce cases like () do frequently occur, and the resulting association can persist just as well in the informal cases as in the more formal ones. A dubbing might be regarded as the origin of the type of causal chain that Kripke () argued underlies the direct referentiality of proper names: the dubbing conventionally associates the name with its intended referent. The other types of proper introduction could then be regarded as the establishment of new links, extending the chain to the new acquaintance so that s/he is familiar with it and may use the name correctly. Thereafter, it is felicitous to use the proper name with that newly introduced acquaintance to refer to the intended bearer of the name. So even proper names bear a familiarity presupposition, albeit of a special type: felicity requires that the addressee be familiar with the existence of a causal chain connecting the name with some original dubbing event, the latter associating the name with its intended referent. All definite NPs tend to carry familiarity presuppositions, though the particular character of these presuppositions may differ between different definite NP types. The classic type of familiarity presupposition is anaphora to an overt preceding NP, wherein the speaker intends that the intended referent of a pronoun, definite description or demonstrative description be retrieved via coreference with the overt antecedent. But there are other means of making the intended referent contextually salient, and hence helping to fix the definite’s meaning. We find the following types of anaphoric uses of third person pronouns, demonstratives, and definite descriptions, reflecting the properties of generalized anaphoricity outlined by Partee ():¹

¹ Whether such definite NPs are always anaphoric is more controversial. For our purposes, what matters is that they do have such uses. See Heim (); Lyons (); and Abbott () for enlightening overviews of the relevant literature.

OUP CORRECTED PROOF – FINAL, 17/1/2019, SPi



 

coreferential with an antecedent NP: ()

A man and a boy were coming down the street. The man looked worried.

()

A man was walking down the street. He looked worried.

()

I saw one quilt which was quite abstract, with lots of asymmetric diagonals. Another one was more traditional, worked in an old Amish pattern. This quilt was less busy than the other, but just as bold.

having a non-linguistically salient referent: ()

[Context: looking together at a house:] The roof needs fixing.

()

[Context: looking together at a house:] It needs a coat of paint.

()

[Context: looking together at a house:] That roof needs fixing.

The non-linguistic salience of the intended referent illustrated for definite descriptions in () is much like that displayed in the canonical use of demonstratives, as in (). Because the neuter singular pronoun it, unlike the rd person masculine and feminine or plural pronouns, has no demonstrative uses (Maclaran ), its use in () cannot be demonstrative. All three of these NP types also have bound variable interpretations, where they have a quantificational or irrealis antecedent: quantificationally bound: ()

At the boy scout camp, every father and son that built a fire together decided that the boy would gather brush and wood while the man made a clearing and laid the fire.

()

Every couple that built a fire together decided that she would gather brush and wood while he made a clearing and laid the fire.

()

Every dog in my neighborhood, even the meanest, has an owner who thinks that that dog is a sweetie.

participating in donkey anaphora:² ()

If a cat and a dog have a fight, the cat usually wins.

()

If John sees a car he likes, he should buy it.

()

If an upwardly mobile yuppie sees that a neighbor has a car, he usually worries that that car is cooler than his.

² These are so-called because of the original examples introduced by Geach (): If a farmer owns a donkey, he beats it, and Every farmer who owns a donkey beats it, the underlined pronouns called ‘donkey pronouns’.

OUP CORRECTED PROOF – FINAL, 17/1/2019, SPi

   



The underlined NPs in ()–() are instances of the donkey pronouns discussed by Geach (), the problems they present for anaphora and semantics admirably explained in the first chapter of Heim () (and see §.). Each takes as its antecedent (the discourse referents for) an arbitrary instance of an entity introduced by an indefinite NP in the conditional antecedent. So-called pure indexicals (Kaplan ) also arguably carry familiarity presuppositions. For example, speaker and addressee are always in some sense salient in a context, serving to satisfy the familiarity presupposition of indexical I and you. The range of familiarity presuppositions exemplified above argues that in the general case, definite NPs so-used require only weak familiarity (Roberts , , ), that is, that the context of utterance accessible to all the interlocutors entail the existence of the relevant entity (and, for proper names, of the relevant causal chain, as above). Thus, cases like ()–() where there is an explicit coreferential antecedent—satisfying strong familiarity—are a special case.

.. Descriptive incompleteness Some definite NPs depend for their successful referential use on their descriptive content, in some cases almost entirely: consider, for example, the nominal complement of the in the owner of  Schermerhorn St. throughout  or the tallest mountain in West Virginia. Assuming there was exactly one owner of said building during the period in question, then by virtue of the descriptive content alone, the first definite seems to refer to that individual. And superlatives of their nature pick out a single individual (assuming some individual does satisfy their descriptive content). But it is not always the case that a definite’s descriptive content is sufficient to pick out the intended referent on a particular felicitous use. Nor do such definites always have an explicit antecedent which in some sense adequately ‘fills out’ the descriptive content. If an NP’s descriptive content is not rich enough to pick out the intended referent, yet the NP evidently felicitously refers, then that content must be in some sense sufficient in the context of utterance. The definite description the man in () has an antecedent, that UPS guy over on the corner, and that seems sufficient to make the description’s content referentially adequate. If the NP doesn’t have an overt antecedent, as in ()–(), the speaker must rely on other contextual factors to make it clear what she intends to refer to. Pronouns, of course, are the limit case, due to their extremely poor descriptive content. Yet uses of pronouns without explicit antecedents, as in (), are quite common and felicitous. The more interesting question theoretically is when they are not acceptable. In canonical uses of demonstrative NPs, accompanied by a deictic gesture, the gesture may be sufficiently clear to pick out a unique entity satisfying the descriptive content of the NP. That might be the case in (), for example. But even deixis needn’t suffice. Consider: () [Police officer, following in the path of another running man through a crowd of people, points toward the runner ahead:] Stop that man!

OUP CORRECTED PROOF – FINAL, 17/1/2019, SPi



 

Even though there are many men in the crowd, that man accompanied by the gesture will probably succeed in referring to the man being chased. But then it does not do so either in virtue of the combination of the descriptive content man and the gesture—in the direction the policeman is pointing there may be many men. Instead, the evident chase itself and the stereotypical character of situations in which police chase criminals both play a crucial role in successful reference.

.. Domain restriction Quantificational operators, including quantificational determiners, adverbs of quantification, tense operators, and modals, are most often intended to be understood relative to some intended implicit restriction on their domain of quantification. In the context suggested for (), the speaker most likely intends to use the vocative everyone to refer to (and thereby call to) the students in that classroom: ()

[in a classroom:] Everyone get out your notebooks.

Though the class of quantificational NPs which includes everyone is not considered referential, domain restriction does come to bear on reference in at least two ways. First, on some analyses, like that of Russell () or Neale () definite descriptions involve existential quantification, the domain explicitly restricted by the nominal complement. Then one way of contending with the problem of incomplete descriptions is to argue that the speaker is understood to intend that they be interpreted with a contextually appropriate domain restriction. For example, the existential operator that is part of the logical form of the man in () would be understood as ranging over the set of men walking down the street with a boy in the situation of utterance, yielding the relevant man. In (), the existential associated with the roof would be understood to range over the set of roofs the interlocutors are considering; in felicitous use this is a singleton set, and the Russellian truth conditions are satisfied. This is one way one might understand so-called E-type (or D-type) analyses of definite descriptions and unbound pronouns (Evans , ; Neale ; Elbourne , )—as investigations of how the domain of the existential gets ‘filled in’. The virtue of this approach to free (unbound) definites across discourse is that it appeals to a phenomenon, domain restriction, which is independently attested and about which there is a significant literature (von Fintel ; Roberts ; Stanley and Szabo ). Another way in which domain restriction may come to bear on the interpretation of definite NPs is illustrated here: ()

A thief broke into the house. The bastard stole the silver.

()

A thief might break into the house. The bastard would steal the silver.

In each of these pairs, the epithet the bastard is clearly intended to be anaphoric to the thief introduced in the first utterance. Let’s assume that the anaphoric relationship is

OUP CORRECTED PROOF – FINAL, 17/1/2019, SPi

   



fairly straightforward in (), where one is reasonably irritated with someone that steals one’s silver, so that the demeaning epithet is understood to apply to the antecedent thief. But there are technical difficulties explaining how the comparable anaphora is licensed in (). This is brought out in comparing () with (): () A thief might break into the house. Yesterday I saw the bastard sneaking into the garage. For many, () is also acceptable, but here the bastard has a different sense. In () we most naturally take the first sentence to speculate about some arbitrary thief—any thief—possibly breaking in. So the bastard refers not to any particular person, but attributes another property to that arbitrary thief. But in (), we can only understand the bastard to be a particular individual, so that a thief must be understood itself to refer to a particular thief the speaker has in mind. This illustrates the following principle (Chierchia and Rooth ; Roberts b): Scope constraint on anaphoric relations: If NP x is to be a potential anaphoric antecedent for NP y, then any quantificational elements which have scope over x must also have scope over y. Prima facie, the Scope Constraint would seem to rule out the type of anaphora observed in (), since both a thief and the bastard take narrow scope relative to the modals in their respective sentences (might, would). Roberts (, b) argues that the kind of interpretation observed in () is licensed by a phenomenon she calls modal subordination, wherein the modal would that takes scope over the anaphoric definite is understood to have its domain restricted to only range over scenarios in which a thief breaks into the house—the sort of situation described in the first utterance. Then the relationship between the two NPs themselves, a thief and the bastard, is indirect, mediated by the familiar thief in the relevant type of scenario restricting the domain of would. Roberts calls a thief in such a case a licensing NP, rather than a true anaphoric antecedent. In (), since there is no modal in the second sentence, modal subordination is not possible. Then because the scope constraint precludes an antecedent taking narrow scope with respect to an operator that doesn’t have scope over the anaphoric element and might cannot take scope over the following sentence, the only reading possible is one where a thief itself is understood to take wide scope relative to might—a specific indefinite interpretation, wherein the speaker is taken to have a particular thief in mind. Some might take the use of the definite under modal subordination, as in (), to be non-referential, since the intended denotation is not a singular individual. Perhaps, but note that modal subordination can set up what seems like clear reference, as in (): () A: A mouse must have gotten into the pasta. It could have gotten in through that hole in the back of the cabinet. B: Yes, I see: Those are the little bastard’s teeth marks on the linguini. Consider how this parallels Strawson’s example (cited in Neale ; discussed in Roberts )

OUP CORRECTED PROOF – FINAL, 17/1/2019, SPi



 

() A: A man jumped off the cliff. B: He didn’t jump, he was pushed. In (A), the speaker intends to refer to someone; in (B), the speaker takes issue with the description provided by A, but still intends to refer to the same person A had in mind, the pronoun understood to mean something like ‘the person who apparently jumped on the occasion in question’. Similarly, in (A), the speaker considers a hypothetical mouse, while in (B) the speaker just assumes that the mouse is real and contributes further information about it. So the understood interpretation of the little bastard in (B) is ‘the (irritating) mouse who got into the pasta’. This content must be contextually retrieved, as is the case in (). In each case, B simply assumes there was an entity of the relevant sort and, in saying something about it, supposes that A can follow along.

.. Shifted perspective Intensional contexts play another role in the interpretation of definite NPs, besides the anaphora resolution we find in modal subordination. If an NP falls within the scope of an attitude predicate or epistemic modal, the context may trigger a shift in the interpretation to an interpretation de dicto, de re, or de se. Quine () tells a famous story, along the following lines:³ Ralph lives in a small New England town, where he’s acquainted with the mayor, a fellow named Ortcutt. Ralph generally takes Ortcutt to be a fine-upstanding citizen. However, while out walking in a rather rough section of town one evening, Ralph observes a strange fellow lurking about in a trench coat, with a broad-brimmed hat pulled down to shade his face. Unbeknownst to Ralph, the man is Ortcutt. Suppose that Ralph is accompanied by his friends Mark and Evan. He tells Mark of his concern about this shady character, who he thinks is a spy, and then runs off to find a policeman. Mark might report Ralph’s concern to Evan by truthfully uttering (): ()

Ralph believes that the man in the broad-brimmed hat is a spy.

Now suppose instead that Ralph is accompanied by his friends Steve and Zack, who both work in the mayor’s office and know His Honor’s habit of going about at night incognito to keep his finger on the pulse of the city. Ralph quickly tells them about his concern and runs off to find a policeman. Steve might turn to Zack and utter (): ()

Ralph believes that the mayor is a spy!

It’s clear in the scenarios described that both () and () are true. Ralph thinks that the person he knows as the man in the broad-brimmed hat is a spy. And the suspicious ³ I have embroidered the story in various ways and added the first examples in () for pedagogical purposes, to try to bring out the point of the puzzle as I believe Quine intended it, and to clarify its relationship to the distinction between de dicto and de re.

OUP CORRECTED PROOF – FINAL, 17/1/2019, SPi

   



attire of the guy that Steve and Zack know to be the mayor has led Ralph to believe that he’s a spy. But of course there’s another sense in which () is false. Ralph wouldn’t describe the suspected spy as the mayor—Steve uses that description in his report of Ralph’s beliefs because he knows that Zack understands who the ‘spy’ really is. The false sense of () is said to involve a de dicto interpretation of the NP the mayor, while on the true interpretation it is understood de re. On the de re interpretation, the speaker uses the NP to pick out the individual that he and the addressee know to uniquely bear that description, and says of that individual–the res–that Ralph believes he is a spy. (See Keshet a; and Keshet and Schwarz (Chapter  of this volume) for discussion of the relevant literature on de dicto and de re interpretations and semantic constraints on their occurrence.) Note that instead of uttering (), Steve might report Ralph’s concern to Zack using the proper name of their boss, as in (): () Ralph believes that Ortcutt is a spy! This, too, seems like a correct report of Ralph’s belief. Again, since Ralph doesn’t know that the man in question is Ortcutt, () can only be true with a de re belief attribution. However, since Ralph also believes of Ortcutt, the man he knows as the mayor, that he is an upstanding citizen, () is true as well: () Ralph believes that Ortcutt is not a spy. One might say that () is true on a de dicto interpretation, since Ralph does hold this belief of the man he knows to be named Ortcutt. However, it is true de re as well: of the man we know as Ortcutt, Ralph believes that he is not a spy. Moreover, although the understanding in question is one in which though Ralph is mistaken about Ortcutt’s identity, he is not being portrayed as irrational—as believing both that p and that not-p. The fact that both () and () seem to be true in this way, both involving de re interpretations of the same proper name, gives rise to a puzzle that Quine calls the “double vision problem”. How can these two reports of Ralph’s beliefs de re both be true without entailing a contradiction in his thinking? Intuitively, the problem arises because Ralph is familiar with the same individual, the res Ortcutt, under two distinct guises, reasonably failing to recognize them as guises of the same individual. This suggests that in intensional contexts, the de re/de dicto distinction in NP interpretation is not sufficiently fine-grained to permit us to characterize all the kinds of interpretations attested. Perhaps de re attitudes involve a relation not to an individual simpliciter, the res, but to an individual under a guise. Then the problem is how to reflect the required distinctions between de re interpretations in intensional contexts. Because proper names are usually taken to be rigid designators, denoting the same individual (the actual bearer) in all possible worlds, the double vision problem is especially evident when proper names occur in attitude complements. Closely related problems of de re belief attribution have been pointed out by Kripke () and many others since (see Aloni : Chapter , for an excellent overview of the relevant literature). These puzzles suggest that the understood meaning of a proper name, at

OUP CORRECTED PROOF – FINAL, 17/1/2019, SPi



 

least in such intensional contexts, may be something richer than just the entity which bears it. Another problem with interpretations under the scope of intensional operators is that of how to characterize interpretations de se. Here is an illustration, after a famous example due to Morgan ():⁴ ()

[Context: The baseball player Ernie Banks gets hit on the head and develops total amnesia. He doesn’t know his name or remember anything about his past, though he is lucid. During his long recuperation, he reads in the newspapers about a baseball player named Ernie Banks, and becomes fascinated with the guy’s career. His social worker reports to a nurse:] Ernie Banks thinks he is one of the greatest shortstops of all time.

This is another type of de re interpretation: Banks has a belief about this guy he reads about, a fellow by the name of Ernie Banks. What he doesn’t recognize is that he is that guy, though clearly the speaker knows they are one and the same individual. But despite the speaker’s intended anaphoric relationship between the underlined coreferential NPs, Banks’s evident lack of self-recognition leads to a difference in truth conditions. Hence, we must assume that this leads to a possible difference in interpretation of he. The interpretation on which Banks realizes he is Banks is called the de se interpretation; the one we see in (), the non-de se. Again, these non-de se interpretations generally arise in attitude reports, where the subject of the matrix verb is co-referential with a pronoun in the complement sentence. But we can see a parallel with the so-called referential interpretations of Donnellan (): ()

Smith’s murderer is insane.

()

The man in the corner drinking a martini is my boss.

Consider () uttered in a courtroom by one observer to another. Suppose he says this nodding toward the person on trial, accused of murdering Smith. The fellow is evidently insane, and most people take him to be the murderer. Then, goes the story, the definite description may successfully pick out the accused, even though either the speaker or the addressee may secretly know that the accused is actually not the murderer—one of them may be! Or consider () in a situation in which there is exactly one man in the corner sipping a colorless liquid from a martini glass. Though the speaker knows that it’s actually just chilled water, still this description might seem like the clearest way to draw the addressee’s attention to the man she wishes to refer to, despite the falsity of the description. She’s not trying to mislead, just to be communicatively effective, to say something about that individual.

⁴ As I finish this chapter in , the real Ernie Banks has just passed away in Chicago. RIP Ernie Banks, and thanks for the memories!

OUP CORRECTED PROOF – FINAL, 17/1/2019, SPi

   



Kripke () offers arguments that the so-called referential use does not arise from ambiguity in the definite article itself, but from pragmatic factors. Abbott (: §.–, ) argues against Kripke. There is as yet no consensus on how this interpretation arises. In () and (), there is a difference between the description the speaker believes will successfully pick out the intended referent from the addressee’s point of view and that referent’s actual properties. This intriguingly parallels a feature of the cases of de re interpretation under attitudes: the relevant description (the man in the broad-brimmed hat) or proper name (Ortcutt, the pronoun’s antecedent Ernie Banks) doesn’t accurately describe the intended referent from the point of view of the agent of the attitude (Ralph, Banks himself), just as the speaker knows that the man . . . drinking a martini . . . doesn’t accurately describe the intended referent in (). But in all of these types of interpretation, the ‘incorrect’ definite does successfully pick out the intended referent from the perspective of the addressee. In all these cases, then, it is a shift in perspective which is essential to understanding the intended interpretation. If we understand a definite description like those in () and () not as contributing to the proposition expressed—the proffered content of the utterance—but as helping the addressee to pick out the res which satisfies the definite’s familiarity presupposition (Heim ), then we can understand the sense in which examples like these are true.

.. Context-sensitive predicates Another general type of context-sensitivity displayed by referential NPs involves the use of predicates which are vague or imprecise, whose application is a matter of personal taste, or which require coercion for appropriate interpretation in a given context. Here are some examples: vagueness: () [To the school photographer:] Take pictures of the tall boys in the kindergarten class. imprecision: () I’m going to visit the hexagonal country in western Europe. personal taste: () [To a friend who doesn’t like sweets:] Give me all of those tasty cookies. coercion: () [Context: In a room with only silk flowers.] Put the flowers on the table. () I listened to the rain. To understand what she’s being asked to do, the addressee in (), (), or () must grasp the intended extension of the predicate tall, tasty, or flower, respectively; this will

OUP CORRECTED PROOF – FINAL, 17/1/2019, SPi



 

require making a judgment about what counts as being tall in that class, whose taste is at issue, and what counts as being a flower—certainly not, in the context in (), being the reproductive organ of a plant. () can mean that the speaker is going to France, but only if we grant an imprecise description of that country as being hexagonal. And the speaker in () surely means that she listened to the sound that the rain made falling on the roof or the pavement, quite unlike the interpretation of I listened to the sound that John made. The interpretation of a complex definite NP may also involve a conversational implicature, as we see in: ()

Some of my colleagues and I come to work every day. I drive, but the fellow who rides his bike always gets to the office before me.

()

[Discussing with a friend the people we heard at Karaoke the previous night:] I was most impressed with the man who sounded rather like someone trying to sing “Sweet Caroline”.

The underlined definite in () is clearly intended to refer to the (presumably unique) fellow who rides his bike to work, so that it is relevant to the proposition expressed in the previous sentence. In (), the man who sounded rather like someone trying to sing “Sweet Caroline” seems like an unnecessarily wordy description of someone singing that song. But it tends to suggest that the fellow did a rather poor job. In fact, one might successfully use this description to pick that poor singer out even if during the course of the evening many people sang “Sweet Caroline”, so long as one was particularly bad. Of course, all the above might be attributed to domain restriction on some accounts, as in the Russellian story about definite descriptions discussed above. But even in such an approach, the kinds of contextual factors one needs to take into account to understand the intended domain restriction are varied and subtle—the resolution of intended standards for vagueness, precision, and taste; implicature generation; etc.

. R  ,   

.................................................................................................................................. The wide range of phenomena illustrated above put meat on the bones of Strawson’s claim: if definite NPs are those that are most frequently referential, and if their interpretations are so frequently dependent on context for resolution of anaphora, retrieval of incomplete descriptive content and domain restriction, perspective, and other pragmatic factors, it becomes clear that instead of the NPs themselves referring, speakers use these NPs to refer in context. That is, with most NPs there is no fixed reference outside of context. Even if we take proper names to be unambiguous rigid designators, we have to explain (a) the familiarity constraint on their felicity, discussed above, and (b) how perspective may come to bear on their interpretations in context.

OUP CORRECTED PROOF – FINAL, 17/1/2019, SPi

   



Even more to the point, definite descriptions are regularly and successfully used to refer, yet on almost all analyses (e.g., Russell ; Heim ; Neale ; Elbourne ) they are not referential per se, but either quantificational or anaphoric. Then it seems reasonable to assume that an adequate theory of reference cannot be developed without coming to grips with what constitutes a meaningful utterance of a linguistic expression. Assuming Bar Hillel’s notion of utterance, answering the question of how a speaker successfully conveys her referential intentions involves addressing two others: the first is what types of contextual factors interlocutors regularly draw on in the course of successful reference, across the range of NP-types that give rise to referential interpretations. Addressing this question involves developing a theory of the notion of context of utterance. The second question, building on the answer to the first, is just how such a context interacts with conventional content in the course of interpretation. Here we can only point to trends in the relevant literature which suggest the state of the art. Especially influential early theories of context were developed by Kaplan (); Stalnaker (); and Lewis (a), and were implemented in compositional theories of interpretation, like that of Montague () and his successors. Context in such theories was considered static, given once and for all at the outset of interpretation of an utterance. Following Kaplan (), context was usually expressed as a set of indices of evaluation which fix the values of specific types of indexical expressions (see the discussions in Lewis ; Cresswell ). These might be limited to the so-called pure indexicals (e.g., I, you, now) whose values are fixed by distinguished elements of the context (the speaker, addressee, time of utterance, etc.). But generally in contemporary philosophy of language and linguistic semantics, indexicality is extended to include other anaphoric elements whose antecedents must be contextually retrieved. If they are not quantificationally bound, pronouns and demonstratives unaccompanied by deixis are treated as free variables in logical form, and the indices of evaluation are taken to include an assignment of values (interpretations, referents) to these variables. In these accounts, canonical uses of demonstratives (accompanied by deixis) are assigned values via a distinct deictic index, involving the same kind of mechanism as the assignment of the actual speaker to be the value of I, etc.—fixing a contextual value for pure indexicals. And definite descriptions are treated as quantificational, more or less along the lines of Russell’s () account, with the descriptive content serving as a domain restriction on the existential operator. Hence, in these classical accounts of context-in-reference, there is no unified account of the way in which demonstratives and definite descriptions, like pronouns, may be anaphoric. Note also that no attempt is made in such theories to account for how the intended values are retrieved in a particular context of utterance; the assignment function and other indices are just a way to funnel pragmatically relevant values into the compositional interpretation. The result of this interpretation is assumed to be the truth conditional content of the utterance—a proposition (for an indicative), question (for an interrogative), or suggestion (for an imperative). But after truth conditional interpretation, the context is understood to interact once more with the

OUP CORRECTED PROOF – FINAL, 17/1/2019, SPi



 

compositionally determined content, to adjust it in ways that lead to Gricean implicatures and the like. Stalnaker () argues for a richer notion of the context of utterance than just a set of indices, taking it to be the interlocutors’ Common Ground, which he models as the set of propositions which they take each other to (purport to) believe. He argues that the Common Ground includes not only the information encoded in Kaplan’s indices and the propositions explicitly asserted (and accepted) in the discourse, but all kinds of implicitly shared information as well—the existence of those entities generally familiar to the interlocutors, cultural information, the information to which the interlocutors jointly have perceptual access in the context of utterance, etc. The experimental work of Clark and his group (summarized in ) and of Tanenhaus and his associates (e.g., Chambers et al. ; Hanna and Tanenhaus ; Tanenhaus et al. ; Hanna and Brennan ) argues that access to the Common Ground, understood to include such non-explicit information, plays a central role in anaphora resolution and other referential tasks, and is implicitly assumed by speakers and addressees alike. It is now generally assumed in linguistic semantics and psycholinguistics that the appropriate notion of context should have at least the richness of the Common Ground—a body of information which interlocutors have reason to believe that each other have access to in the context of utterance. This includes a large body of consensual encyclopedic information, as well as the ability to reason over that information to generate implicatures. A richer notion of context is assumed in many dynamic theories of interpretation, cited below, as well as in work developing the notion of context as a conversational scoreboard, due to Lewis (a) (see Roberts a). If we take anaphora, in the broader sense discussed and illustrated above, to amount to a requirement for mere weak familiarity, then the implicit content in the Common Ground plays a crucial role in resolving the intended referent in a very large percentage of uses of all types of definite NPs. For example, weak familiarity clearly plays an important role in the cases involving coreference with a non-linguistically salient referent, as in ()–() above, and in those cases that Clark and Marshall () called bridging anaphora, as in (): ()

. . . John entered the room. The ceiling was painted blue.

Here we readily understand the ceiling to refer to the ceiling of the room John has just entered. Bridging generally involves an implicitly relational interpretation of the head of the definite: ‘ceiling of x’, where x is anaphoric to some contextually salient entity. The speaker can assume that the addressee will be able to retrieve the intended referent, though the relation is merely implicit, because in our society the fact that rooms have ceilings is encyclopedic information in our shared Common Ground. The encyclopedic information in the Common Ground also plays a crucial role in the resolution of vagueness and imprecision, coercion, and other pragmatic factors in the interpretation of predicates. But like Kaplan’s model, Stalnaker’s conception of context is essentially static, still presenting a two-phase view of the interaction of context with conventionally given

OUP CORRECTED PROOF – FINAL, 17/1/2019, SPi

   



content. These two-phase models of the interaction of context with the conventional contribution of the constituents of an expression uttered encounter a number of problems in addressing the kinds of phenomena considered in the previous sections. Consider first the problem of anaphora resolution. Donkey sentences, like those in ()–() in §.. pose a significant problem for theories of static context. In these, the ‘donkey pronoun’ (it in ()) or other definite (the definite article in (), the demonstrative in ()) occurs in the consequent of a conditional; this is also illustrated in () below. Donkey anaphora can also take place when the donkey pronoun occurs in the scope of a quantificational determiner or other operator in the utterance, while its antecedent is introduced in the conditional’s antecedent or the operator’s overt domain restriction, as exemplified in (). But true quantificational binding cannot take place between the antecedent and consequent of a conditional, or between the restriction and scope of an operator more generally, as we see in () and (), which both contrast with () where the binding is available (under the usual c-command relation between quantificational antecedent and bound element): () If a farmer owns a donkey, he should treat it well. () #If a farmer owns every donkey, he treats it well. () Most farmers who own a donkey treat it well. () #Most farmers who own every donkey treat it well. () Every donkey owned by a farmer appreciates the fact that he treats it well. This means that the variable corresponding to a donkey pronoun is free, not bound, and so takes its antecedent from context. But in these examples, the antecedent is nonspecific; there is no donkey under discussion prior to utterance of () or () which can serve as contextually available antecedent. So the arbitrary farmer’s donkey must become familiar under the scope of the conditional, the context temporarily updated to reflect the scenario hypothetically under discussion. This problem led to work on dynamic theories of interpretation, prominently those of Kamp () and Heim (), with subsequent work due to Groenendijk and Stokhof (); Muskens (); and Martin (), among others. In all these theories, the notion of context is dynamic in that (a) it is updated in the course of interpretation, so that (as noted by Groenendijk and Stokhof ) ϕ∧ψ needn’t have the same interpretation as ψ∧ϕ, as we see in (a) versus (b). ()

a. John loved her and he married Jane. [her presumably refers to someone other than Jane] b. John married Jane and he loved her. [her presumably refers to Jane]

And (b) contextual information may be not only updated from utterance to utterance, but also regularly ‘downdated’, so that information made available in the course of utterance under the scope of an operator may no longer be accessible for the

OUP CORRECTED PROOF – FINAL, 17/1/2019, SPi



 

interpretation of subsequent context-sensitive expressions (respecting the Scope Constraint above). We see downdating in (a): ()

a. If John sees a car he likes, he should buy it. He probably took it for a test drive already. b. John saw a car he likes. He probably took it for a test drive already.

Here the donkey pronoun it in the consequent of the first sentence may take a car as antecedent, illustrating the canonical relationship in a conditional with the indefinite in the antecedent. Thus, the context should be temporarily updated as a consequence of the interpretation of the antecedent, with (hypothetical) information that there’s a car John sees and likes, in order to offer an antecedent to it in the consequent. But on the natural assumption that the indefinite in the conditional is non-specific, after interpretation of the conditional the hypothetically updated context is downdated again so that there is no anaphorically available car in the context of utterance for the second sentence. This explains the fact that, the pronoun in the follow-up He probably took it for a test drive already is odd, either (i) infelicitous because without an antecedent, or else (ii) awkwardly forcing the specific reading of the conditional indefinite found in (b), as we saw in the forced specific reading of the bastard in () above, resulting in a revised interpretation of the conditional. To address such problems, Heim used the metaphor of a File to model the information interlocutors have about a particular entity (real or hypothetical); Kamp talks about his Discourse Representations as mental models containing information that’s shared and updated during interpretation. In their theories, the donkey anaphora (and other types of presupposition satisfaction—see e.g., Partee ) is a function of the dynamic update/downdate conventionally associated with conditionals, quantificational determiners, and other operators. One might wonder whether the phenomenon of donkey anaphora is relevant to the discussion of reference, since the canonical examples like () and () don’t seem to refer to particular individuals in the actual world. But an argument can be made that it is relevant to understanding intended reference in use. The reason is that reference is so often successful as a function of anaphora, as we saw in the first section. We also saw in () that an indefinite (a mouse) under the scope of a modal can set up anaphora with a definite description, thus referring to an actual mouse. We can do the same with donkey anaphora: ()

If a car has a manual transmission, then its driver has to know how to use the gear shift. They used to be on the steering column, but mine is on the floor.

In (), the donkey pronoun it takes as antecedent a car, with the gear shift given an implicit relational interpretation, bridging to the car: ‘the gear shift of the car’. Then both the generic bare plural they (referring to gear-shift-kind) and mine in the subsequent sentence are understood to be anaphorically dependent on the gear shift, mine referring to the particular gear shift in the speaker’s car.

OUP CORRECTED PROOF – FINAL, 17/1/2019, SPi

   



So successful reference is arguably, at least in many cases, partly a function of how the context of use plays a role in the successful resolution of anaphora, itself quite a complex phenomenon. Since it involves dynamic update and downdate, the notion of context needed for the retrieval of intended reference arguably must itself be dynamic, modeled as changing in the course of interpretation, in interaction with the utterance’s conventional content. There is much more to say about the other contextual influences noted above— domain restriction, descriptive incompleteness, and shifted perspective—but for reasons of space, here I can say only very little. Domain restriction is generally regarded as presuppositional, typically as anaphoric; for example, Rooth (, ); von Fintel (); and Stanley and Szabo () all assume that NPs like everyone in () and the bastard in ()–() above are associated with an implicit free variable in logical form, whose value is the (group of) individual(s) over which the associated quantificational operator—the universal or existential quantifier associated with the determiner—is intended to range. This is a way of formally encoding the intuition that a competent speaker using these NPs assumes that, in (), there is some salient group of students, or, in (), some salient individual under discussion whom the speaker might dislike, and presupposes that the addressee can recognize that it is this individual or group of individuals which she intends as the domain of every or the existential associated with the definite article. As with anaphora generally, how addressees retrieve the intended (group of) individual(s) to restrict the domain of an operator probably is related to how we understand who a speaker intends to refer to when she says Mary called yesterday to say that she’ll be coming a day later than she’d planned. The proper name Mary is quite common. But the utterance itself implicates that the Mary in question is one the interlocutors know (implicated by the lack of introductory appositive, etc.), and that perhaps they had discussed Mary’s plan to visit. This suggests that the speaker believes that there’s no more than one Mary familiar in the common ground whom the addressees will take to plausibly have the relevant properties. If there were two Marys that were known to be coming, he’d have to say, for example, Mary Russell or Mary Cummins, in order to make the intended referent clear. Similarly in the examples of domain restriction. Like other presuppositions, in order to be felicitous, implicit domain restriction should be obvious in view of two factors in the context of utterance: the interlocutors’ common ground, and the topic of discussion. If we’re talking about what happened in a particular classroom yesterday, then in order to be relevant, the utterance of () should be relevant to that question, and hence relevant to the individuals in that situation. Roberts (/) would argue that the context offers a formal specification of the question under discussion, which would guide the resolution of this evident presupposition (surely the speaker doesn’t mean that absolutely everyone in the whole world should get out their notebooks). The fact that the utterance is imperative suggests that the speaker is an individual in authority—perhaps a teacher, in keeping with the usual classroom scenario. Then the individuals over whom a teacher has authority in a classroom are usually the students; hence the

OUP CORRECTED PROOF – FINAL, 17/1/2019, SPi



 

intended domain for everyone is the set of students in the classroom. Thus, if a parent is sitting in that day, s/he wouldn’t be automatically included in the presupposed domain. Some combination of the relevant question or situation under discussion, the classroom scenario, plus the plausibility of authority involved in the imperative, together (abductively) yield the intended restriction. The issue of descriptive incompleteness is closely related. In fact, if one adopts a Russellian (Neale ), or Fregean (Heim ; Elbourne ) account of definite descriptions one might simply assume that the ‘missing’ descriptive content is given by the domain restriction on the existential operator in the DP’s Logical Form. Elbourne (, ) adopts a variant of the domain restriction view in a situation semantics. He takes all DP interpretations to be restricted to a contextually given situation, and then the unique entity with the overt descriptive content of the definite must be found in that situation. Roberts () argues for a different view, in which definite descriptions themselves are anaphoric, requiring a discourse referent antecedent (roughly, a ‘file’ in the sense of Heim ). On Roberts’s account, the overt descriptive content of a definite description, much like the person, gender, and number on an English pronoun, only serves to guide anaphora resolution. Then, as with domain restriction, the question under discussion and general constraints on plausibility resolve the anaphoric presupposition just in case the overt descriptive content is sufficient to distinguish the intended referent from all others that are relevant at that point in the discussion. Like anaphora resolution, the problem of shifted perspective in intensional contexts is complex. But there is a recent vein in the literature (Aloni ; Percus and Sauerland a, b) that shows promise of illuminating what is at issue. The leading intuition is that an agent may have multiple perspectives on the same res. In some cases, where the perspectives are quite different and the resulting guises under which the res is known to the agent appear quite distinct, this may result in the agent having inconsistent beliefs about the same res. In such cases, the agent fails to recognize that these different guises are in fact guises of the same individual. Then the intended interpretation of the de re definite NP in an attitude context is shifted by a contextually-given perspective operator in such a way that the agent is reported as having an attitude towards the res-under-the-relevant-guise, from a particular conversationally relevant perspective. This, then, is a pragmatic account, rather than one that makes predicates such as believe ambiguous between the usual relation to a proposition and a semantic three-place operator ‘believe-of ’ (Quine ). One question for further research is how interlocutors come to understand that such a shifted perspective is relevant and appropriate in the interpretation of a proper name or other definite NP. Another is about in which types of cases the perspective-shifting operator is contextually available to influence interpretation. For example, one might take the understood interpretation in examples involving so-called ‘referential’ uses of definite descriptions to involve a similar perspectival interpretation. In examples like () or (), even if the speaker takes someone other than the accused to be the murderer, or knows that the man holding the martini glass is

OUP CORRECTED PROOF – FINAL, 17/1/2019, SPi

   



just drinking water, she might use these descriptions to pick out the intended referents because she thinks that is how the addressee views these individuals, or at least that that is how the addressee thinks she views them. The point, then, is purely pragmatic, as argued by Kripke (): the speaker chooses a description not because it is accurate, but because it is most likely to successfully pick out the individual she intends to refer to, given what she knows about the addressee’s beliefs. The notion of perspective arguably plays a role as well in the cases involving de se interpretation: in () there are two ways Ernie Banks might view Ernie Banks—two guises under which Banks might be familiar with Banks: as his own self, or as the famous Chicago shortstop. It is clear that these are in principle independent. The extra factor in these cases is how to model the subjective notion of self in the semantics of attitude predicates. Drawing on Quine (), Lewis (b) proposed the use of centered worlds to model the denotations of complements of attitude reports like that in ().⁵ Each centered world is an ordered pair of a world and an individual in that world, the individual being the entity that the agent of the attitude takes himself to be in that world. A number of subsequent authors have proposed modifications of this general proposal and linguistic refinements; see Chierchia (); Percus and Sauerland (a, b); Anand (); Stalnaker (); and Pearson () for especially relevant discussions and contributions. Roberts (b) weaves this understanding of de se interpretation and the other threads about perspective noted above into a unified account. One final important point: the range of ways considered here in which context influences the understood reference of a definite noun phrase might be taken as reasons to reconsider prominent theories of the semantics of particular types of definites in which they are considered to be directly referential (Kaplan  on demonstratives and indexicals) or rigid designators (Kripke  on proper names). Recognition of the problems which anaphoric and bound variable uses pose for the usual deictic analysis of demonstratives led to King’s () account of demonstratives, an alternative to Kaplan in which they are quantificational. Roberts () offers a different kind of account, in which demonstratives are anaphoric (subsuming the deictic uses as a special case), while Elbourne () argues for a neo-Russellian (‘Fregean’) account in which existence and uniqueness are presupposed instead of proffered, with demonstratives turning out to be much like definite descriptions plus an optional demonstration. Recent work on languages in which the counterparts of English ‘pure’ indexicals can be shifted under the scope of attitude predicates to refer to the agent of the attitude (Schlenker ; Anand and Nevins ; Sudo ; Deal ; and much ongoing work across a wide range of languages) argues that there are, pace Kaplan, monsters! And Wechsler () argues that there is a special, subjective aspect of the semantics of indexicals across all known human languages that cannot be captured by the direct ⁵ Lewis () argues that such examples show that the complement of an attitude predicate cannot be understood as a proposition, but must be a property instead. However, Stalnaker () argues that this is incorrect, and offers a modification of Lewis’s centered-worlds approach to de se phenomena in which the sentential complements do denote propositions, sets of centered worlds.

OUP CORRECTED PROOF – FINAL, 17/1/2019, SPi



 

reference accounts. Roberts (b) offers a new account of indexicality which aims to account for all these problems without Kaplanian Character or direct reference. Similarly, Cumming () argues for a non-rigid, “variabilist” approach to the semantics of proper names, and the anaphoric character of names sketched in the first section suggests there might be something to that view. It is premature to determine the merit of these proposals, but the problems they raise for the direct (or rigid) accounts cannot be ignored. In any case, as we have often seen in the earlier literature (Geach ; Kripke ; Morgan ; Kaplan ; Evans ; Heim ), a better understanding of the attested interpretations of definite NPs in particular kinds of contexts may also, ultimately, lead us to empirically superior accounts of their conventional contents.

A I am grateful to the editors, Barbara Abbott and Jeanette Gundel, for useful comments on an earlier draft.

OUP CORRECTED PROOF – FINAL, 17/1/2019, SPi

  .............................................................................................................

IMPLICATIONS AND APPLICATIONS: PROCESSING AND A C Q U I S I T I O N OF REFERENCE .............................................................................................................

OUP CORRECTED PROOF – FINAL, 17/1/2019, SPi

OUP CORRECTED PROOF – FINAL, 17/1/2019, SPi

        ......................................................................................................................

REFERENCE AND REFERRING EXPRESSIONS IN FIRST LANGUAGE ACQUISITION ......................................................................................................................

  

. I

.................................................................................................................................. T HE field of acquisition of reference and referring expressions has grown into a rich and stimulating mirror of child development issues. Children’s acquisition and use of referring expressions raise fundamental questions because they affect the morphosyntactic, semantic, and pragmatic aspects of language acquisition, their use and their interfaces. After a first period where studies were focused primarily on narratives and other experimental settings, in the last twenty years there has been a significant development of studies conducted on dialogues, both in natural and experimental settings. All these studies flourished from very different theoretical approaches and brought a radical change in the perception of children’s discursive productions and uses of referring expressions. The result is a rather complex landscape of controversial accounts, due both to the diversity of theoretical and methodological options and to the great variety of factors that affect children’s behavior. Orientations, hypotheses, and interpretations of children’s referring behaviors have also been strongly influenced by the fact that first studies concerned Indo-European languages such as English or French. However differences among languages at least regarding morphosyntactic aspects, such as argument expression or structure of noun phrases, can have a significant influence in the acquisition and use of referring expressions. There are some cross-linguistic studies conducted on narratives (Hickmann ; Hickmann, Hendriks, Roland, and Liang ; Hickmann, Schimke, and Colonna ) or on naturally occurring dialogues (Coene ; Rozendaal and Baker ) as well as some studies on bilingual children (Paradis and Navarro ; Serratrice ). But globally, all the studies on naturally occurring dialogues offer an extended view of cross-linguistic

OUP CORRECTED PROOF – FINAL, 17/1/2019, SPi



  

diversity (Allen ; Allen, Hughes, and Skarabela ; Clancy , ; De Cat a; Guerriero, Oshima-Takane, and Kuriyama ; Huang ; Hughes and Allen ; Salazar Orvig, A., Marcos, H., et al. a; Skarabela ; Skarabela and Allen ). The present chapter will mostly refer to European languages and monolingual studies; examples will be drawn from a specific research in French.¹ The chapter is organized as follows: §. summarizes the milestones of acquisition of referring expressions from the morphological point of view; §. presents research on child reference along two main lines, studies on the expression of arguments and studies on determiners. The chapter concludes with a discussion on the controversial results that characterize the field.

. F  

.................................................................................................................................. Children’s first steps in reference are also their first steps in language. In languages such as English or French first words are very frequently nouns,² and full referring expressions progressively emerge with grammatical and syntactical development. In the first stages of language acquisition, children’s utterances usually consist of one or two words or formulaic expressions. These utterances accomplish referential, predicative, or socio-pragmatic functions. Whereas the latter do not involve reference (e.g., bye bye), the former are used to refer to entities and can be bare nouns (see example ()) or demonstratives (in French ça, as in ()): ()

Daniel³ (;) – MLU . Daniel turns around the camcorder pointing or touching every component. DAN sa + sa ‘ça + ça’ ‘that + that’

‘that’s the cassette’ OBS ça c’est la cassette DAN a ‘ah’ ‘ah’ ‘yes’ OBS oui

Nouns (and proper names) are used both in referential and non-referential (e.g., labeling, vocative, or attributive) functions. In example () the child produces a noun ¹ Empirical data in French is drawn from the DIAREF project. See Appendix (§.) for details. ² In fact, at this stage it is not accurate to speak of ‘nouns’ or ‘verbs’ (or any other grammatical category) as there are no specific morphological paradigms to distinguish them. ³ The captions of examples indicate the name of the child, his/her age (years; months), and the Mean Length of Utterance (MLU) for the cited session. Children’s utterances are transcribed phonetically, the interpretation in French is given in inverted commas. The English translation, in inverted commas, appears in the right-hand column. Braces indicate uncertain transcriptions or alternative interpretations. X stands for uninterpretable or inaudible segments. The sign + indicates a pause. In both interpretations and translations, ‘F’ stands for a filler (Peters ; Veneziano ).

OUP CORRECTED PROOF – FINAL, 17/1/2019, SPi

    



(‘lait’ ‘milk’) to introduce a new referent and another (‘cheval’ ‘horse’) to label the object she hands to her mother. ()

Pauline (;) – MLU . Pauline and her mother are playing with different toys. They are currently handling a doll. MOT et si tu lui donnais à boire ‘and what if you give her something to drink?’ ɔ! lɛ ‘oh! lait’ ‘oh! milk’ MOT elle boit du lait elle aussi ‘she also drinks milk’ PAU wi bwa nəlɛ ‘oui boit ne lait’ ‘yes drink(s) F milk’ Pauline picks up a horse and hands it to her mother. PAU ʃəval ‘cheval’ ‘horse’ ‘yes’ MOT oui

PAU

Other utterances consist of predicative terms with null arguments as in example () Pauline’s ‘boit ne lait’ (‘drink(s) F milk’) or in example () Iris’s ‘tombé’ (‘fell/fallen’) uttered about the toy she is playing with. ()

Iris (;) – MLU . Iris and her father are playing with a Mister Potato Head . ‘there you have the ear’ FAT voilà l’oreille ‘ah! mman’ IRI ‘ah! mman’

®

Mister Potato Head slips from her hands. IRI ɔ! tɔ͂be ‘oh! tombé’ ‘oh! fell’ FAT il est tombé. eh bien ramasse-le ‘it fell. well pick it up’ Morphological development of children presents transitional phenomena as protomorphological filler syllables⁴ (hereafter fillers) that can be observed in front of verbs: ()

Daniel (;) – MLU . Daniel points once again to the camera. DAN  a? + sa? ‘ah? + ça?’ OBS  ça c’est la cassette DAN  ətjə. atu ‘{XX}. F tou(rne)’ OBS  mhm

‘ah? that?’ ‘that’s the cassette’ ‘{XX}. F turns’ ‘hum’

⁴ Filler syllables can also be found in prenominal positions (see §...). For discussion of the proto-morphological status of these filler syllables see Demuth and McCullough ; Dressler, KilaniSchoch, and Klampfer ; Kilani-Schoch and Dressler ; Peters ; Veneziano ; Veneziano and Sinclair .

OUP CORRECTED PROOF – FINAL, 17/1/2019, SPi



  

In this example Daniel produces, before the verb ‘tou(rne)’(‘turn(s)’), a filler syllable [a] which cannot be identified either as a demonstrative or as a clitic pronoun. The presence of fillers suggests that children grasp the need to fill the pre-verbal position before acquiring the pronominal paradigm. During this first period, children produce very few clitic pronouns. They are frequently found in quasi-frozen expressions, such as ‘(i)l est où?’ (‘where is it?) in example (): ()

Lisa (;) – MLU . Lisa and her father are playing with farm animal toys. FAT si tu mettais le cheval dedans? ‘what if you put the horse inside ?’ LIS

leu? ‘(i)l est où?’

‘where is it?’

FAT regarde bien y en a plusieurs de ‘look closely there are several of them, chevaux horses’ One of the main characteristics of linguistic development (MLU over ) is the increase of clitic pronouns among referring expressions: ()

Loli (;) – MLU . LOL mama͂ lɛu papa? ‘maman (i)l est où papa?’

‘Mummy where is Daddy?’

MOT t(u) as d(é)jà demandé. ‘you already asked. he is at work?’ il est à son tra(vail)? LOL

sɔ͂byʁo ‘son bureau’

‘his office’

MOT il est à son bureau oui. c’est quoi le ‘he is at his office yes. what is the bureau? qu’est-ce qu’i(l) fait au office? what does he do at the bureau? office.’ LOL

ilkʁava! ‘il trava(ille)!’

‘he is working!’

In this example, Loli, produces two PPs⁵; one in a dislocated construction (‘maman (i) l est où papa?’, ‘Mummy, where is Daddy?’) and the other as the subject of the sentence (‘il trava(ille)! ’, ‘he is working’). Spoken French language presents another clitic pronoun, the demonstrative ‘c’’ in the ‘C’est +X’ construction. This clitic demonstrative

⁵ rd person (clitic) pronouns will be referred to as PPs. In French data, rd person strong pronouns are not very frequent in child language and they will not be specifically addressed in this chapter.

OUP CORRECTED PROOF – FINAL, 17/1/2019, SPi

    



pronoun is a very frequent device for encoding inanimate referents, as we can see both in the adult’s and child’s utterances in example (): ()

Loli (;) – MLU . During snack time, Loli and her mother are talking about a playground. t(u) as vu le dada! attends ‘did you see the horse? wait!’ MOT a͂sɛʁo LOL ‘han c’est (g)ros’ ‘hum it’s big’ c’est trop gros ‘it’s too big’ MOT

Whereas demonstratives (‘ça’ in French, that in English; see Clark and Sengul ) are the very first grammatical referring devices acquired by children, the personal pronoun paradigm is built all along the grammaticalization process. Most studies⁶ date the emergence of first personal pronouns around two years old and consider that children master the whole paradigm (with all its forms and functions) around thirty months. At the same time, significant variability can be observed, which can be explained either by individual styles, which we cannot address here, or by the influence of genres and activities (see §..). Null forms in child language are a recurrent concern in language acquisition research, particularly in languages that do not license null subjects, such as English or French (Allen ; Guasti ; Valian , ). Studies propose complementary theoretical explanations. In formal approaches inspired by generative theory, authors usually oppose ‘competence and grammar-focused’ hypotheses versus ‘performance focused’ hypotheses. According to the first, children have an inner grammar that allows the production of null arguments (Hyams ; Rizzi /; Wexler ). According to the second, children’s grammar does not differ from the target grammar but their production is limited in terms of processing (Bloom ; Jakubowicz and Rigaut  for French). In another theoretical framework, usage based approaches explain omissions or errors on the basis of the distributional features of child directed speech without postulating an innate grammar (Freudenthal, Pine, and Gobet ; Theakston, Lieven, Pine, and Rowland ). However, as will be discussed below, it is necessary to consider also discourse-pragmatic factors to account for the use of null forms and fillers in the first stages of language acquisition. In parallel with the pronominal paradigm, children progressively acquire the determiner paradigms. The acquisition of determiners is also a gradual process involving omission, filler syllables, fluctuation, and (nearly) systematic presence of adult-like forms in mandatory contexts (between ; and , see Bassano, Maillochon, and Mottet ; Rozendaal and Baker ; Salazar Orvig, Marcos, et al. ). However, the

⁶ For French see Hamann, Rizzi, and Frauenfelder ; Jakubowicz and Rigaut ; Salazar Orvig, Marcos, et al. a; for both French and English Girouard, Ricard, and Gouin Decarie ; PérezLeroux, Pirvulescu, and Roberge , ; Pirvulescu .

OUP CORRECTED PROOF – FINAL, 17/1/2019, SPi



  

rhythm of acquisition varies with languages. It is more precocious in French and in other Romance languages than in Germanic ones (Guasti, Gavarró, de Lange, and Caprin ; Rozendaal and Baker ). In languages that have the definite/indefinite contrast (as English and French), definite and indefinite determiners make up the majority of determiners used by the youngest children (Coene ; Rozendaal and Baker ; Salazar Orvig, Marcos, et al. ) as well as by adults. In the first stages definite determiners precede indefinite ones by far. With linguistic development indefinite determiners grow to the same values as definite ones. As with pronominal acquisition, children’s development is characterized by great variability that can be accounted for in terms of both individual styles and influence of genres and activities.

. F   

.................................................................................................................................. The construction of grammatical paradigms simultaneously involves the acquisition of referential functions. The characteristics and development of argument expression (and choice of nominal phrases) will be presented before addressing the issue of the functional dimension of determiners acquisition.

.. Argument expression How do children express referents? This question is often regarded as reflecting a choice⁷ between overt and covert arguments, on the one hand, and between lexical or grammatical forms, on the other. In a seminal study, Greenfield (Greenfield ; Greenfield and Smith ) showed the role of informativeness in the choice of the situational element that was encoded. In her account, informativeness was to be understood from the perspective of the child. Studies that consider the issue within the framework of a cognitive conception of reference discuss children’s uses in terms of information structure, accessibility, or givenness. Other studies have been inspired by textualist perspectives and very frequently considered reference in child language in terms of cohesion. This dividing line partially coincides with the distinction between studies on naturally occurring dialogues on one side, and studies on narratives on the other; experimental studies may fall within one or the other of these approaches.

⁷ All along this chapter we use some debatable terms “to prefer”, “to choose”, “preferred”, and “choice” that could imply that children, and adults, consciously pick a referring expression and discard others. As we stand for a far more conservative position and do not attribute such conscious behavior to speakers, these terms should be read as describing a purely statistical reality.

OUP CORRECTED PROOF – FINAL, 17/1/2019, SPi

    



... Studies on narratives Studies of narratives by children aged two to five (Bennett-Kastor ; Peterson and Dodsworth ) show that even the youngest ones use the third person pronoun much more frequently to reiterate a referent than to introduce a new one, although the authors also mention inappropriate or ambiguous productions. In many studies on narratives, early uses of pronouns (as well as definite articles, see §...) are described as being deictic or exophoric. Hickmann () and de Weck () considered that as young children mostly speak of present referents, the head of a co-referential chain and the subsequent mentions are necessarily deictic. KarmiloffSmith’s () study on narratives of English and French speaking children stemmed from a similar perspective. Her results showed that before the age of six, children had a deictic strategy which did not involve any discursive representation. Children narrated the story as if they were describing unlinked pictures; pronouns were often associated with pointing gestures; their reference was not discourse but stimulus driven. During the second stage, children adopted a ‘thematic subject strategy’ which corresponds to a first, simplified form of discursive representation, reduced to the main character. Only after nine, do children adopted an adult-like anaphoric strategy where they introduced linguistic contrasts according to the narrative roles of the various discourse objects. At this stage children manage both the relation to the referent and the characters. However in a study on German speaking children, Bamberg () showed that the thematic subject strategy appeared earlier (at ;) when children had previously looked at the picture book.

... The choice of referring expressions in naturally occurring dialogues In the last twenty years there has been a great development of studies on naturally occurring dialogues that has provided evidence for a more subtle understanding of children’s behavior. These studies converge to show that young children predominantly use weak referring expressions (null forms, fillers, unstressed and clitic pronouns) to encode ‘in focus’ or highly accessible referents, and strong referring expressions to encode less accessible referents. This trend has been observed in several languages such as English (Gundel and Johnson ; Gundel, Ntelitheos, and Kowalsky ; Hughes and Allen ), French (Salazar Orvig, Marcos, et al. a, b), Italian (Serratrice ), Inuktitut (Allen ; Skarabela ; Skarabela and Allen ), Japanese (Guerriero et al. ), Korean (Clancy ), and Mandarin (Huang ). Gundel and Johnson () compare the use of rd person and demonstrative pronouns and demonstrative, definite, and indefinite NPs, following the Gundel et al. () Givenness Hierarchy in a corpus of nine children aged between ; and ;. The children used % of the PPs for in-focus referents, whereas demonstrative pronouns were preferred for referents that were activated, but not in focus. With a different methodology, Hughes and Allen () studied (in a corpus of four children

OUP CORRECTED PROOF – FINAL, 17/1/2019, SPi



  

aged between  and  at two developmental moments) the choice between four forms (null/PP/demonstrative pronoun/lexical noun) in the subject function. They showed that PPs and null forms were preferred for prior mentioned referents and/or in the context of joint attention. They also showed an evolution between Time  and : in Time , children produced mostly null subjects when the referent was under joint attention or had been previously mentioned whereas in Time , they prefered PPs. Cross-sectional and longitudinal results for French replicate both Gundel and Johnson’s () and Hughes and Allen’s () for English. Whereas dislocations, nouns, and strong demonstratives are more often used to introduce a new referent (see examples (), ()) or to reactivate a referent mentioned in a previous topical sequence (see example ()), the overwhelming majority of children’s PPs encode a previously mentioned referent (see example (), and ()). Null forms (see example (), fillers () and, to a lesser extent, clitic demonstratives () alternate with PPs in this last context. At the same time, all data show some few cases of PPs or other weak referring expressions for referents that have not been previously mentioned. These cases have been usually considered as evidence of a preferred deictic behavior by children or of the lack of sensitivity towards perceptual availability, for example. But the study of naturally occurring dialogues shows the need to consider also the non-verbal dimension of the ongoing activity (as well as other dialogical factors; see also §. and §.). For instance, when the only criterion to consider that a referent is in focus is the fact that it has been mentioned in the immediate previous utterances (by the child or by her interlocutor), we miss the fact that participants can be jointly attending to non-verbalized entities. Skarabela and colleagues (Skarabela and Allen ; Skarabela, Allen, and ScottPhillips ) have explored the way joint attention affects the choice of referring expressions, independently of discourse givenness. They showed that Inuktitut speaking children tended to omit arguments or use clitic demonstratives in the context of joint attention and to use lexical expressions or independent demonstratives in the absence of joint attention. The same trend can be observed in data from French speaking children. The rare cases when rd person clitic pronouns are used for first mentions correspond to objects that are clearly under joint attention as it can be seen in example (). ()

Lisa (;) – MLU . Lisa is playing with a puzzle, her father is watching her. She places the pieces on the board but does not put them in the right place. Having done so, she looks at the puzzle. LIS alɛfini ‘ah! l’est fini’ ‘ah! it’s finished’ FAT tu crois que c’est fini? tu ‘you think it’s finished? you think crois que c’est comme ça que ça va that’s the way it’s supposed to be, Lisa? Lisa?’

The child uses the form ‘l’ for ‘il’ (‘it’) to first mention the puzzle she has been doing. This puzzle is clearly under joint attention, as can be inferred from the fact that the father does not ask for clarification.

OUP CORRECTED PROOF – FINAL, 17/1/2019, SPi

    



When young children mention referents which are not under joint attention but part of a wider shared knowledge, they prefer nouns and demonstratives (Gundel and Johnson ; Salazar Orvig, Da Silva, et al. ; Skarabela et al. ). In example () the child uses ‘ça’ (‘this’) when she picks up and presents new Lego pieces to her mother.

®

()

Elodie (;) – MLU . Elodie and her mother are playing with Lego . E. picks up various Lego pieces one by one. ELO  ale +++ esa ++ esa ++ isi kɔmsa § + isi sa + esa ‘allez +++ et ça +++ et ça ‘there +++ and this +++ and this ++ ici comme ++ here like that + here this ça + ici ça + et ça’ + and this’

®

Finally, occurrences of PPs for non-retrievable referents are very rare and children do not use pronouns to first mention a non-familiar referent more often than adults do.⁸

... The choice of referring expression in experimental settings Several studies investigated the choice of referring expressions by children in controlled dialogical contexts (Campbell, Brooks, and Tomasello ; Matthews, Lieven, Theakston, and Tomasello ; Salomo, Graf, Lieven, and Tomasello ; Salomo, Lieven, and Tomasello ; Serratrice , ; Wittek and Tomasello ). They explored the influence of discursive factors, such as the previous mention of the referent by the interlocutor, versus the influence of perceptual availability. They also addressed the impact of the type of question the children were asked (see §..). These studies converged to show that discursive context affects children’s choice of nouns versus pronouns and/or null forms. Campbell et al. () showed that children aged ; to ; preferred null forms and/or pronouns to respond to questions (What did X do?) that mentioned the referent, whereas they would use nouns and pronouns after an open question (What happened?). This influence of the preceding discourse context was confirmed for German by Wittek and Tomasello (), who showed that questions about the target referent were answered by children aged ; to ; with null forms or pronouns whereas questions that did not mention the target referent and that revealed no knowledge of it were answered with nouns. However, younger children (around ) were less influenced by the discursive context. On the other hand, these authors showed that children aged ; to ; were not sensitive to the perceptual availability of the referents. Matthews et al. () further explored this issue and confirmed for English speaking children (aged  and ) that children are sensitive to prior mention of the referent and not to its perceptual availability for the interlocutor. They showed that nouns were very frequently used when the referent had not been previously mentioned. Salomo and colleagues (Salomo et al. , ) worked on the expression of the patient in ⁸ For example, .% for adults and children in the DIAREF corpus.

OUP CORRECTED PROOF – FINAL, 17/1/2019, SPi



  

German through action focused questions. In a first study (Salomo et al. ) they showed that when only action changes, children (mean age ;) tended to use only a verb whereas when the patient was the new element they expressed it with a lexical NP. However, they also found that children expressed perceptually new patients more often when the experimenter had provided a discourse context previous to the question. The second study (Salomo et al. ) explored further the role of discursive context and perceptual availability for the expression of patients. Contrary to previous studies they found that children can take the interlocutor’s perspective when a conflict between previous discourse and perception arises.

... Competing referents in discourse One of the reasons why narratives are more challenging for children than everyday dialogue is probably because they require management of several different coreferential chains throughout a long discourse. In French, young children use significantly more PPs, omissions, and fillers than nouns for topical continuity (as defined in Lambrecht ) in subject position (Salazar Orvig, Marcos, et al. b). Studies on older children’s narratives (Hickmann ; Jisa ) showed that pronouns are preferred when the character is maintained as a subject throughout utterances. For younger children (e.g., in the DIAREF data) more than % of the PPs encode a referent maintained in the subject function whereas more than % of the nouns encode a referent maintained in a non-subject function. As later in narratives (Karmiloff-Smith ), very young children use dislocations to contrast two referents in the subject function. Alternation in syntactic function is usually linked to the fact that speakers have to deal with competing referents. This is clearly the case in narratives when several characters alternate at the forefront of events. Adults and older children deal with contrasting topics by shifting weak (zero and pronouns) and strong expressions (stressed pronouns, lexical forms, dislocated and cleft constructions). In Karmiloff-Smith’s model (), this corresponds to the ‘anaphoric strategy’ by older children whereas Bamberg () showed that younger children (;) contrasted pronouns for the main character and nouns for referents that did not contribute to the thematic progression. In naturally occurring dialogues, the contrast might not be managed in the same way. Hughes and Allen () showed for English that, even though adults preferred lexical forms, they also used pronouns (about %) when dealing with competing referents. In their study, children at Time  resorted to a more heterogeneous range of expressions and developed towards more adult-like uses in Time . Dealing with potential ambiguity also depends on whether the referents are human versus non-human or animate versus non-animate. In the DIAREF data, children and adults showed a similar trend: PPs were more often used for animate referents (%) whereas nouns and demonstratives were preferred for inanimate referents (% and % respectively for the children, % and % for the adults). But Allen and her colleagues (Allen, Skarabela, and Hughes, ; Hughes and Allen ) found, for various

OUP CORRECTED PROOF – FINAL, 17/1/2019, SPi

    



languages, great variability among children: some preferred null forms or pronouns for animate referents whereas others preferred lexical forms or demonstratives. In narratives, Bamberg () showed that German speaking children tended to contrast the human character—a boy—and the main second character—a dog—by preferring pronouns for the first and nouns for the second. In contrast, when competing referents share the same features, the narrative role constrains the choice of referring expressions: in that case, pronouns are devoted to main characters (Hickmann ). In an experimental setting, (Serratrice ) showed that children (mean age ;) tended to use pronouns in the absence of competing referents, or when the other referent was inanimate, and to avoid pronouns when the competing referent was also animate. Distance from the previous mention of the referent is another factor affecting the choice of referring expressions (Givón ). For instance, in the DIAREF corpus, pronouns were predominantly used (% for adults and % for children) in the context of immediate continuity (within the last four turns). Studies on French narratives (Hickmann ; Jisa ) have shown that adults and children (aged –, , and ) prefer pronouns for subsequent co-referential mentions whereas nouns are used when the character has not been mentioned in the preceding utterance.

.. Determiners Even if they are preferred for new and reintroduced referents, full lexical NPs appear in various contexts in children’s discourse. A simplistic vision of child communicative experience would suggest that infants and young children only deal with specific, concrete, and present referents, and that other types of reference are late acquisitions. In fact, children experience both referential and non-referential uses of nouns from the onset. One of the first and most frequent non-referential uses of nouns is labeling (Ninio and Bruner ) as shown in example (). () Pauline (;) – MLU . MOT  c’est quoi ça Pauline? MOT  c’est un monsieur. oui c’est un monsieur Pauline shows another figure PAU  møsjø, møsjø ‘monsieur, monsieur’ MOT  non ça c’est la vache PAU  møsjø ‘monsieur’ MOT  oui c’est un monsieur, c’est bien PAU  waty ‘voiture’ MOT  voiture. oui c’est une voiture tu connais bien ça

‘what’s that Pauline?’ ‘it’s a man. yes it’s a man’ ‘man, man’ ‘no. that’s the cow’ ‘man’ ‘yes. It’s a man. it’s fine’ ‘car’ ‘car. yes it’s a car. you know that well’

OUP CORRECTED PROOF – FINAL, 17/1/2019, SPi



  

The focus of the discussion between mother and daughter in this sequence is labeling the figures they are handling. It can be observed, incidentally, that even if the indefinite determiner is preferred for labeling, adults (and children) also use definites (MOT ) (see also Rozendaal and Baker ). Labeling is not the only non-referential use of nouns. Naturally occurring dialogues show various uses of nouns in idioms (prendre un bain ‘take a bath’; avoir peur ‘to be scared’) as well as in attributive functions as in example () where mother and child use the noun banane (‘banana’) to qualify a “petit Suisse” dessert. ()

Julien (;) – MLU . MOT : regarde un petit suisse MOT : ‘oh’ epetisisis JUL : ‘F petit suisse’ MOT : un petit suisse à la banane  JUL : ‘banane’

‘look a “petit Suisse”’ ‘oh’ ‘F “petit Suisse”’ ‘a banana “petit Suisse”’ ‘banana’

Moreover, through the adult’s discourse, children are exposed to non-specific reference and even generic reference, as can be seen in example (). ()

Maxime (;) – MLU . Maxime and his mother have finished playing with a puzzle. They are putting the pieces away in the box. MOT  et les têtes? et les têtes des lapins? ‘and the heads? and the rabbit heads?’ MAX  wɛjelatɛtlapɛ ͂ ‘où est elle la tête lapin?’ ‘where is the rabbit head’ ‘the heads. {XX} ‘that’s the heads’ MOT  les têtes. {XX} c’est ça les têtes MAX  awɛ. la? ‘ah ouais. ‘là?’ ‘oh yeah. here’? MOT  ça a des grandes oreilles un lapin ‘a rabbit has big ears, see?’ hein? MAX  wi ‘oui’ ‘yes’ ‘you, do you have ears’? MOT  toi tas des oreilles? MAX  wi ‘oui’ ‘yes’ ‘where are your ears’ MOT  elles sont où tes oreilles MAX  la++latavy? ‘la++la t’as vu?’ ‘here ++ here did you see’ ‘yes I saw’ MOT  oui j’ai vu

The sequence begins with an utterance about some concrete pieces representing heads of rabbits. The identification of one piece gives the occasion to the mother to refer

OUP CORRECTED PROOF – FINAL, 17/1/2019, SPi

    



generically to the class of rabbits and their ears (ça a des grandes oreilles un lapin hein,⁹ ‘a rabbit has big ears, see? ’). The second step for the mother is to bring up, through questioning, a specific referent, first in a indeterminate mode (toi t’as des oreilles? ‘you, do you have ears?’) before getting back to a specific one (elles sont où tes oreilles, ‘where are your ears’). During this short sequence the mother indirectly shows to the child that the same noun can have three referential statuses. Last, discussions about absent referents do emerge in the first stages of language acquisition (Ninio and Snow ; Veneziano and Sinclair ). Absent referents are very often familiar persons (like father, grandmother, friends), but they can also be less familiar and fictional characters. In example () the child refers to two fictional characters: a prince and the mother of the main character, Poucelina. () Léa (;) – MLU . Léa and her mother are talking about a fictional character, Poucelina (Thumbelina). ɛlaɛ ͂pɛ ͂s LÉA  ‘elle un prince’ ‘she has a prince’ elle est grande Poucélina? ‘is Thumbelina big?’ MOT  ‘m:’ ‘m:’ LÉA  nɔ͂ɛlepətit. LÉA  ʃamama͂ɛetutpətit.pətitʃamama͂ ‘no she is small. her mother ‘non elle est petite. sa maman is really small. small her est toute petite. petite sa maman’ mother’ The mastery of determiners for these different modes of reference has been differently explored in naturally occurring dialogue and in experimental settings. The next sections present first the main results of studies from experimental settings and then studies from naturally occurring dialogues.

... Studies in experimental settings (narratives and other eliciting contexts) First studies on the acquisition of referential values of determiners in the s explored the contrast between definite and indefinite determiners. Maratsos () found that young children do distinguish non-specific reference (marked by an indefinite determiner) from specific reference, but can use the definite determiner for first mention of a referent (see also Power and Dal Martello ; Warden ). Karmiloff-Smith () explored—in a wide range of conditions—both production and comprehension of determiners by French speaking children. She concluded that definite and indefinite determiners develop in parallel and that the definite/indefinite contrast for specific referents is a late acquisition: at the onset, the indefinite is mainly used for labeling and the definite has a deictic value. Further on, definite determiners ⁹ This kind of dislocation is a frequent form to express generic reference in French (Charolles ).

OUP CORRECTED PROOF – FINAL, 17/1/2019, SPi



  

shift to exophoric values. In a third stage, indefinites are used to introduce new specific referents and anaphoric values appear for definites. These studies converge to consider that young children tend to make ‘egocentric’ errors, that is, they do not seem to take their interlocutor’s knowledge into account and overuse definites in contexts where the referent is new for the interlocutor. Schaeffer and Matthewson () who found—in a dialogue elicitation task—that English speaking children tended to use % of definites for referents that were not mutually shared. These results could be consistent with the more general assumption that young children process reference egocentrically or lack a theory of mind (Perner ). However, some studies show contradictory results and raise the question of the influence of the experimental conditions on children’s production (Emslie and Stevenson ) or the interpretation of the acquisition sequence (Zehler and Brewer ). The developing contrast between determiners has also been investigated in narratives. Consistently with the idea that first determiners have exophoric values, Hickmann () considers also that there is not a real contrast between definite and indefinite determiners for specific reference. In her data, four-year-olds used indefinite forms to introduce a new referent in labeling utterances. Other studies investigated whether young children were sensitive to shared knowledge when telling a story. This issue was addressed by contrasting shared knowledge and non-shared knowledge conditions, for example, children telling a story to a blindfolded or to a ‘naïve’ experimenter. Young children tended to use definite determiners to introduce referents even in the absence of shared knowledge (De Cat ; de Weck and Jullien ; Kail and Hickmann ; Kail and Lopez ; Schaeffer and Matthewson ). These results were not confirmed by other studies (Schafer and de Villiers ; van Hout, Harrigan, and de Villiers ) that did not find definite determiners for totally new referents in discourse. Van Hout et al. () found that children did use definite determiners when the referent had been established in the discourse as one of a group. This suggests that the tendency to be coherent interferes with the determined or nondetermined nature of the reference.

... Determiners in naturally occurring dialogues In contrast with studies in experimental contexts, studies on naturally occurring dialogues have shown that children under three exhibit a wide range of referential values and determiner distinctions that are relatively close to those of adult usage (De Cat b; Gundel et al. ; Rozendaal and Baker ; Salazar Orvig, Marcos, et al. , ). Analyses conducted in every day conversational contexts with children under three (Gundel et al. ; Ochs Keenan and Shieffelin  for English, and De Cat b for French) showed that young children acquire determiners with contrasted values. Rozendaal and Baker ()—who compared French, Dutch, and English—found evidence of some degree of sensitivity in French-speaking children—and to a lesser degree in English-speaking children, unlike Dutch ones—to differences in shared knowledge.

OUP CORRECTED PROOF – FINAL, 17/1/2019, SPi

    



When indefinites appear and grow in frequency, they progressively replace fillers and omissions in the contexts of non-specific and non-referential uses as well as in first mentions of specific reference. Moreover, even if all studies show that non-mutually known referents are seldom mentioned in naturally occurring dialogues, Frenchspeaking children seem to prefer indefinites to introduce them (Rozendaal and Baker ; Salazar Orvig, Marcos, et al. ). On the one hand, Salazar Orvig, Marcos, et al. (, ) and Rozendaal and Baker () confirmed that indefinite forms are clearly associated with labeling and nonspecific referencing functions whereas definite ones are assigned the value of discourse given. If we take the example of French acquisition, it appears that, considering omissions and fillers as the first forms produced by children, definite articles emerge as the first determiners. They are mostly used for previously mentioned and mutually known referents. Children do not use definite determiners in contexts where indefinites would have been the only appropriate determiner, as can be observed in example (): () Margaux (;) – MLU . MAR  vaseseɛ ͂sukula ‘va chercher un chocolat’

‘(I’m) going to look for a chocolate’ ‘oh Margaux!’

MÈR  oh, Margaux! MAR  vaseseɛsukula + vasɛsisukula + vasɛsisukula? ‘va chercher Fchocolat+ va chercher ‘(I’m) going to look for Fchocolat + va chercher Fchocolat’ Fchocolate+ going to look for Fchocolate+ going to look for Fchocolate’ ‘well no. you have eaten MÈR  ben non, t’as mangé assez de chocolat enough chocolate’ ‘yes’ MAR  ‘si’ ‘non’ MÈR  non ‘yes {xxx}’ MAR  ‘si {xxx}’ MAR  jaɛfil ‘(il) y a Ffil’ ‘there is a thread’ ‘is there a thread?’ MÈR  y a un fil? MAR  alamɛ ͂ ‘à la main’ ‘in the hand’ ‘is there a thread in the hand?’ MÈR  y a un fil à la main? In this sequence, three different nouns (‘chocolat’, ‘fil’, and ‘main’ - ‘chocolate’, ‘thread’, and ‘hand’- one of which is repeated three times in a second turn) illustrate three kinds of reference. The first corresponds to non-specific reference (any individual of the class of chocolates). In this case, the indefinite determiner alternates with a filler syllable. The second corresponds to the first mention of a referent. Even if the thread is perceptually available to her mother, the child is drawing the attention of her mother to it. The child uses once again a filler and not a definite determiner. Finally, the third is a clearly mutually known referent. The child uses the definite determiner ‘la’.

OUP CORRECTED PROOF – FINAL, 17/1/2019, SPi



  

There are not many studies on the use of demonstrative determiners by children. In French demonstratives are among the last determiners to be acquired. For English, Gundel and colleagues (Gundel and Johnson ; Gundel et al. ) observed on the contrary that demonstrative determiners were acquired before definite and indefinite articles, and that children used more demonstrative determiners than adults. For them, this overuse of demonstrative determiners could be due to less sensitivity of children to relevance and scalar implicature.¹⁰

.. To summarize Quite distinct results can be observed from these different sets of studies. Studies on narratives consider that anaphoric reference is a late acquisition and that first uses of referring expressions and determiners are deictic and/or egocentric. Studies on naturally occurring dialogues have opposite results and show a very early competence in the use of referring expressions and determiners. Experimental settings show the role of prior context and point out the difficulties that children can experience in dealing with perceptual availability. Conflicting results could be due to the differences of dialogical contexts. It has been often stressed that dialogues provide children with a scaffolding context in which they are not faced with challenges such as the absence of mutual knowledge and diverging perspectives which would reveal cognitive difficulties. This issue is addressed in the next sections.

. T   

.................................................................................................................................. A significant factor that can account for the divergence between results in experimental settings and naturally occurring dialogues is the dialogical/interactional nature of the latter, which has so far not been thoroughly explored, either in experimental settings or in familiar interactions. Researchers consider the dialogical dimension of reference in different ways. Cognitive models of reference (Ariel ; Gundel et al. ; Cornish ; Charolles ) assume that reference necessarily involves projecting the interlocutor’s representation of the ongoing discourse and context, his or her knowledge, beliefs, expectations, needs, and intentions. This projection of other’s representation takes place in

¹⁰ As demonstrative determiners are not very frequent in young children’s discourse, this phenomenon is not observed in French naturally occurring dialogues. However, if this analysis were to be transposed to French it would concern possible overuse of definite determiners.

OUP CORRECTED PROOF – FINAL, 17/1/2019, SPi

    



actual interaction. Thus reference like other aspects of discourse is a matter of common construction (Bakhtin ), consisting of the mutual contributions of the interlocutors that ensure the common ground (H. H. Clark ; E.V. Clark ). In the conversation analysis framework (Fox ; Pekarek Doehler ; Pekarek ), referring expressions are considered as resources speakers use in interaction for positioning and/or displaying their understanding of the conversational sequence. Finally, referring expressions are never produced in isolation. They are always produced within activities, that pertain to what Wittgenstein called language games “consisting of language and the actions into which it is interwoven” (Wittgenstein : §). These activities generate specific configurations of language use, usually called discourse genres (Bakhtin ). Our contention is that young children acquire values and uses of referring expressions through these genres and activities. We will successively examine the way these aspects of language use influence young children’s use of referring expressions.

.. The influence of the interlocutors’ discourse In a naturally occurring dialogue corpus (the DIAREF corpus) more than % of the children’s referents were immediately mentioned by the interlocutor in the same topical sequence. This prevalence grew to more than % when it came to PPs. Together with the fact that the referent is in focus, it is necessary to consider the possible influence of its previous encodings. Repetition and question–answer relations are among the most frequent issues investigated. For instance, in dialogues, repetition and answer can account for a significant proportion of children’s referring expressions (more than % in the DIAREF corpus, see Salazar Orvig, de Weck, et al. (in preparation); Salazar Orvig and Morgenstern ).

Questions Cameron-Faulkner, Lieven, and Tomasello () showed that questions account for % of the utterances mothers address to their two-year-old children. As mentioned in §..., some experimental studies have addressed the influence of questions in children’s choice of lexical versus pronominal and/or null forms. Most of these studies (Campbell et al. ; Matthews et al. ; Wittek and Tomasello  for English and for German) contrasted predicate focused questions (What did X do?) and sentence focused questions (What happened?). They showed that English and German speaking children aged ; to ; preferred null forms or pronouns in the first case and explicit or strong arguments in the second. Matthews et al. () showed that even very young children (two years old) were sensitive to the focus of the question (subject or predicate focused) suggesting an early competence for determining topic and focus in discourse.

OUP CORRECTED PROOF – FINAL, 17/1/2019, SPi



  

Considering that previous studies did not allow for disentangling the respective influence of discursive and perceptual cues, Serratrice () designed a study to assess the weight of each factor (for children aged , , and ). Even if six-year-olds still experienced difficulties dealing with perceptual cues when not scaffolded by discourse cues, she confirmed that children (ages , , and ) were sensitive to the type of questions (sentence versus predicate focus question; see also Salomo et al. , ). This influence did not depend on whether or not the target referent was encoded in the question. Let’s turn to the influence of questions in naturally occurring dialogues. In the DIAREF corpus, % of the referring expressions were produced by children when answering a question. However, nouns and pronouns did not appear in the same position: nouns were used when elicited by the question, while pronouns were most frequently used when the referent was the topic of the question, as illustrated in (): ()

Pauline (;) – MLU . MOT qu’est-ce (que) tu veux manger maint(e)nant? PAU yn bʁiɔʃ ‘une brioche’ MOT elle est où la brioche? PAU ɛlɛdədã ‘elle est dedans!’

‘what do you want to eat now?’ ‘a brioche’ ‘where is the brioche’ ‘it’s inside’

The first question elicits a noun ‘brioche’ in focus position, whereas in the second one, the referent is the topic, and the child answers with a simple clitic pronoun.

Repetition Repetition fulfills multiple functions in first dialogues (Benoit ; Clark and Bernicot ; Keenan ; Keenan and Klein ; Salazar Orvig ), including cohesion, ratification, and establishing common ground. ()

Léa (;) – MLU . In a discussion about the child’s toys, her mother shows a baby boy. ‘and that’s?’ MOT  et ça cest? bebe LÉA  ‘bébé’ ‘baby’ MOT  bon ( . . . ) ‘good’( . . . ) the mother shows the doll to the child again, ‘do you think he is clean?’ MOT  est-ce que tu trouves qu’il est propre? LÉA  wi ‘oui’ ‘yes’

OUP CORRECTED PROOF – FINAL, 17/1/2019, SPi

     MOT  non, mais regarde comme il est sale, tu lui as pas donné son bain? hein. u? LÉA  ‘où?’ MOT  il est sale là. + regarde là wi ilesal LÉA  ‘oui, il est sale’



‘no. but look how dirty he is. you haven’t given him a bath, have you?’ ‘where?’ ‘he is dirty there. + look there’ ‘yes he is dirty’

In LÉA  the child reproduces her mother’s use of the PP. However, repetition does not affect the use of nouns and pronouns in the same manner: other repetition accounts for % of the nouns and only % of PPs.

.. Positioning Example () above also shows that repetition has another function in dialogue. In this case, it conveys an acknowledgment of the mother’s assessment after a short disagreement. Shifting referring expressions can also convey other sorts of positioning, like opposition or perspective confrontations. This is the case in the next example. () Cécile (;) – MLU . FAT  est-ce que tu crois que la vache elle va rentrer dans la voiture? CÉC  nɔ͂ ‘non FAT  c’est pas permis ou elle est trop grosse? CÉC  i i i pœpa mɔ͂te ‘i(l) i(l) i(l) peut pas monter’ FAT  ah i(l) peut pas monter CÉC  i ‘i(l)’ FAT  ben c’est pas possible elle est trop grosse CÉC  wi ‘oui’ FAT  il faudrait que ce soit un bébé peut-être. un bébé de la vache peut-être qu’il pourrait rentrer, mais la vache est trop grosse. et le bonhomme tu crois qu’il peut rentrer le bonhomme?

‘do you think that the cow will go into the car?’ ‘no’ ‘it isn’t allowed or is she too big?’ ‘it it it can’t go inside’ ‘oh it can’t go inside’ ‘It’ ‘well it’s not possible. It’s too big’ ‘yes’ ‘it’d better be a baby maybe. a baby cow would probably enter. but the cow is too big. and the man, do you think that he can fit in, the man?’

OUP CORRECTED PROOF – FINAL, 17/1/2019, SPi



   CÉC  øø. wi wi Ipœpa mɔ͂te ləbɔnɔm. bɔnɔm pœʁa͂tʁe ‘euh euh oui oui. il peut pas monter le bonhomme. bonhomme peut rentrer’ FAT  il est un peu gros lui CÉC  ieɛ ͂pœgʁɔ ‘i(l) est un peu gros’

‘hmm. yes yes. he cannot get in, the man. man can fit in’ ‘he is a bit big’ ‘he is a bit big’

In this sequence we can observe how nouns (alone or in dislocation) are used for the two in-focus referents: the cow and the man. In Céc  the child uses the dislocation (‘il . . . le bonhomme’, ‘he . . . the man’) when she gives a negative answer to her father, and once again the noun when adopting the opposite position. On the other hand, when her answers convey agreement (in CÉC  and in CÉC ), she uses a PP. Salazar Orvig, Marcos, et al. (b) investigated the choice of referring expressions for subjects in dialogue. PP, nouns and dislocations and strong demonstratives were categorized according to the move performed by their current utterance with respect to the previous discourse. Results showed that PPs were mainly used for plain continuity whereas lexical forms (alone or dislocated) and strong demonstratives were preferred when the utterance drew a contrast with the preceding forms. This kind of use was also observed by Bittner () for German-speaking children, who used complex definite phrases in the context of emphasis or contrast.

.. Discourse genres and activities The influence of genres and activities in the choice of forms and expressions has been widely addressed (see Fox ; Gundel, Hedberg, and Zacharski ). Children experience language and discourse in the context of these activities and discourse genres (François ; Nelson ). They are immersed from birth in typical interactions that become routines (which Bruner coined under the name of formats (Bruner ). These “transactional contingent patterns” provide a context in which child directed speech makes sense and are the basis for early language acquisition (Ninio and Bruner ). The influence of activities on children’s productions has been demonstrated for the acquisition of speech acts or of the lexicon (Gleason, Phillips, Ely, and Zaretsky ; Leaper and Gleason  inter alia) and syntax (Berman and Nir-Sagiv ; Chenu, Jisa, and Mazur-Palandre ). Most studies on the impact of speech genres and activities on children’s use of referring expressions concern older children (see Mazur-Palandre and Jisa ). Investigation for young children has been conducted in the framework of the DIAREF project (Salazar Orvig, Marcos, Heurdier, and da Silva ; de Weck, Heurdier, Hassan, Klein, and Nashawati submitted). Activities

OUP CORRECTED PROOF – FINAL, 17/1/2019, SPi

    



involve various types of topics and contrasted ways of talking about the referents as in the two following examples: () Olga (;) – MLU . Olga and her mother are playing with a puzzle. OLG  sekwa? ‘c’est quoi?’ ‘what is it?’ ‘that, it’s a watering can’ MOT  ça c’est un arrosoir OLG  ɛ ͂naʁozwaʁ ‘un arrosoir’ ‘a watering can’ MOT  on le met où l’arrosoir? ‘where do we put the watering can?’11 OLG  ela eela laozwa ‘est là Fest là l’arrosoir’ ‘(it) is there, Fis there the watering can’ ‘ok it’s there’ MOT  voilà il est là () Alice (;) – MLU . Alice and her mother are playing with a doll. MOT  hm. voilà elle est assise MOT  tu as vu elle a les yeux ouverts quand elle est assise ALI  nɔ͂ ‘non’ ALI  awi! ‘ah oui!’ ALI  laɛ ͂ʃapo ‘(e)ll(e) a un chapeau’

‘hm. there it is. she is sitting’ ‘you see. she opens her eyes when she is sitting’ ‘no’ ‘oh yes!’ ‘she has a hat’

In example () labeling organizes the game. In the second example () mother and child are involved in the description of a doll. Even if in both cases referents are inanimate objects, neither mother nor child use the same referring expressions. In the first case () they use the clitic demonstrative c’ (in the construction “C’est+X”). Heurdier, da Silva, Le Mené, and Salazar Orvig () showed that this clitic demonstrative was much more frequent in such activities as picture book reading and games on iconic material than in activities involving object manipulation. This high frequency is linked to the preference for labeling (Salazar Orvig et al. ). On the other hand, description was the preferred genre for first uses of PPs (in the DIAREF corpus % of children’s PPs) as can be seen in example (). These results suggest that activities and genres influence the use of language at different levels: they frame a structured order of sequences and speech acts; they determine the structure of the utterances and the forms used, they entail different

¹¹ The object in MOT  is expressed by a dislocated structure, with a clitic pronoun before the verb and the noun after the verb.

OUP CORRECTED PROOF – FINAL, 17/1/2019, SPi



  

communication strategies (Leaper and Gleason ) and specific perspectives on participants and objects that affect topic management. In that sense they frame such factors as accessibility and topic management, in the choice of referential expressions.

.. To summarize Observations on the impact of dialogue suggest that—together with informativeness or accessibility—the choice of referring expressions is determined by dialogical factors. When a speaker encodes a referent, he/she builds on what has already been said and on what is expected, takes a stance with respect to the interlocutors’ discursive contribution, accepts or changes the other’s topic; he/she aligns with the interlocutors’ perspective or not while computing or projecting the interlocutors’ knowledge. The weight of these dialogical factors is as least as significant as in adult’s discourse. Therefore they give us a hint of how children build early competences.

. A   

.................................................................................................................................. The review of children’s uses of referring expressions presents a paradoxical landscape. Most studies of young children show a rather precocious competence in the use of referring expressions, whereas studies on narratives and other experimental settings tend to show that mastering the values and use of referring expressions is not achieved before age nine. This is all the more puzzling given that usually experimental studies involve older children than do naturalistic studies. As the cognitive differences between young children and older children and adults are undeniable, we have to consider that children’s linguistic behaviors in these different situations are not analogous. Several factors¹² can be identified that may account for this paradox, including the conception of reference, the representation of the child’s cognitive development, and the child’s involvement in the co-construction of reference. The conflicting results arise partly from the contrast between theoretical conceptions of reference. Initial studies, mostly on narratives, followed a textualist approach to reference and approached the issue from the perspective of cohesion and text management (de Weck ; Hickmann ; Karmiloff-Smith , ). The extraversus intra-discursive nature of the referent became the main criterion for capturing ¹² These apparent contradictions have been often explained by methodological differences (Hickmann ; Hickmann et al. ; Allen et al. ) that cannot be addressed here. They can concern the experimental design and the implementation of the observation (Bamberg ; De Cat ; Emslie and Stevenson ), as well as linguistic variables: what is included under ‘overt’ arguments? Are fillers considered? What is missed when wider categories group different pronouns or determiners? And so on . . .

OUP CORRECTED PROOF – FINAL, 17/1/2019, SPi

    



the meaning of children’s linguistic devices. Furthermore, textualist studies considered that children build progressively the semantic meaning of referring expressions from a deictic default value of plain ‘grammatical’ devices, to anaphoric values (when children master the management of long discourse and competing referents). In the meantime, co-referential chains are not considered as anaphoric but as the reiteration of deictic reference. On the other hand, studies on naturally occurring dialogues (Gundel and Johnson ; Allen et al. ; Salazar Orvig et al. a inter alia) have been conducted within cognitive conceptions of reference and the cognitive status of the referent is the key feature to assess the way children use referring expressions. Example (), repeated here, can illustrate this distinction: ()

Lisa (;) – MLU . Lisa is playing with a puzzle, her father is watching her. She places the pieces on the board but does not put them in the right place. Having done so, she looks at the puzzle. LIS alɛfini ‘ah! l’est fini’ ‘ah! it’s finished’ FAT tu crois que c’est fini? tu ‘you think it’s finished? you think crois que c’est comme ça que ça va that’s the way it’s supposed to be, Lisa? Lisa?’

Within a textualist approach, the child’s use of the rd person clitic pronoun l’ (‘it’) to first mention the puzzle, would be considered deictic/exophoric, overlooking the fact that this referent had been under joint attention for a long moment. Within a cognitive perspective, the fact that the puzzle was under joint attention would characterize the use of the pronoun as (proto-)anaphoric. This interpretation would be confirmed by the father’s use of the clitic demonstrative c’ (‘it’) showing that reference was intersubjectively shared even before the child’s first mention. Moreover, the fact that, in naturally occurring dialogues, most uses of clitic pronouns for (the extremely rare cases of) first mentions correspond to referents previously under joint attention, is an argument against the conception of deixis as the default value of rd person pronouns (Salazar Orvig, Hasan, et al. ; Salazar Orvig, Marcos, et al. a). From the moment studies on child language adopted a cognitive perspective, the issue at stake became whether children make the appropriate choice of referring expressions, and how and to what extent they assess the interlocutors’ needs, mainly as a function of the perceptual and discursive context. This position assumes that (a) children have the possibility to make referential choices within the same paradigms as adults and (b) referring expressions children use have adult-like meanings. For studies that clearly acknowledge these assumptions (De Cat a, b, ; Gundel and Johnson ; Gundel et al. ; Schaeffer and Matthewson , inter alia), usually in the Universal Grammar theoretical framework, the issue at stake is whether children have the cognitive abilities that allow them to assess interlocutors’ needs. According to

OUP CORRECTED PROOF – FINAL, 17/1/2019, SPi



  

these authors, when children use referring expressions inappropriately, they fail to correctly assess their interlocutor’s memory and attentional state. However, results and interpretations can be controversial. Schaeffer and Matthewson consider that younger children still lack the pragmatic Concept of Non-Shared Assumptions that would allow them to adequately use referring expressions. As De Cat () stressed, following Schafer and de Villiers () and van Hout et al. (), the significant variation in children’s rates of egocentric errors could be accounted for by aspects of their cognitive development and whether they already have a Theory of Mind (Perner ). However, De Cat () found (for determiners) no difference in the choice of referring expressions between children that succeeded at a Theory of Mind test and those who failed. For her, the difficulty rests more on the inability to monitor the difference of perspective between interlocutors. Gundel and Johnson () discuss Schaeffer and Matthewson’s results in two ways. They consider that their explanations are inconsistent with the cognitive status of the referents in the actual context of their experiment and with the fact that children rarely make mistakes using the definite article in naturally occurring dialogues. For Gundel et al. () we need to consider the use of referring expressions at two levels: the first level involves the ability to assess the cognitive status of the referent and the second one the ability to assess how much information is relevant for the interlocutor. This first level only involves implicit first level mind reading abilities such as understanding others’ intentions and mental states. Young children use pronouns appropriately without needing to reason about the epistemic state of the interlocutor and to assess how much information he/she needs. These last abilities correspond to the second stage of development of theory of mind which begins around age four. Moreover, Gundel and Johnson () recall that in naturally occurring dialogues, young children are able to assess the interlocutor’s cognitive states such as familiarity and attention “but they are not yet sensitive to making the choice based on pragmatic inferences that could arise when irrelevant information is provided”. Yet, these explanations do not consider the fact that younger children (under two) express impressive social–cognitive skills and knowledge. Children are sensitive to more than just intentions and emotions. They are sensitive to the interlocutor’s goals and needs, to shared experience, to disengagement and need for update as well as emotions and preferences of the interlocutors (Liszkowski, Carpenter, Striano, and Tomasello ; Moll, Richter, Carpenter, and Tomasello ; O’Neill , inter alia). As O’Neill () puts it, the matter at stake is not really—for infants—whether the information is familiar to their interlocutor but whether it is relevant in the context of the ongoing interaction. These findings on younger children’s non-verbal behavior suggest earlier socio-cognitive competence than experimental studies based on a language test would expect. It seems therefore necessary to examine the interactional conditions in which children learn and first use referring expressions. Three main complementary accounts can be invoked: the first concerns the formal influence of the adults’ discourse; the second considers the influence of discursive sequences; and the third considers the

OUP CORRECTED PROOF – FINAL, 17/1/2019, SPi

    



conditions of learning in dialogue. The influence of previously mentioned referents in the choice of referring expressions in both naturally occurring and experimental dialogues can seem contradictory with the assumption that very young children are not able to assess the informational needs of their interlocutors. Matthews et al. () considered the possibility that the impact of previous mentions in the choice of the referring expressions could be explained as the expression of a mechanical alignment to the interlocutors’ model of situation (Pickering and Garrod ), constructed by priming rather than by an assessment of the interlocutors’ knowledge. In this vein, repetition could be considered as a vehicle of the acquisition of anaphoric reference. However, its relative high frequency cannot account for all children’s uses. Moreover, in an analysis of subject function, Salazar Orvig, Marcos, et al. (b) showed that more than % of PPs did not follow a PP antecedent by the adult. Syntactic priming could not explain children’s uses. Other factors, such as contrast in dialogue, contributed to the choice of referring expressions. Therefore the question remains of how children learn that a pronoun is expected once a referent has been introduced. Observations on the role of genres and activities can give us a hint of how to solve the paradox. If we consider that children learn how to use referring expressions through the various language games in which they are involved, we can hypothesize that they incorporate recurrent discursive frames linking experience and speech. Children could thus begin by grasping pragmatic configurations (Nelson ) that fit routinized situations: for example, describing a referent in a second move after having introduced it. These formats (Bruner ) are privileged contexts that allow change and variation in a secured context. This could account for the fact that at the onset, children do not need to assess the information needed by the interlocutors. And, as children’s involvement in dialogue is not passive, experience of unclear reference seems to help them learn how to adjust the use of referring expressions (Matthews, Butcher, Lieven, and Tomasello ). The development of reference could, thus, consist of generalization, abstraction, and rectification processes from usage based formats towards productive paradigmatic choices according to the interlocutor’s knowledge and perspective.

. A: T   The indication DIAREF corpus in this chapter refers here to the younger children data (under three) of the DIAREF project. It is an open data base which is continuously incremented with new coded sessions. These data come from different previous research projects: crosssectional corpora come from research on early speech acts (Marcos, Ryckebusch, and Rabain-Jamin ), another on the influence of type of childcare on language acquisition (Marcos et al. ), the longitudinal corpus was gathered for research on the development of dialogue (Salazar Orvig, ), the COLAJE project (Morgenstern and Parisse ), and a doctoral dissertation (Nashawati ).

OUP CORRECTED PROOF – FINAL, 17/1/2019, SPi



  

The twenty-eight children currently included in the DIAREF corpus are aged between ; and ;. This corpus includes five longitudinal data (from ; to ;) and twenty-three cross-sectional data at two age groups (; and ;). As the linguistic development is not strictly correlated to age, data is analyzed according to their Mean Length of Utterance (MLU) which gives a hint of linguistic development. Data have been gathered in three groups (GR: from . to ; GR: from . to .; and GR: from . onwards).

A This chapter stems from a collective work within the framework of a larger research project, DIAREF (Acquisition des Expressions Référentielles en Dialogue: approche multidimensionnelle), funded by the French National Agency for Research (ANR--ENFT-), designed to study the development of referring expressions in children aged ; to . More specifically, analyses and interpretation of the data reported here have been accomplished with the collaboration of Christine da Silva, Julien Heurdier, Marine le Mené, Haydée Marcos, and Salma Nashawati, to whom I am fully indebted. I am also very grateful to Jeanette Gundel and Barbara Abbott for their very useful comments and feedback on this chapter and for their careful editing.

OUP CORRECTED PROOF – FINAL, 21/1/2019, SPi

        ......................................................................................................................

REFERENCE RESOLUTION A psycholinguistic perspective ......................................................................................................................

    

. I

.................................................................................................................................. O NE of the key aspects of linguistic communication is referring to already-mentioned people, objects, places, etc. From the perspective of the speaker, the act of referring involves a number of choices, such as what kind of referring expression to use (e.g., a pronoun, a name, a noun; she, Susan, my neighbor, the professor). These choices also have consequences for the listener: if faced with a seemingly under-informative form like ‘she’, how does the listener identify the intended referent? These questions have attracted the attention of researchers with a variety of backgrounds and using different types of research methods and data. In this chapter we discuss experimental research on reference resolution from the perspective of psycholinguistics.¹ In our review of psycholinguistic lab-based research on referring expressions, we mostly focus on the comprehension side, as that has traditionally received more attention in the experimental literature than production. Much of the experimental/psycholinguistic work has historically tended to focus on English, but in recent years there has been an increasing amount of psycholinguistic work on other languages, including Dutch, German, Spanish, French, Italian, Japanese, Korean, and Finnish, as will become clear over the course of this chapter. Language users refer to many different kinds of things, including entities, locations, times, states, and events. Our focus here will be mostly on anaphoric reference to entities, in particular humans. We will also briefly consider cataphoric configurations,

¹ Our focus here is on experimental results from behavioral, lab-based experiments. For a review of work on reference resolution from a neurolinguistic perspective, see e.g., Nieuwland and Van Berkum .

OUP CORRECTED PROOF – FINAL, 21/1/2019, SPi



    

where a pronominal form refers to an upcoming referent. Other related topics—such as introduction of new referents, deixis, and ellipsis—are unfortunately beyond the scope of this chapter. There is also an increasing amount of work on reference resolution by non-native speakers, but in this chapter we focus on the patterns observed with native speakers. Furthermore, although this chapter focuses on the results and outcomes of lab-based experiments as they are traditionally construed in psycholinguistics, it is also important to emphasize that there exists a large tradition of work that, although it does not involve lab-based experiments, nevertheless also makes important contributions to our understanding of the psycholinguistic and cognitive aspects of reference resolution. In the first part of this chapter, after considering the typological wealth of referring expressions in different languages (§..), we review experiments that aim to shed light on what guides a listener’s or reader’s interpretation of overt pronouns (§..). We start by considering the properties of the antecedent: we first discuss effects of topicality-related factors such as subjecthood, givenness, and pronominalization, and then turn to research that suggests that, perhaps surprisingly, focus-related factors can also play a role in guiding pronoun interpretation (§.). In addition to work on overt anaphor forms, we also consider experimental data on languages such as Italian, Spanish, Chinese, and Korean that have both null and overt pronouns (§.). In addition to reviewing claims about how reference resolution is influenced by properties of the antecedent (e.g., its grammatical role, thematic role, linear position, and information status), we also discuss a different view, according to which pronoun interpretation is a by-product of general inferencing/reasoning about the semantic coherence relations between clauses. Recent work has explored possible ways of reconciling these two views, in addition to questions regarding the mental representation of coherence relations and the link to perspective-taking (§.). The fact that our working memory and attentional resources are limited constrains the reference resolution process. In §., we review consequences of processing limitations on the production and interpretation of pronouns, the possibility of ‘shallow processing’ as a means to decrease processing cost, and how the cognitive cost of holding unresolved referential dependencies in working memory impacts comprehension of cataphoric pronouns. §. is the conclusion and also discusses open questions and directions for future work.

.. Cross-linguistic variation in anaphoric forms Before turning to theories of how comprehenders interpret anaphoric forms, let us consider the cross-linguistic variety of referring expressions. In English, the most common referring expressions for human entities are personal pronouns (she, he), definite and indefinite nouns (a man, the woman, the students), and proper names. In other languages, human entities can be referred to with a wider range of expressions.

OUP CORRECTED PROOF – FINAL, 21/1/2019, SPi

 



For example, Finnish, Estonian, Dutch, and German use both personal pronouns and anaphoric demonstratives or other demonstrative-type elements to refer to humans (e.g., Kaiser ; Bosch and Umbach ; Comrie ; Diessel ; and many others), as shown in (). In sentences with canonical subject-object word order, anaphoric demonstratives tend to refer to the preceding object, while personal pronouns often refer to the subject but can—at least in some contexts—also refer to the object (see e.g., Comrie  for a discussion of demonstratives being ‘more strict’ in their referential biases). Traditionally, the anaphoric demonstratives are often described as referring to a non-topical entity, but this may be an oversimplification, as recent work suggests that logophoricity and perspective-taking may also be at play. See Hinterwimmer and Bosch () on German, and Kaiser () on Finnish for recent work on the relevance of logophoricity and perspective-taking. There are also languages with long and short variants of personal pronouns (sometimes called weak and strong), including Dutch (ze/zij ‘she’, e.g., Haeseryn et al. ) and Estonian (tema/ta, gender-neutral ‘she/he’, e.g., Pajusalu ). () a. Finnish: Pekkai halusi pelata tennistä Matinj kanssa, mutta {häni, j/tämäj} oli sairas. Pekka wanted play-INF tennis Matti with, but {s/he/demonstrative} was sick ‘Pekka wanted to play tennis with Matti, but {he/demonstrative} was sick.’ b. German: Peteri wollte mit Paulj Tennis spielen. Doch {eri, j/derj} war krank. Peter wanted with Paul tennis play-INF. but {he/demonstrative} was sick. ‘Peter wanted to play tennis with Paul, but {he/demonstrative} was sick.’ (from Bosch and Umbach ) c. Dutch: Peteri wilde met Paulj gaan tennissen. Maar {hiji, j/diej} was ziek. Peter wanted with Paul go tennis-play-INF. but {he/demonstrative} was sick. ‘Peter wanted to play tennis with Paul, but {he/demonstrative} was sick.’ (from Bosch and Umbach ) In addition to overt pronouns, many languages also use zero (null) pronouns. For example, in Spanish human antecedents can be referred to with overt personal pronouns (and clitics), null pronouns, or demonstrative pronouns. However, not all null pronouns pattern alike: in some languages, rich verbal inflection provides information about the referent (e.g., Spanish, Italian), but other languages have null subjects and objects despite lacking verbal morphology (e.g., Chinese). These two subtypes of null pronouns are frequently referred to as pro-drop and topic-drop respectively. Another type, bound pronouns, occurs in languages like Navajo (see Kibrik  for crosslinguistic discussion). Languages also utilize other types of discourse-governed pronouns, such as the logophoric pronouns in reported speech contexts in a number of African languages, which refer to the person whose thoughts or speech are being reported (e.g., Hagege ;

OUP CORRECTED PROOF – FINAL, 21/1/2019, SPi



    

Kibrik : ; Pearson ). In addition to ‘pure’ logophoric languages which possess specific pronouns that are used only in logophoric contexts, there are also ‘mixed’ languages (Culy ) where forms with logophoric/perspective-sensitive properties also have other functions and/or also occur in non-logophoric contexts. For example, long-distance reflexives in languages such as Chinese pose challenges for many syntactic theories of reflexivization and are associated with perspectival effects (e.g., Huang ). These forms are not always reflexive pronouns. See, for example, Laitinen () and Kaiser () on Finnish. Having discussed the variety of anaphoric forms in human language, it is important to acknowledge that the psycholinguistic research discussed in this chapter does not do justice to the intricacies of the referential devices used in the world’s languages, because most psycholinguistic research has focused on a fairly small handful of languages. For example, Anand et al. () estimate that only ten languages (a very small slice of the world’s languages) account for at least % of the research in psycholinguistic journals and conferences. However, it is very encouraging that psycholinguistic researchers are showing increasing interest in conducting research on a variety of languages (e.g., Jaeger and Norcliffe ; Christianson and Ferreira ; Wagers, Borja, and Chung ). The following sections start by considering a basic question that any theory of reference must address: given that many referring expressions are semantically underspecified, how do listeners and readers identify the intended referent that the speaker/ writer has in mind? In other words, what guides the interpretation of referring expressions? We first review existing theoretical and typological work on this topic, and then turn to psycholinguistic experiments.

.. On the form–function relationship: hierarchies of referring expressions It is widely assumed that there exists a correlation between the type of referential form used and a referent’s² level of topicality or prominence, such that more-reduced anaphoric expressions tend to be used for more salient and/or more topical referents. For example, according to Givón’s topicality-based approach (e.g., Givón ), the more continuous a topic is (based on factors such as the linear distance between two mentions of a referent and the number of intervening referents), the more likely it is to be realized with a reduced referring expression such as a null or unstressed pronoun. Givón’s topicality-based approach is supported by corpus data from a variety of languages, including Chamorro, English, Japanese, and Hausa. ² There exists a distinction between an actual entity in the world and a discourse referent (sometimes called a discourse entity) in the linguistic domain. Thus, strictly speaking, anaphoric expressions take discourse referents/discourse entities as their antecedents. In this chapter, we sometimes use the terms ‘referent’ and ‘entity’ interchangeably.

OUP CORRECTED PROOF – FINAL, 21/1/2019, SPi

 



A related tradition focuses on the notion of cognitive accessibility/activation/attention (e.g., Ariel ; Gundel, Hedberg, and Zacharski ; Chafe ; Kibrik ). The core idea is that when a referent is highly accessible/activated in the addressee’s mind, or in the addressee’s focus of attention, the speaker can use a reduced referring expression. Less activated referents are referred to with fuller referring expressions. A simplified version of part of the ‘standard’ hierarchy of referring expressions is in (), with forms further to the left being used to refer to more accessible/activated referents.³ ()

null forms > (unstressed) pronouns > demonstratives > full nouns . . .

However, different approaches vary in their conceptualization of the notion of activation/accessibility and the nature of the hierarchy. According to Ariel’s () Accessibility Hierarchy, anaphoric expressions directly encode information about the accessibility of their referents in the addressee’s mind (as assessed by the speaker). Under this view, anaphoric forms are ranked according to how much information they encode, how phonologically attenuated they are, and how rigidly they identify entities. Positing a connection between accessible/prominent referents and reduced referring expressions seems very plausible. As noted by Garnham (), “[a]n expression that has little semantic content . . . can contribute little or nothing to the identification process, and can only be used where identification of the referent is either straightforward or not an issue” (Garnham : ). However, not all referring expressions differ in their informativeness. In English, it and that “are indistinguishable with respect to the description they provide for the intended referent (an inanimate object)” (Ariel : ), but according to the hierarchy in (), demonstratives are nevertheless used for less prominent referents than pronouns. Thus, Ariel’s approach makes specific claims about a ranking between linguistic forms that is at least partly independent of their informativity. According to Gundel, Hedberg, and Zacharski (), different referential forms allow the speaker to encode information about the cognitive status of the intended referent, and thus provide procedural information for the listener about how to identify the intended referent. For example, according to Gundel et al.’s Givenness Hierarchy, unstressed pronouns signal that the intended referent is currently in the focus of attention; stressed pronouns and the demonstratives this/that and this N signal that the intended referent is currently in working memory, and that N signals that the intended referent is familiar/in memory.⁴ The Givenness Hierarchy does not provide direct information about degrees of accessibility, and thus differs from Ariel’s ³ The question of how stressed/accented pronouns fit on the hierarchy is under debate. Some (e.g., Ariel ; Kameyama ) argue for an accessibility-oriented approach where stressed pronouns refer to less accessible referents than unstressed pronouns, but others (e.g., de Hoop ) argue for a contrast-based account (see also Venditti et al.  for eye-tracking work on the effects of coherence relations on the interpretation of stressed pronouns, see also Mozuraitis & Heller ). ⁴ As Gundel () notes, the cognitive statuses that form part of the Givenness Hierarchy constitute a “folk psychology theory” (Gundel : ). She notes that the extent to which these cognitive states map onto notions such as attention and working memory as used in the current psychological literature is an empirical question (see Gundel ; Gundel et al. : ).

OUP CORRECTED PROOF – FINAL, 21/1/2019, SPi



    

Accessibility Hierarchy (see Gundel et al. ,  for further discussion). Crucially, the Givenness Hierarchy is an implicational scale, and each cognitive status entails the cognitive statuses below it (e.g., something that is currently in the focus of attention is also in working memory). Gundel et al. support their approach using corpus evidence and native speaker judgments from a range of languages, including English, Chinese, Ojibwe, and Tunisian Arabic (Gundel, Hedberg, and Zacharski ; Gundel, Bassene, Gordon, Humnick, and Khalfaoui ). In sum, there exists a considerable amount of mostly corpus-based work supporting the idea that different referring expressions are associated with referents with different levels of accessibility/prominence. If we assume that different anaphoric forms map on to different levels of salience/accessibility, we are faced with the fundamental question of what influences the accessibility/cognitive status/prominence of a particular entity. From the perspective of the comprehender, we can ask: what makes something a good antecedent for a pronoun (or some other reduced referring expression) that the comprehender needs to resolve? In other words, when a reader or listener encounters a pronoun, what guides their interpretation of that form? In the next sections, we review psycholinguistic experiments that seek to answer this question.

. P        

.................................................................................................................................. Prior experimental work on the processing of pronouns and other referential forms suggests that multiple factors guide reference resolution, including (i) properties of the antecedent such as its grammatical role, discourse status, thematic role, and linguistic form, (ii) properties of the referential form being interpreted, such as its grammatical role and linguistic form, and (iii) properties of the linguistic context, such as discourse structure and the coherence relations between sentences. We start our discussion by considering the properties of the antecedent. The discussion below is divided into two main parts, inspired by information structure: we first discuss effects of topicality-related factors such as subjecthood, givenness, and pronominalization (§..–§..), and then turn to research that suggests that focus-related factors can also play a role in guiding pronoun interpretation (§..).

.. Topicality-related factors: subjecthood, givenness, pronominalization Prior research using both experimental and non-experimental data (e.g., corpus data, native speaker intuitions) suggests that being realized in subject position renders an

OUP CORRECTED PROOF – FINAL, 21/1/2019, SPi

 



entity a good antecedent for a subsequent pronoun (e.g., Brennan, Friedman, and Pollard ; Chafe ; Crawley and Stevenson ; Gordon, Grosz, and Gilliom ; see also Centering Theoretic work e.g., Grosz, Joshi, and Weinstein ). Although this subject bias does not always hold and has been argued to be partially epiphenomenal (e.g., Kehler ; Schumacher et al. ; see also §.), it has played an important role in research on reference resolution; so we start by reviewing it here. More broadly, as we will see in the next few sections, a collection of topicality-related factors—subjecthood among them—appear to be related with an increased likelihood of being interpreted as the antecedent of a subsequent pronoun. In one of the earliest experiments highlighting the special status of subjects, Crawley and Stevenson () asked participants to write continuations for fragments with and without pronoun prompts, exemplified in (). Participants exhibited a preference for referring to the preceding subject in both conditions, and especially when interpreting pronouns (as in b).⁵ ()

a. Shaun led Ben along the path and . . . b. Shaun led Ben along the path and he . . .

Additional evidence for the special status of subjects comes from one of the first reaction-time studies on this topic by Gordon, Grosz, and Gilliom (). Building on Centering Theory (Grosz, Joshi, and Weinstein ), Gordon et al. measured participants’ reading times for sentences with two pronouns, a pronoun and a name, or two names. If an entity had been mentioned with a name in the subject position of one sentence and was mentioned again as the subject of the following sentence, reading times were significantly longer if a name was used again (e.g., Bruno-Bruno), compared to a pronoun (Bruno-he). However, the slowdown did not occur (i.e., there was no processing penalty) if the second occurrence was in a non-subject position. The slowdown with subject-position repeated names is known as the repeated name penalty, and highlights the importance of subjecthood as well as the (potentially counter-intuitive) finding that using a highly explicit form (a name) may in fact slow down language processing. Indeed, this finding that overly informative forms slow down processing goes against Gernsbacher (), who proposes that anaphoric explicitness should facilitate processing. Claims regarding the special status of subjects are not restricted to English, and have also been observed for other languages (e.g., Carminati  on Italian; Alonso-Ovalle et al.  on Spanish; Kaiser and Trueswell  on Finnish), which we discuss in §.. and §... Evidence regarding the impact of subjecthood also comes from visual-world eyetracking paradigms (pioneered in psycholinguistics by Tanenhaus et al. ). In ⁵ This study, like many others, focused on pronouns in subject position. Thus, as Crawley and Stevenson () note, these results may be due to a subject-assignment strategy or to the parallelism strategy, according to which pronouns tend to be interpreted as referring to antecedents in parallel grammatical roles (e.g., Smyth ). See Kertz et al. () for recent claims that neither parallelism effects nor subjecthood effects provide the full picture.

OUP CORRECTED PROOF – FINAL, 21/1/2019, SPi



    

visual-world studies, participants hear linguistic stimuli as they look at visual scenes, while their eye-movements are tracked. Eye-movements provide a measure of which potential antecedents participants consider as they hear referring expressions and what the time-course of this process is. In one of the first visual-world studies on pronoun resolution, Arnold et al. () investigated how gender cues and grammatical role (which in English is largely correlated with order-of-mention⁶) guide the interpretation of pronouns. Participants listened to sentences like (a) while viewing a scene depicting the two mentioned referents. () a. Donald is bringing some mail to {Mickey/Minnie} while a violent storm is beginning. He’s/She’s carrying an umbrella, and it looks like they’re both going to need it. With two different-gender referents (e.g., Donald and Minnie), eye-movements showed a rapid sensitivity to gender cues: participants start to look at the gendermatching referent about ms after the onset of the pronoun. When two characters had the same gender (e.g., Donald and Mickey) and there was no early semantic cue to disambiguate which was the intended referent, a preference for the preceding subject started to emerge within –ms of pronoun onset. This subject preference is strengthened when the subject was re-mentioned in an additional clause before the critical sentence (see (b)). This provides experimental support for the claim that being old (given) information also makes a referent a likely antecedent for a pronoun (e.g., Strube and Hahn , ; see also earlier work by Garrod and Sanford ). () b. Donald is bringing some mail to {Mickey/Minnie}. [He’s sauntering down the hill,]additional clause while a violent storm is beginning. He’s/She’s carrying an umbrella, and it looks like they’re both going to need it. In a follow-up study with four- to five-year-old children, Arnold, Brown-Schmidt, and Trueswell () found that while children’s eye-movements reveal early effects of gender, children did not seem to be sensitive to subjecthood. Somewhat divergent results arise in a preferential-looking study by Song and Fisher (), whose results suggest that as long as there are enough convergent cues (in their study, subjecthood and repeated mention), even three-year-olds can identify the intended referent, though not until ms after the pronoun (see also Pyykkönen, Matthews, and Järvikivi ). A recent paper by Hartshorne, Nappa, and Snedeker () finds that five-year-olds show a preference for subjects, but—in line with Song and Fisher—their results suggest that it takes over ms for this preference to emerge.

⁶ Because English has canonical Subject-Verb-Object order, this study does not separate effects of subjecthood from effects of linear order, and in fact work on English often refers to the subjecthood preference as a “first-mention bias”. We return to this in §.. where we discuss flexible word order languages.

OUP CORRECTED PROOF – FINAL, 21/1/2019, SPi

 



Taken together, these experiments provide evidence for a subject bias in pronoun interpretation and highlight the importance of considering not only the comprehender’s ultimate interpretation but also the time-course of reference resolution.

.. Subjecthood and information structure The considerable body of prior work suggesting that pronouns tend to be interpreted as referring to preceding subjects raises the question of why subjecthood matters: what about subjects makes them good antecedents? In the following sections we consider three possibilities: (i) information structure (subjects tend to be old/given/topical information), (ii) linear position, often related to information structure (subjects tend to occur linearly before objects), and (iii) thematic roles (subjects tend to be agentive). Although this is not an exhaustive list, the studies conducted so far probing these issues suggest that, at least from a cross-linguistic perspective, all three characteristics of subjects can be at play. First, we will consider the information status of subjects: the referents realized in subject position tend to be information that has already been mentioned in the discourse. More specifically, corpus studies have shown that subjects are often givenold information (e.g., Prince ). This raises the question of whether effects of subjecthood in the domain of reference resolution persist when subjecthood is dissociated from information-structural considerations such as givenness. To shed light on this, Kaiser (a) used visual-world eye-tracking to test what happens when subjecthood is explicitly pitted against givenness. Participants heard short stories (see example ()) while viewing scenes depicting the characters in the sentence as well as other objects. Kaiser manipulated whether the subject, the object, or neither of the two was given information (and pronominalized), resulting in three conditions as shown in (a–c). The underlined section shows the givenness/pronominalization manipulation (see Kaiser a for sample images; the images had lookaway objects designed to attract eye-movements during the sentence preceding the critical sentence).⁷ () a. [S=pro, O=name] condition Greg is always very supportive of others. He congratulated John enthusiastically yesterday. The prizes for the best-ranked tennis players were about to be announced, and he was holding a new yellow tennis racket. [Critical sentence] Everyone was in a good mood that day. [Wrap-up] ⁷ Participants were told that if there was a mismatch between the narrative and the image, they should click with the mouse on the region of the image that contains the error. In all targets, the critical sentence was incorrect with respect to both potential referents. For example, in (), both men were holding tennis rackets, but neither racket was yellow. Thus, participants’ clicks provide an offline measure of which character they interpreted as the referent of the pronoun.

OUP CORRECTED PROOF – FINAL, 21/1/2019, SPi



     b. [S=name, O=pro] condition Mike did very well in last month’s tennis tournament. John congratulated him enthusiastically yesterday. The prizes for the best-ranked tennis players were about to be announced, and he was holding a new yellow tennis racket. [Critical sentence] Everyone was in a good mood that day. [Wrap-up] c. [S=name, O=name] condition Greg congratulated John enthusiastically yesterday. The prizes for the best-ranked tennis players were about to be announced, and he was holding a new yellow tennis racket. [Critical sentence] Everyone was in a good mood that day. [Wrap-up]

Participants’ eye-movements showed clear effects of both subjecthood and pronominalization. In conditions with a preceding pronominalized object ([S=name, O=pro], (b)), the pronoun in the critical sentence triggered looks to images of both the character realized with a pronoun in preceding object position (e.g., Mike in (b)) and the character realized with a full name in the preceding subject position (e.g., John in (b)), indicating competition between these two referents. In contrast, the pronouns in the other two conditions ([S=name, O=name] and [S=pro, O=name]) showed a clear subject preference. In other words, a given, pronominalized object ‘competes’ with a preceding subject when the subject is discourse-new information, but when both subject and object are new information, the pronoun triggers more consideration of the subject. Thus, the subject receives special consideration even if it is not given information or pronominalized. Kaiser’s (a) results show that both subjecthood and givenness (strengthened here with pronominalization) have separable effects on real-time pronoun interpretation. The separable effects of subjecthood and givenness show that (agentive) subjects have a privileged status even when they are new information and do not fit the traditional definition of ‘topic.’

.. Subjecthood and linear order In this section we turn to the relation between subjecthood and linear order. English has fairly rigid word order and subjects often linearly precede objects (but see Ward and Birner  on non-canonical word orders). Thus, in the case of English, disentangling the contributions of word order and grammatical role is not easy. However, languages with flexible word order (usually related to information-structural considerations) such as Finnish and German allow us to test whether the subjecthood preference observed in English is due to linear order or grammatical role. Work on Finnish by Kaiser () and Kaiser and Trueswell () used offline sentence completion tasks, corpus analyses, and visual-world eye-tracking to investigate the effects of word order variation on the interpretation of the gender-neutral personal

OUP CORRECTED PROOF – FINAL, 21/1/2019, SPi

 



pronoun hän ‘s/he’ and the proximal demonstrative tämä ‘this’, which can also be used anaphorically to refer to human referents. Recall that according to most hierarchies of referring expressions (§..), demonstrative pronouns are used for less accessible referents than pronouns, or—from a different perspective—provide procedural retrieval instructions to the addressee signaling that the referent may not be in the current focus of attention. Kaiser and Trueswell focused on the interpretation of personal pronouns and demonstrative pronouns, following canonical SVO sentences and non-canonical OVS sentences which were embedded in short narratives. The preverbal argument in both SVO and OVS was old/given information, whereas the post-verbal argument was new information. The results show that personal pronouns and demonstrative pronouns pattern asymmetrically: the personal pronoun hän prefers the preceding subject regardless of word order. In contrast, the demonstrative tämä tended to be interpreted as referring to the post-verbal argument, especially when it was an object. Thus, Kaiser and Trueswell concluded that the interpretation of hän is driven primarily by the syntactic role (or thematic role) of potential antecedents, while tämä exhibits a sensitivity to both word order (correlated with information structure) and to syntactic role. Based on these patterns, Kaiser and Trueswell argue for the form-specific, multiple-constraint approach. According to this approach, referential forms can differ in how sensitive they are to different antecedent properties. They argue that the multiple constraints that play a role in the interpretation of referential forms do not necessarily carry the same weight for all referential forms. These results suggest that questions such as ‘is reference resolution sensitive to subjecthood/linear order/some other factor’ need to be relativized to the referential form at hand, since different forms exhibit different levels of sensitivity to different factors (see also Schumacher et al. ; Brown-Schmidt, Byron, and Tanenhaus ; Filiaci et al. ; and Ueno and Kehler ; Kaiser et al. , for evidence supporting the form-specific, multiple-constraints approach). Related eye-tracking work on Finnish by Järvikivi, van Gompel, Bertram, and Hyönä () found that the personal pronoun hän exhibits an initial subject preference, followed by later effects of both syntactic role and word order. Although these results differ somewhat from Kaiser and Trueswell ()—perhaps due to absence of a discourse context—they are not incompatible with the claim that hän and tämä differ in how sensitive they are to various types of information. Like Finnish, German uses personal and demonstrative pronouns for human antecedents and has flexible word order. Corpus data and native speaker intuitions suggest that, at least with canonical SVO order, German personal pronouns (er, sie ‘he, she’) prefer preceding subjects while demonstratives (der, die; so called d-pronouns) tend to refer to objects (or non-topical antecedents, e.g., Bosch et al. ; Zifoun et al. ; Abraham ), which is similar to the patterns with Finnish SVO sentences. What about the effects of word order? Psycholinguistic studies have led to mixed results. For example, Bosch and Umbach () used a reading-time task and a fill-in-the-blank task and found that demonstratives prefer objects regardless of word order and personal pronouns show a numerical (but non-significant) preference for the subject

OUP CORRECTED PROOF – FINAL, 21/1/2019, SPi



    

regardless of word order. In an offline questionnaire testing personal pronouns, Bouma and Hopp () similarly found a subject preference regardless of word order. A series of experiments by Wilson () using visual-world eye-tracking corroborates the strong subject preference of pronouns, but also finds that demonstratives show a clear preference for post-verbal, non-topical antecedents both in SVO and OVS order. Thus, contrary to Bosch and Umbach, Wilson () observes an asymmetry in the behavior of German personal pronouns and demonstratives, similar to that which Kaiser and Trueswell () found for Finnish. Why have some of the prior experiments on German lead to inconsistent results? A recent study by Kaiser (c) on German investigated the interplay between referring expression type and the coherence relations between sentences (see §. on coherence effects), and found that while (i) interpretation of German personal pronouns can be ‘pushed around’ depending on the semantic relation between two sentences, (ii) German demonstrative pronouns are much more fixed in their preference for the object. In light of this finding, it may be the case that some of the seemingly divergent results in earlier work may be due to differences in coherence relations (e.g., a cause–effect relation versus a temporal relation). Further evidence for discourse-level effects on reference resolution comes from recent work by Ellert (), who found that in comparative constructions in German with no agent–patient asymmetry (e.g., The cupboard is heavier than the table. Pro/ Dem . . . ), personal pronouns preferred the first-mentioned entity and demonstratives preferred the second-mentioned entity. However, when the word order was changed to signal that the second-mentioned referent is focused (e.g., Heavier than the table is the cupboard. Pro/Dem . . . ), pronouns and demonstratives overlap in their referential behavior; both showed a preference for the focused entity. (We will return to the question of how focus impacts pronoun interpretation in §..) As a whole, the findings for German show that pronouns and demonstratives cannot be straightforwardly described as referring to more accessible and less accessible referents respectively, at least not if we define accessibility in the same way for both forms. What about the question of whether the widely-discussed subjecthood preference exhibited by pronouns in English is due to grammatical role or linear order? The data from German and Finnish suggest that (agentive) subjects can be privileged antecedents even if they are not linearly initial. At the same time, the Finnish and German results point to a complex interplay between different factors and highlight the existence of form-specific effects, that is, personal pronouns and demonstrative pronouns can differ in terms of what properties of the antecedent they are most sensitive to.

.. Subjecthood and agentivity As mentioned earlier, much of the prior work on effects of grammatical role has looked at agentive subjects, essentially confounding grammatical roles and thematic roles.

OUP CORRECTED PROOF – FINAL, 21/1/2019, SPi

 



Given that subjects are indeed typically agentive (e.g., Ferreira ), this is by no means an unreasonable structure to focus on. However, to assess whether pronoun interpretation is being influenced by subjecthood, by agenthood, or by both, one should also consider verbs with non-agentive subjects. There is existing research on so-called implicit causality verbs (e.g., Garvey and Caramazza ), which in many cases have subjects that are non-agentive (e.g., the subject of ‘to fear’ is commonly viewed as an experiencer). However, because work on implicit causality verbs has tended to focus specifically on sentences linked by an explanation relation (X frightens Y because . . . , X fears Y because . . . ) which is known to influence pronoun interpretation in specific ways (see §.), it is difficult to draw conclusions directly about the relative contributions of subjecthood versus agentivity from this work. However, recent work in German by Schumacher, Roberts, and Järvikivi () shows that agentivity plays an important role in pronoun resolution, even when dissociated from subjecthood. Schumacher et al. used visual-world eye-tracking to investigate the real-time interpretation of personal pronouns (er ‘he’) and anaphoric d-pronouns (der, see §..) following sentences with agentive subjects (e.g., wash, hit) as well as sentences with non-agentive dative experiencer subjects, as in example () with the verb ‘to please’. Both the canonical order (where the dative experiencer occurs preverbally, at the start of the sentence) and the non-canonical order (where the dative experience occurs after the verb) were tested. () a. DATIVE-NOMINATIVE (canonical order) Dem Gärtner gefällt der Kapitän, der ein Eis isst. the-DAT gardener is-pleasing-to the-NOM skipper who an ice cream eats ‘The skipper who eats ice cream is pleasing to the gardener.’ b. NOMINATIVE-DATIVE (non-canonical order) Der Kapitän gefällt dem Gärtner, der ein Eis isst. The-NOM skipper is-pleasing-to the-DAT gardener who an ice cream eats ‘The skipper is pleasing to the gardener who eats ice cream.’ c. Critical sentence Aber {er/der} redet gerade mit zwei Damen. but {he/demonstrative} talks now with two ladies ‘But he is talking to two ladies right now.’ The results show an interplay between word order and thematic role. Whereas the personal pronoun er ‘he’ mostly triggers looks to the first-mentioned referent (regardless of whether it is the dative experiencer or the nominative argument), the anaphoric d-pronoun der shows a complex time-dependent sensitivity to both word order and thematic role. Based on the results of the experiment with dative experiencer verbs as well as another experiment with agentive-subject verbs, Schumacher et al. conclude that the personal pronoun er ‘he’ prefers proto-agents and the d-pronoun der prefers proto-patients. Thus, they suggest that thematic roles—such as agentivity—play a central role in guiding reference resolution.

OUP CORRECTED PROOF – FINAL, 21/1/2019, SPi



    

.. Topic and focus In the preceding sections, we focused on subjecthood, agentivity, givenness, and linear order—all factors that can be regarded as related to topicality: subjects are often topics and often agentive, topics are usually given information and tend to occur in sentenceinitial position. The general idea that topical referents are cognitively prominent (and thus likely to be interpreted as the antecedents of pronouns) contrasts with another line of research suggesting that entities which are contrastively focused have a privileged cognitive status (e.g., Hornby ; Zimmer and Engelkamp ; Sturt, Sanford, Stewart, and Dawydiak ; Ward and Sturt ). Looking specifically at reference resolution, Almor () found that NPs referring to clefted antecedents are read faster (i.e., processed more easily) than NPs that refer to non-clefted/non-focused antecedents. Furthermore, on the basis of pronoun resolution patterns, Birch, Albrecht, and Myers () argued that focused concepts are more salient and have a “stronger memory trace” (Birch et al. : ) than non-focused concepts. Relatedly, Foraker and McElree () suggested that clefting makes referents more distinctive in memory. To see how focusing and topicality interact in the domain of pronoun use, Arnold () directly compared topical referents to focused referents in a story-continuation production task. She found that participants were more likely to use pronouns to refer back to topical subjects than focused objects. This suggests that given information in subject position is more prominent, at least from the perspective of pronoun production, than a contrastively focused object. Recent work by Colonna, Schimke, and Hemforth () on clefting in French also found a preference for topic over focus (but see also Ellert  on related work on German). However, divergent results are reported by Cowles () and Cowles et al. (). Cowles tested three kinds of antecedents: (i) contrastively focused names in subject position, (ii) given-information names in subject position, and (iii) new-information names in subject position (a–c). () a. Contrastively focused subject: A new movie opened in town. It was Anne who called Sarah. b. Given subject: Anne wanted to see the new movie with Sarah. So, Anne called Sarah. c. New subject: A new movie opened in town. So, Anne called Sarah. d. Pronoun-containing third sentence (same in all conditions): But later that night, she couldn’t go to the movie after all. Participants saw names (e.g., Sarah or Anne) right after the critical pronoun in the third sentences, and had to say them aloud. Naming latencies were shorter when participants had to name the preceding subject (e.g., Anne) than the preceding object (e.g., Sarah), regardless of information structure. In contrast to Arnold, Cowles concludes that topics and foci “appear to have the same psychological effect” (Cowles : ). The divergent conclusions may be due to differences in grammatical role: Arnold’s foci were in object position but Cowles’s foci were in subject position.

OUP CORRECTED PROOF – FINAL, 21/1/2019, SPi

 



To probe these divergent claims, Kaiser (a) tested focused and old/given entities in both subject and object position. Focus was marked prosodically in canonical SVO order or by means of an it-cleft. People listened to dialogues such as example () while viewing clip-art scenes, while their eye-movements were recorded: () Speaker A: I heard that Greg congratulated Mike enthusiastically yesterday. Speaker B: a. No, that’s not quite right. b. (i) He congratulated John. [SVO.Object=focus] (ii) John congratulated him. [SVO.Subject=focus] (iii) It was John that he congratulated. [Cleft.Object=focus] (iv) It was John who congratulated him. [Cleft.Subject=focus] c. The prizes for the best-ranked tennis players were about to be announced, and d. he was holding a new yellow tennis racket [TEST SENTENCE] e. Everyone was in a good mood that day. Participants’ eye-movement show that the pronoun in the test sentence (d) tends to be interpreted as referring to the subject, regardless of whether the subject or object was contrastively focused (in an SVO sentence or a cleft) or pronominalized. This shows that a subject preference can arise even in sentences where the subject is focused and the object is given, pronominalized information. Recall that in a similar configuration, when the subject was new information (but not contrastively focused) and the object was given and pronominalized, Kaiser (a) found competition between the subject and the object (§.., example ()). However, when the subject is discourse-new and contrastively focused, it is preferred over the given object. This shows that when grammatical role is taken into account, we can detect effects of contrastive focus boosting referents’ prominence. As a whole, prior work suggests that topicality and (contrastive) focus—two notions that are often regarded as differing in many informational-structural and pragmatic respects (e.g., Gundel and Fretheim )—can both boost referents’ chances of being interpreted as the antecedent of a following pronoun.

. P          

.................................................................................................................................. In addition to personal pronouns and demonstratives, psycholinguistic experiments have also compared overt personal pronouns and null (or zero) pronouns, as well as strong and weak variants of personal pronouns. In particular, the phenomenon of pro-drop in Romance languages has received considerable attention. Carminati (), based on self-paced reading and questionnaire data for Italian, introduced

OUP CORRECTED PROOF – FINAL, 21/1/2019, SPi



    

the Position of Antecedent Hypothesis (PAH), which states that null (subject) pronouns prefer antecedents in the syntactic spec-IP position—that is, in a preverbal subject position, a syntactically prominent position—whereas overt (subject) pronouns prefer antecedents in less prominent syntactic positions (positions lower in the syntactic tree). Similar patterns have been found in Catalan by Mayol and Clark (), but with a less clear object-bias for the overt pronoun. Alonso-Ovalle et al. () tested the PAH in Iberian Spanish using offline questionnaires and found that null pronouns prefer referents in spec-IP (subjects), whereas overt pronouns are split between subjects and objects—suggesting that overt pronouns in Italian and Spanish do not pattern alike. However, recent work by de la Fuente and Hemforth (: ex. ()) on a range of Spanish dialects suggests that, at least in some contexts, null pronouns prefer subjects and overt pronouns prefer objects (see also related work on Buenos Aires Spanish by Gelormini-Lezama and Almor ; Spanish dialects have been shown to differ in their rate of null versus overt pronouns, e.g., Morales ). A direct comparison between Italian and Spanish null and overt pronouns can be found in Filiaci, Sorace, and Carreiras (). Based on reading-time studies that used parallel experimental designs and materials in (Iberian) Spanish and Italian, Filiaci et al. observed that null pronouns in both languages have a strong preference for subjects. However, overt pronouns differ in the two languages: overt pronouns in Italian incur a processing penalty when they refer to syntactically prominent antecedents (subjects), whereas overt pronouns in Spanish show a significantly weaker processing penalty when referring to subjects.⁸ Filiaci et al. connect their results to Kaiser and Trueswell’s () form-specific multiple-constraints account, since their findings support the idea that informationally equivalent anaphoric expressions can differ in how sensitive they are to “different linguistic determinants of prominence.” Recent work by Runner () on Spanish shows that semantic verb bias effects (in particular, implicit causality effects) as well as passivization and topic-focus distinctions can modulate or override the subject bias of null pronouns and the object bias of overt pronouns. Runner and Ibarra () also find that null and overt pronouns in Spanish differ in how sensitive they are to the given versus new status of preceding subjects. Experiments on so-called topic-drop languages (e.g., Japanese, Korean, Chinese) have not identified a clear division of labor between null and overt pronouns. A series of studies by Yang et al. (, , ) on the interpretation of null and overt pronouns in Mandarin Chinese suggests that reference resolution is influenced by a

⁸ Carminati (; Italian) and de la Fuente and Hemforth (; Spanish) used intrasentential contexts (main clauses and subordinate clauses) whereas Alonso-Ovalle et al. (; Spanish) used two separate sentences, as did Mayol and Clark (; Catalan). However, Filiaci et al. () compared Spanish and Italian directly and used stimuli parallel to those of Carminati in both languages. Thus, the differences observed between Spanish and Italian by Filiaci et al. cannot be attributed to differences in inter- versus intrasentential anaphora.

OUP CORRECTED PROOF – FINAL, 21/1/2019, SPi

 



range of information, structural, and syntactic considerations (e.g., BA and BEI constructions in Chinese) and finds that null and overt pronouns do not exhibit a very clear division of labor (see also Simpson et al.  on overt pronouns in Chinese). Ueno and Kehler (, ) found that Japanese overt pronouns are influenced by semantic-/coherence-related factors such as telicity (similar to English pronouns, see Rohde et al. , see also §.), whereas null pronouns showed an overarching preference for subjects. This asymmetrical behavior is in line with Kaiser and Trueswell’s () form-specific account, which posits that different anaphoric forms may differ in how sensitive they are to different types of information. A study by Kwon and Sturt () on Korean suggests that Korean null pronouns (the most reduced form available in Korean) may have a stronger preference to ‘latch onto’ a preceding discourse topic than English overt pronouns (the most reduced form available in English). Kwon and Sturt conclude that “individual languages might assign different weights to different cues for pronoun resolution” (: ). As a whole, work on languages that have both null and overt pronouns provides evidence for the general idea that subjects tend to be privileged as antecedents, but also supports the form-specific approach (Kaiser and Trueswell ) which argues that different referential forms can differ with regards to which properties of the antecedent they are most sensitive to. A distinction related to the null/overt contrast is the distinction between strong and weak pronouns that exists in some languages (§..). Kaiser (b) conducted a series of experiments on Dutch emphatic ‘strong’ pronouns and non-emphatic ‘weak’ pronouns (and compared them to demonstratives), to see if they pattern as predicted by salience hierarchies such as the one in (). The results indicate that while the differences between demonstrative pronouns and non-emphatic (‘weak’) personal pronouns correlate with the antecedent’s grammatical role, the distinction between strong and weak pronouns cannot be satisfactorily explained by grammatical role. Instead, the results indicate that the strong form is sensitive to the presence of contrast, providing additional evidence that referring expressions can differ in what kind of information they are most sensitive to.

. E         

.................................................................................................................................. So far, our discussion has explored claims about how reference resolution is influenced by properties of the antecedent, such as its grammatical role, thematic role, linear position, and information status. A rather different view is advocated by Hobbs (, ) and Kehler (). They argue for a coherence-based account, where pronoun

OUP CORRECTED PROOF – FINAL, 21/1/2019, SPi



    

use and interpretation depends on semantic coherence relations between clauses. Crucially, this approach views pronoun interpretation as a by-product of general inferencing/reasoning about relations between clauses (e.g., Hobbs ; Kehler ; Kehler, Kertz, Rohde, and Elman ). To see how coherence relations can influence pronoun interpretation, consider example () below. In principle, ‘him’ could refer to Phil or Stan. If the relation between the sentences is semantically parallel (a), people tend to interpret ‘him’ as referring to the parallel argument: Stan was tickled and he was also poked (Kertz, Kehler, and Elman ). However, if the relation between the two clauses is a causal result relation (b), people are more likely to interpret ‘him’ as the subject Phil (Kertz et al. ). () a. Phil tickled Stan, and similarly Liz poked him. Parallel relation: him => bias to object (Stan) b. Phil tickled Stan, and as a result Liz poked him. Result relation: him => bias to subject (Phil) Subject-position pronouns are also sensitive to coherence, as illustrated in (c–d) by the switch from a subject preference in (c)—with a temporal relation between the two clauses—to an object preference in (d), with a result relation between the two clauses (see Kertz et al. ; Kehler ). () c. Phil tickled Stan, and then he laughed at Bob’s joke. Narrative relation: he => bias to subject (Phil) d. Phil tickled Stan, and as a result he laughed uncontrollably. Result relation: he => bias to object (Stan) However, these examples should not be construed as evidence for a fixed mapping between coherence relations and grammatical role: as Kertz et al.’s examples in () show, there is not a fixed relationship between coherence relations and grammatical roles. A subject pronoun in a result relation does not have to refer to the preceding object: both (a) and (b) involve a result relation, and in one case this triggers an object preference (a) and in the other case a subject preference (b). What actually matters is the semantics of the clauses and their relation to each other. (Kehler’s work adopts a neo-Humian set of coherence relations with a basic division between resemblance, cause–effect, and contiguity relations as well as their subtypes, see also Asher and Lascarides () and Mann and Thompson () for somewhat different sets of relations, and Hovy () for a summary of different coherence relations proposed in the literature). () a. Peter snapped at Ethan, and he sulked the rest of the afternoon. Result relation: he => bias to object (example from Kertz et al. ) b. Peter snapped at Ethan, and he felt guilty the rest of the afternoon. Result relation: he => bias to subject (example from Kertz et al. ) There also exists a large body of related work on so-called implicit causality (IC) verbs, starting with Garvey and Caramazza () who observed that some verbs, when

OUP CORRECTED PROOF – FINAL, 21/1/2019, SPi

 



followed by an explanation relation (as signaled by a ‘because’ connective), had strong biases towards a subject continuation (e.g., (c) ‘he’ tends to be interpreted as the prisoner) whereas other verbs mostly elicited object continuations (e.g., (d) ‘she’ tends to be interpreted as the daughter). Thus, IC verbs are often classified as NP (subject-biased, e.g., confess, disappoint, impress) or as NP (object-biased, e.g., admire, detest, punish), or as something in-between (see e.g., Ferstl et al. ; Hartshorne and Snedeker  for norming studies on English verbs). () c. The prisoner confessed to the guard because he . . . d. The mother punished her daughter because she . . . These biases are dependent on the presence of an explanation coherence relation (see Rohde ) and have been found to guide comprehenders’ expectations about upcoming referents early on, immediately as a pronoun is encountered (e.g., Koornneef and Van Berkum ) and even before the causal connective or pronoun is encountered (Pyykkönen and Järvikivi , see related work by Fedele  which also highlights the importance of sentence boundaries). As a whole, a number of studies indicate that a successful account of pronoun interpretation needs to take into account the semantic coherence relations that hold between clauses (e.g., Wolf, Gibson, and Desmet ; Kertz et al. ; Kehler et al. ; see also Kaiser c on German; Simpson et al.  on Mandarin Chinese, and many others).

.. Steps towards reconciling coherence-based and salience-based approaches The findings about the importance of discourse structure seem to be at odds with the results discussed earlier, such as the subjecthood preference (§..). If coherence relations play a deep role in guiding pronoun resolution, how we can reconcile this with the other results suggesting that syntactic factors like subjecthood or semantic theta-roles like agentivity guide pronoun interpretation? This tension can be at least partially resolved if we distinguish (i) likelihood of mention from (ii) likelihood of pronominalization—a distinction which is supported by a growing body of work (see Kehler et al. ; Fukumura and van Gompel ; Kaiser a; Kehler and Rohde . For example, in a series of studies with stimulus-experience and experiencer-stimulus verbs, Fukumura and van Gompel () found that pronoun production was sensitive to subjecthood and distinct from the question of how likely the subject was to be mentioned. Similarly, Kaiser (a) investigated pronoun interpretation and production after it-clefts, and found that when participants were asked to interpret a pronoun or chose to produce a pronoun, they showed a strong bias for the preceding subject (regardless of the presence/absence of focus). However, when participants were able to freely generate different kinds of referring expressions, both

OUP CORRECTED PROOF – FINAL, 21/1/2019, SPi



    

information structure (focus versus topic) and grammatical role influence which referent is mentioned next. In recent work, Kehler and Rohde () explicitly argue for a reconciliation between coherence-based approaches and subjecthood/topicality-based approaches such as Centering Theory (Grosz et al. ). They show that neither coherencedriven nor Centering-driven approaches alone are able to fully capture the results of a series of experiments. Building on ideas introduced in Kehler  and Kehler et al. (), they argue for a Bayesian approach in which pronoun interpretation is influenced both by (i) comprehenders’ expectations about coherence relations and by (ii) production-based constraints that modulate speakers’ decisions about whether to produce a pronoun, a process which is sensitive to topicality effects.

.. Mental representations of coherence relations According to Hobbs (, ), inferences about coherence relations are independent of pronoun interpretation. This suggests that the coherence representations activated during pronoun interpretation are the same as the representations activated during non-pronominal coherence establishment, and raises questions about how these representations relate to more domain-general representations about relations such as causality. Kaiser () used visual-world eye-tracking and priming to investigate these issues. Participants were presented with non-linguistic, visuo-spatial primes: silent video clips that encoded (i) cause-effect/result relations, (ii) similarity relations, or (iii) other/neutral relations (see (a) for a cause–effect/result video prime). In another experiment, the coherence primes were linguistic and encoded different coherence relations (see (b) for a cause–effect/result relation prime). Participants were then shown a target image with three characters and heard a sentence with nonce verbs, an ambiguous ‘and’ connective, and an ambiguous object pronoun (c). Nonce words eliminated effects of verb semantics. The task was to click on the last-mentioned referent (i.e., the antecedent of him/her). () a. Cause-effect relation visuo-spatial prime: In a silent video, a triangle knocks into a circle which then falls off a ledge. b. Cause-effect relation linguistic prime: The patient pressed the red emergency button near the bed and a nurse quickly ran into the room. c. Target sentence: Phil linded Stanley and Kate hepped him. The results show that both linguistic and non-linguistic primes influenced processing of the ambiguous object pronoun in the target sentence. For example, watching a causal event where a triangle knocks a circle off a ledge made people more likely to consider the preceding subject (e.g., Paul) as an antecedent for the object pronoun (e.g., him)

OUP CORRECTED PROOF – FINAL, 21/1/2019, SPi

 



in the unrelated test sentence (c), presumably because the prime influenced construal of the coherence relation in the test sentence. This suggests pronoun interpretation can be primed by coherence relations in preceding linguistic and visual (non-linguistic) input, even when primes and targets are connected only on the level of abstract coherence relations. This is the first direct evidence for the existence of shared representations between non-pronominal coherence-related inferencing and pronoun resolution processes. Evidence for the domain-generality of coherence relations comes from Kaiser (), who investigated whether having participants carry out motor actions (e.g., causal actions such as using a small ball to knock over toy bowling pins) could prime their performance in a subsequent sentence completion task (with prompt sentences that were semantically unrelated to the actions, e.g., Lisa pinched Nancy). The results show priming of causal relations from motor actions to linguistic tasks: carrying out a causal action made people more likely to write causal continuations, as compared to non-causal motor actions. Kaiser also observed priming on the referential level: executing motor actions containing two distinct events made people more likely to produce sentences with two referentially-distinct subjects (topic shift) in the linguistic task, when compared to the execution of a one-event motor action. This suggests that shifting one’s attention from one event to another resembles the act of shifting one’s attention from one referent to another and points to similarities in our mental representations of events and entities.

.. Coherence and perspective-taking In addition to the coherence relations between sentences, pronoun interpretation is also influenced by the perspectival structure by discourse, such as free indirect discourse (FID). Free indirect discourse is a narrative/literary style used to convey a character’s thoughts, for example, “Peter was getting really tired of sleeping on Tom’s couch. How could anyone sleep on that old thing? He would go home tomorrow. Nothing would change his mind about that” (from Kaiser ). In contrast to direct speech and reported speech, free indirect discourse presents a character’s speech or thoughts without embedding or explicit quotation marks (cf. Peter thought, “I will go home tomorrow”). Kaiser () tested sentences like () and (), manipulating whether or not they contained expressions known to signal FID, namely (i) evaluative expressions/epithets (e.g., ‘poor girl’) and (ii) epistemic adverbs (e.g., ‘probably’). Based on narratological work (e.g., Banfield ; McHale ; Fludernik ), epithets and epistemic adverbials should make participants more likely to interpret the second sentence as being from the perspective/point-of-view of one of the characters (here, the subject of the first sentence). Thus, in (b), if participants are sensitive to the FID cue, they

OUP CORRECTED PROOF – FINAL, 21/1/2019, SPi



    

should interpret the second sentence as Mary’s thoughts about Elizabeth. Similarly, if the epistemic adverbial cues FID in (b), participants should interpret the second clause as Luke’s thoughts about Andrew. ()

a. Mary looked woefully at Elizabeth. She was sick. [regular sentence] b. Mary looked woefully at Elizabeth. Poor girl; she was sick. [FID cue]

()

a. Luke glanced at Andrew warily. He’d put toothpaste in the shampoo bottle again. [regular sentence] b. Luke glanced at Andrew warily. He’d probably put toothpaste in the shampoo bottle again. [FID cue]

Test items were designed so that in the plain/non-cued version (a) and (a), the pronoun could refer to the preceding subject or object/oblique argument, perhaps with a slight bias for the subject. However, in the cued versions (b) and (b), if readers recognize a perspective shift to the point-of-view of the subject character, they should interpret the pronoun as referring to the preceding object. If readers interpret the epithet ‘poor girl’ as signaling that the second sentence is Mary’s thoughts about Elizabeth, they should interpret ‘she’ as referring to Elizabeth. Indeed, Kaiser () found that participants are more likely to interpret the pronoun as referring to the preceding object in the presence of epithet/evaluative expressions or adverbs of possibility, as compared to the regular/plain versions. This suggests that participants are sensitive to the perspective shift triggered by FID cues, and that this impacts pronoun resolution. These results pose challenges for theories of pronoun resolution, because current theories offer no straightforward way of capturing the different rates of subject versus object interpretation in the presence/absence of evaluatives/epithets and adverbs of possibility. The finding that subtle cues to perspective-shifting have significant effects on pronoun interpretation highlights the importance of going beyond ‘surface heuristics’ and acknowledging the role of meaning-based factors. Work on coherence relations between clauses (e.g., Kehler and Rohde ) goes beyond an overarching subject bias but still does not straightforwardly predict differences between sentences with and without FID cues. Thus, additional work is needed to capture these perspectival effects.

. C    

.................................................................................................................................. In this section, we consider the consequences of humans’ limited processing abilities: although we are very skilled at producing and interpreting referring expressions, our working memory and attentional resources are nevertheless limited (e.g., Baddeley ; Kahneman ). In this section, we review (i) consequences of processing

OUP CORRECTED PROOF – FINAL, 21/1/2019, SPi

 



limitations on the production and interpretation of pronouns, (ii) the possibility of ‘shallow processing’ as a means to decrease processing cost, and (iii) how the cognitive cost of holding unresolved referential dependencies in working memory impacts comprehension of cataphoric pronouns.

.. Limited attentional resources When people’s attentional resources are limited, this impacts how they interpret and produce pronouns. For example, on the comprehension side, Van Rij et al. () conducted a dual-task self-paced reading experiment in Dutch where participants had to remember a series of numbers while reading four-sentence stories mentioning two same-gender characters. Van Rij et al. manipulated whether or not the stories involved a topic shift. In the no-topic shift stories, the same character was realized in the subject position of the first three sentences, and the last sentence had a subject-position pronoun (e.g., Eric–Eric–Eric–He). In the topic-shift stories, the subject changed between the first and the second sentence, and the last sentence again had a subject-position pronoun (e.g., Eric–Philip–Philip–He). (In sentences  and , the other referent was mentioned in object position). The stories were followed by comprehension questions testing participants’ interpretation of the pronoun (as well as a number recall task). Van Rij et al. also manipulated whether participants had to memorize three numbers (low working memory load) or six numbers (higher working memory load). Importantly, prior work by Hendriks, Koster, and Hoeks () found that in topic-shift configuration contexts without a memory load, adults tended to interpret the ambiguous pronoun as referring to the subject of the immediately preceding sentence (e.g., Philip), whereas children tended to interpret it as referring to the initial subject (e.g., Eric). Hendriks et al. attribute this to children having a lower working memory capacity. The results of Van Rij et al.’s dual-task experiment with adults show that a higher working memory load makes it harder to process a topic shift configuration—in other words, the adults pattern more like children: in the low-load condition, adults show a stronger preference to interpret the pronoun as referring to the immediately preceding subject than in the high-load condition. Van Rij et al. conclude that pronoun interpretation patterns—in this case, a bias for the local subject—can fluctuate depending on the available working memory resources. In addition to studies showing effects of processing limitations on pronoun comprehension, researchers have also identified consequences of attentional resource limits on pronoun production. Arnold and Griffin () found that characters present in the discourse context compete for our (limited) attentional resources and impact the production of referring expressions. In a picture-based storytelling task in English, Arnold and Griffin () manipulated whether one or two characters were present in the pictures (e.g., only Mickey, or both Mickey and Daisy). Participants produced more pronouns in one-

OUP CORRECTED PROOF – FINAL, 21/1/2019, SPi



    

character contexts than in two-character contexts—even though the two characters had different genders, so a pronoun would have been unambiguous. They conclude that people’s attention is divided between the two characters, decreasing the attentional resources that can be allotted to the to-be-mentioned character and resulting in fewer pronouns. This is in line with the idea that character needs to attain a certain threshold of prominence/accessibility in the speaker’s mind before it can be pronominalized (see also Fukumura, Hyönä, and Scholfield  for related work on Finnish, a language that does not mark gender on pronouns). Divergent results come from Vogels, Krahmer, and Maes (), who found that speakers produce more pronouns—not less—when they are under cognitive load and have fewer attentional resources available for the reference-production task. In a production study in Dutch, speakers provided continuations to sentence fragments combined with visual scenes (e.g., A girl was arguing with a boy. The boy got really annoyed. Subsequently . . . ). Vogels et al. manipulated whether speakers were under cognitive load by using a verbal word recall task. In contrast to what one might expect based on Arnold and Griffin (), Vogels et al. found that speakers produced more pronouns when they were under cognitive load, especially when referring to lowersalience characters. This difference in findings may be due to the way in which cognitive load was manipulated—specifically, whether attention is divided between one or more characters or one or more tasks. Interesting questions also emerge regarding the relation (or lack thereof ) between cognitive load and whether a speaker adjusts their speech for the benefit of the address or not (see also Fukumura and van Gompel ; Galati and Brennan  for discussion of audience design).

.. Processing depth An issue related to limited attentional resources is processing depth. Traditionally, researchers have often assumed that pronouns are fully resolved. Recent evidence, however, suggests that sentence processing may be quite shallow and may not result in a fully fleshed-out representation of the sentence. Instead, comprehenders may construct a superficial ‘good enough’ representation that leaves some aspects of the sentence unspecified (e.g., Ferreira et al. ; Sanford and Sturt  on gardenpath sentences). Does this kind of shallow processing occur with pronouns as well? Recent work by Stewart et al. (); Love and McKoon (); and Kaiser and Do () points in that direction (see also Greene et al. ). Stewart et al. () measured whole-sentence reading times for sentences such as (), with ambiguous or unambiguous pronouns. They manipulated processing depth by means of the number and type of comprehension questions: in the deep processing condition, all targets and fillers were followed by comprehension questions which, in the case of targets, required participants to resolve the pronoun (e.g., Who left for the holidays?). In the shallow processing conditions, only a third of items were followed by

OUP CORRECTED PROOF – FINAL, 21/1/2019, SPi

 



questions, and the questions could be answered without resolving the pronoun (e.g., Did Paul lend Rick the CD?): () Paul lent {Rick/Kate} the CD before {he/she} left for the holidays. In the shallow-processing condition, reading times showed no effects of whether pronouns were ambiguous or unambiguous. In contrast, when forced to engage in deep processing, people’s reading times in the ambiguous-pronoun conditions were longer than in unambiguous-pronoun conditions, suggesting that people (when engaged in deep processing) need to invest more processing effort to come to a final interpretation of an ambiguous pronoun. These results—as well as those from a second experiment—suggest that pronoun interpretation is modulated by depth of processing, and that during shallow processing, ambiguous pronouns may be initially underspecified and resolved later if/when disambiguating information is provided (Stewart et al. ). Additional evidence in line with this comes from Love and McKoon (), who used a probe task to measure whether comprehenders resolved unambiguous pronouns in long versus short texts. As in Stewart et al. (), the pronouns were unambiguous due to the presence of two different-gender referents. Love and McKoon’s results suggest that in short texts (which they show in a separate study to be less engaging), unambiguous pronouns were not resolved, in contrast to pronouns in longer, more engaging texts. In subsequent work, Kaiser and Do () investigated whether use of a noncategorical, probabilistic cue—coherence relations—would also be similarly susceptible to processing depth effects. In examples like (), the referent of the pronoun in the second sentence can be inferred to be Elaine, thanks to the semantic cause–effect relation between the two sentences. () Elaine protected Jen. Stephen awarded her a special prize for bravery and everyone cheered happily. To test whether differences in processing depth influence whether comprehenders use verb-based coherence-related inferences to guide pronoun resolution, Kaiser and Do () manipulated (i) the informativity of the verb in the first sentence by using real verbs versus nonce verbs and (ii) the ambiguity of the pronoun in the second sentence, (). The real verbs could be used to infer the intended antecedent of the pronoun; the nonce verbs provided no such information. Processing depth was manipulated with questions, similar to Stewart et al. (). () a. Elaine protected {Joel/Jen}. Stephen awarded her a special prize for bravery and everyone cheered happily. b. Elaine brondled {Joel/Jen}. Stephen awarded her a special prize for bravery and everyone cheered happily. In the second half of the experiment (when processing-depth strategies should be most clearly visible), pronoun ambiguity resulted in a bigger slowdown in the nonce-verb conditions than the real-verb conditions, but only in the deep-processing group. This

OUP CORRECTED PROOF – FINAL, 21/1/2019, SPi



    

suggests that whether people use verb-based probabilistic information for pronoun resolution is modulated by processing depth, and confirms prior findings that under shallow processing, the interpretation of pronouns is not fully automatic. As a whole, processing-depth modulations can impact how readers use both simple, categorical cues and probabilistic cues that require more complex inferencing. This has theoretical implications for our understanding of anaphor resolution as well as methodological implications for experimental design.

.. Cataphora versus anaphora The consequences of processing limitations can also be seen in the domain of cataphora, a configuration where a pronoun appears before potential antecedents (see example ()). With cataphora, the language processing system essentially has to keep an unresolved dependency in mind and to look for a suitable antecedent downstream. ()

a. When Peter came home, he made a sandwich. [anaphora] b. When he came home, Peter made a sandwich. [cataphora]

It was observed early on that real-time comprehension of cataphora is subject to both grammatical and processing-based constraints. Cowart and Cairns () were one of the first to observe that people have a strong preference to link cataphoric pronouns with the first available noun phrase encountered after cataphor. With sentences such as (a) and (b), Cowart and Cairns showed that people try to associate the cataphoric ‘they’ (but not the non-cataphoric second-person ‘you’) with the first available noun phrase, making ‘charming babies’ more likely to be analyzed as a noun phrase than a gerund, despite the lack of semantic plausibility (babies don’t lecture). In contrast, in sentences such as (c) and (d), Principle C of the Binding Theory rules out coreference between ‘they’ and the subsequent noun ‘uncles’ (due to c-command) and comprehenders do not associate ‘they’ with ‘uncles’. In sum, the parsing system tries to interpret a pronoun as referring to the first subsequent noun phrase, but only if it is grammatically accessible. ()

a. b. c. d.

Whenever they lecture during the procedure, charming babies . . . Whenever you lecture during the procedure, charming babies . . . If they want to believe that visiting uncles . . . If you want to believe that visiting uncles . . .

Van Gompel and Liversedge () built on this work using eye-tracking. In sentences such as those in example (), they manipulated whether the pronoun’s gender matches (a) or mismatches (b) the gender of the following subject noun (e.g., the girl). They observe processing difficulties with gender-mismatch sentences, suggesting that comprehenders attempt to associate the cataphoric pronoun to the first noun, initially ignoring the specific gender features of that noun. Broadly speaking, these results indicate that parsers are ‘impatient’ and try to resolve the pronoun as early as possible.

OUP CORRECTED PROOF – FINAL, 21/1/2019, SPi

 



() a. Gender Match: When she was fed up, the girl visited the boy very often. b. Gender Mismatch: When he was fed up, the girl visited the boy very often. Extending these investigations beyond English, we looked at cataphora and anaphora in Italian (Fedele and Kaiser ), to see how the referential biases of null and overt pronouns in Italian (see §.) modulate cataphoric processing (see also Kazanina and Phillips  on cataphora in Russian). We tested anaphora and cataphora will null and overt pronouns, using sentences such as (a) and (b) in a questionnaire task that probed people’s interpretation of null and overt pronouns by means of comprehension questions (e.g., Who talked about the trip to London?). () a. Anaphoric configuration Maria abbraccia Rita, mentre {ø/lei} parla del viaggio a Londra. Maria hugs Rita, while {null/she} speaks about-the trip to London ‘Maria hugs Rita, while she talks about the trip to London.’ b. Cataphoric configuration Mentre {ø/lei} parla del viaggio a Londra, Maria abbraccia Rita. While null/she speaks about-the trip to London, Maria hugs Rita ‘While she talks about the trip to London, Maria hugs Rita.’ Prior work on anaphora found that in Italian, null pronouns generally prefer subjects and overt pronouns prefer non-subjects (§.). We also found this for the anaphoric configurations. However, in the cataphoric configurations, (i) the subject preference of null pronouns becomes even stronger in the cataphoric configuration (compared to the anaphoric configuration), whereas (ii) the object preference of overt pronouns is weakened in the cataphoric configuration (compared to the anaphoric configuration). We suggest that this seemingly counter-intuitive pattern stems from a combination of general processing biases and form-specific referential biases: the processor tries to associate cataphora with the first available entity (e.g., Cowart and Cairns ), which in our sentences is the subject. This bias to ‘latch onto’ the first available referent pushes both null and overt pronouns towards the subject, which is why—in the cataphor configuration—the subject preference of null pronouns is strengthened and the object preference of overt pronouns is weakened. According to this interpretation, processing effects are powerful enough to weaken the inherent referential biases of particular referring expressions (e.g., the overt pronoun’s object bias).

. C,  ,  

.................................................................................................................................. The question of how listeners and readers interpret anaphoric expressions that refer to previously-mentioned entities continues to be an active area of research. It is enriched by an increasing amount of cross-linguistic work, a variety of research methods, and the interaction of different theoretical frameworks. However, there are still many

OUP CORRECTED PROOF – FINAL, 21/1/2019, SPi



    

empirical areas of inquiry that are relatively under-researched from a psycholinguistic perspective, including (but not limited to) patterns of reference to non-human referents, such as inanimate objects, propositions, events, and discourse segments (but see Brown-Schmidt et al. , Çokal et al. ), as well as the interpretation and processing of pronouns that do not pick out ‘concrete’ referents in the discourse model (e.g., e-type pronouns/donkey anaphora; Heim ). Many recent psycholinguistic discussions on reference resolution have also tended to focus largely on singular pronouns, with the consequence that research on processing of plural pronouns is less well-integrated into many existing processing theories than one might expect. As a whole, existing work suggests that reference resolution is a multi-faceted process influenced by various aspects of linguistic structure, discourse structure, limits on human attention, and processing constraints. In future work, it will be important to continue exploring how these different pieces best fit together to help solve one of the key puzzles of human communication.

OUP CORRECTED PROOF – FINAL, 17/1/2019, SPi

        ......................................................................................................................

ACCESSIBILITY AND REFERENCE PRODUCTION The interplay between linguistic and non-linguistic factors ......................................................................................................................

 ,  ,   

. I

.................................................................................................................................. R EFERENCE is an essential part of language. When we speak, we talk about something. The act of referring can be seen as forming a link between the speaker’s mind and the outside world. For example, a speaker asking ‘could you hand me that stapler?’ is expressing her¹ intention to get hold of a physical object in the world by referring to that object with a linguistic expression (in this case, the definite noun phrase ‘the stapler’). The things we refer to are, however, not always physical objects (including people), nor do they need to be part of the outside world. For example, we can refer to objects that are not present in the direct physical environment (‘I left the stapler in the office’), or objects that only exist in our imagination (‘the stapler I dreamt about last night’). We can refer to objects that existed in the past (‘the cake that I ate yesterday’), or will exist in the future (‘the cake that I will bake tomorrow’). We can also refer to events (‘last night’s dinner party’), locations (‘the picturesque town of Tilburg’), and abstract concepts (‘the financial crisis’), to name a few. In none of these situations is the thing that is being referred to (the referent) an object in the directly perceivable world. It would therefore be better to say that we refer to conceptualizations in our minds, rather than to objects in the outside world (e.g., Jackendoff ; Johnson-Laird ). Even in those cases where the referent is present in the world, reference is still mediated ¹ Following common practice, feminine forms will be used to refer to speakers, and masculine forms to refer to hearers.

OUP CORRECTED PROOF – FINAL, 17/1/2019, SPi



 ,  ,   

by a conceptualization of the referent (which may be wrong, as in ‘Could you hand me that stapler?’ ‘That’s not a stapler, that’s a hole punch.’). One of the main questions in linguistic research is how speakers put such conceptualizations into language. In Levelt’s model of language production (Levelt ), a speaker who wants to communicate about a certain entity has to make a number of important decisions. First, she has to decide which information to include in the utterance, that is, she needs to select the content of the message to be expressed. Once a relevant concept has been selected, it has to be put into a grammatical structure. Given that speech proceeds serially, this structure ultimately has to map on a linear order of words. That is, one thing has to be mentioned before another. Hence, a speaker needs to choose a concept that will be referred to first. Second, the speaker has to decide which linguistic form she is going to use to refer to a certain concept. That is, she has to choose a referring expression. This chapter reviews research that is concerned with the factors that influence a speaker’s referential choices in language production, that is, how they choose which referent they mention first, and how they choose what type of referring expression to use for a certain referent. In particular, it explores the small amount of data that exists on how linguistic factors (e.g., grammatical function; topicality) interact with nonlinguistic factors (e.g., perceptual salience; animacy) in the choice of a referring expression in discourse. Research on the choice of referring expressions has been primarily concerned with the role of the linguistic context. By contrast, the influence of nonlinguistic factors on the choice of referent for first mention (or, more generally, choice of word order) has been investigated quite extensively. Therefore, this chapter focuses on the question to what degree such non-linguistic factors also influence the choice of referring expression. The area of interest is the choice of a particular type of referential form in a discourse (e.g., a pronoun or a full noun phrase), rather than the selection of semantic content to include in an expression to describe something (e.g., how speakers choose between ‘the large stapler’, ‘the red stapler’, and ‘the large red stapler’). This chapter is organized as follows: §. briefly discusses some terminological issues associated with the notion of accessibility, which is assumed to underlie many referential choices. §. reviews the most important research on these referential choices. §. discusses recent studies that investigated interactions between linguistic and nonlinguistic factors in the choice of referring expression in discourse. In §., different viewpoints on the nature of accessibility effects described in the literature are discussed, and a tentative proposal of how seemingly contradictory effects might be unified is presented. Finally, §. concludes and provides suggestions for further research.

. A   

.................................................................................................................................. Both the choice of referent for first mention and the choice of referring expression have been related to the degree of activation of the conceptualizations—or mental representations—of referents in memory. The higher the activation of a certain

OUP CORRECTED PROOF – FINAL, 17/1/2019, SPi

   



representation, the more likely it is to appear early in the linguistic structure, and the higher the likelihood that the expression referring to it is more attenuated (e.g., Levelt ). This activation status has been described with a variety of terms, such as accessibility (Ariel ; Bock and Warren ), salience (Osgood ; Sridhar ), cognitive status (Gundel, Hedberg, and Zacharski ), givenness (Chafe ; Gundel et al. ; Prince b), topicality (Givón ) and focus of attention (Grosz, Joshi, and Weinstein ), each with slightly different assumptions and viewpoints. Some of these terms, such as givenness and topicality, emphasize the importance of information structure in the discourse. For example, when a referent was the topic of the preceding sentence, with topic being defined as what the sentence is about (Reinhart ), its representation in memory is likely to be highly activated. Other terms—such as focus of attention and cognitive status—emphasize the importance of cognitive capacities. For example, it seems likely that those referents that are attended to are more activated, since they may be actively maintained in memory (Foraker and McElree ). To remain implicit as to the source of the activation, we will use the more general term accessibility to refer to the activation of mental representations in the memories of speakers and hearers, whatever the cause. For the sake of brevity, we will often use ‘the accessibility of a referent’ as shorthand for ‘the accessibility of the mental representation of a referent’. Crucially, the notion of accessibility concerns activation of non-linguistic representations, rather than activation of lexical items in the mental lexicon (cf. Arnold ). To distinguish the two types of activation, Bock and Warren () speak of conceptual accessibility, which they define as “the ease with which the mental representation of some potential referent can be activated in or retrieved from memory” (: ), as opposed to lexical accessibility, which refers to “the ease with which the representations of word forms can be recovered from memory” (: ). In this chapter, the term accessibility is used to refer to conceptual accessibility, unless explicitly specified otherwise. Furthermore, the term salience is reserved for properties of the referent itself rather than of its representation in memory. These properties can be linguistic, as when the referent is mentioned in a prominent or non-prominent syntactic position,² or non-linguistic (e.g., perceptual), as in the size or color of the physical object that is referred to. They can also be determined by the context, such as the preceding discourse or the physical environment, or they can be intrinsic to the referent, such as animacy. Finally, the terms topicality and givenness are taken to denote factors that contribute to an entity’s (linguistic) salience, while focus of attention is used as a speaker- or hearer-internal factor that might influence accessibility directly. Thus, accessibility is thought to be a determining factor both in the choice of referent for first mention and in the choice of referring expression. However, research has revealed differences in how exactly accessibility affects these choices. Notably, there may be differences in the degree to which the two types of referential choice are affected ² Depending on the language, prominent syntactic positions include the subject, topic, or preverbal position, for example, while object, focus, and post-verbal positions may be considered non-prominent.

OUP CORRECTED PROOF – FINAL, 17/1/2019, SPi



 ,  ,   

by certain factors, linguistic as well as non-linguistic (e.g., Fukumura and van Gompel ; Rohde and Kehler ; Stevenson, Crawley, and Kleinman ). The next section reviews relevant literature on the role of accessibility in the choice of referent for first mention (§..) and in the choice of referring expression (§..).

. F   

..................................................................................................................................

.. Effects of accessibility on the choice of referent for first mention When people speak (or write), seemingly unordered thoughts and concepts have to be put into the linear order that characterizes language. As a first decision, speakers have to choose a starting point from which their utterance is going to unfold (MacWhinney ). The question is how they do this. Research suggests that the salience of the concepts people talk about plays an important role. It has frequently been observed that when people describe simple visual scenes, salient entities are mentioned earlier than less salient entities. For example, entities that are perceptually more prominent are typically found to have a higher likelihood to be produced early in the sentence than perceptually less prominent entities (e.g., Flores d’Arcais ; Gleitman, January, Nappa, and Trueswell ; Osgood ; Sridhar ; Tomlin ). Gleitman et al. (), for instance, presented participants with simple scenes, for example, of a dog chasing a man, and very briefly showed a not consciously noticeable attentional cue (a black square), either on the dog or on the man. They found that the scenes were more likely to be described with active sentences (‘the dog chases the man’) when the cue was on the dog, and more likely to be described with passive sentences (‘the man is chased by the dog’), when the cue was on the man. Another important factor that has been found to influence order of mention is animacy. Human or animate entities tend to be produced before inanimate entities in a sentence (e.g., Branigan, Pickering, and Tanaka ; McDonald, Bock, and Kelly ; Prat-Sala and Branigan ; Van Nice and Dietrich a). For example, in a series of sentence recall experiments by McDonald et al. (), participants were more likely to recall originally passive sentences with an inanimate subject as active sentences with an animate subject. Other studies have concentrated on the production of sentences within a discourse. Here, it has been found that given information is more likely to be mentioned first in a sentence than new information (Bock and Irwin ; Ferreira and Yoshita ). For example, Ferreira and Yoshita () found that participants were more likely to recall Japanese dative constructions such that given information preceded new information, even when the original sentences had the reverse order. In addition, information that is

OUP CORRECTED PROOF – FINAL, 17/1/2019, SPi

   



predictable from the discourse context may be produced early in the sentence (Arnold ; Givón ), although predictable entities may also be postponed to a less prominent position, due to a preference to start an utterance with the most important (i.e., most newsworthy) information (Givón , ; Gundel ). These saliency effects on the choice of referent for first mention have generally been attributed to the conceptual accessibility of referents’ mental representations in the speaker’s memory (e.g., Bock and Warren ; Prat-Sala and Branigan ). If we assume that language production proceeds incrementally (e.g., Kempen and Hoenkamp ; Levelt ), speakers start producing an utterance before the planning of that utterance is completed. Because highly accessible referents are more easily retrieved from memory, they are subsequently mentioned earlier in the sentence. However, the exact relation between degree of accessibility and order of mention is not yet completely clear. With respect to perceptual salience, for example, Gleitman et al. () argued that when a speaker’s visual attention is focused on a referent, this speeds up the retrieval of lemma representations, and hence directly affects order of mention. Other researchers argue instead that in picture descriptions, visual attention does not influence the mentioning of entities directly. Rather, the order in which entities are produced is determined by the global apprehension of the scene, for example, who does what to whom (Bock, Irwin, and Davidson ; Griffin and Bock ). It is in this apprehension stage that visual attention may have an influence (see also Vogels, Krahmer, and Maes a). The relation between accessibility and the positioning of concepts in the sentence may also be mediated by grammatical function or topichood. In English, for example, it has been found that the most accessible concept is typically made the subject (e.g., Bock and Warren ; McDonald et al. ). This suggests that accessibility determines which entity becomes the subject of the sentence, which in turn is preferably produced in the sentence-initial position, but that it does not determine sentence position directly. However, subject and sentence-initial position are highly confounded in English, which makes the exact relation between accessibility and sentence position unclear. In languages in which word order is more free than in English, such as Greek (Branigan and Feleki ), German (Kempen and Harbusch ), Hungarian (É. Kiss ), Italian, and Spanish (Brunetti ), accessibility has been found to affect sentence position independently of grammatical function. However, in such languages, accessibility may instead affect the likelihood of a referent becoming a topic. This referent will in turn be more likely to occupy the topic position, which is often the first position in the sentence (Lambrecht ). In a study of the Algonquian language Odawa, Christianson and Ferreira () were able to disentangle effects on both grammatical function and topichood from those on linear order. They found that accessible entities in Odawa were not directly promoted to the sentence-initial position, but were given prominent syntactic functions via the priming of a particular syntactic structure. Different saliency factors may also interact. Combining picture descriptions with the presence of a discourse context, Prat-Sala and Branigan () found that a referent’s

OUP CORRECTED PROOF – FINAL, 17/1/2019, SPi



 ,  ,   

intrinsic salience (in this case, animacy) and its givenness in the discourse interact in the decision to mention this referent first in an utterance: for given entities, the influence of intrinsic salience was greater than for new entities. This led Prat-Sala and Branigan to distinguish two types of accessibility: a referent’s inherent accessibility refers to activation in memory caused by its intrinsic properties—such as animacy or concreteness— which are assumed to be stable across contexts. Within a discourse, this inherent activation can be supplemented by the referent’s derived accessibility, a temporary level of activation caused by the salience of the referent in the discourse, such as whether it is given or topical. Thus, a referent’s derived accessibility adds to its inherent accessibility. However, if the two types of accessibility run counter to each other, such as when the referent is inanimate but given, derived accessibility may override inherent accessibility in strong enough contexts (Prat-Sala and Branigan ). Central to this chapter is the question whether the non-linguistic factors that have been found to affect the choice of referent for first mention, such as visual salience and animacy, also affect the choice of referring expression. This is the topic of the next subsection.

.. Effects of accessibility on the choice of referring expression Language provides many possible ways to refer to something, ranging, for example in English, from very elaborate expressions such as full definite descriptions with modifiers (e.g., ‘the large old-fashioned red stapler with the little scratch on the top’) to very short ones such as pronouns (e.g., ‘it’). For references to objects in simple scene descriptions (so-called ‘one-shot’ references), it has been found that the perceptual context of the referent affects the choice of a referring expression (e.g., Brown-Schmidt and Tanenhaus ; Ferreira, Slevc, and Rogers ; Osgood ; Sedivy ; Sridhar ). For example, speakers use more specific referring expressions, such as ‘the empty martini glass’, when the visual context contains multiple objects of the same type, as a less specific expression, such as ‘the martini glass’, may be ambiguous. In addition, which objects in the scene are taken into account as possible competitors depends on what part of the scene people are focusing on (e.g., Beun and Cremers ; Brown-Schmidt and Tanenhaus ). Objects outside of the focused area are less likely to be considered (see also Koolen, Krahmer, and Swerts ). Speakers also often overspecify their referring expressions, adding information that reflects salient properties such as color (e.g., Arts, Maes, Noordman, and Jansen ; Engelhardt, Bailey, and Ferreira ; Koolen, Goudbeek, and Krahmer ). For example, in a referential description task described in Koolen et al. (), speakers often used color information when providing definite descriptions for objects in a scene, sometimes even when all objects in the scene had the same color.

OUP CORRECTED PROOF – FINAL, 17/1/2019, SPi

   



A different line of research has focused on the production of referring expressions in discourse. In contrast to research on the choice of referent for first mention, research on the choice of referring expressions in discourse has mainly concentrated on the influence of linguistic factors. For example, in both experimental and corpus studies of discourse, it has been found that referring expressions tend to become shorter when the same object is referred to multiple times. This includes the use of fewer words (e.g., Clark and Wilkes-Gibbs ), acoustic reduction (e.g., Aylett and Turk ; Bard et al. ), and the choice for less specific types of referring expressions, such as pronouns (Ariel ). Discourse factors that have been identified as influencing the choice of referring expression include, among others, recency (e.g., Clark and Sengul ), topicality (e.g., Givón ), first mention (e.g., Gernsbacher and Hargreaves ), grammatical function (e.g., Brennan ; Gordon, Grosz, and Gilliom ), syntactic parallelism (e.g., Arnold ), competition (e.g., Ariel ; Arnold and Griffin ), protagonisthood (e.g., Karmiloff-Smith ; Morrow ), episode shifts (e.g., Anderson, Garrod, and Sanford ; Vonk, Hustinx, and Simons ), and thematic role (e.g., Arnold ; Stevenson et al. ). For example, a referent that has recently been mentioned is likely to be referred to with a pronoun. Hence, in the second sentence of (a) a pronoun will generally be preferred to refer to Alison, who is mentioned in the directly preceding sentence.³ () a. Alisoni was working in the garden. Suddenly, {shei/#Alisoni} was hit by a swishing tree branch. b. Emily was helping Alisoni in the garden. Suddenly, {#shei/Alisoni} was hit by a swishing tree branch. Here, repeating the name could give rise to the impression that the discourse contains two people named Alison. However, in (b) a name would be preferred over a pronoun to refer to Alison, despite the fact that Alison is still the most recently mentioned entity. This is because there is a competing entity, Emily, which is mentioned in subject position and which is more topical (i.e., the sentence is more about Emily than about Alison). Even though research on one-shot references in scene descriptions has revealed the importance of the perceptual context, interactions between linguistic and non-linguistic factors have not received much attention in research on the choice of referring expressions in discourse. The selection of a certain type of expression in discourse has traditionally been seen as a highly addressee-oriented process (e.g., Clark, Schreuder, and Buttrick ). Thus, in theories of reference (e.g., Ariel ; Chafe ; Givón ; Gundel et al. ) speakers are commonly believed to choose referring expressions such that these aid the addressee in retrieving the correct antecedent. Since the linguistic context can generally be assumed to be available for both speaker and

³ Indexes mark coreferentiality. Hashes (#) mark low acceptability given the context.

OUP CORRECTED PROOF – FINAL, 17/1/2019, SPi



 ,  ,   

addressee, speakers are assumed to choose referring expressions in accordance with the information from the discourse that is part of the common ground of speaker and addressee. The different theoretical frameworks all share the idea that, in a discourse, some entities are focused on more than others (both by the speaker and the addressee), and that this has an impact on the choice of referring expression. The degree of accessibility (or topicality/givenness/focus of attention) is presented as a property of mental representations, which is influenced by—but by no means identical to—the salience of referents in the preceding discourse. Indeed, all frameworks acknowledge that referents that have not been mentioned previously can still be accessible, for instance from world knowledge (e.g., ‘the king will visit my hometown tomorrow’) or from the physical context (e.g., ‘that man over there’). For example, expressions such as unheralded pronouns (pronouns without a linguistic antecedent) and deictics (e.g., ‘that one’, often accompanied by a pointing gesture) are dependent on the configuration of objects in the physical environment of the interlocutors (e.g., Clark and Marshall ; Clark et al. ; Greene, Gerrig, McKoon, and Ratcliff ; Jarvella and Klein ; Piwek, Beun, and Cremers ). However, the discourse context has been considered the most important factor driving the activation of mental representations. For example, Ariel () claims that: [ . . . ] it is the discoursal rather than the physical salience of the entities involved which determines the degree of accessibility assigned to particular mental representations [ . . . ]. Although the physical context does affect the discourse model of the speakers, mental representations are a direct product of our discourse model only. Ariel (: )

A similar assumption is reflected in the account of Brennan and Clark (). They argue that while factors such as perceptual salience may influence the choice of a referring expression, what is most important, at least in dialogue, is whether the referent has been mentioned recently or frequently in the discourse. In research on accessibility, referring expressions have therefore been investigated mainly as anaphors, that is, expressions that have an antecedent in the preceding discourse (e.g., Brennan ; Chafe ; Clark and Marshall ; Gundel et al. ; Prince b). For example, when the linguistic context contains two possible referents of the same gender, using a pronoun (e.g., ‘he’ or ‘she’) to point out one of these referents would, at least in English, result in ambiguity for the addressee. Therefore, speakers are likely to avoid this type of ambiguity by choosing a more specific referring expression, something that has also been found experimentally (e.g., Arnold, Eisenband, Brown-Schmidt, and Trueswell ; Arnold and Griffin ; Fukumura, van Gompel, and Pickering ). The influence of non-linguistic factors, such as the visual context and intrinsic properties of referents, is likely to be more subtle, and it is only recently that these factors have been systematically investigated in studies of reference production in discourse. The next section discusses a number of relatively recent psycholinguistic studies that have investigated interactions between linguistic and non-linguistic factors in the choice of referring expression.

OUP CORRECTED PROOF – FINAL, 17/1/2019, SPi

   



. T   -       

.................................................................................................................................. While research on the choice of referring expression in discourse has focused on linguistic factors, perceptually and conceptually salient entities are likely to attract attention (e.g., Coco and Keller ; Henderson and Ferreira ; New, Cosmides, and Tooby ; Pratt, Radulescu, Guo, and Abrams ), and may therefore influence referent accessibility (Arnold and Griffin ). The few existing studies that have investigated how non-linguistic factors interact with linguistic factors in the choice of referring expression in discourse have yielded different results. The most important of these are summarized below. Arnold and Griffin () conducted a story completion experiment in which they varied the number of possible referents in the discourse. Participants were presented with a scene accompanied by a context sentence, after which they had to describe a second scene, referring to a target referent. There were three types of stimuli: () a competitor character was present both visually and linguistically; () no competitor was present; () a competitor was present both visually and linguistically in the first scene but not visually present in the second. Target referent and competitor were referentially unambiguous (i.e., of different genders). Arnold and Griffin found that participants used fewer pronouns to refer to the target character in condition () than in condition (), suggesting that the presence of a competitor decreased the target referent’s accessibility. However, no difference in pronoun use was found between conditions () and (), suggesting that it was the linguistic presence of the competitor in the context sentence that affected accessibility rather than its visual presence during sentence production. Fukumura et al. () argued that the lack of an effect of visual salience in Arnold and Griffin’s () experiment could be due to a too weak manipulation of visual salience or a too strong manipulation of linguistic salience. In addition, because speakers in this experiment were not talking to an addressee, there may have been no incentive to take the visual context into account. Fukumura et al. conducted a new experiment in which they remedied these potential problems. Participants were presented with two consecutive pictures showing either two toy characters (competitor present condition) or one toy character (competitor absent condition), as shown in Figure .. The first picture was combined with a written context sentence in which either only the target character was mentioned (competitor not mentioned condition) or both characters were mentioned (competitor mentioned condition). To reduce the dominance of the linguistic context, the target character was mentioned as the possessor in a genitive phrase and the competitor (if present) in a by-phrase (e.g., ‘The pirate’s carpet had been cleaned (by a prince)’). In the second picture, the target character

OUP CORRECTED PROOF – FINAL, 17/1/2019, SPi



 ,  ,   

 . Example of a stimulus item from Fukumura et al. () in two conditions ((a) a competitor character is visually present; (b) no competitor character is visually present). Reprinted from Fukumura, Kumiko, van Gompel, Roger P. G., and Pickering, Martin J. (). ‘The use of visual context during the production of referring expressions’, The Quarterly Journal of Experimental Psychology (): –. doi: ./. © The Experimental Psychology Society, reprinted by permission of Taylor & Francis Ltd, www.tandfonline.com.

performed a simple action. Participants described this picture to a confederate, who then acted out the description using the real toys. The results showed that participants used fewer pronouns when the competitor had been mentioned, replicating the linguistic salience effect found by Arnold and Griffin (). More importantly, participants also used fewer pronouns to refer to the target referent when the competitor was visually present than when it was not visually present. In addition, the effect of the visual context was larger in the condition where the competitor was linguistically present than in the condition where the competitor was not mentioned at all, indicating that the visual context had a greater effect when the linguistic context was less compelling (i.e., when the competitor was also linguistically present). These results suggest that speakers use both linguistic and visual information simultaneously when choosing a referring expression: the presence of a visually presented competitor referent influences pronoun use independently of the presence of a linguistically presented competitor referent. One explanation for this effect may be

OUP CORRECTED PROOF – FINAL, 17/1/2019, SPi

   



that speakers use more specific expressions because they believe that their addressee has more difficulty identifying the correct referent when multiple possible referents are present (Fukumura et al. ). Alternatively, the effect may be explained by interference from the competitor referent in the retrieval of the target referent’s representation. When multiple referents are present, representations of these have to be kept in memory. Since working memory has a restricted capacity (Baddeley ), the presence of an additional possible referent may decrease the speaker’s memory activation of the target referent, resulting in more specific expressions (cf. Arnold and Griffin ). In addition, such an interference effect might be stronger the more similar the target and competitor are. Indeed, Fukumura, van Gompel, Harley, and Pickering (), found that speakers used more explicit referring expressions (and fewer pronouns) for the target referent when a different-gender competitor character was in a similar situation in the visual context (e.g., both characters were sitting on a horse), than when the competitor was in a different situation (e.g., standing on the ground). A similar effect was found when target and competitor had the same gender (Fukumura, Hyönä, and Scholfield ), which was shown to be independent of a general avoidance of gender-ambiguous pronouns. Fukumura et al.’s () explanation for this finding was that a shared semantic feature that is activated during reference (e.g., sitting on a horse; being male) causes the accessibility of the entities that have this feature to be lower than when only one entity has this feature. This decreased accessibility in turn results in the use of more explicit referring expressions (see also Fukumura et al. ). The reduced salience of a referent due to referential competition (i.e., the presence of another possible referent) may however be a different type of salience than salience caused by a referent’s perceptual properties, such as its size (Flores d’Arcais ; Osgood ; Sridhar ). To investigate the influence of the referent’s visual salience within a scene on the choice of referring expression in a discourse, Vogels, Krahmer, and Maes (a) conducted two story completion experiments in which the relative size of the characters in a scene was manipulated by presenting the target referent either on the foreground or in the background, as shown in Figure .. The scenes always contained two characters, keeping the amount of referential competition constant. Discourse salience was manipulated by either mentioning the referent as the subject of the directly preceding context sentence or in a prepositional phrase in the penultimate context sentence. The results showed that although speakers were more likely to mention visually foregrounded characters first in their utterances, they were not more likely to pronominalize these characters. Discourse salience, on the other hand, had a strong effect on pronoun use: for discourse-salient referents, pronouns were highly preferred, which left little room for visual salience to take effect. A second experiment in which the relative linguistic salience of the two characters was reduced (by mentioning both characters as subjects in both context sentences) again showed no increased likelihood of pronominalization for visually foregrounded characters. Thus, Vogels et al. (a) found no evidence that visual foregrounding of a referent leads to more attenuated referential forms within a linguistic context.

OUP CORRECTED PROOF – FINAL, 17/1/2019, SPi



 ,  ,    A: +linguistically salient; +visually salient

‘Once upon a time there was a woman who had a quarrel with a man. The man was terribly angry.’

‘Therefore…’

C: –linguistically salient; +visually salient

B: +linguistically salient; –visually salient

‘Once upon a time there was a woman who had a quarrel with a man. The man was terribly angry.’

‘Therefore…’

D: –linguistically salient; –visually salient

 . Example of a stimulus item in four conditions in Vogels et al. (a), Experiment . Reprinted from Vogels, Jorrig, Krahmer, Emiel, and Maes, Alfons (a). ‘Who is where referred to how, and why? The influence of visual saliency on referent accessibility in spoken language production’, Language and Cognitive Processes (): –. doi: ./... Reprinted by permission of Taylor & Francis Ltd, www.tandfonline.com.

Finally, a corpus study by Duan, Elsner, and de Marneffe () did reveal some effects of several visual saliency variables on the choice of referring expression. However, these effects were mainly found on the choice between longer and shorter definite descriptions, and not on the choice for a pronoun versus a definite description. This suggests that the use of pronouns may indeed be more dependent on linguistic salience. In sum, it remains unclear how exactly the visual context influences referent accessibility in discourse. However, one consistent finding seems to be that the linguistic context often remains dominant in the choice to produce a pronoun. Another source of accessibility consists of higher-level conceptual properties of referents, such as animacy, individuation, and concreteness (e.g., Brown-Schmidt, Byron, and Tanenhaus ; Dahl and Fraurud ; Fukumura and van Gompel ; Maes ; Yamamoto ). A few studies have investigated the interaction between animacy and linguistic factors in reference production. For example, in a story

OUP CORRECTED PROOF – FINAL, 17/1/2019, SPi

   



completion experiment, Fukumura and van Gompel () found that speakers were more likely to pronominalize animate entities than inanimate entities. Thus, when producing a continuation to the sentences given in (a), participants were more likely to use a pronoun (‘they’) when they referred to ‘the hikers’ than when they referred to ‘the canoes’. The same held when grammatical roles were reversed, such as in (b), suggesting that the effect of animacy on pronominalization was independent of the grammatical function of the antecedent. Fukumura and van Gompel also found that animacy affected the choice of referent: participants were more likely to refer to the animate NP than to the inanimate NP in their continuations. () a. The hikers carried the canoes downstream. Sometimes . . . b. The canoes carried the hikers downstream. Sometimes . . . Vogels, Krahmer, and Maes (b) additionally showed that the animacy effect on the choice of referring expression is likely to be driven by the conceptualization of the referent in the (perceptual) context, rather than by the lexical semantics of the antecedent: In an animation retelling experiment, simple geometrical figures were made perceptually animate or perceptually inanimate. For example, Figure . shows a series of stills from an animation in which a green circle moves in such a way that it appears to be climbing up a slope. Entities that were made to look animate in this way were more likely to be pronominalized than perceptually inanimate entities (e.g., a circle that rolled down a slope). This effect was independent of the entity’s lexical animacy (e.g., whether it was called a ‘stone’ or a ‘hiker’). Interestingly, whether speakers mentioned the entity as the subject of their utterances seemed to be influenced more by its lexical than by its perceptual animacy. This suggests that animacy may have differential effects on the choice of referent and the choice of referring expression. At the very least, these results indicate that the speaker’s choice of referring expression in discourse is not only affected by purely linguistic factors: the referent’s saliency based on perceptual or conceptual factors also has an influence. Still, the perceptual animacy effect in Vogels et al. (b) was only present when the referent was not salient in the discourse, again suggesting that linguistic salience is a stronger factor than nonlinguistic salience. In sum, while it seems clear that the choice of referring expression in discourse is strongly affected by the referent’s salience in the linguistic context, as measured by factors such as grammatical function and topichood (e.g., Fukumura and van Gompel ; Rohde and Kehler ), there is some evidence that it can also be influenced by non-linguistic factors, although the findings are not always straightforward. Why do some studies find clear effects of non-linguistic factors on referential choices, independently of linguistic factors, while others find no effects? One reason might be the influence of the nature of the task. For example, with respect to the influence of the visual context, the way scenes are presented may have an effect on accessibility. In a series of picture description studies by Van Nice and Dietrich (b), participants

OUP CORRECTED PROOF – FINAL, 17/1/2019, SPi



 ,  ,    D

C

B

A

“steen”

 . Four stills taken from an animation used in Vogels et al. (b), Experiment . Letters indicate order; arrows indicate movement. Both were not shown during the experiment. From Vogels, Jorrig, Krahmer, Emiel, and Maes, Alfons (b). ‘When a stone tries to climb up a slope: the interplay between lexical and perceptual animacy in referential choices’, Frontiers in Psychology : . doi: ./fpsyg...

either produced descriptions with the scene in view or were only allowed to start describing after the picture had disappeared. The results showed an interaction between the animacy of the agent character and the animacy of the patient character on word order when speakers had to speak from memory (i.e., when the picture disappeared): there were more passive sentences when the patient character was animate than when it was inanimate, but only when the agent character was inanimate. Crucially, this interaction was absent when participants described the pictures in view. Van Nice and Dietrich argued that when people speak from memory instead of describing things that are present in the immediate context, they process information from multiple referents simultaneously, allowing different types of information to interact. Hence, it might be the case that the effects of more subtle, non-linguistic factors on referential choices become more manifest in cases of more compacted processing. Montag and MacDonald () investigated interactions between task context, visual salience, and animacy on the production of relative clauses in picture descriptions. Participants watched scenes containing multiple objects—both animate and

OUP CORRECTED PROOF – FINAL, 17/1/2019, SPi

   



inanimate—in various degrees of visual salience (e.g., size), and answered questions inviting them to describe a target object, such as ‘What is red?’ Montag and MacDonald found that whereas animate target objects were almost always described with a passive relative clause (e.g., ‘the man that is being thrown’), inanimate target objects were described with both active (e.g., ‘the ball that the man is throwing’) and passive (e.g., ‘the ball that is being thrown’) relative clauses. This choice was influenced by the object’s visual salience. Interestingly, inanimate objects that were more visually salient were more likely to be described with active relative clauses. According to Montag and MacDonald, this is surprising, because the standard hypothesis is that more salient referents are assigned to higher grammatical functions, such as subject. In active relative clauses, the target referent is the object of the relative clause, so these clause types should actually be less preferred for salient objects according to this hypothesis. Montag and MacDonald argued that this effect could be explained by assuming that differences in visual salience changed the nature of the task: less salient objects take longer to locate, and this longer search time increases the chance that other objects will be focused on as well. This may lead speakers to more explicitly contrast the target object from other objects, for which a passive relative clause—with the target object as the subject and the possibility to leave out the agent character from the description— may be best suited. Applied to the choice of referring expression, this argument would predict that speakers also use more specific referring expressions when search times to locate the referent are longer, since such expressions may more explicitly contrast possible referents from each other. This reasoning would be in line with Arnold and Griffin’s () proposal that more specific expressions are used when the speaker has to distribute her attention over multiple possible referents. Alternatively, speakers may also use visual information to determine whether to disambiguate referring expressions for an addressee. For example, Brown-Schmidt and Tanenhaus () found in an unscripted referential communication task speakers were more likely to use ambiguous referring expressions (e.g., ‘the green rectangle’ when there are two green rectangles) when the target referent was more salient in the visual context than its possible competitor referents. Salience was mainly determined by the proximity of the competitors to the target and whether the competitors were relevant to the task. Eye-tracking analyses revealed that these were the same factors that addressees used to interpret ambiguous referring expressions. For example, they were less likely to take into account competitors that were far away or irrelevant. This suggests that speakers use less specific expressions when they believe that their addressee is focusing on a restricted referential domain that excludes possible referents that are not relevant to the task. Indeed, the role of the addressee is likely to be an important factor in reference production. In another experiment by Brown-Schmidt and Tanenhaus (), the experimenter occasionally interrupted the referential communication task to recalibrate the addressee’s eye-tracker. During this calibration, the addressee was asked to look at different objects that sometimes started with the same sounds and were therefore temporarily ambiguous (e.g. ‘clown’/‘cloud’). As predicted, the results showed that in

OUP CORRECTED PROOF – FINAL, 17/1/2019, SPi



 ,  ,   

these calibration sessions, addressees were considering a much larger visual domain for the identification of the referent than in the referential communication task, where the domain was restricted by the previous interactions with the speaker. If speakers are sensitive to the specific role the addressee is taking up in a task, this should influence their production of referring expressions. Studies of reference production have indeed provided evidence that the addressee’s degree of engagement makes a difference in the use of referring expressions (e.g., Arts et al. ; Maes, Arts, and Noordman ; Rosa, Finch, Bergeson, and Arnold ). Especially when it is clear to the speaker what the addressee’s task is (e.g., whether he has to operate an alarm clock or assist a surgeon in a life-saving operation; Arts et al. ), speakers may be more motivated to take this information into account in their referring expressions. After all, in most everyday communicative situations speakers have the intention to convey a message with a particular goal in mind (see also Montag and MacDonald ). With respect to the choice of referring expressions in anaphoric contexts, it has been found that just the presence or absence of an addressee affects the use of pronouns (Kantola and van Gompel ). This may explain some of the divergent findings in research on the effect of non-linguistic factors on the choice of referring expression. For example, no addressee was present in the studies by Arnold and Griffin () and Vogels et al. (a), in which no effect of the visual context was found. By contrast, speakers in the study by Fukumura et al. () produced their story continuations for an addressee, who had to recreate the stories based on these continuations. Here, the visual context did have an effect on the type of referring expression. A straightforward interpretation of these findings is that speakers are more likely to take the visual context into account when there is an addressee for whom this visual information might be helpful. Such an explanation would suggest that visual salience effects on the choice of referring expression are related to audience design rather than to speakerinternal cognitive constraints (cf. Arnold and Griffin ; Fukumura et al. ). Still, there is no conclusive evidence yet as to the role of addressee-oriented processes in the effects of non-linguistic factors on reference production. For example, even when no addressee is present, speakers sometimes show linguistic behavior that would otherwise be considered a form of audience design (e.g., Koolen, Gatt, Goudbeek, and Krahmer ; Van der Wege ). This makes it difficult to determine whether certain referential choices are really adaptations towards an addressee. In addition, given that the studies discussed above manipulated visual salience in different ways (i.e., varying the number of characters in the scene versus varying the perceptual prominence of the characters in the scene), different effects of visual salience on reference production may be difficult to compare. Future research should investigate the degree to which the nature of the task and the role of the addressee influence the role of non-linguistic accessibility factors in the choice of referring expressions. To sum up, when speakers refer to an entity they appeal to a mental representation of the referent. This representation is assumed to have a certain degree of activation, which can be determined by many different factors, either linguistic (e.g., structural

OUP CORRECTED PROOF – FINAL, 17/1/2019, SPi

   



properties of a previous mention of the referent in the discourse) or non-linguistic (e.g., the referent’s visual salience or animacy). The precise influence of these factors is probably highly dependent on the nature of the task. Still, it generally holds that the higher the activation of a representation, the more accessible the referent, that is, the easier it can be retrieved from memory. Assuming that spoken language production proceeds incrementally, the element that is retrieved quickest is produced first (within grammatical constraints). Thus, more accessible referents are more likely to be chosen for first mention in an utterance. Because highly accessible referents do not need a lot of linguistic encoding to be retrieved by an addressee, speakers are more likely to use more attenuated expressions, such as pronouns, to refer to such referents.

. D          

.................................................................................................................................. So far, accessibility has been defined as the degree of activation of a referent’s mental representation, which may be influenced by many different factors. However, as has become clear, research on the choice of referent for first mention and research on the choice of referring expressions have generally focused on a different set of factors. This raises the question whether the two referential choices are actually driven by the same notion of accessibility. The choice of what to mention first has generally been considered a speaker-internal process of activating and selecting (lexical) concepts (Levelt ). As discussed above, a concept is assumed to have a high activation, for example, when it has been mentioned recently (Ferreira and Yoshita ), when it is perceptually salient (Gleitman et al. ), or when it is inherently salient (e.g., animate; Branigan et al. ). As soon as this activation reaches a certain threshold, the concept is passed down to the level of linguistic formulation, where it is assigned a prominent syntactic position (either directly, or through the assignment of grammatical functions or information structure). Thus, salient entities are preferred to be mentioned first in the sentence because they are highly accessible for the speaker. Givón () proposes that the different saliency factors—such as animacy and givenness—combine to form a hierarchy of topicality. Since people tend to talk about animate agents, for example, such entities are likely to be the topic of the sentence, and hence to occur in a prominent (e.g., sentence-initial) position. Alternatively, what these saliency factors may have in common is that they attract the speaker’s attention (e.g., Gleitman et al. ; Myachykov, Garrod, and Scheepers ; Tomlin ). Perceptual attention may be captured by, for example, large, foregrounded, animate, or moving objects

OUP CORRECTED PROOF – FINAL, 17/1/2019, SPi



 ,  ,   

(e.g., Flores d’Arcais ; Mazza, Turatto, and Umiltà ; New et al. ; Pratt et al. ). In a discourse, elements in a prominent syntactic function (e.g., subject) may be in the focus of attention (e.g., Grosz et al. ). Because what is attended to is easier to retrieve, it is more likely to be talked about first. By contrast, the choice for a certain type of expression has traditionally been seen as a highly addressee-oriented process (e.g., Ariel ; Clark et al. ; Gundel et al. ). A speaker chooses a referring expression such that her addressee is able to readily resolve what is being referred to. Therefore, the speaker has to make assumptions about what information is in common ground with her addressee. Speakers then attenuate their expressions when they believe that the referent is highly accessible for their addressee. Accessibility in this view is primarily determined by properties of the referent in the discourse, which are most likely to be available to both speaker and addressee. In sum, how speakers choose a referent for first mention has often been explained in terms of the speaker’s focus of attention, or the topic the speaker wants to talk about; conversely, the choice of referring expression has been attributed to the speaker’s assumption about the addressee’s focus of attention, or what the addressee expects the speaker to be talking about. However, if both referential choices (what to refer to first and how to refer to it) are determined by the same concept of accessibility, one would predict that those factors that make a referent accessible will increase the probability that it will be mentioned first as well as the probability that it will be referred to with an attenuated expression. This implies that what is accessible is what is most likely to be talked about next, that is, most predictable. Indeed, Givón () argues that more accessible topics are more likely to be talked about again and hence are more predictable. Because predictable entities need less explicit encoding, they can be referred to with more attenuated expressions (see also Jaeger ). This idea is also reflected in centering theory (Grosz et al. ). One of the assumptions in centering theory is that speakers seek to produce a maximally coherent discourse to minimize inferences addressees have to make about the relations between consecutive utterances. To this end, speakers try to avoid too many shifts to a different backward-looking center (the entity that is currently in the focus of attention) across utterances, that is, they tend to keep talking about the same thing. Speakers are also assumed to choose certain referring expressions to signal whether they continue to talk about the same thing. If any entity in the current utterance is pronominalized, this should at least be the backward-looking center, that is, the entity that is most likely to be talked about again. Whereas in centering theory it is primarily the subject of the preceding sentence that is most likely to be talked about again, and hence to be pronominalized, this does not always have to be the case. For example, in (), ‘he’ most likely refers to Stan, despite John being the subject of the preceding sentence, because it is likely that the second sentence is providing the reason why Stan was admired. ()

John admired Stani. Hei could devour an entire cake in twenty seconds.

OUP CORRECTED PROOF – FINAL, 17/1/2019, SPi

   



Thus, in a predictability account, Stan is the most accessible entity in the first sentence of (), because he is most likely to be mentioned next. Therefore, a subsequent reference to Stan is most likely to be in the form of a pronoun. That accessibility is related to predictability has also been proposed by Arnold (, ) in her expectancy hypothesis: accessibility correlates with the probability that a referent will be mentioned again (Arnold ). Hearers make predictions about what speakers are going to say based on the preceding discourse and on their experience with what types of entities (e.g., subjects; animates) speakers are likely to keep on talking about. Hearers will be more likely to perceive these entities as being accessible for the speaker, who in turn will be more likely to use attenuated expressions to refer to them (for speaker-internal reasons). In this way, the speaker’s choice of referring expression aids the hearer’s comprehension (whether so intended or not), since it signals whether the information was predictable or not. Evidence for this view was found by Arnold (), who used a story continuation task to elicit references to entities with either a source or a goal role, counterbalanced for grammatical function (e.g., ‘LisaSOURCE gave the leftover pie to BrendanGOAL’ or ‘MargueriteGOAL caught a cold from EduardoSOURCE two days before Christmas’). In their continuations, participants were more likely to refer to the goal entity than to the source entity. While grammatical subjects were most likely to be pronominalized, participants were also more likely to use pronouns to refer to the goal entity than to the source entity, at least for non-subject referents. Arnold () concluded that goal entities are more accessible than source entities because they are more likely to be mentioned next. Hence, speakers will (inadvertently) signal this accessibility to their addressees by using attenuated expressions. Other researchers question the view that what the speaker refers to first, and how she refers to it, are both determined by the predictability of a referential event. For example, like Arnold (), Stevenson et al. () found in a story completion study that participants were most likely to continue the story with the character that was in a goal role (e.g., ‘JohnSOURCE passed the comic to BillGOAL’), a patient role (e.g., ‘JosephAGENT hit PatrickPATIENT’), or either a stimulus or experiencer role (e.g., ‘KenSTIMULUS impressed GeoffEXPERIENCER’) depending on the connective that followed the introductory sentence (e.g., ‘because’ or ‘so’). However, the choice of whether to use a pronoun or not was only affected by the referent’s syntactic position, with participants being more likely to use a pronoun when referring to the first mentioned character than to the second mentioned character. More recently, additional evidence that the choice of referring expression is independent from the likelihood that the referent will be mentioned next was obtained by several recent studies (e.g., Chiriacescu ; Fukumura and van Gompel ; Kaiser, Li, and Holsinger ; Karimi, Fukumura, Ferreira, and Pickering ; Rohde and Kehler ; Vogels et al. a). For example, Fukumura and van Gompel () found a strong bias for continuing sentences containing either Stimulus-Experiencer verbs such as ‘scare’ or Experiencer-Stimulus verbs such as ‘fear’ with the character that was the stimulus rather than the experiencer. However, as in Stevenson et al. (), pronoun use was only affected by the antecedent’s syntactic position. These findings

OUP CORRECTED PROOF – FINAL, 17/1/2019, SPi



 ,  ,   

suggest that accessibility-related factors affect the choice of referent and the choice of referential form differently (cf. Rohde and Kehler ). This does not seem to support an account in which the accessibility of referents equals their predictability in context (e.g., Arnold , ; Givón ). The reason is that if effects of accessibility could be explained by predictability alone, it should have been found that what is most predictable, that is, what is most likely to be mentioned next, is also most likely to be referred to with more attenuated expressions. Kehler et al. () offer a Bayesian approach to pronoun production and interpretation that captures the dissociation between the choice to pronominalize and the likelihood of next mention, presented in (). ()

PðreferentjpronounÞ ¼

PðpronounjreferentÞPðreferentÞ PðpronounÞ

In this account, P(referent|pronoun) represents the probability that a pronoun refers to a certain referent. This probability is dependent on both the probability that given a certain referent, a pronoun is used to refer to it (P(pronoun|referent)) and the probability that that referent is mentioned (P(referent)). Crucially, P(pronoun|referent) is taken to be independent of P(referent), enabling both referential choices to be determined by different factors. Based on the opposition between what speakers refer to first and how they refer to it, some researchers have concluded that accessibility is only related to structural linguistic properties of the referent’s previous mention in the discourse, and not affected by predictability based on semantic factors (e.g., Fukumura and van Gompel ; Rohde and Kehler ). However, given that non-linguistic factors may affect the choice of referring expression, as discussed above, it is not clear how these effects would fit in such an account, since they would not be related to accessibility in this view. Other researchers propose instead that accessibility is a multifaceted notion, capturing different constraints that may interact and sometimes work against each other (e.g., Arnold ). For example, Kaiser and Trueswell () propose that referents can be accessible on different levels of representation. On the one hand, a referent can be accessible in a syntactico-semantic representation of the preceding sentence, which includes information about linguistic factors such as grammatical function and thematic role. On the other hand, it can be accessible in a mental model of the discourse, which involves a global representation about the event being described (Johnson-Laird ; Kintsch and Van Dijk ). Kaiser and Trueswell argue for a form-specific multiple constraints approach to reference, in which different referential choices may be either more or less sensitive to different accessibility-related factors. This means that there can be differences between specific referential forms in the degree to which they are sensitive to certain factors (e.g., Brown-Schmidt et al. ; Kaiser and Trueswell ). For example, while the choice between pronouns and more elaborate referring expressions may be affected by discourse salience, the choice between full and reduced pronouns in languages such as Dutch and Estonian may be more related to

OUP CORRECTED PROOF – FINAL, 17/1/2019, SPi

   



contrastiveness or importance (Kaiser b, b; Kaiser and Trueswell ; Vogels, Maes, and Krahmer ). Here, salient (e.g., given; animate) referents may actually be more likely to be referred to with the full forms than with the reduced forms. Such findings are not in accordance with the claim that the accessibility scales proposed by Ariel (), Givón (), and Gundel et al. () map on a hierarchy of referential forms that goes from less to more attenuated expressions. Rather, they suggest that the choice of referring expression may not be a unified phenomenon, as a linear accessibility scale makes it appear. Which type of referring expression is chosen is not just a function of its association with a particular accessibility status; multiple factors may play a role, which may each have different effects on different expressions. In the same vein, some factors may be more important for the choice of referent for first mention, while others are more important for the choice of referring expression. Indeed, even though both linguistic and non-linguistic saliency factors are assumed to underlie the accessibility of mental representations, and hence referential choices, recent research suggests that these factors may affect these referential choices differently. For example, while Vogels et al. (a) found no convincing evidence for visually salient entities being pronominalized more often, visual salience did affect which referent speakers were more likely to mention first, as the subject of their utterance. With respect to discourse salience, whether referents were explicitly introduced as the topic of the discourse influenced the choice of referent for first mention, whereas pronominalization was mainly affected by whether referents were local topics. These findings support a division of accessibility effects along the line of a global mental model versus a local syntactico-semantic representation, as proposed by Kaiser and Trueswell (). In this view, the selection of a referent for first mention may be largely driven by the speaker’s global conceptualization of the discourse, perhaps in the form of a mental model (Johnson-Laird ; Van Nice and Dietrich a). In a narrative discourse, this model may contain information about, for example, who or what the main or most important character is and about who does what to whom and why (Griffin and Bock ). It may also contain contextual information about the event being described and the communicative situation (e.g., ‘who is my addressee?’; Galati and Brennan ). Who or what is conceptualized as the main character may in turn be influenced by whether the character was linguistically introduced as such (e.g., ‘Once upon a time there was a little dragon . . . ’), by whether it is perceptually salient (e.g., visually foregrounded), or by whether it is conceptually salient (e.g., human agents are more likely to be main characters than stones or snowflakes). These factors make a referent accessible in a global representation of the discourse, although they can be in conflict, for example when the discourse topic is not a human agent. The choice of a particular referring expression, on the other hand, may depend more on a local model of the discourse, involving primarily linguistic factors (cf. Grosz ; Grosz et al. ). For example, pronouns are likely to be used when the antecedent was the subject or the topic of the directly preceding sentence. This is supported by the fact that even studies that took into account non-linguistic factors consistently found

OUP CORRECTED PROOF – FINAL, 17/1/2019, SPi



 ,  ,   

that the linguistic context had a stronger effect. However, this does not mean that there is no room for more global factors to play a role. For example, speakers may still be more likely to pronominalize those entities that they have conceptualized as animate or human based on non-linguistic information. This might be part of speakers’ mental model of the event, in which entities higher in animacy are more accessible, for example because they are more likely to be main characters and more important for the event to be described. The distinction between local and global accessibility may partly overlap with the distinction between derived and inherent accessibility proposed by Prat-Sala and Branigan (): inherent properties of referents such as animacy are likely to influence a global model of the discourse, while properties derived from context such as topicality may be more likely to affect the local discourse model. However, Vogels et al. (b) showed that this distinction is not always easy to make, because inherent properties such as animacy may also be influenced by the context. As mentioned above, they found that entities that had animate names (but were not necessarily perceived as animate) had no increased probability of pronoun use when the entity’s discourse salience was controlled for, but that they were more likely candidates for the subject role than entities with inanimate names. The opposite pattern was observed for perceptual animacy, where entities that were perceived as animate were more likely to be pronominalized than entities that were perceived as inanimate, although they were not more likely to be mentioned as the subject of the sentence. These findings suggest that the differences in accessibility effects between different referential choices cannot easily be explained by a distinction between inherent and derived accessibility. After all, it holds for animacy as well as for other factors that what determines accessibility is not the properties of entities per se, but how the entities are conceptualized (Jackendoff ). Another possible division of accessibility is to distinguish a referent’s accessibility for the speaker from the referent’s accessibility for the addressee (e.g., Arnold ; Bard et al. ; Galati and Brennan ). Indeed, research suggests a difference between the choice of referent and the choice of referring expression in the degree to which they are influenced by speaker- and addressee-oriented factors. For example, studies on the linearization of constituents have suggested that the choice of referent for first mention is determined by the degree to which the speaker attends to the referent (e.g., Gleitman et al. ; Tomlin ). However, other findings (e.g., Vogels, Krahmer, and Maes ) suggest that entities receiving less attention from the speaker are not necessarily referred to with less attenuated expressions. Rather, the choice of referring expression may depend on a model of the discourse the speaker maintains to be able to select referring expressions that a (hypothetical) addressee can interpret correctly (e.g., Hendriks, Koster, and Hoeks ). In other words, the choice of a referent for first mention may be mainly driven by the referent’s accessibility for the speaker, while the choice of referring expression is more strongly affected by the speaker’s assumptions

OUP CORRECTED PROOF – FINAL, 17/1/2019, SPi

   



about the referent’s accessibility for the addressee. This would be in line with how these referential choices are traditionally viewed in the literature. Still, such a division between speaker- and addressee-oriented accessibility would be highly confounded with a division between local and global accessibility. After all, the local discourse context constitutes highly concrete evidence for the speaker about the knowledge of the addressee: in most cases, a speaker may safely assume that her addressee has internalized a record of the immediately past discourse with this speaker that is similar to her own. The addressee’s global representation of the discourse is probably more abstract and subjective, which makes it harder for the speaker to build assumptions on it. Therefore, it might be the case that effects of non-linguistic factors such as animacy and visual salience are more related to accessibility in the speaker’s own mind. The distinction between a local and a global model of the discourse in referential choices might be visualized as in Figure . (Vogels ). As shown by the arrows, both models influence the choice of referent as well as the choice of referring expression. However, the amount of influence differs, in accordance with the idea that accessibility-related factors affect both referential choices differently. This could be represented by giving weight values to the connections. For example, if referent

Local discourse model - grammatical function - topichood

Choice of referent

Global discourse model - animacy - visual salience - protagonisthood

Choice of referring expression type reduced pronoun full pronoun demonstrative pronoun full NP

Grammatical encoding

Mental lexicon

 . Visualization of a possible unified model of effects of accessibility from local and global discourse models on the choice of referent and the choice of referring expression type. Solid lines mean ‘provides input for’; dashed lines mean ‘influences’.

OUP CORRECTED PROOF – FINAL, 17/1/2019, SPi



 ,  ,   

accessibility in the global discourse model is more influential in the choice of referent, the weight of the corresponding connection would be higher than that of the arrow departing from the local discourse model. Conversely, if local discourse factors strongly affect the choice of certain types of referring expression, the link from the local discourse model to particular referring expression types should have a higher weight than the link from the global discourse model. Different types of referring expressions—such as full and reduced pronouns—may also differ in their sensitivity to both local and global discourse salience, in line with Kaiser and Trueswell’s () form-specific multiple-constraints approach. In addition, the same type of referring expression may be sensitive to both global and local sources of accessibility to differing degrees. Again, global factors affecting the choice of referring expression, such as the referent’s animacy, may run counter to a classical accessibility account (e.g., Ariel ), such that referents that should count as more accessible are actually preferred to be referred to with the less reduced forms. The model in Figure . extends this account to the choice of referent for first mention: what is most likely to be mentioned next may be affected by both levels of accessibility, but the degree to which these levels are involved differs, and is not the same as for the choice of a particular referring expression. How this tentative model might be integrated into a general model of language production (e.g., Levelt ) is shown by the grey colored connections in Figure .. As soon as a speaker has chosen a referent for next mention, it is probably assigned a grammatical function in the grammatical encoding stage (Levelt ). In this stage, the referent will also be assigned the selected type of referring expression. The lexical item required to express the selected referent is retrieved from the mental lexicon. The ease of retrieval of lexical items (lexical accessibility) may also influence grammatical encoding, for example, whether the item is produced as the subject of the sentence. The assignment of grammatical functions in the grammatical encoding stage may in turn result in an update of the local discourse model, with more prominent grammatical functions increasing the referent’s accessibility in this model. The influence of factors affecting the global mental model will often be in line with local factors. For example, the global discourse topic is likely to also be the topic of individual utterances (e.g., Givón ). Global factors may also affect the local discourse model indirectly: a globally accessible entity is more likely to be mentioned first in a particular utterance, which in turn increases its local accessibility (via the connection between grammatical encoding and the local discourse model in Figure .). This will affect the referential form with which this entity is referred to in the next utterance. In other cases, the two types of factors may be in conflict, such as when the local topic is inanimate (see Vogels et al. b), or when the local topic is not the same as the global discourse topic or the protagonist (e.g., Poesio, Stevenson, Di Eugenio, and Hitzeman ; Van Vliet ). In sum, although local discourse factors might dominate in the choice to pronominalize, more global conceptual factors can still play a role.

OUP CORRECTED PROOF – FINAL, 17/1/2019, SPi

   



. O 

.................................................................................................................................. The model presented in the previous section is still somewhat rough and imprecise, and in its current state does not make specific predictions about how a certain referential choice will be influenced by different factors. Clearly, further research has to be conducted to investigate the exact role of various global and local factors in referential choices. In this section, some questions that still await an answer are discussed briefly. One question is what the locus of accessibility effects in the human mind is. Accessibility is believed to be a property of representations in memory (e.g., Ariel ; Bock and Warren ), but the exact relation between reference and memory often remains implicit (cf. Chafe ; Van Nice and Dietrich b). One idea is that the degree of accessibility relates to whether a referent is represented in long-term or in short-term memory (e.g., Gundel et al. ): referents are more accessible when they are in short-term memory. Alternatively, accessibility may only be a property of working memory. Van Rij, van Rijn, and Hendriks () propose a computational model of the production and comprehension of pronouns, in which accessibility is modeled as a combination of base and spreading activation of entities in working memory. The entity with the highest activation is pronominalized. However, if referents can be accessible on multiple levels (e.g., global versus local), as suggested above, it is not clear how these different levels of accessibility would connect to a single memory representation. It is therefore an open question whether it is possible for these different levels to be subsumed under a single notion of accessibility, or whether they actually represent different cognitive processes. Since the role of accessibility in reference production has been studied mostly by looking at the output (i.e., the referring expressions produced), few direct measures of accessibility are available. Future research in this field may therefore benefit from studies that tap into the cognitive processes underlying reference production themselves, for example by using eye-tracking or manipulating memory load. Second, as discussed above, it is becoming clearer that the influence of different factors on reference production is highly dependent on the specific communicative situation in which an utterance is produced, on the type of task, the speaker’s goals and the role of the addressee. Therefore, these environmental variables cannot be ignored when investigating referential choices. Given that researchers in reference production use a variety of production tasks, this poses a challenge when one wants to compare different studies. For example, many studies use written completion tasks (e.g., Anderson et al. ; Fukumura and van Gompel ; Kaiser and Trueswell ; Stevenson et al. ; Vonk et al. , while other studies focused on spoken language production (cf. Arnold ; Arnold and Griffin ; Fukumura et al. ; Kaiser et al. ; Vogels et al. a, b). Written tasks might not elicit the same kind of results as would spoken language production experiments. For instance, when people speak, they have less opportunity to reflect on what they are saying than

OUP CORRECTED PROOF – FINAL, 17/1/2019, SPi



 ,  ,   

when they write. Hence, spoken tasks might reveal more about people’s initial, automatic linguistic choices than written tasks (Arnold ). In addition, written texts generally have a less clear addressee than spoken communication, which may also influence referential choices. Third, studies differ in the degree to which they resemble naturalistic communicative situations. Many story completion tasks are not embedded within a visual context, while everyday communicative situations often are. If visual context is used in an experiment, it is often highly artificial or cartoonish. On the other hand, naturalistic scenes quickly become cluttered, which may influence reference production (e.g., Coco and Keller ; Koolen et al. ). For instance, it increases the number of factors that may affect referential choices, which may obscure effects of other factors. Moreover, people often talk about things that are not immediately present in the perceptual context. Therefore, speaking from memory rather than describing scenes might actually reflect the more frequently occurring type of language production (Christianson and Ferreira ). Completion tasks also suffer from shifts in modality or breaks in the speech flow. It would be more naturalistic to have participants tell their own stories than to first have them listen to or read (aloud) part of the story and then have them switch to speaking from their own imagination, which also involves a switch from comprehension to production. However, this latter setup may be necessary to systematically investigate effects of the linguistic context. Therefore, we would not argue for or against one or the other method. However, we would argue that the nature of the task is something that should be taken into account when discussing the outcomes of any experiment. It would be interesting to conduct experiments in highly naturalistic, unconstrained contexts for comparison, since it is conceivable that in such contexts non-linguistic cues play a more important role. These contexts may also include other types of discourse than narratives, as well as longer discourses than the ones typically used in story completion experiments (cf. Rohde and Kehler ). In addition, experimental research might be complemented with data from corpus studies or other naturally collected data to see whether the results would generalize to language use outside the lab (Brown-Schmidt and Tanenhaus ). Reference production should also be studied in other groups of participants than university students. Work that has already been done includes research on young children (e.g., Hendriks et al. ; Matthews, Lieven, Theakston, and Tomasello ), elderly people (e.g., Hendriks et al. ), Alzheimer patients (Almor, Kempler, MacDonald, Andersen, and Tyler ), and people with autism (e.g., Arnold, Bennetto, and Diehl ). Finally, most research on referring expression production has focused on the use of third person singular personal pronouns versus more specific types of referring expression (e.g., full NPs) in references to persons and objects. The range of possible referents and referring expressions is, however, much larger than that. To name a few, one could investigate references to spatial locations or points in time (e.g., ‘here’, ‘there’, ‘tomorrow’), to events and other abstract entities (e.g., ‘yesterday’s thunderstorm’, ‘my dream’), to substances (e.g., ‘the mud’) to parts of objects (e.g., ‘the sheep’s nose’) or to sounds (e.g., ‘that noise’); one could also investigate the use of different types of definite

OUP CORRECTED PROOF – FINAL, 17/1/2019, SPi

   



and indefinite descriptions (e.g., with specific or generic referents), deictics (e.g., ‘this’, ‘that’), second person pronouns (e.g., ‘you’), plural expressions or zero anaphora. For all these types of referents and referring expressions, the influence of linguistic and non-linguistic variables on accessibility may be different. In addition, accessibility effects can be different in other dimensions than the choice of referential form, such as acoustic reduction, disfluency, and gesture. Recent work in psycholinguistics has already started to explore some of these directions (e.g., Arnold and Tanenhaus ; Hoetjes, Koolen, Goudbeek, Krahmer, and Swerts ; Kaiser et al. ; Watson, Arnold, and Tanenhaus ). Insights from a broader view on referential choices, for example, one that takes into account non-linguistic factors as well as all kinds of environmental factors, may also be beneficial for computational models of reference production. Most models of referring expression generation (REG) focus on content selection for initial definite descriptions used for object identification outside of a discourse context (e.g., Dale and Reiter ). Existing models of generating referring expressions in discourse generally implement only a basic account of linguistic salience, for example based on centering theory (e.g., Kibble ; Krahmer and Theune ; Passonneau ), although algorithms for generating pronouns in texts have been developed that also take into account factors such as the antecedent’s ontological type (Strube and Wolters ) and global discourse structure (Callaway and Lester ; McCoy and Strube ; Salmon-Alt and Romary ). Recently, the GREC challenges program (Generating Referring Expressions in Context; Belz, Kow, Viethen, and Gatt ) has started to evaluate natural language generation systems that produce referring expressions in discourse, including pronominal expressions. One of the aims of these systems is to generate human-like references within a context, making use of psycholinguistic data (see also Krahmer and Van Deemter, Chapter  of this volume). For human-like generation of referring expressions, future models should also take into account influences of factors that go beyond the local discourse context, incorporating the perceptual context, speaker as well as addressee goals, and the nature of the task. In conclusion, this chapter concerned the question of how speakers choose what they mention first in an utterance and how they refer to it, as well as the role the accessibility of mental representations plays therein. Traditionally, effects of accessibility on the choice of a referent for first mention have been ascribed to the amount of attention the speaker has allocated to a referent, which may be influenced by linguistic, perceptual, or conceptual factors. By contrast, effects of accessibility on the choice of referential form have been explained as the speaker’s effort to choose expressions in accordance with the information that is in common ground between speaker and addressee, which is largely determined by linguistic properties of the directly preceding discourse. The research reviewed here suggests that the choice of referring expression may also be influenced by non-linguistic factors, interacting with linguistic factors, although the effects may be dependent on the task. At the same time, it is clear that there are indeed differences in the effects of accessibility on both referential choices. This has consequences for our understanding of the notion of accessibility, because it

OUP CORRECTED PROOF – FINAL, 17/1/2019, SPi



 ,  ,   

suggests that there are multiple types or levels of accessibility, to which different language production processes can be sensitive in different ways. These findings must be incorporated into both theoretical and computational models of reference.

A This chapter is based on parts of the first author’s PhD thesis, which was written under supervision of the second and third author. The first author is the main contributor to the text of this chapter. The authors are grateful to Josefin Lindgren for her comments on an earlier version of this chapter.

OUP CORRECTED PROOF – FINAL, 21/1/2019, SPi

        ......................................................................................................................

WHAT CAN NEUROSCIENCE TELL US ABOUT REFERENCE? ......................................................................................................................

 

. I

.................................................................................................................................. W HEN seen against the background of traditional formal semantics, the question of whether neuroscience can tell us anything about reference is easily answered. Cognitive neuroscience is concerned with how the brain processes information. Within formal semantics, answers to semantic questions have no bearing on brain processing or mental representations. Although traditional semantic frameworks, such as Montague Grammar, use a representational language, the level of representation is not supposed to reflect mental representation and is, in fact, dispensable (Montague b). So, if traditional formal semantics provides the best account of natural language, then neuroscience can only tell us about the psychology of language processing. Linguists and philosophers, however, are gradually starting to come to terms with the fact that formal semantics cannot provide an adequate semantics for natural language. Many linguistic phenomena depend on larger non-linguistic and linguistic contexts, including quantifier scope, anaphora, VP-ellipsis, presuppositions, and resolution of lexical ambiguity. Such discourse-dependent phenomena cannot be adequately accounted for by traditional formal semantics but require a dynamic semantic framework that can model discourse updates. Two of the most influential dynamic semantic theories and also the first on the linguistic scene are Discourse Representation Theory (DRT), developed by Hans Kamp () and file change semantics, developed independently by Irene Heim (). Here I shall focus primarily on DRT but most of what I say will apply more broadly. Traditional formal semantics took the primary focus of semantic theory to be reference, truth, and satisfaction. The meaning of a sentence is its truth-conditions.

OUP CORRECTED PROOF – FINAL, 21/1/2019, SPi



 

Within DRT and related dynamic approaches, the most central concepts are not those of reference, truth, and satisfaction, but that of information, or representation. The meaning of a sentence consists of both its potential to change a given discourse representation structure (DRS) into a new one and its truth-conditional import in the representation that results from a potential update (Kamp et al. ). In DRT a sentence thus has meaning only in a derivative sense. As an illustration of the difference, consider Chomsky’s grammatical but nonsensical “Colorless green ideas sleep furiously”. Whereas this sentence would be associated with a truth-condition in formal semantics, viz. that green ideas sleep furiously, the sentence is unlikely to have any potential to change a given discourse representation and hence will have no truthconditional import in the representation that results. It is therefore meaningless in the context in which it occurs. While reference and truth are not the most central notions in DRT, they are nonetheless important notions. Proper names, indefinite and definite descriptions, certain other quantifiers, demonstratives, and pronouns can all function as referential expressions that contribute a discourse referent and a condition to the discourse representation. Discourse referents are a special kind of variable that are bound by discoursewide existential quantifiers. So, referential expressions refer to an individual when there is an individual that can serve as a value of the variable introduced by the expression. Reference requires satisfying not only the conditions introduced by the referring expression but also those introduced by co-referring expressions. For example, ‘he’ as it occurs in ‘A delegate arrived. He was dressed in black’ contributes a discourse referent x and the condition that the referent must be male. But assuming that ‘he’ and ‘a delegate’ both occur in the updated discourse representation and that ‘he’ refers back to ‘a delegate’, ‘a delegate’, and ‘he’ refer to a particular individual S only if there is a mapping from ‘male(x) and delegate(x)’ to S. Since discourse representations are continually updated, an expression that doesn’t refer may come to refer after a discourse update. For example, if the hearer learns that the delegate who arrived is a woman, then ‘a delegate’ doesn’t refer to anything, as there is no mapping from ‘male(x) and delegate(x)’ to an individual. But the new information may result in the revised discourse representation A delegate arrived. She was dressed in black, in which case ‘a delegate’ acquires an actual referent. Since traditional formal semantics is not concerned with how the brain processes information, neuroscience cannot provide evidence for or against traditional formal semantics. Evidence from neuroscience only becomes semantically significant once it has been determined on independent grounds that dynamic semantics is required to provide an adequate account of linguistic phenomena in natural language. Neuroscience can then shed light on how the human brain actually interprets natural language. In what follows I will first review some of the arguments for thinking that DTR or a related dynamic framework is required to account for various linguistic phenomena in natural language. I will then review whether the neuroscientific findings lend support to reference-related aspects of the theory. Finally, I will consider some methodological concerns that have been raised about the existing neuroscientific approaches to natural language interpretation.

OUP CORRECTED PROOF – FINAL, 21/1/2019, SPi

       ?



. DRT   

.................................................................................................................................. Kamp’s Discourse Representation Theory (DRT) and Irene Heim’s () related file change semantics have been enormously successful as models of linguistic phenomena that depend on larger linguistic contexts, including quantifier scope, anaphora, VPellipsis, presuppositions, and resolution of lexical ambiguity. Anaphora has proven particularly difficult to account for within traditional formal semantics. For example, traditional semantics seems unable to provide an account of anaphoric discourse of what has become known as ‘donkey anaphora’: ()

If Pedro owns a donkey, he beats it

One might attempt the following analysis of (): ()

∃x: (donkey(x) & owns(pedro, x)) ! beats(pedro, x)

But this is not a well-formed formula, as the variable in the consequent ‘beats(pedro, x)’ is outside the scope of the existential quantifier. The only option is an analysis, using a universal quantifier: ()

∀x: (donkey(x) & owns(pedro, x)) ! beats(pedro, x)

But the structure of this analysis is different from that of () and interprets the existential quantifier as a universal, which is not the natural interpretation of (). In DRT, an interpretation is a mental representation that consists of discourse referents and conditions that put constraints on the values of the discourse referents. Unlike the indefinite description ‘someone named “Pedro”’, the name ‘Pedro’ carries a presupposition to the effect that Pedro(x) is already part of the existing DRS. If Pedro has not already been introduced, the presupposition is accommodated by adding the discourse referent and condition introduced by the name. In Kamp’s () model, this is represented by placing discourse referents for proper names outside of the scope of the conditional to reflect that the discourse referent is supposed to be part of the already processed parts of the text. So, () can be represented as: ()

x: Pedro x & [donkey y & owns (x, y) ! beats(x, y)]

Because the discourse referent for ‘donkey’ occurs in the antecedent of a conditional, any donkey can serve as a value, so the truth-condition comes out as ‘For every donkey Pedro owns, Pedro beats it’. Donkey anaphora is not a decisive factor in choosing a dynamic semantic theory over traditional semantics. Although traditional semantics cannot account for donkey anaphora using standard first-order logic, alternative accounts have been developed that can account for the basic cases. An alternative analysis is Stephen Neale’s (: ) D-type analysis. On Neale’s view, donkey pronouns are numberless

OUP CORRECTED PROOF – FINAL, 21/1/2019, SPi



 

definite descriptions constructed from the content of the antecedent clause. On this view, () comes out as: ()

If Pedro owns a donkey, he beats the donkey(s) he owns.

While this analysis does not require dynamic resources, it is unclear that it can be extended to all cases of donkey anaphora. Adverbs of quantification introduce new levels of difficulties (Brogaard ). For example, ‘if Pedro owns a donkey, he usually beats it’ can be read as saying that Pedro beats most of the donkeys he owns. However, the sentence ‘if Pedro owns a donkey, he usually beats the donkey(s) he owns’ can only be given the temporal interpretation that if Pedro owns any donkey, then he beats all of the donkeys he owns most of the time. DRT also has an advantage in terms of accounting for the difference in the meaning of sentences that are rendered truth-conditionally equivalent by standard semantics. For example, even though ‘one of the ten marbles is not in the bag’ and ‘nine of the ten marbles are in the bag’ are truth-conditionally equivalent, according to standard semantics, only the former licenses anaphora: ()

a. One of the ten marbles is not in the bag. It is under the sofa. b. *Nine of the ten marbles are in the bag. It is under the sofa.

Because DRT predicts that only ‘one of the ten marbles is not in the bag’ introduces a discourse reference, it can account for why the two sentences differ in meaning despite being truth-conditionally equivalent. A further problem for traditional semantics is that it cannot straightforwardly account for the fact that the interpretation of one sentence can depend on what is uttered at a later time. Consider the following example, taken from Asher et al. (): ()

A nurse saw every patient. Dr. Smith did too.

Although there is a scope-ambiguity in the first sentence, the preferred reading after the second sentence is processed is one where ‘a nurse’ takes wide scope, viz. ∃x(∀y (patient y ! x saw y)). Asher et al. () argue that the more specific common themes between sentences is preferred over the wider one. The common theme between the second sentence and the wide-scope reading of the first sentence is more specific than the common theme that would result from a reading where ‘every patient’ takes wide scope. Traditional semantics stipulates that the first sentence in () should have the meaning it does independently of discourse that is added later.¹ DRT, on the other hand, predicts that discourse that is uttered later can affect the meaning of the existing DRS. DRT is thus able to provide a more comprehensive account of a multiplicity of linguistic phenomena than standard semantics. Moreover, because DRT equates meaning with mental representation, it is suitable as an object of investigation from a neuroscientific perspective. The aspects of reference that have been most extensively ¹ This is not to say that traditional semantics does not recognize that natural language often is highly ambiguous.

OUP CORRECTED PROOF – FINAL, 21/1/2019, SPi

       ?



investigated by neuroscience are anaphoric reference and presupposition accommodation. While philosophers of language have traditionally drawn a sharp distinction between anaphoric reference and reference to extra-mental entities, the two are not clearly separable from a language-processing perspective. Empirical studies of reference to extra-mental entities might examine which entity in a visually presented narrative a speaker takes an expression to refer to. Analogously, studies of anaphora might examine which entities in a verbally presented narrative a speaker takes an expression to refer to. While there might be crucial differences between the two approaches, it seems clear that studies of anaphora and related phenomena should be able to shed light on reference to extra-mental entities. After reviewing some general neurophysiological studies of discourse representation I will review the evidence pertaining to anaphoric reference and presupposition accommodation and then turn to the reference of proper names, a topic that has been of particular interest to philosophers of language.

. E    

.................................................................................................................................. Electrophysiological measurements of event-related brain potentials (ERPs) have been used to investigate how the brain processes discourse. ERPs are average brain responses to sensory, cognitive, or motor events measured with electroencephalography (EEG), which record continuous electrical activity in the brain via electrodes placed on the scalp. ERP signals take the form of a temporal sequence of negative and positive voltage deflections compared to a pre-stimulus baseline. These deflections vary in polarity (negativity or positivity), amplitude, latency, duration, and distribution over the scalp, and these variations can be used to distinguish different cognitive processes on the basis of their ERP signature. Because EEG can provide fine-grained temporal information, EEG (and the newer, but related, MEG) provides a more accurate picture of how the brain processes discourse in real time compared to most other forms of neuroimaging and behavioral studies of language. For example, studies that compare response times to different stimuli are end-state measures and hence are unable to track language processing in real time (Kutas and Federmeier ). The most extensively studied linguistic phenomena are syntactically (e.g., ‘the man prepared herself for the operation’) and semantically (e.g., ‘he spread the warm bread with socks’) anomalous sentences and discourses. Syntactic violations (e.g., ‘the man prepared herself for the operation’), syntactic ambiguities (e.g., ‘a nurse visited every patient’), syntactic complexity (e.g., ‘the boat sailed down the river sank’) and new lexical information elicit a P and sometimes an early left anterior negative (ELAN) effect or a left anterior negative (LAN) effect (Osterhout and Mobley ; Callahan ). A P effect is a positive-going deflection in frontal or parietal brain regions,

OUP CORRECTED PROOF – FINAL, 21/1/2019, SPi



 

starting around  milliseconds after stimulus onset, peaking at  milliseconds, and lasting for approximately another – milliseconds. The ELAN effect is a negativegoing wave that peaks around  milliseconds or less after stimulus onset and is most frequently elicited by stimuli that violate phrase structure (e.g., ‘he is the in office’), whereas the LAN effect is a negative-going wave that peaks around – ms post word onset and is most frequently observed in cases of word-category violations (e.g. ‘she likes to food’). The first ERP study that successfully manipulated semantic variables was conducted by Kutas and Hillyard (). They found a correlation between large-amplitude negative ERP components with the semantic plausibility of a word given the preceding sentence context. Consider, for example: ()

a. He spread the warm bread with butter. b. He spread the warm bread with socks.

Relative to baseline (a), sentences such as (b) trigger a large-amplitude negative ERP component in the centroparietal region starting around  ms after word onset, peaking at  ms, and lasting for approximately another  ms. This is also known as the N effect. Even sentences with congruent sentence final words elicit some degree of N activity but the amplitude is significantly larger with incongruent sentence final words than with congruent sentence final words. This relative negative shift elicited by inappropriate words compared to appropriate ones was not observed with other unexpected occurrences, such as variation of the physical attributes of the stimulus (e.g., larger font or different font type). For example, ‘I shaved off my mustache and BEARD’, with unexpected capital letters at the end, did not elicit an N effect but only the more general P that is observed in response to unexpected stimuli more generally. The N effect has subsequently been found in a wide range of variations on the original experimental setup. It was observed for spoken words, American Sign Language, and non-linguistic meaningful stimuli, such as pictures and drawings, when primed by linguistic contexts. But it was not elicited by other structured stimuli, such as music (Kutas and Federmeier ). Kutas and Hillyard () proposed on the basis of their original results that the N may be an electrophysiological marker of the interpretation of semantically anomalous information. However, it was later found that semantic anomaly is not required for the N effect to be elicited. For example, it was found when one word is unexpected relative to another plausible word (Kutas and Hillyard ). Although (a) and (b) are both plausible, (b) elicits a larger N amplitude than (a): ()

a. John turned on the faucet and poured water in his glass. b. John turned on the faucet and poured beer in his glass.

Rather than reflecting semantic anomaly, the N amplitude is now commonly taken to reflect the level of difficulty of integrating new semantic information into an existing semantic representation (Hagoort et al. ) or the level of difficulty of identifying a discourse referent for a definite expression (Burkhardt ). When understood in this way, the N effect can also be treated as an indicator of the extent to which a DRS

OUP CORRECTED PROOF – FINAL, 21/1/2019, SPi

       ?



needs to be revised. When a hearer receives the information ‘John turned on the faucet and poured . . . ’, a model is generated in which John is pouring water in his glass. The unexpected noun requires a revision of the model, which yields a larger N amplitude compared to baseline. In (b) ‘beer’ is unexpected relative to background information rather than previous linguistic context. Various other empirical studies have found that a word or sentence that is unexpected given world knowledge yields the same effect as a word or sentence that is unexpected given the linguistic context. For example, Hagoort et al. () found that ‘The Dutch trains are white and very crowded’ and ‘The Dutch trains are sour and very crowded’ elicited the same N effect in Dutch speakers who know that trains are yellow. The finding that integration of new information into both existing lexico-semantic knowledge and world knowledge occurs within  ms, suggests that the brain does not integrate lexico-semantic knowledge prior to world knowledge. Hagoort et al. () argue that this suggests that the existing DRS consists of both lexico-semantic information, expected information, and background information. New semantic information elicits an N response when it clashes with the existing information and an update is required. The update can consist in a rational reconstruction of the DRS or a rejection of new information if the information is considered incoherent in the context in which it occurs (Kamp et al. ). In line with this suggestion, Baggio et al. () found that discourse models represent the outcome of inferences anticipating the goal state of actions and events prior to receiving information about whether they take place. They compared sentences of the following kind: () a. The girl was writing a letter when her friend spilled coffee on the table cloth. b. The girl was writing a letter when her friend spilled coffee on the paper. Spilling coffee on the tablecloth is likely neutral with respect to the writing activity. So, from (a) the reader would be expected to conclude that the girl will accomplish her goal. The reader should thus be expected to assent to ‘The girl has written a letter’. However, the inference to the goal state is defeasible or non-monotonic. As spilling coffee on the paper is likely to lead to a termination of the writing activity, the inference from that to the accomplishment of writing the letter will likely be suppressed in (b). The reader should thus be expected to assent to ‘The girl has not written a letter’. Because ‘tablecloth’ is less semantically expected in this context, the authors predicted that (a) would evoke a greater N response. This prediction was borne out. Nouns elicited a larger N effect in neutral than in disabling clauses. They further found that the amplitude of the ERP effect evoked by disabling clauses is correlated with the frequency with which readers inferred that the goal state was not attained. The results suggest that the default expectation is that the actual world is an inertial world, that is, a world that is identical to the actual world up until the present moment and that continues in the way that is most compatible with the history of the world up until the present moment (Dowty ). The disabling clause thwarts that expectation. Inference-dependent N attenuations have also been observed with multisentence text (Van Berkum ). Consider:

OUP CORRECTED PROOF – FINAL, 21/1/2019, SPi

 ()

  a. Mark and John were having an argument. Mark began to hit John hard. b. Mark and John were having an argument. Mark got more and more upset. c. Mark and John were gambling at the casino. They won every game of blackjack. d. The next morning John had many bruises.

A smaller N effect was observed when (d) follows the explicitly supportive text in (a) compared to the unsupportive text in (c). However, the implicitly supportive text in (b) also elicited a smaller N effect than the unsupportive text in (c), indicating that the hearer fills in anticipatory information. The ERP studies thus appear to confirm the tenet underlying DRT to the effect that language interpretation proceeds by incremental updating of an ever-growing discourse model.

. A 

.................................................................................................................................. The majority of neuroscientific research related to reference has focused on the question of how speakers determine anaphoric reference. In cases of movement, the speaker encounters a phrase in a non-canonical position and initiates a search in upcoming context for a corresponding anaphor, that is, any pronoun or pro-form with little descriptive content (e.g., ‘there is a girl in our class who is so endearing that every boy adores her’). In co-referential anaphora, the speaker encounters an anaphor and initiates a search in previous context for a linguistic antecedent (Callahan ). Several constraining conditions of the anaphor guide the search for an antecedent, including gender, number, and descriptive information. The difficulty of selecting an antecedent and integrating an anaphor into the existing DRS is determined by these features as well as the distance between the anaphor and the antecedent, the syntactic roles of anaphor and antecedent, the salience of the antecedent, and the number of competitors. A LAN effect or an N effect is elicited when it is more difficult to select and retrieve an antecedent from working or long-term memory, whereas a P is elicited when it is easier to select an antecedent. For example, a pronoun referring to a less frequent antecedent elicits a larger P effect, reflecting that the low frequency of the antecedent makes it easier to retrieve from working memory (Heine et al. ). Difficulties integrating an anaphor owing to increased distance between the anaphor and the antecedent are associated with a larger N amplitude. For example, Streb et al. () found that a more distant antecedent (b) elicited a larger N amplitude compared to the closer antecedent (a), reflecting the difficulty of integrating the anaphor into the previous discourse context. () a. Beate besitzt eine kleine Tierpension. Überall im Haus sind Tiere. Tomi ist ein alter Kater. Heute hat Tom/eri der Frau die Tür zerkratzt. ‘Beate has a small boarding-home for animals. Everywhere in the house are animals. Tom is an old cat. Today Tom/it scratched the door of the woman.’

OUP CORRECTED PROOF – FINAL, 21/1/2019, SPi

       ?



b. Lisai schlender über einen Basar. Peter verkauft Edelsteine an Touristen Die Steine sind hervorragend geschliffen. Nun wird Lisa/siei dem Händler einen Diamenten abkaufen. ‘Lisa strolls across a bazaar. Peter sells gems to tourists. The gems are cut excellently. Then Lisa/she will buy a diamond from the trader.’ A similar N effect has also been found when the antecedent plays a difficult syntactic role compared to the anaphor. For example, the dative ‘ihm’ in (a) yields a more prominent N effect than the nominative ‘er’, indicating greater difficulty of integrating an anaphor that plays a different syntactic role compared to the antecedent. () a. Peteri besucht Julia in der Klinik. Dort hat Peter/eri dem Arzt eine Frage gestellt. ‘Peter visits Julia in the hospital. There Peter asked the doctor a question.’ b. Peteri besucht Julia in der Klinik. Dort hat die Schwester Peter/ihmi das Zimmer gezeigt. ‘Peter visits Julia in the hospital. There the nurse showed Peter/him the room.’ The difficulty of integrating an anaphor into an existing DRS has also been found to vary with whether the anaphor is a surface or a deep anaphor. The distinction between surface and deep anaphora was introduced by Hankamer and Sag (). Surface anaphors require a linguistic antecedent, whereas deep anaphors do not. Consider: () a. Scenario: A man sees a woman about to jump off a bridge. Man: ‘Don’t do it!’ (‘do it’ - deep anaphor) *Man: ‘Don’t do so!’ (‘do so’ - surface anaphor) b. Man: ‘A woman was about to jump off a bridge . . . . . . and I told her not to do it.’ . . . and I told her not to do so.’ Unlike ‘do it’, ‘do so’ is acceptable only with a linguistic antecedent, as shown in (a)–(b). Because it requires a linguistically represented antecedent, it is a surface anaphor. Sag and Hankamer () proposed that unlike deep anaphors, surface anaphors are sensitive to syntactic parallelism, between the antecedent and the anaphor. They offered the following example: () The children asked to be squirted with the hose, so a. they were []. (VPE, surface) b. *we did []. (VPE, surface) c. we did it. [sentential it, deep] Because the antecedent phrase is passive, the surface anaphor is acceptable only when it is also passive (a). Hankamer and Sag furthermore predicted that surface anaphora, unlike deep anaphors, are sensitive to intervening discourse. Consider:

OUP CORRECTED PROOF – FINAL, 21/1/2019, SPi

 ()

  a. John raked the leaves in the back yard. (i) Later, Bill did too. (surface) (ii) Later, Bill did it too. (deep) b. John raked the leaves in the back yard. This was much more fun than studying for exams. (i) ?Later, Bill did too. (surface) (ii) Later, Bill did it too. (deep)

In (b), in which there is intervening discourse between the anaphor and the antecedent, the surface anaphor is unacceptable. Hankamer and Sag () speculated that these and other differences between surface and deep anaphors are due to a difference in how surface and deep anaphors are processed. In the case of deep anaphors hearers access the antecedents using a nonlinguistic discourse-level interpretation of the antecedent. In the case of surface anaphors, hearers access the antecedents at a linguistic level determined by surface syntactic structure. They do this by assigning the antecedent’s VP structure to the site of the null anaphoric VP. For example, ‘John told her not to do so’ is interpreted at a linguistic level as ‘John told her not to jump off the bridge’. Subsequent empirical studies, however, do not support this mechanism for surface anaphora. Woodbury () carried out two fMRI studies combined with naturalness ratings and reading time measurements to test Sag and Hankamer’s claims that distance affects surface anaphora but not deep anaphora and that syntactic parallelism, particularly surface word order, is required for surface anaphora but not for deep anaphora. She found that increased distance between the antecedent and the anaphor affected surface anaphora more than deep anaphora. Increased distance resulted in increased activity in several language-related areas of the brain for surface anaphors but had no added effect on those areas for deep anaphora. One explanation of this may be that the representation of surface syntax decays over intervening discourse while semantic or pragmatic information does not, which explains why only surface anaphors are unacceptable in the case of intervening discourse. The second study of semantic parallelism (i.e., word order) did not confirm the prediction that surface word order has a greater adverse effect on surface anaphora compared to deep anaphora. The most likely explanation for this is that neither surface anaphora nor deep anaphora depend on exact word order for correct processing. As Woodbury () points out, this is consistent with a model of surface anaphora, according to which the syntactic form of the antecedent is not copied into the location of the anaphor, but is processed via the antecedent’s deep-level syntactic representation in semantic memory. This would explain why surface anaphora is sensitive to certain factors that affect syntactic information but not semantic or pragmatic information, yet is not sensitive to surface syntax.²

² It should be noted that since the copying analysis—like syntactic analyses in general—was not intended to a have a direct analogue in processing, the above finding is not a direct contradiction of Hankamer and Sag’s proposal analysis. Their primary contribution was in noting the distinction between surface and deep anaphora.

OUP CORRECTED PROOF – FINAL, 21/1/2019, SPi

       ?



Together the neuroimaging data provide some indication that the treatment of anaphora provided by traditional formal semantics needs to be modified or augmented. Anaphoric reference cannot always be resolved in the same sentence. Moreover, syntactic models of anaphora that predict that the antecedent is copied into the location of the anaphor does not seem to be empirically supported by neuroscience. The empirical data suggests that anaphora is resolved by eliciting a search for an antecedent in the previous discourse context and that integration is affected by factors, such as the prominence and frequency of the antecedent and the distance between the anaphor and the antecedent.

. D    

.................................................................................................................................. Definite descriptions can be considered a special case of anaphora in that they often seem to require an existing discourse referent to refer back to. Consider: () a. John talked to a bouncer b. John talked to the bouncer Unlike (a), which introduces a new discourse referent, (b) seems to require that the existing DRS provides a discourse referent for ‘the bouncer’ to refer back to (Heim ).³ The definite description thus triggers the presupposition that an antecedent for it can be found in the previous discourse context (existence/familiarity). Furthermore, it triggers a uniqueness presupposition to the effect that there is only one salient bouncer in the previous discourse context (uniqueness) (Roberts ; Abbott a). By triggering these presuppositions the definite description acts like a pointer to information provided earlier. When the existing DRS does not provide a discourse referent for a definite description to refer back to, the standard view among discourse representation theorists is that a discourse referent is added via an inferential process, known as ‘presupposition accommodation’. This process updates the DRS by adding the required discourse referent (Lewis a; Heim ). Presupposition accommodation makes definite descriptions different from anaphoric pronouns. While the latter also trigger presuppositions, these presuppositions are not as easily accommodated. Compare the following discourse fragments, taken from Kamp et al. (): () a. Bill is a donkey owner. The donkey is not happy. b. Bill is a donkey owner. ?It is not happy. The first sentence implies that Bill owns one or more donkeys. In (a) ‘the donkey’ triggers a uniqueness presupposition to the effect that Bill owns only one donkey. Since ³ The referent could, of course, also be provided by non-linguistic context. But I shall set that aside here.

OUP CORRECTED PROOF – FINAL, 21/1/2019, SPi



 

there is no single discourse referent for ‘the donkey’ to refer back to, this presupposition is accommodated by adding a discourse referent to the discourse context. In (b) the anaphoric pronoun ‘it’ also triggers a uniqueness and existence presupposition. However, in spite of the fact that ‘it’ introduces the condition ‘non-person(x)’, which rules out that the pronoun refers back to Bill, the presupposition triggered by it is not accommodated. So, (b) is infelicitous. The presuppositions triggered by definite descriptions can be accommodated when there is a plausible inference from the previous discourse context to the presupposition. If, however, the presupposition is not a plausible inference from the DRS, then it cannot easily be accommodated by the DRS. For example, if the context of (b) is that John went to the local elementary school, then (b) is much harder to interpret, as in (b): ()

a. John went to a local club. He talked to the bouncer. b. John went to the local elementary school. ?He talked to the bouncer. c. John went to the local elementary school. He talked to one of the other parents, a bouncer from the local club. The bouncer said that he was there to complain about his kid’s new teacher.

There is evidence from neuroscience supporting the theory of presupposition accommodation. A larger N amplitude has been found for definite descriptions that require accommodation (a) compared to definite descriptions for which there is a salient discourse referent to refer back to (c).⁴ The N effect is most prominent when there is no plausible inference from the immediate discourse context that would allow construction of a new discourse referent for the definite description (b) (Burkhardt ). A P effect has been observed both when an indefinite description is used to introduce a new discourse referent (c) and when presupposition accommodation is required to process a definite description (a) but not when a discourse referent was already available (Burkhardt ; Schumacher ). Van Berkum et al. () examined the uniqueness presupposition of definite descriptions. Participants were presented with sentences that contained a singular definite description embedded in stories that varied in the number of suitable referents they introduced for a singular definite, as in: () a. De aardige reus werd onderweg vergezeld door een elfje (een fee) en een kabouter. Het elfje (de fee) had zich vastgeklampt aan zijn bovenarm, de kabouter had zich genesteld in een comfortabele broekzak. De reus waarschuwde het elfje dat ze niet moest vallen.

⁴ This shows that discourse referents that have been linguistically introduced (and are still in shortterm memory) require less effort to associate with a definite form than those that require an extra inferential process which makes an indirect association. But it would follow equally well from a theory that simply requires association with a unique referent. Constructing a new unique reference by inference would also require more effort than identification with a recently introduced referent that is still in short-term memory; so while these facts do support the theory of presupposition accommodation, they don’t support it over alternative theories of definiteness that require unique identifiability, not previous familiarity.

OUP CORRECTED PROOF – FINAL, 21/1/2019, SPi

       ?



‘On the road, the gentle giant was accompanied by an elf (fairy) and a goblin. The elf (fairy) had clung [itself] to his upper arm, the goblin had ensconced itself in a comfortable trouser-pocket. The giant warned the elf that she shouldn’t fall off.’ b. De aardige reus werd onderweg vergezeld door twee elfjes (feeën). Het ene elfje (de ene fee) had zich vastgeklampt aan zijn bovenarm, het andere (de andere) had zich genesteld in een comfortabele broekzak. De reus waarschuwde het elfje dat ze niet moest vallen. ‘On the road, the gentle giant was accompanied by two elves (fairies). One of the elves (fairies) had clung [itself] to his upper arm, the other had ensconced itself in a comfortable trouser-pocket. The giant warned the elf that she shouldn’t fall off.’ The results showed that the ambiguous two-referent condition triggered a sustained negative deflection, broadly distributed over the scalp but with a pronounced frontal component, and starting at about  ms post word onset but without a well-defined peak. This effect has come to be known as the ‘Nref effect’ and has been taken to indicate difficulties identifying a unique discourse reference for definite descriptions and pronouns to refer back to (Nieuwland et al. ; Van Berkum et al. ). The original data did not establish whether the Nref effect was related to the lack of salience of a unique discourse referent or would emerge as long as two discourse referents had been introduced in the previous context. However, Nieuwland et al. () subsequently looked at whether the Nref effect would disappear if the ambiguity was eliminated prior to the occurrence of the target definite description. Participants were presented with stories in which one of the potential discourse referents of a definite description was made more salient by having one individual leave the scene of the protagonist, for example: () David had asked the two girls to clean up their room before lunchtime. But one of the girls had been sunbathing in the front yard all morning, and the other had actually just driven off in his car for some serious downtown shopping. As he gazed at the empty driveway, David told the girl . . . The Nref effect was found to disappear when one discourse referent was made salient. In (), for example, there is only one salient discourse referent available for the definite description ‘the girl’ to refer back to. This shows that the Nref effect is not triggered simply by the initial availability of two potential discourse referents but is triggered when there is not a unique salient discourse referent for it to refer back to. The data from neuroscience thus give us some reason to think that definite descriptions trigger both a familiarity and a uniqueness presupposition. Speakers accommodate plausible familiarity presuppositions that are not already satisfied by the previous discourse. When a presupposition is implausible given the existing discourse representation, it is more difficult for the reader to accommodate it. At this point the speaker

OUP CORRECTED PROOF – FINAL, 21/1/2019, SPi



 

may accommodate the presupposition or choose not to integrate the new information into the existing discourse representation. Difficulties integrating a definite description can also arise because of lack of saliency of the discourse referent, which provides some evidence for the hypothesis that definite descriptions trigger both a familiarity and a uniqueness presupposition.

. T    

.................................................................................................................................. In philosophy the majority of debates about reference have centered around the reference of proper names. Until the s the dominant theory was the descriptive theory, according to which proper names refer indirectly via descriptions or concepts. Saul Kripke’s () lectures “Naming and Necessity” won many philosophers over on the side of direct referentialism. On the ‘new’ standard view, proper names refer directly to an individual by virtue of a causal-historical connection, to their uses, and an original baptismal event. One of the most influential arguments for direct reference theories of proper names is the modal argument. Consider the following two sentences: ()

a. Aristotle is Aristotle b. Aristotle is the teacher of Alexander the Great

(a) and (b) have different modal profiles. As Aristotle could not have failed to be Aristotle, (a) is metaphysically necessary. By contrast, since Aristotle might not have been the teacher of Alexander the Great, (b) is metaphysically contingent. As (a) and (b) have different modal profiles, they cannot be semantically equivalent. So, ‘Aristotle’ and ‘the teacher of Alexander the Great’ cannot be semantically equivalent. But the same argument can be made, regardless of which descriptions we propose are semantically equivalent to ‘Aristotle’. So, ‘Aristotle’ is not semantically equivalent to any description. Or so the argument goes. There have been many subsequent replies to this argument on behalf of descriptive theories of reference. For example, identifying proper names with rigidified descriptions (e.g., ‘the actual teacher of Alexander the Great’) or clusters of descriptions seems to avoid this particular problem (Reimer ; Abbott ). But it is fair to say that none of the replies to Kripke’s modal argument has gained widespread support. There is, indeed, still a considerable number of philosophers who adhere to a direct reference theory of proper names (http://philpapers.org/surveys/metaresults.pl). In DRT, proper names are treated as predicates (Kamp and Reyle ).⁵ The first occurrence of a proper name either introduces a discourse referent x and a condition N(x) or refers back to a previously introduced discourse referent, as in ‘They elected a ⁵ In Kamp (), proper names are treated as individual constants and introduce the condition x is identical to N.

OUP CORRECTED PROOF – FINAL, 21/1/2019, SPi

       ?



new chair. His name was “Kelvin”. Kelvin chaired the committee for two years’. Additional occurrences of the name refer back to the discourse referent, which refers to an actual individual just in case that individual satisfies all the conditions introduced by co-referring expressions. Despite the semantic differences between the direct reference theories and DRT, the condition introduced by names, N(x), might turn out to map onto an individual only if there is a causal-historical relation between the occurrence of ‘N’ and that individual (this is indeed an implication of Kamp’s notion of anchoring). However, DRT inflicts the further requirement that the individual must satisfy any other constraints imposed by co-referring expressions. For example, a discourse referent introduced by a proper name (e.g., ‘Obama’) and subsequently referred to by a definite description (e.g., ‘the president’) can have a particular value only if the constraining conditions of definite description (e.g., being the president) are satisfied. This requirement imposed by DRT has the virtue that even if a name turns out to have no actual referent, it still has a meaning contributed by the name-qua-predicate and other constraining conditions. In traditional formal semantics, proper names are individual constants. So, if they have no actual referent, they are ‘empty’ and therefore meaningless. Whether a causal-historical relation between a name and an individual is required in order for the name to map onto the individual is an open question but not one that has been directly investigated from the perspective of neuroscience. Other empirical approaches have been used to shed light on this question. In the standard paradigm of experimental philosophy, Machery et al. () presented a version of Kripke’s Gödel case to forty undergraduates at Rutgers University and forty undergraduates from the University of Hong Kong. The specific case they used was the following: Suppose that John has learned in college that Gödel is the man who proved an important mathematical theorem, called the incompleteness of arithmetic. John is quite good at mathematics and he can give an accurate statement of the incompleteness theorem, which he attributes to Gödel as the discoverer. But this is the only thing that he has heard about Gödel. Now suppose that Gödel was not the author of this theorem. A man called ‘Schmidt’, whose body was found in Vienna under mysterious circumstances many years ago, actually did the work in question. His friend Gödel somehow got hold of the manuscript and claimed credit for the work, which was thereafter attributed to Gödel. Thus, he has been known as the man who proved the incompleteness of arithmetic. Most people who have heard the name ‘Gödel’ are like John; the claim that Gödel discovered the incompleteness theorem is the only thing they have ever heard about Gödel. When John uses the name ‘Gödel’, is he talking about: (A) the person who really discovered the incompleteness of arithmetic? or (B) the person who got hold of the manuscript and claimed credit for the work? Machery et al. (: B) The team found that intuitions varied across culture. The Westerners were significantly more likely than the Chinese to give causal-historical responses (response A).

OUP CORRECTED PROOF – FINAL, 21/1/2019, SPi



 

The researchers speculate that the difference in intuitions may be grounded in differences in cognition between Westerners and Chinese. The research casts doubt on a certain methodology used in the philosophy of language, known as ‘intuitions about cases’, but it does not provide any obvious starting point for neuroscience to investigate the reference of proper names. It remains an open question how most speakers in John’s situation who learned in college that Gödel is the man who proved the incompleteness theorem and who then subsequently learn that a man named ‘Schmidt’ discovered the incompleteness theorem would update the existing DRS (i.e., the mental representation) to reflect this fact. If speakers know nothing else about Gödel, it is plausible that they would integrate the name ‘Schmidt’ into the DRS as another name for ‘Gödel’. In that case, their uses of ‘Gödel’ and ‘Schmidt’ would both fail to refer to any actual individuals, as Gödel and Schmidt are different people. So, answering this sort of question about brain processing would not tell us anything useful about the reference of proper names. If participants learned that the man who was baptized ‘Gödel’ didn’t discover the incompleteness theorem but that a man named ‘Schmidt’ discovered it, they would in all likelihood reject the information they learned in college but this sort of discourse revision would merely reflect their knowledge of how we name people. This type of revision would be consistent with the condition ‘Gödel(x)’ mapping onto an individual only if he is named ‘Gödel’. It wouldn’t demonstrate a required historical-causal relation between ‘Gödel’ and that individual. The most promising neuroscientific approach to the reference of proper names may be to look at how updating takes place with proper names in the scope of modals (see Roberts ; and Kamp et al.  for some testable predictions). This type of research should look not only at proper names in alethic modal contexts but also embedded under epistemic operators such as ‘belief ’. Kripke’s causal-historical requirement is considerably less plausible for names embedded in epistemic modal contexts than in alethic modal contexts. For example, ‘I told John that Allen Stewart Konigsberg is Woody Allen but he doesn’t believe me. He thinks Allen Stewart Konigsberg is Woody Allen’s father’ cannot introduce the same discourse referent for all occurrences of ‘Allen Stewart Konigsberg’ and ‘Woody Allen’. Occurrences of names embedded in nonalethic modal contexts may thus require their own discourse referents (Roberts ; Kamp et al. ).

. M 

.................................................................................................................................. One of the main complaints about electrophysiological studies of semantics has been that they don’t study linguistic information. Pylkkänen et al. (), for example, point out that the traditional way of studying linguistic phenomena has been to identify grammatically ill-formed sentences on the basis of the judgments of native speakers and then generate linguistic theories to explain why sentences have the grammar they do. The problem with the electrophysiological studies that claim to be studying

OUP CORRECTED PROOF – FINAL, 21/1/2019, SPi

       ?



semantics, Pylkkänen et al. argue, is that the alleged semantically ill-formed stimuli are, in fact, not ill-formed in the formal sense of linguistic theory. They do not violate syntactic rules but simply conflict with world knowledge. To back up this claim, they point to the original studies performed by Kutas and Hillyard (). “He spread the warm bread with socks” is not grammatically ill-formed. It’s a perfectly grammatical sentence. The reason it gives rise to an N effect is that the participants know that socks are neither edible nor spreadable but this fact is a fact about the world, not a fact about language. They add that they are unaware of any semantic theory that would claim that typical N sentences are semantically ill-formed. The idea that the N effect is a semantics-related ERP, they say, is not conducive to the project of generating theoretically grounded models of how the brain processes language. However, this criticism seems to presuppose that a model of how the brain processes language must be grounded only in syntax or knowledge of truth-conditions. But such a model would provide a rather limited picture of language processing, and would have little to do with semantics, more broadly speaking. It is widely agreed that perfectly grammatical sentences can be semantically ill-formed, as witnessed by Chomsky’s ‘colorless green ideas sleep furiously’. Likewise, sentences that have possible truthconditions may be meaningless to speakers (e.g., “he spread the warm bread with socks”). In DRT a sentence is semantically ill-formed when it cannot coherently be added to the existing discourse representation. Outside of the realm of fiction, there are few contexts that would be able to accommodate a sentence like “he spread the warm bread with socks”, which ordinarily makes the sentence semantically ill-formed. Now, it is true that in the original ERP studies, unlike in later studies, only single sentences were presented in the experimental setting. But the participants used in the study were not blank slates. They came into the experimental setting with existing discourse representation structures, and DRT predicts that sentences are analyzed and integrated relative to existing representations. Furthermore, in DRT the discourse representation is supposed to serve not only as a mental model of the discourse but also as a mental model of the world when supplemented by discourse-independent information. So when seen against the background of DRT or other dynamic theories, the N effect can indeed be reliably interpreted as a semantics-related ERP. That said, no one is claiming that the N effect is exclusively semantic. As Van Berkum et al. () point out, there are no ERP effects that correspond uniquely to the exact level of language representation manipulated in experiments (syntax, semantics, reference, pragmatics). Many regularities have been unearthed, for example, difficulties of semantic and syntactic integration tend to be manifested as N and P effects respectively. However, language processing occurs at multiple levels of language representation, and what experimental researchers define as processing at one level (e.g., semantic) can affect processing at a different level (e.g., syntactic) and may even sometimes be manifested only at that level. A different methodological question is that of whether the standard distinction between a level of semantics and a level of pragmatics can be maintained within dynamic semantic theories. There are, of course, numerous linguistic phenomena

OUP CORRECTED PROOF – FINAL, 21/1/2019, SPi



 

that depend on pragmatic factors for their interpretation. But the question is whether there are good grounds for thinking that the brain processes semantic and pragmatic information by the same processing sites and in a particular order. There are studies that suggest that the brain processes semantic and pragmatic information differently (Hunt et al. ; Politzer-Ahles et al. ). Hunt et al. (), for example, had participants read existentially quantified sentences that had both a semantic meaning (‘at least one’) and a pragmatic meaning (‘not all’) presented next to an image with content related to the sentence. Sentences like ‘The boy cut some of the steaks in this story’ were presented together with images that made both the semantic and pragmatic interpretations true (when the boy cut some but not all), images that made neither true (when the boy cut none), or images that made the semantic interpretation true but the pragmatic one false (e.g., when the boy cut all the steaks). The largest N effects were found when the sentence was false both on its semantic interpretation and its pragmatic interpretation. When the sentence was false on its pragmatic interpretation but true on the semantic interpretation, there was an intermediary N amplitude. The smallest amplitude was observed for pictures that made both the semantic and the pragmatic interpretation true. Using a similar paradigm Politzer-Ahles et al. () further found that pragmatically false but semantically true sentences elicited a sustained posterior negative component that was distinct from the N effect. The researchers propose that the sustained negativity reflects that the integration requires cancellation of the pragmatic inference and retrieval of the semantic meaning. While these results indicate that the brain processes semantic and pragmatic aspects of meaning differently, the pragmatic aspects appear to be calculated fast enough to affect subsequent interpretation. The findings suggest that although the brain is capable of distinguishing pragmatic and semantic interpretations, new information that is integrated into the existing DRS may be the result of pragmatic inferences and the fact that the pragmatically inferred information is sometimes preferred, unless this information is inconsistent with the existing discourse representation. This is consistent with the basic tenet of DRT, which does not necessitate a rigid distinction between semantics and pragmatics. This is so insofar as it captures how we process what the speaker appears to attempt to convey to us rather than the semantic properties of language per se (cf. Recanati , ).

. C

.................................................................................................................................. Discourse Representation Theory (DRT) and other dynamic semantic theories have been considerably more successful than traditional semantics in accommodating linguistic phenomena, such as anaphora, presupposition, and event order. DRT predicts that discourse gives rise to a mental representation that is ever growing and continually revised. On this view, sentences have meanings only derivatively, depending on their capacity to change the existing discourse representation. The mental

OUP CORRECTED PROOF – FINAL, 21/1/2019, SPi

       ?



representations, not sentences, have a truth-conditional impact. A discourse representation consists of discourse referents and conditions introduced by linguistic expressions. New linguistic information either introduces new discourse referents into the discourse representation or refers back to discourse referents in the existing representation. The discourse referents refer to entities in the external world when these entities satisfy the referent and the conditions imposed on it throughout the entire discourse representation. Because DRT takes discourse representation to be the most central notion in semantics, it is suitable as an object of investigation for neuroscience. Since Kutas and Hillyard () discovered that difficulty of integrating new linguistic information into an existing discourse context has its own ERP signature—the N effect, a large body of research has looked at the ERP signatures associated with different linguistic phenomena. The majority of neuroscientific studies relating to reference have been concerned with anaphoric reference and related phenomena. The results of these studies lend evidence to the general framework of DRT. DRT implies that a discourse referent introduced by a referring expression can refer to a single individual only if that individual satisfies all the conditions imposed by co-referential expressions. Although the focus here has primarily been on DRT and discourse referents, however, it is quite plausible that speakers very often manage to refer to extra-mental entities even when most of the constraining conditions in the discourse representation are not satisfied. To borrow the classical example of Donnellan (), a speaker may succeed in picking out a man who is drinking water using the description ‘the man with the martini’. This sort of speaker reference is useful in contexts in which the purpose of the description is not to describe an entity but to pick it out for further conversation (Kripke ). The bulk of empirical work in this area has been in developmental psychology and computer science looking at how speakers generate referring expressions for singling out entities in a visual scene (see e.g., Dale and Reiter ). However, the ERP paradigms for studying discourse could easily be adjusted to look at eventrelated brain potentials in individuals generating referring expressions for picking out objects in a visual scene.

A I am grateful to Barbara Abbott and Jeanette Gundel for invaluable comments on an earlier draft of this paper.

OUP CORRECTED PROOF – FINAL, 17/1/2019, SPi

  ......................................................................................................................

   An electrophysiological perspective ......................................................................................................................

    

. I

.................................................................................................................................. S INCE the pioneering work of Kutas and Hillyard (, ), there has been a proliferation of event-related potential (ERP) studies investigating how language comprehension unfolds in real time. During an ERP language experiment, participants are presented with linguistic stimuli while their continuous electroencephalogram (EEG) is recorded. Upon completion of the experiment, the brain’s response to multiple instances of the same experimental condition is averaged to yield waveforms known as ERPs. These voltage deflections, time-locked to the stimulus of interest, differ in terms of their polarity (positive or negative), latency, distribution at the scalp, and morphology. Though the nature of the physiological processes measured by EEG largely precludes any specific claims about the anatomical sources of these responses, the multi-dimensional nature and high temporal resolution of ERPs does enable researchers to draw many types of inferences (see, e.g., Rugg and Coles  and Van Berkum et al.  for lengthy discussions). For the purposes of this chapter, we will focus largely on what have been referred to as sensitivity, timing, and identity inferences (Van Berkum et al. ). Put generally, we will be discussing studies that take advantage of the high temporal resolution of the ERP technique to determine what the brain is sensitive to and when that sensitivity occurs. Perhaps more importantly, we will also evaluate claims about the identity of brain responses to different types of anaphoric relations (i.e., the underlying cognitive processes that they reflect), and the consequences that these claims have for the architecture of the language system and the functional specificity of brain responses to language input. In the early years of linguistic ERP research, researchers focused primarily on semantic (e.g., Kutas and Hillyard ) and syntactic violations (e.g., Neville et al. ; Osterhout

OUP CORRECTED PROOF – FINAL, 17/1/2019, SPi

  



and Holcomb ; Kluender and Kutas a; Friederici et al. ) and their attendant electrophysiological indices, the N and eLAN/LAN/P, respectively. However, recent research has expanded in scope, and much progress has been made in terms of both mapping the antecedent conditions under which language-related ERP components are elicited, and understanding the myriad factors that influence online language comprehension (see Kutas et al.  for a comprehensive review of language-related ERP research). Long-distance relations, in which two non-adjacent elements depend on each other for their interpretation, have received a great deal of emphasis in the literature because of the unique processing challenges that they pose to the comprehension system. While it is true that somewhat more attention has been paid to purely syntactic relations effected by displacement of sentence elements from their canonical positions (e.g., relative clauses, wh-questions, and scrambling constructions), there has nonetheless also been a steady and productive stream of investigations into referential long-distance relations. While many of the early studies addressed the outright violation of well-formedness conditions on anaphoric relations, there arose a separate line of research that focused on the resolution of referential ambiguity and the type of brain responses this elicited. A little over ten years ago, a comprehensive review of ERP studies on anaphoric constructions appeared (Callahan ) that has unfortunately received little attention since, and is as a consequence not widely known today. We are however long-time admirers of the scope, thoughtfulness, and theoretical lucidity of this work, and therefore felt it both prudent and convenient to revive and utilize Callahan’s review as a starting point, in order to assess how things have developed over the decade since its publication. We highly recommend Callahan’s original review for its historical interest, but by way of introduction, we will briefly summarize the relevant points here before moving on to a discussion of more recent research. Callahan outlines three different types of constraints on anaphoric dependencies: (a) syntactic constraints on the various environments in which different types of anaphors (noun phrases versus pronouns versus reflexives/reciprocals) may occur, (b) discourse constraints on the choice of reference form (often referred to as ‘accessibility’ or ‘givenness’; cf. Ariel ; Gundel et al. ), and (c) processing constraints on both the identification of possible antecedents for an anaphor (‘bonding’) and the determination of which of these possible antecedents is most appropriate (‘resolution’; cf. Garrod and Terras ). Callahan likewise proposes three basic configurations in which anaphoric dependencies occur: (i) co-indexation of an overt antecedent with an overt anaphor (e.g., pronouns, reflexives, repeated noun phrases, and synonyms), (ii) co-indexation of an overt antecedent with a null anaphor (e.g., VP-ellipsis and verb gapping), and (iii) co-indexation between a displaced overt phrase associated with a null anaphor in the original underlying sentence position (e.g., relative clauses, wh-questions, and scrambling constructions). She does not consider cases in which either overt (e.g., cataphoric elements) or null anaphors (e.g., in head-final relative clauses; see §...) precede their antecedents, for example, ‘Even though shei realized the risk shei was taking in doing soj, Hillaryi went ahead and wrote her tellall bookj.’. We return to these types of relations in §...

OUP CORRECTED PROOF – FINAL, 17/1/2019, SPi



    

In cases of simple co-indexation, the anaphor is asymmetrically dependent on the antecedent for its discourse interpretation, regardless of whether it is overt or null, or whether it precedes (cataphora) or follows (anaphora) the antecedent. This can nonetheless lead to difficulties of identification, especially in VP-ellipsis, when antecedent and anaphor are not strictly identical to each other in morphological form, as in the example above, and in this recent attested example: “The problem is that OMB [the Office of Management and Budget] didn’t listen and we are.” (http://www.cnn.com/ ///politics/congress-trump-war-power/index.html) In cases of syntactic displacement, on the other hand, anaphor and antecedent are always symmetrically and mutually dependent on each other for their interpretation— again, independent of linear order. In these cases, however, problems of interpretation can arise when a null anaphor precedes its antecedent, as happens sometimes in headfinal languages. In such cases, an anaphor may not be identified as such immediately, because it could alternatively be an argument that was simply dropped for discourse reasons, as we will discuss in §.... Callahan raises the reasonable question of whether the processing of anaphoric relationships differs on the basis of whether displacement is involved or not, or on the basis of whether anaphors are overt or null. As will become clear during the course of our discussion, in the ensuing years we have begun to discern some of the answers to these questions. Callahan also provides a classification of the processing operations required by each type of anaphoric relationship, but with the / hindsight provided by a decade of subsequent research, these can now be fleshed out in more fine-grained detail than was possible at the time of her original review. In the case of mere co-indexation of an antecedent with an overt anaphor (i.e., when no displacement is involved, such as in a pronoun–antecedent relationship), the anaphor must simply be associated with the antecedent by retrieving the latter from working memory. When the co-indexed anaphor is null, however, it must first be identified as the absence of a required sentence constituent. Then, just as in the case of an overt anaphor, the antecedent must be retrieved from working memory and associated with the anaphor. In the case of displacement (as in a wh-question), the antecedent (in this case, the displaced element) must first be recognized as such, so that it can be encoded and later associated with the null anaphor (the gap site) once the latter is identified. When the null anaphor is located, again as indicated by the absence of a required sentence constituent, the antecedent must be retrieved from working memory. In all three cases, the discourse referent indexed by the two separate referential forms eventually needs to be integrated into the sentence- and discourse-level context. These operations can be distilled into the following inventory ranging across all anaphoric relationship types discussed thus far: () Identification and encoding of an antecedent that is out of its canonical position. () Identification of a null anaphor by registering the absence of a required sentence constituent.

OUP CORRECTED PROOF – FINAL, 17/1/2019, SPi

  



() Association of any anaphor with its antecedent by retrieving the latter from working memory. () Integration of the referent indexed by the two reference forms into the sentence context and ongoing discourse representation. While this inventory diverges somewhat from Callahan’s, we follow her in suggesting that each of these operations may be associated with a particular ERP response. It is this platform that we will use as a springboard for our own discussion of subsequent ERP investigations of anaphoric relationships. There seems to be relatively broad consensus in the literature with regard to operations () and (): both are indexed by anterior negativity, often left-lateralized, and either phasic/transient (i.e., of short duration) or sustained (i.e., ongoing), depending on the time window analyzed. Overt anaphors, whether repeated noun phrases and synonyms (Anderson and Holcomb ) or pronouns (Streb, Rösler, and Hennighausen ), elicit phasic effects of (left) anterior negativity—often abbreviated as LAN—when preceded by coreferential antecedents (). Such effects will be the focus of §., §.., and §... It also seems relatively clear that both the encoding of a displaced antecedent in working memory () and its retrieval upon identification of the associated null anaphor () are indexed by (left) anterior negativity. Between antecedent and anaphor, sustained negativity distributed over the front of the head that persists for several seconds has consistently been observed (though see McKinnon and Osterhout ; Harris ; and Kaan et al.  for counterexamples to this generalization). Interestingly, when this overall response is reanalyzed by re-baselining each individual word in the multiword average (i.e., the ERP waveforms in response to both conditions are realigned at the beginning of each word occurring between antecedent and anaphor), there is no additional effect. In other words, the ERPs to the condition involving displacement and its control do not differ at intermediate positions (King and Kutas ; Phillips, Kazanina, and Abada ; Ueno and Kluender ). The LAN effect is thus elicited only when the antecedent that needs to be encoded in working memory is encountered, and once more when the null anaphor necessitating its retrieval is encountered. For this reason, the LAN effect is now thought to index only memory operations of encoding and retrieval, and not of storage, as was originally thought to be the case (Kluender and Kutas a; King and Kutas ). This is why there is no reference to storage (in the form of activation) in the operations listed above. With regard to theoretical processing models of the working memory system, these findings are not consistent with the idea that words intervening between an antecedent and a null anaphor add cumulatively to working memory load (Gibson , ), but are more in line with models that emphasize content-addressable, cue-based retrieval instead (Gordon et al. ; McElree et al. ; Lewis and Vasishth ). Identifying the ERP response indexing the identification of null anaphors as given in () is more difficult, as the responses observed in these contexts are more varied. It has long been recognized that effects of late positivity are elicited near the end of a

OUP CORRECTED PROOF – FINAL, 17/1/2019, SPi



    

dependency created via displacement when it is compared to a control condition in which no displacement has occurred (Kaan et al. ; Fiebach, Schlesewsky, and Friederici ; Felser, Clahsen, and Mūnte ; Ueno and Kluender ; Phillips, Kazanina, and Abada ; Hagiwara et al. ). These effects have been interpreted as indexing processes of syntactic integration related to the cognitive operations in () above. However, there are two problems with this interpretation: one empirical and one conceptual. First, the original evidence used to support this interpretation was not entirely convincing, relying as it did on marginal interactions at two electrodes, in the absence of any significant interactions in the overall analysis (Kaan et al. ). Second, these late positive responses are typically elicited at the word immediately preceding the null anaphor (i.e., the gap position), whereas LAN effects are more typically elicited at the word following the null anaphor. If anterior negativity indexes some aspect of retrieval from working memory, as we postulated above, interpreting late positive responses occurring prior to a gap as indexing syntactic integration makes little sense, as retrieval necessarily precedes integration. Our position, therefore, is that late positive responses elicited at pre-gap positions instead index identification of the null anaphor itself. If this is the case, then effects of late positivity at the end of a dependency align more closely with the identificational processes outlined in (), rather than with the integrative processes described in (). This is very much in line with research on late positive responses, so-called P effects, the amplitude of which are well known to index identification of task-relevant target stimuli in contexts in which one’s mental model of the environment must be updated based on new input (Donchin and Coles ). While we do not have the space to buttress this specific and admittedly somewhat controversial claim any further here, as it is orthogonal to the aims of this chapter, it does raise a separate issue that will be a general theme of our discussion, namely the classification and definition of ERP responses in language contexts. Let us therefore briefly discuss the past history of component nomenclature in the linguistic ERP literature using two specific examples. Late positive responses in sentence contexts were initially labeled the P component (Osterhout and Holbcomb ). Note that this terminology relies solely on the physical parameters of the response, and that no claims are made about the underlying cognitive processes that the component is assumed to reflect. Subsequent to this initial naming convention, however, late positive responses in sentence contexts began to be labeled as the “syntactic positive shift” (SPS; Hagoort, Brown, and Groothusen ). Note that in this case, rather than relying on physical parameters such as polarity and latency, the label makes explicit reference to the cognitive (namely syntactic) processes that the response is assumed to index. We will refer to this type of label for an ERP response as a ‘functional definition’, a concept that will become important below during our discussion of the brain’s response to ambiguous contexts in which establishing an anaphoric relation is problematic. With these two competing labels in circulation, researchers were for years obliged to refer to the response as the ‘P/SPS’ to avoid sociopolitical sins of omission.

OUP CORRECTED PROOF – FINAL, 17/1/2019, SPi

  



The SPS label, however, eventually fell into disuse as it was convincingly demonstrated that the response was not in fact specific to syntactic contexts, but rather could also be elicited by, for example, spelling errors (Münte et al. ; see also Vissers, Chwilla, and Kolk ). At the same time, the objection was made that the P was in fact simply a P in linguistic clothing (Coulson, King, and Kutas a). This claim led to an extended debate (Coulson, King, and Kutas b; Osterhout ) that has yet to be entirely resolved to anyone’s satisfaction. For this reason, many researchers now simply refer to this type of response in language contexts as late positivity, or a late positive complex (LPC), in order to remain agnostic.¹ We will follow this convention in the remainder of this chapter. The second example of a premature, overly narrow functional interpretation involves early negativities, sometimes also left-lateralized and therefore abbreviated as eLAN. While the name of this ERP response appropriately refers to its polarity, latency (to distinguish it from ordinary LAN effects), and scalp topography, it has been claimed specifically to reflect failure of the first-pass operations of an automatic, serial parser that builds syntactic phrase structure on the basis of word category information alone (Friederici, Pfeifer, and Hahne ; Friederici ). This interpretation has come under increasing fire in recent years as the stimulus paradigms typically used to elicit eLAN effects, as well as the methodology used to analyze them, have been called into question (Steinhauer and Drury ). In addition, early negativity can be elicited even when the unexpected word is of the same grammatical category as the expected word (Roll, Horne, and Lindgren ; Zhang and Zhang ), and has been elicited merely by mismatches between the expected physical form of an upcoming word and the form of the actual word presented (Dikker, Rabagliati, Farmer, and Pylkkänen ), including in highly constrained semantic contexts (Dikker and Pylkkänen ; see also Besson and Macar ). This excursus on the nomenclature and specific functional definition of ERP components elicited in language contexts provides the point of departure for our discussion of electrophysiological studies of referential anaphora. Namely, we adopt the position that all ERP components are most profitably assumed to be maximally cognitive in nature until proven otherwise. In other words, in the absence of the type of overwhelming supporting evidence that has yet to be demonstrated for any response, components should preferably not be assigned overly narrow language-specific functional interpretations. It seems that prematurely narrowing the range of operations that a linguistic ERP component might possibly index—especially when the interpretation is based exclusively or primarily on the type of manipulation undertaken and/or the

¹ One problem with the P label itself is that there was actually no consistent clear peak at  msec. in the broad response of the original Osterhout and Holcomb () report, nor has there been a recognized peak reported in the response since then. To be fair, this is also true of the P. Contrast this with the stable, reliable peak of the N between about  and  msec. post-stimulus onset.

OUP CORRECTED PROOF – FINAL, 17/1/2019, SPi



    

specific stimulus materials involved—can lead the field as a whole down a rabbit hole from which it can take decades to re-emerge. In the interest of optimally efficient and speedy progress as a field, we advocate taking a longer view: that is, a more cautious, conservative approach of systematically eliminating broader interpretations on the basis of carefully controlled studies manipulating the relevant variables, as has been the case for example in the decades-long attempt to elaborate the exact functional interpretation of the N. It is this approach that we will recommend and attempt to apply to referential processing in this chapter.

. T     (N)

.................................................................................................................................. In the previous section, using the review in Callahan () as a starting point, we outlined a broad classification of the cognitive operations involved in processing different types of anaphoric relations, and attempted to establish a mapping between these operations and corresponding brain responses. In parallel to the avenue of research focused largely on working memory operations, a series of studies emerged focusing on factors that influence the processing of anaphors with ambiguous reference. In this section we discuss these studies in the order in which they were published, focusing on functional interpretations of what came to be known as the referentially induced frontal negativity (Nref ), an effect observed in response to anaphors that have multiple available candidate antecedents. Discussing this work in chronological order will allow us to trace the development of thought regarding the functional interpretation of these responses, as well as to track the consequences that this development has had for the ERP literature on language processing as a whole.

.. Van Berkum, Brown, and Hagoort () Jos Van Berkum and colleagues () reported the results of an ERP study designed to investigate the processes involved in establishing reference in multi-sentence discourse contexts. At the time, the brain signatures of forming referential relations (more specifically, the difficulties associated with this process) had yet to be fully described. Aiming to fill this gap in knowledge, the study was designed to yield information on the immediacy and incrementality of referential processing, and the manner in which discourse-level relations guide processing at other linguistic levels. The basic idea was to determine the brain response elicited in linguistic contexts that make establishing reference problematic. Collapsing across tangential details of the

OUP CORRECTED PROOF – FINAL, 17/1/2019, SPi

  



experimental design for ease of exposition, the authors had subjects read short stories of the following types: ()

-referent condition David had asked the boy and the girl to clean up their room before lunchtime. But the boy had stayed in bed all morning, and the girl had been on the phone all the time. David told the girl that had been on the phone to hang up.

()

-referent condition David had asked the two girls to clean up their room before lunchtime. But one of the girls had stayed in bed all morning, and the other had been on the phone all the time. David told the girl that had been on the phone to hang up.

In the -referent condition, compared to its -referent counterpart, the NP anaphor the girl has multiple candidate antecedents (the two girls) and is therefore referentially ambiguous. If this information is taken into account rapidly by the parser, one would predict immediate “effects of referential ambiguity” (Van Berkum et al. : ) in response to the girl (i.e., before the relative clause restricts the reference to the relevant girl in question). This would suggest that referential processing is immediate and incremental, and that it occurs in parallel with the semantic and syntactic processes indexed by the N, [e]LAN, and late positivity. This is exactly the pattern of results that was observed: compared to the -referent condition, the -referent condition elicited a larger negativity with an early onset (around  msec.), that was significant between – msec. and visibly (but not statistically) maximal over anterior electrodes. Relying largely on the literature on syntactic long-distance dependencies discussed in §., the authors interpreted this effect as reflecting the additional use of memory resources required to process an anaphor with two potential antecedents. This increase in processing cost, they argued, could be because the ambiguous anaphor requires two potential antecedents to be maintained in memory, or because it triggers additional attempts at antecedent retrieval. This account of the effect is in broad agreement with our own, namely that anterior negativities elicited in referential contexts index domain-general memory operations. We will return to this issue in §.. and §... For the purposes of this chapter, it is crucial to point out the broad scope of this functional interpretation. Van Berkum et al. () did not claim that the effect was specific to reference, or even to language. Instead, because they felt that it likely indexed well-known memory processes of maintenance and search, they drew on similarities in latency, polarity, and scalp distribution between their effect and negativities observed in response to other construction types known to tax working memory. As such, the cognitive operations that this response was argued to index could, in theory, be easily incorporated in the classification scheme laid out in §.. As we discuss the research that followed these initial findings, however, we will observe a shift in interpretation towards functional specificity that was apparently a simple consequence of including a functional description (namely, ‘referentially induced’) in the label attached to the response (the Nref ).

OUP CORRECTED PROOF – FINAL, 17/1/2019, SPi



    

.. Van Berkum, Brown, Hagoort, and Zwisterlood () In Van Berkum et al. (), the authors addressed two methodological concerns associated with Van Berkum et al. (). In the first experiment, they replicated their earlier findings by showing similar effects of (anterior) negativity in response to ambiguous anaphors embedded in natural connected speech, showing that this effect was not modality dependent. The effect was significant in both the – and – msec. time windows, indicating a more sustained effect than had been originally reported in the () experiment using visual presentation. In the second experiment, the proportion of trials containing ambiguous anaphors was reduced from % to .% in order to determine whether the effect of referential ambiguity was dependent on the probability of encountering an ambiguous anaphor. Though the reduced stimulus-to-noise ratio resulting from this manipulation led to a null main effect of ambiguity at midline electrodes, a distributional analysis revealed a main effect of ambiguity at left and right anterior, and right posterior electrodes between – msec. This effect remained significant at left and right anterior sites between – msec. These findings were again discussed in terms of the immediacy and incrementality of referential processing, and linked to the literature on syntactic long-distance relationships: “Although within the context of this work we sometimes refer to the observed negativity elicited by referentially ambiguous nouns as the ‘referentially induced negative shift’ or the ‘referentially induced ERP effect,’ it is important to reemphasize that we do not claim that this ERP effect is uniquely associated with referential ambiguity in language comprehension.” Van Berkum et al. (: )

In other words, the authors make clear that they are not assigning a novel and narrow functional significance to these responses. However, as will soon become clear, even choices made for expository convenience can have lasting consequences for component identity.

.. Van Berkum, Zwitserlood, Bastiaansen, Brown, and Hagoort () Further pursuing this line of research, Van Berkum et al. () sought to determine whether (a) all antecedent-less anaphors elicit anterior negativity, and (b) whether the brain responds similarly to all referential problems. They compared sentences of the following types: ()

David shot at Linda as he jumped over the fence.

()

David shot at John as he jumped over the fence.

()

Anna shot at Linda as he jumped over the fence.

OUP CORRECTED PROOF – FINAL, 17/1/2019, SPi

  



Relative to its unambiguous counterpart in (), the ambiguous pronoun in () elicited an effect of sustained negativity between – msec., similar to that observed in response to the NP anaphors in Van Berkum et al. (, ). Note that in contrast to the previous studies, these materials do not contain a post-anaphor disambiguating relative clause, showing that the observed effect does not merely reflect the anticipation of disambiguating material. Additionally, the antecedent-less pronoun in () elicited a late positive effect compared to the unambiguous pronoun in (), argued to index so-called “referential failure.” As such, this work showed that the elicitation of sustained negativity during reference assignment does not depend on the type of anaphor (NP versus pronoun), and that the brain responds differently in cases when a pronoun has too few available antecedents (late positivity) compared to when it has too many (anterior negativity). It is worth noting that the – msec. time window chosen for analysis was based on visual inspection of the waveforms rather than on predictions stemming from previous findings. In Van Berkum et al. (, ), in which responses to referential ambiguity were interpreted as being similar to LAN effects in the literature on syntactic dependencies, observed negativities were primarily analyzed between – msec., the same time window that LAN effects are often analyzed in. In some sense then, adopting a new time window begins the process of severing the link between negativities observed in referential contexts and well-established LAN effects, a process that began in earnest in subsequent work.

.. Nieuwland and Van Berkum () Nieuwland and Van Berkum () went on to investigate the roles that individual differences in language processing skills (measured by the reading span task, a widelyused test of working memory) and contextual bias (the degree to which a candidate antecedent is in focus in the discourse model) play in the resolution of referentially ambiguous anaphors. The -referent condition in these materials consisted of a story describing an interaction between two main characters of the same gender (the candidate antecedents), and the critical word consisted of an ambiguous pronominal anaphor. In some sentences the discourse properties of the anaphor strongly biased towards either the first or second antecedent, and in some sentences there was no such bias and each antecedent was an equally likely candidate. The result was a main effect of referential ambiguity (collapsing across all levels of contextual bias) that was significant between – msec. and maximal over anterior electrodes. Moreover, the magnitude of the ambiguity effect was significantly correlated with reading span score (with higher scores associated with larger amplitude effects), and a median split showed a significant effect only in high-span subjects. The authors observed that their reading span results were also consistent with those in King and Kutas (), Müller et al. (), and Münte et al. (), all studies that showed relationships between language processing abilities and the amplitude of sustained negativities.

OUP CORRECTED PROOF – FINAL, 17/1/2019, SPi



    

.. Nieuwland, Otten, and Van Berkum () Nieuwland et al. () conducted a further study to determine whether the negativities elicited in response to ambiguous anaphors reflect superficial text-based ambiguity or deep ambiguity at the level of the situation model. The materials for this experiment contained three types of stories: unambiguous controls, and two types of stories in which the critical NP was formally ambiguous and could in theory refer to two possible antecedents. The crucial manipulation was that—in one ambiguous story type—two antecedents were equally accessible in the situation model (for example, a story of a family gathering in which an uncle is talking to two nephews, the candidate antecedents, about different topics), while in the other type, only one antecedent was accessible at the discourse-level (a story similar to the one above, except that one of the nephews leaves the gathering and goes elsewhere). There was a sustained effect of negativity significant from – msec. in response to ambiguous anaphors, but only in cases in which both antecedents were equally accessible in the situation model (i.e., both nephews still present at the family gathering). These effects were again discussed in the context of the memory burdens imposed by (truly) ambiguous anaphors.

.. Van Berkum, Koornneef, Otten, and Nieuwland () Van Berkum et al. () reported the results of an experiment examining the role of implicit causality in reference resolution, but as these results are tangential to our current purposes, we focus on the literature review found in this paper, and refer the interested reader to the original work for details of the experiment. In the introduction, the authors do something quite important: they introduce the Nref acronym to refer to the effects under discussion. Though the authors point out that this term is used for “ease of reference,” we believe that this terminological decision has had important consequences. To this point, negativities elicited by referential ambiguity had been linked to sustained effects of anterior negativity observed in response to various constructions that tax working memory. Note that the constructions known to elicit such effects vary both within and across languages, and the entirely descriptive label applied to them does not tie the response to any one construction type or linguistic level of analysis. By introducing the Nref label, for expositional convenience or otherwise, Van Berkum et al. () reified a distinction between their effects and those that had previously been reported in the literature, and the consequences of this shift can be seen in subsequent discussions of ERP responses to referential ambiguity. In any case, regardless of the many similarities between previously reported effects of sustained anterior negativity in the literature and these referentially induced responses, both in terms of their physical parameters and the cognitive operations they have been

OUP CORRECTED PROOF – FINAL, 17/1/2019, SPi

  



argued to reflect, the anterior negativities elicited in referential contexts now have a special name: the Nref. We next focus on papers from Van Berkum and Nieuwland’s group published from  onwards. Though their own research on referential ambiguity has slowed to some extent,  appears to have been something of a turning point in terms of the way responses to ambiguity are interpreted. A quick Google search shows how broadly this terminology was adopted in the primary literature: the  paper discussed above has been cited  times (as of this writing) with the overwhelming majority adopting the Nref label and the narrow functional interpretation that began to be associated with it.

.. Nieuwland and Van Berkum (b) Nieuwland and Van Berkum (b) investigated the interaction between referential ambiguity and semantic incoherence. Results showed an effect of sustained negativity in response to ambiguous anaphors between – msec. (notably now directly referred as an Nref response, “the known signature of referential ambiguity,” rather than an anterior negativity), while anaphors that were both ambiguous and semantically incoherent elicited late positivity rather than a sustained negativity. These results were interpreted as showing that semantic incoherence actually precludes the reader from experiencing an anaphor as ambiguous, in the same way that verb bias or discourse context can. Though the authors focus more on their N/ late positivity findings, it is worth noting that they state, “as expected, ambiguous anaphors elicited a sustained, frontal negative ERP effect that is typically associated with referential ambiguity (Nref effect).” (: ). There is no mention of (L)AN effects or memory operations. The implication is clear: the effect is no longer considered to be part of a larger family of anterior negativities (plausibly) indexing working memory operations; it is instead an Nref that is specifically associated with referential ambiguity.

.. Nieuwland () Nieuwland () investigated the processing of pronouns that mismatch a single licensed antecedent in gender (“The boy thought that [s]he . . . ”) while varying the task across three experiments. In the first, participants were instructed to “think of a new referent that would fit the described scenario most plausibly, but were not instructed about who this referent should be” (: ). This manipulation was used in order to discourage participants from processing mismatching pronouns as syntactic anomalies. In the second and third experiments, participants were told just to read for comprehension (and to answer comprehension questions in the third experiment).

OUP CORRECTED PROOF – FINAL, 17/1/2019, SPi



    

In all three cases, mismatching pronouns elicited sustained negative effects (with a left anterior maximum in the first two experiments and a widespread negativity over the entire scalp in the third); there was a late positive effect in response to mismatching pronouns only in the first experiment with the explicit task manipulation. Replicating Nieuwland and Van Berkum (), the magnitude of these sustained negative effects again correlated with reading span score in Experiments  and . But instead of tying this correlation directly to more general working memory operations, these negativities were functionally circumscribed with an interpretation highly specific to not just referential processing, but to referential ambiguity resolution in particular: “The Nref effect may index only the initial perception of referential ambiguity from not knowing who the referent was. Alternatively, the Nref perhaps reflected an ongoing attempt to overcome the perceived ambiguity, perhaps by inferring the intended referent.”

This illustrates a further peril of assigning functional labels to ERP responses: the functional label can begin to drive the interpretation of results by forcing a line of reasoning based on the type of manipulation that originally elicited the response— regardless of how well that description fits the response to the manipulation under consideration. Recall that initial interpretations of the Nref were based on the additional memory operations presumably required by the processing of anaphors with multiple candidate antecedents. On the other hand, gender mismatches between antecedent and anaphor were instead found to elicit effects of late positivity, which was interpreted as an index of referential failure. In this study of gender mismatches, the effects of interest were referred to as anterior negativities in the results section and in the figures, in apparent recognition of the fact that they were unexpected in this context, and their amplitude was reported to correlate with reading span scores, a recognized measure of working memory capacity. Yet they were nevertheless treated as Nref effects in the discussion section, on the apparent assumption that because the study addresses referential processing difficulty in general, the response must be an Nref or, at minimum, more similar to an Nref than to a LAN.

.. How different is different? This raises an important question: is there sufficient evidence to posit the existence of both (L)AN and Nref effects as ontologically separate components with different functional identities? If so, what might this evidence look like? Or could the two ‘types’ of effects be more profitably and parsimoniously viewed as belonging to the same family of anterior negativities? This immediately raises the question of how to determine whether ERP responses are “qualitatively distinct.” Van Berkum () has the following to say about this.

OUP CORRECTED PROOF – FINAL, 17/1/2019, SPi

  



“It is important to know that, by and large, the ERP research community interprets non-identical polarity, morphology, and scalp distribution as reliable cues to the presence of qualitatively different cognitive processes. Furthermore, if two ERP effects are deemed identical in terms of these three features, then (reasonable) differences in their time course or magnitude are most commonly interpreted as reflecting quantitative variations.”

The point is well taken: if we truly want to posit the existence of two classes of ERP responses, ideally there should be minimal variation within a class and large differences between classes. The question is, then, do observed differences in the physical parameters of LAN and Nref effects satisfy the criteria for them to be classified as two qualitatively distinct components? In answer to this question, it has been repeatedly pointed out in the literature that Nref effects are both more sustained and more bilaterally distributed than LAN effects. But if differences in time course are quantitative in nature, and cannot be used to distinguish ERP responses in a qualitative sense, then duration should not be a deciding factor. Scalp distribution could be, but anterior negativities observed in response to the second element in a syntactic dependency can not only persist over multiple word positions (Barkley et al. ), but can also have a bilateral distribution (King and Kutas ; Ueno and Kluender ). Crucially, Nref effects show similar differences in scalp distribution across studies: central (Van Berkum et al. ), anterior (Nieuwland and Van Berkum ), left anterior (Nieuwland : Experiments  and ), and widely distributed over the entire scalp (Nieuwland : Experiment ). Thus if LAN effects differ from Nref effects along a certain qualitative parameter, but Nref responses elicited in different studies differ along the same qualitative parameter, then it becomes difficult to draw a qualitative distinction between the two. Typically, distinct ERP components will differ in more than one qualitative parameter. For example, the LAN and the N share the same timecourse, but differ clearly in scalp distribution (anterior versus posterior) and usually in morphology as well. In light of these observations, it is our opinion that the strict criteria that need to be satisfied in order to classify two ERP responses as belonging to qualitatively different classes are not met for LAN and Nref responses: both classes of response show substantial variation class-internally, and there is a large degree of overlap in the physical parameters of the two². To be clear, we are not claiming that the functional interpretation of ERP effects should not evolve over time in response to new evidence or that, in the presence of truly novel results, that researchers should never posit the ² It is worth noting that there is some suggestive evidence in the fMRI literature on anaphoric processing that LAN and putative Nref effects may be tied to activity in non-overlapping brain areas. Nieuwland et al. () localized Nref effects to medial frontal areas, while studies on syntactic LANeliciting dependencies typically elicit responses in more left-lateralized language areas (see e.g., Just et al. () and Caplan et al. ()). However, if the generators of these responses are truly anatomically distinct, one might well expect consistently different scalp distributions for LAN and Nref effects and, as discussed above, this is not the case.

OUP CORRECTED PROOF – FINAL, 17/1/2019, SPi



    

existence of a new class of brain responses. This is in fact a necessary, and positive, consequence of ERP research. However, we are aware of no study whose interpretation relies on the Nref for which an equally satisfactory interpretation could not be provided in which the effect is classified as a LAN. As a consequence, we feel that the number of ERP primitives necessary to account for the processes involved in language comprehension has unnecessarily been increased by one.

. M     

.................................................................................................................................. In this section, we review more recent ERP studies of referential processing that address issues beyond the questions of how referential ambiguity is resolved, or what happens when reference fails. These studies in turn shed light on some of the issues we raised in §. regarding the scope of the functional interpretation of ERP responses elicited by referential manipulations and, as we shall see, will highlight the need to revise our notions of the operations involved in relating anaphors to their antecedents, as outlined in §..

.. A direct comparison of syntactic and referential anaphoric dependencies Throughout our discussion, we have suggested that syntactic and purely referential dependencies might be processed in similar ways. We addressed this question directly in English in Barkley, Kluender, and Kutas (). Based on our review of the referential processing literature (see §.), it became apparent that no one had investigated how the brain processes simple, grammatical antecedent-pronoun relations devoid of violations or ambiguity. We therefore decided to undertake a straightforward comparison of the brain responses elicited by English object relative clauses with those elicited by simple pronoun–antecedent relationships. The comparisons undertaken were as follows. First, we adopted materials from King and Kutas () in order to compare object relative clause sentences with a control condition consisting of simple conjoined clauses with the same lexical content.³ ³ Our aim here was to mirror the design of our pronoun comparison, in which a sentence with a clear (intrasentential) referential dependency () was compared to a sentence with no such dependency (). For this reason, we opted not to use subject relative clauses as the control condition. With / hindsight, this was perhaps not the most perspicacious decision, because when replicating an effect, it is always best to leave the materials entirely unchanged. As things turned out, the LAN effect we elicited in response to the main clause verb of the object relative sentences had more of a left lateral distribution

OUP CORRECTED PROOF – FINAL, 17/1/2019, SPi

  



()

The soldier [who the sailor roughly pushed . . . ] smashed a bottle against the bar.

()

The soldier roughly pushed the sailor and smashed a bottle against the bar.

Based on previous results, we expected that this comparison would yield LAN effects at the main clause verb (smashed). This response was to serve as a baseline against which to compare the brain responses elicited by the anaphor in simple pronoun-antecedent relations. In addition, we presented participants with sentences containing a straightforward antecedent–pronoun relation () to sentences in which no antecedent preceded the pronoun (). ()

After a covert mission that deployed Will for nine terrible months, he longed for home.

()

After a covert mission that required deployment for nine terrible months, he longed for home.

Note in particular that there is nothing ambiguous about the referential relationship in (): the pronoun he has a single antecedent (Will) with which it unambiguously co-refers. It is also worth noting that the critical pronoun in () is not rendered antecedent-less through the use of a gender manipulation that introduces a (morpho-) syntactic violation (as in Van Berkum et al. ; see §..). By constructing the materials in this way, we were able to make the simple comparison between responses elicited by a pronoun with an antecedent compared to one without, without being concerned that these effects might be obscured by brain responses to (morpho-) syntactic ill-formedness. One of the operations involved in the processing of anaphoric dependencies outlined in §. is that anaphors need to be associated with their antecedents, the latter of which must be retrieved from working memory. Recall that the electrophysiological index of this process across studies—particularly those involving dependencies caused by syntactic displacement—is a LAN effect. We had therefore predicted that a pronoun preceded by an intrasentential antecedent () should also elicit a standard LAN effect in comparison to a pronoun with no antecedent in the same sentence (), as long-distance co-referential relationships pose the same processing challenges as long-distance syntactic dependencies (i.e., ‘association at a distance’). This turned out to be the case, and was the most striking of our results: the main clause subject pronoun ‘he’ in () elicited a clear LAN effect relative to the main clause subject pronoun in (), which lacked an intrasentential antecedent.

instead of the expected left anterior scalp topography. However, when we used the exact same object relative and conjoined clause materials in a subsequent experiment with a different set of participants, the object relative sentences elicited the expected standard LAN effect with a canonical left anterior distribution (Barkley, Kluender, and Kutas in preparation).

OUP CORRECTED PROOF – FINAL, 17/1/2019, SPi



    

This LAN effect moreover exhibited a latency (– msec.)⁴ and anterior scalp distribution very similar to that of a classic Nref effect. This is consistent with our suggestion in §. that the Nref response may merely be a member of a broader class of anterior negativities, indexing increased difficulty in retrieving an appropriate antecedent from working memory when multiple candidates are present. Just as the attempt to associate an anaphor with one of two possible antecedents presents greater processing difficulty than associating an anaphor with only one plausible antecedent, as indexed by an effect of anterior negativity in Van Berkum et al. (, ), we propose that the necessity of associating an anaphor with one available antecedent incurs a greater processing cost than when this association process is not required at all. Furthermore, since it was the condition in which there was a clear and available intrasentential antecedent for the pronoun that elicited this effect, the effect of anterior negativity that this comparison elicited cannot be interpreted as an index of processing referential ambiguity—there is nothing ambiguous about the anaphoric relationship in ().⁵ These results provide evidence that ambiguity is not a necessary condition for the elicitation of anterior negativity in referential processing contexts. We consider these results to at a minimum suggest that Nref and LAN responses may be cut from the same cloth.

.. The effect of reference form on retrieval operations A recent study of referential processing in Karimi, Swaab, and Ferreira () was able to relate anterior negativities elicited in referential contexts directly to memory retrieval operations, providing further support for the claims made in Barkley et al. (). This study was based on the results of a number of prior studies suggesting that antecedents that are more specific in reference are more deeply encoded in working memory and therefore retrieved more easily. For example, semantically elaborated clefted antecedent fillers (“It was a[n alleged Venezuelan] communist who the members of the club banned __ from ever entering the premises”) are initially read more

⁴ There was in fact a time window (–msec.) in which the effect was non-significant, indicating the potential presence of two independent effects that were confounded in the full –msec. analysis window. In reference to the anaphoric processing notions of bonding and resolution (Garrod and Terras ) discussed in §., we conjecture that the initial phasic LAN effect (–msec.) might index the bonding process, and subsequent sustained anterior negativity (–msec.) in the resolution stage. However, this obviously remains a topic for future research. ⁵ Note that if a pronoun for which no antecedent can be located in the same sentence () were to present processing difficulty, for example by forcing the reader to assume and introduce a novel referent into the discourse representation, then the negative brain response to the pronoun in () should have been larger than the response to the parallel pronoun in (), for which an antecedent was readily available; see the discussion of Nieuwland () in §... Instead, the opposite was the case.

OUP CORRECTED PROOF – FINAL, 17/1/2019, SPi

  



slowly (i.e., upon encoding), but result in faster reading times at the corresponding gap position (i.e., upon retrieval; Hofmeister ). Similarly, noun phrases modified by restrictive relative clauses (“The actor [who was frustrated and visibly upset about the night’s performance] walked away from the actress”) are more likely to result in subsequent pronominal reference in production studies (Karimi, Fukumura, Ferreira, and Pickering ), more looks to the modified referent in a visual world paradigm, and increased resolution of ambiguous pronouns in favor of the modified referent (Karimi and Ferreira ) than simple definite descriptions. Karimi, Swaab, and Ferreira () utilized this same design in an ERP study. In view of the retrieval facilitation afforded by referential specificity, the hypothesis was that pronouns with more referentially specific antecedents (i.e., modified by a relative clause) should elicit reduced amplitude anterior negativity compared to pronouns with simple definite descriptions as antecedents. By and large, this turned out to be the case, and Karimi et al. used this as evidence to argue that anterior negativities elicited in response to referential ambiguity are sensitive to the same factors that modulate ease of retrieval in other contexts, and therefore indices of working memory operations. There are some caveats related to this study, however, which at the time of this writing are still in progress. One is that the time course of the negativity observed in response to referential ambiguity was not consistent across the two experiments run in the study: in one experiment, the response was significant between – msec. and again between – msec., but only in the latter window in the second experiment. However, as we indicated in §., we are not convinced that differences in the timing of ERP effects alone are sufficient for drawing qualitative rather than mere quantitative distinctions between them. In other words, we do not believe that the timing discrepancy disqualifies this response from being treated as equivalent to an Nref for purposes of determining whether or not it is sensitive to retrieval operations, especially in view of the variability in the latency of the Nref itself (see §..). The other caveat is that the response to referential ambiguity was not robustly replicated in this study: there was no main effect of referential ambiguity when this was manipulated (a classic ‘Nref effect’), and the interactions of referential ambiguity with referential specificity were somewhat counterintuitive. The significance of these issues is that in order to claim that anterior negativities elicited by referential manipulations index retrieval operations related to working memory, these effects need to be replicated under the usual eliciting conditions of referential ambiguity. Whether we agree with the interpretations of negativities found in the literature on referential processing, the fact remains that they are reliably elicited in referentially ambiguous contexts. As such, the fact that no response to ambiguity is observed in this study is problematic. Hopefully this uncertainty will be cleared up in future research. However, none of this alters the fact that the effect elicited in this study was a difference in left anterior negativity (with roughly the same duration as ‘Nref effects’) that seemed to be sensitive to differences in ease of retrieval of antecedents from working memory.

OUP CORRECTED PROOF – FINAL, 17/1/2019, SPi



    

.. The processing of cataphoric relationships A cataphor is simply an anaphor that precedes its antecedent, thereby reversing the typical linear order of these two elements. Cataphoric relationships are thus interesting from a processing perspective because the challenges they pose for the comprehension system are also turned on their head. Recall from §. that the following were among the operations that we posited were required for the processing of anaphoric relations: () Identification and encoding of an antecedent that is out of its canonical position. () Association of any anaphor with its antecedent by retrieving the latter from working memory. The existence of cataphoric relationships upends this tidy scheme. While the antecedent in a cataphoric relationship still needs to be identified and encoded, there is no need for it to be subsequently retrieved. Because the cataphor comes first, it is this element that instead may need to be encoded prior to subsequent retrieval when the antecedent is encountered. This then raises the following empirical questions. First, is there electrophysiological evidence that a cataphor is encoded in working memory for subsequent retrieval? And second, is there electrophysiological evidence that is in fact retrieved at the position of the antecedent? These are the questions that the following studies were designed to address.

... Overt cataphors in English In addition to a direct comparison of syntactic and referential dependencies, Barkley et al. () also attempted to shed some light on cataphoric processing. The stimulus conditions designed to test this were as follows: ()

After a covert mission that deployed him for nine terrible months, Will longed for home.

()

After a covert mission that required deployment for nine terrible months, Will longed for home.

()

Will joined a mission that deployed him for nine terrible months, and longed for home.

The main clause subject antecedent Will in () is preceded by a cataphor while the same antecedent in (), its control, is not. This comparison was intended to elicit an electrophysiological response indicating that the cataphor is perhaps retrieved at the position of the antecedent Will. The anaphor him in () was designed to serve as the control for the cataphor him in () in order to test for electrophysiological evidence of encoding of the cataphor in working memory. What was surprising about the results we obtained from this experiment was the lack of expected effects. First, we expected that there would be some type of noticeable response to the cataphoric object pronoun him in the sentence-initial adjunct of

OUP CORRECTED PROOF – FINAL, 17/1/2019, SPi

  



()—and there actually may well have been. However, the control condition () that we created especially for this comparison was confounded in that the object pronoun him in () was itself preceded by an antecedent, which, as discussed in §.., was the major factor in eliciting a LAN effect in our experiment. In other words, since him in () was also preceded by an antecedent, it was likely to have elicited its own LAN response when the antecedent was retrieved. If, as we had predicted, a cataphoric pronoun should also elicit a LAN response when it is first encoded in working memory, this would account for the fact that when we compared the object pronouns him in () and (), there were no significant differences between the two in the LAN time window, or for that matter at any other point in the epoch. The second surprise was the lack of any effect whatsoever in response to the antecedent Will in (). We had predicted that even a proper noun preceded by a cataphor should elicit a LAN effect if the cataphor needs to be retrieved at the antecedent position in order to establish co-reference. This turned out not to be the case: there was no difference whatsoever in the response to the main clause subject Will whether preceded by a cataphor () or not (). We attributed this lack of response to the extreme position of proper names on accessibility (Ariel ) and Givenness Hierarchies (Gundel, Hedberg, and Zacharski ): proper names are referentially independent, preferentially introduce new referents, and are forward (rather than backward) looking in terms of discourse reference (see Silverstein , for further discussion). The possible advantage of this latter null effect is that it may help to shed further light on exactly what processing operations are indexed by LAN effects in referential contexts. While we reliably elicited a LAN effect in response to pronouns preceded by an antecedent (§..), it is not possible to determine whether that response (a) was an index of retrieval of the antecedent or (b) merely reflected the association of the second element of any anaphoric dependency with the first. The lack of a LAN difference at the proper noun antecedent in a cataphoric dependency, when compared to a proper noun that is in no such relationship, suggests that the LAN response cannot be tied to the mere association of any two elements in an anaphoric dependency: clearly the referential status of the second element appears to play a role in whether the second element elicits a LAN effect. Pronouns do, but proper names in the same anaphoric relationship and serial sentence position do not. This suggests that the LAN response is in fact an index of retrieval operations. We argued in Barkley et al. ()—independently of the line of reasoning pursued here— that our proper name comparison elicited no ERP differences because it did not trigger any retrieval operations. Specifically, we pointed out that in cue-based, contentaddressable models of working memory (Lewis et al. ; McElree et al. ), features of words entering the parse trigger the retrieval of associated items in working memory exhibiting similar features, based on the number of features that they share. Pronouns and gaps are dependent forms—they lack fully specified referential content—and as such carry features that need to be filled in along various parameters. Proper nouns carry fully specified features, and therefore require no retrieval to reach

OUP CORRECTED PROOF – FINAL, 17/1/2019, SPi



    

their full complement of referential specification. Thus our suggestion, based on the lack of a LAN effect in response to proper nouns preceded by cataphors in strictly referential dependencies, was that they simply do not trigger retrieval operations. If this is true, then it is possible to conclude that when LAN responses to second elements do occur, they are in fact indices of retrieval operations—the difficulty of which is modulated by factors such as referential specificity (Karimi et al. ; see §..).

... Overt cataphors in Dutch An investigation of cataphoric processing in terms of active search mechanisms was undertaken by Pablos, Doetjes, Ruijgrok, and Cheng (), whose Dutch stimulus materials were based on the English stimulus materials used in a reading time study by Kazanina, Lieberman, Yoshida, and Phillips (): “His/Her assistants found out that Lodewijk [m.] Boer had selected no prizewinner, but Mirjam [f.] had no interest in the gossip.” When critical masculine proper names like Lodewijk were preceded by feminine possessive pronouns, they elicited greater anterior negativity approximately – msec. post stimulus onset⁶ than did masculine proper names preceded by appropriately masculine possessive pronouns.⁷ This negativity had the classic scalp distribution and morphology of a LAN effect. In fact, however, based on earlier research on forward (i.e., antecedent–pronoun) anaphoric dependencies (Van Berkum et al. ; see §.), Pablos et al. () had predicted that this manipulation would lead to referential failure at a proper name mismatching in gender—the potential second element of a referential dependency— and therefore elicit late positivity instead. Recall from §.. that Nieuwland () also reported LAN-like effects in response to a pronoun mismatching in gender with the only previously available antecedent in the sentence (“The boy thought that [s]he . . . ”). In other words, there is a difference in LAN amplitude when a potential second element of a referential dependency is reached and no available referential element matches its features, regardless of whether the dependency is anaphoric or cataphoric. These results cast doubt on late positivity as a reliable marker of ‘referential failure,’ but do point to similarities in the processing of the second element in both cataphoric and anaphoric dependencies. Moreover, the fact that in both of these studies—as well as in response to the anaphoric pronouns in Barkley et al. (); ⁶ Due to extremely low statistical power—only nine trials per condition and only twenty-four participants—the exact window for finding statistically significant effects depended on the type of analysis undertaken. ⁷ No such LAN effect was elicited when the proper names were preceded by mismatching pronouns in nominative case, i.e., c-commanding cataphors (“He/She found out that Lodewijk [m.] Boer had selected no prizewinner, but Mirjam [f.] had no interest in the gossip.” In this instance, structural considerations apparently prevented the active search mechanism from attempting to associate the cataphor with the first possible antecedent. As in Kazanina et al. (), this was interpreted as demonstrating the immediate on-line influence of Principle C of the binding theory (Chomsky ) on anaphoric processing—one of the syntactic constraints in Callahan’s taxonomy of constraints on anaphors (see §.).

OUP CORRECTED PROOF – FINAL, 17/1/2019, SPi

  



§..—effects of anterior negativity were elicited in unambiguous referential contexts underscores our contention that ERP effects of this type should not be functionally tied to referential ambiguity. Our interpretation of these effects is that they have nothing to do with the gender mismatch per se. On this point, we agree with Pablos et al. () and their suggestion that “when the parser encounters the first potential antecedent position, it expects to find a matched antecedent” (Pablos et al. : ). It seems plausible to us that when this turns out not to be the case, the parser is forced to retrieve the cataphor from working memory to recheck its features, in order to confirm that this does not constitute an appropriate match. This results in a classic LAN effect. We would likewise suggest that when a pronoun is unexpectedly found to differ in gender from the only available antecedent preceding it in the sentence (Nieuwland ), the latter is retrieved for the same reason: to double check its features. But what happens when a cataphor does match the features of a following antecedent? Wouldn’t such a cataphor likewise need to be retrieved from working memory in order to confirm that its features do match those of its potential antecedent? Here, due to the paucity of currently available studies addressing this question, the answers are less clear. In our discussion in §... of straightforward, referentially unproblematic cataphor–antecedent relationships in English (Barkley et al. ), we concluded—based on the lack of effects at the antecedent position—that in such cases, retrieval of the cataphor from working memory may not be required because a proper name is referentially independent. We could therefore speculate that effects of anterior negativity at the antecedent position arise in cataphoric relationships only when some interpretational conflict necessitates retrieval of the cataphor, as appeared to be the case in the results of Pablos et al. () discussed above. However, it is unfortunately impossible to verify whether this conjecture holds true based on these data alone, as no comparisons were undertaken at the downstream matching antecedent Mirjam, where one might have been able to test this hypothesis— though even here such a comparison might also have been influenced by the presence of a prior mismatching antecedent. The only other possibility afforded by the design in Pablos et al. () would have been to add another stimulus condition in order to allow the direct comparison of proper names that either do or do not figure into a (straightforward) cataphoric relationship: for example, “His/The assistants found out that Lodewijk [m.] . . . ” This would help to determine whether retrieval of the cataphor is required in such cases or not. If so, Lodewijk following his should elicit a LAN effect relative to Lodewijk following the. In support of this idea, we note that the brain response to “His assistants found out that Lodewijk [m.] . . . ” has the exact same LAN morphology as the response to “Her assistants found out that Lodewijk [m.] . . . ” The two responses appear to differ solely in amplitude over the front of the head: namely, the LAN response to Lodewijk following her looks the same as that to Lodewijk following his, but is simply larger (Pablos et al. : , Figures  and ). As pointed out at the beginning of §.., the inventory of anaphoric processing operations outlined in §. needs to undergo revision in any case in order to

OUP CORRECTED PROOF – FINAL, 17/1/2019, SPi



    

accommodate cataphoric relations of this type. In particular, the antecedent can no longer be identified—let alone stored—until the end of the dependency is reached. Moreover, since the cataphor occurs first, it must be identified and presumably encoded in working memory, lacking as it is in referential features, so that its antecedent can be identified and the coreferential relationship can be formed. Pablos et al. (: ) suggest that “the parser prefers to start a search [for an appropriate antecedent] as soon as a [cataphoric] pronoun is encountered” as part of an active search mechanism. Unfortunately, however, the design that Pablos et al. () chose for their experiment precluded valid comparisons at the position of the sentence-initial cataphor that might have provided evidence for this suggestion. Recall that design issues also precluded a valid comparison at the cataphor position in Barkley et al. (); see §.... It is therefore impossible to know at this point whether there is any electrophysiological indication that such a search process is initiated in a backward referential dependency with an overt cataphor. Before trying to sort these issues out any further, we first turn to one final case study in yet another language, after which we conclude with a discussion of remaining outstanding issues.

... Null cataphors in East Asian languages The electrophysiological processing of null cataphora has to our knowledge been investigated exclusively in the context of head-final relative clauses in East Asian languages (Japanese: Ueno and Garnsey ; Korean: Kwon, Kluender, Polinsky, and Kutas ).⁸ In this regard, the gaps in head-final relative clauses present the same processing challenges as do overt cataphors: the association of a null cataphor with its antecedent needs to be delayed until a potential antecedent (i.e., the head noun) is detected. And just as with any gapped position, the relative clause gap must be identified by registering the absence of a required sentence constituent. Beyond this, however, there is further indeterminacy to some degree in Japanese, and especially in Korean relative clauses. Korean is extraordinarily tolerant of missing arguments in both subject (.%) and object position (.%; Y.-J. Kim ). Therefore, when a gap is identified by the absence of a constituent in a sentenceinitial Korean relative clause, there is virtually no way to determine the exact role that this null anaphor will play in the sentence. It could be any number of things: an ⁸ These studies compared object relative clauses to subject relative clauses. Japanese and Korean are both head-final, and the question of interest was whether subject or object relatives would be more difficult to process in these languages, as indexed by LAN effects at the end of the relative clause. At stake was the prediction of distance models of working memory—either linear (Gibson , ) or temporal (Lewis and Vasishth )—that subject relative clauses [[__i [OV]] headi] should tax working memory more than object relative clauses [[S[__i V]] headi] in SOV languages, and therefore be more difficult to process. This turned out not to be the case, however, suggesting that hierarchical distance (O’Grady ) is a more reliable metric here. Note that there is a long-standing debate among Korean syntacticians whether the gaps in Korean relative clauses are traces of movement versus bound pronominals; see Han () for a summary overview.

OUP CORRECTED PROOF – FINAL, 17/1/2019, SPi

  



argument of a main or embedded clause dropped purely for discourse reasons, a gap in a main or embedded clause created by scrambling, a gap in a relative clause, or even a gap in a noun complement clause. In fact, because Korean uses the same adnominal marker on the sentence-final verb at the right edge of both relative clauses and noun complement clauses, one cannot be entirely sure what kind of null anaphor this is until the head noun position itself is reached. As outlined in the introduction to §.., the electrophysiological processing questions at issue in any cataphoric relationship are whether there is evidence that a cataphor is encoded in working memory for subsequent retrieval, and whether it is in fact then retrieved at the antecedent position. As to the latter question, we can say the following with certainty: the end of a head-final relative clause elicits a LAN effect, at the relative clause-final verb in Japanese (Ueno and Garnsey ), and at the head noun position in Korean (Kwon et al. ).⁹ This raises the related question of what kind of processing operation this LAN effect might index. Does it index the retrieval of the null anaphor itself (to the extent that retrieving a null element is a plausible cognitive operation) or retrieval of the entire relative clause string, in order to identify the relative clause gap as such? Or does it merely reflect the association of the second element of any anaphoric dependency with the first (i.e., this head noun must be associated with that earlier gap position)? Unfortunately, the answer to this question can once again not be determined on the basis of the available evidence. However, our conjecture regarding Dutch cataphora in the previous section—namely that anterior negativity is elicited at the antecedent position of a cataphoric relationship only when some interpretational conflict necessitates retrieval of the cataphor—would seem to apply to Korean relative clauses as well, exhibiting as they do rather extensive interpretational ambiguity. This suggests that some form of retrieval may indeed be at play in these constructions. As to the first question of whether a null cataphor is ever encoded in working memory in the first place, definitive electrophysiological evidence is hard to come by, as the only studies involving null cataphora thus far have contrasted head-final subject and object relative clauses, in which both conditions contain a null cataphor.¹⁰ Therefore, unfortunately, we are once again unable to say whether the appearance of a (null) cataphor triggers encoding of the cataphor, or of its surrounding context, in working memory. This is unfortunate because it also impacts whether the clear LAN effect

⁹ The difference in the exact sentence position at which the LAN effect is elicited is likely due to typological differences between Japanese and Korean. Japanese does not morphologically mark the right edge of the relative clause on the relative clause-final verb, while Korean does, and Korean drops object arguments more liberally than does Japanese, creating greater indeterminacy in the parse. As we said above in the main text, one cannot know exactly what kind of null cataphor one is dealing with in a sentence-initial Korean relative clause until the head noun position is reached. ¹⁰ There could still be a processing difference between the encoding of a cataphoric object gap and a cataphoric subject gap, of course. Kwon et al. () reported a difference of anterior negativity in response to object relatives in the relative clause region, but attributed it to mere strategic processing caused by the design of the experiment.

OUP CORRECTED PROOF – FINAL, 17/1/2019, SPi



    

elicited at the head noun in Korean (or relative clause-final verb in Japanese) is indeed an index of retrieval, as opposed to some more general look-back function needed to associate the second element in an anaphoric dependency with whatever the previously occurring first element was.

... Putting it all together Obviously, cataphoric processing constitutes a major obstacle to broad empirical coverage of the cognitive operations that we laid out in §., repeated below for convenience. We now briefly consider the challenges posed by the above studies of cataphoric processing for these operations. () Identification and encoding of an antecedent that is out of its canonical position. () Identification of a null anaphor by registering the absence of a required sentence constituent. () Association of any anaphor with its antecedent by retrieving the latter from working memory. () Integration of the referent indexed by the two reference forms into the sentence context and ongoing discourse representation. With regard to operation (), the problem is that the cataphor needs to be identified first—and encoded in working memory—while the antecedent only needs to be identified subsequently. This is, in other words, the mirror image of an anaphoric configuration. With regard to the identification and encoding of a cataphor, the methodological problem thus far has been that no design has allowed for the relevant comparison to be made (Kwon et al. ; Barkley et al. ; Pablos et al. ). The further theoretical complication is that what exactly would be encoded at the position of a null cataphor is unclear, particularly when the language in question allows extensive ambiguity, as is the case in Korean (Kwon et al. ). Thus, while it seems as if operation () should apply equally to anaphoric and cataphoric configurations without difficulty, there is the added complication that the absence of a required sentence constituent with no preceding antecedent does not automatically signal the presence of a null cataphor. As we saw in §..., in a language with a high tolerance for dropped arguments, the cataphor may not be identified as such until much later when the antecedent itself is encountered (Kwon et al. ). With regard to operation (), the issue is that, while we have ample evidence that an anaphor triggers retrieval of its antecedent, as indexed by an effect of left anterior negativity (Barkley et al. ), the electrophysiological evidence as to whether an antecedent triggers retrieval of an associated preceding cataphor is equivocal. Based on the electrophysiological studies of cataphoric processing currently available, we have tentatively concluded that when interpretational conflict or indeterminacy is involved, there appears to be evidence of active retrieval of the cataphor in the form of a LAN effect (Kwon et al. ; Pablos et al. ), but not when the relationship of cataphor to antecedent is straightforward (Barkley et al. ).

OUP CORRECTED PROOF – FINAL, 17/1/2019, SPi

  



What about operation ()? We have not really addressed the question of what electrophysiological response might be its index since we questioned received wisdom in §. that late positivity in response to gap positions is an index of syntactic integration. There we suggested that late positivity might rather be an index of identification, that is, operation (). On the basis of our review of the anaphoric processing literature, however, we can now say at a minimum that late positivity is unlikely to be the true index of (), as it has been reliably reported in only one specific context, namely in response to anaphors immediately preceded by two antecedents mismatching in gender, as in “Anna shot at Linda as he jumped over the fence” (Van Berkum et al. ; see §..). When there is only one antecedent that mismatches the anaphor (or cataphor) in gender, the more consistent response is anterior negativity (Nieuwland ; Pablos et al. ; see §.. and §...). In sum, most of the outstanding issues in the electrophysiological literature on anaphoric processing have to do with the encoding of the first element in an anaphoric dependency. This is hardly surprising, given that purely anaphoric dependencies, unlike syntactic dependencies, are asymmetric in nature: one element in an anaphoric dependency, and most typically the first one, is referentially independent, while the other is completely dependent on the first for its reference. With regard to storage, while the jury is still out on this issue, the available electrophysiological evidence seems inconsistent with an active storage mechanism, at least for linguistic input. However, the broad outlines of how the second element in an anaphoric relationship is processed seem remarkably consistent across studies. As we have argued throughout, the second element triggers a LAN effect when it unambiguously signals that retrieval of the first element is necessary. This operation appears to be reliably modulated by retrieval difficulty, which can be influenced by factors such as referential specificity (§..) and the number of candidate antecedents in the discourse (§.). In any case, we feel that these operations, first laid out clearly by Callahan (), however much they may require further modification going forward, still continue to provide a useful framework for understanding the operations involved in anaphoric processing and concomitant brain signatures.

. C

.................................................................................................................................. At the beginning of this chapter, we posed the following question: from the brain’s perspective, is there anything special about processing referential anaphora? Based on our review and interpretation of the ERP literature on anaphoric relationships, our answer is no. This is perhaps ultimately not that surprising, as the processing challenges posed by long-distance referential relations are no different than those posed by the multitude of other types of long-distance anaphoric dependencies found across languages. We proposed a broad inventory of the cognitive operations necessary for processing these types of dependencies, argued that the same processes appear to be

OUP CORRECTED PROOF – FINAL, 17/1/2019, SPi



    

required regardless of whether an anaphoric relationship is syntactic or referential in nature, and linked specific brain responses to each operation. We then showed that, by committing to a working memory-based interpretation of the brain’s response to referential ambiguity, even these responses can be mapped onto this general cognitive classification scheme. However, we simultaneously noted that the processing of cataphoric relationships poses a number of challenges to this scheme that will require further research in order to be resolved. Our understanding of the cognitive underpinnings of anaphoric processing thus continues to evolve.

OUP CORRECTED PROOF – FINAL, 21/1/2019, SPi

        ......................................................................................................................

     An updated survey ......................................................................................................................

     

I , in an article in the journal Computational Linguistics, we summarized the state of the art of a research area known as Referring Expressions Generation ().¹ Our survey introduced the  problem, described a range of computational approaches to the problem, and discussed how  algorithms have been experimentally tested. Despite the recency of our survey, a substantial amount of work has come to light in the intervening years, broaching new avenues of research, involving new methods, new collaborations, and new research questions; let us briefly hint at what we see as some of the recent highlights. One new method in  is the use of Bayesian methods, which are successfully reviving the old idea that speakers’ choice of referring expressions may be driven primarily by rationality (Frank and Goodman ; Qing and Franke ; Franke and Degen ); this line of work has attracted considerable attention from philosophers and formal semanticists. In related developments, more and more  algorithms are informed by Machine Learning, with notable additions to the literature such as FitzGerald, Artzi, and Zettlemoyer () and Garoufi and Koller (). An example of a new interdisciplinary collaboration is the increasing interaction between researchers in vision and language (e.g., Mitchell et al. ), which has included attempts to model the influence of visual perception on the production of referring expressions (e.g., Kazemzadeh, Ordonez, Matten, and Berg ), giving rise to new practical applications. Also notable is the intensity with which computational researchers in this area have started to collaborate with psychologists and other ¹ The main body of this chapter is a reprint of Krahmer and van Deemter () and appears here with permission from MIT Press.

OUP CORRECTED PROOF – FINAL, 21/1/2019, SPi



     

cognitive scientists, and the associated tendency to view, and to test,  algorithms as “computational models of human referring” (van Deemter ), using these algorithms as a way to express insights into human language use. In some of this work,  researchers have turned to new phenomena and new research questions. For example, substantial efforts are now being made to model the ways in which speakers and hearers collaborate to ensure together that they attend to the same referent (e.g., Garoufi, Staudte, Koller, and Crocker ), a phenomenon that has long been known to psychologists (e.g., Clark and Wilkes-Gibbs ), but that had attracted relatively little attention from theoretical and computational linguists until recently. The survey below outlines the foundations on which these and other recent developments have built. We hope that it will offer readers a basis that will help them understand and appreciate these developments.

. I

.................................................................................................................................. Suppose one wants to point out a person in the scene in Figure . to an addressee. Most speakers have no difficulty in accomplishing this task, by producing a referring expression such as ‘the man in a suit’, for example. Now imagine a computer being

 . A simple visual scene. Photograph by Harold Miesen

OUP CORRECTED PROOF – FINAL, 21/1/2019, SPi

    



confronted with the same task, aiming to point out individual d₁. Assuming it has access to a database containing all the relevant properties of the people in the scene, it needs to find some combination of properties which applies to d₁, and not to the other two. There is a choice though: there are many ways in which d₁ can be set apart from the rest (‘the man on the left’, ‘the man with the glasses’, ‘the man with the tie’), and the computer has to decide which of these is optimal in the given context. Moreover, optimality can mean different things. It might be thought, for instance, that references are optimal when they are minimal in length, containing just enough information to single out the target. But, as we shall see, finding minimal references is computationally expensive, and it is not necessarily what speakers do, nor what is most useful to hearers. So, what is Referring Expression Generation? Referring expressions play a central role in communication, and have been studied extensively in many branches of (computational) linguistics, including Natural Language Generation (NLG). NLG is concerned with the process of automatically converting non-linguistic information (e.g., from a database) into natural language text, which is useful for practical applications ranging from generating weather forecasts to summarizing medical information (Reiter and Dale ). Of all the subtasks of NLG, Referring Expression Generation () is among the ones that have received most scholarly attention. A survey of implemented, practical NLG systems shows that virtually all of them, regardless of their purpose, contain a  module of some sort (Mellish et al. ). This is hardly surprising in view of the central role that reference plays in communication. A system providing advice about air travel (White, Clark, and Moore ), needs to refer to flights (‘the cheapest flight’, ‘the KLM direct flight’); a Pollen forecast system (Turner, Sripada, Reiter, and Davy ) needs to generate spatial descriptions for areas with low or high pollen levels (‘the central belt and further North’), and a robot dialogue system that assembles construction toys together with a human user (Giuliani et al. ), needs to refer to the components (‘insert the green bolt through the end of this red cube’).  “is concerned with how we produce a description of an entity that enables the hearer to identify that entity in a given context” (Reiter and Dale : ). Since this can often be done in many different ways, a  algorithm needs to make a number of choices. According to Reiter and Dale (), the first choice concerns what form of referring expression is to be used; should the target be referred to, for instance, using its proper name, a pronoun (‘he’) or a description (‘the man with the tie’). Proper names have limited applicability because many domain objects do not have a name that is in common usage. For pronoun generation, a simple but conservative rule is discussed by Reiter and Dale (), similar to one proposed by (Dale : –): use a pronoun if the target was mentioned in the previous sentence, and if this sentence contained no reference to any other entity of the same gender. Reiter and Dale () concentrate mostly on the generation of descriptions. If the NLG system decides to generate a description, two choices need to be made: which set of properties distinguishes the target (content selection), and how can the selected properties be turned into natural language (linguistic realization). Content selection is a complex balancing act: we need to say enough to enable identification of the intended referent, but not too much.

OUP CORRECTED PROOF – FINAL, 21/1/2019, SPi



     

A selection of information needs to be made, and this needs to be done quickly. Reiter and Dale discuss various strategies that try to manage this balancing act, based on Dale and Reiter (), an early survey article that summarizes and compares various influential algorithms for the generation of descriptions. Why a survey on , and how to read it? , like NLG in general, has changed considerably after the overviews presented in Dale and Reiter () and Reiter and Dale (), owing largely to an increased use of empirical data, and a widening of the class of referring expressions studied. Moreover, a gradual shift has taken place towards extended application domains, different input and output formats, and more flexible interactions with the user, and this shift is starting to necessitate the use of new  techniques. Examples include recent systems in areas such as weather forecasting (Turner, Sripada, and Reiter ) and medical care (Portet et al. ) where complex references to spatial regions and time periods abound. The results of recent  research are scattered over proceedings, books, and journals. The current survey offers a compact overview of the progress in this area and an assessment of the state of the art. The concept of reference is difficult to pin down exactly (Searle ; Abbott ). Searle therefore suggests that the proper approach is “to examine those cases which constitute the center of variation of the concept of referring and then examine the borderline cases in light of similarities and differences from the paradigms” (Searle : –). The “paradigms” of reference in Reiter and Dale () are definite descriptions whose primary purpose is to identify their referent. The vast majority of recent  research also subscribes to this view. Accordingly, these paradigmatic cases will also be the main focus of this survey, although we shall often have occasion to discuss other types of expressions. However, to do full justice to indefinite or attributive descriptions, proper names, and personal pronouns would, in our view, require a separate, additional survey. In §. we offer a brief overview of  research up to , discussing some classic algorithms. Next, we zoom in on the new directions in which recent work has taken  research: extension of the coverage of algorithms, to include, for example, vague, relational, and plural descriptions (§.), exploration of different computational frameworks, such as Graph Theory and Description Logic (§.) and collection of data and evaluation of  algorithms (§.). §. highlights open questions and avenues for future work. §. summarizes our findings.

. A     -  

.................................................................................................................................. The current survey focuses primarily on the progress in  research in the twenty-first century, but it is important to have a basic insight into pre-  research and how it laid the foundation for much of the current work.

OUP CORRECTED PROOF – FINAL, 21/1/2019, SPi

    



.. First beginnings  can be traced back to the earliest days of Natural Language Processing; Winograd (: §.., Naming Objects and Events), for example, sketches a primitive “incremental”  algorithm, used in his SHRDLU program. In the s, researchers such as Appelt and Kronfeld set themselves the ambitious task of modeling the human capacity for producing and understanding referring expressions in programs such as KAMP and BERTRAND (Appelt ; Appelt and Kronfeld ; Kronfeld ). They argued that referring expressions should be studied as part of a larger speech act. KAMP (Appelt ), for example, was conceived as a general utterance planning system, building on Cohen and Levesque’s () formal speech act theory. It used logical axioms and a theorem prover to simulate an agent planning instructions such as ‘use the wheelpuller to remove the flywheel, which contains two referring expressions, as part of a larger utterance. Like many of their contemporaries, Appelt and Kronfeld’s hoped to gain insight into the complexities of human communication. Doug Appelt (p.c.): “[ . . . ] the research themes that originally motivated our work on generation were the outgrowth of the methodology in both linguistics and computational linguistics at the time that research progress was best made by investigating hard, anomalous cases that pose difficulties for conventional accounts.” Their broad focus allowed these researchers to recognize that although referring expressions may have identification of the referent as their main goal, a referring expression can also add information about a target. By pointing to a tool on a table, while saying ‘the wheelpuller’, the descriptive content of the referring expression may serve to inform the hearer about the function of the tool (Appelt and Kronfeld ). They also observed that referring expressions need to be sensitive to the communicative context in which they are used and that they should be consistent with the Gricean maxims (see §..), which militate against overly elaborate referring expressions (Appelt ). It is remarkably difficult, after twenty years, to find out how these programs actually worked, since code was lost and much of what was written about them is pitched at a high level of abstraction. Appelt and Kronfeld were primarily interested in difficult questions about human communication, but they were sometimes tantalizingly brief about humbler matters. Here, for instance, is how Appelt (: ) explains how KAMP would attempt to identify a referent: “KAMP chooses a set of basic descriptors when planning a describe action to minimize both the number of descriptors chosen, and the amount of effort required to plan the description. Choosing a provably minimal description requires an inordinate amount of effort and contributes nothing to the success of the action. KAMP chooses a set of descriptors by first choosing a basic category descriptor (see [Rosch ]) for the intended concept, and then adding descriptors from those facts about the object that are mutually known by the speaker and the hearer, subject to the constraint that they are all linguistically realizable in the current noun

OUP CORRECTED PROOF – FINAL, 21/1/2019, SPi



     

phrase, until the concept has been uniquely identified. [ . . . ] Some psychological evidence suggests the validity of the minimal description strategy; however, one does not have to examine very many dialogues to find counter-examples to the hypothesis that people always produce minimal descriptions.”

This quote contains the seeds of much later work in , given its skepticism about the naturalness of minimal descriptions, its use of Rosch ()-style basic categories, and its acknowledgment of the role of computational complexity. Broadly speaking, it suggests an incremental generation strategy, compatible with the ones described below, although it is uncertain what exactly was implemented. In recent years, the Appelt–Kronfeld line of research has largely given way to a new research tradition which focused away from the full complexity of human communication, with notable exceptions such as Heeman and Hirst (); Stone and Webber (); O’Donnell, Cheng, and Hitzeman (); and Koller and Stone ().

.. Generating distinguishing descriptions In the early nineties, a new approach to  started gaining currency, when Dale and Reiter re-focused on the problem of determining what properties a referring expression should use if identification of the referent is the central goal (Dale , ; Reiter ; Reiter and Dale ). This line of work culminated in the seminal Dale and Reiter (). Like Appelt (), Dale and Reiter are concerned with the link between the Gricean maxims and the generation of referring expressions. They discuss the following pair of examples: ()

Sit by the table.

()

Sit by the brown wooden table.

In a situation where there is only one table, which happens to be brown and wooden, both the descriptions in () and () would successfully refer to their target. However, if you hear () you might make the additional inference that it is significant to know that the table is brown and wooden; why else would the speaker mention these properties? If the speaker merely wanted to refer to the table, your inference would be an (incorrect) ‘conversational implicature’, caused by the speaker’s violation of Grice’s (: ) Maxim of Quantity (“Do not make your contribution more informative than is required.”). Dale and Reiter () ask how we can efficiently compute which properties to include in a description, such that it successfully identifies the target while not triggering false conversational implicatures. For this, they zoom in on a relatively straightforward problem definition, and compare a number of concise, welldefined algorithms solving the problem. Problem definition. Dale and Reiter () formulate the  problem as follows. Assume we have a finite domain D of objects with attributes A. In our example scene

OUP CORRECTED PROOF – FINAL, 21/1/2019, SPi

    



Table . Tabular representation of some information in our example scene Object

type

clothing

position

d₁ d₂ d₃

man woman man

wearing suit wearing t-shirt wearing t-shirt

left middle right

(Figure .), D = {d₁, d₂, d₃} and A = {type, clothing, position, . . . }. The type attribute has a special status in Dale and Reiter () since it represents the semantic content of the head noun. Alternatively, we could have defined an attribute gender, stating that it should be realized as the head noun of a description. Typically, domains are represented in a knowledge base such as Table ., where different values are clustered together because they are associated with the same attribute. Left, right, and middle, for example, belong to the attribute position, and are said to be three values that this attribute can take. The objects of which a given attribute–value combination (or ‘property’) is true are said to form its denotation. Sometimes we will drop the attribute, writing man, rather than ⟨type, man⟩, for instance. The  task is now defined by Dale and Reiter () through what may be called identification of the target: given a target (or referent) object r ∈ D, find a set of attribute–value pairs L whose conjunction is true of the target but not of any of the distractors (i.e., D
{r}, the domain objects different from the target). L is called a distinguishing description of the target. In our simple example, suppose that d₁ is the target (and hence {d₂, d₃} the set of distractors), then L could, for example, be either {⟨type, man⟩,⟨clothing, wearing suit⟩} or {⟨type, man⟩,⟨position, left⟩}, which could be realized as ‘the man wearing a suit’ or ‘the man to the left’. If identification were all that counted, a simple, fast, and fault-proof  strategy would be to conjoin all the properties of the referent: this conjunction will identify the referent if it can be identified at all. In practice, Dale and Reiter—and others in their wake—include an additional constraint which is often left implicit: that the referring expressions generated should be as similar to human-produced ones as possible. In §.., §.. (evaluation sections), and §. (conclusion section), we return to this ‘human-likeness’ constraint (and to variations on the same theme). Full Brevity and Greedy Heuristic. Dale and Reiter () discuss various algorithms which solve the  task. One of these is the   algorithm (Dale ), which deals with the problem of avoiding false conversational implicatures in a radical way, by always generating the shortest possible distinguishing description. Originally, the   algorithm was meant to generate both initial and subsequent descriptions, by relying on a previous step that determines the distractor set based on which objects are currently salient. Given this set, it first checks whether there is a

OUP CORRECTED PROOF – FINAL, 21/1/2019, SPi



     

single property of the target that rules out all distractors. If this fails, it considers all possible combinations of two properties, and so on: . Look for a description L that distinguishes target r using one property. If success then return L. Else go to . . Look for a description L that distinguishes target r using two properties. If success then return L. Else go to . . Etcetera Unfortunately, there are two problems with this approach. First, the problem of finding a shortest distinguishing description has a high complexity—it is NP hard (see e.g., Garey and Johnson )—and hence is computationally very expensive, making it prohibitively slow for large domains and descriptions. Second, Dale and Reiter note that human speakers routinely produce descriptions that are not minimal. This is confirmed by a substantial body of psycholinguistic research (Olson ; Sonnenschein ; Pechmann ; Engelhardt, Bailey, and Ferreira ). An approximation of   is the   algorithm (Dale , ), which iteratively selects the property that rules out most of the distractors not previously ruled out, incrementally augmenting the description based on what property has most discriminatory power at each stage (as a result, it does not always generate descriptions of minimal size). The   algorithm is a more efficient algorithm than the   one, but it was soon eclipsed by another algorithm (Reiter and Dale ; Dale and Reiter ), which turned out to be the most influential algorithm of the pre- era. It is this later algorithm that came to be known as ‘the’ Incremental Algorithm (IA). The Incremental Algorithm. The basic idea underlying the IA is that speakers ‘prefer’ certain properties over others when referring to objects, an intuition supported by the experimental work of, for instance, Pechmann (). Suppose you want to refer to a person ten meters away from you. You might mention the person’s gender. If this is insufficient to single out the referent, you might be more likely to make use of the color of the person’s coat than to the color of her eyes. Less preferred attributes, such as eye color, are only considered if other attributes do not suffice. It is this intuition of a preference order between attributes that the IA exploits. By making this order a parameter of the algorithm, a distinction can be made between domain/genre dependent knowledge (the preferences), and a domain-independent search strategy. As in the   algorithm, descriptions are constructed incrementally; but unlike the  , the IA checks attributes in a fixed order. By grouping properties into attributes, Dale and Reiter predict that all values of a given attribute have the same preference order. Ordering attributes rather than values, may be disadvantageous, however. A simple shape (e.g., a circle), or a size that is unusual for its target (e.g., a tiny whale) might be preferred over a subtle color (purplish grey). Also, some values of a given attribute might be difficult to express, and ‘dispreferred’ for this reason (kind of like a UFO shape with a christmas tree sticking out the side).

OUP CORRECTED PROOF – FINAL, 21/1/2019, SPi

     1. 2. 3. 4. 5. 6. 7. 8. 9. 10. 11. 12. 13.



Incremental Algorithm ({r}, D, Pref ) { L←0 C ← D – {r} for each Ai in list Pref do V = Value (r, Ai) if C ∩ RulesOut (〈Ai,V〉) ≠ 0 then L ← L ∪ {〈Ai,V〉} C ← C – RulesOut (〈Ai,V〉) endif if C ≠ 0 then return L endif return failure }

 . Sketch of the core Incremental Algorithm.

Figure . contains a sketch of the IA in pseudo code. It takes as input a target object r, a domain D, consisting of a collection of domain objects, and a domainspecific list of preferred attributes Pref (). Suppose we apply the IA to d₁ of our example scene, assuming that Pref = type > clothing > position. The description L is initialized as the empty set (), and the context set C of distractors (from which d₁ needs to be distinguished) is initialized as D–{d₁} (). The algorithm then iterates through the list of preferred attributes (), for each one looking up the target’s value (), and checking whether this attribute–value pair rules out any of the distractors not ruled out so far (). The function RulesOut (⟨Ai, V⟩) returns the set of objects which have a different value for attribute Ai than the target object has. If one or more distractors are ruled out, the attribute–value pair ⟨Ai, V⟩ is added to the description under construction () and a new set of distractors is computed (). The first attribute to be considered is type, for which d₁ has the value man. This would rule out d₂, the only woman in our domain, and hence the attribute–value pair ⟨type, man⟩ is added to L. The new set of distractors is C = {d₃}, and the next attribute (clothing) is tried. Our target is wearing suit, and the remaining distractor is not, so the attribute–value pair ⟨clothing, wearing suit⟩ is also included. At this point all distractors are ruled out (), a set of properties has been found which uniquely characterize the target {⟨type, man⟩,⟨clothing, wearing suit⟩} (‘the man wearing a suit’), and we are done (). If we had reached the end of the list without ruling out all distractors, the algorithm would have failed (): no distinguishing description for our target was found. The sketch in Figure . simplifies the original algorithm in a number of respects. First, Dale and Reiter always include the type attribute, even if it does not rule out any distractors, because speakers use type information in virtually all their descriptions. Second, the original algorithm checks, via a function called UserKnows, whether a given property is in the common ground, to prevent the selection of properties which the addressee might not understand. Unlike Appelt and Kronfeld, who discuss detailed examples that hinge on differences in common ground, Dale and Reiter () treat UserKnows as a function that returns ‘true’ for each true proposition, assuming that all relevant information is shared. Third, the IA can take some ontological information into account via subsumption hierarchies. For instance, in a dog-and-cat domain, a pet

OUP CORRECTED PROOF – FINAL, 21/1/2019, SPi



     

may be of the chihuahua type, but chihuahua is subsumed by dog, and dog in turn is subsumed by animal. A special value in such a subsumption hierarchy is reserved for the so-called basic level values (Rosch ); dog in this example. If an attribute comes with a subsumption hierarchy, the IA first computes the best value for that attribute, which is defined as the value closest to the basic level value, such that there is no more specific value that rules out more distractors. In other words, the IA favors dog over chihuahua, unless the latter rules out more distractors. The IA is conceptually straightforward and easy to implement. In addition, it is computationally efficient, with polynomial complexity: its worst-case run time is a constant function of the total number of attribute–value combinations available. This computational efficiency is due to the fact that the algorithm does not perform backtracking: once a property has been selected, it is included in the final referring expression, even if later additions render it superfluous. As a result, the final description may contain redundant properties. Far from seeing this as a weakness, Dale and Reiter (: ) point out that this makes the IA less “psycholinguistically implausible” than its competitors. It is interesting to observe that while Dale and Reiter () discuss the theoretical complexity of the various algorithms in detail, later research has tended to attach more importance to empirical evaluation of the generated expressions (§.).

.. Discussion Appelt and Kronfeld’s work, founded on the assumption that  should be seen as part of a comprehensive model of communication, started to lose some of its appeal in the early nineties, because it was at odds with the emerging research ethos in computational linguistics that stressed simple, well-defined problems allowing for measurable results. The way current  systems are shaped is largely due to developments summarized in Dale and Reiter (), which focuses on a specific aspect of , namely determining which properties serve to identify some target referent. Dale and Reiter’s work aimed for generating human-like descriptions, but was not yet coupled with systematic investigation of data.  as search. The algorithms discussed by Dale and Reiter () can be seen as different instantiations of a general search algorithm (Bohnet and Dale ; Gatt ). They all basically search through the same space of states, each consisting of three components: a description that is true of the target, a set of distractors, and a set of properties of the target that have not yet been considered. The initial state can be formalized as the triple ⟨Ø, C, P⟩ (no description for the target has been constructed, no distractors have been ruled out, and all properties P of the target are still available), and the goal state as ⟨L, Ø, P′⟩, for certain L and P′: a description L has been found, which is distinguishing—the set of distractors is empty. All other states in the search space are intermediate ones, through which an algorithm might move depending on its search strategy. For instance, when searching for a distinguishing description for d₁ in our

OUP CORRECTED PROOF – FINAL, 21/1/2019, SPi

    



example domain, an intermediate state could be s = ⟨{⟨type, man⟩}, {d₃}, {⟨clothing, wearing suit⟩, ⟨position, left⟩}⟩. The algorithms discussed earlier differ in terms of their so-called expand method, determining how new states are created, and their queue method, which determines the order in which these states are visited (i.e., how states are inserted into the queue). Full Brevity, for example, uses an expand method that creates a new state for each attribute of the target not checked before (as long as it rules out at least one distractor). Starting from the initial state and applied to our example domain, this expand method would result in three new states, creating descriptions including type, clothing, and position information respectively. These states would be checked using a queue method which is breadth-first. The IA, by contrast, uses a different expand method, each time creating a single new state in accordance with the pre-determined preference order. Thus, in the initial state, and assuming (as before) that type is the most preferred attribute, the expand method would create a single new state: s above. Since there is always only one new state, the queue method is trivial. Limitations of pre- . In the IA and related algorithms, the focus is on efficiently computing which properties to use in a distinguishing description. However, these algorithms rest on a number of implicit simplifications of the  task. () The target is always just one object, not a larger set (hence, plural noun phrases are not generated). () The algorithms all assume a very simple kind of knowledge representation, consisting of a set of atomic propositions. Negated propositions are only represented indirectly, via the Closed World Assumption, so an atomic proposition that is not explicitly listed in the database is false. () Properties are always ‘crisp’, never vague. Vague properties such as small and large are treated as Boolean properties, which do not allow borderline cases and which keep the same denotation, regardless of the context in which they are used. () All objects in the domain are assumed to be equally salient, which implies that all distractors have to be removed, even those having a very low salience. () The full  task includes first determining which properties to include in a description, and then providing a surface realization in natural language of the selected properties. The second stage is not discussed, nor is the relation with the first. A substantial part of recent  research is dedicated to lifting one or more of these simplifying assumptions, although other limitations are still firmly in place (as we shall discuss in §.).

. E  

..................................................................................................................................

.. Reference to sets Until recently,  algorithms aimed to produce references to a single object. But references to sets are ubiquitous in most text genres. In simple cases, it takes only a slight modification to allow classic  algorithms to refer to sets. The IA, for example,

OUP CORRECTED PROOF – FINAL, 21/1/2019, SPi



     

can be seen as referring to the singleton set {r} that contains the target r and nothing else. If in line  (Figure .), {r} is replaced by an arbitrary set S, and line  is modified as saying C ← D – S instead of C ← D – {r}, then the algorithm produces a description that applies to all elements of S. Thus, it is easy enough to let these algorithms produce expressions such as ‘the men’ or ‘the t-shirt wearers’, to identify {d₁, d₃} and {d₂, d₃} respectively. Unfortunately, things are not always so simple. What if we need to refer to the set {d₁, d₂}? Based on the properties in Table . alone this is not possible, because d₁ and d₂ have no properties in common. The natural solution is to treat the target set as the union of two smaller sets, {d₁} ∪ {d₂}, and refer to both sets separately (e.g., ‘the man who wears a suit, and the woman’). Once unions are used, it becomes natural to also allow set complementation, as in ‘the people who are not on the right’. Note that set complementation may also be useful for single referents. Consider a situation where all cats except one are owned by Mary, while the owner of the remaining one is unknown or non-existent. Complementation allows one to refer to ‘the cat not owned by Mary’. We shall call the resulting descriptions Boolean. As part of a more general logical analysis of the IA, van Deemter () made a first stab at producing Boolean descriptions, using a two-stage algorithm whose first stage is a generalization of the IA, and whose second stage involves the optimization of the possibly lengthy expressions produced by the first phase. The resulting algorithm is logically complete in the following sense: if a given set can be described at all using the properties available then this algorithm will find such a description. With intersection as the only way to combine properties,  cannot achieve logical completeness. The first stage of the algorithm starts by conjoining properties (man, left) (omitting attributes for the sake of readability) in the familiar manner of the IA; if this does not suffice for singling out the target set then the same incremental process continues with unions of two properties (e.g., man ∪woman, middle ∪left; that is, properties expressing that a referent is a man or a woman, in the middle or on the left), then with unions of three properties (e.g., man ∪wearing suit ∪woman), and so on. The algorithm terminates when the referent (individual or set) is identified (success) or when all combinations of properties have been considered (failure). Figure . depicts this in schematic form, where n represents the total number of properties in the domain, and P+/– denotes the set of all literals (atomic properties such as man, and their complements ¬man). Step () 1. [Length 1.] Run IA using all properties of the form P+/– If success then return L else go to (2) to add new properties to L. 2. [Length 2.] Run IA using all properties of the form P+/– ∪ P+/– If success then return L else go to (3) to add new properties to L. 3. [Length 3.] Run IA using all properties of the form P+/– ∪ P+/– ∪ P+/– If success then return L else go to (4) to add new properties to L. 4. Etcetera, up to unions of length n.  . Outline of the first stage of van Deemter’s () Boolean  algorithm.

OUP CORRECTED PROOF – FINAL, 21/1/2019, SPi

    



generalizes the original IA allowing for negated properties and target sets. As before, L is the description under construction. It will consist of intersections of unions of literals such as (woman ∪ man) ∩ (woman ∪ ¬wearing suit) (in other words, L is in Conjunctive Normal Form, CNF). Note that this first stage is not only incremental at each of its n steps, but also as a whole: once a property has been added to the description, later steps will not remove it. This can lead to redundancies, even more than in the original IA. The second stage removes the most blatant of these, but only where the redundancy exists as a matter of logic, rather than world knowledge. Suppose, for example, that step  selects the properties P ∪ S and P ∪ R, ruling out all distractors. L now takes the form (P ∪ S) ∩ (P ∪ R) (e.g., ‘the things that are both (women or men) and (women or wearing suits)’). The second phase uses logic optimization techniques, originally designed for the minimization of digital circuits (McCluskey ), to simplify this to P ∪ (S ∩ R) (‘the women, and the men wearing suits’). Variations and extensions. Gardent () drew attention to situations where this proposal produces unacceptably lengthy descriptions; suppose, for example, the algorithm accumulates numerous properties during steps  and , before finding one complex property (a union of three properties) during step  which, on its own would have sufficed to identify the referent. This will make the description generated much lengthier than necessary, because the properties from steps  and  are now superfluous. Gardent’s take on this problem amounts to a reinstatement of Full Brevity embedded in a reformulation of  as a constraint satisfaction problem (see §..). The existence of fast implementations for constraint satisfaction alleviates the problems with computational tractability to a considerable extent. But by re-instating  , algorithms like Gardent’s could run into the empirical problems noted by Dale and Reiter, given that human speakers frequently produce non-minimal descriptions (see Gatt ) for evidence pertaining to plurals. Horacek () makes a case for descriptions in Disjunctive Normal Form (DNF; unions of intersections of literals). Horacek’s algorithm first generates descriptions in CNF, then convert these into DNF, skipping superfluous disjuncts. Consider our example domain (Table .). To refer to {d₁, d₂}, a CNF-oriented algorithm might generate (man ∪ woman) ∩ (left ∪ middle) (‘the people who are on the left or middle’). Horacek converts this, first, into DNF: (man ∩ left) ∪ (woman ∩ middle) ∪ (man ∩ middle) ∪ (woman ∩ left), after which the last two disjuncts are dropped, because there are no men in the middle, and no women on the left. The outcome could be worded as ‘the man on the left and the woman in the middle’. Later work has tended to agree with Horacek in opting for DNF instead of CNF (Gatt ; Khan, van Deemter, and Ritchie ). Perspective and coherence. Recent studies have started to bring data-oriented methods to the generation of references to sets (Gatt ; Gatt and van Deemter ; Khan et al. ). One finding is that referring expressions benefit from a ‘coherent’ perspective. For example, ‘the Italian and the Greek’ is normally a better way to refer to two

OUP CORRECTED PROOF – FINAL, 21/1/2019, SPi



     

people than ‘the Italian and the cook’, since the former is generated from one coherent perspective (i.e., nationalities). Two questions need to be addressed, however. First, how should coherence be defined? Gatt () opted for a definition that assesses the coherence of a combination of properties using corpus-based frequencies as defined by Kilgarriff (), which in turn is based on D. Lin (). This choice was supported by a range of experiments (although the success of the approach is less well attested for descriptions that contain adjectives). Secondly, what if full coherence can only be achieved at the expense of brevity? Suppose a domain contains one Italian and two Greeks. One of the Greeks is a cook, while the other Greek and the Italian are both IT consultants. If this is all that is known, the generator faces a choice between either generating a description that is fully coherent but unnecessarily lengthy (‘the Italian IT consultant and the Greek cook’), or brief but incoherent (‘the Italian and the cook’). Simply saying ‘the Italian and the Greek’ would not be distinguishing. In such cases, coherence becomes a tricky—and computationally complex—optimization problem (Gatt ; Gatt and van Deemter ). Collective plurals. Reference to sets is a rich topic, where many issues on the borderline between theoretical, computational, and experimental linguistics are waiting to be explored. Most computational proposals, so far, use properties that apply to individual objects. To refer to a set, in this view, is to say things that are true of each member of the set. Such references may be contrasted with collective ones (e.g., ‘the lines that run parallel to each other’, ‘the group of four people’) which are more complicated from a semantic point of view (Scha and Stallard ; Lønning , among others). For initial ideas about the generation of collective plurals, we refer to Stone ().

.. Relational descriptions Another important limitation of most early  algorithms is that they are restricted to one-place predicates (e.g., ‘being a man’), instead of relations involving two or more arguments. Even a property like ‘wearing a suit’ is modeled as if it were simply a oneplace predicate without internal structure (instead of a relation between a person and a piece of clothing). This means that the algorithms in question are unable to identify one object via another, as when we say ‘the woman next to the man who wears a suit’, and so on. One early paper does discuss relational descriptions, making a number of important observations about them (Dale and Haddock ). First, it is possible to identify an object through its relations to other objects without identifying each of these objects separately. Consider a situation involving two cups and two tables, where one cup is on one of the tables. In this situation, neither ‘the cup’ nor ‘the table’ is distinguishing, but ‘the cup on the table’ succeeds in identifying one of the two cups. Secondly, descriptions of this kind can have any level of ‘depth’: in a complex situation, one might say ‘the white cup on the red table in the kitchen’, and so on. To be avoided, however,

OUP CORRECTED PROOF – FINAL, 21/1/2019, SPi

    



are the kinds of repetitions that can arise from descriptive loops, since these do not add information. It would, for example, be useless to describe a cup as ‘the cup to the left of the saucer to the right of the cup to the left of the saucer . . . ’. We shall return to this issue in §., where we shall ask how suitable each of a number of representational frameworks is for the proper treatment of relational descriptions. Various researchers have attempted to extend the IA by allowing relational descriptions (Horacek ; Krahmer and Theune ; Kelleher and Kruijff ), often based on the assumption that relational properties (like ‘x is on y’) are less preferred than non-relational ones (like ‘x is white’). If a relation is required to distinguish the target x, the basic algorithm is applied iteratively to y. It seems, however, that these attempts were only partly successful. One of the basic problems is that relational descriptions—just like references to sets, but for different reasons—do not seem to fit in well with an incremental generation strategy. In addition, it is far from clear that relational properties are always less preferred than non-relational ones (Viethen and Dale ). Viethen and Dale suggest that even in simple scenes, where objects can easily be distinguished without relations, participants still use relations frequently (in about one third of the trials). We return to this in §.. On balance, it appears that the place of relations in reference is only partly understood, with much of the iceberg still under water. If -place relations can play a role in , then surely so can n-place relations for larger n, as when we say ‘the city that lies in between the mountains and the sea’ (n = ). No existing proposal has addressed n-place relations in general, however. Moreover, human speakers can identify a man as the man who ‘kissed all women’, ‘only women, or ‘two women’. The proposals discussed so far do not cover such quantified relations, but see Ren, van Deemter, and Pan ().

.. Context-dependency, vagueness, and gradability So far we assumed that properties have a crisply defined meaning, which is fixed, regardless of the context in which they are used. But many properties fail to fit this mould. Consider the properties young and old, for example. In Figure ., it is the leftmost male who looks the older of the two. But if we add an old-aged pensioner to the scene then suddenly he is the most obvious target of expressions like ‘the older man’ or ‘the old man’. Whether a man counts as old or not, in other words, depends on what other people he is compared with: being old is a context-dependent property. The concept of being ‘on the left’ is context-dependent too: suppose we add five people to the right of the young man in Figure .; now all three characters originally depicted are suddenly on the left, including the man in the t-shirt who started out on the right. Concepts like ‘old’ and ‘left’ involve comparisons between objects. Therefore, if the knowledge base changes, the objects’ descriptions may also change. But even if the knowledge base is kept constant, the referent may have to be compared against different objects, depending on the words in the expression. The word ‘short’ in ‘John is

OUP CORRECTED PROOF – FINAL, 21/1/2019, SPi



     

a short basketball player’, for example, compares John’s height with that of the other basketball players, whereas ‘John is a short man’ compares its referent with all the other men, resulting in different standards for what it means to be short. ‘Old’ and ‘short’ are not only context dependent but also gradable, meaning that you can be more or less of it (older, younger, shorter, taller) (Quirk, Greenbaum, Leech, and Svartvik ). Gradable words are extremely frequent, and in many NLG systems they are of great importance, particularly in those that have numerical input, for example in weather forecasting (Goldberg, Driedger, and Kittredge ) or medical decision support (Portet et al. ). In addition to being context dependent, they are also vague, in the sense that they allow borderline cases. Some people may be clearly young, others clearly not, but there are borderline cases in between for whom it is not quite clear whether they were included. Context can help to diminish the problem, but it won’t go away: in the expression ‘short basketball player’, the noun gives additional information about the intended height range, but borderline cases still exist. Generating vague references. , as we know it, lets generation start from a Knowledge Base (KB) whose facts do not change as a function of context. This means that contextdependent properties like a person’s height need to be stored in the KB in a manner that does not depend on other facts. It is possible to deal with size adjectives in a principled way, by letting one’s KB contain a height attribute with numerical values. Our running example can be augmented by giving each of the three people a precise height, for example: height(d₁) = cm, height(d₂) = cm, and height(d₃) = cm (here the height of the woman d₂ has been increased for illustrative purposes). Now imagine we want to refer to d₃. This target can be distinguished by the set of two properties {man, height = cm}. Descriptions of this kind can be produced by means of any of the classic  algorithms. Given that type and height identify the referent uniquely, this set of properties can be realized simply as ‘the man who is cm tall’. But other possibilities exist. Given that cm is the greatest height of all men in this KB, the set of properties can be converted into {man, height = maximum}, where the exact height has been pruned away. The new description can be realized as ‘the tallest man’ or simply as ‘the tall man’ (provided the referent’s height exceeds a certain minimum value). The algorithm becomes more complicated when sets are referred to (because the elements of the target set may not all have the same heights), or when two or more gradable properties are combined (e.g., ‘the strong, tall man in the expensive car’; van Deemter ). Variations and extensions. Horacek () integrates vagueness with other types of uncertainty. Horacek could be said to depict a  algorithm as, essentially a gambler who wants to maximize the chance of the referent being identified on the basis of the generated expression. Other things being equal, for example, it may be safer to identify a dog as being ‘owned by John’, than as being ‘tall’, because the latter involves borderline cases. A similar approach can be applied to perceptual uncertainty (as when it is uncertain whether the hearer will be able to observe a certain property), or to the uncertainty associated with little-known words (e.g., will the hearer know what a

OUP CORRECTED PROOF – FINAL, 21/1/2019, SPi

    



basset hound is?) Quantifying all types of uncertainties could prove problematic in practice, yet by portraying a generator as a gambler, Horacek has highlighted an important aspect of reference generation which had so far been ignored. Crucially, his approach makes the success of a description a matter of degrees. The idea that referential success is a matter of degrees appears to be confirmed by recent applications of  to geospatial data. Here there tend to arise situations in which it is simply not feasible to produce a referring expression that identifies its target with absolute precision (though good approximations may exist). Once again, the degree of success of a referring expression becomes gradable. Suppose you were asked to describe that area of Scotland where the temperature is expected to fall below zero on a given night, based on some computer forecast of the weather. Even if we assume that this is a well-defined area with crisp boundaries, it is not feasible to identify the area precisely, because listing all the thousands of data points that make up the area separately is hardly an option. Various approximations are possible, including: ()

Roads above  metres will be icy.

()

Roads in the Highlands will be icy.

Descriptions of this kind are generated by a system for road gritting, where the decision which roads to treat with salt depends on the description generated by the system (Turner et al. ): roads where temperatures are predicted to be icy should be treated with salt; others should not. The two descriptions above are arguably only partially successful in singling out the target area. Generally speaking, one can distinguish between false positives and false negatives: the former are roads that are covered by the description but should not be (because the temperature there is not predicted to fall below zero), the latter are icy roads that will be left ungritted. Turner and colleagues decided that it would be unacceptable to have even one false negative. In other situations, safety (from accidents) and environmental damage (through salt) might be traded off in different ways, for example by associating a finite cost with each false positive and a possibly different cost with each false negative, and choosing the description that is associated with the lowest total cost (van Deemter : –). Again, a crucial and difficult part is to come up with the right cost figures.

.. Degrees of salience and the generation of pronouns When we speak about the world around us, we do not pay equal attention to all the objects in it. In a novel, for example, a sentence such as ‘Smiley saw the man approaching’ does not mean that Smiley saw the only man: it simply means that Smiley saw the man who is most salient at this stage of the novel. Passonneau () and Jordan () have shown how algorithms such as the IA may produce reasonable referring expressions ‘in context’, by limiting the set of salient objects in some sensible way, for example, to those objects mentioned in the previous utterance. Salience, in

OUP CORRECTED PROOF – FINAL, 21/1/2019, SPi



     

these works, was treated as a two-valued, ‘black-or-white’ concept. But perhaps it is more natural to think of salience—just like height or age—as coming in degrees. Existing theories of linguistic salience do not merely separate what is salient from what is not. They assign referents to different salience bands, based on factors such as recency of mention and syntactic structure (Gundel, Hedberg, and Zacharski ; Hajičová ; Grosz, Joshi, and Weinstein ). Salience and context-sensitive . Early  algorithms (Kronfeld ; Dale and Reiter ) assumed that salience could be modeled by means of a focus stack (Grosz and Sidner ): a referring expression is taken to refer to the highest element on the stack that matches its description, see also DeVault et al. (). Krahmer and Theune () argue that the focus stack approach is not flexible enough for context-sensitive generation of descriptions. They propose to assign individual salience weights (sws) to the objects in the domain, and to reinterpret referring expressions like the man as referring to the currently most salient man. Once such a gradable notion of salience is adopted, we are back in the territory of §... One simple way to generate contextsensitive referring expressions is to keep the algorithm of Figure . exactly as it is, but to limit the set of distractors to only those domain elements whose salience weight is at least as high as that of the target r. Line  (Figure .) becomes: ’. C ← {x|sw(x) ≥ sw(r)} – {r} To see how this works, consider the knowledge base of Table . once again, assuming that sw(d₁) = sw(d₂) = , while sw(d₃) =  (d₁ and d₂ are salient, for example, because they were just talked about, and d₃ was not). Suppose we keep the same domain and preference order as before. Now if d₁ is the target, then, according to the new definition ’, C = {d₁, d₂} – {d₁} = {d₂} (i.e., d₂ is the only distractor which is at least as salient as the target, d₁). The algorithm will select ⟨type, man⟩, which rules out the sole distractor d₂, leading to a successful reference (‘The man’). If, however, d₃ would be the target then C = {d₁, d₂, d₃} – {d₃} = {d₁, d₂}, and the algorithm would operate as normal, producing a description realizable as ‘the man in the t-shirt’. Krahmer and Theune chose to graft a variant of this idea onto the IA, but application to other algorithms is straightforward. Krahmer and Theune () compare two theories of computing linguistic salience— one based on the hierarchical focus constraints of Hajičová (), the other on the centering constraints of Grosz et al. (). They argue that the centering constraints, combined with a gradual decrease in salience of non-mentioned objects (as in the hierarchical focus approach) yields the most natural results. Interestingly, the need to compute salience scores can affect the architecture of the  module. In Centering Theory, for instance, the salience of a referent is co-determined by the syntactic structure of the sentence in which the reference is realized; it matters whether the reference is in subject, object, or another position. This suggests an architecture in which  and syntactic realization should be interleaved, a point to which we return below. Variations and extensions. Once salience of referring expressions is taken into account, and they are no longer viewed as de-contextualized descriptions of their referent, a number of questions come up. When, for example, is it appropriate to use

OUP CORRECTED PROOF – FINAL, 21/1/2019, SPi

    



a demonstrative (‘this man’, ‘that man’), or a pronoun (‘he’, ‘she’)? As for demonstratives, it has proven remarkably difficult to decide when these should be used, and even harder to choose between the different types of demonstratives (Piwek et al. ). Concerning pronouns, Krahmer and Theune suggested that ‘he’ abbreviates ‘the (most salient) man’, and ‘she’ ‘the (most salient) woman’. In this way, algorithms for generating distinguishing descriptions might also become algorithms for pronoun generation. However, such an approach to pronoun generation is too simple, since additional factors are known to determine whether a pronoun is suitable or not (McCoy and Strube ; Henschel, Cheng, and Poesio ; Callaway and Lester ; Kibble and Power ). Based on analyses of naturally occurring texts, McCoy and Strube (), for example, emphasized the role of topics and discourse structure for pronoun generation, and pointed out that the changes in time scale are a reliable cue for this. In particular, they found that in certain places a definite description was used where a pronoun would have been unambiguous. This happened, for example, when the time frame of the current sentence differed from that of the sentence in which the previous mention occurred, as can be signaled, for example, by a change in tense or a cue phrase such as ‘several months ago’. Kibble and Power (), in an alternative approach, use Centering Theory as their starting point in a constraint-based text generation framework, taking into account constraints such as salience, cohesion, and continuity for the choice of referring expressions. Many studies on contextual reference take text as their starting point (Poesio and Vieira ; Belz, Kow, Viethen, and Gatt , among others), unlike the majority of  research discussed so far, which uses standard knowledge representations of the kind exemplified in Table . (or some more sophisticated frameworks, see §.). An interesting variant is presented by Siddharthan and Copestake (), who set themselves the task of generating a referring expression at a specific point in a discourse, without assuming that a knowledge base (in the normal sense of the word) is available: all their algorithm has to go by is text. For example, a text might start saying ‘The new president applauded the old president’. From this alone, the algorithm has to figure out whether, in the next sentence, it can talk about ‘the old president’ (or some other suitable noun phrase) without risk of misinterpretation by the reader. The authors argue that standard  methods can achieve reasonable results in such a setting, particularly (as we shall see next) with respect to the handling of lexical ambiguities that arise when a word can denote more than one property. Lexical issues such as these transcend the selection of semantic properties. Clearly, it is time for us to consider matters that lie beyond Content Determination. Before we do this, however, we would like to mention that differences in salience can also be caused by non-linguistic factors: some domain objects, for example, may be less salient because they are further removed from the hearer than others. Paraboni et al. () demonstrated experimentally that such situations may require substantial deviations from existing algorithms, to avoid causing unreasonable amounts of work to the reader. To see the idea, consider the way we refer to a (non-salient) address on a map: we probably don’t say ‘Go to house number  in Aberdeen’, even if only one

OUP CORRECTED PROOF – FINAL, 21/1/2019, SPi



     

house in Aberdeen has that number and this description is thus perfectly distinguishing. It is more likely we say something like, ‘Go to house number  So-and-so Road, in the West End of Aberdeen’, adding logically redundant information specifically to aid the hearer’s search.

.. Beyond content determination In many early  proposals, Lexical Choice and Surface Realization follow Content Determination, in the style of a pipeline, with most of the actual research focusing predominantly on Content Determination. One might have thought that good results are easy to achieve by sending the output of the Content Determination module to a generic realizer (that is: a program converting meaning representations into natural language). With hindsight, any such expectations must probably count as naive. Some  studies have taken a different approach, interleaving Content Determination and Surface Realization (Horacek ; Stone and Webber ; Krahmer and Theune ; Siddharthan and Copestake ), running counter to the pipeline architecture (Mellish et al. ). In this type of approach, syntactic structures are built up in tandem with semantic descriptions: when ⟨type, man⟩ has been added to the semantic description, a partial syntactic tree is constructed for a noun phrase, whose head noun is man. As more properties are added to the semantic description, appropriate modifiers are slotted into the syntax tree; finally, the noun phrase is completed by choosing an appropriate determiner. Even in these interleaved architectures, it is often assumed that there is a one-to-one correspondence between properties and words; but often a property can be expressed by different words, one of which may be more suitable than the other, for example because it is unambiguous whereas the other is not (Siddharthan and Copestake ). One president may be ‘old’ in the sense of former, while another is ‘old’ in the sense of aged, in which case ‘the old president’ can become ambiguous between the two people. To deal with the choice between ‘old’ and ‘former’, Siddharthan and Copestake propose to look at discriminatory power, the idea being that in this case ‘former’ rules out more distractors than ‘old’ (both presidents are old). One wonders, however, to what extent readers interpret ambiguous words ‘charitably’: suppose two presidents are aged, while only one is the former president. In this situation, ‘the old president’ seems clear enough, because only one of its two interpretations justifies the definite article (namely the one where ‘old’ is to be understood as ‘former’). Clearly, people’s processing of ambiguous expressions is an area where there is still a lot to explore. If we turn away from Siddharthan and Copestake’s setup, and return to the situation where generation starts from a non-textual knowledge base, similar problems with ambiguities may arise. In fact, the problem is not confined to Lexical Choice: ambiguities can also arise during Surface Realization. To see this, suppose Content Determination has selected the properties ‘man’ and ‘with telescope’ to refer to a person, and the

OUP CORRECTED PROOF – FINAL, 21/1/2019, SPi

    



result after Surface Realization and Lexical Choice is ‘John saw the man with the telescope’ then, once again, the clarity of the semantic description can be compromised by putting the description in a larger context, causing an attachment ambiguity, which may sometimes leave it unclear what man is the intended referent of the description. The generator can save the day by choosing a different realization, generating ‘John saw the man who holds the telescope’ instead. Similar ambiguities occur in conjoined references to plurals, as in ‘the old men and women’, where ‘old’ may or may not pertain to the women. These issues have been studied in some detail as part of a systematic study of the ambiguities that arise in coordinated phrases of the form ‘the Adjective Noun and Noun’, asking when such phrases give rise to actual comprehension problems, and when they should be avoided by a generator (Chantree, Kilgarriff, de Roeck, and Willis ; Khan et al. ). When the generated referring expressions are realized in a medium richer than plain text, for instance in the context of a virtual character (Gratch et al. ), another set of issues comes into play. It needs to be decided, then, which words should be emphasized in speech, possibly in combination with visual cues such as eyebrow movements and other gestures. Doing full justice to the expanding literature on multimodal reference is beyond the scope of this survey, but a few pointers may be useful. Various early studies looked at multimodal reference (Lester, Voerman, Towns, and Callaway ). One account, where pointing gestures directly enter the Content Determination module of , is presented by van der Sluis and Krahmer (van der Sluis and Krahmer ), who focus on the trade-off between gestures and words. Kopp et al. () are more ambitious, modeling different kinds of pointing gestures and integrating their approach with the generation strategy of Stone et al. ().

.. Discussion Early  research made a number of simplifying assumptions, and as a result the early  algorithms could only generate a limited variety of referring expressions. When researchers started lifting some of these assumptions, this resulted in  algorithms with an expanded repertoire, being able to generate, for instance, plural and relational descriptions. However, this move created a number of new challenges. For instance, the number of ways in which one can refer to a set of target objects increases, so choosing a good referring expression is also more difficult. Should we prefer, for example, ‘the men not wearing an overcoat’, ‘the young man and the old man’, or ‘the men left of the woman’. In addition, from a search perspective, the various proposals result in a larger search space, making computational issues more pressing. For some of the extensions (e.g., where boolean combinations of properties are concerned), the complexity of the resulting algorithm is substantially higher than that of the base IA. Moreover, researchers have often zoomed in on one extension of the IA, developing a new version that lifts one particular limitation. Combining all the different extensions into one algorithm which is

OUP CORRECTED PROOF – FINAL, 21/1/2019, SPi



     

capable of, say, generating references to salient sets of objects, using negations and relations and possibly vague properties, is a non-trivial enterprise. To give just one example, consider what happens when we combine salience with (other) gradable properties (cf. §.. and §..). Should ‘the old man’ be interpreted as ‘the oldest of the men that are sufficiently salient’ or ‘the most salient of the men that are sufficiently old’? Expressions that combine gradable properties can easily become unclear, and determining when such combinations are nevertheless acceptable is an interesting challenge. Some simplifying assumptions have only just begun to be lifted, through extensions that are only in their infancy, particularly in terms of their empirical validation. Other simplifying assumptions are still in place. For instance, there is a dearth of work that addresses functions of referring expressions other than mere identification. Similarly, even recent proposals tend to assume that it is unproblematic to determine what information is shared between speaker and hearer. We return to these issues in §..

.  

.................................................................................................................................. Most early  algorithms represent knowledge in a very basic way, specifically designed for . This may have been justified at the time, but years of research in Knowledge Representation (KR) suggest that such a carefree attitude towards the modeling of knowledge may not be wise in the long run. For example, when well-established KR frameworks are used, it may become possible to re-use existing algorithms for these frameworks, which have often been optimized for speed, and whose computational properties are well understood. Depending on the choice of framework, many other advantages can ensue. Since research that couples  with KR is relatively new, and technical properties of the frameworks themselves can be easily found elsewhere, we shall be comparatively brief. For each framework, we focus on three questions: (a) How is domain information represented? (b) How is the semantic content of a referring expression represented? (c) How can distinguishing descriptions be found?

..  using graph search One of the first attempts to link  with a more generic mathematical formalism was the proposal by Krahmer, van Erk, and Verleg (), who used labeled directed graphs for this purpose. In this approach, objects are represented as the nodes (vertices) in a graph, and the properties of and relations between these objects are represented as edges connecting the nodes. Figure . shows a graph representation of our example domain. One-place relations (i.e., properties) such as ‘man’ are modeled as loops (edges beginning and ending in the same node), while -place relations such as ‘left of ’ are modeled as edges between different nodes.

OUP CORRECTED PROOF – FINAL, 21/1/2019, SPi

     older-looking

younger-looking

younger-looking

wears suit

wears t-shirt

wears t-shirt

man

woman

man

left of 1

right of



left of 2

right of

3

 . Representation of our example scene as a labeled directed graph. wears t-shirt

wears suit man

man

man

woman

left of

 . Some referring graphs for target d₁.

Two kinds of graphs play a role: a scene graph representing the knowledge base, and referring graphs representing the content of referring expressions. The problem of finding a distinguishing referring expression can now be defined as a comparison between graphs. More specifically, it is a graph search problem: given a target object (i.e., a node in the scene graph), look for a distinguishing referring graph that is a subgraph of the scene graph and uniquely characterizes the target. Intuitively, such a distinguishing graph can be ‘placed over’ the target node with its associated edges, and not over any other node in the scene graph. The informal notion of one graph being ‘placed over’ another corresponds with a subgraph isomorphism (Read and Corneil ). Figure . shows a number of referring graphs which can be placed over our target object d₁. The leftmost, which could be realized as ‘the man’, fails to distinguish our target, since it can be ‘placed over’ the scene graph in two different ways (over nodes  and ). Krahmer et al. () use cost functions to guide the search process and to give preference to some solutions over others. They assume that these cost functions are monotonic, so extending a graph can never make it cheaper. Graphs are compatible with many different search algorithms, but Krahmer et al. () employ a simple    algorithm for finding the cheapest distinguishing graph for a given target object. The algorithm starts from the graph containing only the node representing the

OUP CORRECTED PROOF – FINAL, 21/1/2019, SPi



     

target object and recursively tries to extend this graph by adding adjacent edges: edges starting from the target, or in any of the other vertices added later on to the referring graph under construction. For each referring graph, the algorithm checks which objects in the scene graph it may refer to, other than the target; these are the distractors. As soon as this set is empty, a distinguishing referring graph has been found. At this point, only alternatives that are cheaper than this best solution found so far need to be inspected. In the end, the algorithm returns the cheapest distinguishing graph which refers to the target, if one exists, otherwise it returns the empty graph. One way to define the cost function would be to assign each edge a cost of one point. Then the algorithm will output the smallest graph that distinguishes a target (if one exists), just as the   algorithm would. Alternatively, one could assign costs in accordance with the list of preferred attributes in the IA, making more preferred properties cheaper than less preferred ones. A third possibility is to compute the costs of an edge e in terms of the probability P(e) that e occurs in a distinguishing description (which can be estimated by counting occurrences in a corpus), making frequent properties cheap and rare ones expensive: cost(e) = –log₂(P(e)) Experiments with stochastic cost functions have shown that these enable the graph-based algorithm to capture a lot of the flexibility of human references (Krahmer, Theune, Viethen, and Hendrickx ; Viethen, Dale, Krahmer, Theune, and Touset ). In the graph-based perspective, relations are treated in the same way as individual properties, and there is no risk of running into infinite loops (the cup to the left of the saucer to the right of the cup . . . ’). Unlike Dale and Haddock () and Kelleher and Kruijff (), no special measures are required, since a relational edge is either included in a referring graph or not: including it twice is not possible. Van Deemter and Krahmer () show that many of the proposals discussed in §. can be recast in terms of graphs. They argue, however, that the graph-based approach is ill-suited for representing disjunctive information. Here, the fact that directed graphs are not a fully fledged KR formalism makes itself felt. Whenever a  algorithm needs to reason with complex information, heavier machinery is required.

..  using Constraint satisfaction Constraint satisfaction is a computational paradigm that allows efficient solving of NP hard combinatoric problems such as scheduling (van Hentenryck ). It is among the earliest frameworks proposed for  (Dale and Haddock ), but in later years, this approach has seldom been emphasized—with a few notable exceptions, such as Stone and Webber ()—until Gardent () showed how constraint programming can be used to generate expressions that refer to sets. She proposed to represent a description L for a target set S as a pair of set variables:

OUP CORRECTED PROOF – FINAL, 21/1/2019, SPi

    



Ls ¼ hPs þ ; Ps
i; where one variable (Ps+) ranges over sets of properties that are true of the elements in S and the other (Ps
) over properties that are false of the elements in S. The challenge— taken care of by existing constraint solving programs—is to find suitable values (i.e., sets of properties) for these variables. To be ‘suitable’, values need to fulfil a number of -style constraints: .

All the properties in P+s are true of all elements in S.

.

All the properties in Ps
are false of all elements in S.

.

For each distractor d there is a property in P+s which is false of d, or there is a property in Ps
which is true of d.

The third clause says that every distractor is ruled out by either a positive property (i.e., a property in Ps+) or a negative property (i.e., a property in Ps
), or both. An example of a distinguishing description for the singleton target set {d₁) in our example scene would be ⟨{man},{right}⟩, since d₁ is the only object in the domain who is both a man and not on the right. The approach can be adapted to accommodate disjunctive properties to enable reference to sets (Gardent ). Constraint satisfaction is compatible with a variety of search strategies (Kumar ). Gardent opts for a ‘propagate-and-distribute’ strategy, which means that solutions are searched for in increasing size, first looking for single properties, next for combinations of two properties, etc. This amounts to the Full Brevity search strategy, of course. Accordingly, Gardent’s algorithm yields a minimal distinguishing description for a target, provided one exists. Given the empirical questions associated with Full Brevity, it may well be worthwhile to explore alternative search strategies. The constraint approach allows an elegant separation between the specification of the  problem and its implementation. Moreover, the handling of relations is straightforwardly applicable to relations with arbitrary numbers of arguments. Gardent’s approach does not run into the aforementioned problems with infinite loops, because a set of properties (being a set) cannot contain duplicates. Yet, like the labeled graphs, the approach proposed by Gardent has significant limitations, which stem from the fact that it does not rest on a fully developed KR system. General axioms cannot be expressed, and hence cannot play a role in logical deduction. We are forced to revisit the question of what is the best way for  to represent and reason with knowledge.

..  using modern Knowledge Representation To find out what is missing, let us see what happens when domains scale up. Consider a furniture domain, and suppose every chair is in a room, that every room is in an apartment, and every apartment in a house. Listing all relevant relations between

OUP CORRECTED PROOF – FINAL, 21/1/2019, SPi



     

individual objects separately (‘chair a is in room b’, ‘room b is in apartment c’, ‘chair a is in apartment c’, ‘apartment c is in house d’) is onerous, error prone, spaceconsuming, and messy. Modern KR systems solve this problem by employing axioms (e.g., expressing transitivity of the ‘in’ relation; if x is in y, and y is in z, then x is in z). Logical inference allows the KR system to derive implicit information. For example, from ‘chair a is in room b’, ‘room b is in apartment c’, and ‘apartment c is in house d’, the transitivity of ‘in’ allows us to infer that ‘chair a is in house d’. This combination of basic facts and general axioms allows a succinct and insightful representation of facts. Modern KR comes in different flavors. Recently, two different KR frameworks have been linked with , one based on Conceptual Graphs (Croitoru and van Deemter ), the other on Description Logics (Gardent and Striegnitz ; Areces, Koller, and Striegnitz ). The first have their origin in Sowa () and were greatly enhanced by Baget and Mugnier (). The latter grew out of work on KL-ONE (Brachman and Schmolze ) and became even more prominent in the wider world of computing when they came to be linked with the ontology language OWL, which underpins current work on the semantic web (Baader, Calvanese, McGuinness, Nardi, and Patel-Schneider ). Both formalisms represent attempts to carve out computationally tractable fragments of First-Order Predicate Logic for defining and reasoning about concepts, and are closely related (Kerdiles ). For reasons of space, we focus on Description Logic. The basic idea is that a referring expression can be modeled as a formula of Description Logic, and that  can be viewed as the problem of finding a particular kind of formula, namely one that denotes (i.e., refers to) the target set of individuals. Let us revisit our example domain, casting it as a logical model M, as follows: M = ⟨D, ||.||⟩ where D (the domain) is a finite set {d₁, d₂, d₃} and ||.|| is an interpretation function which gives the denotation of the relevant predicates (thus: ||man|| = {d₁, d₃}, ||left-of || = {⟨d₁, d₂⟩, ⟨d₂, d₃⟩} etc.). Now the  task can be formalized as: given a model M and a target set S ⊆ D, look for a Description Logic formula φ such that ||φ|| = S. The following three expressions are the Description Logic counterparts of the referring graphs in Figure .: (a)

man

(b) man ⊓ wears suit (c)

man ⊓ ∃ left-of. (woman ⊓ wears t-shirt)

The first, (a), would not be distinguishing for d₁ (since its denotation includes d₃), but (b) and (c) would. Note that ⊓ represents the conjunction of properties, and ∃ represents existential restriction. Negations can be added straightforwardly, as in ‘man ⊓ ¬ wears suit’, which denotes d₃. Areces et al. () search for referring expressions in a somewhat non-standard way. In particular, their algorithm does not start with one particular target referent: it simply attempts to find the different sets that can be referred to. They start from the observation that  can be reduced to computing the similarity set of each domain

OUP CORRECTED PROOF – FINAL, 21/1/2019, SPi

    



object. The similarity set of an individual x is the set of those individuals that have all the properties that x has. Areces et al. () present an algorithm, based on a proposal by Hopcroft (), which computes the similarity sets, along with a Description Logic formula associated with each set. The algorithm starts by partitioning the domain using atomic concepts such as ‘man’ and ‘woman’, which splits the domain in two subsets ({d₁, d₃} and {d₂} respectively). At the next stage, finer partitions are made by making use of concepts of the form ∃R.AtomicConcept (e.g., ‘men left of a woman’), and so on, always using concepts established during one phase to construct more complex concepts during the next. All objects are considered in parallel, so there is no risk of infinite loops. Control over the output formulae is achieved by specifying an incremental preference order over possible expressions, but alternative control strategies could have been chosen.

.. Discussion Even though the role of KR frameworks for  has received a fair bit of attention in recent years, one can argue that this constitutes just the first step of a longer journey. The question of which KR framework suits  best, for example, is still open; which framework has the best coverage, which allows all useful descriptions to be expressed? Moreover, can referring expressions be found quickly in a given framework, and is it feasible to convert these representations into adequate linguistic realizations? Given the wealth of possibilities offered by these frameworks, it is remarkable that much of their potential is often left unused. In Areces et al.’s proposal, for example, generic axioms do not play a role, nor does logical inference. Ren, van Deemter, and Pan () sketch how  can benefit if the full power of KR is brought to bear, using Description Logic as an example. They show how generic axioms can be exploited, as in the example of the furniture domain, where a simple transitivity axiom allows a more succinct and insightful representation of knowledge. Similarly, incomplete information can be used, as when we know that someone is either Dutch or Belgian, without knowing which of the two. Finally, by making use of more expressive fragments of Description Logic, it becomes possible to identify objects that previous  algorithms were unable to identify, as when we say ‘the man who owns three dogs’, or ‘the man who only kisses women’, referring expressions that were typically not considered by previous  algorithms. Extensions of this kind also raise new empricial questions. It is an open question, for instance, when human speakers would be inclined to use such complex descriptions. These problems existed even in the days of the classic  algorithms (when it was already possible to generate lengthy descriptions) but they have become more acute now that it is also possible to generate structurally complex expressions. There is a clear need for empirical work here, which might teach us how the power of these formalisms ought to be constrained.

OUP CORRECTED PROOF – FINAL, 21/1/2019, SPi



     

. E 

.................................................................................................................................. Pre-  research gave little or no attention to the empirical evaluation of algorithms. More recently, however,  evaluation studies have started to be carried out more and more often. It appears that most of these were predicated on the assumption (debated in §.) that  algorithms should try to generate expressions that are optimally similar to these produced by human speakers or writers, even though, importantly, this assumption was seldom made explicit. The dominant method at the moment is, accordingly, to measure the similarity between generated expressions and the ones in a suitable corpus of referring expressions.  came late to corpus-based evaluation (compared to other parts of computational linguistics) because suitable data sets are hard to come by. In this section, we discuss what criteria a data set should meet to make it suitable for  evaluation, and we survey which collections are currently available. In addition, we discuss how one is to determine the performance of a  algorithm on a given data set. We shall see that while a lot of work has been done in recent years, there are still significant open questions, particularly regarding the relation between automatic metrics and human judgments.

.. Corpora for  evaluation Text corpora are full of referring expressions. For evaluating the realization of referring expressions, such corpora are very suitable, and various researchers have used them, for instance to evaluate algorithms for modifier orderings (Shaw and Hatzivassiloglou ; Malouf ; Mitchell ). Text corpora are also important for the study of anaphoric links between referring expressions. The texts that make up the GNOME corpus (Poesio, Stevenson, di Eugenio, and Hitzeman ), for instance, contain descriptions of museum objects and medical patient information leaflets, with each of the two subcorpora containing some  NPs. A lot of information is marked up, including anaphoric links. Yet, text corpora of this kind are of limited value for evaluating the content selection part of  algorithms. For that, one needs a corpus that is fully ‘semantically transparent’ (van Deemter, van der Sluis, and Gatt ): a corpus that contains the actual properties of all domain objects as well as the properties that were selected for inclusion in a given reference to the target. Text corpora such as GNOME do not meet this requirement, and it is often difficult or impossible to add all necessary information, because of the size and complexity of the relevant domains. For this reason, data sets for content selection evaluation are typically collected via experiments with human participants in simple and controlled settings. Broadly speaking, two kinds of experimental corpora can be distinguished: corpora specifically collected with reference in mind, and corpora collected wholly or partly for other purposes, but which have nevertheless been analyzed for the referring expressions in them. We will briefly sketch some corpora of the latter kind, after which we shall discuss the former in more detail.

OUP CORRECTED PROOF – FINAL, 21/1/2019, SPi

    



General-purpose corpora. One way to elicit ‘natural’ references is to let participants perform a task for which they need to refer to objects. An example is the corpus of socalled pear stories of Chafe (), in which people were asked to describe a movie about a man harvesting pears, in a fluent narrative. The resulting narratives featured such sequences as ‘And he fills his thing with pears, and comes down and there’s a basket he puts them in. [ . . . ] And then a boy comes by, on a bicycle, the man is in the tree, and the boy gets off his bicycle [ . . . ]’, where a limited set of individuals come up several times. The referring expressions in a subset of these stories were analysed in Passonneau (), who asked how the form of the re-descriptions (such as ‘he’, ‘them’, and ‘the man’) in these narratives might best be predicted, comparing ‘informational’ considerations (which form the core of most algorithms in the tradition started by Dale and Reiter, as we have seen) with considerations based on Centering Theory (Grosz et al. ). Passonneau, who tested her rules on  noun phrases, found support for an integrated model, where centering constraints take precedence over informational considerations. The well-known MapTask corpus (Anderson et al. ) is another example of a corpus in which reference plays an important role. It consists of dialogues between two participants; both have maps with landmarks indicated, but only one (the instruction giver) has a route on the map and he or she instructs the other (the follower) about this particular route. Referring expressions are routinely produced in this task to refer to the landmarks on the maps (‘the cliff ’). Participants use these not only for identification purposes but also, for instance, to verify whether they understood their dialogue partner correctly. In the original MapTask corpus, the landmarks were labeled with proper names (‘cliff ’), making them less suitable for studying content determination. To facilitate the study of reference, the iMap corpus was created (Guhe and Bard ), a modified version of the MapTask corpus where landmarks are not labeled, and systematically differ along a number of dimensions, including type (e.g., owl, penguin, etc.), number (singular, plural) and color; a target may thus be referred to as ‘the two purple owls’. Since participants may refer to targets more than once, it becomes possible to study initial and subsequent reference (Viethen, Zwarts, Dale, and Guhe ). Yet another example is the Coconut corpus (Di Eugenio, Jordan, Thomason, and Moore ), a set of task-oriented dialogues in which participants negotiate which furniture items they want to buy on a fixed, shared budget. Referring expressions in this corpus (‘a yellow rug for  dollars’) do not only contain information to identify a particular piece of furniture, but also include properties which directly refer to the task at hand (e.g., how much money is still available for a particular furniture item and what the state of agreement between the negotiators is). An attractive aspect of these corpora is that they represent fairly realistic communication, related to a more or less natural task. However, in these corpora, the identification of objects tends to be mixed with other communicative tasks (verification, negotiating). This does not mean that the corpora in question are unsuitable for the study of reference, of course. More specifically, they have been used for evaluating  algorithms, to compare the performance of traditional algorithms

OUP CORRECTED PROOF – FINAL, 21/1/2019, SPi



     

with special-purpose algorithms that take dialogue context into account (Passonneau ; Jordan and Walker ; Gupta and Stent ). For example, when the speaker attempts to persuade the hearer to buy an item, Jordan’s   algorithm selects those properties of the item that make it a better solution than a previously discussed item. In yet other situations—for example, when a summarization is offered—all mutually known properties of the item are selected. Jordan’s algorithm outperforms traditional algorithms, which is perhaps not surprising given that the latter were not designed to deal with references in interactive settings (Jordan ). Dedicated corpora. In recent years, a number of new corpora have been collected, specifically focusing on the types of referring expressions that we are focusing on in this survey. A number of such corpora are summarized in Table .. In some ways, these corpora are remarkably similar. Reflecting the prevalent aims of research on , for example, they focus on descriptions that aim to identify their referent ‘in one shot’, disregarding the linguistic context of the expression, that is, in the ‘null context’, as it is sometimes called (Viethen and Dale ). In all these corpora, participants were asked to refer to targets in a visual scene also containing the distractors. This setup means that the properties of target objects and their distractors are known, which makes it comparatively easy to make these corpora semantically transparent by annotating the references that were produced. In addition, most corpora are also ‘pragmatically transparent’, meaning that the communicative goals of the participants were known (typically identification). An early example is the Bishop corpus (Gorniak and Roy ). For this data set, participants were asked to describe objects in various computer generated scenes. Each of these scenes contained up to thirty objects (‘cones’) randomly positioned on a virtual surface. All objects had the same shape and size, and hence targets could only be distinguished using their color (either green or purple) and their location on the surface (‘the green cone at the left bottom’). Each participant was asked to identify targets in one shot, for the benefit of an addressee who was physically present but did not interact with the participant.

Table . Overview of dedicated Referring Expression corpora (alphabetical), with for each corpus a representative reference, an indication of the domain, and the number of participants and collected distinguishing descriptions Corpus Name Reference

Domain

Bishop Drawer GRED iMap TUNA

Colored cones in D scene Drawers in filing cabinet Spheres, Cubes in D scene Various objects on a map Furniture, People

Gorniak and Roy () Viethen and Dale () Viethen and Dale () Guhe and Bard () van Deemter et al. ()

Participants Descriptions     

    

OUP CORRECTED PROOF – FINAL, 21/1/2019, SPi

    



The Drawer corpus, collected by Viethen and Dale (), has a similar objective, but here targets are real, being one of sixteen colored drawers in a filing cabinet. On different occasions, participants were given a random number between one and sixteen and asked to refer to the corresponding drawer for an onlooker. Naturally, they were asked not to use the number; instead they could refer to the target drawers using color, row, and column, or some combination of those. In this corpus, referring expressions (‘the pink drawer in the first row, third column’) once again solely serve an identification purpose. Viethen and Dale () also collected another corpus (GRED) specifically looking at when participants use spatial relations. For this data collection, participants were presented with D scenes (made with Google SketchUp) containing three simple geometric objects (spheres and cubes of different colors and sizes, and in different configurations), of which one was the target. Viethen and Dale () found that spatial relations were frequently used (‘the ball in front of the cube’), even though they were never required for identification. Whether this generalizes to other visual scenes (in which spatial relations are less immediately ‘available’) is an interesting question for future research. The TUNA corpus (Gatt, van der Sluis, and van Deemter ; van Deemter, Gatt, van der Sluis, and Power ) was collected via a web-based experiment, in which singular and plural descriptions were gathered by showing participants one or two targets, where the plural targets could either be similar (same type) or dissimilar (different type). Targets were always displayed with six distractors, and the resulting domain objects were randomly positioned in a  x  grid, with targets surrounded by a red border. Example trials are shown in Figure .. The corpus contains two different domains: a furniture and a people domain. The first domain is based on pictures of furniture and household items, taken from the Object Databank (see http://www.tarrlab.org/). These were manipulated so that besides type (chair, desk, fan) also color, orientation, and size could systematically be varied. The number of possible attributes and values in the people domain is much larger (and more difficult to pin down); this domain consists of a set of black and white photographs of people (all famous mathematicians) used in an earlier study of van der Sluis and Krahmer (). Properties of these photographs include gender, head orientation,

 . Example trials from the TUNA corpus, a singular trial for the furniture domain (“the small blue fan”, left) and a plural trial for the people domain (“the men with glasses”, right).

OUP CORRECTED PROOF – FINAL, 21/1/2019, SPi



     

age, beard, hair, glasses, suit, shirt, and tie. It is interesting to realize that the TUNA corpus was designed to have one shortest description for each target, while in other data sets, such as Viethen and Dale’s () drawer corpus, a single shortest description does not always exist. The TUNA corpus has formed the basis of three shared  challenges, to which we turn below.

.. Evaluation metrics How to compare human descriptions with those produced by a  algorithm? When looking for measures that compute the content overlap, one source of inspiration may come from biology and information retrieval (van Rijsbergen ). One measure used in these fields is the Dice () coefficient, which was originally proposed to quantify ecologic association between species, and was first applied to  by Gatt et al. (). The Dice coefficient—which is not dissimilar to the ‘match’ function used by Jordan ()—is computed by scaling the number of elements that two sets have in common, by the size of the two sets combined: DiceðA; BÞ ¼

2  jA∩Bj jAj þ jBj

ð1Þ

The Dice measure ranges from  (no agreement, i.e., no elements shared between A and B) to  (complete agreement; A and B share all elements). For , A and B can be understood as attributes (e.g., ‘type’) or as attribute–value pairs (properties; ⟨type, man⟩). The former option tends to be used in earlier work, but has the somewhat counterintuitive consequence that two descriptions which express different values of the same attribute (‘the man’ and ‘the woman’, say, or ‘the dog’ and ‘the chihuahua’, in the earlier discussed cats-and-dogs example) have a Dice score of . Hence, in the discussion below we shall measure overlap in terms of properties. An alternative to Dice that is sometimes used is the MASI (Measuring Agreement on Set-valued Items) metric of Passonneau (): MASI ðA; BÞ ¼ δ 

jA ∩ Bj jA ∪ Bj

ð2Þ

This is basically an extension of the well-known Jaccard () metric with a weighting function δ which biases the score in favor of similarity where one set is a sub- or a superset of the other: 8 1; if A ¼ B > > < 2=3; if A ⊂ B or B ⊂ A δ¼ ð3Þ 0; if A ∩ B ¼ ø > > : 1=3; otherwise Dice and MASI are straightforward measures for overlap, but they do have their disadvantages. For example, they assume that all properties are independent and that

OUP CORRECTED PROOF – FINAL, 21/1/2019, SPi

    



all are equally different from each other. Suppose a human participant referred to d₁ in our example domain as ‘the man in the suit next to a woman’, and consider the following two references produced by a  algorithm: ‘the man in the suit’ and ‘the man next to a woman’. Both omit one property from the human reference and thus have the same Dice and MASI scores. But only the former reference is distinguishing; the latter is not. This problem could be solved, for example, by adopting a binary weighted version of the metrics which multiply the resulting score with  for a distinguishing description and with  for a non-distinguishing one. A more general issue with these overlap metrics can be illustrated with an example from Richard Power (pc.). Consider the two (roughly equivalent) expressions ‘the palomino’ and ‘the horse with the gold coat and white mane and tail’. Straightforward counting of attribute–value pairs would result in an overlap score of zero, which would be misleading, since the two descriptions express essentially the same content, with the latter description combining, in one property, all properties expressed in the former. This problem clearly calls for a more principled approach to representing and counting properties. During evaluations, Dice or MASI scores are typically averaged over references for different trials and produced by different human participants, making them fairly rough measures. It could be that an algorithm’s predictions match the descriptions of some participants very well, but those of other participants not at all. To partially compensate for this, sometimes also the proportion of times an algorithm achieves a perfect match with a human reference is reported. This measure is known, somewhat confusingly, as Recall (Viethen and Dale ), the Perfect Recall Percentage (PRP) (Gatt et al. ) and Accuracy (Gatt, Belz, and Kow ). The measures discussed so far do not take the actual linguistic realization of the referring expressions into account. For these, string distance metrics are obvious candidates, since these have proven their worth in various other areas of computational linguistics. One well-known string distance metric, which has also been proposed for  evaluation, is the Levenshtein () distance: the minimal number of insertions, deletions, and substitutions needed to convert one string into another, possibly normalized with respect to length (Bangalore, Rambow, and Whittaker ). The BLEU (Papineni, Roukos, Ward, and Zhu ) and NIST (Doddington ) metrics, which have their origin in machine translation evaluation, have also been proposed for  evaluation. BLEU measures n-gram overlap between strings; for machine translation n is often set to , but given that referring expressions tend to be short, n =  seems a better option for  evaluation (Gatt, Belz, and Kow ). NIST is a BLEU variant giving more importance to less frequent (and hence more informative) n-grams. Finally, Belz and Gatt () also use the rouge- and rouge-su measures (C.-Y. Lin and Hovy ), originally proposed for evaluating automatically generated summaries. An obvious benefit of these string metrics is that they are easy to compute automatically, while property-based evaluation measures such as Dice require an extensive manual annotation of selected properties. However, the added value of string-based metrics for  is relatively unclear. It is not obvious, for instance, that a smaller

OUP CORRECTED PROOF – FINAL, 21/1/2019, SPi



     

Levenshtein distance is always to be preferred over a longer one; the expressions ‘the man wearing a t-shirt’ and ‘the woman wearing a t-shirt’ are at a mere  Levenshtein distance from each other, but only the former would be a good description for target d₃. On the other hand, ‘the male person on the right’ is at a Levenshtein distance of  from ‘the man wearing a t-shirt’, and both are perfect descriptions of d₃. Alternatively, referring expressions could also be evaluated by human judges, although this obviously is more time consuming than an automatic evaluation. Gatt et al. () collected judgments of Adequacy (‘How clear is this description? Try to imagine someone who could see the same grid with the same pictures, but didn’t know which of the pictures was the target. How easily would they be able to find it, based on the phrase given?’) and Fluency (‘How fluent is this description? [ . . . ] Is it good, clear English?’). One may also be interested in the extent to which references are useful for addressees. This can be evaluated in a number of different ways. Belz and Gatt (), for example, first showed participants a generated description for a trial. After participants read this description, a scene appeared and participants were asked to click on the intended target. This allowed them to compute three extrinsic evaluation metrics: the reading time, the identification time, and the error rate, which they defined as the number of incorrectly identified targets.

.. Discussion Three lessons can be learnt from the recent work on evaluation. First, the emergence of transparent corpora after  has greatly facilitated the empirical evaluation of  algorithms, particularly for content selection. Focusing on reference in simple situations, a number of studies based on transparent corpora found that the IA outperformed the   and   algorithms (Viethen and Dale ; van Deemter et al. ). There is an important catch, however: as demonstrated in van Deemter et al. (), the performance of the IA crucially depends on the chosen preference order: the best preference order outperforms the other two algorithms, but many other preference orders perform far worse. This is a problem, since no procedure for finding a good preference order is known. (For n attributes, there are n! preference orders to consider, so trial and error is not an option except in extremely simple cases.) Perhaps most controversially, the authors argue that the evidence is starting to stack up in favor of the thesis that the  algorithm—or variants of the  algorithm that choose properties on the basis of more than just their discriminatory power—might be superior to algorithms that use the same preference order all the time. Second, evaluations suggest that human-produced descriptions differ from automatically generated ones in a number of ways. Human references often include redundant information, making the descriptions overspecified in ways that were not accounted for by standard  algorithms. An additional problem is that there appears

OUP CORRECTED PROOF – FINAL, 21/1/2019, SPi

    



to be considerable individual variation, both within and between speakers, which is something that existing  algorithms do not model (Dale and Viethen ). Third, it is still somewhat unclear what the best  evaluation metrics are. The three  Challenges based on the TUNA set-up offer a wealth of information in this respect (Gatt and Belz ). In each of these challenges, a number of research teams submitted one or more  generation systems, allowing detailed statistical analyses over the various metrics. It was found that Dice, MASI, and PRP are very highly correlated (all r > .). Interestingly, these metrics correlate negatively with the proportion of references that are minimally specified (Gatt et al. ); in other words, systems that produce more overspecified references tend to do better in terms of Dice and other overlap metrics. Concerning the surface realization metrics, it was found that—when comparing different realizations of a given set of attributes—the NIST and BLEU string metrics correlate strongly with each other (r = .), as one might expect, but neither correlates well with Levenshtein distance (Gatt et al. ). As for the extrinsic measures, Gatt et al. () only report a significant correlation between reading time and identification time, which suggests that slow readers are also slow identifiers, or that referring expressions that are hard to read also make it harder to identify the intended referent. Gatt et al. () let participants listen to expressions that were produced either automatically or by human speakers, and found a strong correlation between identification accuracy and adequacy, suggesting that more adequate references also have more correct identifications. Also, they found a negative correlation between fluency and identification time, implying that more fluent descriptions reduce the identification time. It is notable that essentially no correlations were found between these extrinsic task performance measures and the automatic metrics for human-likeness (Belz and Gatt ; Gatt and Belz ). Different explanations are possible for this lack of correlation. Gatt and Belz () note that the nature of the TUNA data could be partly responsible. The TUNA data collection was done in a web-based and relatively unrestricted manner, and idiosyncratic references do occur in it (‘a red chair, if you sit on it, your feet would show the south east’). It is possible that for another, more controlled corpus, a correlation between the two kinds of metrics would show up. It could also be that people are not always very good at designing their utterances in a way that is optimal for hearers (Horton and Keysar ) (see also §.), so producing descriptions that resemble human-produced ones is not the same as producing descriptions that are of optimal use for hearers. This suggests that the two sets of metrics measure different things, and that they correspond with two different aims that the designer of a  algorithm might have: one set of metrics could be used if the aim is to mimic speakers, another if the aim is to produce optimal benefits for hearers. So far, experimental evaluation has mostly been limited to the simplest of situations, focusing on algorithms that produce singular descriptions, expressing conjunctions of basic properties in small and artificial domains. Most of the extensions discussed in §. have not been evaluated systematically. Moreover, tasks such as the one on which the TUNA corpus is based can be argued not to be ‘ecologically valid’: human

OUP CORRECTED PROOF – FINAL, 21/1/2019, SPi



     

participants produce type-written expressions for an imaginary audience on the basis of abstract visual scenes. The effects of these limitations on the descriptions produced are partly unknown, although some reassuring results have recently been obtained. It has been shown, for example, that, speakers who address an imaginary audience refer in similar ways to those who address an audience that is physically present (van der Wege ). Similarly, Koolen et al. () show that speaking rather than typing has no effect on the kind and number of attributes in the referring expressions that are produced, although speakers tend to use more words than typists to convey the same amount of information. It would nevertheless be valuable to evaluate  algorithms in the context of a specific application, so the added value of different  algorithms for a real-life application can be gauged (Gatt et al. ). Two recent evaluation challenges seem promising for these reasons. GREC (Belz et al. ) focuses on the task of deciding which form a referring expression should take in a textual context, which is important for generating coherent texts such as summaries (see also §.). GIVE (Koller et al. ) focuses on generating directions in a virtual D environment, where reference is only one task among a number of others. This new challenge has so far not included a separate test of  algorithms employed in the systems submitted, but it seems likely that GIVE will cause  research to focus on harder tasks, including reference in discourse context, reference to sets, and references that are spread out over several utterances (Denis ).

. O 

.................................................................................................................................. In the previous sections we have discussed three main dimensions in which  research has moved beyond the state-of-the-art of . Along the way, various loose ends have been identified. For example, not all simplifying assumptions of early  work have been adequately addressed, and the enterprise of combining extensions is still in its infancy (§.). It is unclear whether complex referring expressions can always be found quickly, particularly where the generation process relies on theorem proving, and it is similarly unclear whether it is always feasible to turn the representations into fluent natural language expressions (§.). Empirical data has only been collected for the simplest referring expressions, and it is still unclear what the proper evaluation metrics are (§.). In this section, we suggest six further questions for future research. . How to match a  algorithm to a particular domain and application? Evaluation of classic  algorithms has shown that with some preference orders, the IA outperformed the   and   algorithms, but with others it performed much worse than these (van Deemter et al. ). The point is that the IA, as it stands, is underdetermined, because it does not contain a procedure for finding a preference order. Sometimes psycholinguistic experiments come to our aid, for

OUP CORRECTED PROOF – FINAL, 21/1/2019, SPi

    



instance Pechmann’s () study showing that speakers have a preference for absolute properties (color) over relative ones (size). Unfortunately, for most other attributes, no such experiments have been done. It seems reasonable to assume that frequency tells us something about preference: a property that is used frequently is also more likely to be high on the list of preferred properties (Gatt and Belz ; van Deemter et al. ). But suitable corpora to determine preferences are rare, as we have seen, and their construction is time consuming. This raises the question about how much data would be needed to make reasonable guesses about preferred properties; this could be studied, for instance, by drawing learning curves where increasingly large proportions of a transparent corpus are used to estimate a preference order and the corresponding performance is measured. The IA is more drastically underdetermined than most other algorithms: the   and the   algorithms are specified completely up to situations where there is a tie: a tie between two equally lengthy descriptions in the first case, and a tie between two properties that have the same discriminatory power in the second. To resolve such ties frequency data would clearly be helpful. Similar questions apply to other generation algorithms. For instance, the - algorithm as described by Krahmer et al. () assigns one of three different costs to properties (they can be free, cheap, or somewhat expensive), and frequency data is used to determine which costs should be assigned to which properties (properties that are almost always used in a particular domain can be for free, etc.). A recent experiment (Theune, Koolen, Krahmer, and Wubben ) suggests that training the - algorithm on a corpus with a few dozen items may already lead to a good performance. In general, knowing how much data is required for a new domain to reach a good level of performance is an important open problem for many  algorithms. . How to move beyond the ‘paradigms’ of reference? A substantial amount of  research focuses on what we referred to in the Introduction as the ‘paradigms’ of reference: ‘first-mention’ distinguishing descriptions consisting of a noun phrase starting with ‘the’, which serve to identify some target, and which do so without any further context. But how frequent are these ‘paradigmatic’ kinds of referring expressions? Poesio and Vieira (), in one of the few systematic attempts to quantify the frequency of different uses of definite descriptions in segments of the Wall Street Journal corpus, reported that ‘first mention definite descriptions’ are indeed the most frequent in these texts. These descriptions often do not refer to visual objects in terms of perceptual properties but to more abstract entities. One might think that it matters little whether a description refers to a perceivable object or not; a description like ‘the third quarter’ rules out three quarters much like ‘the younger-looking man’ in our example scene rules out the older-looking distractor. It appears, however, that the representation of the relevant facts in such cases tends to be a more complicated affair, and it is here particularly that more advanced knowledge representation formalisms of the kind discussed in §. come into their own (a point to which we return below).

OUP CORRECTED PROOF – FINAL, 21/1/2019, SPi



     

Even though first-mention definite descriptions are the most frequent in Poesio and Vieira’s sample, other uses abound, including anaphoric descriptions and bridging descriptions, whose generation is studied by Gardent and Striegnitz (). Pronouns come to mind as well. The content determination problem for these other kinds of referring expressions may not be overly complex, but deciding where in a text or dialogue each kind of referring expression should be used is hard. Still, this is an important issue for, for example, automatic summarization. One of the problems of extractive summaries is that co-reference chains may be broken, resulting in less coherent texts. Regeneration of referring expressions is a potentially attractive way of regaining some of the coherence of the source document (Nenkova and McKeown ). This issue is even more pressing in multi-document summarization, where different source documents may refer to a given person in different ways; see (Siddharthan, Nenkova, and McKeown ) for a machine-learning approach to this problem.  research often works from the assumption that referents can not be identified through proper names. (If proper names were allowed, why bother inventing a description?) But in real text, proper names are highly frequent. This does not only raise the question when it’s best to use a proper name, or which version of a proper name should be used (is it ‘Prince Andrei Nikolayevich Bolkonsky’, ‘Andrei Bolkonsky’, or just ‘Andrei’?), but also how proper names can occur as part of a larger description, as when we refer to a person using the description ‘the author of Primary Colours’, for example, where the proper name ‘Primary Colours’ refers to a well known book (whose author was long unknown). Surely, it is time for  to turn proper names into first-class citizens. Generation of referring expressions in a text is studied in the GREC (Generating Referring Expressions in Context) challenges (Belz, Kow, Viethen, and Gatt ). A corpus of wikipedia texts (for cities, countries, rivers, persons, and mountains) was constructed, and in each text all elements of the coreference chain for the main subject were removed. For each of the resulting reference gaps, a list of alternative referring expressions, referring to the subject, was given (including the ‘correct’ reference, i.e., the one that was removed from the text). One well-performing entry (Hendrickx, Daelemans, Luyckx, Morante, and Asch ) predicted the correct type of referring expression in % of the cases, using a memory-based learner. These results suggest that it is feasible to learn which type of referring expression is best in which instance. If so,  in context could be conceived of as a two-stage procedure where first the form of a reference is predicted, after which the content and realization are determined.  algorithms as described in the present survey would naturally fit into this second phase. It would be interesting to see if such a method could be developed for a data collection such as that of Poesio and Vieira (). . How to handle functions of referring expressions other than identification? Target identification is an important function of referring expressions, but it is not the only one. Consider the following example, which Dale and Reiter () discuss to illustrate the limits of their approach:

OUP CORRECTED PROOF – FINAL, 21/1/2019, SPi

     ()



Sit by the newly painted table.

Here, ‘the newly painted table’ allows the addressee to infer that it would be better not to touch the table. To account for examples such as this one, a  algorithm should be able to take into account different speaker goals (e.g., to identify, to warn, etc.) and allow these goals to drive the generation process. These issues were already studied in the plan-based approach to  of Appelt and Kronfeld (§..), and more recent work addresses similar problems using new methods. Heeman and Hirst (), for example, present a plan-based, computational approach to  where referring is modeled as goal-directed behavior. This approach accounts for the combination of different speaker goals, which may be realized in a single referring expression through ‘information overloading’ (Pollack ). Context is crucial here: a variant such as ‘What do you think of the newly painted table?’ does not trigger the intended ‘don’t touch’ inference. In another extension of the plan-based approach to reference, Stone and Webber () use overloading to generate references that only become distinguishing when the rest of the sentence is taken into account. For example, we can say ‘Take the rabbit from the hat’ if there are two rabbits, as long as only one of them is in a hat. Plan-based approaches to natural language processing are not as popular as they were in the eighties and early nineties, in part because they are difficult to develop and maintain. However, Jordan and Walker () show that a natural language generator can be trained automatically on features inspired by a plan-based model for  (Jordan ). Jordan’s   model incorporates multiple communicative and task-related problem solving goals, besides the traditional identification goal. Jordan supports her model with data from the Coconut corpus (discussed in §..) and shows that traditional algorithms such as the IA fail to capture which properties speakers typically select for their references, not only because these algorithms focus on identification, but also because they ignore the interactive setting (see below). In short, it seems possible to incorporate different goals into a  algorithm, even without invoking complex planning machinery. However, this calls for a close coupling of  with the generation of the carrier utterance, containing the generated expression. What impact this has on the architecture of an NLG system, what the relevant goals are, how combinations of different goals influence content selection and linguistic realization, and how such expressions are best evaluated is still mostly unexplored. Answers might come from studying  in the context of more complex applications, where the generator may need to refer to objects for different reasons. . How to generate suitable referring expressions in interactive settings? Ultimately, referring expressions are generated for some addressee, yet most of the algorithms we have discussed are essentially ‘addressee-blind’ (Clark and Bangerter ). To be fair, some researchers have paid lip service to the importance of taking the addressee into account (cf. Dale and Reiter’s UserKnows function), but it is still largely an open question to what extent the classical approaches to  can be used in interactions. In fact, there are good reasons to assume that most current 

OUP CORRECTED PROOF – FINAL, 21/1/2019, SPi



     

algorithms cannot directly be applied in an interactive setting. Psycholinguistic studies on reference production, for example, show that human speakers do take the addressee into account when referring—an instance of ‘audience design’ (Clark and Murphy ). Some psycholinguists have argued that referring is an interactive and collaborative process, with speaker and addressee forming a ‘conceptual pact’ on how to refer to some object (Brennan and Clark ; Metzing and Brennan ; Heeman and Hirst ). This also implies that referring is not necessarily a ‘one shot’ affair; rather a speaker may quickly produce a first approximation of a reference to some target, which may be refined following feedback from the addressee. Others have argued that conversation partners automatically ‘align’ with each other during interaction (Pickering and Garrod ). For instance, Branigan et al. () report on a study showing that if a computer uses the word ‘seat’ instead of the more common ‘bench’ in a referring expression, the user is also subsequently more likely to use ‘seat’ instead of ‘bench’. This kind of lexical alignment takes place at the level of linguistic realization, and there is at least one NLG realizer that can mimic this process (Buschmeier, Bergmann, and Kopp ). Goudbeek and Krahmer () found that speakers in an interactive setting also align at the level of content selection; they present experimental data showing that human speakers may opt for a ‘dispreferred’ attribute (even when a preferred attribute would be distinguishing) when these were salient in a preceding interaction. The reader may want to consult Arnold () for an overview of studies on reference choice in context, Clark and Bangerter () for a discussion of studies on collaborative references, or Krahmer () for a confrontation of some recent psycholinguistic findings with  algorithms. Psycholinguistic studies suggest that traditional  algorithms which rely on some predefined ranking of attributes cannot straightforwardly be applied in an interactive setting. This is confirmed by the findings of Jordan and Walker () and Gupta and Stent (), who studied references in dialogue corpora discussed in §.. They found that in these data-sets, traditional algorithms are outperformed by simple strategies which pay attention to the referring expressions produced earlier in the dialogue. A more recent machine learning experiment on a larger scale, using data from the iMap corpus, confirmed the importance of features related to the process of alignment (Viethen, Dale, and Guhe ). Other researchers have started exploring the generation of referring expressions in interactive settings as well. Stoia et al. (), for example, presented a system that generates references in situated dialogues, taking into account both dialogue history and spatial visual context, defined in terms of which distractors are in the current field of vision of the speakers and how distant they are from the target. Janarthanam and Lemon () present a method which automatically adapts to the expertise level of the intended addressee (using ‘the router’ when communicating with an expert user, and ‘the black block with the lights’ while interacting with a novice). This line of research fits in well with another, more general, strand of research concentrating on choice optimization during planning based on user data (Walker, Stent, Mairesse, and Prasad ; White et al. ).

OUP CORRECTED PROOF – FINAL, 21/1/2019, SPi

    



Interactive settings seem to call for sophisticated addressee modeling. However, detailed reasoning about the addressee can be computationally expensive, and some psychologists have argued, based on clever experiments in which speakers and addressees have slightly different information available, that speakers only have limited capabilities for considering the addressee’s perspective (Horton and Keysar ; Keysar, Lin, and Barr ; Lane, Groisman, and Ferreira ). Some of the studies mentioned above, however, emphasize a level of cooperation that may not require conscious planning: the balance of work on alignment, for example, suggests that it is predominantly an automatic process which does not take up much computational resource. Recently, Gatt et al. () proposed a new model for interactive , consisting of a preference-based search process based on the IA, which selects properties concurrently and in competition with a priming-based process, both contributing properties to a limited capacity working memory buffer. This model offers a new way to think about interactive , and the role therein for  algorithms of the kind discussed in this survey. . What is the impact of visual information? In this survey, we have often discussed references to objects in shared visual scenes, partly because this offers a useful way to illustrate an algorithm. So far, however, relatively few  studies have taken visual information seriously. Most real-life scenes contain a multitude of potential referents. Just look around you: every object in your field of vision could be referred to. It is highly unlikely that speakers would take all these objects into account when producing a referring expression. Indeed, there is growing evidence that the visual system and the speech production system are closely intertwined (Meyer, Sleiderink, and Levelt ; Spivey, Tyler, Eberhard, and Tanenhaus ; Hanna and Brennan ), and that human speakers employ specific strategies when looking at real-world scenes (Itti and Koch ; Wooding, Muggelstone, Purdy, and Gale , among others). Wooding and colleagues, for instance, found that certain properties of an image, such as changes in intensity and local contrasts, determine viewing patterns to a large extent. Top-down strategies also play a role: for instance, areas that are currently under discussion are looked at more frequently and for longer periods of time. Yet, little is known about how scene perception influences the human production of referring expressions, and how  algorithms could mimic this. When discussing visual scenes, most  researchers assume that some of the relevant visual information is stored in a database (compare our visual example scene in Figure . and its database representation in Table .). Still, the conversion from one to the other is far from trivial. Clearly, the visual scene is much more informative than the database; how do we decide which visual information we store in the database and which we ignore? And, how do we map visual information to symbolic labels? These are difficult questions, which have received very little attention so far. A partial answer to the first question can be found in the work of John Kelleher and colleagues, who argue that visual and linguistic salience co-determine which aspects of a scene are relevant for the understanding and generation of referring

OUP CORRECTED PROOF – FINAL, 21/1/2019, SPi



     

expressions (Kelleher, Costello, and van Genabith ; Kelleher and Kruijff ). A partial answer to the second question is offered by Deb Roy and colleagues (Roy and Pentland ; Roy ) who present a computational model for automatically grounding attributes based on sensor data, and by Gorniak and Roy () who apply such a model to referring expressions. One impediment to progress in this area is the lack of relevant human data. Most, if not all, of the dedicated data-sets discussed in §. were collected using artificial visual scenes, either consisting of grids of unrelated objects not forming a coherent scene, or of coherent scenes of unrealistic simplicity. Generally speaking, the situation in psycholinguistics is not much better. Recently, some studies started exploring the effects of more realistic visual scenes on language production. An example is Coco and Keller (), who photoshopped a number of (more or less) realistic visual scenes, manipulating the visual clutter and number of actors in each scene. They found that more clutter and more actors resulted in longer delays before language production started, and that these factors also influenced the syntactic constructions that were used. A similar paradigm could be used to collect a new corpus of human-produced references, with targets being an integral part of a visual scene (rather than being randomly positioned in a grid). When participants are subsequently asked to refer to objects in these scenes, eye tracking can be used to monitor where they are looking before and during the production of particular references. Such data would be instrumental for developing  algorithms which take visual information seriously. . What Knowledge Representation framework suits  best? Recent years have seen a strengthening of the link between  and knowledge representation frameworks (see §.). There is a new emphasis on questions involving () the expressive power of the formalism in which domain knowledge is expressed (e.g., does the formalism allow convenient representation of n-place predicates or quantification?), () the expressive power of the formalism in which ontological information is expressed (e.g., can it express more than just subsumption between concepts?), () the amount of support available for logical inference, and () the mechanisms available for controlling the output of the generator. To illustrate the importance of expressive power and logical inference, consider the type of examples discussed in Poesio and Vieira (). What would it take to generate an expression like ‘the report on the third quarter of ’? It would be cumbersome to represent the relation between all entities separately, saying that  has a first quarter, which has a report, and the same for all other years. It would be more elegant and economical to spell out general rules, such as ‘Every year has a unique first quarter’, ‘Quarter  of a given year precedes Quarter  of any later year’, ‘The relation “precede” is transitive’, and so on. As NLG is starting to be applied in large-scale applications, the ability to capture generalizations of this kind is bound to become increasingly important. It is remarkable that most  research has distanced itself so drastically from other areas of Artificial Intelligence, by limiting itself to atomic facts in the knowledge base. If

OUP CORRECTED PROOF – FINAL, 21/1/2019, SPi

    



 came to be linked with modern knowledge representation formats—as opposed to the simple attribute–value structures exemplified in Table .—then atomic formulas are no longer the substance of the knowledge base but merely its seeds. In many cases, resources developed for the semantic web—ontology languages such as OWL, reasoning tools, and even the ontologies themselves—could be reused in .  could even link up with ‘real AI’, by tapping into models of common-sense knowledge, such as Lenat () or Lieberman et al. (). The new possibilities raise interesting scientific and strategic questions. For example, how do people generate referring expressions of the kind highlighted by the work of Poesio and colleagues? Is this process best modeled using a knowledge-rich approach using general axioms and deduction, or do other approaches offer a more accurate model? Is it possible that, when  starts to focus a bit less on identification of the referent, the result might be a different, and possibly less logic-oriented problem? What role could knowledge representation play in these cases? Here, as elsewhere in , we see ample space for future research.

. G   

.................................................................................................................................. After preparatory work in the s by Appelt and Kronfeld, and the contributions summarized in Dale and Reiter (), the first decade of the new millennium has seen a new surge of interest in referring expression generation. Progress has been made in three related areas which have been discussed extensively in this survey. First, researchers have lifted a number of simplifications present in the work of Dale and Reiter () and others, thereby considerably extending coverage of  algorithms to include, for instance, relational, plural, and vague references (§.). Second, proposals have been put forward to recast  in terms of existing and well-understood computational frameworks, such as labeled directed graphs and Description Logic, with various attractive consequences (§.). Third, there has been a shift towards data collection and empirical evaluation; this has made it possible to empirically evaluate  algorithms, which is starting to give us an improved understanding of the strengths and weaknesses of existing work (§.). As a result of these developments,  is now one of the best developed subfields of NLG. How should the current state of the art in  be assessed? The good news is that current  algorithms can produce natural descriptions, which may even be more helpful than descriptions produced by people (Gatt et al. ). However, this is only true when certain simplifying assumptions are made, as in the early  research typified by Dale and Reiter (). When  leaves this limited ‘comfort zone’, the picture changes drastically. While in recent years the research community has gained a much better understanding of the challenges that face  in that wider arena, many of these challenges are still waiting to be met (§.).

OUP CORRECTED PROOF – FINAL, 21/1/2019, SPi



     

New complexities. Recent  research has revealed various new complexities. Some of these pertain to the nature of the target. Sets are difficult to refer to, for example, and algorithms designed to deal with them achieve a lower human-likeness when referring to sets than to individual objects (van Deemter et al. ). Recent efforts to let  algorithms refer to spatial regions suggest that in large, realistic domains, precise identification of a target is a goal that can be approximated, but seldom achieved (Turner et al. , ). Little work has been done so far on reference to events, or to points and intervals in time (e.g., ‘When Harry met Sally’, ‘the moment after the impact’), and references to abstract and other uncountable entities (e.g., water, democracy) are beyond the horizon. Where domain knowledge derives from sensor data— with unavoidably uncertain and noisy inputs—this is bound to cause problems not previously addressed by . It is in such domains that salience (especially in the nonlinguistic sense) becomes a critical issue. When reference takes place in real life—as opposed to a typical psycholinguistics experiment—it is often unclear what their salience depends on. It might be that salience is partly in the eye of the beholder, and that this is one of the causes of the considerable individual variation that exists between different human speakers (Dale and Viethen ). Human-likeness and evaluation. In early  research, including (Dale and Reiter ), it was often remarkably unclear what exactly the proposed algorithms aimed to achieve. It was only when evaluation studies were starting to be conducted that researchers had to ‘show their cards’ and say what they regarded as their criterion for success. In most cases, they used a form of human-likeness as their success criterion, by comparing the expressions generated by an algorithm with those in a corpus. The human-likeness criterion dictates that  algorithms are to mimic humans ‘warts and all’: if speakers produce unclear descriptions, then so should algorithms. But of course, human-likeness is not the only yardstick that can be used. In NLG systems whose main aim is to be practically useful, for example, it may be more important for referring expressions to be clear than to be human-like in all respects. The difference is important because psycholinguists have shown that human speakers have only limited capabilities for taking the addressee into account, frequently producing expressions that cannot be interpreted correctly by an addressee, for example when they are under time pressure (Horton and Keysar ). If usefulness or successfulness (Garoufi and Koller ), rather than human-likeness is the yardstick for success then a different type of evaluation test needs to be used. Possible tests include, for example, speed and accuracy of task completion (i.e., how often and how fast do readers find the referent?). A variety of hearer-oriented tests is starting to be used in recent  research (Paraboni et al. ; Khan et al. ), but evaluation of  algorithms (and of NLG in general) remains difficult (see e.g., Oberlander ; Belz ; and Gatt and Belz ). Arguably, a central problem is that many different evaluations metrics are conceivable, and a  algorithm (like an NLG system in general) may well score high on some and poorly on other metrics.

OUP CORRECTED PROOF – FINAL, 21/1/2019, SPi

    



Hearer-oriented experiments may also be useful for evaluating referring expressions that are logically complex (cf. §..). It is one thing for a  algorithm to use logical quantification to generate a fairly simple description, such as ‘the woman who feeds four cats’, but quite another to generate a highly complex description (‘the woman who owns four cats that are chased by between three and six dogs each of which is fed only by men’), which can be generated using the same methods. There are difficult methodological questions to be answered here, about whether the aim of the generator is to model human competence or human performance. And if it is performance that is to be modeled, then this raises the question what types of complexities are exploited by human speakers, and what types of complexities are understandable to human hearers. Such questions can only be answered by new empirical studies. Widening the scope of  algorithms. Much  research has concentrated on the main ‘paradigms’ of reference (Searle ). Early work on  treated reference as emphatically part of communication, as we have seen (§..: ‘First beginnings’). But after the refocusing that went on in the s, many  algorithms operate as if describing objects were a goal unto itself, instead of a part of communication. Still, when referring expressions occur in their natural habitat—in text or dialogue—then the reference game becomes subtly different, with factors such as salience and adaptation playing important (and partly unknown) roles. In these natural contexts, it is also not always necessary to identify a referent ‘in one shot’: in dialogue, and identification of the referent can be seen as a joint responsibility of both dialogue partners. Even in monologue, an entire sequence of utterances may guide a hearer towards the referent. In casual conversation, it is even unclear whether exact identification of the referent is a requirement at all, in which case all existing algorithms are wrongfooted. Reference in real life is also characterized by domains that are much larger and complicated than the ones usually studied (at least until they have been narrowed down by means of some salience metric): the set of people, for example that we are able to refer to in daily life is almost unlimited, and the properties that we can use to refer to them seem almost unbounded, including not only their physical appearance and location, but their ideas, actions, and so on. Evaluation challenges such as TUNA, , GREC, and GIVE have helped to bring the research community together, focusing on small domains and, predominantly, on simple types of referring expressions. We believe that it is time for evaluation studies to extend their remit and look at the types of complex references that more recent  research has drawn attention to. Such studies would do well, in our view, to pay considerable attention to the question which referring expressions have the greatest benefit for readers or hearers. One day, perhaps, all these issues will have been resolved. If there is anything that a survey of the state of the art in  makes clear it is that, for all the undeniable progress in this growing area of NLG, this holy grail is not within reach yet.

OUP CORRECTED PROOF – FINAL, 21/1/2019, SPi



     

A The order of authors was determined by chance; both contributed equally. Emiel Krahmer thanks The Netherlands Organization for Scientific Research (NWO) for VICI grant ‘Bridging the Gap between Computational Linguistics and Psycholinguistics: The Case of Referring Expressions’ (––). Kees van Deemter thanks the EPSRC’s Platform Grant ‘Affecting People with Natural Language’. We both thank the anonymous reviewers for their constructive comments, and Doug Appelt, Johan van Benthem, Robert Dale, Martijn Goudbeek, Helmut Horacek, Ruud Koolen, Roman Kutlak, Chris Mellish, Margaret Mitchell, Ehud Reiter, Advaith Siddharthan, Matthew Stone, Mariët Theune, Jette Viethen and especially Albert Gatt, for discussions and/or comments on earlier versions of this text. Thanks to Jette Viethen for her extensive  bibliography. Thanks to Harold Miesen for producing the image in Figure . and for allowing us to use it.

OUP CORRECTED PROOF – FINAL, 21/1/2019, SPi

        ......................................................................................................................

                  A Givenness Hierarchy theoretic approach ......................................................................................................................

    

. I

.................................................................................................................................. A robots become increasingly prevalent in our society, it becomes increasingly important to endow them with natural language capabilities. Natural language capabilities are especially important for robots designed to operate in domains such as eldercare robotics, education robotics, space robotics, and urban search-and-rescue robotics. In eldercare robotics and education robotics, it may simply be too cognitively burdensome for the target population to learn to interact with their would-be caregiving or educational assistants through some other modality. In space robotics and urban search-and-rescue robotics, it may be too physically burdensome for the target population to interact with their would-be assistants or rescuers, due to, for example, lack of gravity, or trapped limbs. In urban search-and-rescue environments, victims are also not likely to have the time or inclination to learn another control modality to interact with their would-be rescuers. It is thus important that robots operating in these and other domains be taskable through control modalities, such as natural language, that the general public is already familiar and proficient with. A crucial aspect of natural language communication is the ability to refer: the capability to which this volume is dedicated. Robots must thus be able to both understand so-called referring expressions, and generate them as well. In this chapter we will focus on the task of referring expression understanding. There are a number of unique challenges that present themselves to robots seeking to understand referring expressions due to robots’ status as situated agents: agents (entities capable of autonomously acting to achieve their own goals (Jennings )) that are embedded in an environment that is perceivable and manipulable by themselves and other agents with whom they can interact (Smith and Gero ).

OUP CORRECTED PROOF – FINAL, 21/1/2019, SPi



    

While a software entity operating within a non-situated domain such as text mining or document summarization may need to associate entities referenced in a text with previous portions of that text, a robot must instead associate entities referenced in dialogue with its own mental representations resulting not only from dialogue and inference, but also from interpretation of sensory data gathered by its perceptual systems. The robot’s knowledge of perceived entities will almost certainly be uncertain (as robots do not have perfect perception of the world) and incomplete (as robots cannot presume to be familiar with every entity in the world). In robotics, the problem of identifying what real-world entities are the referents of referring expressions goes by many names, including language grounding (Steels and Hild ), reference resolution (Popescu-Belis, Robba, and Sabah ), and entity resolution (Meyer ). While these names are sometimes used to denote the same concept, they carry different connotations. Specifically, reference resolution connotes association of a referring expression with a discretely represented entity, while language grounding further connotes grounding that discrete representation to continuously represented perceptual data (Harnad ). In our work, we are specifically interested in reference resolution that is domain independent (i.e., in which it is not assumed that referents will be of one particular type (cf. spatial reference resolution)) and open world (i.e., in which it is is not assumed that all candidate referents are perceivable or otherwise known at resolution time); assumption relaxations that are particularly important for realistic human–robot interaction scenarios. Imagine a search-andrescue scenario, in which a supervisor says to a robot: ()

The east wing needs to be evacuated. Please tell that to all personnel.

This example contains three referring expressions: ‘The east wing’, ‘that’, and ‘all personnel’. We propose that a domain-independent algorithm is needed to resolve these references, which are of different types, and that an open-world algorithm is needed as the agent should be able to understand the utterance even if it did not previously know that the building being discussed had an ‘east wing’. The need for open-world reference resolution algorithms underlies our decision to study only the reference resolution half of the language grounding problem. In closed-world resolution, at resolution time all entities are either perceivable or have been previously perceived, and thus continuous perceptual data is available to adjudicate the fitness of all candidate referents. Because this is not the case in an open world, it is important to develop reference resolution algorithms that do not require symbols to be ground to perceptual data, as do language grounding algorithms. Furthermore, we argue that it is important to focus on the two halves of the language grounding problem separately, as a robot is unlikely to generate subsymbolic representations for entities learned of through dialogue. In order to facilitate domaindependent open-world reference resolution, we have developed an algorithm which makes use of the Givenness Hierarchy (GH) (Gundel, Hedberg, and Zacharski ), which provides an elegant linguistic framework for reasoning about notions of reference in human discourse.

OUP CORRECTED PROOF – FINAL, 21/1/2019, SPi

  



In the next section (§.), we provide an overview of the GH and discuss previous GH-theoretic approaches to reference resolution. We then describe in §. our own GH-theoretic approach, the - algorithm, and suggest future refinements of our algorithm with respect to the theoretical commitments of the GH. In §., we go on to briefly survey other prominent approaches to reference resolution in robotics, and discuss how these compare to our approach. Finally, in §. we conclude with a discussion of possible directions for future work.

. T G H

.................................................................................................................................. The GH (Gundel et al. ) is comprised of six hierarchically nested tiers of cognitive status, as seen in Figure .. If a candidate referent is marked as having one of these statuses, the hierarchical nature of this framework means that it also has all statuses lower in the hierarchy. For example, a candidate referent that is familiar is also uniquely identifiable, referential, and type identifiable. It is possible that the candidate referent is also activated, or even in focus, but a higher status cannot be inferred from a lower status. Each level of the GH is ‘cued’ by a set of linguistic forms, as seen in Table . for English. For example, the second row of the table shows that when the definite ‘this’ is used, one can assume that the speaker assumes the referent of ‘this’ to be at least activated for their interlocutor.

In focus Activated Familiar Uniquely identifiable Referential Type identifiable

 . The Givenness Hierarchy.

Table . Cognitive status and form in the GH Cognitive Status

Linguistic Form

In focus Activated Familiar Uniquely identifiable Referential Type identifiable

it this, that, this N that N the N indefinite this N aN

OUP CORRECTED PROOF – FINAL, 21/1/2019, SPi



    

The GH is attractive to computational researchers not only because it suggests a clear mapping between linguistic form and cognitive status, but because, due to its focus on means of access rather than salience, each status evokes particular mnemonic actions (i.e., actions involving selecting or creating mental representations) upon an agent’s cognitive structures. When the linguistic form of an expression explicitly signals that its referent is type identifiable or referential (but not necessarily uniquely identifiable), this suggests the action of hypothesization: creating a new mental representation, and then selecting that representation as the target referent. When the linguistic form of an expression signals that its referent can also be uniquely identified (but is not necessarily familiar), this suggests either the action of hypothesizing a referent or selecting an existing referent from memory. When the linguistic form of an expression signals that its referent is also familiar, this suggests that the referent should be able to be found by searching through memory and selecting an existing representation. When the linguistic form of an expression signals that its referent is also activated or in focus, this suggests that the referent should be able to be found by searching through a subset of memory (the subset of activated entities and the subset of activated entities that are in focus, respectively) and selecting a referent from that subset. The GH can directly solve certain computational problems: to determine the cognitive status ascribed to a candidate referent, one need only check which forms explicitly encode which statuses on the GH in a given language (see also the Coding Protocol provided by Gundel et al. ()). And, when Speaker S uses linguistic form F to refer to entity E when speaking to hearer H, it is easy to determine the most restrictive status that H can rationally assume S to ascribe to E. For example, when S uses ‘it’, we can assume that S believes E to be in the subset of H’s memory that is in focus: any information that could not plausibly be in focus can be ruled out, as such an interpretation would not be possible given the cognitive status conventionally signaled by ‘it’; when S uses ‘this’, we can assume that S believes E to be at least in the subset of H’s memory that is currently activated. E may also be in the subset of those entities that are in focus, but we can not assume this; and in fact, it is unlikely that S believes E to be in that subset, as otherwise S could have used the more informative ‘it’. Furthermore, information that could not plausibly be in the activated subset of H’s memory can be ruled out, as such an interpretation would not be possible given the cognitive status conventionally signaled by ‘this’. However, within the GH framework, choices among referents that meet cognitive status restrictions are made through interaction of the GH with general pragmatic principles operative in language interpretation, such as Grice’s maxims or Relevance theory. As a result, the GH can only facilitate, but not directly produce solutions for the aforementioned computational problems of reference resolution (i.e., determining— when S uses linguistic form F when speaking to H—what entity E is most likely being referenced) and referring expression generation (i.e., determining—when S wishes to refer to E when speaking to H—what linguistic form F should be used (cf. Krahmer and

OUP CORRECTED PROOF – FINAL, 21/1/2019, SPi

  



Gesture Focus Visible Others

 . Chai’s Modified Hierarchy

Van Deemter ; Van Deemter )). As previously discussed, this chapter will focus on a GH-theoretic approach to the reference resolution problem. There have been several previous partial implementations of the GH (e.g., Kehler ; Chai, Prasov, and Qu ) for use in reference resolution algorithms, the most extensive of which is that presented by Chai et al. (). Chai et al. were interested in handling multimodal referring expressions within the context of multimodal user interfaces, and combined the GH with ideas from Grice’s theory of Conversational Implicature (Grice ) to produce the reduced four-tier hierarchy seen in Figure .. In this hierarchy, ‘Focus’ subsumes the GH’s in focus and activated tiers; ‘Visible’ subsumes the familiar and uniquely identifiable tiers; and ‘Others’ subsumes the referential and type identifiable tiers, but crucially, does not appear to be used. These three tiers are topped by a new ‘Gesture’ tier which specifically handles gestured-towards entities. In previous work (Williams, Acharya, Schreitter, and Scheutz ), we discuss how this reduced hierarchy—and its accompanying algorithm presented by Chai et al. ()—do not address all aspects of reference resolution found in typical human– robot dialogues. To summarize our concerns, the approach taken by Chai et al. () assumes complete certainty in the properties of entities, () appears to only handle references to objects, not locations, people, or less concrete entities (e.g., utterances), () operates under the closed-world assumption, () does not account for the GH’s preference for lower tiers over higher tiers (e.g., in ‘Could you repeat that’, ‘that’ is more likely to refer to an activated entity (e.g., the previous utterance) than an in-focus entity (e.g., the topic of the previous utterance)), () cannot differentiate between the tiers that they join together, and () may be prone to errors and inefficiencies when resolving complex noun phrases. To address our concerns, we presented the - algorithm (Williams et al. ), which we outline in the next section.

. T - 

.................................................................................................................................. The - reference resolution algorithm dictates how the referent of a referring expression should be searched for, given a memory model organized in a specific, hierarchical way that parallels the organization of the GH. In this section, we will first discuss the memory structure used in our approach. Next, we will discuss the

OUP CORRECTED PROOF – FINAL, 21/1/2019, SPi



    

between-structure processes by which - algorithm chooses which structures to search. We will then discuss the within-structure processes by which - selects suitable referents from a given structure. Finally, we will discuss the shortcomings of this approach which we are seeking to address in our current work. In all cases we choose to describe our approach at a high level, and do not provide pseudocode or optimization details. The interested reader should consult our previous work for such details, as cited throughout the section.

.. The GH-POWER memory model The memory model used by GH-POWER aligns well with Nelson Cowan’s conceptualization of working memory (Cowan ). According to Cowan, working memory and long-term memory are not disjoint structures. Rather, working memory can be regarded as the subset of entities in long term memory that are currently activated. Cowan further posits an additional substructure, the focus of attention, which is a subset of those activated entities that is limited in size to at most four elements, comprised of those items of which an agent is consciously aware. There is clearly a strong parallel between Cowan’s Focus of Attention ⊂ Set of Activated Entities ⊂ Long Term Memory structures and Gundel’s In Focus ⊂ Activated ⊂ Familiar statuses, and observing this connection will facilitate understanding the connection between our own memory structure and the statuses of the GH. Our approach consists of four hierarchical data structures: the Focus of Attention (FOA), Set of Activated Entities (ACT), Set of Familiar Entities (FAM) and Long Term Memory (LTM). These four data structures are hierarchically organized such that FOA ⊂ ACT ⊂ FAM ⊂ LTM. At the computational level of analysis (Marr ), the FOA, ACT, and LTM data structures are identical to Cowan’s three memory structures. But in a robot architecture, all of a robot’s knowledge is not typically located in a single, monolith knowledge base. Instead, it may be distributed across a set of knowledge bases that may be located on different machines, may use different knowledge representation schemes, and may have different ways of accessing and modifying the knowledge contained within them. Thus, at the algorithmic level, our LTM data structure is really a set of domain-specific distributed, heterogeneous knowledge bases. Because LTM is not a single coherent knowledge base, the FOA and ACT also must differ at the algorithmic level; instead of being literal subsets of the mental representations distributed across LTM, the FOA and ACT instead contain memory traces that allow access to certain of those mental representations. Note that these three structures are not intended to serve as the agent’s actual cognitive structures; instead, they serve to model what an interlocutor might believe to be in those structures, and thus as a model of common ground. Finally, for the sake of convenience and efficiency, we introduce the FAM structure, a minor point of departure from both the GH and Cowan’s model of Working

OUP CORRECTED PROOF – FINAL, 21/1/2019, SPi

  



Memory, which we make for practical rather than theoretical reasons. FAM contains memory traces for entities in LTM that are likely to be referenced, such as entities mentioned at some point in the robot’s current dialogue, recently visited locations, and recently visited objects, including all entities in ACT (and by extension, in the FOA). Because searching all of LTM is potentially expensive, when LTM needs to be searched for an entity that matches some criteria, that search is preempted by a search of FAM: if a match can be found there, LTM need not be searched. To summarize, our model consists of four hierarchically nested data structures: a distributed LTM data structure containing mental representations of known entities, and three smaller data structures that contain memory traces allowing fast access to entities in LTM (i.e., FOA, ACT, and FAM). These three data structures are populated periodically (e.g., after an utterance is processed) according to rules inspired by the GH Coding Protocol. In the next two sections, we will describe how these structures are used during reference resolution: in §.. we discuss how the linguistic form of a referring expression is used by the - algorithm to determine which of these structures to examine; in §.., we discuss how - chooses whether a particular candidate referent within one of those structures is the target referent.

.. Between-structure processes The GH alone does not specify how cognitive structures are selected for perusal during reference resolution. For example, suppose Speaker S uses the pronoun‘that’ to refer to entity E when speaking with Hearer H. The GH suggests that H can assume that S assumes that E is at least in H’s ACT, and thus may or may not also be in H’s FOA. Several strategies could be used to search ACT and the FOA. The agent could consider entities in the FOA, then out-of-focus entities in short-term memory (a top down approach), or she could consider out-of-focus entities in ACT, then in-focus entities (a bottom up approach). While some previous approaches (e.g., Chai et al. ) have used a global top-down approach, this may violate certain predictions of the GH. For example, the Givenness Hierarchy framework (i.e., the GH when working in conjunction with general cognitive principles such as Grice’s maxim of Quantity) suggests that in the example above, while the referent of ‘that’ could be assumed to be in H’s FOA, it is more likely to be assumed to be in H’s ACT but not in H’s FOA, as otherwise S could have used ‘it’ to refer to the referent. If a purely top-down approach is used, this effect may not be captured. On the other hand, consider the utterance ‘Pick up the box’. The bottom-up approach would inappropriately prioritize inactive boxes from LTM over an activated box in front of the listener. Since neither a purely top-down or purely bottom-up approach seems adequate, we developed a hybrid approach, in which a unique search strategy is used for each GH tier. These strategies, refinements of those presented in (Williams et al. ) are seen in Table .. In that table, FOA denotes a search through memory

OUP CORRECTED PROOF – FINAL, 21/1/2019, SPi



    

Table . Search plans for complete GH Level

Search Plan

In focus Activated Familiar Uniquely identifiable Referential Type identifiable

FOA ACT ! FOA ACT ! FOA ! FAM ! LTM ACT ! FOA ! FAM ! LTM ! HYP ACT ! FOA ! HYP HYP

traces found in the FOA; ACT denotes a search through memory traces found in ACT but not in the FOA; FAM denotes a search through memory traces found in FAM but not in ACT. LTM denotes a search through all of LTM; HYP denotes hypothesization. We will now explain the rationale for each strategy.

... In focus In the case of an ‘in focus’ cueing form (e.g., ‘it’), we only search the FOA, as it would be otherwise inappropriate to use such a form.

... Activated entities In the case of an ‘activated’ cueing form (e.g., ‘this’), search is expanded to include outof-focus entities in ACT. For the reasons discussed above, we proceed bottom-up, first searching the out-of-focus entities in ACT, then searching the FOA. However, this process is modified in the case of ‘this N’, as we discuss below.

... Familiar entities In the case of a ‘familiar’ cueing form (e.g., ‘that N’), search is expanded to include all entities in memory. As it is inappropriate to prioritize entities in LTM over those in ACT, we still perform our search through ACT and the FOA first, and then move on to search through LTM. As previously discussed, we first search through FAM, the subset of most probable referents in LTM (not including those referents also found in ACT), and only search all of LTM if this search fails, using the - algorithm described in (Williams and Scheutz ). - is a distributed extension of the cognitive model proposed in (Williams and Scheutz a) and computationalized in (Williams and Scheutz b). This algorithm has two main features relevant to . First, it is able to simultaneously resolve all parts of a complex definite description. The second feature will be discussed in the following subsection.

... Uniquely identifiable In the case of a ‘uniquely identifiable’ cueing form (e.g., ‘the N’), search is extended to allow for the possibility that the speaker is referencing a previously unknown entity. This begins by searching through the four tiers of the - memory model, as

OUP CORRECTED PROOF – FINAL, 21/1/2019, SPi

  



performed with familiar entities. However, when searching through LTM, we take advantage of -’s second important feature -’s ‘hypothesization mode’. When run in this mode, if - is unable to find a mental representation that satisfies all semantic criteria of a definite description, it attempts to find a subset of that description that it can successfully resolve, and automatically hypothesizes representations for remaining entities.

... Referential Gundel et al. () suggest that the indefinite form of ‘this N’ (as in ‘This dog I saw was enormous!’) cues the ‘referential’ tier,¹ resulting in the hypothesization of a representation. As a simplification (i.e., so that we do not need to decide whether each use of ‘this N’ is definite or indefinite), - deals with both forms at the referential tier. To do so, we begin with the standard ‘activated’ search strategy (i.e., a bottom-up search starting from ACT), and hypothesize a representation only if this search fails. We acknowledge that there may be cases in which this does not produce the correct behavior. For example, if one says ‘This dog I saw was enormous!’ while standing in front of a dog, ‘This dog’ may be incorrectly resolved to the co-present canine.

... Type identifiable In the case of a linguistic form that only cues the Type Identifiable tier (e.g., ‘a N’), we immediately hypothesize a representation in the same way as is performed in the previous subsection. Note that this does not imply that the robot does not yet have a representation for the intended referent. For example, suppose the robot is looking at a box, and its interlocutor says to it remotely, ‘You should see a box: bring it to me.’ In this case, the robot’s interlocutor actually intends to refer to a particular box, and the robot in fact already knows of this box. Even in such a case, we still create a new mental representation for a new box. It will be up to subsequent processing stages to recognize the meaning of the sentence, find the two representations, verify that they match, consolidate them into a single representation, and of course, bring the box to the interlocutor.

... Complex referring expressions The GH framework also does not specify how to resolve syntactically complex referring expressions, that is, referring expressions containing multiple referents described in relation to each other, such as those in example (): ()

Scene: A table upon which sits a large green block and a large blue block (towards the front of the table), and a greenish-yellow block on a bluish-purple block (in a far corner of the table). a. Pick up the green block that is on the blue block. b. Pick up the one on the blue block.

¹ In fact, this form, which is only used colloquially, is the only form in English that overtly cues the referential status.

OUP CORRECTED PROOF – FINAL, 21/1/2019, SPi



    

Table . Sample joint search plan table Y ACT ACT FOA FOA HYP HYP

X ACT FOA ACT FOA ACT FOA

Chai et al. () resolve references of this sort using a greedy algorithm in which locally optimal choices are sequentially made for each sub-expression. However, in cases like that seen in example (a), this is likely to incorrectly resolve whichever referential expression is considered first, due to the decreased salience, prototypicality, and proximity of the targets. Greedily resolving example (b) will likely be even less successful due to the underspecification of ‘the one’. We would thus argue that syntactically complex referring expressions should not be considered greedily in a GH-theoretic reference resolution algorithm. How, then, should search plans (i.e., from Table .) for an expression’s constituent parts be jointly examined? We decided to handle this problem by ‘crossing’ the search plans for the constituent parts, that is, considering all possible combinations of search plans sorted in search plan order. For example, crossing ACT ! FOA ! HYP with ACT ! FOA yields Table .. The rows of this table are successively examined until a sufficiently probable solution is found or the table is exhausted. Here, for example, - would begin by searching for a pair of memory traces contained in ACT which fit the given description. If no such pair can be found, - will proceed to the next line of the table, and search for a pair of memory traces, one from ACT and one from the FOA, which fit the given description, and so forth. Two decisions were made in designing this subroutine. First, rows are considered in left-to-right order. For example, when searching for a pair of referents on the second line, - would first try to find candidate referents from ACT to associate with Y, and then try to find candidate referents from FOA to associate with X given the set of restricted candidates for Y. Second, the action of hypothesization (denoted HYP) is postponed until the search process is successfully terminated; a new representation should only be generated if sufficiently probable referents are found for all other entries in a row, halting the search process. For example, when line five is considered, - will begin by associating Y with a dummy referent ‘?’. A new representation will be created for this referent if and only if a sufficiently probable referent can be found in ACT to associate with X.

OUP CORRECTED PROOF – FINAL, 21/1/2019, SPi

  



.. Within-structure processes The GH does not specify how candidates are selected from within cognitive structures during reference resolution. Despite what is often assumed (cf. Brown-Schmidt, Byron, and Tanenhaus ), Gundel () state that the GH is not a hierarchy of salience or accessibility, and that it is necessary to model salience independently of tier of cognitive status. We will now describe how the proposed model handles degree of salience and uncertainty, and how these measures are used to select candidates.

... Focus of attention and activated entities In order to account for salience without relying on, e.g., a dedicated gestural tier (cf. Chai et al. ), - uses a multimodal salience score to assign a ‘degree of activation’ to entities contained in the FOA and ACT. The entities returned by the assess methods associated with the FOA and ACT structures are then the set of all sufficiently probable entities within those tiers, ordered by activation such that the most salient candidate will be chosen if multiple are available.

... Familiar entities and long-term memory In the proposed model, the Set of Familiar Entities is equivalent to a ‘highly salient’ LTM cache; we would argue that the ‘Familiar’ and ‘Uniquely Identifiable’ tiers can be viewed as different means of accessing the same structures, with different worst-case conditions. This is consistent with Gundel’s claim (Gundel ) that: “[F]orms that encode cognitive status on the GH are not markers of degree of accessibility. Rather, they provide procedural information about manner of accessibility, how and where to mentally access an appropriate representation.” The entities returned by the assess method associated with the FAM are its sufficiently probable entities, ordered in reverse chronological order; the entities returned by the assess method associated with LTM are its sufficiently probable entities, ordered in decreasing order of likelihood.

.. Discussion - was able to resolve the majority of references occurring in a corpus of human–human and human–robot team tasks (Williams et al. ). While, there were a number of cases that - was unable to handle, it was able to capture several aspects of the GH missing from previous GH-theoretic approaches. Consider, for example, the following example presented by Gundel (): ()

a. Alice: I failed my linguistics course. b. Bob: Can you repeat that?

OUP CORRECTED PROOF – FINAL, 21/1/2019, SPi



    

Before resolving ‘that’, the referent of ‘my linguistics course’ should be in the agent’s FOA, while the utterance itself should be in the agent’s ACT, but not in the agent’s FOA since, as Gundel et al. () note, speech acts are activated, but not brought into focus just by being uttered. Gundel et al. () suggest that if Bob had meant to refer to the course, he would have used it instead of that, because ‘it’ explicitly picks out an in focus referent, whereas ‘that’ only signals that the referent is activated and therefore could be in focus, and thus the course should be dispreferred to the sentence itself. This effect is captured through -’s between-structure processes: When ‘that’ is used, ACT is first checked; and because the utterance is in ACT, it is chosen. FOA is not even examined, because any options residing therein should be dispreferred. However, consider example (): ()

Scene: A table on which sits a black box and a white box a. Bob: Look at the white box b. Bob: Pick that up

Before resolving ‘that’, the white box should be in the agent’s FOA, and the black box is likely to only be in the agent’s ACT, as depicted in Figure . Following the logic of example (), if Bob had meant to refer to the white box, he could have used ‘it’ instead of ‘that’, and thus the white box should be dispreferred. Yet while ‘it’ would have been more natural in referring to the white box, choosing the white box as the referent of ‘that’ is clearly not wrong and probably preferred in this situation in the absence of any gesture indicating a shift in attention. In this scenario, - errs for two reasons. First, it treats hierarchical preference as absolute, whereas dispreferred entities should be just that: dispreferred, not removed from consideration. Second, - does not take conversational relevance into account. These factors were initially overlooked because the GH does not specify how relevance factors influence search. - checks whether resolution

FOA white box ACT black box interlocutor table last utterance FAM […] LTM […]

 .. Contents of -’s cognitive structures during hypothetical algorithm run. Structures are arranged to depict their hierarchical nature (i.e., an entity in one structure is also in all lower structures). [ . . . ] indicates the wide variety of entities contained in the set of familiar entities and in long term memory which are not immediately relevant to this example.

OUP CORRECTED PROOF – FINAL, 21/1/2019, SPi

  



candidates are suitable, that is, whether they satisfy all described properties, and only moves on to consider entities in another cognitive structure if no resolution candidates in the current structure are deemed suitable. In this case, however, this is insufficient. In order for - to perform correctly in this scenario, it should recognize not only that both boxes are suitable, but that the white box is conversationally more relevant than the black box; there is no clear reason why the agent would be asked to look at the white box and then pick up the black box. In order to address this issue, - should operate in the following way: when ACT is examined, the low conversational relevance of the black box should be noted. This should result in search extending to the FOA while retaining the black box as a resolution candidate. The white box should then be selected using an equation that takes relevance, suitability, and other factors into account. To be precise, at least three factors must be considered in the within-structure processes of future versions of -: () suitability (i.e., the agent’s certainty that a candidate holds all described properties), () relevance, (i.e., the agent’s certainty that reference to a candidate would not violate, for example, Grice’s Maxim of Relevance (Grice )), and () common-sense judgments (here, e.g., the agent’s certainty that a candidate can be picked up). Note that each of these factors may be used differently: while a candidate must score highly on all three factors for the search to cease, only low suitability will likely result in a candidate’s complete removal from consideration. Furthermore, to respect the Theory-of-Mind considerations of the GH, this process must consider the extent to which the speaker would have been cognizant of each of these factors. We are currently in the process of developing a new algorithm that takes these factors into account.

. R 

.................................................................................................................................. In the previous sections, we presented our own approach to reference resolution, and described how it compared to other GH-theoretic approaches (which had been developed within the context of multimodal user interfaces). In this section, we will discuss how our approach relates to other approaches to reference resolution within robotics. While our approach is open-world in nature, there has been very little work in that area², and thus we will be primarily discussing closed-world approaches to reference resolution. We will also not discuss approaches that assume that a unique name (i.e., a rigid designator) is always provided (e.g., Khayrallah, Trott, and Feldman ). Reference resolution has been a topic of interest in robotics for nearly half a century, beginning with Winograd’s SHRDLU system (Winograd ). In the SHRDLU ² Cf. work in open-world directive grounding (Matuszek, Herbst, Zettlemoyer, and Fox ; MacMahon, Stankiewicz, and Kuipers ), in which natural language utterances are translated directly into action sequences, bypassing the need to ground constituent noun phrases.

OUP CORRECTED PROOF – FINAL, 21/1/2019, SPi



    

system, a simulated robot could be issued an array of natural language commands in order to move blocks around in its simulated environment. Winograd introduced a procedural semantics view of natural language understanding in which each lexical item was associated with a short procedure to be executed: adjectives and nouns, for example, were associated with procedures which considered each entity in the scene and decided whether or not that entity fit the property denoted by the lexical item. SHRDLU was also able to perform anaphora resolution: anaphoric lexical items were associated with procedures that considered the plausibility of anaphoric reference to each entity in SHRDLU’s world, giving preference to elements considered to be ‘in focus’ (see also Mitkov ). Decades later, SHRDLU continues to inspire researchers. Gorniak and Roy, for example, employ a similar approach (Gorniak and Roy ; Roy, Hsiao, Mavridis, and Gorniak ). In their work, a simpler approach to anaphora resolution is used, but they make important steps forward in other respects: whereas SHRDLU’s procedural attachments consulted a knowledge base populated with hand-assigned symbolic properties, Gorniak and Roy’s procedural attachments effect the incremental and greedy application of composable visual models. That is, properties of objects are assessed by consulting the continuous perceptual features of those objects, thus grounding internal symbols to the physical world. This thus represents a solution to the full language grounding problem. In our work, we only address the reference resolution half of the language problem, leaving the symbol grounding half to POWER’s distributed knowledge bases. But unlike Gorniak and Roy’s approach, - is domain-independent, and operates in uncertain and open worlds. Just as Gorniak and Roy incrementally execute procedures associated with particular lexical items, Kruijff et al. (Kruijff, Lison, Benjamin, Jacobsson, and Hawes ) incrementally employ a set of comparators that can make true/false judgments as to whether certain entities satisfy certain properties. But while Gorniak and Roy focus on grounding, do not address deixis, and only narrowly address anaphora, Kruijff et al. () treat grounding as a separate process, and address many aspects of deixis and anaphora. Furthermore, Kruijff et al. () use a central symbolic knowledge base that is informed by subsymbolic perceptual systems, which is not dissimilar from our own decision to use a set of distributed knowledge bases. Several other researchers have used knowledge-based approaches similar to Kruijff et al. (), differing in the way in which their knowledge bases are queried. Lemaignan et al. () use semantic parsing techniques to translate utterances into knowledge base queries. These authors handle anaphora and deixis to a limited extent: anaphoric expressions are replaced by the last entity in the dialogue history that matches animacy and gender constraints; deictic expressions are resolved to the most recent focus of simultaneous eye gaze and gesture. Like that of Kruijff et al., this approach does not handle uncertain or open worlds. Zender et al. () take a similar approach, but apply their approach to the spatial domain of large-scale topological spaces (e.g., rooms and hallways) rather than the

OUP CORRECTED PROOF – FINAL, 21/1/2019, SPi

  



object domain used by all previous approaches (Zender, Kruijff, and Kruijff-Korbayová ). Zender et al. also use semantic parsing techniques to translate utterances into knowledge base queries. As Zender et al. () are specifically targeting references to large-scale locations, they did not attempt to handle deixis and eye gaze. They do, however, handle anaphora through a dedicated co-reference resolution³ pre-processing step, similar to other approaches we will examine. Meyer, for example, performs co-reference resolution to resolve anaphora; but this step is tightly coupled with his Markov Logic theoretic reference resolution system such that anaphoric and other referential expressions can be resolved as part of a joint model (Meyer ). Deictic expressions are not handled by this approach, and referents are restricted to objects. A slightly different approach is taken in recent work by Chai and colleagues (Fang, Liu, and Chai ; Liu, Fang, She, and Chai ; Chai et al. ) (in work distinct from their GH-theoretic approach discussed above). As dialogue unfolds, Chai et al. () build up a graph representing the relations between discussed entities. When each utterance is heard, Chai et al. () use a graph matching algorithm to find the best partial overlapping region between the dialogue graph and a separate vision graph representing the relations between currently observed entities. Anaphora is handled in this approach by a co-reference resolution pre-processing step; and while deictic expressions are not discussed in this work, Chai et al. show in earlier work how gestural information can be integrated into their dialogue graph structure (Chai, Hong, and Zhou ). Furthermore, while more recent publications do not discuss it, older work suggests that they are able to handle uncertain properties (Fang et al. ). The previous approaches we have discussed assume that sensors provide straightforward true-or-false judgments on whether entities in the world have certain properties; but in realistic situated interactions, an agent may not always be certain whether certain entities in the world hold certain properties. Chai et al.’s approach begins to address this, by incorporating extent of compatibility into their graph-matching scoring functions. While Chai et al.’s approach is, like -, able to handle uncertain properties, it is unable to handle open worlds. Another approach that begins to address the uncertain nature of reality is the Semantic Fields work presented by Fasola and Matarić (). In that work, reference resolution at the lowest level is quite simple: when nouns are used, a knowledge base is examined to determine whether it contains a unique entity with that label. If so, that entity is chosen as the target referent. If not, their approach attempts to disambiguate using semantic field models of spatial prepositions. This approach does not handle deixis, and seems to be restricted to operation in environments in which knowledge of objects is provided a priori. It does handle some anaphoric expressions however, by ³ A co-reference resolution procedure attempts to identify whether a referring expression refers to the same referent as a previous referring expression. If so, the new RE is added to the previous RE’s co-reference cluster; a unique identifier is used to identify the presumed referent of each co-reference cluster during subsequent language processing steps.

OUP CORRECTED PROOF – FINAL, 21/1/2019, SPi



    

choosing as the target referent the most recent entity that matches gender and animacy constraints (Fasola and Matarić ). While the reference resolution portion of this work may not have all the capabilities of other approaches, and is limited to closed worlds, we believe it would be interesting to integrate the preposition models used by this approach into the - framework. It is important to recognize that  is not necessarily incompatible with all of the examined approaches. Many of these approaches present innovative ways for grounding or evaluating certain properties, and as long as these methods can be adapted to produce probability values, they can be integrated into the - framework. In contrast to the approaches examined thus far stand a number of recent Bayesian approaches, which seek to more formally handle uncertainty. Kennington and Schlangen present an incremental Bayesian model: as each word in a sentence is heard, the probability of each entity in a scene being the target referent is modulated based on learned probabilistic models that associate lexical units with observable properties (Kennington and Schlangen ). This approach handles deixis and gaze by linearly combining the probability of reference given the utterance, the probability of reference given the speaker’s gaze, and the probability of reference given the speaker’s gestures. Anaphora is handled by attributing a ‘selected’ property to entities which become selected through dialogue: pronouns are statistically associated with this property through the same learning process used for other lexical units. While this approach does handle uncertainty with respect to the relationship between words and features, it does not handle uncertainty with respect to the relationships between features and objects: each object in the scene has a set of properties which are known a priori to be true of that object. This approach is much more cognitively plausible than the other approaches examined, including our own; and with this cognitive plausibility come a number of computational advantages, the foremost being the increased speed of processing inherent to the incremental approach. But this approach is unable to handle uncertain and open worlds—a limitation that is not entirely shared by the other Bayesian approaches we will examine. In the Bayesian Generalized Grounding Graph approach taken by Tellex, Kollar, et al. (), utterance structures are used to instantiate probabilistic graphical models where certain nodes are associated with certain words in the utterance (Tellex et al. ). This approach operates in a manner similar to our own (cf. Williams and Scheutz b). Deixis and gaze do not appear to be handled in this approach, and like the previous approach, the uncertainty of objects’ properties is not represented. Anaphora is handled through a co-reference resolution pre-processing step (Tellex et al. ), and it appears to be usable in an incremental fashion (Manek and Tellex ). This framework does improve upon previous approaches, however, in an important way: it is not limited to handling just objects, or just locations, but handles both. It appears to handle references to any physically extant entities located at particular points in space. Furthermore, recent work building on this framework has begun to address open world resolution. The work by Duvallet et al. () allows a robot to hypothesize a new object described with respect to another object. This

OUP CORRECTED PROOF – FINAL, 21/1/2019, SPi

  



hypothesization is, however, limited to spatially situated objects. Through -, we are interested in the hypothesization of not only spatially situated objects, but other entities such as agents and locations. Similarly, Chung, Propp, Walter, and Howard () extend Generalized Grounding Graphs to produce the Hierarchical Distributed Correspondance Graph approach, which uses utterance structures to instantiate probabilistic graphical models of a similar form. While this approach is more nascent, and thus has not been incrementalized and cannot yet handle deixis, gaze, or anaphora, it improves on previous approaches in that it begins to pay attention to what entities are considered. Considering all entities in the world when performing reference resolution may be feasible when you need only consider a small number of entities in a visual scene; but it will likely be computationally intractable in larger, more realistic settings. In Chung et al.’s approach, only the set of entities matching the correct type indicated by each noun phrase are considered as possible referents for that noun phrase. We believe that the GH provides a powerful alternative—in the majority of cases, - need only consider the limited subset of entities in ACT and the FOA. Like Chung et al., we are interested in restricting the search space considered during reference resolution—but through -, we are able to do so under uncertainty, and in a context-sensitive manner. Finally, Matuszek, Fitzgerald, Zettlemoyer, Bo, and Fox () present an approach similar to both those of Tellex and Kollar, Gorniak and Roy (), and Kennington and Schlangen (). In this approach, a semantic parser is connected to a set of visual classifiers used to identify objects. As with Kennington and Schlangen, deixis and gaze are handled through a linear combination of probabilities. Unlike the majority of previous approaches, however, Matuszek et al.’s () approach is able to represent the robot’s uncertainty in the properties of the objects it detects, based on classifier confidences. This is an important step towards enabling operation in natural, realistic settings. Like the majority of previous approaches, however, Matuszek et al.’s approach is limited to handling references to objects, and is restricted to operation in a closed world.

. C

.................................................................................................................................. In this chapter, we began by outlining the language grounding problem, and its constituent parts: reference resolution and symbol grounding. We then described -, in which the task of symbol grounding is considered the purview of the distributed heterogeneous knowledge bases that comprise long term memory, and in which the task of reference resolution is performed by a GH-theoretic algorithm that makes use of the information distributed across those knowledge bases. Next, we discussed some theoretical concerns which provide motivation for future work, and discussed - in relation to other approaches to reference resolution within robotics. In addition to the modifications proposed in previous sections, there are a number of directions for future work within our framework. Our algorithm should be extended to

OUP CORRECTED PROOF – FINAL, 21/1/2019, SPi



    

handle references to sets, and references to non-discrete entities (e.g., vague regions of space). We should integrate common-sense affordance-based reasoning capabilities (Chambers, Tanenhaus, and Magnuson ) and incrementalize and parallelize our algorithm, to come in line with psycholinguistic literature (Eberhard, SpiveyKnowlton, Sedivy, and Tanenhaus ), similar to previous work from our lab (Scheutz, Eberhard, and Andronache ) and others (Kennington and Schlangen ). We are also interested in using this approach to generate referring expressions in a GH-theoretic manner. And we are interested in more deeply integrating - with other components within our architecture (e.g., Vision Processing) so that our within-structure processes can better account for eye gaze and gesture. Finally, and more generally, it is our hope that the framework discussed in this paper will serve as a jumping-off point for much further study of the interaction of language, memory, and attention, not only for algorithmic purposes in the development of integrated systems, but for cognitive modeling purposes as well.

A This work was in part funded by grant N--- from the US Office of Naval Research.

OUP CORRECTED PROOF – FINAL, 21/1/2019, SPi

        ......................................................................................................................

    Complications of information sharing ......................................................................................................................

  

. I

.................................................................................................................................. W Artificial Intelligence was first conceived, the construction of computer programs that understand and produce language was paramount among its aims (Turing ). As soon as the first modest successes in these areas were achieved, referring expressions () were seen as centrally important (e.g., Winograd ). Over time, more and more computational work began to focus on reference itself, often emphasizing the generation—rather than the interpretation—of referring expressions (Davey ; Appelt ; Dale and Reiter ; van Deemter ; Stone et al. ; Krahmer et al. ). In this Chapter, we examine some fundamental limitations that algorithms in this area—known as Referring Expressions Generation () algorithms—tend to have, and discuss ways in which these limitations can be addressed.¹ We focus on issues that arise from the situation in which the  is uttered. Some readers may want to jump immediately to our discussion of concrete communicative situations in §.–§.; others may be interested in reading about the  task and the role of Information Sharing in it. These days,  algorithms are starting to be utilized by practically useful computer systems that convert data into ordinary language for practical applications (Reiter and

¹ I am indebted to Albert Gatt, Rodrigo Gomes de Oliveira, Imtiaz Hussain Khan, Agnieszka Konopka, Emiel Krahmer, Roman Kutlák, Judith Masthoff, Chris Mellish, Ivandré Paraboni, Ehud Reiter, Graeme Ritchie, and Yaji Sripada for contributions to this chapter.

OUP CORRECTED PROOF – FINAL, 21/1/2019, SPi



  

Dale ; more recently, Perera and Nand ; Gatt and Krahmer ).² In this chapter, we will take a different perspective, regarding  algorithms as putative models of human language production (van Deemter ), and discussing referential scenarios that put such models to the test. In the s, a perspective on  came to the fore that I will call the classic model. The essence of this model will not lie in any particular algorithmic strategy, but in a set of assumptions concering the aim of these algorithms, discussed in §.. In recent years, limitations of this model have attracted attention. I will point out some of these limitations and discuss attempts to get rid of them. In doing so, I will borrow freely from Part IV of van Deemter (). I focus on a family of issues that arise not from the logical or linguistic structure of the  but from the situation in which the  is uttered. The structure of this chapter is as follows: §. offers a theoretical reconstruction of the classic model and sketches the limitations of the model with a broad brush. Later sections turn to specific limitations: §. discusses breakdowns of shared information; §. discusses issues arising from large domains; §. concerns situations in which it is not feasible to identify the referent with precision; and §. discusses situations in which even the aims of reference amount to something other than the identification of a target referent. We focus on reference ‘in one shot’: pronouns and collaboration between speakers and hearers are left aside. We focus on the semantic content of a referring expression, leaving words and syntactic structure aside.

. T C M  

.................................................................................................................................. Let me start by proposing a rational reconstruction of the perspective on referring that came to be accepted widely in the s. At its heart lies the notion of Information Sharing, the process whereby speaking turns private information into information that is common knowledge between the speaker and the hearer. Information Sharing has not often been emphasized in connection with , but later sections will bring Information Sharing sharply into focus.

.. Information Sharing Philosophers of language have long debated the analysis of sentences that contain a referring expression (). Bertrand Russell famously analyzed ‘The so-and-so is P’ as ² The web pages of the following companies demonstrate the state of the art in practical Natural Language Generation: Arria (http://www.arria.com), Narrative Science (https://www.narrativescience. com), and Yseop (http://yseop.com/EN/home).

OUP CORRECTED PROOF – FINAL, 21/1/2019, SPi

   



making two statements, namely () ‘There exists exactly one so-and-so’, and () ‘All so-and-so’s are P’. Although this analysis makes some excellent predictions about the interpretations that complex sentences can have, many linguists have come to prefer the analysis of Peter Strawson, which regard ‘The so-and-so is P’ as consisting of two parts, which contribute to communication in different ways: the referring expression, ‘The so-and-so’ makes use of information that is shared (also: given), the remainder of the sentence presents new (also: privileged) information. This analysis—which can be traced to a proposal by Strawson ()—can be discerned at the heart of many computational accounts. Suppose I have new information about an animal a: it is in a cage. If I want to communicate this information to you, and if our shared information is represented in the Knowledge Base depicted in Table ., I can say ‘the Kenyan lion is in a cage’, exploiting our shared knowledge. After this utterance, my privileged information has shrunk, but our shared information has increased because the fact that a is in the cage is now part of our shared information. Crucially, a very different  may be chosen, for example, ‘The animal that is a lion, weighs kg, and has injuries in its teeth’ to name just one. Each  has its own peculiarities, for instance the latter one might be wrongly taken to imply that the lion has no other injuries. Choosing the most natural, or perhaps the most effective,  is what  is concerned with. Most authors on  have written about Information Sharing as hinging on what knowledge the speaker knows the hearer to possess (e.g., Dale and Reiter ). Although this is an important part of it, theoreticians have argued convincingly that information p is only truly shared between a set of agents (in which case it is also said to be ‘common knowledge’, or ‘in common ground’) if these agents know that p and that p is shared. (Note the recursion inherent in this assertion.) To borrow an example from Clark and Marshall (), suppose I utter the  ‘the movie showing at the Roxy tonight’. If you and I believe this is movie x, and I believe you believe it is x, but you believe that I believe it is movie y, then this will make you misunderstand me, because you will think I’m referring to y. A proposition p is only shared between you and me if I know that p, you know that p, I know that you know that p, you know that I know that p, and so on, using epistemic embeddings of arbitrary depth. Many researchers have contributed to our understanding of Information Sharing (Beaver ). Philosophers, for example, have argued that a felicitous utterance should normally be consistent with shared information while also adding new information to it (Stalnaker ; Lewis a). More importantly from the perspective of Table . Information shared by speaker and hearer Identifier

Species

Origin

Weight

Injuries

a b c

lion lion tiger

Kenya China China

kg kg kg

paws, teeth paws back

OUP CORRECTED PROOF – FINAL, 21/1/2019, SPi



  

the present chapter, the extent to which speakers and hearers are able to determine what information is shared between them is the subject of the egocentricity debate between psychologists. On one side are those who emphasize this ability (e.g., Clark and Wilkes-Gibbs ; Brennan and Clark ); on the other side are those whose experiments have sowed doubt about our ability to reason about other minds (Horton and Keysar ; Keysar et al. ). The outcome of this debate is still uncertain (e.g., Heller et al. ). The truth is likely to lie somewhere in the middle, with reasoning about other people’s knowledge playing a crucial role despite sometimes being compromised (e.g., under time pressure; for children; and possibly for people on the autism spectrum (Nadig et al. )). Let’s see how the reference task can be formulated according to the classic model. In line with linguistic-philosophical tradition, we call the set of elements that have a property P the denotation of P, abbreviated ⟦P⟧. Given is a finite domain involving a set M of entities that is shared information between the speaker and the hearer. M contains an element r∊M, the target referent. Given are also one or more other domain elements, the distractors; also, a finite set. P of atomic properties, whose denotation is a subset of M, and whose denotation is shared information between the speaker and the hearer.  can now be defined as follows, where the enterprise has one of two aims (a or b):  The classic  task. If there exist properties P₁, . . . , Pn in the set P such that the intersection of their extensions equals r (more precisely, ⟦P₁⟧∩ . . . ∩⟦Pn⟧={r}), then  needs to find such a set. The algorithm needs to make sure that (a) the properties P₁, . . . , Pn are as similar as possible to properties found in s produced by human speakers, or that (b) the properties permit the generation of an  that is optimal for human recipients.

The formula ⟦P₁⟧∩ . . . ∩⟦Pn⟧ is the Logical Form of the . It will be convenient to use the terms ‘’ and ‘description’ to denote either these Logical Forms or the English noun phrases that express them; sometimes we will speak explicitly of a distinguishing description (Dale ), to make explicit that the description refers unambiguously to its intended referent. Note that, without the (a)/(b) addition, the above-defined  task would be trivial: it could be met by intersecting (i.e., logically conjoining) all the properties true of r. Task (a) is a classic computational modeling task, whose aim is to mimic human language production; task (b) is what the designer of a Human–Computer Interface may have in mind, aiming to generate  that can be understood quickly and reliably by users.

.. Cracks in the classic model Algorithms using the classic model were tested extensively on simple domains. In the most popular testing paradigm, human speakers are shown a number of visual scenes,

OUP CORRECTED PROOF – FINAL, 21/1/2019, SPi

   



where one thing in the scene (e.g., a red cup) is marked as the intended referent. Each speaker is asked to produce an  that refers to the intended referent. The result is a data-text corpus in which each  is coupled with a referent. Next, the same domains, and the same intended referents, are given to a  algorithm, which then generates an  that aims to refer to the intended referent (e.g., Gatt and Belz ; Viethen and Dale ; van Deemter et al. ). In simple situations, many algorithms have been shown to be capable of producing s that are similar to human-produced s. However, a number of fundamental limitations built into the classic model itself have come to the fore. These issues are discussed at length in van Deemter (); it is the last of the four categories below that the remainder of this chapter will be concerned with. A. Limitations in the types of referring expressions generated. s constructed following the classic model are conjunctions of atomic properties. Proper names are not a part of this scheme, for example, which therefore omits s such as ‘Sir Walter Scott’, ‘The author of Waverley’ (where ‘Waverley’ is a proper name), etc. Likewise, these algorithms are unable to produce ‘the man who is not wearing a hat’, or ‘the man who feeds two felines’ (which uses a numerical quantifier). Plural s are out of reach, because the referent is always one single individual; consequently such expressions as ‘the (two) lions’, or ‘the Kenyan lion and the tiger’ are never generated. To address these limitations, there is a need for algorithms that are able to generate a broad variety of logical forms. Algorithms of this kind have been proposed, using a variety of computational formalisms. Reference to sets can be addressed using a straightforward set-theoretic approach van Deemter (), though this throws up difficult empirical questions (Gardent ; Horacek ; Gatt ; FitzGerald et al. ) and does not deal with sets as collectives (Stone ). Logically more expressive approaches have used Description Logic, which permits the construction of the formula Man ⊓ ≥ feed.Feline, denoting the man who feeds at least two felines. Algorithms search the space of all well-formed formulas of a suitable fragment of Description Logic to find a formula that denotes the intended referent (Areces et al. ; Ren et al. ). B. Limitations in the type of Knowledge Representation employed. The vast majority of work on  makes use of the simple type of Knowledge Representation exemplified by Table . above, which models human knowledge as if it were a simple table of atomic facts. This makes it impossible to model logically complex facts, such as the fact that a given animal is Indian or Pakistani (which involves disjunction), the fact that every animal is housed in a cage (which involves universal quantification), the fact that every person has exactly one father, and so on. To address these limitations, and to exploit the information contained in logically complex knowledge, there is a need for serious Knowledge Representation and automated reasoning. Researchers have sometimes tackled problems A and B together, as part of one package. Description Logic, for example, permits the representation of logically complex shared information and this might permit it to offer an attractive model of

OUP CORRECTED PROOF – FINAL, 21/1/2019, SPi



  

attributive descriptions (e.g., Donnellan ). Suppose, for example, the Knowledge Base asserts that every person has exactly one father. Once we know about the existence of an individual John, then an  whose semantic content says ∃father.{John} (literally: the set of all individuals x such that x is a physical father of at least one element of the singleton set {John}) refers uniquely to a person who is not listed in the Knowledge Base but whose existence can be inferred (van Deemter : Chapter ). C. The limitation to deterministic algorithms. The classic model does not require  algorithms to be deterministic, but in practice they have tended to be: given an input (defined as a referential domain, an intended referent, and a set of properties over the domain), such an algorithm will always produce the same . This makes such an algorithm a bad model of the set of all speakers (because different speakers may choose different s), and even of one individual speaker (because speakers choose different s on different occasions). To remedy this limitation, some recent  algorithms try to model a probability distribution over s (van Deemter : Chapter ). Essentially, these algorithms roll a loaded dice to make decisions. For example, they may choose one type of property twice as often as another because the former occurs twice as often in a corpus. Additionally, human-generated s are often overspecified (Ford and Olson ; Engelhardt et al. ), therefore a  algorithm may also use a loaded die to choose whether or not to produce an overspecified  (e.g., van Gompel et al. ), making the algorithm probabilistic in two different ways (see also Mitchell et al. ). D. Limitations in the type of information sharing. Yet another type of limitation, discussed in §.–§., comes from the inability of the classic model to acknowledge complications in Information Sharing. In one class of situations, discussed in §., the speaker may be uncertain what the hearer knows; for example when a speaker addresses an unknown audience (e.g., in writing, or on TV), very little information may initially be shared. In such situations, the classic model cannot explain how we are referring. In §., we will see that in many situations the hearer’s knowledge is only implicit: he can find out about it, but he does not have it at his finger tips all the time; in situations of this kind, something needs to be done to make information become shared. In yet other situations, discussed in §., the referent cannot be singled out, because it is not feasible to remove all distractors without also removing some intended referents. Finally, as we shall see in §., there are situations in which identification of the referent is not the goal of a description, and the main challenge is for us to understand what it is that the speaker is trying to do in the first place. The logic of Shared Information can be confusing. Suppose the hearer does not know whether c is a tiger or a lion (Table .). This time around, c’s being a tiger is not shared between the speaker and the hearer. Yet if the speaker said, ‘The tiger is in the cage’, the hearer could combine the information in the  with the information in the knowledge base and figure out which of the three animals this is about, because it can be neither a nor b (which he knows to be lions). In yet other situations, the speaker may humor the hearer; speaking to a child, for example, he might refer to a whale-like

OUP CORRECTED PROOF – FINAL, 21/1/2019, SPi

   



Table . A modification of the Knowledge Base of Table ., with uncertainty over what species c belongs to Identifier

Species

Origin

Weight

Injuries

a b c

lion lion tiger or lion

Kenya China China

kg kg kg

paws, teeth paws back

sea animal as ‘the fish’ because this is what the child (incorrectly) believes the animal to be.³ Our discussion will not focus on situations of this kind, but on situations in which the speaker adheres to the facts as she understands them to be. Some of the algorithms discussed in the following sections employ a simple monotonic approach to . The idea is to start with an empty , in which the properties available for referring are inspected one by one, and added to the  if they are useful; once added, a property is never retracted. The difference between these algorithms lies principally in the order in which the properties are inspected. Dale and Reiter’s Incremental Algorithm, for example, (see Algorithm  below) employs a fixed ‘Preference Order’, which is assumed given (e.g., on psycholinguistic grounds), (Dale and Reiter ); other algorithms choose properties based on their discriminatory power (i.e., the number of distractors they remove). Later sections will indicate how algorithms of this kind may be modified to address situations in which Information Sharing is compromised. Algorithm  The Incremental  Algorithm Input: A domain of objects containing a target referent r and a non-empty set of distractors M. A set  of properties true of r, listed in Preference Order. Output: A distinguishing description  of r that uses conjunctions of properties in , if such a distinguishing description exists. : Start out with an empty  : while not all distractors have been ruled out and  ≠ ø do : Select a new property p from , choosing the one that appears next in the Preference Order : if p is false of some distractors then : Add p to  : Remove p from  : Remove from M all distractors ruled out by p

³ For other deviations from the truth in Natural Language Generation, see van Deemter and Reiter ().

OUP CORRECTED PROOF – FINAL, 21/1/2019, SPi



  

. B   

..................................................................................................................................

The psycholinguistic literature abounds with situations where speakers have all the necessary information about their hearer, yet they struggle to take this information into account when they speak (Horton and Keysar ) (see §..). But in many situations, speakers are in an even more difficult situation, because they lack the necessary information: The ‘Who is?’ scenario. The publishing industry comes up with a new reading gadget. This gadget allows the reader of a document to select a proper name x that occurs in the document, asking “Who (or what) is x?” Its aim is to answer these questions in an optimally useful way. How should they be answered, assuming that space does not allow an entire Wikipedia page to be displayed?

The scenario starts from a situation in which the hearer does not know who x is. Situations of this kind occur frequently, because proper names are conventional in a way that descriptive properties are not: once I’ve learned what red means, I can apply the word to any red object, but learning your name does not teach me how to apply this name to anyone else. This has interesting implications: if I use a proper name x to refer to the referent r, I implicitly convey that I consider the fact that x is a name of r to be in our common ground. This is significant, for example, when x is a nickname: by saying to you that The Gunners have just won a football match (this is a nickname for the North-London football club Arsenal), I’m conveying that this nickname is in our common ground. Effectively, I am saying that we both belong to a social circle of people who are in the know about the club (van Deemter : Chapter ). But let us turn to the task of the person who answers the ‘Who is?’ question in the scenario. This task was discussed by the philosophers Boër and Lycan () and addressed computationally by Roman Kutlák (). This task poses difficult challenges to , compounded by the fact that some hearers may not be familiar with the intended referent of the . Mindful of the difference between these two kinds of readers—knowing and unknowing readers, as we shall say—Kutlák focused on the question of what information is most widely known about the referent. As for knowing readers, information that is widely known is relatively likely to be known by them as well; hence this information stands a good chance of triggering their memories of the referent. As for unknowing readers, information that is widely known is what these readers might be most interested in acquiring, because it helps to close the information gap between them and the people who do know the referent. Both types of hearers benefit if the  contains information that is widely known. The challenge is for the computer to find out what information is widely known. Kutlák hypothesized that this problem may be solved if the internet is used as a window on hearers’ knowledge. Many people read information on the world-wide web. If we do not know anything specific about a particular hearer’s knowledge state, then

OUP CORRECTED PROOF – FINAL, 21/1/2019, SPi

   



this hearer might be modeled as having been exposed to a large but unknown fragment of the world-wide web. From this perspective, if an item i of information occurs on the web more frequently than another item j, then the hearer is more likely to have been exposed to i than to j, so i is put earlier in the Preference Order than j. Several versions of this idea were implemented and tested, the simplest of which is this: to find out how likely it is that a reader knows that Albert Einstein was a physicist, search for documents on the world-wide web that contain the name ‘Albert Einstein’ and the property ‘physicist’. The more documents match this search, the more widely known this proposition is hypothesized to be. More sophisticated methods make use of Pointwise Mutual Information (PMI; Fano ), which is based on a comparison between the joint probability of two events and the probability of co-occurrence of the two events if n and p are probabilistically independent of each other. In the present case, n is the occurrence of a particular name in a document and p is the occurrence of a particular property in the same document: PMI ðn; pÞ ¼ log

Pðn; pÞ : PðnÞ PðpÞ

ð1Þ

If this was all there is to Kutlák’s problem, and if the set of distractors is known (e.g., by extracting a large set of famous people from the internet), then the monotonic approach to , introduced in §.. might be resurrected to shape these ideas (Algorithm ). Algorithm   algorithm based on an assessment of hearers’ knowledge Input: A domain of objects, containing a target referent r and a non-empty set M of distractors. A set  of properties of r. A metric (see text) that estimates how well known a property is in connection with r. Output: A distinguishing description  of r if one exists. : Start out with an empty  : while not all distractors have been ruled out and  ≠ ø do : Select a new property p from , choosing the best-known one : if p is false of some distractors then : Update , , and M In the ‘Who is?’ scenario, however, it is unknown what distractors exist in the hearer’s mind, and what properties he ascribes to each of them. The set of distractors is unknown; hence it is not possible to ascertain whether all distractors have been ruled out (clause ) and whether P is false of some distractors (clause ), so a different algorithm needs to be used, which does not rely on the distractor set. Kutlák hypothesized that, once again, a crowd-sourced information resource like the internet can shed light, essentially by estimating the Discriminatory Power of an  (as opposed to that of a single property). Once again, documents that express a certain conjunction of properties are retrieved from the world-wide web. One option is to linguistically ‘realize’ this conjunction by expressing it as an English , and to use this  to search the internet. If the proportion of documents returned that also contain the

OUP CORRECTED PROOF – FINAL, 21/1/2019, SPi



  

name of the referent is small, then the  is considered not to be complete yet, in which case the next-best known property is added. For example, if the description of Einstein composed so far is ‘German-born physicist’ and fewer than, say, % of the documents returned by the  contain the name ‘Albert Einstein’, then the algorithm should continue to add properties to the Logical Form. The generation algorithm terminates when the number of documents retrieved that contain the name of the referent exceeds the threshold of %. Variants of these ideas were first tested in pilot experiments. The two most promising ones, identical except for their termination heuristic, were then employed to generate expressions that refer to a number of famous people, based on the facts about these people represented in DBpedia, a huge database extracted from Wikipedia (Bizer et al. ). The output of algorithms was assessed via an experiment based on Mechanical Turk, in which participants were shown an English  and given the instructions in Figure .. In response to the , participants were asked to move a slider to the left or the right, depending on the extent of their agreement or disagreement with a given statement. Evaluation was performed in terms of the number of people correctly identified by participants who were shown a number of s that were generated, and in terms of the two other evaluation questions (Figure .). Kutlák’s algorithms were compared

 . Kutlák’s final evaluation experiment.

OUP CORRECTED PROOF – FINAL, 21/1/2019, SPi

   



against a number of competitors, including a version of the Incremental Algorithm, and a set of specially crafted (‘gold standard’) s produced by human authors. The results were encouraging. In terms of correct identification and quality (i.e., answers to the question Suppose you did not know this person, how good would you find the description?), both algorithms performed considerably better than the Incremental Algorithm. In terms of naturalness (i.e., participants’ answer to the question How natural does the description read to you?) there was no statistically significant difference. Unsurprisingly, gold-standard descriptions (i.e., s) composed by human authors outperformed their competitors in all respects (Kutlák et al. ). Some examples of the descriptions compared can be found in Table .. Kutlák’s work shows how the challenge of the ‘Who Is?’ scenario can be addressed by using open, crowd-sourced, information resources. Kutlák used the internet as the main source of information; and as his search queries were performed in English, the documents returned were also in English. These and similar methods could also be used with different languages and other audiences. Many communicative situations combine features of the different scenarios discussed here. For example, Barton Lipman () described an imaginary game that features a speaker who needs an acquaintance to be picked up from an airport. Lipman, who was intrigued by the fact that human language makes such frequent use of vague expressions (e.g., ‘tall’, ‘large’, ‘grey’) asked under what circumstances it would be beneficial, for the speaker and/or the hearer, if the speaker uses vaguely defined words (as opposed to exact measurements) as part of an . Here is a scenario inspired by Lipman’s: The Airport scenario. A speaker asks a hearer to go to the airport to pick up an acquaintance of the speaker. Unlike the speaker, the hearer does not know what the acquaintance looks like. There will be other people at the airport, but their number and features are not known in advance. What should the speaker say? Table . REs generated by one of Kutlák’s main algorithms, the Incremental Algorithm with optimal Preference Order (IA), and a human-authored RE (Human). Notochthalamus, in descriptions of Darwin, is a species of barnacle Referent

Algorithm Description

Charles Darwin

Human IA Kutlák

Billie Holiday

Human IA Kutlák

This person is considered the father of the modern theory of evolution due to his book On the Origin of Species. This person was a British scientist who was popularized by Alvar Ellegard. This person died in Downe, was known for On the Origin of Species and named Notochthamalus. This person was a noted jazz singer, famous for her songs ‘God Bless the Child’, ‘Lady Sings the Blues’ and ‘Strange Fruit’. This person is a musician and is the author of A Mothers Gift. This person wrote ‘Our Love Is Different’, was also known as Lady Day, and has homepage http . . .

OUP CORRECTED PROOF – FINAL, 21/1/2019, SPi



  

As in the Direction Giving scenario, the hearer will have to inspect the domain to find the facts. But this time, the speaker cannot know what information the hearer will find there. As in the ‘Who is?’ scenario, the speaker does not know what distractors the hearer will find. Lipman argued, using a Game-Theoretical perspective, that in situations of this kind, speakers who want to minimize the chance of misunderstandings should be as informative as possible. In Lipman’s scenario this meant avoiding vagueness and rounding; in our own Airport scenario, it means conveying all those properties that the hearer will be able to check—a strategy that tends to lead to an extremely lengthy , of course. Lipman noted, however, that his arguments did not take into account that lengthy s may be difficult to produce and to interpret, creating a potential trade-off between clarity and production and/or comprehension effort. Further discussion of these issues would be out of place here, but it would be interesting to explore what is the optimal amount of information that the speaker should pack into her s and compare this with actual speaker behavior. The scenarios discussed so far in this chapter show that the referential situations usually studied in  fail to address some of the most commonly occurring types of reference, because they make simplistic assumptions about the epistemic nature of the reference task. In these scenarios, reference is nonetheless possible. We now turn to a type of scenario where this is less clear, because it is not feasible to separate the target referent from all possible distractors.

. C   

.................................................................................................................................. A subtler set of problems arises when we delve into what it means to know something. Research on  has often started from neatly stratified knowledge represented in a table that is especially arranged to highlight the commonalities and differences between entities (e.g., Table .). But real life is seldom so accommodating: we often have to work hard to find out how the facts line up, and this affects the way in which referring works. Consider the following scenario: The Direction Giving scenario. Suppose Lewes Road is the longest street in Brighton. I invite you to come to number , the highest number in the street. You don’t know Brighton and have no map. I tell you, in a cafe in London, to ‘come visit me in Brighton tomorrow, at number ’. Because other streets in Brighton only have numbers lower than , this is a distinguishing description. Yet it does not help you much to find the house. A description like ‘ Lewes Road’ (‘ . . . in the Moulsecoomb area, right at the end of the street’) could have saved you a huge amount of time.

From earlier work, we knew that two types of situations are particularly problematic, which we termed Lack of Orientation () and Dead End (), both of which are closely related to the knowledge of the hearer (Paraboni et al. ); we decided to test

OUP CORRECTED PROOF – FINAL, 21/1/2019, SPi

   



the applicability of these concepts to spatial domains by studying the virtual environments provided by the  evaluation challenge (Koller et al. ), a game-like setup designed to test language generation algorithms that forced contributors to refer clearly in spatially complex situations. This setup allowed us to measure search effort in a variety of ways, by logging the time taken and the distance traveled by a hearer who is searching for the referent. A  ‘world’ consists of a D virtual space containing rooms with doors, tables, chairs, etc., while rooms contained buttons of various colors. Some buttons are hidden behind the larger landmarks (e.g., plants). It is natural to refer to the less accessible object (e.g., the button) via the accessible landmark, thereby facilitating the hearer’s search task: ()

the blue button, behind the plant

Suppose there is only one blue button. In this case, mentioning the plant is not necessary to permit identification, yet the simpler  that just says ‘the blue button’ may make it difficult to find the referent. Where this problem occurs, we speak of Lack of Orientation (). Now let us modify the scenario: there are  blue buttons, so the reference to the plant is necessary for disambiguation (Figure .). This time, even () will cause problems if the hearer comes across another plant before reaching the intended plant: seeing this other plant, the reader may be puzzled not to find a blue button behind it. This is what we call Dead End (). Search would be facilitated if some additional, logically redundant information was added, for example: ()

the blue button behind a plant, on your right.

As expected, s generated by existing  algorithms caused readers to spend much more time, and to travel much further, than if the  is overspecified judiciously target b3

b4 c2

p2

p1

c1

DE b1

b2 start

 . Schematic view of a domain in the  game. For ease of reference, identifiers have been added; the space has been divided into five rooms.

OUP CORRECTED PROOF – FINAL, 21/1/2019, SPi



  

(Paraboni and van Deemter ). The idea is to monitor the s that a given  algorithm produces, and to add information if the monitoring reveals an  or  problem. Overspecification means the s in question contain properties that are not logically necessary for identifying the referent. Despite the name, there is nothing wrong with overspecification per se: an overspecified  can sometimes be the shortest  that any reasonable speaker would produce. The hearer begins his search in the position start (see Figure .). If we abstract away from the relations (‘above’, ‘near’, etc.) between objects, we can model a description with one landmark as a sequence of two pairs, 〈(x₁, P₁), (x₂, P₂)〉, where x₁ is r, and x₁ is interpreted by means of the landmark x₂. For example, x₁ may be a button, P₁ may be the conjunctive property of being blue and a button, and P₂ may be the property of being a plant. If the relation is ‘behind’, then D encodes the  ‘the blue button behind the plant’.  (Judicious Overspecification) starts from an initial description of this kind. P is the set of all search paths that the reader may choose, going from where he is, all the way up to xn. P is a set because the reader’s search path is not fully known. In Paraboni and van Deemter () we proposed and tested a simple model of human search, called Nearest-First Search (). Once P is computed,  monitors the . If the initial description leads to  or , then an additional property P is included, and this process of adding properties continues until the risk of / has been averted. Algorithm  : Judicious Overspecification Input: A spatial domain containing a hearer location l and another location where the target referent r is positioned. A set  of properties defined on the domain. In a description of the form  = 〈(x₁, P₁), (x₂, P₂) . . . (xn, Pn)〉, where n ≥ , the referent r equals x₁, and xi+₁ is a landmark that helps to locate xi (as in ‘the button behind the plant’, where the button is xi and the plant xi+₁). Output: an  , which may contain overspecification. The notion of a ‘problematic’ path (and hence the choice of whether or not to overspecify) is explained below. : Let an existing REG algorithm produce an initial description  = 〈(x₁, P₁), (x₂, P₂) . . . (xn, Pn)〉, where n ≥  : P: = the set of -compatible search paths from l to xn : while some of the elements of P are problematic do : Expand  by adding a suitable new location property to Pn : return  Suppose the algorithm is trying to identify an entity xn and has ascribed the conjunction Pn of properties to xn. Now a search path O ∈ P is problematic in two possible situations. A -type problem arises if O contains an entity of the same type as xn (yet different from xn) for which Pn holds true, and which may consequently be confused with xn. An  problem arises if the path contains a pair (xn, Pn) where xn is not located in Room (xn–) yet there is no information in Pn that makes this clear. The intuition is that xn is too far away to get noticed by the hearer, and there is nothing in the  that points towards xn. A -type problem arises in the domain depicted in Figure .

OUP CORRECTED PROOF – FINAL, 21/1/2019, SPi

   



if the starting instruction says ‘push the green button’: there is only one green button (b₁), but it is in a different room, therefore  considers it to be too far away to be easily found. Example. Consider Figure ., where the aim is to steer the player from ‘start’ towards the button b₄. Initially, a distinguishing description is produced, for example, ‘the blue button behind a plant’. The reference ‘a plant’ to p₂ is tested for /. If the hearer finds the nearest plant first (as the  search model of Paraboni and van Deemter () predicts), then this creates a  problem. The fact that the wrong plant p₁ may be encountered first is sufficient reason for  to amplify the initial Logical Form by adding location information (e.g., ‘on the right’). The expanded  that results from this does not cause  any longer, so this overspecified  is generated. The Direction Giving scenario demonstrates that existing  algorithms can produce s that are strikingly unhelpful, because they fail to distinguish between information that the hearer possesses and information that the hearer can acquire. A very similar set of issues, likewise relating to the  game, addresses the problem using a Machine Learning algorithm based on maximum entropy (Garoufi and Koller ). Other relevant work, also focusing on complex domains, includes Coco and Keller (); Barclay (); Koolen et al. (); and Clarke et al. (). Psycholinguistic research has taught us not to assume too readily that speakers design their s optimally for hearers (§..). The work discussed above highlights a situation in which the lack of any audience design would tend to cause disaster. The work reported in Paraboni and van Deemter ()—which focuses on the effects that s have on recipients—does not tell us what speakers actually do, but Paraboni et al. () contains a small study suggesting that human authors of written texts are good at this type of audience design when they are encouraged to think about the choice of  in a text. It may well be that human speakers are frequently more egocentric (cf. the discussion in §..), especially when they are under time pressure or when referential success is not of crucial importance. Where speakers do engage in the type of behavior suggested by the  model, they can be seen as performing a kind of self-monitoring: they examine an , and if it falls short, they add more information to it; selfmonitoring processes of this kind are well attested in human language production (see e.g., the famous model of Levelt (): Chapter ).

. A 

.................................................................................................................................. The classic model assumes that identification of the referent is a requirement in all cases; moreover, all distractors should be removed from the hearer’s attention, and one distractor cannot be more important than another. In many situations, however, these assumptions are not justified.

OUP CORRECTED PROOF – FINAL, 21/1/2019, SPi



  

The Olive Oil scenario. You are preparing a meal in a friend’s house, and you wish to obtain, from your own kitchen, a bottle of Italian extra virgin olive oil. You phone home to ask your young son to bring it round for you. You know that also in your kitchen cupboard are some distractors: one bottle each of Spanish extra virgin olive oil, Italian non-extra virgin olive oil, cheap vegetable oil, linseed oil (for varnishing) and camphorated oil (which is medicinal). It is imperative that you do not get the linseed or camphorated oil, and preferable that you receive olive oil. The expression ‘Italian extra virgin olive oil’ guarantees clarity but may overload your helper’s abilities. A very short expression, ‘oil’, is risky. Perhaps you should settle for the intermediate ‘olive oil’.

Discussing this scenario, Graeme Ritchie, Imtiaz Khan, and I once sketched an approach to  that operates by searching for the  that has the lowest cost (Khan et al. ). The cost of a Logical Form S is defined in terms of a combination of the length of an  (the Brevity cost, fB(S), with B for Brevity) and the number of distractors that the  fails to remove (the Clarity cost, fC(S)), for example, Cost(S) = fB(S) + fC(S). Search for the lowest-cost  could be greedy, always selecting the property that reduces Cost the most, or it could work in such a way that a minimalcost  is always found. Clarity is essentially a form of utility, of course, hence Clarity cost is a form of negative utility. A failure to remove linseed oil should bear a higher cost than a failure to remove Spanish extra virgin olive oil. Therefore, instead of a simple linear function of the size of the set of distractors that an  fails to remove, there is a curve where the cost drops more steeply as the more undesirable distractors are excluded. For example, each object could be assigned a numerical rating of how undesirable it is, with the target having a score of zero; the brevity cost function fB(S) could still be a linear function (Khan et al. ). An algorithm along these lines will sometimes terminate before all distractors have been removed. Approximate s also come up when the aim is to refer to a spatial region. Consider the following scenario as an example. As in the Olive Oil scenario, it is difficult to remove all distractors, but this time the difference in status between different distractors does not come to the generator’s aid. The Dirty Floor scenario. Mr X has dropped a piece of fish on the floor, then removes it. He would now like Mrs X to wipe the area clean. The fish doesn’t leave a visible stain, so he has to explain where it was. It appears that there is no such thing as a distinguishing description (except the insufficiently informative ‘where I dropped the fish’), although Mr X can arbitrarily increase precision by adding further information, for instance ‘near the table’, ‘on your right’, and so on.

An ideal description would cover the dirty area and nothing more. On the other hand, a larger area will also do. The domain is defined as all conceivable sub-areas of the floor, so the target is one element of the domain (i.e., one sub-area). The function fC(S) should now assess the aptness of any potential  S ensuring that the area described by S contains the target (i.e., the contaminated area), and that it does not contain too much else.

OUP CORRECTED PROOF – FINAL, 21/1/2019, SPi

   



Computerized weather forecasts (e.g., Turner et al. ) resemble the Dirty Floor scenario. These forecasts can assess which roads are likely to be icy, and hence dangerous. Systems of this kind inform road gritters concerning the condition of the roads, to help them decide which ones require treatment with salt to avoid traffic accidents. The Road Gritting scenario. Suppose one weather warning says ‘Roads in the Highlands will be icy’ whereas another says ‘Roads above  metres will be icy’. The first summary may have ten false positives (i.e., roads gritted unnecessarily) and ten false negatives (i.e., dangerous roads not gritted); the second has  false positives and only two false negatives. Which of the two references to the intended geographical area is preferable?

The weather warning is based on sensor readings from a finite set of points on the map. But although this makes it possible in principle to list precisely which points are predicted to be icy, this would lead to enormously lengthy s. Crucially, one and the same area can be approximated in different ways. For example, one might say ‘The American Midwest’ (describing a given area as a whole), ‘The Great Lakes region and the Eastern Great Plains region’ (splitting the same area in two), or one might list all the States that cover the area: ‘Indiana, Iowa, Kansas, Michigan, Minnesota, Missouri, Nebraska, North Dakota, Ohio, South Dakota, and Wisconsin’. Each summary will lead to a different set of roads being treated with salt (i.e., gritted). Salt is bad for the environment, on the other hand failing to grit a slippery road can cause traffic accidents. Faced with these difficulties, Turner and colleagues decided to disallow false negatives (i.e., every road predicted to be icy must be reported as such). In other words, their generator would select a third weather warning, which covers the entire target area but may have numerous false positives (e.g., ‘roads outside coastal areas will be icy’). Turner’s cautious approach may have been justified given the potential gravity of traffic accidents, but in other situations, a more flexible approach is called for. Once again, one could associate each utterance with a numerical utility. For example, one might reason that if a false positive has a negative utility (i.e., cost) of . and a false negative has a cost of ., then the first summary is preferable to the second; for the first summary only has a cost of (.) + (.) = , whereas the second has a cost of (.) + (.) = . The choice of the utility weights associated with the different kinds of error is crucial, of course, and can be tricky to justify. Moreover, although our emphasis on utility is only able to shed light on the question what s are best for hearers, it is as yet unclear to what extent speakers are sensitive to this type of utility.

. I     ?

.................................................................................................................................. When we describe an entity, identification of that entity is not always our main aim. To gain an understanding of what other aims a description can have, and how these aims

OUP CORRECTED PROOF – FINAL, 21/1/2019, SPi



  

may be modeled computationally, we examine some scenarios in which identification of the referent is merely one factor alongside others. Consider the  experiment, in which pairs of participants buy furniture together. The dialogues saw participants proposing a piece of furniture, persuading their partner to buy it, changing an earlier proposal, confirming an earlier proposal, and so on. The Intentional Influences model (Jordan ) contained rules that guarantee identification of the referent. Additionally, the model contains rules such as the following, that serve a different purpose: In a persuade context, select the properties that make the item a good solution to the problem.

This rule applies when there is a need to convince the hearer of the merits of a particular furniture item. These rules show how the identification of a referent can go hand in hand with other communicative goals. Recent years have seen work on recommender systems that follows a very similar logic when a system figures out what to say about an item to a particular user (Carenini and Moore ). Students at Aberdeen studied a similar type of situation, focusing on reference in a sales situation in which the preferences of the user are not known: The Camera Adviser scenario: You are discussing cameras with a sales adviser. You are looking for a high-end compact camera and have discussed some options with the adviser, who has mentioned five different cameras to you. At a loss from so much choice, you ask ‘Which camera should I buy?’ The answer should refer to a specific camera, in such a way that the choice for this camera is motivated.

Common though this type of scenario is, existing  algorithms are not well placed to handle it. For although the adviser is trying to identify a camera, identification is only a part of what she is trying to do when she says ‘The Nikon with the full-frame sensor’: her aim is to let the most important features of the camera get across to you. If your interests, as a customer, are very well understood—you may be particularly interested in a camera’s ability to take good pictures in low light, for example—then these should be taken into account. But what if not enough is known about the customer’s interests? In this scenario, interestingness is relatively easy to predict, because quantitative attributes, such as price, weight, and pixel count, are pervasive in it. It might be reasonable to regard extreme values of a quantitative attribute as inherently interesting. Suppose, for example, the cameras range in price from  to  dollars, with most cameras costing somewhere around ; then the property of costing  dollars is less ‘interesting’ than the property of costing, say,  dollars, even if the property of costing (precisely)  dollars has huge Discriminatory Power. Extremity can be measured statistically as a Z score (i.e., comparing a value’s distance to the mean of the set with the standard deviation over the set); in a pilot study, this approach appeared to give plausible results in relation to the camera scenario, producing s that express numerical information that camera buyers might care about.

OUP CORRECTED PROOF – FINAL, 21/1/2019, SPi

   



The interestingness of a property can be used as yet another way to populate the monotonic  algorithm scheme of §.., which we have seen a number of times now. Consider Algorithm  below, for example, which also takes on board the idea from §., that properties are added regardless of whether they remove any distractors. It may be possible to rank properties in terms of interestingness in advance: if interestingness is understood simply in terms of extreme values, then it can be computed once and for all for a given r. If interestingness is understood in terms of how surprising a property is, however, then it should depend on the other properties assigned to r. For example, a price tag of  dollars may be extreme for a camera, but if the camera is known to be a luxury brand, it is no longer surprising and it is therefore less notable. Algorithm   algorithm based on ‘interestingness’ Input: A domain of objects, containing a target referent r and a non-empty set M of distractors. A set  of properties of r. A metric (see text) that tells us how interesting a property is in connection with r. Output: A distinguishing description  of r if one exists. : Start out with an empty  : while  ≠ ø do : Select a new property P from , choosing the most interesting one : if P is false of some distractors then : Update , , and M Unfortunately, the notion of interestingness outlined above is only applicable to quantitative variables; in other cases it is meaningless. Can we predict how interesting a property is when the property is not quantitative? In their survey of metrics for data mining, Geng and Hamilton () suggested that interestingness may be determined by a variety of factors, including novelty, generality, and reliability, each of which one may attempt to measure. One of the most important factors is what the authors call surprisingness or unexpectedness, defined as patterns that contradict existing knowledge or expectations. In the study discussed in §., Kutlák applied this idea to  by hypothesizing that unexpected properties of a person are properties that are rare for a person to have. For example, being awarded the Nobel prize is unexpected because only a handful of people receive this prize. Formalizations of unexpectedness tend to have conditional probability at their core. For example, when we compute the extent to which x’s having won the Nobel Prize is unexpected, we want to know how likely it is for someone to win this prize. But what should we assume as given? The fact that x is a physicist? The fact that x is female? Other statistical constructs may be relevant too. For example, being born on  March has low probability, yet someone’s being born on that date is not very unexpected or newsworthy. Perhaps this is because in the case of someone’s birthday, where information-theoretic entropy is high, we do not build up an expectation, hence  March is not unexpected. But in a case where entropy is lower, as in the date

OUP CORRECTED PROOF – FINAL, 21/1/2019, SPi



  

when someone starts their PhD (which is often around the start of the academic year),  March might be unexpected. At the time of writing, these ideas are only starting to be explored (see Kutlak  and van Deemter ). Moving away even further from the classic reference task, consider: The Unnamed Company scenario. Two people in my research group have founded a company, Arria . When colleagues ask who they visited last week, they may say ‘(we visited) a large oil company whom we are hoping to get as a customer’. Their description is not designed to allow us to identify the referent, but to give us an insight into the reasons for their visit. In a variant of this scenario, there may be business reasons why the two are not able to disclose the identity of their prospective customer.

The speaker’s task is to provide information about a specific company. But this time, he does not try to identify the referent, because he considers this uninteresting. In the variant scenario, the plan is not to permit identification of the referent. Far from being unusual, this is precisely what is done when researchers in medical and other sensitive areas do when they anonymize data. A research article will tend to discuss specific individuals—for example, it may be important that the patient who underwent a given operation is the same as the one who developed certain complications—yet researchers have to ensure that what they say about a patient does not identify the patient to a reader; to ensure anonymity, computational techniques very similar to the ones discussed in this chapter are brought to bear (Raghunathan ). What all these examples emphasize is that speakers and authors can have very different aims with their descriptions, and that the boundaries of reference are not always clear.

. C   

.................................................................................................................................. This chapter started by carving out a view of reference production that most of the classic  algorithms have in common; I have called this the classic model of referring. We then listed a number of limitations of the classic model, after which we concentrated on a set of challenges to the model, which arise when Information Sharing is not a simple interplay between given and new information (in which the  singles out the target referent by exploiting information that is shared between the speaker and the hearer), but something more complicated. Discussion of these challenges has taught us that the purpose of a description can vary considerably: it can be to identify the entity precisely, to identify it approximately (as in the Road Gritting scenario), or to explain what’s striking about it (as in the “who is?” scenario), or combinations of the above (as in the Camera Adviser scenario). In addition to its discriminatory power (i.e., the number of distractors removed by the

OUP CORRECTED PROOF – FINAL, 21/1/2019, SPi

   



property), the decision of whether or not to include a property into a description may be influenced by its ability to recommend an object, the extent to which a quantitative property is extreme, the extent to which a property is surprising or unexpected, or the extent to which a property is perceptually salient. With the exception of discriminatory power, these factors were ignored until very recently. One way in which the algorithms discussed further improved, is by taking into account that speakers and hearers frequently collaborate to ensure that the identity of the referent ends up in shared information: the speaker may start out with a description that is not clear enough, but further interactions with the hearer may clinch the deal. Clearly, one of the functions of collaboration is precisely to iron out epistemic problems such as the ones we have discussed: in many of the scenarios discussed, it would have been possible for the speaker to start out with a first, tentative ; if this  fails to identify the referent for the hearer, the hearer may ask for some missing information (e.g., ‘You mean the one on the right?’), after which the speaker adds information to complete the . This implies a modification of the process of Information Sharing that has been studied widely by psycholinguists (Clark and WilkesGibbs ), and increasingly by computational linguists as well (Heeman and Hirst ; Engonopoulos et al. ; Garoufi and Koller ; Fang et al. ). The computational challenges discussed in this chapter echo themes that have long been prominent in the psycholinguistics of language production. This is most evident from our discussion of shared information in §., which focuses on issues that were defined by psycholinguists, and that are still hotly debated in their ‘egocentricity’ debate. The same holds for our discussion of large domains in §., given that the theme of overspecification has always been paramount (Ford and Olson ; Engelhardt et al. ; Koolen et al. ). Self-monitoring is another such theme, as we have seen in §.. This convergence between disciplines illustrates a joining of forces that can also be seen in other areas of computational modelling (e.g., Sun ), and that is starting to turn the computational modeling of human referring into a genuine area of Cognitive Science.

OUP CORRECTED PROOF – FINAL, 21/1/2019, SPi

OUP CORRECTED PROOF – FINAL, 21/1/2019, SPi

R EFERENCES .....................................

Abbott, Barbara (). A Study of Referential Opacity: PhD dissertation. Berkeley, CA: University of California. Abbott, Barbara (). ‘A pragmatic account of the definiteness effect in existential sentences’, Journal of Pragmatics : –. Abbott, Barbara (). ‘Some remarks on specificity’, Linguistic Inquiry : –. Abbott, Barbara (). ‘Doing without a partitive constraint’, in Jacob Hoeksema (ed.), Partitives: Studies on the Syntax and Semantics of Partitive and Related Constructions, Berlin: Mouton de Gruyter, –. Abbott, Barbara (). ‘Support for a unique theory of definite descriptions’, in Tanya Matthews and Devon Strolovitch (eds), Proceedings from Semantics and Linguistic Theory IX, Ithaca: Cornell University, –. Abbott, Barbara (). ‘Presuppositions as nonassertions’, Journal of Pragmatics : –. Abbott, Barbara (). ‘Donkey demonstratives’, Natural Language Semantics (): –. Abbott, Barbara (). ‘A reply to Szabó’s “Descriptions and uniqueness” ’, Philosophical Studies : –. Abbott, Barbara (). ‘Definiteness and indefiniteness’, in Laurence R. Horn and Gregory Ward (eds), Handbook of Pragmatics, Oxford: Blackwell, –. Abbott, Barbara (a). ‘Issues in the semantics and pragmatics of definite descriptions in English’, in Jeanette K. Gundel and Nancy Hedberg (eds), Reference: Interdisciplinary Perspectives, Oxford: Oxford University Press, –. Abbott, Barbara (b). ‘Presuppositions and common ground’, Linguistics and Philosophy : –. Abbott, Barbara (). Reference. Oxford University Press. Abbott, Barbara (). ‘Reference: foundational issues’, in Claudia Maienborn, Paul Portner, and Klaus von Heusinger (eds), Semantics: An International Handbook of Natural Language Meaning, Berlin: Mouton de Gruyter, –. Abbott, Barbara (). ‘Speaker’s reference: Smith’s murderer’, in Lisa Matthewson, Cécile Meier, Hotze Rullmann, and Ede Zimmermann (eds), Companion to Semantics, Wiley. Abbott, Barbara (). ‘An information packaging approach to presuppositions and conventional implicatures’, Topoi : –. Abbott, Barbara and Laurence R. Horn (). ‘Nonfamiliarity and indefinite descriptions’, paper presented at the annual meeting of the Linguistic Society of America, Pittsburgh, PA, https://www.msu.edu/~abbottb/Nonfam&Indef.pdf. Abney, Steven P. (). ‘The English Noun Phrase in its Sentential Aspect’, PhD dissertation, MIT. Abraham, W. (). ‘Discourse binding: DP and pronouns in German, Dutch, and English’, in Werner Abraham, Elisabeth Leiss, and Elisabeth Stark (eds), Nominal Determination: Typology, Context Constraints, and Historical Emergence, Amsterdam: Benjamins, –.

OUP CORRECTED PROOF – FINAL, 21/1/2019, SPi



REFERENCES

Abusch, D. (). ‘The scope of indefinites’, Natural Language Semantics (): –. Acton, Eric and Potts, Christopher (). ‘That straight talk: Sarah Palin and the sociolinguistics of demonstratives’, Journal of Sociolinguistics : –. Adams, F. and Stecker, R. (). ‘Vacuous singular terms’, Mind and Language (): –. Aissen, Judith (). ‘Differential object marking: iconicity vs. economy’, Natural Language and Linguistic Theory : –. Allen, S. E. M. (). ‘A discourse-pragmatic explanation for argument representation in child Inuktitut’, Linguistics (): –. Allen, S. E. M. (). ‘Formalism and functionalism working together? Exploring roles for complementary contributions in the domain of child null arguments’, in R. Slabakova, S. A. Montrul, and P. Prévost (eds), Inquiries in Linguistic Development: In Honor of Lydia White, Amsterdam/Philadelphia: John Benjamins, –. Allen, S. E. M., Hughes, M., and Skarabela, B. (). ‘The role of cognitive accessibility in children’s referential choice’, in L. Serratrice and S. E. M. Allen (eds), The Acquisition of Reference, Amsterdam: John Benjamins, –. Allen, S. E. M., Skarabela, B., and Hughes, M. (). ‘Using corpora to examine effects in syntax’, in H. Behrens (ed.), Corpora in Language Acquisition Research. History, Methods, Perspectives, Amsterdam: John Benjamins Publishing Company, –. Almog, Joseph (). ‘The Subject-predicate class II’, Noûs (): –. Almog, Joseph and Leonardi, Paolo (). Having in Mind. Oxford: Oxford University Press. Almor, Amit (). ‘Noun-phrase anaphora and focus: the informational load hypothesis’, Psychological Review (): –. Almor, Amit, Kempler, Daniel, MacDonald, Maryellen C., Andersen, Elaine S., and Tyler, Lorraine K. (). ‘Why do Alzheimer patients have difficulty with pronouns? Working memory, semantics, and reference in comprehension and production in Alzheimer’s disease’, Brain and Language (): –. doi: ./brln... Aloni, Maria (). Quantification under Conceptual Covers. PhD dissertation, ILLC, Amsterdam. Alonso-Ovalle, L., Fernandes-Solera, S., Frazier, L., and Clifton, C. (). ‘Null vs. overt pronouns and the Topic-Focus articulation in Spanish’, Rivista di Linguistica (): –. Anand, Pranav (). De De Se. PhD dissertation, MIT, Cambridge, MA. Anand, Pranav, Chung. S., and Wagers, M. (). ‘Widening the net: challenges for gathering linguistic data in the digital age’, response to NSF SBE : Future Research in the Social, Behavioral, & Economic Sciences. Online: http://www.nsf.gov/sbe/sbe_/. Anand, Pranav and Andrew Nevins (). ‘Shifty operators in changing contexts’, in K. Watanabe and R. B. Young (eds) Proceedings of SALT . Ithaca, NY: CLC Publications. Anderson, Anne H., Garrod, Simon C., and Sanford, Anthony J. (). ‘The accessibility of pronominal antecedents as a function of episode shifts in narrative text’, The Quarterly Journal of Experimental Psychology Section A (): –. doi: ./. Anderson, J. C. (). ‘Misreading like a lawyer: cognitive bias in statutory interpretation’, SSRN Scholarly Paper ID , Social Science Research Network, Rochester, NY. URL: http://papers.ssrn.com/abstract=. Anderson, Anne, Miles Bader, Ellen Bard, Elizabeth Boyle, Gwyneth Doherty, Simon Garrod, Stephen Isard, Jacqueline Kowtko, Jan McAllister, Jim Miller, Catherine Sotillo, and Henry Thompson (). ‘The HCRC map task corpus’, Language and Speech : –. Anderson, J. and Holcomb, P (). ‘An electrophysiological investigation of the effects of coreference on word repetition and synonymy’, Brain and Language : –.

OUP CORRECTED PROOF – FINAL, 21/1/2019, SPi

REFERENCES



Appelt, D. (). ‘Planning English referring expressions’, Artificial Intelligence : –. Appelt, D. and Kronfeld, A. (). ‘A computational model of referring’, in Proceedings of the th International Joint Conference on Artificial Intelligence (IJCAI), –. Areces, C., Koller, A., and Striegnitz, K. (). ‘Referring expressions as formulas of Description Logic’, in Proceedings of the th International Natural Language Generation Conference (INLG), –, Salt Fork, Ohio. Ariel, Mira (). ‘Referring and accessibility’, Journal of Linguistics : –. Ariel, Mira (). Accessing Noun-phrase Antecedents. London: Routledge. Ariel, Mira (). ‘The Linguistic Status of the Here and Now’, Cognitive Linguistics , –. Ariel, Mira (). ‘Accessibility theory: an overview’, in T. Sanders, J. Schilperoord, and W. Spooren (eds), Text Representation, Linguistic and Psycholinguistic Aspects, Amsterdam/Philadelphia: John Benjamins, –. Arnold, Jennifer E. (). Reference Form and Discourse Patterns. Stanford University, Palo Alto, CA. [Retrieved from www.unc.edu/~jarnold/papers/diss/fulldissertation.doc]. Arnold, Jennifer E. (). Marking Salience: The Similarity of Topic and Focus. Unpublished manuscript, University of Pennsylvania. Arnold, Jennifer E. (). ‘The effect of thematic roles on pronoun use and frequency of reference continuation’, Discourse Processes (): –. doi: ./SDP_. Arnold, Jennifer E. (). ‘Reference production: production-internal and addresseeoriented processes’, Language and Cognitive Processes (): –. doi: ./ . Arnold, Jennifer E. (). ‘How speakers refer: the role of accessibility’, Language and Linguistics Compass (): –. doi: ./j.-X...x. Arnold, Jennifer E. and Griffin, Z. (). ‘The effect of additional characters on choice of referring expression: everyone competes’, Journal of Memory and Language : –. Arnold, Jennifer E., Eisenband, J. G., Brown-Schmidt, S, and Trueswell, J. C. (). ‘The immediate use of gender information: eyetracking evidence of the time-course of pronoun resolution’ Cognition : B–B. Arnold, Jennifer E., Brown-Schmidt, S., and Trueswell, J. C. (). ‘Children’s use of gender and order-of-mention during pronoun comprehension’, Language and Cognitive Processes (): –. Arnold, Jennifer E. and Tanenhaus, Michael K. (). ‘Disfluency effects in comprehension: how new information can become accessible’, in Edward A. Gibson and Neal J. Pearlmutter (eds), The Processing and Acquisition of Reference, Cambridge, MA: The MIT Press, –. doi: ./mitpress/... Arnold, Jennifer E., Bennetto, Loisa, and Diehl, Joshua J. (). ‘Reference production in young speakers with and without autism: effects of discourse status and processing constraints’, Cognition (): –. doi: ./j.cognition.... Arnold, Lynnette (). ‘Dialogic embodied action: using gesture to organize sequence and participation in instructional interaction’, Research on Language and Social Interaction : –. Arts, Anja, Maes, Alfons, Noordman, Leo, and Jansen, Carel (). ‘Overspecification in written instruction’, Linguistics : –. doi: ./LING... Arundale, Robert (). ‘Against (Gricean) intentions at the heart of human interaction’, Intercultural Pragmatics : –.

OUP CORRECTED PROOF – FINAL, 21/1/2019, SPi



REFERENCES

Asher, Nicholas, Hardt, D., and Busquets J. (). ‘Discourse parallelism, ellipsis, and ambiguity’, Journal of Semantics (): –. Asher, Nicholas and Lascarides, A. (). Logics of Conversation. Cambridge: Cambridge University Press. Asher, Nicholas and Lascarides, Alex (). ‘Strategic conversation’, Semantics and Pragmatics : –. Atlas, Jay D. and Stephen C. Levinson (). ‘It-clefts, informativeness’, in Peter Cole (ed.) Radical Pragmatics, –. New York: Academic Press. Austin, J. L. (). How to Do Things With Words. Cambridge, Massachusetts: Harvard University Press. Aylett, Matthew and Turk, Alice (). ‘The smooth signal redundancy hypothesis: a functional explanation for relationships between redundancy, prosodic prominence, and duration in spontaneous speech’, Language and Speech (): –. doi: ./ . Baader, F., Calvanese, D., McGuinness, D., Nardi, D., and Patel-Schneider, P. (). The Description Logic Handbook: Theory, Implementation, and Applications. Cambridge, UK: Cambridge University Press. Bach, Kent (a). ‘What’s in a name’, Australasion Journal of Philosophy : –. Bach, Kent (b). ‘Referential/attributive’, Synthese : –. Bach, Kent (/). ‘Failed reference and feigned reference: much ado about nothing’, Grazer Philosophische Studien : –. Bach, Kent (). ‘Intentions and demonstrations’, Analysis (): –. Bach, Kent (). ‘Conversational impliciture’, Mind & Language (): –. Bach, Kent (). ‘Giorgione was so-called because of his name’, Philosophical Perspectives : –. Bach, Kent (). ‘Descriptions: points of reference’, in Marga Reimer and Anne Bezuidenhout (eds), Descriptions and Beyond, Oxford: Clarendon Press, –. Bach, Kent (). ‘Replies to my critics’, Croatian Journal of Philosophy : –. Bach, Kent and Harnish, Robert (). Linguistic Communication and Speech Acts. Cambridge, MA: The MIT Press. Baddeley, A. (). Working Memory. Oxford, England: Clarendon Press. Baget, J.-F. and Mugnier, M.-L. (). ‘Extensions of simple conceptual graphs: the complexity of rules and constraints’, Journal of Artificial Intelligence Research : –. Baggio, G., van Lambalgen, M., and Hagoort, P. (). ‘Computing and recomputing discourse models: An ERP study’, Journal of Memory and Language : –. Baker, C. Leory (). unpublished. Definiteness and indefiniteness in English. MA thesis. Chicago, IL: University of Illionois. Bakhtin, M. (). The Dialogic Imagination. Austin: University of Texas Press. Bakhtin, M. (). Speech Genres and Other Late Essays. Austin: University of Texas Press. Bamberg, M. (). ‘A functional approach to the acquisition of anaphoric relationships’, Linguistics : –. Banfield, A. (). ‘Narrative style and the grammar of direct and indirect speech’, Foundations of Language (): –. Bangalore, S., Rambow, O., and Whittaker, S. (). ‘Evaluation metrics for generation’, in Proceedings of the st International Conference on Natural Language Generation (INLG), –, Mitzpe Ramon.

OUP CORRECTED PROOF – FINAL, 21/1/2019, SPi

REFERENCES



Barclay, M. (). Reference Object Choice in Spatial Language: Machine and Human Models. University of Exeter: Exeter. Bard, Ellen Gurman, Anderson, Anne H., Sotillo, Catherine, Aylett, Matthew, DohertySneddon, Gwyneth, and Newlands, Alison (). ‘Controlling the intelligibility of referring expressions in dialogue’, Journal of Memory and Language (): –. doi: ./ jmla... Bar-Hillel, Yehoshua (). Pragmatics of Natural Language. Dordrecht: Reidel. Barker, Chris (). ‘Partitives, double genitives, and anti-uniqueness’, Natural Language and Linguistic Theory : –. Barker, Chris (). ‘Definite possessives and discourse novelty’, Theoretical Linguistics : –. Barker, Chris (). ‘Possessive weak definites’, in Ji-Yung Kim, Yury A. Lander, and Barbara H. Partee (eds), Possessives and Beyond: Semantics and Syntax. University of Massachusetts Working Papers in Linguistics . Amherst: Massachusetts: GLSA. –. Barkley, C., Kluender, R., and Kutas, M. (). ‘Referential processing in the human brain: an event-related potential (ERP) study’, Brain Research : –. Barkley, C., Kluender, R., and Kutas, M. (in preparation). Elicitation of early negativity in sentence processing contexts depends on attentional efficiency. Barlew, Jefferson (). ‘The deictic motion verb come, discourse centers, and the representation of mental states’, in Proceedings of Sinn und Bedeutung . Barwise, Jon and Cooper, R. (). ‘Generalized quantifiers and natural language’, Linguistics and Philosophy : –. Bassano, D., Maillochon, I., and Mottet, S. (). ‘Noun grammaticalization and determiner use in French children’s speech: A gradual development with prosodic and lexical influences’, Journal of Child Language : –. Bäuerle, R. (). ‘Pragmatischsemantische Aspekte der NP-Interpretation’, in M. Faust, R. Harweg, W. Lehfeldt, and G. Wienold (eds), Allgemeine Sprachwissenschaft, Sprachtypologie und Textlinguistik, Tübingen: Narr, –. Beaumont, Ronald C. (). She Shashishalhem, the Sechelt Language: Language, Stories and Sayings of the Sechelt Indian People of British Columbia. Penticton, British Columbia: Theytus Books. Beaver, D. (). ‘Presupposition’, in van Benthem, J. and ter Meulen, A. (eds), Handbook of Logic and Language, pages, North Holland, –. Belz, Anja (). ‘That’s nice . . . what can you do with it? (last words)’, Computational Linguistics : –. Belz, Anja and A. Gatt (). ‘Intrinsic vs. extrinsic evaluation measures for referring expression generation’, in Proceedings of the th annual meeting of the association for computational linguistics (acl). Columbus, OH. Belz, Anja, E. Kow, J. Viethen, and A. Gatt (). ‘The GREC challenge : overview and evaluation results’, in Proceedings of the th International Natural Language Generation Conference (INLG), –. Belz, Anja, Eric Kow, Jette Viethen, and A. Gatt (). ‘Generating referring expressions in context: The GREC task evaluation challenges’, in Emiel Krahmer and Mariët Theune (eds), Empirical Methods in NLG, NLAI  Methods in Natural Language, Springer Verlag, Berlin/Heidelberg, –. doi: ./----_. Bennett-Kastor, T. (). ‘Noun phrases and coherence in child narratives’, Journal of Child Language : –.

OUP CORRECTED PROOF – FINAL, 21/1/2019, SPi



REFERENCES

Benoit, P. J. (). ‘Formal coherence production in children’s discourse’, First Language , Part  : –. Benthem, J. F. A. K. v. (). ‘Tense logic and standard logic’, Logique et Analyse : –. Berman, R. A. and Nir-Sagiv, B. (). ‘Linguistic indicators of inter-genre differentiation in later language development’, Journal of Child Language (): –. Besson, M. and Macar, F. (). ‘An event-related potential analysis of incongruity in music and other non-linguistic contexts’, Psychophysiology : –. Beun, Robbert-Jan and Cremers, Anita H. M. (). ‘Object reference in a shared domain of conversation’, Pragmatics & Cognition (½): –. Bianchi, Andrea (ed.) (). On Reference. Oxford: Oxford University Press. Birch, S. L., Albrecht, J. E., and Myers, J. L. (). ‘Syntactic focusing structures influence discourse processing’, Discourse Processes : –. Birner, Betty J. (). ‘Discourse entities and the referential/attributive distinction’, Paper presented at the Annual Meeting of the Linguistic Society of America, Chicago. Birner, Betty J. (). Introduction to Pragmatics. Wiley-Blackwell. Birner, Betty and Gregory Ward (). ‘Uniqueness, familiarity, and the definite article in English’, Proceedings of the Twentieth Annual Meeting of the Berkeley Linguistics Society: –. Birner, Betty and Gregory Ward (). Information Status and Noncanonical Word Order in English. Philadelphia: John Benjamins. Bittner, D. (). ‘Early functions of definite determiners and DPs in German First Language Acquisition’, in E. Stark, E. Leiss, and W. Abraham (eds), Nominal Determination: Typology, Contrast, Constraints, and Historical Emergence (Vol. ), Amsterdam: John Benjamins, –. Bizer, C., Lehmann, J., Kobilarov, G., Auer, S., Becker, C., Cyganiak, R., and Hellmann, S. (). ‘Dbpedia - a crystallization point for the web of data’, Web Semantics: Science, Services, and Agents on the World Wide Web (): –. Bloom, P. (). ‘Subjectless sentences in child language’, Linguistic Inquiry : –. Bock, J. Kathryn and Irwin, David E. (). ‘Syntactic effects of information availability in sentence production’, Journal of Verbal Learning and Verbal Behavior (): –. doi: ./S-()-. Bock, J. Kathryn and Warren, Richard K. (). ‘Conceptual accessibility and syntactic structure in sentence formulation’, Cognition (): –. doi: ./-()-X. Bock, J. Kathryn, Irwin, David E., and Davidson, Douglas J. (). ‘Putting first things first’, in John M. Henderson and Fernanda Ferreira (eds), The Interface of Language, Vision, and Action. Eye Movements and the Visual World, New York/Hove: Psychology Press, –. Boër, S. E. and Lycan, W. G. (). Knowing Who. MIT Press: Cambridge, Mass. Bohnet, B. and Dale, R. (). ‘Viewing referring expression generation as search’, in Proceedings of the th International Joint Conference on Artificial Intelligence (IJCAI), –, Edinburgh. Bolden, Galina (). ‘Multiple modalities in collaborative turn sequences’, Gesture : –. Bolinger, Dwight (). Syntactic Diffusion and the Indefinite Article. Bloomington: Indiana University Linguistics Club. Bolinger, Dwight (). Intonation and its Parts: Melody in Spoken English. Stanford, CA: Stanford University Press. Borthen, Kaja (). Norwegian Bare Singulars. PhD dissertation, Norwegian University of Science and Technology.

OUP CORRECTED PROOF – FINAL, 21/1/2019, SPi

REFERENCES



Borthen, Kaja (). ‘The distribution and interpretation of Norwegian “bare superlatives” ’, Working papers Isk . Department of Languages and Communication Studies, Norwegian University of Science and Technology. –. Bosch, Peter (). Agreement and Anaphora: A Study of the Role of Pronouns in Syntax and Discourse. New York: Academic Press. Bosch, Peter and Umbach, C. (). ‘Reference determination for demonstrative pronouns’, in Dagmar Bittner and Natalia Gargarina (eds), Intersentential Pronominal Reference in Child and Adult Language, Zentrum für Allgemeine Sprachwissenschaft: Berlin, Germany, –. Bosch, Peter, Rozario, T., and Zhao, Y. (). ‘Demonstrative pronouns and personal pronouns. German der vs. er.’, Proceedings of the EACL. Budapest. Workshop on The Computational Treatment of Anaphora, Springer: Berlin/Heidelberg, Germany. Bouma, G. and Hopp, H. (). ‘Coreference preferences for personal pronouns in German’, ZAS Papers in Linguistics : –. Bowdle, Brian F. and Gregory Ward (). ‘Generic demonstratives’, Proceedings of the Twenty-First Annual Meeting of the Berkeley Linguistics Society: –. Bozickovic, Vojislav (). ‘The Semantic insignificance of referential intentions’, Grazer Philosophische Studien : –. Brachman, R. J. and Schmolze, J. G. (). ‘An overview of the KL-ONE knowledge representation system’, Cognitive Science (): –. Branigan, Holly P. and Feleki, Eleonora (). ‘Conceptual accessibility and serial order in Greek speech production’, in M. Hahn and S. C. Stoness (eds), Proceedings of the Twenty First Annual Conference of the Cognitive Science Society, –. Branigan, Holly P., Pickering, Martin J., and Tanaka, Mikihiro (). ‘Contributions of animacy to grammatical function assignment and word order during production’, Lingua (): –. doi: ./j.lingua.... Branigan, Holly P., Pickering, M. J., Pearson, J., and McLean, J. F. (). ‘Linguistic alignment between people and computers’, Journal of Pragmatics : –. Bratman, Michael (). Intention, Plans, and Practical Reason. Cambridge, MA: Harvard University Press. Bratman, Michael (). ‘What is intention?’, in P. R. Cohen, J. Morgan, and M. E. Pollack (eds), Intentions in Communication. Cambridge, MA: MIT Press, –. Bratman, Michael (). ‘I Intend that we J’, in M. Bratman, Faces of Intention. Cambridge: Cambridge University Press, –. Braun, David (). ‘Empty names’, Noûs : –. Braun, David (). ‘Katz on Names Without Bearers’, Philosophical Review : –. Braun, David (). ‘Empty names, fictional names, mythical names’, Noûs (): –. Braun, David (). ‘Indexicals’, in E. N. Zalta (ed.), The Stanford Encyclopedia of Philosophy (Summer  Edition), The Metaphysics Research Lab, Center for the Study of Language and Information (CSLI), Stanford University, Stanford, California. . Breheny, R. (). ‘A lexical account of implicit (bound) contextual dependence’, in R. B. Young and Y. Zhou (eds), Proceedings of SALT XIII, CLC Publications: Cornell University, Ithaca, NY, –. Brennan, Susan E. (). ‘Centering attention in discourse’, Language and Cognitive Processes (): –. doi: ./. Brennan, Susan E. and Clark, H. H. (). ‘Conceptual pacts and lexical choice in conversation’, Journal of Experimental Psychology (): –.

OUP CORRECTED PROOF – FINAL, 21/1/2019, SPi



REFERENCES

Brennan, Susan E., Friedman, M. A., and Pollard, C. J. (). ‘A centering approach to pronouns’, in Proceedings of the th Annual Meeting of the Association for Computational Linguistics, Stanford, CA: Association for Computational Linguistics, –. Brogaard, B. (). ‘Descriptions: predicates or quantifiers?’, Australasian Journal of Philosophy : –. Brogaard, B. (Forthcoming). ‘An Empirically Informed Cognitive Theory of Propositions’, Canadian Journal of Philosophy. Bromberger, Sylvain and Halle, Morris (). ‘The Contents of phonological signs: a comparison between their use in derivational theories and in optimality theories’, in I. Roca (ed.), Derivations and Constraints in Phonology. Oxford: Oxford University Press, –. Brown, Gillian (). Speakers, Listeners, and Communication: Explorations in Discourse Analysis. Cambridge: Cambridge University Press. Brown-Schmidt, Sarah and Tanenhaus, Michael K. (). ‘Watching the eyes when talking about size: an investigation of message formulation and utterance planning’, Journal of Memory and Language (): –. doi: ./j.jml.... Brown-Schmidt, Sarah and Tanenhaus, Michael K. (). Real-Time Investigation of Referential Domains in Unscripted Conversation: A Targeted Language Game Approach, Cognitive Science, Vol. , Elsevier: New York, New York. doi: ./. Brown-Schmidt, Sarah, Byron, D. K, and Tanenhaus, M. (). ‘Beyond salience: interpretation of personal and demonstrative pronouns’, Journal of Memory and Language : –. Brown-Schmidt, Sarah, Christine Gunlogson, and Michael Tanenhaus (). ‘Addressees distinguish shared from private information when interpreting questions during interactive conversation’, Cognition : –. Bruner, J. S. (). Child’s Talk; Learning to use Language. New York: W. W. Norton & Company. Bruner, J. S. (). ‘Contexts & formats’, in M. Moscato and G. Piérault-Le Bonniec (eds), Le langage. Construction et Actualisation, Rouen: Université de Rouen, –. Brunetti, Lisa (). ‘On the semantic and contextual factors that determine topic selection in Italian and Spanish’, The Linguistic Review (–): –. doi: ./tlir... Burge, T. . ‘Reference and proper names’, Journal of Philosophy : –. Büring, D. (). ‘Crossover situations’, Natural Language Semantics (): –. Büring, D. (). Binding Theory. Cambridge University Press, Cambridge. Burkhardt, P. (). ‘Inferential bridging relations reveal distinct neural mechanisms: evidence from event-related brain potentials’, Brain and Language : –. Burton-Roberts, N. (). The Limits to Debate. Cambridge, Cambridge University Press. Buschmeier, H., Bergmann, K., and Kopp, S. (). ‘An alignment-capable microplanner for natural language generation’, in Proceedings of the th European Workshop on Natural Language Generation (ENLG), –. Cable, S. (). ‘A new argument for lexical decomposition: transparent readings of verbs’, Linguistic Inquiry (): –. URL: http://muse.jhu.edu/journals/linguistic inquiry/v/ ..cable.html. Callahan, S. M. (). ‘Processing anaphoric constructions: insights from electrophysiological studies’, Journal of Neurolinguistics : –. Callaway, Charles B. and Lester, James C. (). ‘Pronominalization in generated discourse and dialogue’, in Proceedings of the th Annual Meeting on Association for Computational Linguistics, –. doi: ./..

OUP CORRECTED PROOF – FINAL, 21/1/2019, SPi

REFERENCES



Cameron-Faulkner, T., Lieven, E. V. M., and Tomasello, M. (). ‘A construction based analysis of child directed speech’, Cognitive Science : –. Campbell, A. L., Brooks, P., and Tomasello, M. (). ‘Factors affecting young children’s use of pronouns as referring expressions’, Journal of Speech, Language, and Hearing Research : –. Campbell, John (). Reference and Consciousness. Oxford: Oxford University Press. Candlish, Stewart and Wrisley, George (). ‘Private language’, Stanford Encyclopedia of Philosophy, fall  edition, The Metaphysics Research Lab, Center for the Study of Language and Information (CSLI), Stanford University, Stanford California. Accessed at http://plato.stanford.edu/archives/fall/entries/private-language/ Caplan, B. (). ‘Creatures of fiction, myth, and imagination’, American Philosophical Quarterly : –. Caplan, D., Alpert, N., and Waters, G. (). ‘Effect of syntactic structure and propositional number on patterns of regional cerebral blood flow’, Journal of Cognitive Neuroscience : –. Capuano, Antonio (). ‘From having in mind to direct reference’, in W. P. Kabasenche, M. O’Rourke, and M. H. Slater (eds), Reference and Referring, Cambridge, MA: MIT Press, –. Carenini, G. and D. Moore (). ‘An empirical study of the influence of user tailoring on evaluative argument effectiveness.’ In Proceedings of the th International Joint Conference on Artificial Intelligence, Seattle WA, –. Carlson, Greg N. (). Reference to Kinds in English. New York: Garland Publishing. Carminati, M. N. (). The Processing of Italian Subject Pronouns. PhD dissertation. University of Massachusetts at Amherst. Carston, Robyn (). ‘Implicature, explicature, and truth-theoretic semantics’, in S. Davis (ed.), Pragmatics: A Reader, New York: Oxford University Press, –. Carston, Robyn (). Thoughts and Utterances: the Pragmatics of Explicit Communication. London, Blackwell. Cartwright, R. (). ‘Negative existentials’, Journal of Philosophy : –. Castañeda, H. (). ‘Indicators and quali-indicators’, American Philosophical Quarterly : –. Chafe, W. (). ‘Givenness, contrastiveness, definiteness, subjects, topics, and point of view’, in C. Li (ed.), Subject and Topic, New York: Academic Press, –. Chafe, Wallace L. (ed.) (). The Pear Stories: Cognitive, Cultural, and Linguistic Aspects of Narrative Production. Norwood, NJ: Ablex. Chafe, Wallace L. (). ‘Cognitive constraints on information flow,’ in Russell Tomlin (ed.), Coherence and Grounding Discourse, Amsterdam: John Benjamins, –. Chafe, W. (). Discourse, Consciousness, and Time: The Flow and Displacement of Conscious Experience in Speaking and Writing. Chicago: The University of Chicago Press. Chai, J., Hong, P., and Zhou, M. X. (). ‘A probabilistic approach to  reference resolution in multimodal user interfaces’, in Proceedings of the th International Conference on Intelligent User Interfaces, –. Chai, J., Prasov, Z., and Qu, S. (). ‘Cognitive principles in robust multimodal interpretation’, Journal of Artificial Intelligence Research : –. Chai, J., She, L., Fang, R., Ottarson, S., Littley, C., Liu, C., et al. (). ‘Collaborative effort towards common ground in situated human–robot dialogue’, in Proceedings of the  ACM/IEEE International Conference on Human-Robot Interaction, –).

OUP CORRECTED PROOF – FINAL, 21/1/2019, SPi



REFERENCES

Chambers, C. G., Tanenhaus, M., and Magnuson, J. (). ‘Actions and affordances in syntactic ambiguity resolution’, Journal of Experimental Psychology: Learning, Memory, and Cognition (): . Chambers, C. G., Tanenhaus, M. K, Eberhard, K. M., Filip, H., and Carlson, G. N. (). ‘Circumscribing referential domains in real-time sentence comprehension’, Journal of Memory and Language : –. Chantree, F., Kilgarriff, A., de Roeck, A., and Willis, A. (). ‘Disambiguating coordinations using word distribution information’, in Proceedings of the International Conference on Recent Advances in Natural Language Processing (RANLP). Borovets, Bulgaria. Charolles, M. (). La Référence et Les Expressions Référentielles En Français. Paris: Ophrys. Chastain, Charles (). ‘Reference and context’, in Keith Gunderson (ed.), Minnesota Studies in the Philosophy of Science, Volume : Language Mind and Knowledge, Minneapolis: University of Minnesota Press, –. Chenu, F., Jisa, H., and Mazur-Palandre, A. (). ‘Développement de la connectivité syntaxique à travers deux types de textes à l’oral et à l’écrit’, Paper presented at the SHS Web of Conferences : Congrès Mondial de Linguistique Française - CMLF . Chierchia, Gennaro (). ‘Anaphora and attitudes de se’, in R. Bartsch, J. van Benthem, and van Emde Boas, (eds), Semantics and Contextual Expression, Dordrecht: Foris, –. Chierchia, Gennaro (). ‘A puzzle about indefinites’, in C. Checchetto and G. Chierchia (eds), Semantic Interfaces: Reference, Anaphora, and Aspect. Stanford, CA: CSLI Publications, –. Chierchia, Gennaro (). ‘Definites, ocality, and intentional identity’, in G. N. Carlson and F. J. Pelletier (eds), Reference and Quantification: The Partee Effect. Stanford, CA: CSLI Publications, –. Chierchia, Gennaro and Mats Rooth (). ‘Configurational notions in Discourse Representation Theory’, in Charles Jones and Peter Sells (eds) Proceedings of NELS , Graduate Linguistics Student Association, University of Massachusetts, Amherst. Chiriacescu, Sofiana Iulia (). ‘Effects of reference form on frequency of mention and rate of pronominalization’, in I. Hendricks, S. L. Davi, A. Branco, and R. Mitkov (eds), Anaphora Processing and Applications, th Discourse Anaphora and Anaphor Resolution Colloquium, DAARC, Springer: Berlin/Heidelberg, Germany, –. Chiriacescu, Sofiana (). The discourse structuring potential of indefinite noun phrases. Special markers in Romanian, German, and English. PhD dissertation. Stuttgart: University of Stuttgart. Chiriacescu, Sofiana and Klaus von Heusinger (). ‘Discourse prominence and pe-marking in Romanian’, International Review of Pragmatics (): –. Chomsky, Noam (). ‘Deep structure, surface structure, and semantic interpretation’, in Danny Steinberg and Leon Jakobovits (eds), Semantics: An Interdisciplinary Reader in Linguistics, Philosophy, and Psychology, Cambridge: Cambridge University Press, –. Chomsky, Noam (). Lectures on Binding and Government. Foris Publications, Dordrecht. Christianson, Kiel and Ferreira, Fernanda (). ‘Conceptual accessibility and sentence production in a free word order language (Odawa)’, Cognition (): –. doi: ./j. cognition.... Christophersen, Paul (). The Articles: A Study of their Theory and Use in English. Copenhagen: Munksgaard. Chung, I., Propp, O., Walter, M. R., and Howard, T. M. (). ‘On the performance of hierarchical distributed correspondence graphs for efficient symbol grounding of robot

OUP CORRECTED PROOF – FINAL, 21/1/2019, SPi

REFERENCES



instructions’, in Proceedings of the  IEEE/RSJ International Conference on Intelligent Robots and Systems, –). Chung, Sandra and Ladusaw, William (). Restriction and Saturation. Cambridge, MA: MIT Press. Clancy, P. (). ‘Discourse motivations for referential choice in Korean acquisition’, in H. Sohn and J. Haig (eds), Japanese/Korean Linguistics (Vol. ), Standford: CSLI, –. Clancy, P. (). ‘The discourse basis of constructions: some evidence from Korean acquisition’, Paper presented at the nd Stanford Child Language Research Forum, Standford. Clapp, L. (). ‘The problem of negative existentials does not exist: a case for dynamic semantics’, Journal of Pragmatics : –. Clark, Andy (). Mindware: An Introduction to the Philosophy of Cognitive Science. nd edition. Oxford: Oxford University Press. Clark, E. V. (). ‘Pragmatics in acquisition’, Journal of Child Language : –. doi: ./S. Clark, E. V. and Bernicot, J. (). ‘Repetition as ratification: how parents and children place information in common ground’, Journal of Child Language (): –. Clark, E. V. and Sengul, C. (). ‘Strategies in the acquisition of deixis’, Journal of Child Language (): –. Clark, H. H. (). ‘Bridging’, in R.C. Schank and B.L. Nash-Webber (eds) Theoretical Issues in Natural Language Processing. Association for Computing Machinery, New York. Reprinted in P. N. Johnson-Laird and P. C. Wasow (eds) ()Thinking. Cambridge University Press, –. Clark, H. H. (). Using Language. Cambridge: Cambridge University Press. Clark, H. H. (). ‘Coordinating with each other in a material world’, Discourse Studies : –. Clark, H. H. and Bangerter, A. (). ‘Changing ideas about reference’, in I. A. Noveck and D. Sperber (eds), Experimental Pragmatics, Basingstoke: Palgrave Macmillan, –. Clark, H. H. and Haviland, S. E. (). ‘Comprehension and the given-new contract’, in R. O. Freedle (ed.), Discourse Production and Comprehension, Erlbaum, –. Clark, H. H. and Henetz, Tania (). ‘Working together’, in T. Holtgraves (ed.), Oxford Handbook of Language and Social Psychology. Oxford: Oxford University Press, –. Clark, H. H. and Marshall, C. (). ‘Definite reference and mutual knowledge’, in A. K. Joshi et al. (eds), Elements of Discourse Understanding, Cambridge University Press, –. Clark, H. H. and Murphy, G. (). ‘Audience design in meaning and reference’, in J. F. L. Ny and W. Kintsch (eds), Language and Comprehension, North Holland, –. Clark, H. H., and Sengul, C. J. (). ‘In search of referents for nouns and pronouns’, Memory & Cognition : , –. doi: ./BF. Clark, H. H. and Wilkes-Gibbs, Deanna (). ‘Referring as a collaborative process’, Cognition (): –. doi: ./-()-. Clark, H. H. and Wilkes-Gibbs, Deanna (). ‘Referring as a collaborative process’, in P. R. Cohen, J. Morgan, and M. E. Pollack (eds), Intentions in Communication. Cambridge, MA: MIT Press, –. Clark, H. H., Schreuder, Robert, and Buttrick, Samuel (). ‘Common ground at the understanding of demonstrative reference’, Journal of Verbal Learning and Verbal Behavior : –. doi: ./S-()-. Clarke, A., Elsner, M., and Rohde, H. (). ‘Where’s wally: the influence of visual salience on referring expression generation’, Frontiers in Psychology ().

OUP CORRECTED PROOF – FINAL, 21/1/2019, SPi



REFERENCES

Coco, Moreno I. and Keller, Frank (). ‘The impact of visual information on reference assignment in sentence production’, in Niels A. Taatgen and Hedderik van Rijn (eds), st Annual Conference of the Cognitive Science Society, –. doi: ./j....x. Coco, Moreno I. and Keller, Frank (). ‘Sentence production in naturalistic scenes with referential ambiguity’, in S. Ohlsson and R. Catrambone (eds), Proceedings of the nd Annual Conference of the Cognitive Science Society, The Cognitive Science Society, Inc., Austin, Texas, –. Coene, M. (). ‘On the acquisition of the indefinite article: a cross-linguistic study of French, Italian, Romanian, and Spanish child speech’, in S. Vogeleer (ed.), Bare Plurals, Indefinites, and Weak–Strong Distinction, Amsterdam: John Benjamins, –. Cohen, P. R. and Levesque, H. J. (). ‘Speech acts and rationality’, in Proceedings of the rd Annual Meeting of the Association of Computational Linguists (ACL), –, Chicago, Illinois. Çokal, D., Sturt, P., and Ferreira, F. (). ‘Processing of it and this in written narrative discourse’, Discourse Processes, Taylor & Francis: Oxfordshire. http://dx.doi.org/./ X... Colonna, S., Schimke, S., and Hemforth, B. (). ‘Le rôle de la structure informationnelle dans l’interprétation d’une anaphore pronominale inter-phrastique en français’, in F. Neveu, Muni Toke V., Durand J., Klingler T., Mondada L., and Prévost S. (eds), Congrès Mondial de Linguistique Française, Paris, France, –. Comrie, B. (). ‘Pragmatic binding: demonstratives as anaphors’ [in Dutch], in M. L. Juge and J. L. Moxley (eds), Proceedings of the Twenty-third Annual Meeting of the Berkeley Linguistics Society, Berkeley, CA: Berkeley Linguistics Society, –. Cornish, F. (). Anaphora, Discourse, and Understanding. New York: Oxford University Press. Coulson, S., King, J., and Kutas, M. (a). ‘Expect the unexpected: event-related brain response to morphosyntactic violations’, Language and Cognitive Processes (): –. Coulson, S., King, J., and Kutas, M. (b). ‘ERPs and domain specificity: beating a straw horse’, Language and Cognitive Processes : –. Cowan, N. (). Attention and Memory: An Integrated Framework. Oxford University Press. Cowart, W. and Cairns, H. (). ‘Evidence for an anaphoric mechanism within syntactic processing: some reference relations defy semantic and pragmatic constraints’, Memory & Cognition : –. Cowles, H. W. (). Processing information structure: Evidence from comprehension and production. PhD dissertation, UCSD, San Diego, CA. Cowles, H. W., Walenski, M., and Kluender, R. (). ‘Linguistic and cognitive prominence in anaphor resolution: topic, constrastive focus, and pronouns’, Topoi : –. Crawley, R. J. and Stevenson, R. J. (). ‘Reference in single sentences and in texts’, Journal of Psycholinguistic Research (): –. Cresswell, M. J. (). Logics and Languages. London: Methuen. Cresswell, M. J. (). Entities and Indices. Kluwer: Dordrecht. Cresti, Diana (). ‘Extraction and reconstruction’, Natural Language Semantics : –. Croitoru, M. and van Deemter, K. (). ‘A conceptual graph approach to the generation of referring expressions’ in Proceedings of the th International Joint Conference on Artificial Intelligence (IJCAI), –, Hyderabad, India. Culy, C. (). ‘Aspects of logophoric marking’, Linguistics : –.

OUP CORRECTED PROOF – FINAL, 21/1/2019, SPi

REFERENCES



Cumming, Samuel (). ‘Variabilism’, Philosophical Review (): –. Cumming, Samuel (). ‘Names’, Stanford Encyclopedia of Philosophy. Dahl, Deborah A. (). ‘Recognizing specific attributes’, paper presented at the Annual Meeting of the Linguistic Society of America, Baltimore. Dahl, Deborah A. (). ‘Focusing and reference resolution in Pundit’, paper presented at AAAI-, Philadelphia. Dahl, Östen and Fraurud, Kari (). ‘Animacy in grammar and discourse’, in Thorstein Fretheim and Jeanette Gundel (eds), Reference and Referent Accessibility, Amsterdam: John Benjamins, –. Dale, R. (). ‘Cooking up referring expressions’, in Proceedings of the th Annual Meeting of the Association for Computational Linguistics (ACL), –. Dale, R. (). Generating Referring Expressions: Constructing Descriptions in a Domain of Objects and Processes. Cambridge, Massachusetts: The MIT Press. Dale, R. and Haddock, N. (). ‘Generating referring expressions involving relations’, in Proceedings of the th Conference of the European Chapter of the Association of Computational Linguists (EACL), –, Berlin. Dale, R. and Reiter, E. (). ‘Computational interpretations of the Gricean maxims in the generation of referring expressions’, Cognitive Science (): –. Dale, R. and Viethen, J. (). ‘Attribute-centric referring expression generation’, in E. Krahmer and M. Theune (eds), Empirical Methods in Natural Language Generation, Berlin: Springer Verlag, –. Davey, A. (). The Formalisation of Discourse Production. PhD thesis, University of Edinburgh, Edinburgh, Scotland. Davidson, D. (). ‘Truth and meaning’, Synthese : –. De Cat, C. (a). ‘Early “pragmatic” competence and the null subject phenomenon’, in R. Bok-Bennema, B. Hollebrandse, B. Kampers-Manhe, and P. Sleeman (eds), Romance Language and Linguistic Theory . Selected Papers from Going Romance , Amsterdam: John Benjamins, –. De Cat, C. (b). ‘A fresh look at how young children encode new referents’, International Review of Applied Linguistics, : –. De Cat, C. (). ‘Egocentric definiteness errors and perspective evaluation in preschool children’, Journal of Pragmatics, : –. de Hoop, Helen (). ‘On the characterization of the weak–strong distinction’, in E. Bach, E. Jelinek, A. Kratzer, and B. Partee (eds), Quantifcation in Natural Languages, Dordrecht: Kluwer Academic Publishers, –. de Hoop, Helen (). ‘On the interpretation of stressed pronouns’, in M. Weisgerber (ed.), Proceedings of the th Sinn und Bedeutung conference, Germany: Universität Konstanz, –. de la Fuente, I. and Hemforth, B. (). ‘Effects of clefting and left-dislocation on subject and object pronoun resolution in Spanish’, in J. Cabrelli Amaro et al. (eds), Selected Proceedings for the th Hispanic Linguistics Symposium, Somerville, MA: Cascadilla Proceedings Project, –. de Weck, G. (). La cohésion dans les textes d’enfants. Etude du développement des processus anaphoriques. Neuchâtel: Delachaux et Niestlé. de Weck, G. and Jullien, S. (). ‘How do French-speaking children with Specific Language Impairment first mention a referent in storytelling? Between reference and grammar’, Journal of Pragmatics, : –.

OUP CORRECTED PROOF – FINAL, 21/1/2019, SPi



REFERENCES

de Weck, G., Heurdier, Hassan, J., Klein, J., and Nashawati, S. (submitted). ‘Activities, contexts, and the construction of reference’, in A. Salazar Orvig, G. de Weck, R. Hassan, and A. Rialland (eds), The Acquisition of Referring Expressions: a Dialogic Approach. Deal, Amy Rose (). ‘Nez Perze embedded indexicals’, in Proceedings of SULA. Amherst: GLSA. Dechaine, Rose-Marie and Mireille Tremblay (). ‘Deriving nominal reference’, paper presented at the Western Conference on Linguistics (WECOL), Department of Linguistics, University of California, Fresno, California. http://www.fresnostate.edu/artshum/linguis tics/wecolproceedings.html Deichsel, Annika (). The Semantics and Pragmatics of the Indefinite Demonstrative Dieser in German. PhD dissertation. Stuttgart: University of Stuttgart. Deichsel, Annika and von Heusinger, Klaus (). ‘The cataphoric potential of indefinites in German’, in I. Hendrickx, S. Lalitha Devi, A. Branco, and R. Mitkov (eds), Anaphora and Reference Resolution. th Discourse Anaphora and Anaphor Resolution Colloquium, DAARC  Faro, Portugal, – October , Revised Selected Papers. Heidelberg: Springer, –. Dekker, Paul (). ‘Meaning and use of indefinite expressions’, Journal of Logic, Language, and Information : –. Dekker, Paul (). ‘The pragmatic dimension of indefinites’, Research on Language and Computation : –. Demuth, K. and McCullough, E. (). ‘The prosodic (re)organization of children’s early English articles’, Journal of Child Language : –. Denis, A. (). ‘Generating referring expressions with reference domain theory’, in Proceedings of the th International Natural Language Generation Conference (INLG). Trim, Ireland. DeVault, D., Rich, C., and Sidner, C. L. (). ‘Natural language generation and discourse context: Computing distractor sets from the focus stack’, in Proceedings of the th International Meeting of the Orida Artificial Intelligence Research Society (AIRS). Miami Beach. Di Eugenio, B., Jordan, P. W., Thomason, R. H., and Moore, J. D. (). ‘The agreement process: an empirical investigation of human–human computer-mediated collaborative dialogs’, International Journal of Human-Computer Studies : –. Dice, L. R. (). ‘Measures of the amount of ecologic association between species’, Ecology, , –. Dickson, W. Patrick (). ‘Two decades of referential communication research: a review and meta-analysis’, in C. Brainerd and M. Pressley (eds), Verbal Processes in Children. Berlin: Springer-Verlag, –. Diesing, Molly (). Indefinites. Cambridge, MA: MIT Press. Diessel, Holger (). Demonstratives: Form, Function, and Grammaticalization. [Typological Studies in Language ]. Amsterdam: John Benjamins. Diessel, Holger (). ‘Deixis and demonstratives’, in C. Maienborn, K. von Heusinger, and P. Portner (eds), Semantics: An International Handbook of Natural Language Meaning, Volume  (HSK .), Berlin/Boston: Mouton de Gruyter, – [Chapter ]. Dikker, S. and Pylkkänen, L. (). ‘Before the N: effects of lexico-semantic violations in visual cortex’, Brain and Language (–): –. Dikker, S., Rabagliati, H., Farmer, T., and Pylkkänen, L. (). ‘Early occipital sensitivity to syntactic category is based on form typicality’, Psychological Science (): –. Doddington, G. (). ‘Automatic evaluation of machine translation quality using n-gram co-occurrence statistics’, in Proceedings of the nd International Conference on Human Language Technology Research (HLT), –.

OUP CORRECTED PROOF – FINAL, 21/1/2019, SPi

REFERENCES



Donchin, E. and Coles, M. (). ‘Is the P component a manifestation of context updating?’, Behavioral and Brain Sciences (): –. Donnellan, Keith (). ‘Reference and definite descriptions’, Philosophical Review : –. Donnellan, Keith S. (). ‘Putting Humpty Dumpty back together again’, The Philosophical Review (): –. Donnellan, Keith S. (). ‘Proper names and identifying descriptions’, Synthese : –. Donnellan, Keith (). ‘Speaking of nothing’, Philosophical Review : –. Donnellan, Keith S. (). ‘Speaker reference, descriptions, and anaphora’, in P. A. French, T. E. Uehling, Jr., and H. K. Wettstein (eds), Contemporary Perspectives in the Philosophy of Language. Minneapolis: University of Minnesota Press, –. Doran, Ryan and Gregory Ward (). ‘Proximal demonstratives in predicate NPs’, Proceedings of the Forty-First Annual Meeting of the Berkeley Linguistics Society: –. Doran, Ryan, and Gregory Ward (in preparation). Bridging demonstratives. Dowty, D. (). Word Meaning and Montague Grammar. Dordrecht: Reidel. Dressler, W. U., Kilani-Schoch, M., and Klampfer, S. (). ‘How does a child detect morphology? Evidence from production’, in W. Bisang, H. H. Hock, and W. Winter (eds), Trends in Linguistics. Studies and Monographs: Vol. , Berlin/New York: Mouton de Gruyter, –. Dretske, Fred (). Explaining Behavior. Cambridge, MA: MIT Press. Du Bois, John W. (). ‘Beyond definiteness: the trace of identity in discourse’, in Wallace L. Chafe (ed.), The Pear Stories: Cognitive, Cultural, and Linguistic Aspects of Narrative Production, Norwood, NJ: Ablex, –. Duan, Manjuan, Elsner, Micha, and de Marneffe, Marie-Catherine (). ‘Visual and linguistic predictors for the definiteness of referring expressions’, Proceedings of the th Workshop on the Semantics and Pragmatics of Dialogue (SemDial). Dutilh Novaes, C. (). ‘A medieval reformulation of the de dicto/de re distinction’, in Libor Behounek (ed.), Logica Yearbook, College Publications: London, –. URL: http://sta_.science.uva.nl/dutilh/articles/A%Medieval%Reformulation%of%the %de%Dicto%. Duvallet, F., Walter, M. R., Howard, T., Hemachandra, S., Oh, J., Teller, S., et al. (). ‘Inferring maps and behaviors from natural language instructions’, in Proceedings of the  International Symposium on Experimental Robotics, –). É. Kiss, Katalin (). The Syntax of Hungarian. Cambridge University Press. Eberhard, K. M., Spivey-Knowlton, M. J., Sedivy, J. C., and Tanenhaus, M. K. (). ‘Eye movements as a window into real-time spoken language comprehension in natural contexts’, Journal of Psycholinguistic Research (): –. Ebert, Christian, Cornelia Ebert, and Stefan Hinterwimmer (). ‘The interpretation of the German specificity markers bestimmt and gewiss’, in C. Ebert and S. Hinterwimmer (eds), Different Kinds of Specificity across Languages (Studies in Linguistics and Philosophy ). Heidelberg/Berlin: Springer, –. Ebert, Christian, Cornelia Endriss, and Stefan Hinterwimmer (). ‘Embedding topiccomment structures results in intermediate scope readings’, in A. Schardl, M. Walkow, and M. Abdurrahman (eds), Proceedings of North Eastern Society of Linguistics (=NELS) . Amherst, MA: GLSA Publications, –. Ebert, Cornelia and Hinterwimmer, Stefan (eds) (). Different Kinds of Specificity across Languages. Dordrecht: Springer. Egli, Urs and von Heusinger, Klaus (). ‘The epsilon operator and e-type pronouns’, in U. Egli, P. Pause, C. Schwarze, A. von Stechow, and G. Wienold (eds), Lexical Knowledge

OUP CORRECTED PROOF – FINAL, 21/1/2019, SPi



REFERENCES

in the Organization of Language, Amsterdam: John Benjamins Publishing Company, –. Elbourne, Paul D. (). Situations and Individuals. MIT Press. Elbourne, Paul D. (). ‘Demonstratives as individual concepts’, Linguistics and Philosophy : –. Elbourne, Paul D. (). Definite Descriptions. Oxford: Oxford University Press. Ellert, M. (). ‘Information structure affects the resolution of the subject pronouns er and der in spoken German discourse’, Discours : doi: ./discours.. Emslie, H. C. and Stevenson, R. J. (). ‘Pre-school children’s use of the articles in definite and indefinite referring expressions’, Journal of Child Language : –. Enç, Mürvet (). ‘Toward a referential analysis of temporal expressions’, Linguistics and Philosophy : –. Enç, Mürvet (). ‘The semantics of specificity’, Linguistic Inquiry : –. Endriss, Cornelia (). Quantificational Topics. A Scopal Treatment of Exceptional Wide Scope Phenomena. Berlin: Springer. Enfield, N. J. (). The Anatomy of Meaning: Speech, Gesture, and Composite Utterances. Cambridge: Cambridge University Press. Engelhardt, Paul E., Bailey, Karl G. D., and Ferreira, Fernanda (). ‘Do speakers and listeners observe the Gricean Maxim of Quantity?’, Journal of Memory and Language (): –. doi: ./j.jml.... Engonopoulos, N., Villalba, M., Titov, I., and Koller, A. (). ‘Predicting the resolution of referring expressions from user behavior’, in Proceedings of the Conference on Empirical Methods in Natural Language Processing (EMNLP), Short Papers, Seattle. Enqi, Morvet. (). ‘The semantics of specificity’, Linguistic Inquiry (): –. Epley, Nicholas, Carey Morewedge, and Boaz Keysar (). ‘Perspective taking in children and adults: equivalent egocentrism but differential correction’, Journal of Experimental Social Psychology : –. Erkü, Feride and Gundel, Jeanette K. (). ‘The pragmatics of indirect anaphors’, in Verschueren, J. and Bertuccelli, M. (eds), The Pragmatic Perspective: Selected Papers from the  International Pragmatics Conference. Amsterdam: John Benjamins. –. Evans, Gareth (). ‘Pronouns, quantifiers, and relative clauses (I)’, Canadian Journal of Philosophy : –. Evans, Gareth (). ‘Pronouns’, Linguistic Inquiry (): –. Evans, Gareth (). Varieties of Reference. [published posthumously, edited by John McDowell] Oxford: Oxford University Press. Evans, Gareth (/). ‘The causal theory of names’, in Collected Papers. Oxford: Oxford University Press, –. Everett, A. (). ‘Empty names and ‘gappy’ propositions’, Philosophical Studies (): –. Everett, A. (). ‘Pretense, existence, and fictional objects’, Philosophy and Phenomenological Research : –. Fang, R., Doering, M., and Chai, J. Y. (). ‘Collaborative models for referring expression generation in situated dialogue’, in Proceedings of the Twenty-Eighth AAAI Conference on Artificial Intelligence (AAAI), Quebec City. Fang, R., Liu, C., and Chai, J. (). ‘Integrating word acquisition and referential grounding towards physical world interaction’, in Proceedings of the th ACM International Conference on Multimodal Interaction, –.

OUP CORRECTED PROOF – FINAL, 21/1/2019, SPi

REFERENCES



Fano, R. (). Transmission of Information: A Statistical Theory of Communications. MIT Press: Cambridge, Mass. Farkas, Donka (). ‘Quantifier scope and syntactic islands’, in R. Hendrick, C. Masek, and M. F. Miller (eds), Papers from the Seventeenth Regional Meeting Chicago Linguistics Society (=CLS) . Chicago, IL: University of Chicago, –. Farkas, Donka (). ‘Specificity and scope’, in L. Nash, and G. Tsoulas (eds), Actes du Premier Colloque Langues & Grammaire . Paris: Université Paris-, –. Farkas, Donka (). ‘Specificity distinction’, Journal of Semantics : –. Farkas, Donka and von Heusinger, Klaus (). Stability of Reference and Object Marking in Romanian. Ms. Stuttgart: University of Stuttgart. Fasola, J. and Matarić, M. J. (). ‘Using semantic fields to model dynamic spatial relations in a robot architecture for natural language instruction of service robots’, in IEEE/RSJ International Conference on Intelligent Robots and Systems, –. Fasola, J. and Matarić, M. J. (). ‘Interpreting instruction sequences in spatial language discourse with pragmatics towards natural human–robot interaction’, in IEEE International Conference on Robotics and Automation, –. Fedele, E. (). Discourse Level Processing and Pronoun Interpretation. PhD Dissertation, University of Southern California. Fedele, E. and Kaiser, E. (). ‘Looking back and looking forward: anaphora and cataphora in Italian’, Proceedings of the  th Annual Penn Linguistics Colloquium : –. Felser, C., Clahsen, H., and Münte, T. (). ‘Storage and integration in the processing of filler-gap dependencies: an ERP study of topicalization and wh-movement in German’, Brain and Language (): –. Ferreira, F. (). ‘Choice of passive voice is affected by verb type and animacy’, Journal of Memory and Language (): –. Ferreira, F. (). ‘The misinterpretation of noncanonical sentences’, Cognitive Psychology, , –. Ferreira, F., Bailey, K. G. D., and Ferraro, V. (). ‘Good-enough representations in language comprehension’, Current Directions in Psychological Science : –. Ferreira, Victor S. and Yoshita, Hiromi (). ‘Given-new ordering effects on the production of scrambled sentences in Japanese’, Journal of Psycholinguistic Research (): –. doi: ./A:. Ferreira, Victor S., Slevc, L. Robert, and Rogers, Erin S. (). ‘How do speakers avoid ambiguous linguistic expressions?’, Cognition (): –. doi: ./j.cognition.... Ferstl, E. C., Garnham, A., and Manouilidou, C. (). ‘Implicit causality bias in English: a corpus of  verbs’, Behavior Research Methods (): –. Fiebach, C. J., Schlesewsky, M., and Friederici, A. D. (). ‘Separating syntactic memory costs and syntactic integration costs: the processing of German wh-questions’, Journal of Memory and Language (): –. Filiaci, F. (). ‘Null and overt subject biases in Spanish and Italian: a cross-linguistic comparison’, in C. Borgonovo et al. (eds), Selected Proceedings for the th Hispanic Linguistics Symposium, Somerville, MA: Cascadilla Proceedings Project, –. Filiaci, F., Sorace, A., and Carreiras, M. (). ‘Anaphoric biases of null and overt subjects in Italian and Spanish: a cross-linguistic comparison’, Language and Cognitive Processes (): –. Fillmore, C. J. (). On the Syntax of Preverbs, Glossa Vol , –. New York: Glossa Society.

OUP CORRECTED PROOF – FINAL, 21/1/2019, SPi



REFERENCES

Fillmore, Charles J. (). Santa Cruz lectures on deixis. Bloomington: Indiana Linguistics Club. Fillmore, Charles J. (). ‘Toward a descriptive framework for spatial deixis’, in R. J. Jarvella and W. Klein, Speech, Place, and Action, Chichester, NY: John Wiley & Sons, –. Fine, K. (). ‘The problem of non-existents. I. Internalism’, Topoi : –. FitzGerald, N., Artzi, Y., and Zettlemoyer, L. (). ‘Learning distributions over logical forms for referring expression generation’, in Proceedings of the  Conference on Empirical Methods in Natural Language Processing (EMNLP), –, Seattle, Washington. Flores d’Arcais, Giovanni B. (). ‘Some perceptual determinants of sentence construction’, in G. B. Flores d’Arcais (ed.), Studies in Perception: Festschrift for Fabio Metelli, Milan, Italy: Martello-Guinti, –. Fludernik, M. (). The Fictions of Language and the Languages of Fiction: The Linguistic Representation of Speech and Consciousness. London: Routledge. Fodor, J. (). The Linguistic Description of Opaque Contexts. PhD thesis, Massachusetts Institute of Technology. Fodor, Janet (). The Linguistic Description of Opaque Contexts. PhD dissertation. Bloomington, IN: Linguistic Club. Fodor, Janet Dean and Sag, Ivan A. (). ‘Referential and quantificational indefinites’, Linguistics and Philosophy : –. Fodor, Jerry (). Psychosemantics: The Problem of Meaning in the Philosophy of Mind. Cambridge, MA: MIT Press. Foraker, Stephani and McElree, Brian (). ‘The role of prominence in pronoun resolution: active versus passive representations’, Journal of Memory and Language : –. doi: ./j.jml.... Ford, W. and Olson, D. (). ‘The elaboration of the noun phrase in children’s object descriptions’, Journal of Experimental Child Psychology : –. Fossard, Marion, Garnham, Alan, and Cowles, H. Wind (). ‘Between anaphora and deixis . . . the resolution of the demonstrative noun phrase “that N”’, Language and Cognitive Processes (): –. Fox, B. A. (). Discourse Structure and Anaphora. Cambridge: Cambridge University Press. Fox, Barbara (). ‘Interactional reconstruction in real-time language processing’, Cognitive Science : –. François, F. (). ‘Problèmes et esquisse méthodologique’, in F. François, C. Hudelot, and E. Sabeau-Jouannet, Conduites Linguistiques chez le Jeune Enfant, Paris: PUF, –. Frank, M. and Goodman, N. (). ‘Predicting pragmatic reasoning in language games’, Science (): . Franke, M. and Degen, J. (). ‘Reasoning in reference games: individual vs. population-level probabilistic modeling’, PLoS ONE (): –. Fraurud, Kari (). ‘Definiteness and the processing of noun phrases in natural discourse’, Journal of Semantics : –. Frege, G. (). Begriffsschrift, eine der arithmetischen nachgebildeten Formelsprache des reinen Denkens. Halle: Nebert. Frege, G. (). ‘Über Sinn und Bedeutung’, Zeitschrift für Philosophie und philosophische Kritik : –. [Translated as ‘On Sense and Reference’ in Geach and Black (: –)]. Frege, G. (). The Foundations of Arithmetic. [Translated by J. L. Austin] Evanston, Illinois: Northwestern University Press.

OUP CORRECTED PROOF – FINAL, 21/1/2019, SPi

REFERENCES



Freudenthal, D., Pine, J. M., and Gobet, F. (). ‘Understanding the developmental dynamics of subject omission: The role of processing limitations in learning’, Journal of Child Language (): . Friederici, A. D. (). ‘Towards a neural basis of auditory sentence processing’, Trends in Cognitive Sciences (): –. Friederici, A. D., Pfeifer, E., and Hahne, A. (). ‘Event-related brain potentials during natural speech processing: effects of semantic, morphological, and syntactic violations’, Cognitive Brain Research (): –. Fukumura, K., Hyönä, J., and Scholfield, M. (). ‘Gender affects semantic competition: the effect of gender in a non-gender-marking language’, Journal of Experimental Psychology: Learning, Memory, and Cognition (): –. Fukumura, Kumiko and van Gompel, Roger P. G. (). ‘Choosing anaphoric expressions: do people take into account likelihood of reference?’, Journal of Memory and Language (): –. doi: ./j.jml.... Fukumura, Kumiko and van Gompel, Roger P. G. (). ‘The effect of animacy on the choice of referring expression’, Language and Cognitive Processes (): –. doi: ./ ... Fukumura, K. and van Gompel, R. P. G. (). ‘Producing pronouns and definite noun phrases: do speakers use the addressee’s discourse model?’, Cognitive Science : –. Fukumura, Kumiko, Hyönä, Jukka, and Scholfield, Merete (). ‘Gender affects semantic competition: the effect of gender in a non-gender-marking language’, Journal of experimental psychology. Learning, memory, and cognition : , –. doi: ./ a. Fukumura, Kumiko, van Gompel, Roger P. G., and Pickering, Martin J. (). ‘The use of visual context during the production of referring expressions’, Quarterly Journal of Experimental Psychology () (): –. doi: ./. Fukumura, Kumiko, van Gompel, Roger P. G., Harley, Trevor, and Pickering, Martin J. (). ‘How does similarity-based interference affect the choice of referring expression?’, Journal of Memory and Language (): –. doi: ./j.jml.... Fuller, Judith and Gundel, Jeanette K. (). ‘Topic prominence in inter-language’, Language Learning : –. Galati, Alexia and Brennan, Susan E. (). ‘Attenuating information in spoken communication: for the speaker, or for the addressee?’, Journal of Memory and Language (): –. doi: ./j.jml.... Gardent, C. (). ‘Generating minimal definite descriptions’, in Proceedings of the th Annual Meeting of the Association for Computational Linguistics (ACL), –, Philadelphia. Gardent, C. and Striegnitz, K. (). ‘Generating bridging definite descriptions’, in H. Bunt and R. Muskens (eds), Computing Meaning [Volume ], Studies in Linguistics and Philosophy, Springer Publishers, –. Garey, M. R., and Johnson, D. S. (). Computers and Intractability: A Guide to the Theory of NP-Completeness. New York: W.H. Freeman. Garnham, A. (). Mental Models and the Interpretation of Anaphora. Hove, UK: Psychology Press. Garnham, A., Oakhill, J., and Cruttenden, H. (). ‘The role of implicit causality and gender cue on the interpretation of pronouns’, Language and Cognitive Processes, , –.

OUP CORRECTED PROOF – FINAL, 21/1/2019, SPi



REFERENCES

Garoufi, K. and Koller, A. (). ‘Combining symbolic and corpus-based approaches for the generation of successful referring expressions’, in Proceedings of the th European Workshop on Natural Language Generation (ENLG). Nancy, France. Garoufi, K. and Koller, A. (). ‘Generation of effective referring expressions in situated context’, Language, Cognition and Neuroscience (): –. Garoufi, K., Staudte, M., Koller, A., and Crocker, M. (). ‘Exploiting listener gaze to improve situated communication in dynamic virtual environments’, Cognitive Science (): –. Garrod, S. C. and Sanford, A. J. (). ‘The mental representation of discourse in focussed memory system: Implications for the interpretation of anaphoric noun phrases’, Journal of Semantics : –. Garrod, S. and Sanford, A. J. (). ‘Thematic subjecthood and cognitive constraints on discourse structure’, Journal of Pragmatics : –. Garrod, S. and Terras, M. (). ‘The contribution of lexical and situational knowledge to resolving discourse roles: bonding and resolution’, Journal of Memory and Language : –. Garvey, C. and Caramazza, A. (). ‘Implicit causality in verbs’, Linguistic Inquiry (): –. Gatt, A. (). Generating Coherent References to Multiple Entities. University of Aberdeen: Unpublished PhD thesis. Gatt, A. and Belz, A. (). ‘Introducing shared task evaluation to NLG: The TUNA shared task evaluation challenges’, in E. Krahmer and M. Theune (eds), Empirical Methods in Natural Language Generation, Berlin: Springer Verlag, –. Gatt, A. and Krahmer, E. (). ‘Survey of the state of the art in natural language generation: Core tasks, applications and evaluation’, in press. Gatt, A. and van Deemter, K. (). ‘Lexical choice and conceptual perspective in the generation of plural referring expressions’, Journal of Logic, Language, and Information : –. Gatt, A., Belz, A., and Kow, E. (). ‘The TUNA challenge : Overview and evaluation results’, in Proceedings of the th International Conference on Natural Language Generation (INLG). Salt Fork, Ohio. Gatt, A., Belz, A., and Kow, E. (). ‘The TUNA-REG challenge : Overview and evaluation results’, in Proceedings of the th European Workshop on Natural Language Generation (ENLG), –, Athens, Greece. Gatt, A., Krahmer, E., and Goudbeek, M. (). ‘Attribute preference and priming in reference production: Experimental evidence and computational modeling’, in Proceedings of the rd Annual Meeting of the Cognitive Science Society (COGSCI), –, Boston, Massachusetts. Gatt, A., van der Sluis, I., and van Deemter, K. (). ‘Evaluating algorithms for the generation of referring expressions using a balanced corpus’, in Proceedings of the th European Workshop on Natural Language Generation (ENLG), –, Schloss Dagstuhl, Germany. Gazdar, Gerald (). Pragmatics: Implicature, Presupposition, and Logical Form. New York: Academic Press. Geach, P. and Black, M. (eds) (). Translations from the Philosophical Writings of Gottlob Frege, Oxford: Basil Blackwell. Geach, Peter (). Reference and Generality. Cornell University Press, Ithaca, New York.

OUP CORRECTED PROOF – FINAL, 21/1/2019, SPi

REFERENCES



Geach, P. T. (). ‘Intentional identity’, Journal of Philosophy : –. Gelormini-Lezama, C. and A. Almor (). ‘Repeated ames, overt pronouns, and null pronouns in Spanish’, Language and Cognitive Processes (): –. Geng, L. and Hamilton, H. (). ‘Interestingness measures for data mining: a survey’, ACM Computing Surveys (): –. Gernsbacher, M. A. (). ‘Mechanisms that improve referential access’, Cognition : –. Gernsbacher, Morton Ann and Hargreaves, David J. (). ‘Accessing sentence participants: the advantage of first mention’, Journal of Memory and Language (): –. Gernsbacher, Morton Ann and Shroyer, Suzanne (). ‘The cataphoric use of the indefinite this in spoken narratives’, Memory & Cognition (): –. Gibson, E. (). ‘Linguistic complexity: locality of syntactic dependencies’, Cognition (): –. Gibson, E. (). ‘Dependency locality theory: a distance-based theory of linguistic complexity’, in A. Marantz, Y. Miyashita, and W. O’Neil (eds), Image, Language, Brain. Boston: MIT Press. Gillon, Carrie (). The Semantics of Determiners: Domain Restriction in Skwxwumesh. PhD Dissertation. University of British Columbia. Girbau, Dolors (). ‘Children’s referential communication failure: the ambiguity and abbreviation of message’, Journal of Language and Social Psychology : –. Girouard, P. C., Ricard, M., and Gouin Decarie, T. (). ‘The acquisition of personal pronouns in French-speaking and English-speaking children’, Journal of Child Language : –. Giuliani, M., Foster, M. E., Isard, A., Matheson, C., Oberlander, J., and Knoll, A. (). ‘Situated reference in a hybrid human–robot interaction system’, in Proceedings of the th International Natural Language Generation Conference (INLG), –. Givón, Talmy (). ‘Topic, pronoun, and grammatical agreement’, in C. N. Li (ed.), Subject and Topic, London: Academic Press, –. Givón, Talmy (ed.) (). Topic Continuity in Discourse: A Quantitative Cross-language Study. Amsterdam: John Benjamins. Givón, Talmy (). ‘The pragmatics of word order: predictability, importance, and attention’, in M. Hammond, E. A. Moravcsik, and J. R. Wirth (eds), Studies in Syntactic Typology (Vol. ), Amsterdam/Philadelphia: John Benjamins. Givón, Talmy (). Syntax: A functional-typological introduction, vol. . Amsterdam: John Benjamins. Givón, Talmy (). ‘Coherence in text vs. coherence in mind’, in M. A. Gernsbacher and T. Givón (eds), Coherence in Spontaneous Text, Amsterdam: John Benjamins, –. Glanzberg, Michael (). ‘Context, content and relativism’, Philosophical Studies : –. Gleason, J. B., Phillips, B. C., Ely, R., and Zaretsky, E. (). ‘Alligators all around: The acquisition of animal terms in English and Russian’, in J. Guo, E. Lieven, N. Budwig, S. Ervin-Tripp, Keiko Nakamura, and S. Ozcaliskan (eds), Crosslinguistic Approaches to the Psychology of Language: Research in the Tradition of Dan Isaac Slobin, New York: Psychology Press, –. Gleitman, Lila R., January, David, Nappa, Rebecca, and Trueswell, John C. (). ‘On the give and take between event apprehension and utterance formulation’, Journal of Memory and Language (): –. doi: ./j.jml.... Glucksberg, Samuel, Robert Krauss, and Robert Weisberg (). ‘Referential communication in nursery school children: method and some preliminary findings’, Journal of Experimental Child Psychology : –.

OUP CORRECTED PROOF – FINAL, 21/1/2019, SPi



REFERENCES

Goldberg, E., Driedger, N., and Kittredge, R. (). ‘Using natural language processing to produce weather forecasts’, IEEE Expert (): –. Goldin-Meadow, Susan (). ‘The role of gesture in communication and thinking’, Trends in Cognitive Sciences : –. Goodwin, Charles (). ‘Environmentally coupled gestures’, in S. Duncan, J. Cassell, and E. Levy, (eds), Gesture and the Dynamic Dimensions of Language. Amsterdam/Philadelphia: John Benjamins, –. Gordon, P., Hendrick, R., and Johnson, M. (). ‘Memory interference during language processing’, Journal of Experimental Psychology: Learning, Memory, and Cognition (): –. Gordon, Peter C., Grosz, Barbara J., and Gilliom, Laura A. (). ‘Pronouns, names, and the centering of attention in discourse’, Cognitive Science : –. doi: ./ scog_. Görgülü, Emrah (). ‘On definiteness and specificity in Turkish’, in Shibagaki, R. and R. Vermeulen (eds), MIT Working Papers in Linguistics . Proceedings of the th Workshop on Formal Altaic Linguistics. Cambridge, MA: MIT. Gorniak, P. and Roy, D. (). ‘Grounded semantic composition for visual scenes’, Journal of Artificial Intelligence Research : –. Goudbeek, M. and Krahmer, E. (). ‘Preferences versus adaptation during referring expression generation’, in Proceedings of the th Annual Meeting of the Association for Computational Linguistics (ACL), –, Uppsala, Sweden. Graff, D. (). ‘Descriptions as predicates’, Philosophical Studies : –. Graff, Gordon (). ‘Better bubbles’, Popular Science, (), . Grafton, Sue (). U is for Undertow: A Kinsey Millhone Mystery. New York: Berkley. Gratch, J., Rickel, J., André, E., Badler, N., Cassell, J., and Petajan, E. (). ‘Creating interactive virtual humans: some assembly required’, IEEE Intelligent Systems : –. Green, Georgia M. (). Pragmatics and Natural Language Understanding. Hillsdale, NJ: Lawrence Erlbaum. Greene, S. B., McKoon, G., and Ratcliff, R. (). ‘Pronoun resolution and discourse models’, Journal of Experimental Psychology: Learning, Memory, and Cognition : –. doi:./ -.... Greene, Steven B., Gerrig, Richard J., McKoon, Gail, and Ratcliff, Roger (). ‘Unheralded pronouns and management by common ground’, Journal of Memory and Language (): –. doi: ./jmla... Greenfield, P. M. (). ‘Informativeness, presupposition, and semantic choice in singleword utterances’, in E. Ochs and B. Schieffelin (eds), Developmental Pragmatics, New York: Academic Press, –. Greenfield, P. M. and Smith, J. H. (). The Structure of Communication in Early Language Development. New York: Academic Press. Grice, H. Paul (). ‘Meaning’, The Philosophical Review (): –. Grice, H. P. (). ‘Logic and Conversation, William James Lectures, Harvard’, Published in Donald Davidson and Gilbert Harman (eds), , The Logic of Grammar. Reprinted in Grice (), Dickenson Publishing Company. Grice, H. Paul (). ‘Utterer’s meaning and intention’, The Philosophical Review (): –. Grice, H. P. (). ‘Logic and conversation’, Syntax and semantics : –. Grice, H. P. (). ‘Logic and conversation’, in Peter Cole and Jerry Morgan (eds), Speech acts, New York: Academic Press, –.

OUP CORRECTED PROOF – FINAL, 21/1/2019, SPi

REFERENCES



Grice, H. Paul (). Studies in the Way of Words. Cambridge, MA: Harvard University Press. Griffin, Zenzi M. and Bock, J. Kathryn (). ‘What the eyes say about speaking’, Psychological Science (): –. doi: ./-.. Groenendijk, Jeroen and Martin Stokhof (). ‘Dynamic Montague Grammar’, in L. Kálman and L. Pólos (eds), Papers from the Second Symposium on Logic and Language, Budapest: Adakémiai Kiadó, –. Groenendijk, Jeroen and Martin Stokhof (). ‘Dynamic interpretation’, in The Encyclopaedia of Language and Linguistics, London: Pergamon Press. Grosz, Barbara J. (). ‘The representation and use of focus in a system for understanding dialogs’, Proceedings of Fifth International Joint Conference on Artificial Intelligence (IJCAI’), –. Grosz, Barbara J. (). ‘Focusing and description in natural language dialogues’, in Joshi, A., Webber, B. L., and Sag, I. A. (eds), Elements of Discourse Understanding, Cambridge, Cambridge University Press: London, –. Grosz, Barbara J. and Candace L. Sidner (). ‘Attention, intentions, and the structure of discourse’, Computational Linguistics (): –. Grosz, Barbara J. and Candace L. Sidner (). ‘Plans for discourse’, in P. R. Cohen, J. Morgan, and M. E. Pollack (eds), Intentions in Communication, Cambridge, MA.: MIT Press, –. Grosz, Barbara, Aravind K. Joshi, and Scott Weinstein (). ‘Providing definite noun phrases in discourse’, Proceedings of the st Annual Meeting of the Association for Computational Linguistics : –. Grosz, B. J., Joshi, A. K., and Weinstein, S. (). ‘Centering: a framework for modeling the local coherence of discourse’, Computational Linguistics : –. Guasti, M. T. (). The Growth of Grammar. Cambridge, MA: MIT Press. Guasti, M. T., Gavarró, A., de Lange, J., and Caprin, C. (). ‘Article omission across child languages’, Language Acquisition (): –. Guerriero, A. M. S., Oshima-Takane, Y., and Kuriyama, Y. (). ‘The development of referential choice in English and Japanese: a discourse-pragmatic perspective’, Journal of Child Language (): –. Guhe, M. and Bard, E. G. (). ‘Adapting referring expressions to the task environment’, in Proceedings of the th Annual Conference of the Cognitive Science Society (COGSCI), –, Austin, TX. Guindon, Raymonde (). ‘Anaphora resolution: Short-term memory and focusing’, Proceedings of the Annual Meeting of the Association for Computational Linguistics : –. Gundel, Jeanette K. (). ‘Stress, pronominalization, and the given–new distinction’, University of Hawaii Working Papers in Linguistics (): –. Gundel, Jeanette K. (). ‘Shared knowledge and topicality’, Journal of Pragmatics : –. Gundel, Jeanette K. (). ‘Universals of topic-comment structure’, in M. Hammond, E. A. Moravcsik, and J. R. Wirth (eds), Studies in Syntactic Typology (Vol. ), Amsterdam/Philadelphia: John Benjamins, –. Gundel, Jeanette (). ‘Topic, focus, and the grammar–pragmatics interface’, in J. Alexander, N. Han, and M. Minnick (eds), Proceedings of the rd Annual Penn Linguistics Colloquium (University of Pennsylvania Working Papers in Linguistics ). Philadelphia, PA: Penn Libraries.

OUP CORRECTED PROOF – FINAL, 21/1/2019, SPi



REFERENCES

Gundel, Jeanette K. (). ‘Children’s use of referring expressions: what can it tell us about Theory of Mind?’, Cognitive Critique : –. Gundel, Jeanette K. (). ‘Reference and accessibility from a Givenness Hierarchy perspective’, International Review of Pragmatics (): –. Gundel, Jeanette K. (). ‘Child language, Theory of Mind, and the role of procedural markers in identifying referents of nominal expressions’, in Escandell-Vidal, V., M. Leonetti and A. Aherm (eds), Procedural Meaning: Problems and Perspectives. Emerald Group Publishing, –. Gundel, Jeanette K. and Fretheim, T. (). ‘Topic and focus’, in G. Ward and L. Horn (eds), Handbook of Pragmatics (Blackwell Handbooks in Linguistics), Oxford: Blackwell, –. Gundel, Jeanette K. and Johnson, K. (). ‘Children’s use of referring expressions in spontaneous discourse: implications for theory of mind development’, Journal of Pragmatics : –. Gundel, Jeanette K., Hedberg, N., and Zacharski, R. (). ‘On the Generation and Interpretation of Demonstrative Expressions.’, paper presented at the Proceedings of the XIIth International Conference on Computational Linguistics, Budapest. Gundel, Jeanette K., Hedberg, N., and Zacharski, R. (). ‘Cognitive status and the form of referring expressions in discourse’, Language : –. Gundel, Jeanette K., Ntelitheos, D., and Kowalsky, M. (). ‘Children’s use of referring expressions: some implications for theory of mind’, ZAS Papers in Linguistics . Gundel, Jeanette K., Bassene, M., Gordon, B., Humnick, L., and Khalfaoui, A. (). ‘Testing predictions of the Givenness Hierarchy framework: a crosslinguistic investigation’, Journal of Pragmatics (): –. Gundel, Jeanette K., Hedberg, N., and Zacharski, R. (). ‘Underspecification of Cognitive Status in Reference Production: The Grammar–Pragmatics Interface’, Pre-Cog Sci  Workshop, Bridging the Gap between Computational, Empirical, and Theoretical Approaches to Reference. Boston, Massachusetts,  July . Gundel, Jeanette K., Hedberg, N., and Zacharski, R. (). ‘Underpecification of cognitive status in reference production: some empirical predictions’, Topics in Cognitive Science (): – [Issue on the Production of referring expressions: Bridging the gap between computational and empirical approaches]. Gundel, Jeanette K., Nancy Hedberg, and Ron Zacharski (). ‘Definite descriptions and cognitive status in English: why accommodation is unnecessary’, English Language and Linguistics : –. Gundel, Jeanette K. and Nancy Hedberg (). ‘Reference and cognitive status: scalar inference and typology’, in M. M. Jocelyne Fernandez-Vest and Robert D. Van Valin (eds), Information Structure in Spoken Language from a Cross-linguistic Perspective, Berlin: Mouton de Gruyter, –. Gundel, Jeanette K., Nancy Hedberg, and Ron Zacharski (). ‘Pronouns without NP antecedents: How do we know when a pronoun is referential?’, in A. Branco, T. McEnery, and R. Mitkov (eds), Anaphora Processing: Linguistic, Cognitive, and Computational Modelling. Amsterdam/Philadelphia: John Benjamins, –. Gundel, Jeanette K., Nancy Hedberg, Ron Zacharski, Ann Mulkern, Tonya Custis, Bonnie Swierzbin, Amel Khalfoui, Linda Humnick, Bryan Gordon, Mamadou Bassene, and Shana Watters (). Coding Protocol for Statuses on the Givenness Hierarchy. http://www.sfu.ca/ ~hedberg/Coding_for_Cognitive_Status.pdf.

OUP CORRECTED PROOF – FINAL, 21/1/2019, SPi

REFERENCES



Gundel, Jeanette K, Nancy Hedberg, and Ron Zacharski (). ‘Givenness, implicatures, and demonstrative expressions in English discourse’, Chicago Linguistic Society /.–. Gundel, Jeanette K, Nancy Hedberg, and Ron Zacharski (). ‘Givenness, implicature, and the form of referring expressions in discourse’, Berkeley Linguistics Society .–. Gupta, S. and Stent, A. (). ‘Automatic evaluation of referring expression generation using corpora’, in Proceedings of the st Workshop on Using Copora in Natural Language Generation (UCNLG), –, Brighton, UK. Haeseryn, W., Romijn, K., Geerts, G., de Rooij, J., and van den Toorn, M. C. (). Algemene Nederlandse Spraakkunst, Band  [General Dutch grammar, Vol. ]. Groningen: Martinus Nijhoff. Hagege, C. (). ‘Les pronoms logophoriques’, Bulletin de la Société de Linguistique de Paris : –. Hagiwara, H., Soshi, T., Isihara, M., and Imanaka, K. (). ‘A topographical study on the event-related potential correlates of scrambled word order in Japanese complex sentences’, Journal of Cognitive Neuroscience (): –. Hagoort, P., Baggio, G., and Willems, R. M. (). ‘Semantic unification’, in M. S. Gazzaniga (ed.), The Cognitive Neurosciences, Boston, MA: MIT Press. th ed, –. Hagoort, P., Brown, C., and Groothusen, J. (). ‘The syntactic positive shift (SPS) as an ERP measure of syntactic processing’, Language and Cognitive Processes (): –. Hagoort, P., Hald, L., Bastiaansen, M., and Petersson, K. M. (). ‘Integration of word meaning and world knowledge in language comprehension’, Science : –. Hajičová, Eva (). ‘Focussing—a meeting point of linguistics and artificial intelligence’, in Ph. Jorrand and V. Sgurev (eds), Artificial Intelligence II: Methodology, systems, applications, Amsterdam: North-Holland, –. Hajičová, Eva (). Issues of Sentence Structure and Discourse Patterns - Theoretical and Computational Linguistics, Vol. . Prague: Charles University. Halliday, M. A. K. (). ‘Notes on transitivity and theme in English II’, Journal of Linguistics : –. Halliday, M. A. K. and Ruqaiya Hasan (). Cohesion in English. London: Longman. Hamann, C., Rizzi, L., and Frauenfelder, U. H. (). ‘On the acquisition of subjects and object clitics in French’, in H. Clahsen (ed.), Generative Perspectives on Language Acquisition, Amsterdam/Philadelphia: John Benjamins, –. Hamzelou, Jessica (). ‘Brains wired up to work together’, New Scientist, : –. Han, C. (). ‘On the syntax of relative clauses in Korean’, Canadian Journal of Linguistics (): –. Hankamer, J. and Sag, T. (). ‘Deep and surface anaphors’, Linguistic Inquiry : –. Hanks, Peter (). Propositional Content. Oxford: Oxford University Press. Hanna, Joy E. and Susan E. Brennan (). ‘Speakers’ eye gaze disambiguates referring expressions early during face-to-face conversation’, Journal of Memory and Language : –. Hanna, J. E. and Tanenhaus, M. K. (). ‘Pragmatic effects on reference resolution in a collaborative task: evidence from eye movements’, Cognitive Science : –. Hanna, Joy, Michael Tanenhaus, and John Trueswell (). ‘The effects of common ground and perspective on domains of referential interpretation’, Journal of Memory and Language : –.

OUP CORRECTED PROOF – FINAL, 21/1/2019, SPi



REFERENCES

Harnad, S. (). ‘The symbol grounding problem’, Physica D: Nonlinear Phenomena (–): –. Harris, A. (). Electrophysiological Indices of Syntactic Processing Difficulty. PhD Thesis: MIT. Hartshorne, J. K. and Snedeker, J. (). ‘Verb argument structure predicts implicit causality: the advantages of finer-grained semantics’, Language and Cognitive Processes (): –. Hartshorne, Joshua K., Rebecca Nappa, and Jesse Snedeker (). ‘Development of the firstmention bias’, Journal of Child Language : –. Haspelmath, Martin (). Indefinite Pronouns. Oxford: Oxford University Press. Haugh, Michael (). ‘On understandings of intention: a response to Wedgewood’, Intercultural Pragmatics : –. Hawkins, John A. (). Definiteness and Indefiniteness. Atlantic Highland, NJ: Humanities Press. Hawkins, John A. (). ‘A note on referent identifiability and copresence’, Journal of Pragmatics : –. Hawkins, John A. (). ‘On (in)definite articles: implicatures and (un)grammaticality prediction’, Journal of Linguistics : –. Hawthorne, John and Manley, David (). The Reference Book. Oxford: Oxford University Press. Hedberg, Nancy Ann (). Discourse Pragmatics and Cleft Sentences in English. University of Minnesota dissertation. Hedberg, Nancy Ann, Emrah Görgülü, and Morgan Mameni (). ‘On definiteness and specificity in Turkish and Persian’, Proceedings of the  Annual Conference of the Canadian Linguistic Association. http://homes.chass.utoronto.ca/~cdla-acl/actes/actes.html. Heeman, P. A. and Hirst, G. (). ‘Collaborating on referring expressions’, Computational Linguistics (): –. Heim, Irene (). The Semantics of Definite and Indefinite Noun Phrases. Doctoral Dissertation, University of Massachusetts, Amherst, MA. Heim, Irene (). ‘File change semantics and the familiarity theory of definiteness’, in Rainer Bauerle, Christoph Schwarze, and Arnim von Stechow (eds), Meaning, Use and the Interpretation of Language, Berlin: Walter de Gruyter, –. Heim, Irene (). ‘E-Type pronouns and donkey anaphora’, Linguistics and Philosophy : –. Heim, Irene (). ‘Artikel und Definitheit’, in A. von Stechow and D. Wunderlich (eds), Semantik: Ein Internationales Handbuch der Zeitgenössischen Forschung, Bd. . Berlin: Mouton de Gruyter, –. Heim, Irene (). ‘Presupposition projection and the semantics of attitude verbs’, Journal of Semantics : –. Heim, Irene (). ‘Definiteness and indefiniteness’, in K. von Heusinger, C. Maienborn, and P. Portner (eds), Semantics. An International Handbook of Natural Language Meaning, vol. . Berlin: Mouton de Gruyter, –. Heim, Irene and Kratzer, A. (). Semantics in Generative Grammar. Oxford: Blackwell. Heine, A., Tamm, S., Hofmann, M., Hutzler, F., and Jacobs, A. M. (). ‘Does the frequency of the antecedent noun affect the resolution of pronominal anaphors? An ERP study’, Neuroscience Letters : –. Heller, Daphna and Lynsey Wolter (). ‘That is Rosa: identificational sentences as intensional predication’, Proceedings of Sinn und Bedeutung : –.

OUP CORRECTED PROOF – FINAL, 21/1/2019, SPi

REFERENCES



Heller, Daphna, Skovbroten, K., and Tanenhaus, M. (). ‘To name or to describe: shared knowledge affects referential form’, Topics in Cognitive Science (): –. Heller, Daphna, Daniel Grodner, and Michael Tanenhaus (). ‘The role of perspective in identifying domains of reference’, Cognition : –. Helm, Irene R. (). The semantics of definite and indefinite noun phrases. Amherst: University of Massachusetts dissertation. Henderson, John M. and Ferreira, Fernanda (eds) (). The Interface of Language, Vision, and Action. Eye Movements and the Visual World. New York/Hove: Psychology Press. Hendrickx, I., Daelemans, W., Luyckx, K., Morante, R., and Asch, V. V. (). ‘CNTS: Memory-based learning of generating repeated references’, in Proceedings of the th International Natural Language Generation Conference (INLG), –. Hendriks, P., Koster C., and Hoeks, J. (). ‘Referential choice across the lifespan: why children and elderly adults produce ambiguous pronouns’, Language, Cognition, and Neuroscience (): –. Henschel, R., Cheng, H., and Poesio, M. (). ‘Pronominalisation revisited’, in Proceedings of the th International Conference on Computational Linguistics (COLING), –, Saarbrücken, Germany. Heurdier, J., da Silva, C., Le Mené, M., and Salazar Orvig, A. (). ‘Premiers usages de «c’est»: formes et valeurs sémantico-discursives’, paper presented at the AFLS  Conference: “Le français à travers le temps: acquisition, changement, variation”, Newcastle. Heycock, Caroline and Anthony Kroch (). ‘Inversion and equation in copular sentences’, ZAS Papers in Linguistics, Volume . Berlin: Zentrum für Allgemeine Sprachwissenschaft, –. Hickmann, M. (). ‘The pragmatics of reference in child language: some issues in developmental theory’, in M. Hickmann (ed.), Social and Functional Approaches to Language and Thought, Orlando: Academic Press, –. Hickmann, M. (). ‘The development of discourse cohesion: some functional and crosslinguistic issues’, in G. Piérault-Le Bonniec and M. Dolitsky (eds), Language Bases . . . Discourse Bases, Amsterdam: Benjamins. Hickmann, M. (). Children’s Discourse: Person, Time, and Space Across Languages. Cambridge: Cambridge University Press. Hickmann, M., Hendriks, H. T., Roland, F. O., and Liang, J. (). ‘The marking of new information in children’s narratives: A comparison of English, French, German, and Mandarin Chinese’, Journal of Child Language (): –. Hickmann, M., Schimke, S., and Colonna, S. (). ‘From early to late mastery of reference: multifunctionality and linguistic diversity’, in L. Serratrice and S. E. M. Allen (eds), The Acquisition of Reference, Amsterdam: John Benjamins, –. Higginbotham, James (). ‘Indefinites and predication’, in E. Reuland and A. ter Meulen (eds), The Representation of (In)definiteness, Cambridge, MA: MIT Press, –. Higgins, Roger F. (). The Pseudo-cleft Construction in English. New York: Garland. Hinds, John (). ‘Anaphoric demonstratives in Japanese’, Journal of the Association of Teachers of Japanese : –. Hinterwimmer, S. and Bosch, P. (). ‘Demonstrative pronouns and perspective’ [with Peter Bosch], in Patel, P. and P. Patel-Grosz (eds), The Impact of Pronominal Form on Interpretation, De Gruyter (Studies in Generative Grammar): Berlin/New York, –. Hintikka, Jaakko (). ‘The semantics of a certain’, Linguistic Inquiry (): –. Hirschberg, Julia (). A Theory of Scalar Implicature, Philadelphia: University Pennsylvania dissertation.

OUP CORRECTED PROOF – FINAL, 21/1/2019, SPi



REFERENCES

Hobbs, J. (). ‘Coherence and coreference’, Cognitive Science : –. Hobbs, J. (). Literature and Cognition [Lecture notes ]. Stanford, CA: CSLI. Hoetjes, Marieke, Koolen, Ruud, Goudbeek, Martijn, Krahmer, Emiel, and Swerts, Marc (). ‘Reduction in gesture during the production of repeated references’, Journal of Memory and Language –: –. doi: ./j.jml.... Hofmeister, P. (). ‘Representational complexity and memory retrieval in language comprehension’, Language and Cognitive Processes : –. Hopcroft, J. (). ‘An n log(n) algorithm for minimizing states in a finite automaton’, in Z. Kohave (ed.), Theory of Machines and Computations, Academic Press. Horacek, H. (). ‘A new algorithm for generating referring expressions’, in Proceedings of the th European Conference on Artificial Intelligence (ECAI), –, Budapest, Hungary. Horacek, H. (). ‘An algorithm for generating referential descriptions with exible interfaces’, in Proceedings of the th Annual Meeting of the Association for Computational Linguistics (ACL), –, Madrid. Horacek, H. (). ‘On referring to sets of objects naturally’, in Proceedings of the rd International Conference on Natural Language Generation (INLG), –, Brockenhurst, UK. Horacek, H. (). ‘Generating referential descriptions under conditions of uncertainty’, in Proceedings of the th European Workshop on Natural Language Generation (ENLG), –, Aberdeen, UK. Horn, Laurence R. (). On the semantic properties of logical operators in English. Los Angeles: UCLA dissertation. Horn, Laurence R. (). ‘Toward a new taxonomy for pragmatic inference: Q-based and R-based implicature in context’, in Deborah Schiffrin (ed.), Meaning, form, and use in context, Washington, DC: Georgetown University Press, –. Horn, Laurence R. (). ‘Metalinguistic Negation and Pragmatic Ambiguity’, Language, , –. Horn, Laurence R. (). A Natural History of Negation. Chicago, University of Chicago Press. Horn, Laurence R. (). ‘Toward a Fregean pragmatics: Voraussetzung, Nebengedanke, Andeutung’, in Istvan Kecskes and Laurence R. Horn (eds), Explorations in Pragmatics: Linguistic, Cognitive, and Intercultural Aspects, New York: Mouton de Gruyter, –. Horn, Laurence R. and Barbara Abbott (). ‘: (In)definiteness and implicature’, to appear in Joseph Keim Campbell, William Kabasenche, and Michael O’Rourke (eds), Reference and Referring, Topics in Contemporary Philosophy, vol. . Cambridge, MA: MIT Press. Horn, Laurence R. and Barbara Abbott (). ‘: (In)definiteness and implicature’, in William P. Kabasenche, Michael O’Rourke, and Matthew H. Slater (eds), Reference and Referring (Topics in Contemporary Philosophy, vol. ), Cambridge, MA: MIT Press, –. Hornby, P. A. (). ‘Surface structure and presupposition’, Journal of Verbal Learning and Verbal Behavior : –. Horton, W. and Keysar, B. (). ‘When do speakers take into account common ground?’, Cognition : –. Hovy, E. H. (). ‘Parsimonious and profligate approaches to the question of discourse structure relations’, Proceedings of th International Workshop on Language Generation. Pittsburgh, PA. Huang, C.-C. (). ‘Referential choice in Mandarin child language: A discourse-pragmatic perspective’, Journal of Pragmatics (): –.

OUP CORRECTED PROOF – FINAL, 21/1/2019, SPi

REFERENCES



Huang, Y. (). Anaphora: A Cross-Linguistic Study. Oxford; New York, NY: Oxford University Press. Hughes, M. E. and Allen, S. E. (). ‘The effect of individual discourse-pragmatic features on referential choice in child English’, Journal of Pragmatics : –. Hunt, L. III, Politzer-Ahles, S., Gibson, L., Minai, U., and Fiorentino, R. (). ‘Pragmatic inferences modulate N during sentence comprehension: evidence from picturesentence verification’, Neurosci Lett. (): –. Hutchins, Edwin (). Cognition in the Wild. Cambridge, MA: MIT Press. Hyams, N. M. (). Language Acquisition and the Theory of Parameters. D. Reidel Publishing Company Dordrecht. Ionin, Tania (). ‘This is definitely specific: specificity and definiteness in article systems’, Natural Language Semantics : –. Ionin, Tania, Ora Matushansky, and E. G. Ruys (). ‘Parts of speech: toward a unified semantics for partitives’, in Chris Davis, Amy Rose Deal, and Youri Zabbal (eds), Proceedings of NELS , Amherst, MA: University of Massachusetts GLSA, –. Ioup, G. (). ‘Some universals for quantifier scope’, Syntax and Semantics : –. Isard, Stephen (). ‘Changing the context’, in Edward L. Keenan (ed.), Formal semantics of natural language, Cambridge: Cambridge University Press, –. Itti, L. and Koch, C. (). ‘A saliency-based search mechanism for overt and covert shifts in visual attention’, Vision Research : –. Jaccard, P. (). ‘Étude comparative de la distribution orale dans une portion des alpes et des jura’, Bulletin de la Société Vaudoise des Sciences Naturelles : –. Jackendoff, Ray (). Semantic Interpretation in Generative Grammar. Cambridge: MIT Press. Jackendoff, Ray (). X-bar Syntax: A Study of Phrase Structure. Cambridge, MA: MIT Press. Jackendoff, Ray (). Foundations of Language: Brain, Meaning, Grammar, Evolution. Oxford/New York: Oxford University Press. Jaeger, T. Florian (). ‘Redundancy and reduction: speakers manage syntactic information density’, Cognitive Psychology (): –. doi: ./j.cogpsych.... Jaeger, F. and Norcliffe, E. (). ‘The cross-linguistic study of sentence production’, Language and Linguistics Compass : –. Jakubowicz, C. and Rigaut, C. (). ‘L’acquisition des clitiques nominatifs en français’, in A. Zribi-Hertz (ed.), Les Pronoms. Morphologie, Syntaxe, et Typologie, Saint-Denis: Presses Universitaires de Vincennes, –. Janarthanam, S. and Lemon, O. (). ‘Learning lexical alignment policies for generating referring expressions for spoken dialogue systems’, in Proceedings of the th European Workshop on Natural Language Generation (ENLG), –, Athens, Greece. Jarvella, Robert J. and Klein, Wolfgang (eds) (). Speech, Place, and Action: Studies of Deixis and Related Topics. New York: Wiley. Järvikivi, J., van Gompel, R. P. G., Bertram, R., and Hyönä, J. (). ‘Ambiguous pronoun resolution: contrasting the first-mention and subject preference accounts’, Psychological Science : –. Jennings, N. R. (). ‘On agent-based software engineering’, Artificial Intelligence (): –. Jisa, H. (). ‘Increasing cohesion in narratives: a developmental study of maintaining and reintroducing subjects in French’, Linguistics (): –.

OUP CORRECTED PROOF – FINAL, 21/1/2019, SPi



REFERENCES

Johnson-Laird, P. N. (). Mental Models: Towards a Cognitive Science of Language, Inference, and Consciousness, Cognitive Science Series. Cambridge, MA: Harvard University Press. Jordan, P. W. (). Intentional Inuences on Object Redescriptions in Dialogue: Evidence from an Empirical Study (Unpublished doctoral dissertation). University of Pittsburgh. Jordan, P. W. (). ‘Contextual inuences on attribute selection for repeated descriptions’, in K. van Deemter and R. Kibble (eds), Information Sharing: Reference and Presupposition in Language Generation and Interpretation, Stanford, CA: CSLI Publications. Jordan, P. W. and Walker, M. (). ‘Learning content selection rules for generating object descriptions in dialogue’, Journal of Artificial Intelligence Research : –. Joshi, Aravind K. (). ‘Mutual beliefs in question–answer systems’, in N. Smith (ed.), Mutual Knowledge, New York: Academic Press, –. Joshi, Aravind K., Bonnie L. Webber; and Ivan A. Sag (eds) (). Elements of Discourse Understanding. Cambridge: Cambridge University Press. Just, M., and Carpenter, P. (). ‘A capacity theory of comprehension: Individual differences in working memory’, Psychological Review, , –. Just, M., Carpenter, P., Keller, T., Eddy, W., and Thulborn, K. (). ‘Brain activation modulated by sentence comprehension’, Science : –. Kaan, E., Harris, T., Gibson, E., and Holcomb, P. (). ‘The P as an index of syntactic integration difficulty’, Language and Cognitive Processes (): –. Kahneman, D. (). Attention and Effort. Englewood Cliffs, NJ: Prentice-Hall. Kail, M. and Hickmann, M. (). ‘French children’s ability to introduce referents in narratives as a function of mutual knowledge’, First Language (): –. Kail, M. I. and Lopez, I. S. Y. (). ‘Referent introductions in Spanish narratives as a function of contextual constraints: a crosslinguistic perspective’, First Language (): –. Kaiser, Elsi (). The Quest for a Referent: A Crosslinguistic Look at Reference Resolution. PhD dissertation, University of Pennsylvania. Kaiser, Elsi (). ‘Effects of anaphoric dependencies and semantic representations on pronoun interpretation’, in S. L. Devi, A. Branco, and R. Mitkov (eds), Anaphora Processing and Applications (Selected papers from the th Discourse Anaphora and Anaphor Resolution Colloquium), Heidelberg: Springer, –. Kaiser, Elsi (a). ‘Investigating the consequences of focus on the production and comprehension of referring expressions’, International Review of Pragmatics (): –. Kaiser, Elsi (b). ‘Effects of contrast on referential form: investigating the distinction between strong and weak pronouns’, Discourse Processes (): –. doi: ./ . Kaiser, Elsi (a). ‘Focusing on pronouns: consequences of subjecthood, pronominalisation, and contrastive focus’, Language and Cognitive Processes (): –. Kaiser, Elsi (b). ‘Salience and contrast effects in reference resolution: the interpretation of Dutch pronouns and demonstratives’, Language and Cognitive Processes (): –. Kaiser, Elsi (c). ‘On the relation between coherence relations and anaphoric demonstratives in German’, in Ingo Reich et al. (eds), Proceedings of Sinn & Bedeutung , Saarland University Press: Saarbrücken, Germany, –. Kaiser, Elsi (). ‘Taking action: a cross-modal investigation of discourse-level representations’, Frontiers in Psychology : . Kaiser, Elsi (). ‘Perspective-shifting and free indirect discourse: experimental investigations’, in Sarah D’Antonio, Mary Moroney, and Carol Rose Little (eds), Proceedings of Semantics and Linguistic Theory  (SALT ), Stanford University, –.

OUP CORRECTED PROOF – FINAL, 21/1/2019, SPi

REFERENCES



Kaiser, E. (). Pronoun use in Finnish reported speech and free indirect discourse: Effects of logophoricity. In Pritty Patel-Grosz, Patrick Grosz and Sarah Zobel (eds), Pronouns in Embedded Contexts, –. Springer Studies in Linguistics and Philosophy. Kaiser, Elsi and Do, M. (). ‘Taking a look beneath the surface: effects of processing depth on pronoun interpretation’, in Iris Hendrickx, Antonio Branco, Sobha Lalitha Devi, and Ruslan Mitkov (eds), Proceedings of the th Discourse Anaphora and Anaphor Resolution Colloquium (DAARC), Edicoes Colibri: Lisbon, Portugal, –. Kaiser, Elsi and Trueswell, John C. (). ‘The referential properties of Dutch pronouns and demonstratives: is salience enough?’, in Cécile Meier and Matthias Weisgerber (eds), Proceedings of the Conference ‘Sub - Sinn und Bedeutung’, Arbeitspapier Nr. , –. Kaiser, Elsi and Trueswell, J. (). ‘Interpreting pronouns and demonstratives in Finnish: evidence for a form-specific approach to reference resolution’, Language and Cognitive Processes : –. Kaiser, Elsi, Runner, J., Sussman, R., and Tanenhaus, M. (). ‘Structural and semantic constraints on the resolution or pronouns and reflexives’, Cognition : –. Kaiser, Elsi, Li, David Cheng-huan, and Holsinger, Edward (). ‘Exploring the lexical and acoustic consequences of referential predictability’, in I. Hendricks, A. Branco, S. L. Devi, and R. Mitkov (eds), Anaphora Processing and Applications, th Discourse Anaphora and Anaphor Resolution Colloquium, DAARC, –. Kameyama, Megumi (). Zero anaphora: The case of Japanese. Stanford, CA: Stanford University dissertation. Kameyama, Megumi (). ‘Stressed and unstressed pronouns: complementary preferences’, in P. Bosch and R. van der Sandt (eds), Focus. Linguistic, Cognitive and Computational Perspectives, Cambridge: Cambridge University Press, –. Kamp, Hans (). ‘To the memory of Arthur Prior formal properties of now’, Theoria (): –. URL: http://onlinelibrary.wiley.com/doi/./j.-..tb.x/ abstract. Kamp, Hans (). ‘A theory of truth and semantic representation’, in J. Groenendijk, T. Janssen, and Martin Stokhof (eds), Formal Methods in the Study of Language, Proceedings of the Third Amsterdam Colloquium. Amsterdam: Mathematical Center, –. [Reprinted in: J. Groenendijk, T. Janssen, and Martin Stokhof (eds), Truth, Interpretation and Information, Selected Papers from the Third Amsterdam Colloquium. Dordrecht: Foris, , –; and reprinted in: K. von Heusinger, and A. ter Meulen (eds), The Dynamics of Meaning and Interpretation. Selected Papers of Hans Kamp. Leiden: Brill, , –]. Kamp, Hans (). ‘Prolegomena to a structural theory of belief and other attitudes’, in C. A. Anderson and J. Owens (eds), Propositional Attitudes. Stanford, CA: CSLI Publications, –. [Reprinted in: K. von Heusinger and A. ter Meulen (eds), The Dynamics of Meaning and Interpretation. Selected Papers of Hans Kamp. Leiden: Brill, , –]. Kamp, Hans (). ‘Dividing the province of indefinite noun phrase uses into three parts’, unpublished manuscript. Universität Stuttgart/University of Texas at Austin. Kamp, Hans (). ‘Using proper names as intermediaries between labelled entity representations’, Erkenntnis : –. Kamp, Hans and Bende-Farkas, Agnes (). Specific Indefinites: Anchors and Functional Readings. Stuttgart: University of Stuttgart. Kamp, Hans and Reyle, U. (). From Discourse to Logic. Dordrecht: Kluwer. Kamp, Hans, van Genabith, J., and Reyle, U. (). ‘Discourse Representation Theory’, Handbook of Philosophical Logic Volume : –.

OUP CORRECTED PROOF – FINAL, 21/1/2019, SPi



REFERENCES

Kantola, Leila and van Gompel, Roger P. G. (). ‘Does the addressee matter when choosing referring expressions?’, in Kees van Deemter, Albert Gatt, Roger P. G. van Gompel, and Emiel Krahmer (eds), Proceedings of PRE-Cogsci: Bridging the Gap Between Computational, Empirical and Theoretical Approaches to Reference. Kaplan, David (). ‘Demonstratives’, draft #. Unpublished manuscript, UCLA Philosophy Department. [Revised and published as ‘Demonstratives: an essay on the semantics, logic, metaphysics, and epistemology of demonstratives and other indexicals,’ in Joseph Almog, John Perry, and Howard Wettstein (eds) Themes from Kaplan, Oxford University Press, , –]. Kaplan, David (). ‘Dthat’, in P. Cole (ed.), Syntax and Semantics, vol. . New York: Academic Press, –. Kaplan, David (a). ‘Demonstratives’, in J. Almog, J. Perry, and H. Wettstein (eds), Themes from Kaplan, New York: Oxford University Press, –. Kaplan, David (b). ‘Afterthoughts’, in J. Almog, J. Perry, and H. Wettstein (eds), Themes from Kaplan, New York: Oxford University Press, –. Karimi, Hossein and Ferreira, F. (). ‘Informativity renders a referent more accessible: evidence from eyetracking’, Psychonomic Bulletin & Review (): . Karimi, Hossein, Fukumura, Kumiko, Ferreira, Fernanda, and Pickering, Martin J. (). ‘The effect of noun phrase length on the form of referring expressions’, Memory and Cognition (): –. doi: ./s---. Karimi, Hossein, Swaab, T., and Ferreira, F. (). ‘Electrophysiological evidence for memory retrieval during sentence comprehension’, Annual Meeting of the Society for the Neurobiology of Language, Baltimore, MD. Karmiloff-Smith, Annette (). A Functional Approach to Child Language: A Study of Determiners and Reference. Cambridge: Cambridge University Press. Karmiloff-Smith, Annette (). ‘The grammatical marking of thematic structure in the development of language production’, in W. Deutsch (ed.), The Child’s Construction of Language, London: Academic Press, –. Karmiloff-Smith, Annette (). ‘Language and cognitive processes from a developmental perspective’, Language and Cognitive Processes (): –. Karttunen, Lauri (). ‘What do Referential Indices Refer to?’, paper prepared for the Linguistics Colloquium, University of California, Los Angeles,  April . Karttunen, Lauri (). ‘Discourse referents’, in International Conference on Computational Linguistics COLING , Stockholm: Research Group For Quantitative Linguistics, –. [Reprinted in: James D. McCawley (ed.) (). Syntax and Semantics, Volume : Notes from the Linguistic Underground, New York: Academic Press, –]. Karttunen, Lauri (). ‘Discourse referents’, in James D. McCawley (ed.), Syntax and Semantics, Volume : Notes from the Linguistic Underground, Academic Press, New York, –. Katz, J. (). ‘Has the description theory of names been refuted?’ in Boolos, G. (ed.), Meaning and Method: Essays in Honor of Hilary Putnam, Cambridge: Cambridge University Press. Katz, J. (). ‘Names without bearers’, Philosophical Review : –. Kazanina, N. and Phillips, C. (). ‘Differentials effects on constraints in the processing of Russian cataphora’, The Quarterly Journal of Experimental Psychology : –.

OUP CORRECTED PROOF – FINAL, 21/1/2019, SPi

REFERENCES



Kazanina, N., Lau, E. F., Lieberman, M., Yoshida, M., and Phillips, C. (). ‘The effect of syntactic constraints on the processing of backwards anaphora’, Journal of Memory and Language : –. Kazemzadeh, S., Ordonez, V., Matten, M., and Berg, T. L. (). ‘Referit game: referring to objects in photographs of natural scenes’, in Proceedings of Conference on Empirical Methods in Natural Language Processing (EMNLP-), Doha, Qatar. Keenan, Edward L. (). ‘Names, quantifiers, and a solution to the sloppy identity problem’, Papers in Linguistics /. Keenan, Edward L. and Jonathan Stavi (). ‘A semantic characterization of natural language determiners’, Linguistics and Philosophy : –. Keenan, Edward O. (). ‘Making it last: repetition in children’s discourse’, in S. ErvinTripp and C. Mitchell-Kernan (eds), Child Discourse, New York: Academic Press, –. Keenan, Edward O. and Klein, E. (). ‘Coherency in children’s discourse’, Journal of Psycholinguistic Research (): –. Kehler, A. (). ‘Cognitive status and form of reference in multimodal human–computer interaction’, in Proceedings of the th AAAI Conference on Artificial Intelligence, –. Kehler, A. (). Coherence, Reference, and the Theory of Grammar. Stanford: CSLI Publications. Kehler, A. and Rohde, H. (). ‘A probabilistic reconciliation of coherence-driven and centering-driven theories of pronoun interpretation’, Theoretical Linguistics : –. Kehler, A., Kertz, L., Rohde, H., and Elman, J. (). ‘Coherence and coreference revisited’, Journal of Semantics (Special Issue on Processing Meaning) : –. Kelleher, J. and Kruijff, G.-J. (). ‘Incremental generation of spatial referring expressions in situated dialog’, in Proceedings of the st International Conference on Computational Linguistics (COLING) and th Annual Meeting of the Association for Computational Linguistics (ACL), –, Sydney, Australia. doi: ./.. Kelleher, J., Costello, F., and van Genabith, J. (). ‘Dynamically structuring, updating and interrelating representations of visual and linguistics discourse context’, Artificial Intelligence : –. Kempen, Gerard and Harbusch, Karin (). ‘A corpus study into word order variation in German subordinate clauses: Animacy affects linearization independently of grammatical function assignment’, in T. Pechmann and C. Habel (eds), Multidisciplinary Approaches to Language Production, Berlin: Mouton de Gruyter. Kempen, Gerard and Hoenkamp, E. (). ‘An incremental procedural grammar for sentence formulation’, Cognitive Science : –. doi: ./S-()-X. Kennington, C. and Schlangen, D. (). ‘A simple generative model of incremental reference resolution for situated dialogue’, Computer Speech & Language : –. Kennison, S., Fernandez, E., and Bowers, J.M. (). ‘Processing differences for anaphoric and cataphoric pronouns: Implications for theories of discourse processing’, Discourse Processes, , –. Kerdiles, G. (). Saying it with Pictures: A Logical Landscape of Conceptual Graphs. ILLC, Amsterdam: Unpublished PhD thesis. Kertz, L., Kehler, A., and Elman, J. (). ‘Grammatical and coherence-based factors in pronoun interpretation’, in Proceedings of the th Annual Conference of the Cognitive Science Society, Mahwah, NJ: Lawrence Erlbaum Associates, Inc, –.

OUP CORRECTED PROOF – FINAL, 21/1/2019, SPi



REFERENCES

Keshet, Ezra (a). Good Intensions: Paving Two Roads to a Theory of Good Intensions: Paving Two Roads to a Theory of the De re/De dicto Distinction. PhD thesis, Massachusetts Institute of Technology, Cambridge, MA. Keshet, Ezra (b). ‘Infinitival complements and tense’, Proceedings of SuB, University of Oslo, Norway, –. Keshet, Ezra (c). ‘Only the strong: restricting situation variables’, in T. Friedman and S. Ito (eds), Proceedings of SALT XVIII, Cornell University: Ithaca, NY, –. Keshet, Ezra (). ‘Situation economy’, Natural Language Semantics (): –. URL: http://www.springerlink.com/content/uqm/ Keshet, Ezra (). ‘Split intensionality: a new scope theory of de re and de dicto’, Linguistics and Philosophy (). URL: http://www.springerlink.com/content/d/ Keysar, Boaz (). ‘Egocentric processes in communication and miscommunication’, in I. Kecskes and J. Mey (eds), Intention, Common Ground, and the Egocentric Speaker– Hearer. Berlin: Mouton de Gruyter, –. Keysar, Boaz, Dale Barr, Jennifer Balin, and Jessica Brauner (). ‘Taking perspective in conversation: the role of mutual knowledge in comprehension’, Psychological Science : –. Keysar, Boaz, S. Lin, and D. J. Barr (). ‘Limits on theory of mind use in adults’, Cognition : –. Khan, I. H., Ritchie, G., and van Deemter, K. (). ‘The clarity–brevity trade-off in generating referring expressions’, in Proceedings of the th International Conference on Natural Language Generation (INLG), –, Sydney, Australia. Khan, I. H., van Deemter, K., and Ritchie, G. (). ‘Generation of referring expressions: managing structural ambiguities’, in Proceedings of the th International Conference on Computational Linguistics (COLING), Manchester, UK. Khayrallah, H., Trott, S., and Feldman, J. (). ‘Natural language for human–robot interaction’, in Proceedings of the Workshop on Human-Robot Teaming at the th ACM/IEEE International Conference on Human-Robot Interaction. Kibble, Rodger (). ‘Cb or not Cb? Centering theory applied to NLG’, Proceedings of the ACL workshop on the Relation of Discourse/Dialogue Structure and Reference, –. Kibble, Rodger and Power, R. (). ‘Optimizing referential coherence in text generation’, Computational Linguistics : –. Kibrik, A. (). ‘Anaphora in Russian narrative discourse: a cognitive calculative account’, in B. Fox (ed.), Studies in Anaphora, Amsterdam: Benjamins, –. Kibrik, A. (). Reference in discourse. Oxford: Oxford University Press. Kilani-Schoch, M. and Dressler, W. U. (). ‘Are fillers as precursors of morphemes relevant for morphological theory? A case story from the acquisition of French’, in W. U. Dressler, O. E. Pfeiffer, M. A. Pöchtrager, and J. R. Rennison (eds), Morphological Analysis in Comparison, Amsterdam/Philadelphia: John Benjamins, –. Kilgarriff, A. (). ‘Thesauruses for natural language processing’, in Proceedings of the International Conference on Natural Language Processing and Knowledge Engineering (NLPK), –. Kim, Y.-J. (). ‘Subject/object drop in the acquisition of Korean: a cross-linguistic comparison’, Journal of East Asian Linguistics : –. King, J. W. and Kutas, M. (). ‘Who did what and when? Using word- and clause-level ERPs to monitor working memory usage in reading’, Journal of Cognitive Neuroscience (): –.

OUP CORRECTED PROOF – FINAL, 21/1/2019, SPi

REFERENCES



King, Jeffrey C. (). ‘Are indefinite descriptions ambiguous?’, Philosophcal Studies : –. King, Jeffrey C. (). Complex Demonstratives: A Quantificational Account. Cambridge: MA: MIT Press. King, Jeffrey C. (). ‘Supplementives, the coordination account, and conflicting intentions’, Philosophical Perspectives: Philosophy of Language : –. King, Jeffrey C. (a). ‘The Metasemantics of contextual sensitivity’, in A. Burgess and B. Sherman (eds) New Essays on Metasemantics. Oxford: Oxford University Press, –. King, Jeffrey C. (b). Speaker Intentions in Context. Noûs (): –. Kintsch, Walter and van Dijk, Teun A. (). ‘Toward a model of text comprehension and production’, Psychological Review (): –. doi: ./-X.... Kita, Sotaro (). ‘Cross-cultural variation of speech-accompanying gesture’, Language and Cognitive Processes : –. Kitagawa, Chisato (). ‘A note on sono and ano’ in George Bedell, E. Kobayashi, and M. Muraki (eds), Explorations in linguistics: Papers in honor of Kazuko Inoue, Tokyo: Kenkyusha, –. Kluender, R. and Kutas, M. (). ‘Bridging the gap: Evidence from ERPs on the processing of unbounded dependencies’, Journal of Cognitive Neuroscience (): –. Klyne, G., and Carroll, J. J. (). Resource Description Framework (RDF): Concepts and Abstract Syntax. Kneale, W. (). ‘Modality de dicto and de re’, in P. S. Ernest Nagel and A. Tarski (eds), Studies in Logic and the Foundations of Mathematics, Vol. Volume  of Logic, Methodology and Philosophy of Science Proceeding of the  International Congress, Elsevier, –. URL: http://www.sciencedirect.com/science/article/pii/ SX. Koller, A. and Stone, M. (). ‘Sentence generation as a planning problem’, in Proceedings of the th Annual Meeting of the Association for Computational Linguistics Conference Proceedings (ACL), –, Prague. Koller, A., Striegnitz, K., Byron, D., Cassell, J., Dale, R., Moore, J., and Oberlander, J. (). ‘The first challenge on generating instructions in virtual environments’, in Krahmer, E. and Theune, M. (eds), Empirical Methods in Natural Language Generation, volume , LNAI, Springer. Koolen, Ruud, Gatt, A., Goudbeek, M., and Krahmer, E. (). ‘Need I say more? On factors causing referential overspecification’, in Proceedings of the Workshop on Production of Referring Expressions: Bridging Computational and Psycholinguistic Approaches (PRECOGSCI’). Koolen, Ruud, Gatt, Albert, Goudbeek, Martijn, and Krahmer, Emiel (). ‘Factors causing overspecification in definite descriptions’, Journal of Pragmatics (): –. doi: ./j.pragma.... Koolen, Ruud, Goudbeek, Martijn, and Krahmer, Emiel (). ‘The effect of scene variation on the redundant use of color in definite reference’, Cognitive Science : –. doi: ./cogs.. Koolen, Ruud, Krahmer, Emiel, and Swerts, Marc (). ‘How distractor objects trigger referential overspecification: testing the effects of visual clutter and distractor distance’, Cognitive Science ().

OUP CORRECTED PROOF – FINAL, 21/1/2019, SPi



REFERENCES

Koornneef, A. W. and Van Berkum, J. J. A. (). ‘On the use of verb-based implicit causality in sentence comprehension: evidence from self-paced reading and eye tracking’, Journal of Memory and Language (): –. doi:./j.jml.... Kopp, S., Bergmann, K., and Wachsmuth, I. (). ‘Multimodal communication from multimodal thinking. towards an integrated model of speech and gesture production’, Semantic Computing : –. Korta, Kepa and Perry, John (). Critical Pragmatics: An Inquiry into Reference and Communication. Cambridge: Cambridge University Press. Krahmer, Emiel (). ‘What computational linguists can learn from psychologists (and vice versa)’, Computational Linguistics : –. Krahmer, Emiel and Theune, Mariët (). ‘Efficient context-sensitive generation of referring expressions’, in Kees van Deemter and Rodger Kibble (eds), Information Sharing: Reference and Presupposition in Language Generation and Interpretation. Stanford, CA: CSLI Publications, –. Krahmer, Emiel and van Deemter, K. (). ‘Computational generation of referring expressions: a survey’, Computational Linguistics (): –. Krahmer, Emiel, Theune, M., Viethen, J., and Hendrickx, I. (). ‘Graph: the costs of redundancy in referring expressions’, in Proceedings of the International Conference on Natural Language Generation (INLG), –, Salt Fork, Ohio. Krahmer, Emiel, van Erk, S., and Verleg, A. (). ‘Graph-based generation of referring expressions’, Computational Linguistics (): –. Kratzer, Angelika (). ‘Conditionals’, Chicago Linguistics Society (): –. Kratzer, Angelika (). ‘An investigation of the lumps of thought’, Linguistics and Philosophy (): –. Kratzer, Angelika (). ‘Scope or pseudoscope? Are there wide-scope indefinites?’, in S. D. Rothstein (ed.), Events and grammar, Dordrecht: Kluwer, –. Kratzer, Angelika (). A note on choice functions in context. Unpublished Ms. University of Massachusetts at Amherst. Kratzer, Angelika (). ‘Covert quantifier restrictions in natural languages’, talk given at Palazzo Feltrinelli in Gargnano  June . Kratzer, Angelika (). ‘Situations in natural language semantics’, in E. N. Zalta (ed.), Stanford Encyclopedia of Philosophy, CSLI: Stanford. Krauss, Robert, Yihsiu Chen, and Purnima Chawla (). ‘Nonverbal behavior and nonverbal communication: what do conversational hand gestures tell us?’, in M. Zanna (ed.), Advances in Experimental Social Psychology. San Diego, CA: Academic Press, –. Krifka, Manfred, Francis Jeffry Pelletier, Greg Carlson, Alice ter Meulen, Gennaro Chierchia, and Godehard Link (). ‘Introduction’, in G. Carlson and F. Pelletier (eds), The Generic Book, Chicago: University of Chicago Press, –. Kripke, Saul (). ‘Semantical analysis of modal logic I’, Zeitschrift fr Mathematische Logik und Grundlagen der Mathematik , –. Kripke, Saul A. (). ‘Naming and necessity’, in D. Davidson and G. Harman (eds), Semantics of Natural Languages, Reidel: Dordrecht. [Revised and enlarged revision published in  by Blackwell, Oxford]. Kripke, Saul (). ‘Speaker reference and semantic reference’, in French, Uehling, and Wettstein (eds), Contemporary Perspectives in the Philosophy of Language, Minneapolis: University of Minnesota Press, –.

OUP CORRECTED PROOF – FINAL, 21/1/2019, SPi

REFERENCES



Kripke, Saul (). ‘A puzzle about belief ’, in A. Margalit (ed.) Meaning and Use, Reidel: Dordrecht, –. Kripke, Saul (). Naming and Necessity. Cambridge, MA: Harvard University Press. Kripke, Saul (). Naming and Necessity. Cambridge, Massachusetts: Harvard University Press. Kripke, Saul (). Reference and Existence. Oxford, Oxford University Press. Kronfeld, Amichai (). The Referential/Attributive Distinction and the Conceptual/ Descriptive Approach to the Problem of Reference. PhD dissertation. University of California, Berkeley. Kronfeld, Amichai (). ‘Donnellan’s distinction and a computational model of reference’, Proceedings of the th Meeting of the Association for Computational Linguistics, New York, –. Kronfeld, Amichai (). Reference and Computation. Cambridge: Cambridge University Press. Kruijff, G.-J. M., Lison, P., Benjamin, T., Jacobsson, H., and Hawes, N. (). ‘Incremental, multi-level processing for comprehending situated dialogue in human–robot interaction’, in Symposium on Language and Robots. Kumar, V. (). ‘Algorithms for constraint satisfaction problems: a survey’, Artificial Intelligence Magazine : –. Kuno, Susumu (). The Structure of the Japanese Language. Cambridge, MA: The MIT Press. Kuno, Susumu (). Functional Syntax: Anaphora, Discourse, and Empathy. Chicago: University of Chicago Press. Kuroda, S.-Y. (). Generative Grammatical Studies in the Japanese Language. MA: MIT dissertation. Kusumoto, K. (). ‘On the quantification over times in natural language’, Natural Language Semantics (): –. Kutas, M. and Federmeier, K. D. (). ‘Thirty years and counting: finding meaning in the N component of the Event-Related Brain Potential (ERP)’, Annu. Rev. Psychol. (): –. Kutas, M. and S. Hillyard (). ‘Reading senseless sentences: brain potentials reflect semantic incongruity’, Science : –. Kutas, M. and Hillyard, S. (). ‘Event-related brain potentials to syntactic errors and semantic anomalies’, Memory and Cognition (): –. Kutas, M. and S. Hillyard (). ‘Brain potentials during reading reflect word expectancy and semantic association’, Nature : –. Kutas, M., Kluender, R., Barkley, C., and Amsel, B. (). ‘Language’, in Cacioppo, J., Tassinary, L., and Berntson, G. (eds), Handbook of Psychophysiology ( th edition), Cambridge: Cambridge University Press. Kutlák, R. (). Generation of Referring Expressions for an Unknown Audience. PhD thesis, University of Aberdeen. Kwon, N. and Sturt, P. (). ‘Null pronominal (pro)-resolution in Korean, a discourseoriented language’, Language and Cognitive Processes (): –. Kwon, N., Kluender, R., Kutas, M., and Polinsky, M. (). ‘Subject/object processing asymmetries in Korean relative clauses: evidence from ERP data’, Language (): –.

OUP CORRECTED PROOF – FINAL, 21/1/2019, SPi



REFERENCES

Ladusaw, William A. (). ‘Semantic constraints on the English partitive construction’, Proceedings of the West Coast Conference on Formal Linguistics (WCCFL) : –. Laitinen, L. (). ‘Hän, the third speech act pronoun in Finnish’, in Ritva Laury (ed.), Minimal Reference. The Use of Pronouns in Finnish and Estonian, Helsinki: Finnish Literature Society, –. Lakoff, Robin (). ‘Remarks on this and that’, in M. LeGaly, R. Fox, and A. Bruck (eds), Papers from the Tenth Regional Meeting of the Chicago Linguistics Society, Chicago: Chicago Linguistics Society, –. Lambrecht, Knud (). Information Structure and Sentence Form. Cambridge: Cambridge University Press. Lambrecht, Knud (). Topic, Focus, and the Grammar of Spoken French. Berkeley: University of California dissertation. Landman, F. (). Indefinites and the Type of Sets. Blackwell Publishers: Oxford. Lane, L. W., Groisman, M., and Ferreira, V. S. (). ‘Don’t talk about pink elephants! speakers’ control over leaking private information during language production’, Psychological Science (): –. Leaper, C. and Gleason, J. B. (). ‘The relationship of play activity and gender to parent and child sex-typed communication’, International Journal of Behavioral Development (): –. Lee-Goldman, Russell (). Context in Constructions. PhD dissertation, University of California, Berkeley. Lemaignan, S., Ros, R., Alami, R., and Beetz, M. (). ‘What are you talking about? Grounding dialogue in a perspective-aware robotic architecture’, in Proceedings of the th IEEE International Symposium on Robot and Human Interactive Communication, –. Lenat, D. (). ‘CYC: A large-scale investment in knowledge infrastructure’, Communication of the ACM : –. Lester, J., Voerman, J., Towns, S., and Callaway, C. (). ‘Deictic believability: coordinating gesture, locomotion, and speech in lifelike pedagogical agents’, Applied Artificial Intelligence : –. Levelt, Willem J. M. (). Speaking: From Intention to Articulation. Cambridge, MA: The MIT Press. Levenshtein, V. I. (). ‘Binary codes capable of correcting deletions, insertions, and reversals’, Soviet Physics Doklady : –. Levinson, Stephen C. (). Pragmatics. Cambridge: Cambridge University Press. Levinson, Stephen C. (). ‘Pragmatics and the grammar of anaphora: A partial pragmatic reduction binding and control phenomena’, Journal of Linguistics : –. Levinson, Stephen C. (). ‘Deixis’, in L. R. Horn and G. Ward (eds), Handbook of Pragmatics, Oxford: Blackwell. Lewis, David (). ‘General semantics’, Synthese : –. [Reprinted in Davidson and Harman (eds) Semantics of Natural Language, Dordrecht: Reidel]. Lewis, David (). ‘Adverbs of quantification’, in E. L. Keenan (ed.), Formal Semantics of Natural Language, Cambridge University Press: Cambridge, –. URL: http://onlinelibrary. wiley.com/doi/./.ch/summary Lewis, David (a). ‘Scorekeeping in a language game’, Journal of Philosophical Logic : –. Lewis, David (b). ‘Attitudes de dicto and de se’, The Philosophical Review (): –.

OUP CORRECTED PROOF – FINAL, 21/1/2019, SPi

REFERENCES



Lewis, Karen S. (). ‘Speaker’s reference and anaphoric pronouns’, Philosophical Perspectives : –. Lewis, R. and Vasishth, S. (). ‘An activation-based model of sentence processing as skilled memory retrieval’, Cognitive Science : –. Lewis, R., Vasishth, S., and Van Dyke, J. (). ‘Computational principles of working memory in sentence comprehension’, Trends in Cognitive Sciences (): –. Lezama, C. and Almor, A. (). ‘Repeated names, overt pronouns, and null pronouns in Spanish’, Language and Cognitive Processes (): –. Li, Charles N. and Thompson, Sandra A. (). Mandarin Chinese: A functional reference grammar. Berkeley: University of California Press. Lieberman, H., Liu, H., Singh, P., and Barry, B. (). ‘Beating common sense into interactive applications’, AI Magazine Winter : –. Lin, C.-Y. and Hovy, E. (). ‘Automatic evaluation of summaries using N-gram cooccurrence statistics’, in Proceedings of the Human Language Technology Conference of the North American Chapter of the Association for Computational Linguistics (HLTNAACL), –, Edmonton, Canada. Lin, D. (). ‘An information-theoretic de_nition of similarity’, in Proceedings of the th International Conference on Machine Learning (icml), –, Madison, Wisconsin. Linde, Charlotte (). ‘Focus of attention and the choice of pronouns in discourse’, in T. Givón (ed.), Discourse and syntax. New York: Academic Press, –. Lipman, B. (). ‘Why is language vague?’, in Working Papers, Department of Economics, Boston University. Liszkowski, U., Carpenter, M., Striano, T., and Tomasello, M. (). ‘- and -month-olds point to provide information for others’, Journal of Cognition and Development (): –. Litman, Diane J. and Allen, James F. (). ‘Discourse processing and commonsense plans’, in P. R. Cohen, J. Morgan, and M. E. Pollack (eds), Intentions in Communication. Cambridge, MA: MIT Press, –. Liu, C., Fang, R., She, L., and Chai, J. (). ‘Modeling collaborative referring for situated referential grounding’, in Proceedings of the  SIGDIAL Conference, –. Lloyd, Peter, Humbert Boada, and Maria Forns (). ‘New directions in referential communication research’, British Journal of Developmental Psychology : –. Löbner, Sebastian (). ‘Definites’, Journal of Semantics : –. Löbner, Sebastian (). ‘Polarity in natural language: predication, quantification, and negation in particular and characterizing sentences’, Linguistics and Philosophy , –. Löbner, Sebastian (). Definite Associative Anaphora. Düsseldorf: Heinrich-HeineUniversität, Ms. Lønning, J. T. (). ‘Plurals and collectivity’, in J. van Benthem and A. ter Meulen (eds), Handbook of Logic and Language, Amsterdam: Elsevier, –. Love, J. and McKoon, G. (). ‘Rules of engagement: incomplete and complete pronoun resolution’, Journal of Experimental Psychology: Learning, Memory, and Cognition : –. Ludlow, Peter (). ‘Referential semantics for I-languages?’, in L. M. Antony and N. Hornstein (eds), Chomsky and His Critics. Oxford: Blackwell, –. Ludlow, Peter and Neale, Stephen (). ‘Indefinite descriptions: in defense of Russell’, Linguistics and Philosophy (): –. Lyon, John (). Predication and Equation in Okanagan Salish: The Syntax and Semantics of Determiner Phrases. PhD dissertation. University of British Columbia.

OUP CORRECTED PROOF – FINAL, 21/1/2019, SPi



REFERENCES

Lyon, John (). ‘Okanagan determiner phrases and domain restriction’, International Journal of American Linguistics (): –. Lyons, Christopher (). Definiteness. Cambridge Textbooks in Linguistics, Cambridge University Press. Lyons, John (). ‘Deixis as the source of reference’, in E. Keenan (ed.), Formal Semantics for Natural Language, Cambridge: Cambridge University Press, –. Machery, E., Mallon, R., Nichols, S., and Stich, S. P. (). ‘Semantics, cross-cultural style’, Cognition : B–B. MacKay, Alfred F. (). ‘Mr. Donnellan and Humpty Dumpty on referring’, The Philosophical Review (): –. Maclaran, Rose (). ‘On the two asymmetric uses of English demonstratives’, Linguistics : –. Maclaran, Rose (). The Semantics and Pragmatics of the English Demonstratives. PhD dissertation. Ithaca: Cornell University. MacMahon, M., Stankiewicz, B., and Kuipers, B. (). ‘Walk the talk: connecting language, knowledge, and action in route instructions’, in Proceedings of the st National Conference on Artificial intelligence, –). MacWhinney, Brian (). ‘Starting points’, Language (): . doi: ./. Maes, Alfons (). ‘Referent ontology and centering in discourse’, Journal of Semantics (): –. doi: ./jos/... Maes, Alfons, Arts, Anja, and Noordman, Leo (). ‘Reference management in instructive discourse’, Discourse Processes (): –. Malouf, R. (). ‘The order of prenominal adjectives in natural language generation’, in Proceedings of the th Annual Meeting of the Association for Computational Linguistics (ACL), –. Malone, Thomas, et al. (). ‘Intelligent information sharing systems’, Communications of the ACM, (): . Manek, G. and Tellex, S. (). ‘Incrementally identifying objects from referring expressions using spatial object models’, in Proceedings of the  RSS Workshop on Model Learning for Human–Robot Communication. Mann, W. and Thompson, S. (). ‘Rhetorical structure theory: toward a functional theory of text organization’, Text : –. Maratsos, M. P. (). The Use of Definite and Indefinite Reference in Young Children: An Experimental Study of Semantic Acquisition. Cambridge: Cambridge University Press. Marcos, H., Ryckebusch, C., and Rabain-Jamin, J. (). ‘Adult’s responses to young children’s communicative gestures: joint achievement of speech acts’, First Language (): –. Marr, D. (). Vision: A Computational Investigation. San Francisco: W. H. Freedman and Company. Marslen-Wilson, William, Elene Levy, and Lorraine K. Tyler (). ‘Producing interpretable discourse: The establishment and maintenance of reference’, in Robert J. Jarvella and Wolfgang Klein (eds), Speech, place, and action. Chichester, NY: John Wiley & Sons, –. Martin, Scott (). The Dynamics of Sense and Implicature. PhD dissertation, The Ohio State University. Matthews, D., Butcher, J., Lieven, E., and Tomasello, M. (). ‘Two- and four-year-olds learn to adapt referring expressions to context: effects of distracters and feedback on referential communication’. Topics in Cognitive Science (): –.

OUP CORRECTED PROOF – FINAL, 21/1/2019, SPi

REFERENCES



Matthews, D., Lieven, E. V. M., Theakston, A. L., and Tomasello, M. (). ‘The effect of perceptual availability and prior discourse on young children’s use of referring expressions’, Applied Psycholinguistics : –. Matthewson, Lisa (). Determiner Systems and Quantificational Strategies: Evidence from Salish. The Hague: Holland Academic Graphics. Matuszek, C., Fitzgerald, N., Zettlemoyer, L., Bo, L., and Fox, D. (). ‘A joint model of language and perception for grounded attribute learning’, in Proceedings of the th International Conference on Machine Learning, –. Matuszek, C., Herbst, E., Zettlemoyer, L., and Fox, D. (). ‘Learning to parse natural language commands to a robot control system’, in Proceedings of the th International Symposium on Experimental Robotics, –. May, R. (). The Grammar of Quantification. PhD thesis, MIT. Mayol, L. and Clark, R. (). ‘Pronouns in Catalan: games of partial information and the use of linguistic resources’, Journal of Pragmatics : –. Mazur-Palandre, A. and Jisa, H. (). ‘Introduire et développer l’information: une acquisition tardive?’, CogniTextes . Mazza, Veronica, Turatto, Massimo, and Umiltà, Carlo (). ‘Foreground-background segmentation and attention: a change blindness study’, Psychological Research-Psychologische Forschung (): –. doi: ./s---. McCluskey, E. J. (). Introduction to the Theory of Switching Circuits. New York: McGraw-Hill. McCoy, Kathleen F. and Strube, Michael (). ‘Generating anaphoric expressions: pronoun or definite description?’, in Proceedings of the ACL Workshop on The Relation of Discourse/ Dialogue Structure and Reference, –. McDonald, Janet L., Bock, J. Kathryn, and Kelly, Michael H. (). ‘Word and world order: semantic, phonological, and metrical determinants of serial position’, Cognitive Psychology (): –. doi: ./cogp... McElree, B., Foraker, S., and Dyer, L. (). ‘Memory structures that subserve sentence comprehension’, Journal of Memory and Language (): –. McGinn, Colin (). ‘The mechanism of reference’, Synthese (): –. McHale, B. (). ‘Free indirect discourse: a survey of recent accounts’, PTL: A Journal for Descriptive Poetics and Theory of Literature : –. McKinnon, R. and Osterhout, L. (). ‘Constraints on movement phenomena in sentence processing: evidence from event-related brain potentials’, Language and Cognitive Processes (): –. McNally, Louise (). ‘Existential sentences without existential quantification’, Linguistics and Philosophy (): –. Meinong. A. (). ‘Über Gegenstände höherer Ordnung und deren Verhältniss zur inneren Wahrnehmung’ in Zeitschrift für Psychologie und Physiologie der Sinnesorgane, : –. [Translated as ‘The Theory of Objects’, trans. Isaac Levi, D. B. Terrell, and Roderick Chisholm, in Roderick Chisholm. Atascadero (ed.), Realism and the Background of Phenomenology, CA: Ridgeview, , –]. Meinong, A. (ed.) (). Untersuchung zur Gegenstandstheorie und Psychologie. Leibzig. Mellish, C., Scott, D., Cahill, L., Paiva, D., Evans, R., and Reape, M. (). ‘A reference architecture for natural language generation systems’, Natural Language Engineering : –.

OUP CORRECTED PROOF – FINAL, 21/1/2019, SPi



REFERENCES

Metzing, C. A. and Brennan, S. E. (). ‘When conceptual pacts are broken: partner effects on the comprehension of referring expressions, Journal of Memory and Language : –. Meyer, A. S., Sleiderink, A. M., and Levelt, W. J. (). ‘Viewing and naming objects: eye movements during noun phrase production’, Cognition : B–B. Meyer, F. (). Grounding Words to Objects: A Joint Model for Co-reference and Entity Resolution Using Markov Logic for Robot Instruction Processing. Unpublished doctoral dissertation, TUHH. Mikkelsen, Line (). Specifying Who: On the Structure, Meaning, and Use of Specificational Copular Clauses. PhD Dissertation, University of California, Santa Cruz. Mill, J. S. (). A System of Logic. Honolulu: University Press of the Pacific. Millikan, Ruth Garrett (). Language, Thought, and Other Biological Categories. Cambridge, MA: MIT Press. Millikan, Ruth (). Varieties of Meaning: The  Jean Nicod Lectures. Cambridge, MA: MIT Press. Milsark, Gary (). Existential Sentences in English. PhD thesis, Massachusetts Institute of Technology, Cambridge, MA. Milsark, Gary (). ‘Toward an explanation of certain peculiarities of the existential construction in English’, Linguistic Analysis : –. Mitchell, M. (). ‘Class-based ordering of prenominal modifiers’, in Proceedings of the th European Workshop on Natural Language Generation (ENLG), –, Athens, Greece. Mitchell, M., Han, X., Dodge, J., Mensch, A., A., G., Berg, A., Daume, H., et al. (). ‘Midge: Generating image descriptions from computer vision detections’, in Proceedings of the th Conference of the European Chapter of the Association for Computational Linguistics (EACL), –, Stroudsburg, PA. Mitchell, M., van Deemter, K., and Reiter, E. (). ‘Generating expressions that refer to visible objects’, in Proceedings of the North American Chapter of the Association for Computational Linguistics (NAACL). Mitkov, R. (). Anaphora Resolution: The State of the Art. School of Languages and European Studies, University of Wolverhampton. Moll, H., Richter, N., Carpenter, M., and Tomasello, M. (). ‘Fourteen-month-olds know what ‘we’ have shared in a special way’, Infancy (): –. Montag, Jessica L. and MacDonald, Maryellen C. (). ‘Visual salience modulates structure choice in relative clause production’, Language and Speech (): –. doi: ./ . Montague, R. (a). ‘English as a formal language’, in B. Visentini et al. (eds), Linguaggi nella Societa et nella Technica, Milan: Edizioni di Communita, –. Montague, R. (b). ‘Universal grammar’, Theoria : –. Montague, R. (). ‘The proper treatment of quantification in ordinary English’, in K. J. J. Hintikka, J. M. E. Moravcsik, and P. Suppes (eds), Approaches to Natural Language (Synthese Library ), Dordrecht: Reidel, –. Montague, R. (). ‘The proper treatment of quantification in ordinary english’, in Formal Philosophy: Selected Papers of Richard Montague, Yale University Press, New Haven. Morales, A. (). ‘La hipótesis funcional y la aparición de sujeto no nominal: El español de Puerto Rico’, Hispania (): –. Morgan, J. L. (). Presupposition and the Representation of Meaning: Prolegomena. PhD thesis, University of Chicago. Morgenstern, A. and Parisse, C. (). ‘The Paris Corpus’, French Language Studies (): –.

OUP CORRECTED PROOF – FINAL, 21/1/2019, SPi

REFERENCES



Morrow, Daniel G. (). ‘Prominent characters and events organize narrative understanding’, Journal of Memory and Language (): –. Mount, Allyson (). ‘Intentions, gestures, and salience in ordinary and deferred demonstrative reference’, Mind & Language (): –. Mozuraitis, M. & Heller, D. (). Discourse coherence and the interpretation of accented pronouns. Dialogue & Discourse, () –. Müller, H., King, J., and Kutas, M. ().’ Event-related potentials elicited by spoken relative clauses’, Cognitive Brain Research : –. Münte, T., Schlitz, S., and Kutas, M. (). ‘When temporal terms belie conceptual order’, Nature : –. Musan, R. (). On the Temporal Interpretation of Noun Phrases, PhD thesis, Massachusetts Institute of Technology, Cambridge. Muskens, Reinhart (). ‘Combining Montague semantics and Discourse Representation’, Linguistics and Philosophy : –. Myachykov, Andriy, Garrod, Simon C., and Scheepers, Christoph (). ‘Attention and syntax in sentence production: a critical review’, Discours . doi: ./discours.. Nadig, A., Seth, S., and Sasson, M. (). ‘Global similarities and multifaceted differences in the production of partner-specific referential pacts by adults with autism spectrum disorders’, Frontiers in Psychology (): –. Nashawati, S. (). Le développement des expressions référentielles chez le jeune enfant: noms et pronoms dans des dialogues mère-enfant. (PhD), Université Sorbonne Nouvelle-Paris , Paris. Neale, Stephen (). Descriptions. MIT Press (Bradford Books), Cambridge, MA. Neale, Stephen (). ‘This, that, and the other’, in M. Reimer and A. Bezuidenhout (eds), Descriptions and Beyond. Oxford: Oxford University Press, –. Nelson, K. (). Young Minds in Social Worlds: Experience, Meaning, and Memory. Cambridge, MA.: Harvard University Press. Nenkova, A. and McKeown, K. R. (). ‘References to named entities: A corpus study’, in Proceedings of the Human Language Technology (HLT) Conference, Companion Volume, –. Neville, H., Nicol, J., Barss, A., Forster, K., and Garrett, M. (). ‘Syntactically based sentence processing classes: evidence from event related brain potentials’, Journal of Cognitive Neuroscience (): –. New, Joshua, Cosmides, Leda, and Tooby, John (). ‘Category-specific attention for animals reflects ancestral priorities, not expertise’, Proceedings of the National Academy of Sciences of the United States of America (): –. doi: ./ pnas.. Nieuwland, M. S. (). ‘ “Who’s he?” Event-related brain potentials and unbound pronouns’, Journal of Memory and Language : –. Nieuwland, M. S. and Van Berkum, J. (). ‘Individual differences and contextual bias in pronoun resolution’, Brain Research : –. Nieuwland, M. S. and Van Berkum, J. (a). ‘The interplay between semantic and referential aspects of anaphoric noun phrase resolution: evidence from ERPs’, Brain and Language : –. Nieuwland, M. S. and Van Berkum, J. (b). ‘The neurocognition of referential ambiguity in language comprehension’, Language and Linguistics Compass (): –. doi:./ j.-x...x.

OUP CORRECTED PROOF – FINAL, 21/1/2019, SPi



REFERENCES

Nieuwland, M. S., Otten, M., and Van Berkum, J. (). ‘Who are you talking about? Tracking discourse-level referential processing with ERPs’, Journal of Cognitive Neuroscience : –. Nieuwland, M. S., Petersson, K.M., and Van Berkum, J. (). ‘On Sense and Reference: Examining the Functional Neuroanatomy of Referential Processing’, NeuroImage, , –. Ninio, A. and Bruner, J. S. (). ‘The achievement and antecedents of labelling’, Journal of Child Language : –. Ninio, A. and Snow, C. (). Pragmatic Development. Boulder: Westview Press. Nunberg, Geoffrey (). The Pragmatics of Reference. Bloomington: Indiana University Linguistics Club. O’Grady, W. (). Syntactic Development. Chicago: University of Chicago Press. O’Rourke, Michael (). Understanding Descriptions. PhD dissertation, Stanford University. O’Rourke, Michael (). ‘The scope argument’, The Journal of Philosophy : –. Oberlander, J. (). ‘Do the right thing . . . but expect the unexpected’, Computational Linguistics : –. Ochs Keenan, E. and Shieffelin, B. (). ‘Topic as discourse notion: a study of topic in the conversations of children and adults’, in C. N. Li (ed.), Subject and Topics, New York: Academic Press, –. O’Donnell, M., Cheng, H., and Hitzeman, J. (). ‘Integrating referring and informing in NP planning’, in Proceedings of the Acl Workshop on the Computational Treatment of Nominals, –, Montreal, Canada. Olson, D. R. (). ‘Language and thought: aspects of a cognitive theory of semantics’, Psychological Review : –. Onea, Edgar (). Potential Questions at the Semantics-Pragmatics Interface. Leiden: Brill. Onea, Edgar and Geist, Ljudmilla (). ‘Indefinite determiners and referential anchoring’, International Review of Pragmatics (): –. O’Neill, D. K. (). ‘Two-year-old children’s sensitivity to a parent’s knowledge state when making requests’, Child Development (): –. O’Neill, D. K. (). ‘Talking about “new” information: the given/new distinction and children’s developing theory of mind’, in J. W. Astington and J. A. Baird (eds), Why Language Matters for Theory of Mind, Oxford: Oxford University Press, –. Osgood, Charles E. (). ‘Where do sentences come from?’, in D. D. Steinberg and L. A. Jakobovits (eds), Semantics: An Interdisciplinary Reader in Philosophy, Linguistics and Psychology, London: Cambridge University Press, –. Osterhout, L. (). ‘A superficial resemblance does not necessarily mean you are a part of the family: counterarguments to Coulson, King, and Kutas () in the P-SPS-P debate’, Language and Cognitive Processes (): –. Osterhout, L. and Holcomb, P. (). ‘Event-related brain potentials elicited by syntactic anomaly’, Journal of Memory and Language (): –. Osterhout, L. and Mobley, L. A. (). ‘Event-related brain potentials elicited by failure to agree’, Journal of Memory and Language : –. Overstreet, Maryann, and Yule, George (). ‘Locally Contingent Categorization’, Discourse Processes : –. Öztürk, Balkz (). Case, Referentiality, and Phrase Structure. Amsterdam: John Benjamins. Pablos, L., Doetjes, J., Ruijgrok, B., and Cheng, L. (). ‘Active search for antecedents in cataphoric pronoun resolution’, Frontiers in Psychology : –.

OUP CORRECTED PROOF – FINAL, 21/1/2019, SPi

REFERENCES



Pajusalu, R. (). ‘Eesti pronoomeneid I. Ühiskeele see, too, tema/ta [Estonian pronouns I. See, too, tema/ta in common Estonian]’, Keel ja Kirjandus : –; : –. Papineni, K., Roukos, S., Ward, T., and Zhu, W.-J. (). ‘BLEU: a method for automatic evaluation of machine translation’, in Proceedings of the th Annual Meeting of the Association for Computational Linguistics (ACL), –, Philadelphia, PA. Paraboni, I. and van Deemter, K. (). ‘Reference and the facilitation of search in spatial domains’. Language, Cognition and Neuroscience (). Paraboni, I., van Deemter, K., and Masthoff, J. (). ‘Generating referring expressions: making referents easy to identify’, Computational Linguistics (): –. Paradis, J. and Navarro, S. (). ‘Subject realization and crosslinguistic interference in the bilingual acquisition of Spanish and English: what is the role of the input?’, Journal of Child Language (): –. Parsons, T. (). Nonexistent Objects. New Haven: Yale University Press. Partee, Barbara (). ‘Opacity, coreference, and pronouns’, Synthese : –. Partee, Barbara H. (). ‘Opacity, coreference, and pronouns’, in Donald Davidson and Gilbert Harman (eds), Semantics of Natural Language, Dordrecht: Reidel, –. Partee, Barbara H. (). ‘Nominal and temporal anaphora’, Linguistics and Philosophy : –. Partee, Barbara H. (). ‘Noun phrase interpretation and type-shifting principles’, in Jeroen Groenendijk, Dick de Jongh, and Martin Stokhof (eds), Studies in Discourse Representation Theory and the Theory of Generalized Quantifiers, Dordrecht: Foris, –. Partee, Barbara H. (). ‘Noun phrase interpretation and type-shifting principles’, in J. A. G. Groenendijk, D. de Jongh, and M. J. B. Stokhof (eds), Studies in Discourse Representation Theory and the Theory of Generalized Quantifiers, Dordrecht: Foris, –. Partee, Barbara H. (). ‘A note on Mandarin possessives, demonstratives, and definiteness’, in Drawing the Boundaries of Meaning: Neo-Gricean Studies in Pragmatics and Semantics in Honor of Laurance R. Horn, Philadelphia: John Benjamins, –. Passonneau, Rebecca J. (). ‘Using centering to relax Gricean informational constraints on discourse anaphoric noun phrases’, Language and Speech (/): –. Passonneau, Rebecca J. (). ‘Measuring agreement on set-valued items (MASI) for semantic and pragmatic annotation’, in Proceedings of the th International Conference on Language Resources and Evaluation (LREC). Genoa, Italy. Pearson, Hazel (). The Sense of Self: Topics in the Semantics of De Se Expressions. PhD dissertation, Harvard University. Pearson, Hazel (). ‘The interpretation of the logophoric pronoun in Ewe’, Natural Language Semantics : –. Pechmann, T. (). ‘Incremental speech production and referential over-specification’, Linguistics : –. Pekarek, S. (). ‘Linguistic forms and social interaction: why do we specify referents more than is necessary for their identification?’, in J. Verschueren (ed.), Pragmatics in . Selected Papers from the th International Pragmatics Conference (Vol. ), Reims: IPRA, –. Pekarek Doehler, S. (). ‘Referential processes as situated cognition: pronominal expressions and the social co-ordination of talk’, in T. Eniko Nemeth (ed.), Cognition in Language use. Selected Papers from the th International Pragmatics Conference (Vol. ), Budapest: IPRA, –.

OUP CORRECTED PROOF – FINAL, 21/1/2019, SPi



REFERENCES

Percus, O. (). ‘Constraints on some other variables in syntax’, Natural Language Semantics (): –. Percus, Orin and Uli Sauerland (a). ‘On the LFs of attitude reports’, in Proceedings of the Conference “Sub - Sinn Und Bedeutung”, University of Konstanz, –. Percus, Orin and Uli Sauerland (b). ‘Pronoun movement in dream reports’, in NELS : Proceedings of the rd Annual Meeting of the North East Linguistic Society, MIT, Cambridge MA, –. Perera, R. and Nand, P. (). ‘Recent advances in natural language generation: a survey and classification of the empirical literature’, Computing and Informatics . Pérez-Leroux, A. T., Pirvulescu, M., and Roberge, Y. (). ‘Null objects in child language: syntax and the lexicon’, Lingua : –. Pérez-Leroux, A. T., Pirvulescu, M., and Roberge, Y. (). ‘Topicalization and object omission in child language’, First Language (): –. Perlman, A. (). ‘This as a third article in American English’, American Speech : –. Perner, J. (). ‘Theory of mind’, Developmental Psychology: Achievements and Prospects –. Perry, John (). Frege on Demonstratives. The Philosophical Review (): –. Perry, John (). ‘The essential indexical,’ Noûs : –. Perry, John (). ‘Cognitive significance and new theories of reference,’ Noûs : –. Perry, John (). ‘Indexicals and demonstratives’, in R. Hale and C. Wright (eds), Companion to the Philosophy of Language. Oxford: Blackwell, –. Perry, John (). Reference and Reflexivity. Stanford, CLSI. Perry, John (). ‘Directing intentions’, in J. Almog and P. Leonardi (eds), The Philosophy of David Kaplan. Oxford: Oxford University Press, –. Perry, John (). Reference and Reflexivity, nd ed. Stanford: CSLI Publications. Pesetsky, David (). ‘Wh-in-situ: movement and unselective binding’, in E. J. Reuland and A. G. B. ter Meulen (eds), The Representation of (In)definiteness, Cambridge, MA: MIT Press, –. Peters, A. M. (). ‘Filler syllables: what is their status in emerging grammar?’, Journal of Child Language : –. Peters, Stanley and Dag Westerståhl (). Quantifiers in Language and Logic. Oxford: Clarendon Press. Peterson, C. and Dodsworth, P. (). ‘A longitudinal analysis of young children’s cohesion and noun specification in narratives’, Journal of Child Language : –. Phillips, C., N. Gazanina, and A. Abada (). ‘ERP effects of the processing of syntactic long-distance dependencies’, Cognitive Brain Research : –. Pica, Teresa, Dom Berducci, Lloyd Holliday, Nora Lewis, and Jeanne Newman (). ‘Language learning through interaction: what role does gender play?’, Penn Working Papers in Educational Linguistics : –. Pickering, M. and Garrod, S. (). ‘Toward a mechanistic psychology of dialogue’, Behavioural and Brain Sciences : –. Pirvulescu, M. (). ‘Theoretical implications of object clitic omission in early French: spontaneous vs. elicited production’, Catalan Journal of Linguistics : –. Piwek, Paul, Beun, Robbert-Jan, and Cremers, Anita H. M. (). ‘ “Proximal” and “distal” in language and cognition: evidence from deictic demonstratives in Dutch’, Journal of Pragmatics (): –. doi: ./j.pragma....

OUP CORRECTED PROOF – FINAL, 21/1/2019, SPi

REFERENCES



Poesio, Massimo and Renata Vieira (). ‘A corpus-based investigation of definite description use’, Computational Linguistics : –. Poesio, Massimo, Stevenson, Rosemary, Di Eugenio, Barbara, and Hitzeman, Janet (). ‘Centering: a parametric theory and its instantiations’, Computational Linguistics (): –. doi: ./. Politzer-Ahles, S., Fiorentino, R., Jiang, X., and Zhou, X., (). ‘Distinct neural correlates for pragmatic and semantic meaning processing: an event-related potential investigation of scalar implicature processing using picture-sentence verification’, Brain Research (): –. Pollack, Martha E. (). ‘Plans as complex mental attitudes’, in P. R. Cohen, J. Morgan, and M. E. Pollack (eds), Intentions in Communication. Cambridge, MA: MIT Press, –. Pollack, Martha E. (). ‘Overloading intentions for efficient practical reasoning’, Noûs : –. Popescu-Belis, A., Robba, I., and Sabah, G. (). ‘Reference resolution beyond coreference: a conceptual frame and its application’, in Proceedings of the th International Conference on Computational Linguistics, –. Portet, F., Reiter, E., Gatt, A., Hunter, J., Sripada, S., Freer, Y., and Sykes, C. (). ‘Automatic generation of textual summaries from neonatal intensive care data’, Artificial Intelligence : –. Portner, Paul (). ‘Topicality and (non-)specificity in Mandarin’, Journal of Semantics : –. Portner, Paul and Yabushita, Katsuhiko (). ‘Specific indefinites and the information structure theory of topics’, Journal of Semantics : –. Power, R. J. D. and Dal Martello, M. F. (). ‘The use of definite and indefinite articles by Italian preschool children’, Journal of Child Language : –. Prat-Sala, Mercè and Branigan, Holly P. (). ‘Discourse constraints on syntactic processing in language production: a cross-linguistic study in English and Spanish’, Journal of Memory and Language (): –. doi: ./jmla... Pratt, Jay, Radulescu, Petre V., Guo, Ruo Mu, and Abrams, Richard A. (). ‘It’s alive! animate motion captures visual attention’, Psychological Science (): –. doi: ./. Prince, Ellen (a). ‘On the inferencing of indefinite-this NPs’, in B. L. Webber, A. K. Joshi, and I. A. Sag (eds), Elements of Discourse Understanding, Cambridge: Cambridge University Press, –. Prince, Ellen (b). ‘Toward a taxonomy of given-new information’, in Peter Cole (ed.), Radical Pragmatics, New York: Academic Press, –. Prince, Ellen (). ‘Fancy syntax and “shared knowledge” ’, Journal of Pragmatics : –. Prince, Ellen (). ‘The ZPG letter: subjects, definiteness, and information-status’, in S. Thompson and W. Mann (eds), Discourse Description: Diverse Analyses of a Fund Raising Text, Philadelphia/Amsterdam: John Benjamins, –. Prior, A. N. (). ‘Modality de dicto and modality de re’, Theoria (): –. URL: http:// onlinelibrary.wiley.com.proxy.lib.umich.edu/doi/./j.-..tb.x/abstract Prior, A. N. (). Time and Modality. Clarendon Press, Oxford. Prud’Hommeaux, E., Seaborne, A., et al. (). ‘SPARQL Query Language for RDF’, WC recommendation, . Pylkkänen, L., Brennan, J., and Bemis, D. K. (). ‘Grounding the cognitive neuroscience of semantics in linguistic theory’, Language and Cognitive Processes (): –.

OUP CORRECTED PROOF – FINAL, 21/1/2019, SPi



REFERENCES

Pyykkönen, P. and Järvikivi, J. (). ‘Activation and persistence of implicit causality information in spoken language comprehension’, Experimental Psychology (): –. Pyykkönen, P., Matthews, D., and Järvikivi, J. (). ‘Three-year-olds are sensitive to semantic prominence during online language comprehension: a visual world study of pronoun resolution’, Language and Cognitive Processes : –. Qing, C. and Franke, M. (). ‘Variations on a bayesian theme: comparing Bayesian models of referential reasoning’, in H.-C. Schmitz and H. Zeevat (eds), Bayesian Natural Language Semantics and Pragmatics, Heidelberg: Springer, –. Quine, Willard van Orman (). ‘On what there is’, Review of Metaphysics. [Reprinted in Quine ()]. Quine, Willard van Orman (). Methods of Logic. Cambridge: Harvard University Press. Quine, Willard van Orman (). From a Logical Point of View. Cambridge: Harvard University Press. Quine, Willard van Orman (). ‘Quantifiers and propositional attitudes’, The Journal of Philosophy (): –. Quine, Willard van Orman (). Word and Object. Cambridge, MIT Press. Quine, Willard van Orman (). ‘Propositional objects’, in Quine, Ontological Relativity and Other Essays. New York: Columbia University Press. Quirk, R., Greenbaum, S., Leech, G., and Svartvik, J. (). A Grammar of Contemporary English (Ninth Impression). Burnt Mill, Harlow, Essex: Longman. Raghunathan, B. (). The Complete Book of Data Anonymization: From Planning to Implementation. CRC Press, Boca Raton, London, and New York. Read, R. C. and Corneil, D. G. (). ‘The graph isomorphism disease’, Journal of Graph Theory (): –. Recanati, F. (). ‘Domains of discourse’, Linguistics and Philosophy (): –. Recanati, F. (). Literal Meaning. Cambridge: Cambridge University Press. Recanati, F. (). ‘Descriptions and situations’, in M. Reimer and A. Bezuidenhout (eds), Descriptions and Beyond, Clarendon Press, Oxford, –. Recanati, F. (). Mental Files. Oxford: Oxford University Press. Reimer, Marga (). ‘Do demonstrations have semantic significance?’, Analysis (): –. Reimer, Marga (). ‘Three views of demonstrative reference’, Synthese : –. Reimer, Marga (). ‘The problem of empty names’, Australasian Journal of Philosophy (): –. Reimer, Marga (). ‘Reference’, The Stanford Encyclopedia of Philosophy (Spring  Edition), E. N. Zalta (ed.), URL=. Reinhart, Tanya (). ‘Pragmatics and linguistics: an analysis of sentence topics’, Philosophica (): –. Reinhart, Tanya (). Anaphora and Semantic Interpretation. London: Croom Helm. Reinhart, Tanya (). ‘Quantifier scope: how labor is divided between QR and choice functions’, Linguistics and Philosophy : –. Reiter, E. (). ‘The computational complexity of avoiding conversational implicatures’, in Proceedings of the th Annual Meeting of the Association for Computational Linguistics (ACL), –. Reiter, E. and Dale, R. (). ‘A fast algorithm for the generation of referring expressions’, in Proceedings of the th international conference on computational linguistics (coling), –, Nantes, France.

OUP CORRECTED PROOF – FINAL, 21/1/2019, SPi

REFERENCES



Reiter, E. and Dale, R. (). Building Natural Language Generation Systems. Cambridge University Press. Ren, Y., van Deemter, K., and Pan, J. (). ‘Charting the potential of Description Logic for the generation of referring expressions’, in Proceedings of the th International Natural Language Generation Conference (INLG), –. Richard, M. (). ‘Commitment’, Philosophical Perspectives : –. Rivero, Marìa-Luisa (). ‘Referential properties of spanish noun phrases’, Language : –. Rizzi, L. (/). ‘Some notes on linguistic theory and language development: the case of root infinitives’, Language Acquisition (): –. Roberts, Craige (). ‘Modal subordination and pronominal anaphora in discourse’, Linguistics and Philosophy (): –. [Reprinted in Javier Gutierrez-Rexach (ed.) Semantics: Critical Concepts in Linguistics, Routledge, ]. Roberts, Craige (). ‘Domain selection in dynamic semantics’, in Emmon Bach, Eloise Jelinek, Angelika Kratzer, and Barbara H. Partee (eds), Quantification in Natural Languages, Kluwer, Dordrecht. Roberts, Craige (/). ‘Information structure in discourse: towards an integrated formal theory of pragmatics’, in Jae Hak Yoon and Andreas Kathol (eds) Ohio State University Working Papers in Linguistics Volume . [Reprinted in the  version with a new Afterword in Semantics and Pragmatics, Volume , ]. Roberts, Craige (b). ‘Anaphora in intensional contexts’, in Shalom Lappin (ed.) The Handbook of Contemporary Semantic Theory, Basil Blackwell, . Roberts, Craige (). ‘Demonstratives as definites’, in K. van Deemter and R. Kibble (eds), Information Sharing: Reference and Presupposition in Language Generation and Interpretation, Stanford: CSLI, –. Roberts, Craige (). ‘Uniqueness in definite noun phrases’, Linguistics and Philosophy : –. Roberts, Craige (). ‘Context in dynamic interpretation’, in L. R. Horn and G. Ward (eds), The Handbook of Pragmatics. Melden, MA: Blackwell, –. Roberts, Craige (). ‘Pronouns as definites’, in M. Reimer and A. Bezuidenhout (eds), Descriptions and Beyond, Oxford University Press. Roberts, Craige (a). ‘Accommodation in a language game’, in Berry Loewer and Jonathan Schaffer (eds), A Companion to David Lewis, Wiley-Blackwell, Hobeken, NJ. Roberts, Craige (b). Indexicality: A De Se Semantics and Pragmatics. Unpublished Manuscript. The Ohio State University. Roberts, Lawrence D. (). ‘How demonstrations connect with referential intentions’, Australasian Journal of Philosophy (): –. Rochemont, Michael and Peter Culicover (). English focus constructions and the theory of grammar. Cambridge: Cambridge University Press. Rohde, H. (). Coherence-Driven Effects in Sentence Discourse Processing. PhD dissertation. University of California, San Diego. Rohde, H., and Kehler, Andrew (). ‘Grammatical and Information-Structural Influences on Pronoun Production’, Language, Cognition and Neuroscience : , –. doi: ./... Rohde, H. and Kehler, A. (). ‘Grammatical and information-structural influences on pronoun production’, Language, Cognition, and Neuroscience : –. Rohde, H., Kehler, A., and Elman, J. L. (). ‘Event structure and discourse coherence biases in pronoun interpretation’, in R. Sun (ed.), Proceedings of the th Annual Conference of the Cognitive Science Society, Mahwah, NJ: Lawrence Erlbaum Associates, –.

OUP CORRECTED PROOF – FINAL, 21/1/2019, SPi



REFERENCES

Roll, M., Horne, M., and Lindgren, M. (). ‘Object shift and event-related brain potentials’. Journal of Neurolinguistics : –. Roman Kutlak and Kees van Deemter and Chris Mellish (). Production of referring expressions for an unknown audience: a computational model of communal common ground. Frontiers of Psychology,  Aug. . Romoli, J. and Sudo, Y. (). ‘De re/de dicto ambiguity and presupposition projection’, in A. Riester and T. Solstad (eds), Proceedings of Sinn und Bedeutung , University of Stuttgart, –. Rooth, Mats E. (). Association with Focus. PhD dissertation, University of Massachusetts at Amherst. Rooth, Mats E. (). ‘A theory of focus interpretation’, Natural Language Semantics : –. Rosa, Elise C., Finch, Kayla H., Bergeson, Molly, and Arnold, Jennifer E. (). ‘The effects of addressee attention on prosodic prominence’, Language and Cognitive Processes (/): –. doi: ./... Rosch, E. (). ‘Principles of categorization’, in E. Rosch and B. L. Lloyd (eds), Cognition and Categorization, Hillsdale, NJ: Erlbaum, –. Rothschild, Daniel (). ‘Presuppositions and scope’, Journal of Philosophy : –. Roy, D. (). ‘Grounding words in perception and action: computational insights’, Trends in Cognitive Sciences (): –. Roy, D. and Pentland, A. (). ‘Learning words from sights and sounds: a computational model’, Cognitive Science : –. Roy, D., Hsiao, K.-Y., Mavridis, N., and Gorniak, P. (). ‘Ripley, Hand Me The Cup! Sensorimotor representations for grounding word meaning, in Proceedings of the International Conference of Automatic Speech Recognition and Understanding. Rozendaal, M. and Baker, A. (). ‘A cross-linguistic investigation of the acquisition of the pragmatics of indefinite and definite reference in two-year-olds’, Journal of Child Language : –. Rubio-Fernández, Paula (). ‘On the automaticity of egocentricity: a review of the egocentric anchoring and adjustment model of perspective taking’, UCL Working Papers in Linguistics : –. Rugg, M. and Coles, M. (). ‘The ERP and cognitive psychology: conceptual issues’, in M. D. Rugg and M. G. H. Coles (eds), Electrophysiology of Mind: Event Related Brain Potentials and Cognition (Vol. ), Oxford: Oxford University Press, –. Runner, J. (). ‘Syntactic structure, information structure, and lexical effects on null and overt subject comprehension in Spanish’, talk given at the th Annual Conference of the German Linguistic Society (DGfS), Workhop on “Information Structural Evidence in the Race for Salience,” March , Potsdam, Germany. Runner, J. and Ibarra, A. (). ‘Information structure effects on null and overt subject comprehension in Spanish’, in Anke, Holler and Katja Suckow (eds), Empirical Perspectives on Anaphora Resolution, De Gruyter, –. Russell, Bertrand (). The Principles of Mathematics. Cambridge: Cambridge University Press. Russell, Bertrand (). ‘On denoting’, Mind (): –. Russell, Bertrand (). ‘The Relation of Sense Data to Physics’, Scientia, , –. Russell, Bertrand (). ‘The Philosophy of Logical Atomism’, in Russell, Bertrand (), –.

OUP CORRECTED PROOF – FINAL, 21/1/2019, SPi

REFERENCES



Russell, Bertrand (). Introduction to Mathematical Philosophy. London: George Allen & Unwin. Russell, Bertrand (). ‘Logic and Knowledge’, in R.C. Marsh (ed.), London: Allen & Unwin, . Ruys, Eduard (). ‘Unexpected wide scope phenomena’, in M. Everaert and H. van Riemsdijk (eds), The Blackwell Companion to Syntax, vol. , Oxford: Blackwell, –. Sadrai, Mahmoud (). Cognitive Status and RA-marked Referents of Nominal Expressions in Persian Discourse. PhD Dissertation. University of Minnesota. Sæbø, Kjell Johan (). ‘Reports of specific indefinites’, Journal of Semantics : –. Sag, I. (). Deletion and Logical Form. PhD thesis, Massachusetts Institute of Technology. Sag T. and Hankamer J. (). ‘Toward a theory of anaphoric processing’, Linguistics & Philosophy : –. Salazar Orvig, A. (). ‘La reprise aux sources de la construction discursive’, Langages : –. Salazar Orvig, A. (). ‘L’inscription dialogique du jeune enfant: évolution, diversité, et hétérogénéité’ TRANEL –: –. Salazar Orvig, A., Marcos, H., Morgenstern, A., Hassan, R., Leber-Marin, J., and Parès J. (). ‘Entre gramática y pragmática: adquisición de los determinantes en francés’, in Luis Miranda (ed.), Actas del V Congreso Nacional de Investigaciones Lingüístico-Filológicas. Cátedra UNESCO para la Lectura y la Escritura en América Latina (Sede Perú): Lima, –. Salazar Orvig, A., Marcos, H., Morgenstern, A., Hassan, R., Leber-Marin, J., and Parès, J. (a). ‘Dialogical beginnings of anaphora: the use of third person pronouns before the age of ’, Journal of Pragmatics (): –. Salazar Orvig, A., Marcos, H., Morgenstern, A., Hassan, R., Leber-Marin, J., and Parès, J. (b). ‘Dialogical factors in toddlers’ use of clitic pronouns’, First Language (–): –. Salazar Orvig, A. and Morgenstern, A. (). ‘Acquisition and uses of pronouns in a dialogic perspective’, in L. Serratrice and S. E. M. Allen (eds), The Acquisition of Reference, Amsterdam: John Benjamins, –. Salazar Orvig, A., Da Silva, C., Fox, G., Heurdier, J., Le Mené, M., and Marcos, H. (). ‘First Uses of French Demonstrative “ça”, the Development of Deictic Reference’, Paper presented at the Child Language Seminar , Manchester. Salazar Orvig, A., de Weck, G., Hassan, R., and Rialland, A. (eds). (in preparation). The Acquisition of Referring Expressions: A Dialogic Approach. Salazar Orvig, A., Hasan, R., Leber-Marin, J., Marcos, H., Morgenstern, A., and Parès, J. (). ‘Peut-on parler d’anaphore chez le jeune enfant ?’, Langages : –. Salazar Orvig, A., Marcos, H., Caët, S., Corlateanu, C., da Silva, C., Hassan, R., Heurdier, J., Le Mené, M., Leber-Marin, J., and Morgenstern, A. (). ‘Definite and indefinite determiners in French-speaking toddlers: distributional features and pragmatic-discursive factors’, Journal of Pragmatics : –. Salazar Orvig, A., Marcos, H., Heurdier, J., and da Silva, C. (). ‘Referential features, speech genres, and activity types’, in M. Hickmann, H. Jisa, and E. Veneziano (eds), Sources of Variation in First Language Acquisition: Languages, Contexts, and Learners, Amsterdam: Benjamins, –. Salmon, N. (). Frege’s Puzzle. Cambridge: MIT Press. Salmon, N. (). ‘Nonexistence’, Noûs : –.

OUP CORRECTED PROOF – FINAL, 21/1/2019, SPi



REFERENCES

Salmon, N. (). ‘Mythical objects’, in Campbell, J., O’Rourke, M., and Shier, D. (eds), Meaning and Truth: Investigations in Philosophical Semantics, New York, Seven Bridges Press, –. Salmon-Alt, Susanne and Romary, Laurent (). ‘Generating Referring Expressions in Multimodal Contexts’, International Natural Language Generation Conference. Salomo, D., Graf, E., Lieven, E., and Tomasello, M. (). ‘The role of perceptual availability and discourse context in young children’s question answering’, Journal of Child Language (): –. Salomo, D., Lieven, E., and Tomasello, M. (). ‘Young children’s sensitivity to new and given information when answering predicate-focus questions’, Applied Psycholinguistics (): –. Sandford, John (). Silken Prey. New York: Berkley. Sanford, A. J. and Sturt, P. (). ‘Depth of processing in language comprehension: not noticing the evidence’, Trends in Cognitive Sciences : –. Scha, R. and Stallard, D. (). ‘Multi-level plurals and distributivity’, in Proceedings of the th Annual Meeting of the Association for Computational Linguistics (ACL), –, Buffalo, NY. Schaeffer, J. and Matthewson, L. (). ‘Grammar and pragmatics in the acquisition of article systems’, Natural Language & Linguistic Theory : –. Schafer, R. and de Villiers, J. (). ‘Imagining Articles: What “a” and “the” can tell us about the emergence of DP’, Paper presented at the BUCLD . Scheutz, M., Eberhard, K., and Andronache, V. (). ‘A real-time robotic model of human reference resolution using visual constraints’, Connection Science Journal : –. Schlenker, Philippe (). ‘A plea for monsters’, Linguistics and Philosophy : –. Schober, Michael (). ‘Speakers, addressees, and frames of reference: whose effort is minimized in conversations about locations?’, Discourse Processes , –. Schumacher, P. (). ‘Definiteness marking shows late effects during discourse processing: evidence from ERPs’, in S. Lalitha Devi, A. Branco, and R. Mitkov (eds), Lecture Notes in Artificial Intelligence, Vol. : Anaphora Processing and Applications, Heidelberg: Springer, –. Schumacher, P., Roberts, L., and Järvikivi, J. (). ‘Agentivity drives real-time pronoun resolution: evidence from German er and der’, Lingua : –. Schwarz, Bernhard (). ‘Two kinds of long distance indefinites’, in R. van Rooy and M. Stokhof (eds), Proceedings of the Thirteenth Amsterdam Colloqium, Amsterdam: University of Amsterdam, –. Schwarz, F. (). Two Types of Definites in Natural Language. PhD thesis, University of Massachusetts Amherst, Amherst, MA. Schwarz, F. (). ‘Situation pronouns in determiner phrases’, Natural Language Semantics (): –. Schwarzschild, Roger (). ‘Singleton indefinites’, Journal of Semantics (): –. Scott, Kate J. (). The Relevance of Referring Expressions: The Case of Diary Drop in English. PhD dissertation. University College London. Searle, John R. (). ‘Proper names’, Mind : –. Searle, John (). Speech Acts: An Essay in the Philosophy of Language. Cambridge, UK: Cambridge University Press. Searle, John R. (). Speech Acts: An Essay in the Philosophy of Language. Cambridge: Cambridge University Press. Searle, John (a). Expression and Meaning. Cambridge: Cambridge University Press. Searle, John (b). ‘Referential and attributive’. Monist : –. Searle, John R. (). Intentionality. Cambridge: Cambridge University Press.

OUP CORRECTED PROOF – FINAL, 21/1/2019, SPi

REFERENCES



Searle, John R. (). ‘Collective intentions and actions’, in P. R. Cohen, J. Morgan, and M. E. Pollack (eds), Intentions in Communication. Cambridge, MA: MIT Press, –. Sedivy, Julie C. (). ‘Pragmatic versus form-based accounts of referential contrast: evidence for effects of informativity expectations’, Journal of Psycholinguistic Research (): –. doi: ./A:. Serratrice, L. (). ‘The role of discourse pragmatics in the acquisition of subjects in Italian’, Applied Psycholinguistics : –. Serratrice, L. (). ‘Null and overt subjects at the syntax–discourse interface. Evidence from monolingual and bilingual acquisition’, in S. Baauw, J. van Kampen, and M. Pinto (eds), The Acquisition of Romance Languages. Selected papers from The Romance Turn II , Utrecht: LOT, –. Serratrice, L. (). ‘The role of discourse and perceptual cues in the choice of referential expressions in English preschoolers, school-age children, and adults’, Language Learning and Development (): –. Serratrice, L. (). ‘The role of number of referents and animacy in children’s use of pronouns’, Journal of Pragmatics : –. Sgall, Petr, Hajičová, Eva, and Benešová, Eva (). Topic, Focus, and Generative Semantics. Kronberg: Scriptor Verlag. Sgall, Petr, Hajičová, Eva, and Panevová, Jarmila (). The Meaning of the Sentence in its Semantic and Pragmatic Aspects. Dordrecht: Reidel. Sharvy, Richard (). ‘A more general theory of definite descriptions’, Philosophical Review : –. Shaw, J. and Hatzivassiloglou, V. (). ‘Ordering among premodifiers’, in Proceedings of the th Annual Meeting of the Association for Computational Linguistics (ACL), –. Sheldon, A. (). ‘The role of parallel function in the acquisition of relative clauses in English’, Journal of Verbal Learning and Verbal Behavior, , –. Siddharthan, A. and Copestake, A. (). ‘Generating referring expressions in open domains’, in Proceedings of the nd Annual Meeting of the Association for Computational Linguistics (ACL), –, Barcelona, Spain. Siddharthan, A., Nenkova, A., and McKeown, K. (). ‘Information status distinctions and referring expressions: an empirical study of references to people in news summaries’, Computational Linguistics (): –. Sidner, Candace L. (). ‘Focusing in the comprehension of definite anaphor’, in Michael Brady (ed.), Computational models of discourse. Cambridge, MA: MIT Press, . Silverstein, M. (). ‘Hierarchy of features and ergativity’, in R. M. W. Dixon (ed.), Grammatical Categories in Australian Languages, New Jersey: Humanities Press. Simchen, Ori (a). ‘Necessity in reference’, in W. P. Kabasenche, M. O’Rourke, and M. H. Slater (eds), Reference and Referring. Cambridge, MA: MIT Press, –. Simchen, Ori (b). Necessary Intentionality. Oxford: Oxford University Press. Simpson, A., Wu, Z., and Li, Y. (). ‘Grammatical roles, coherence relations, and the interpretation of pronouns in Chinese’, Lingua Sinica (): –. Skarabela, B. (). ‘Signs of early social cognition in children’s syntax: the case of joint attention in argument realization in child Inuktitut’, Lingua (): –. Skarabela, B. and Allen, S. E. (). ‘The role of joint attention in argument realization in child Inuktitut’, Paper presented at the Proceedings of the Annual Boston University Conference on Language Development. Skarabela, B., Allen, S. E., and Scott-Phillips, T. C. (). ‘Joint attention helps explain why children omit new referents’, Journal of Pragmatics : –.

OUP CORRECTED PROOF – FINAL, 21/1/2019, SPi



REFERENCES

Smith, G. J. and Gero, J. S. (). ‘What does an artificial design agent mean by being ‘situated’?’ Design Studies (): –. Smyth, R. (). ‘Grammatical determinants of ambiguous pronoun resolution’, Journal of Psycholinguistic Research : –. Soames, S. (). ‘How presuppositions are inherited: a solution to the projection problem’, Linguistic Inquiry : –. Soames, S. (). ‘Direct reference, propositional attitudes, and semantic content’, Philosophical Topics (). Soames, S. (). Beyond Rigidity: The Unfinished Semantic Agenda of Naming and Necessity. Oxford: Oxford University Press. Song, H. and Fisher, C. (). ‘Who’s ‘she’? Discourse prominence influences preschoolers’ comprehension of pronouns’, Journal of Memory and Language : –. Sonnenschein, S. (). ‘The effect of redundant communication on listeners: why different types may have different effects’, Journal of Psycholinguistic Research : –. Sowa, J. (). Conceptual Structures: Information Processing in Mind and Machine. Addison-Wesley. Sperber, Dan and Deirdre Wilson (). Relevance. Cambridge, MA: Harvard University Press. Sperber, Dan and Deidre Wilson (). Relevance: Communication and Cognition, Second edition. Oxford: Blackwell. Spivey, M., Tyler, M., Eberhard, K., and Tanenhaus, M. (). ‘Linguistically mediated visual search’, Psychological Science : –. Sridhar, Shikaripur N. (). Cognition and Sentence Production: A Cross-linguistic Study. New York: Springer-Verlag. Stainton, Robert (). Words and Thoughts: Subsentences, Ellipsis, and the Philosophy of Language. Oxford: Oxford University Press. Stalnaker, Robert C. (). ‘Pragmatic presuppositions’, in Milton K. Munitz and Peter K. Unger (eds), Semantics and Philosophy, New York: New York University Press, –. Stalnaker, R. (). ‘Pragmatic presuppositions’, in Munitz, M. and Unger, P. (eds), Semantics and Philosophy, New York University Press, New York, –. Stalnaker, Robert (). ‘Assertion’, in Peter Cole (ed.), Syntax and Semantics, New York: Academic Press, –. Stalnaker, Robert (). ‘On the representation of context’, Journal of Logic, Language, and Information : –. Stalnaker, Robert C. (). ‘Common ground’, Linguistics and Philosophy : –. Stalnaker, Robert C. (). Our Knowledge of the Internal World. Oxford University Press, Chapter : ‘Locating ourselves in the world’. Stalnaker, Robert C. (). Context. Oxford University Press. Stanley, Jason and Zoltan Gendler Szabo (). ‘On quantifier domain restriction’, Mind & Language (–): –. Steedman, Mark. (). ‘Structure and intonation’, Language : –. Steedmann, Mark (). Taking Scope. The Natural Semantics of Quantifiers. Cambridge, MA: MIT Press. Steels, L. and Hild, M. (). Language Grounding in Robots. Springer Science & Business Media. Steinhauer, K. and Drury, J. (). ‘On the early left anterior negativity (eLAN) in syntax studies’, Brain and Language (): –.

OUP CORRECTED PROOF – FINAL, 21/1/2019, SPi

REFERENCES



Sternefeld, W. (). ‘Wide scope in situ’, in T. Hanneforth and G. Fanselow (eds), Language and Logos. Studies in Theoretical and Computational Linguistics, Akademie-Verlag, Berlin, –. Stevenson, Rosemary J., Crawley, Rosalind A., and Kleinman, David (). ‘Thematic roles, focus and the representation of events’, Language and Cognitive Processes Vol. . doi: ./. Stewart, A., Holler, J., and Kidd, E. (). ‘Shallow processing of ambiguous pronouns; evidence for delay’, The Quarterly Journal of Experimental Psychology : –. Stewart, A., Pickering, M., and Sanford, A. (). ‘The role of implicit causality in language comprehension: focus versus integration accounts’, Journal of Memory and Language, , –. Stockwell, Robert P., Paul Schachter, and Barbara Hall Partee (). The Major Syntactic Structures of English. New York: Holt, Rinehart, and Winston. Stoia, L., Byron, D. K., Shockley, D. M., and Fosler-Lussier, E. (). ‘Noun phrase generation for situated dialogs’, in Proceedings of the th International Natural Language Generation Conference (INLG), –. Stone, M. (). ‘On identifying sets’, in Proceedings of the st International Conference on Natural Language Generation (INLG), –, Mitzpe Ramon. Stone, M. (). ‘Communicative intentions and conversational processes in human–human and human–computer dialogue’, in J. C. Trueswell and M. K. Tanenhaus (eds), Approaches to Studying World-Situated Language Use. Cambridge, MA: MIT Press, –. Stone, M. and Webber, B. (). ‘Textual economy through close coupling of syntax and semantics’, in Proceedings of the th International Workshop on Natural Language Generation (INLG), –, Niagara-on-the-Lake, Ontario. Stone, M., Doran, C., Webber, B., Bleam, T., and Palmer, M. (). ‘Microplanning with communicative intentions: The SPUD system’, Computational Intelligence (): –. Strawson, P. (). ‘On referring’, Mind : –. Strawson, P. (). Introduction to Logical Theory. London: Methuen. Streb, J., Hennighausen, E., and Rösler, F. (). ‘Different anaphoric expressions are investigated by event-related brain potentials’, Journal of Psycholinguistic Research (): –. Streb, J., Rösler, F., and Hennighausen, E. (). ‘Event-related responses to pronoun and proper name anaphors in parallel and nonparallel discourse structures’, Brain and Language : –. Strube, M. and Hahn, U. (). ‘Functional centering’, in Proceedings of the th Annual Meeting of the Association for Computational Linguistics, Santa Cruz, CA, –. Strube, M. and Hahn, U. (). ‘Functional centering: grounding referential coherence in information structure’, Computational Linguistics (): –. Strube, Michael and Wolters, Maria (). ‘A probabilistic genre-independent model of pronominalization’, Proceedings of the st Conference of the North American Chapter of the Association for Computational Linguistics, –. Sturt, P. (). ‘The time-course of the application of binding constraints in reference resolution’, Journal of Memory and Language, , –. Sturt, P., Sanford, A. J., Stewart, A., and Dawydiak, E. (). ‘Linguistic focus and goodenough representations: an application of the change-detection paradigm’, Psychonomic Bulletin and Review : –.

OUP CORRECTED PROOF – FINAL, 21/1/2019, SPi



REFERENCES

Sudo, Yasutada (). On the Semantics of Phi Features on Pronouns. PhD dissertation, MIT Linguistics. Sun, R. (). The Cambridge Handbook of Computational Psychology. Cambridge University Press, Cambridge. Szabo, Z. G. (). ‘Specific, yet opaque’, in M. Aloni, H. Bastiaanse, T. de Jager, and K. Schulz (eds), Proceedings of the th Amsterdam Colloquium, Springer, Berlin: –. Tanenhaus, Michael K., Craig G. Chambers and Joy E. Hanna (). ‘Referential domains in spoken language comprehension: Using eye movements to bridge the product and action traditions’, in J.M. Henderson and F. Ferreira (eds), The Interface of Language, Vision, and Action: Eye Movements and the Visual World. New York: Psychology Press, –. Tanenhaus, M. K., Spivey-Knowlton, M., Eberhard, K. M., and Sedivy, J. C. (). ‘Integration of visual and linguistic information in spoken language comprehension’, Science : –. Taylor, Kenneth A. (). ‘Sex, breakfast, and descriptus interruptus’, Synthese : –. Tellex, S., Kollar, T., Dickerson, S., Walter, M. R., Banerjee, A. G., Teller, S., et al. (). ‘Approaching the symbol grounding problem with probabilistic graphical models’, AI Magazine (): –. Tellex, S., Thaker, P., Deits, R., Simeonov, D., Kollar, T., and Roy, N. (). Toward a Probabilistic Approach to Acquiring Information from Human Partners Using Language (Tech. Rep.). MIT. Theakston, A. L., Lieven, E. V., Pine, J. M., and Rowland, C. F. (). ‘The role of performance limitations in the acquisition of verb-argument structure: an alternative account’, Journal of Child Language (): –. Theune, M., Koolen, R., Krahmer, E., and Wubben, S. (). ‘Does size matter: how much data is required to train a REG algorithm?’, in Proceedings of the th Annual Meeting of the Association for Computational Linguistics: Human Language Technologies (ACL-HLT), –, Portland, Oregon. Thomason, Richmond H. (). ‘Accommodation, meaning, and implicature: interdisciplinary foundations for pragmatics’, in P. Cohen, J. Morgan, and M. Pollack (eds), Intentions in Communication, Cambridge, MA: MIT Press, –. Thomasson, A. (). Fiction and Metaphysics. Cambridge: Cambridge University Press. Thomasson, A. (). ‘Speaking of fictional characters’, Dialectica : –. Tomasello, Michael (). ‘Reference: intending that others jointly attend’, Pragmatics & Cognition (/): –. Tomlin, Russell S. (). ‘Mapping conceptual representations into linguistic representations: the role of attention in grammar’, in J. Nuyts and E. Pederson (eds), Language and Conceptualization, New York: Cambridge University Press, –. Tuomela, Raimo (). The Philosophy of Sociality. Oxford: Oxford University Press. Turing, A. (). ‘Computing machinery and intelligence’, Mind (): –. Turner, R., Sripada, S., and Reiter, E. (). ‘Generating approximate geographic descriptions’, in Proceedings of the th European Workshop on Natural Language Generation (ENLG), –, Athens, Greece. Turner, R., Sripada, S., Reiter, E., and Davy, I. P. (). ‘Using spatial reference frames to generate grounded textual summaries of georeferenced data’, in Proceedings of the th International Natural Language Generation Conference (INLG), –. Ueno, M. and Garnsey, S. (). ‘An ERP study of the processing of subject and object relative clauses in Japanese’, Language and Cognitive Processes (): –.

OUP CORRECTED PROOF – FINAL, 21/1/2019, SPi

REFERENCES



Ueno, M. and Kehler, A. (). ‘The interpretation of null and overt pronouns in Japanese: grammatical and pragmatic factors’, in S. Ohlsson and R. Catrambone (eds), Proceedings of the nd Annual Conference of the Cognitive Science Society. Cognitive Science Society, –. Ueno, M. and Kehler, A. (). ‘Grammatical and pragmatic factors in the interpretation of Japanese null and overt pronouns’, Linguistics (): –. Ueno, M. and Kluender, R. (). ‘Event-related brain indices of Japanese scrambling’, Brain and Language : –. Ueno, M. and Kluender, R. (). ‘On the processing of Japanese wh-questions’, Brain Research : –. Valian, V. (). ‘Null subjects: a problem for parameter-setting models of language acquisition’, Cognition (): –. Valian, V. (). ‘Syntactic subjects in the early speech of American and Italian children’, Cognition (): –. Van Berkum, J. J. A. (). ‘Sentence comprehension in a wider discourse: can we use ERPs to keep track of things?’, in M. Carreiras, Jr. and C. Clifton (eds), The Online Study of Sentence Comprehension: Eyetracking, ERPs and Beyond, New York: Psychology Press. Van Berkum J. J. A. (). ‘The neuropragmatics of “simple” utterance comprehension: an ERP review’, in U Sauerland and K Yatsushiro (eds), Semantics and Pragmatics: From Experiment to Theory, –. Basingstoke, UK: Palgrave Macmillan, –. Van Berkum, J. J. A. (). ‘The electrophysiology of discourse and conversation’, in M. J. Spivey, K. McRae, and M. F. Joanisse (eds), The Cambridge Handbook of Psycholinguistics, New York: Cambridge University Press, –. Van Berkum, J. J. A., Brown, C. M., and Hagoort, P. (). ‘Early referential context effects in sentence processing: evidence from event-related brain potentials’, Journal of Memory and Language : –. Van Berkum, J., Brown, C., Hagoort, P., and Zwisterlood, P. (). ‘Event-related brain potentials reflect discourse-referential ambiguity in spoken language comprehension’, Psychophysiology : –. Van Berkum, J. J. A., Koornneef, A. W., Otten, M., and Nieuwland, M. S. (). ‘Establishing reference in language comprehension: an electrophysiological perspective’, Brain Research : –. Van Berkum, J. J. A., Zwitserlood, P., Bastiaansen, M. C. M., Brown, C. M., and Hagoort, P. (). ‘So who’s “he” anyway? Differential ERP and ERSP effects of referential success, ambiguity and failure during spoken language comprehension’, Annual meeting of the Cognitive Neuroscience Society, San Francisco, – April. van Deemter, K. (). ‘Generating referring expressions: Boolean extensions of the Incremental Algorithm’, Computational Linguistics (): –. van Deemter, K. (). ‘Generating referring expressions that involve gradable properties’, Computational Linguistics (): –. van Deemter, K. (). Not Exactly: In Praise of Vagueness. Oxford, UK: Oxford University Press. van Deemter, K. (). Computational Models of Referring: A Study in Cognitive Science. Cambridge, Mass.: MIT Press. van Deemter, K. and Krahmer, E. (). ‘Graphs and Booleans: on the generation of referring expressions’, in H. Bunt and R. Muskens (eds), Computing Meaning [Volume ], Studies in Linguistics and Philosophy, Springer Publishers, –.

OUP CORRECTED PROOF – FINAL, 21/1/2019, SPi



REFERENCES

van Deemter, K. and Reiter, E. (). ‘Lying and computational linguistics’, in J. Meibauer (ed.), The Oxford Handbook of Lying, Oxford University Press. van Deemter, K., Gatt, A., van der Sluis, I., and Power, R. (). ‘Generation of referring expressions: assessing the Incremental Algorithm’, Cognitive Science to appear. van Deemter, K., Gatt, A., van der Sluis, I., and Power, R. (). ‘Generation of referring expressions: Assessing the incremental algorithm’, Cognitive Science (): –. van Deemter, K., van der Sluis, I., and Gatt, A. (). ‘Building a semantically transparent corpus for the generation of referring expressions’, in Proceedings of the th International Conference on Natural Language Generation (INLG), –, Sydney, Australia. van der Sluis, I., and Krahmer, E. (). ‘Generating multimodal referring expressions’, Discourse Processes (): –. Van der Wege, Mija M. (). ‘Lexical entrainment and lexical differentiation in reference phrase choice’, Journal of Memory and Language (): –. doi: ./j. jml.... van Geenhoven, Veerle (). Semantic Incorporation and Indefinite Descriptions. Semantic and Syntactic Aspects of Noun Incorporation in West Greenlandic. Standford, CA: CSLI Publications. van Geenhoven, Veerle and McNally, Louise (). ‘On the property analysis on opaque complements’, Lingua : –. van Gompel, R. and Liversedge, S. (). ‘The influence of morphological information on cataphoric pronoun assignment’, Journal of Experimental Psychology: Learning, Memory, and Cognition : –. van Gompel, R., Gatt, A., Krahmer, E., and van Deemter, K. (). ‘Overspecification in reference: modelling size contrast effects’, in Proceedings of the Architectures and Mechanisms for Language Processing (AMLaP) Conference [abstract], Edinburgh. van Hentenryck, P. (). Constraint Satisfaction in Logic Programming. Cambridge, MA: The MIT Press. van Hout, A., Harrigan, K., and de Villiers, J. (). ‘Comprehension and Production of Definite and Indefinite Noun Phrases in English Preschoolers’, Paper presented at the rd Conference on Generative Approaches to Language Acquisition North America (GALANA ), Somerville, MA. Van Inwagen, P. (). ‘Fiction and metaphysics’, Philosophy and Literature : –. Van Inwagen, P. (). ‘Creatures of fiction’, in Ontology, Identity, and Modality, Cambridge: Cambridge University Press, –. Van Nice, Kathy Y. and Dietrich, Rainer (a). ‘Animacy effects in language production: from mental model to formulator’, in H. Härtl and H. Tappe (eds), Mediating Between Concepts and Grammar, Berlin/New York: Mouton de Gruyter, –. Van Nice, Kathy Y. and Dietrich, Rainer (b). ‘Task sensitivity of animacy effects: evidence from German picture descriptions’, Linguistics (): –. doi: ./ ling... Van Rij, Jacolien, Van Rijn, Hedderik, and Hendriks, Petra (). ‘How WM load influences linguistic processing in adults: a computational model of pronoun interpretation in discourse’, Topics in Cognitive Science : –. doi: ./tops.. van Rijsbergen, C. (). Information Retrieval [nd ed.]. London: Butterworths. van Rooij, Robert (). ‘Exhaustivity in dynamic semantics: referential and descriptive pronouns’, Linguistics and Philosophy : –. Van Vliet, Sarah (). Proper Nouns and Pronouns. The Production of Referential Expressions in Narrative Discourse. Utrecht: LOT.

OUP CORRECTED PROOF – FINAL, 21/1/2019, SPi

REFERENCES



Venditti, Jennifer, Matt Stone, Preetham Nanda, and Paul Tepper (). ‘Discourse constraints on the interpretation of nuclear-accented pronouns’, Proceedings of the  International Conference on Speech Prosody. Aix-en-Provence, France. Vendler, Zeno (). ‘Verbs and times’, The Philosophical Review : –. Veneziano, E. (). ‘The emergence of noun and verb categories in the acquisition of French’, Psychology of Language and Communication (): –. Veneziano, E. and Sinclair, H. (). ‘Functional changes in early child language: the appearance of references to the past and of explanations’, Journal of Child Language : –. Veneziano, E. and Sinclair, H. (). ‘The changing status of “filler syllables” on the way to grammatical morphemes’, Journal of Child Language : –. Viethen, J. and Dale, R. (). ‘Algorithms for generating referring expressions: do they do what people do?’, in Proceedings of the th International Conference on Natural Language Generation (INLG), –, Sydney, Australia. Viethen, J. and Dale, R. (). ‘Evaluation in natural language generation: lessons from referring expression generation’, Traitement Automatique des Langues : –. Viethen, J., and Dale, R. (). ‘The use of spatial relations in referring expressions’, in Proceedings of the th International Natural Language Generation Conference (INLG), –. Viethen, J. and Dale, R. (). ‘Speaker-dependent variation in content selection for referring expression generation’, in Proceedings of the th Australasian Language Technology Workshop, pages –, Melbourne. Viethen, J., Dale, R., and Guhe, M. (). ‘Generating subsequent reference in shared visual scenes: computation versus re-use’, in Proceedings of the Conference on Empirical Methods in Natural Language Processing (EMNLP), –, Edinburgh, Scotland, UK. Viethen, J., Dale, R., Krahmer, E., Theune, M., and Touset, P. (). ‘Controlling redundancy in referring expressions’, in Proceedings of the th Language Resources and Evaluation Conference (LREC). Marrakech, Morocco. Viethen, J., Zwarts, S., Dale, R., and Guhe, M. (). ‘Dialogue reference in a visual domain’, in Proceedings of the th Language Resources and Evaluation Conference (LREC). Valetta, Malta. Vissers, C., Chwilla, D., and Kolk, H. (). ‘Monitoring in language perception: the effect of misspellings of words in highly constraining sentences’, Brain Research (): –. Vlach, F. (). ‘Now’ and ‘Then’: A Formal Study in the Logic of Tense Anaphora. PhD thesis, UCLA. Vogels, Jorrig (). Referential choices in language production: The role of accessibility. TiCC Ph.D. Series, Vol. . Vogels, Jorrig, Krahmer, Emiel, and Maes, Alfons (a). ‘Who is where referred to how, and why? The influence of visual saliency on referent accessibility in spoken language production’, Language and Cognitive Processes (): –. doi: ./... Vogels, Jorrig, Krahmer, Emiel, and Maes, Alfons (b). ‘When a stone tries to climb up a slope: the interplay between lexical and perceptual animacy in referential choices’, Frontiers in Psychology : . doi: ./fpsyg... Vogels, J., Krahmer, E. J., and Maes, A. (c). ‘Effect of cognitive load on the choice of referential form’, Preceedings Workshop PRE-CogSci , Berlin, –. Vogels, Jorrig, Krahmer, Emiel, and Maes, Alfons (). ‘How cognitive load influences speakers’ choice of referring expression’, Cognitive Science (): –. doi: ./cogs.. Vogels, Jorrig, Maes, Alfons, and Krahmer, Emiel (). ‘Choosing referring expressions in Belgian and Netherlandic Dutch: effects of animacy’, Lingua : –. von Fintel, Kai (). Restrictions on Quantifier Domains. PhD dissertation, UMass/ Amherst.

OUP CORRECTED PROOF – FINAL, 21/1/2019, SPi



REFERENCES

von Fintel, K. and Heim, I. (). ‘Intensional semantics lecture notes’, Notes for class taught at MIT. von Heusinger, Klaus (). ‘Salience and definiteness’, The Prague Bulletin of Mathematical Linguistics : –. von Heusinger, Klaus (). ‘Specificity and definiteness in sentence and discourse structure’, Journal of Semantics : –. von Heusinger, Klaus (). ‘Specificity’, in K. von Heusinger, C. Maienborn, and P. Portner (eds), Semantics. An International Handbook of Natural Language Meaning, vol. . Berlin: Mouton de Gruyter, –. von Heusinger, Klaus, and Chiriacescu, Sofiana (). ‘Pe-marked definite NPs in Romanian and discourse prominence‘, in N. Pomino, and E. Stark (eds), Proceedings of the V Nereus International Workshop “Mismatches in Romance”. Arbeitspapier . Fachbereich Sprachwissenschaft. Konstanz: Universität Konstanz, –. von Stechow, A. (). ‘Comparing semantic theories of comparison’, Journal of Semantics (–): –. von Stechow, Arnim (). ‘Some remarks on choice functions and LF-movement’, in K. von Heusinger and U. Egli (eds), Reference and Anaphoric Relation. Dordrecht: Kluwer, –. von Wright, G. H. (). An Essay in Modal Logic. North-Holland, Amsterdam. Vonk, Wietske, Hustinx, Lettica G. M. M., and Simons, Wim H. G. (). ‘The use of referential expressions in structuring discourse’, Language and Cognitive Processes (–): –. doi: ./. Wagers, M., Borja, M. F., and Chung S. (). ‘The real-time comprehension of WH-dependencies in a WH-agreement language’, Language (): –. Walker, M., Stent, A., Mairesse, F., and Prasad, R. (). ‘Individual and domain adaptation in sentence planning for dialogue’, Journal of Artificial Intelligence Research : –. Walton, K. (). Mimesis as Make-Believe. Cambridge: Harvard University Press. Walton, K. (). ‘Restricted quantification, negative existentials, and fiction’, Dialectica, :–. Ward, Gregory (). ‘Brave new would’, CLS : –. Ward, Gregory and Betty J. Birner (). ‘Definiteness and the English existential’, Language (): –. Ward, Gregory, Richard Sproat, and Gail McKoon (). ‘A pragmatic analysis of so-called anaphoric islands’, Language : –. Ward, P. and Sturt, P. (). ‘Linguistic focus and memory: An eye-movement study’, Memory and Cognition : –. Warden, D. A. (). ‘The influence of context on children’s use of identifying expressions and references’, British Journal of Psychology (): –. Watson, Duane G., Arnold, Jennifer E., and Tanenhaus, Michael K. (). ‘Tic Tac Toe: effects of predictability and importance on acoustic prominence in language production’, Cognition (): –. doi: ./j.cognition.... Webber, Bonnie L. (). A Formal Approach to Discourse Anaphora. Garland Publishing. Webber, Bonnie Lynn (). ‘So what can we talk about now’, in Michael Brady and Robert C. Berwick (eds), Computational Models of Discourse, Cambridge, MA: MIT Press, –. Webber, Bonnie L. (). ‘So what can we talk about now?’, in B. J. Grosz, K. S. Jones and B. L. Webber (eds), Readings in Natural Language Processing. Morgan Kaufmann. –. Webber, Bonnie Lynn (). ‘Discourse deixis and discourse processing’, Technical Report. Philadelphia: University of Pennsylvania Department of Computer and Information Science.

OUP CORRECTED PROOF – FINAL, 21/1/2019, SPi

REFERENCES



Webber, Bonnie, Matthew Stone, Aravind Joshi, and Alistair Knott (). ‘Anaphora and discourse semantics’, Computational Linguistics (): –. Wechsler, Stephen (). ‘What ‘you’ and ‘I’ mean to each other: person indexicals, selfascription, and theory of mind’, Language : –. Wettstein, Howard (). ‘Demonstrative reference and definite descriptions’, Philosophical Studies : –. Wettstein, Howard (). ‘How to bridge the gap between meaning and reference’, Synthese : –. Wettstein, Howard (). The Magic Prism: An Essay in the Philosophy of Language. New York: Oxford University Press. Wexler, K. (). ‘Very early parameter setting and the unique checking constraint: a new explanation of the optional infinitive stage’, Lingua (): –. White, M., Clark, R., and Moore, J. (). ‘Generating tailored, comparative descriptions with contextually appropriate intonation’, Computational Linguistics : –. Wilder, C. (). ‘Phrasal movement in LF: de re readings, VP-ellipsis, and binding’, Proceedings of NELS : –. Williams, T. and Scheutz, M. (a). ‘A domain-independent model of open-world reference resolution’, in Proceedings of the th Meeting of the Cognitive Science Society, –. Williams, T. and Scheutz, M. (b). ‘POWER: A domain-independent algorithm for probabilistic, open-world entity resolution’, in IEEE/RSJ International Conference on Intelligent Robots and Systems, –. Williams, T. and Scheutz, M. (). ‘A framework for resolving open-world referential expressions in distributed heterogeneous knowledge bases’, in Proceedings of the Thirtieth AAAI Conference on Artificial Intelligence, –. Williams, T., Acharya, S., Schreitter, S., and Scheutz, M. (). ‘Situated open-world reference resolution for human–robot dialogue’, in Proceedings of the th ACM/IEEE Conference on Human-Robot Interaction, –. Wilson, Deidre and Dan Sperber (). ‘Relevance Theory’, in Laurence R. Horn and Gregory Ward (eds), The Handbook of Pragmatics. Oxford: Blackwell. –. Wilson. F. (). Processing at the syntax–discourse interface in second language acquisition. PhD dissertation, University of Edinburgh. Winograd, Terry (). ‘Procedures as a Representation for Data in a Computer Program for Understanding Natural Language’ (Tech. Rep.). DTIC Document. Winograd, Terry (). Understanding Natural Language. New York: Academic Press. Winter, Yoad (). ‘Choice functions and the scopal semantics of indefinites’, Linguistics and Philosophy : –. Wittek, A. and Tomasello, M. (). ‘Young children’s sensitivity to listener knowledge and perceptual context in choosing referring expressions’, Applied Psycholinguistics (): –. Wittgenstein, L. (). Tractatus Logico-Philosophicus. [Translated by C. K. Ogden] London: Routledge. Wittgenstein, L. (). Philosophical Investigations. [Translated by G. E. M. Anscombe] Oxford: Basil Blackwell. Woisetschlaeger, Erich (). ‘On the question of definiteness in “an old man’s book” ’, Linguistic Inquiry : –. Wolf, F., Gibson, E., and Desmet, T. (). ‘Discourse coherence and pronoun resolution’, Language and Cognitive Processes (): –. Wolter, Lynsey (). That’s That: The Semantics and Pragmatics of Demonstrative Noun Phrases. PhD Dissertation, University of California, Santa Cruz.

OUP CORRECTED PROOF – FINAL, 21/1/2019, SPi



REFERENCES

Wolter, Lynsey (). ‘Demonstratives in philosophy and linguistics’, Philosophy Compass (): –. Woodbury, R. (). Behavioral and Neural Correlates of Deep and Surface Anaphora. Doctoral Dissertation, Massachusetts Institute of Technology. Wooding, D., Muggelstone, M., Purdy, K., and Gale, A. (). ‘Eye movements of large populations’, Behavior Research Methods, Instruments, and Computers : –. Wray, Alison (). Formulaic Language and the Lexicon. Cambridge: Cambridge University Press. Wright, Susan and Talmy Givón (). ‘The pragmatics of indefinite reference: quantified text-based studies’, Studies in Language (): –. Yamamoto, Mutsumi (). Animacy and Reference: A Cognitive Approach to Corpus Linguistics. Amsterdam/Philadelphia: John Benjamins. Yang, C. L., Gordon, P. C., Hendrick, R., and Hue, C. W. (). ‘Constraining the comprehension of pronominal expressions in Chinese’, Cognition : –. Yang, C. L., Gordon, P. C., Hendrick, R., and Wu, J. T. (). ‘Comprehension of referring expressions in Chinese’, Language and Cognitive Processes (–): –. Yang, C. L., Gordon, P. C., Hendrick, R., Wu, J. T., and Chou, T. L. (). ‘The processing of coreference for reduced expressions in discourse integration’, Journal of Psycholinguistic Research : –. Yeom, Jae-Il (). A Presuppositional Analysis of Specific Indefinites: Common Grounds as Structured Information States. New York: Garland Publishing. Yule, George (). Referential Communication Tasks. Mahwah, NJ: Lawrence Erlbaum Associates. Zalta, E. (). Abstract Objects: An Introduction to Axiomatic Metaphysics. Dordrecht: Reidel. Zalta, E. (). Intensional Logic and the Metaphysics of Intentionality. Cambridge: MIT Press. Zeevat, Henk (). ‘A compositional approach to Discourse Representation Theory’, Linguistics and Philosophy : –. Zehler, A. M. and Brewer, W. F. (). ‘Sequence and principles in article system use: an examination of “a”, “the”, and “null” acquisition’, Child Development (): –. Zender, H., Kruijff, G.-J. M., and Kruijff-Korbayová, I. (). ‘Situated resolution and generation of spatial referring expressions for robotic assistants’. in Proceedings of the st International Joint Conference on Artifical intelligence, –. Zhang, Y. and Zhang, J. (). ‘Brain responses to agreement violations of Chinese grammatical aspect’, NeuroReport (): –. Zifonun, Gisela, Ludger Hoffman, and Bruno Strecker (). Grammatik der Deutschen Sprache. Berlin: De Gruyter. Zimmer, H. D. and Engelkamp, J. (). ‘The given-new structure of cleft sentences and their influence on picture viewing’, Psychological Research : –. Zimmermann, T. E. (). ‘On the proper treatment of opacity in certain verbs’, Natural Language Semantics : –.

OUP CORRECTED PROOF – FINAL, 21/1/2019, SPi

I .......................

a/an vs. the –,  Abbott, B. , n, n,  Abbott, B. and L. R. Horn , – Abelard, P.  accessibility effect on choice of referring expression –, – local vs. global  and memory  of mental representations – studies on effects – and word order – Accessibility Hierarchy (Ariel) – accommodation –,  ACT (Set of Activated Entities) –, – action sequences n activated cognitive status –, ,  activated entities , ,  Acton, E. and C. Potts  acts of reference see speech acts of reference Adams, F. and R. Stecker n Adequacy judgments  adverbial quantifiers, de re/de dicto distinction – agentivity  and subjecthood – aiming vs. having in mind ,  Airport scenario – alignment, lexical  Allen, S. E. M. et al. – Almog, J. n Almor, A.  Alonso-Ovalle, L. et al.  ambiguity negation  of pronouns  Anand, P. et al. 

anaphora – and antecedents ,  bridging  vs. cataphora – cross-linguistic variation – deep vs. surface – definite descriptions – dependencies  donkey –, – and familiarity presuppositions – in formal semantics  neuroscience studies – null vs. overt – psycholinguistic perspective – syntactic and referential – anaphoric demonstratives – anaphoric dependencies, recent studies – anaphoric reference, neuroscience studies – anaphoric relations – Anderson, J. C. – animacy, salience  animacy effect – Antecedent Contained Deletion n antecedents and anaphora ,  anterior negativities –,  Appelt, D. – Appelt, D. and A. Kronfeld –, , ,  approximative reference – Aquinas, T. – Areces, C. et al. – argument expression – argument positions, numbers of  Ariel, M. –,  Aristotle – Arnold, J. E. , , 

OUP CORRECTED PROOF – FINAL, 21/1/2019, SPi





Arnold, J. E. and Z. Griffin –, –, ,  Arnold, J. E., S. Brown-Schmidt, and J. C. Trueswell  Asher, N. and A. Lascarides ,  Asher, N. et al.  assertions  assumptions in definiteness – Atlas, J. D. and S. C. Levinson  attentional resources, limited – attitude verbs, de re/de dicto distinction – attitudes, referential intentions as – attributive use vs. referential use –, , n, – audience, internal (talking to oneself )  audience-directed referential intentions , n Austin, J. L. n,  Bach, K. –, , , n, n, n Baget, J.-F. and M.-L. Mugnier  Baggio, G. et al.  Baker, C. L. n Bamberg, M. ,  Bar-Hillel, Y. ,  bare nominals – Barker, C. n,  Barkley, C., R. Kluender, and M. Kutas , , , ,  Barwise, J. and R. Cooper –, – basic joint activities  Bäuerle, R. n,  Bayesian model ,  Beaumont, R. C. – Belz, A. and A. Gatt ,  between-structure processes – Bianchi, A.  Binding Theory , n Birch, S. L., J. E. Albrecht, and J. L. Myers  Birner, B. J.  Bishop corpus  Bittner, D.  BLEU metric ,  Bock, J. K. and R. K. Warren  Boër, S. E. and W. G. Lycan  Bolinger, D. 

Boolean descriptions – Borthen, K. –, – Bosch, P. and C. Umbach – Bouma, G. and H. Hopp  bound variables – Bowdle, B. F. and G. Ward – branch and bound algorithm – Branigan, H. P. et al.  Bratman, M. ,  Braun, D. , – Brennan, S. E. and H. H. Clark  bridging anaphora  bridging inferences  Brown, G. n, n, – Brown-Schmidt, S. and M. K. Tanenhaus  Bulu (Bantu language)  Büring, D.  c-relation , n Cable, S.  Callahan, S. M. –, ,  Camera Adviser scenario  Cameron-Faulkner, T., E. V. M. Lieven, and M. Tomasello  Campbell, A. L. et al.  Carlson, G. N.  Carminati, M. N. –, n Carnap/Agnew case (Kaplan) , n,  Cartwright, R. n Castañeda, H.  Catalan, Position of Antecedent Hypothesis (PAH)  cataphora vs. anaphora – cataphoric relationships – Centering Theory (Givón) , , –, , ,  centred worlds  Chafe, W. L. –,  Chai, J. et al. , ,  chess analogy – children choice of referring expressions – private languages n referential communication tasks  theory of mind (ToM) n working memory task 

OUP CORRECTED PROOF – FINAL, 21/1/2019, SPi

 Chinese bare nominals  cognitive statuses , –, , , , , n, , n,  null vs. overt pronouns – personal pronouns  Chiriacescu, S.  Chiriacescu, S. and K. von Heusinger  choice function approaches , – choice negation  choice of referring expression – accessibility – dissociation from choice of referent – factors influencing – non-linguistic factors , – Chomsky, N. ,  Christianson, K. and F. Ferreira  Christophersen, P.  Chung, I., O. Propp, M. R. Walter, and T. M. Howard  Chung, S. and W. Ladusaw  Clark, H. H. ,  Clark, H. H. and A. Bangerter  Clark, H. H. and T. Henetz  Clark, H. H. and C. Marshall ,  Clark, H. H. and D. Wilkes-Gibbs  classic model of referring – clefting  clitics, in first language acquisition –,  closed world approaches vs. open-world approaches  co-indexation of anaphora  co-reference resolution  Coco, M. I. and F. Keller  Coconut corpus –,   experiment  code model of communication –, – cognition – cognitive fix – cognitive load, effect on pronoun production  cognitive states  cognitive statuses – correlation with linguistic forms –,  in the familiarity scale –



in the Givenness Hierarchy, universality –, – Cohen, P. R. and H. J. Levesque  coherence – coherence relations – collective plurals  Colonna, S., S. Schimke, and B. Hemforth  Common Ground  communication in joint actions  referential – communication intentions n, ,  communication plans –, n communication tasks, joint reference , –, – comparatives  complex embedding environments – complex referring expressions – compositionality  computational modeling task  computational models of referring – Concept of Non-Shared Assumptions  conceptual accessibility  Conceptual Graphs  conceptualizations/mental representations accessibility – reference to – Conjunctive Normal Form (CNF)  Constraint satisfaction – Content Determination – context of utterances –,  Common Ground  descriptive incompleteness –, – domain restriction –, – familiarity presuppositions – shifted perspective – two-phase models – context-dependence –,  Context Principle (Frege) – context sensitivity , – contextually bound choice function approach to specificity (Kratzer) – contrastive focus ,  controlling intentions –, n conversation, intentions in  Conversational Implicature (Grice) n, , n, 

OUP CORRECTED PROOF – FINAL, 21/1/2019, SPi





conversational intentions  coordinating joint activities  Coordination Account (King) –, , – corpora,  evaluation – Cowan, N. – Cowart, W. and H. Cairns  Cowles, H. W.  Cowles, H. W. et al.  Crawley, R. J. and R. J. Stevenson  Cresswell, M. J. ,  cross-linguistic variation, anaphoric forms – Cumming, S. , – d-linking (discourse linking)  D-type analysis – Dahl, D. A.  Dale, R. and N. Haddock  Dale, R. and E. Reiter , –, , –, ,  see also Reiter, E. and R. Dale De Cat, C.  de Hoop, H.  de la Fuente, I. and B. Hemforth  de re readings constraints – context of utterances –, ,  de re/de dicto distinction , – empirical phenomena – multiple embeddings – non-noun phrases – practical applications – and scope theory –, –, – de se interpretations – de sensu  de Weck, G.  Dead End () ,  Dechaine, R.-M. and M. Tremblay n deep processing – deferential intentions  definite articles n,  negation n definite descriptions, presupposition accommodation – definite NPs – definiteness n ambiguity 

assumptions – definitions – effect  familiarity theories –, – vs. indefiniteness – negation , – partitives – quantificational status  and referentiality – strength – uniqueness – Deichsel, A.  deictic demonstratives – deictic gestures – deixis , ,  in children’s narrative , ,  demonstrative determiners , –, –, –, –, –,  demonstrative pronouns , , , , , –, –,  demonstrative reference  demonstratives anaphoric –,  based on private shared knowledge – bound variable uses – cognitive statuses – deictic – discretionary n in first language acquisition , – generic uses – indefinite-this – inferable – in joint reference  kind-referring uses – in negative existentials  predicative use – quantificational uses – restrictive that – stereotypical uses with proper names – ‘true’  uniqueness  uses –, – derived accessibility vs. inherent accessibility  Description Logic –,  description theory 

OUP CORRECTED PROOF – FINAL, 21/1/2019, SPi

 descriptive conditions  descriptive incompleteness – descriptive theory of reference  Determiner Phrases (DPs) vs. Noun Phrases (NPs) n determiners, first language acquisition –, – deterministic algorithms  dialogue, influence in language acquisition – DIAREF corpus n, –, , – Dice coefficient –,  dies-indefinites  Diesing, M.  differential object marking – direct reference, negative existentials – directing intentions  Direction Giving scenario – Dirty Floor scenario – discourse/spoken language choice of referring expressions – spontaneity  discourse goals, or questions under discussion  discourse intentions –, – see also referential intentions discourse models, local vs. global – discourse prominence theories of specificity  discourse referents  discourse representation structure (DRS) , , , , –,  Discourse Representation Theory (DRT) , , – vs. formal semantics – methodologies – proper names – discourse-relevant constituents – discourse salience ,  discretionary indexicals  and demonstratives n discriminating cognitive fix  Discriminatory Power ,  Disjunctive Normal Form (DNF)  - algorithm  doctrine of acquaintance – domain goals 



domain restriction – domain-independent algorithm  donkey anaphora , –, – Donnellan, K. , , , n, , –, n, –, , –, ,  Doran, R. and G. Ward  downdating – Drawer corpus  dual-task experiment  Duan, M., M. Elsner, and M.-C. de Marneffe  dubbing intentions  Dutch cataphoric relationships – first language acquisition  personal pronouns  presupposition accommodation – strong vs. weak pronouns  Duvallet, F. et al. – dynamic semantic theories  E-type analysis of definite descriptions  East Asian languages – see also Chinese; Japanese; Korean Ebert, C., C. Ebert, and S. Hinterwimmer  egocentricity debate  ELAN (early left anterior negative) effect ,  Elbourne, P. D. , ,  electroencephalography (EEG) ,  electrophysiological studies – methodologies – Ellert, M.  emotional deixis n empty sets  Enç, M. n, n, , , ,  Entity Representation approach to specificity (Kamp)  entity resolution  Epistemic Constraint  epistemic contrasts in specificity – epistemic specificity  Evans, G. , , n, n, n event-related brain potentials (ERPs) –, , –, – Everett, A. 

OUP CORRECTED PROOF – FINAL, 21/1/2019, SPi





exceptional scope theories of specificity – exclusion negation  existential theories of specificity  existentials – Existing Fictional Entities view – expectancy hypothesis  Expectation Constraint , n experimental settings children’s referring expressions –, – influence of dialogue in language acquisition – explanatory failure – expressives  extra-mental entities  extrinsic task performance measures  eye-tracking studies , –, , , –, –, – false belief task n FAM (Set of Familiar Entities) –,  familiar cognitive status , , , ,  familiar entities , ,  familiarity implicature – familiarity presuppositions – familiarity scale – familiarity theories of definiteness –, – familiarity theories of specificity  Farkas, D. , ,  Fasola, J. and M. J. Matarić – feature tokens  Ferreira, V. S. and H. Yoshita  fictional entities – fictional realism  file change semantics ,  Filiaci, F., A. Sorace, and M. Carreiras  filler syllables – finite clauses –, , – Finnish linear word order – personal pronouns  first mention –, , , , –, –, –, –, – first-mention bias n flexible scope theory 

Fluency judgments  fMRI studies  focus – explicit vs. implicit  and topicality – Focus of Attention (FOA) , –, , – focus shifts – focus stack approach  Fodor, J. –, ,  Fodor, J. D. and I. A. Sag , , , –, , , , –,  Foraker, S. and B. McElree  form–function relationship – form-specific approach  formal semantics vs. Discourse Representation Theory (DRT) – neuroscience approach – forms – and cognitive statuses –, –,  formulaic language  Fraurud, K.  Frederickson tapes n free indirect discourse (FID) – Frege, G. , , , n, n Context Principle – French clefting  language acquisition –, –, , –, –, –,  frequency  Fukumura, K. and R. P. G. van Gompel , –,  Fukumura, K., R. P. G. van Gompel, and M. J. Pickering –,  Fukumura, K., R. P. G. van Gompel, T. Harley, and M. J. Pickering  fulfillment conditions    algorithm –, , , , – Fuller, J. and J. K. Gundel n functional intentions  Game-Theoretical perspective  Gappy Proposition view (Braun) – Gardent, C. , –

OUP CORRECTED PROOF – FINAL, 21/1/2019, SPi

 Gardent, C. and K. Striegnitz  Garnham, A.  Garrod, S. C. and A. J. Sanford – Garvey, C. and A. Caramazza – Gatt, A.  Gatt, A. et al. , , ,  Gazdar, G. n gaze , , ,  Geach, P.  gender matching studies –, – general intentions , – Generalization X (Percus) –,  Generalization Y (Percus) n Generalization Z (Keshet) – Generalized Grounding Graph approach – generalized quantifiers  generative grammar n generic uses of demonstratives – Geng, L. and H. Hamilton  genres and activities, influence on dialogue –,  genuine names  Genuine Singular Terms , , , –, , ,  German agentivity  ambiguity in definiteness  anaphoric reference – dies-indefinites  first language acquisition –,  linear word order – personal pronouns  referential anchoring  Gernsbacher, M. A.  Gernsbacher, M. A. and S. Shroyer – gestures, deictic – - algorithm –, – between-structure processes – cognitive structures – memory model – within-structure processes  Gillon, C.  GIVE evaluation challenge ,  given information vs. new information – givenness 



Givenness Hierarchy (GH) –, –, –, – bare nominals – first language acquisition  and Grice’s maxim of quantity – and natural language discourse – referential/attributive distinction – sense of ‘referentiality’ – universality – Givón, T. , , – Gleitman, L. R. et al. ,  Glucksberg, S. et al.  GNOME corpus  Gordon, P. C., B. J. Grosz, and L. A. Gilliom  Görgülü, E.  Gorniak, P. and D. Roy ,  Goudbeek, M. and E. Krahmer  gradability  grammatical form vs. logical form – graph matching algorithm  Graph search – - algorithm  GRE₃D₃ corpus  GREC (Generating Referring Expressions in Context) challenges , ,    algorithm , , –,  Green, G. M. n Greenfield, P. M.  Grice, H. P. n, , , –, , ,  Grosz, B. J. and C. L. Sidner  Grosz, B. et al.  grounding  Gundel, J. K. , n,  Gundel, J. K. and K. Johnson n, , ,  Gundel, J. K., D. Ntelitheos, and M. Kowalsky  Gundel, J. K., N. Hedberg, and R. Zacharski , –, , , , , , –, ,  Gupta, S. and A. Stent  Hagoort, P. et al.  Hajičová, E.  Hamzelou, J. n

OUP CORRECTED PROOF – FINAL, 21/1/2019, SPi





Hankamer, J. and T. Sag – Hartshorne, J. K., R. Nappa, and J. Snedeker  Haugh, M. n having in mind vs. aiming ,  Hawkins, J. A. , n, n, , , ,  Hawthorne, J. and D. Manley n HCRC Map Task Corpus , ,  Hedberg, N. A., E. Görgülü, and M. Mameni , ,  Heeman, P. A. and G. Hirst  Heim, I. n,, , , , , , , , , ,  Heim, I. and A. Kratzer  Hendriks, P., C. Koster, and J. Hoeks  Hickmann, M. ,  Hierarchical Distributed Correspondence Graph approach  hierarchical structures ,  hierarchies of referring expressions – Higginbotham, J.  Hinterwimmer, S. and P. Bosch  Historical Block view (Donnellan) – Hobbs, J. –,  Hopcroft, J.  Horacek, H. , – Horn, L. R. , n, n Horn, L. R. and B. Abbott  Horn Scale  Hovy, E. H.  Hughes, M. E. and S. E. Allen –,  Human–Computer interface  human-likeness criterion – human processing limitations – Hunt, L. III et al.  Hutchins, E. –, – hyperintentionalism – hypothesization , – identifiability n, ,  identification of entities – if-clauses –, , ,  illocutionary acts –, ,  illocutionary intentions ,  iMap corpus  implicit causality (IC) verbs , –, 

implicit partitives  in focus cognitive status –, , , , , , , ,  Incentive Constraint – Incremental Algorithm (IA) –, , , , –, ,  indefinite articles – cognitive statuses – indefinite determiner this –, , , , –, , – indefinite proximal demonstrative forms – indefinites vs. definites – with definite NPs – epistemic contrasts – exceptional scope theories of specificity – expressing specificity – German dies  in opaque contexts – partitive specificity  referential, specific, and non-specific  scopal specificity – specificity – as temporary state  uniqueness-neutral – topical contrasts – indexicality theories of specificity – indexicality theory (Fodor and Sag) – indexicals  automatic and intentional n discretionary n intentional and discretionary  pure or automatic n Inferables –, , , – Information Sharing – breakdown – limitations – information structure – informativeness  inherent accessibility vs. derived accessibility  initiating intentions , n intend that vs. intend to –, – Intensional Functional Application (IFA)  intensional operators –

OUP CORRECTED PROOF – FINAL, 21/1/2019, SPi

 intention assemblies – Intention Generation Principle  intentional indexicals  Intentional Influences model (Jordan) ,  intentions general vs. specific – primary vs. secondary  see also referential intentions interactive settings, referring expressions in – interestingness  intermediate intentions  Intersective Predicate Generalization (IPG)  Inuktitut language, first language acquisition  Ionin, T. , ,  Ionin, T. et al. – Italian cataphora vs. anaphora  null vs. overt pronouns  pro-drop – Jackendoff, R.  Janarthanam, S. and O. Lemon  Japanese bare nominals  cataphoric relationships  cognitive statuses , , –, , , n, , – null vs. overt pronouns  Järvikivi, J., R. P. G. van Gompel, R. Bertram, and J. Hyönä  John Locke Lectures (Kripke)  joint actions –,  referential communication – joint intentions  joint reference – negotiating  referential communication tasks , –, – successful and unsuccessful – Jordan, P. W. ,  Jordan, P. W. and M. Walker ,   (Judicious Overspecification) – Kaiser, E. , –, , , , – Kaiser, E. and M. Do 



Kaiser, E. and J. C. Trueswell –, , , , ,  Kamp, H. , , , , , n Kamp, H. and A. Bende-Farkas ,  Kamp, H. et al.  KAMP program – Kaplan, D. n, n, , , , , nn,, ,  Karimi, H., T. Swaab, and F. Ferreira – Karmiloff-Smith, A. ,  Karttunen, L. , ,  Katz, J. n, n, – Kazanina, N., E. F. Lau, M. Lieberman, M. Yoshida, and C. Phillips  Keenan, E. L. and J. Stavi – Kehler, A. – Kehler, A. and H. Rohde  Kehler, A. et al.  Kelleher, J. and G.-J. Kruijff  Kelleher, J. et al. – Kennington, C. and D. Schlangen  Kertz, L. et al. n,  Keshet, E. n, –, –, ,  Keysar, B. – Khan, I. H. et al.  Kibble, R. and R. Power  kind-referring uses of demonstratives – King, J. C. n, , –,  Coordination Account –, , – King, J. W. and M. Kutas  Kneale ,  knowing and unknowing readers – Knowledge Base (KB) , , ,  Knowledge Representation (KR) frameworks , – limitations –  – Koolen, R. et al.  Kopp, S. et al.  Korean cataphoric relationships – null pronouns  Korta, K. and J. Perry  Krahmer, E.  Krahmer, E. and M. Theune ,  Krahmer, E., S. van Erk, and A. Verleg – Krahmer, E. et al. 

OUP CORRECTED PROOF – FINAL, 21/1/2019, SPi





Kratzer, A. –, ,  Krifka, M. et al.  Kripke, S. , , , , , , –, nn,, –, , –, , , , , – Kronfeld, A. , n, ,  Kruijff, G.-J. M. et al.  Kusumoto, K.  Kutas, M. and S. Hillyard , ,  Kutlák, R. –,  Kwon, N. and P. Sturt  L learners see second language acquisition labeling  Lack of Orientation () ,  Lakoff, R. , , n LAN (left anterior negative) effect , , , , –, – Landman, F.  language acquisition argument expression – determiners –, – first referring expressions – influence of dialogue – paradox of referring expressions – language grounding algorithms  language grounding problem  language production model (Levelt)  late positive complex (LPC)  Lemaignan, S. et al.  Lenat, D.  Levelt, W. J. M.  Levenshtein distance metric , ,  Levinson, S. C. ,  Lewis, D. , , , ,  lexical accessibility  lexical alignment  Lexical Choice – Li, C. N. and S. A. Thompson , n Lieberman, H. et al.  Lillooet language  linear order – linguistic reference ,  vs. speech acts of reference – linguistic salience – Lipman, B. – listeners, referential intentions 

Löbner, S. , , , , –,  locally contingent categorization  locutionary intentions ,  logical form vs. grammatical form – logophoric languages – long term memory (LTM) –,  long-distance relations  long-distance scope shift theory  Love, J. and G. McKoon  Lyon, J. n m-reference n Machery, E. et al.  Machine Learning algorithm ,  MacKay, A. F.  Maclaran, R. –, ,  Mandarin Chinese see Chinese Mann, W. and S. Thompson  Maori language, specificity  map tasks (referential communication tasks) –, n, – MapTask corpus  Maratsos, M. P.  Markov Logic theoretic reference resolution system  MASI (Measuring Agreement on Set-valued Items) –,  Matthews, D. et al. , ,  Matthewson, L.  Matuszek, C., N. Fitzgerald, L. Zettlemoyer, L. Bo, and D. Fox  Maxim of Quantity (Grice)  and Givenness Hierarchy – Maxim of Relevance (Grice)  maximum entropy  May, R. –,  Mayol, L. and R. Clark  McCoy, K. F. and M. Strube  McDonald, J. L. et al.  Mechanical Turk  Meinong, A. –, – Meinongian population explosion , – memory ,  and accessibility  long term (LTM) –,  working –

OUP CORRECTED PROOF – FINAL, 21/1/2019, SPi

 mental representations  accessibility – of coherence relations – reference to – robots  Meyer, F.  Milsark, G. , , – mind-to-mind communication  minimal references  mistaken references , , – modal operators  modal subordination  modals, de re/de dicto distinction  Modified Hierarchy (Chai)  monotonic approach to   Montag, J. L. and M. C. MacDonald – Montague, R. ,  Montague Grammar  Morgan, J. L.  Moroccan Arabic, bare nominals  morphemes as proper names  speaker intentions  most as determiner  motivational potential of intentions  Mount, A.  mythical entities ,  N effect –, , –,  naming n narrative relations vs. result relations  narratives, children –, – narrow scope – natural dialogue, influence in language acquisition – natural language, importance in robotics –,  natural language discourse, and the Givenness Hierarchy – Natural Language Generation (NLG) , n natural language generation studies  Natural Language Processing  naturally occurring dialogue, referring expressions – children’s referring expressions –, , , –



Neale, S. , , n, n, n, , – Nearest-First Search ()  negation ambiguity  definite articles , –, n negative existentials direct reference – elimination of all genuine singular terms – Existing Fictional Entities view – Gappy Proposition view – Historical Block view – No Such Proposition view – problem of – Pure Metalinguistic Description theory (PMT) – theory of descriptions – neuroscience anaphora –, – approach to formal semantics – methodologies – presupposition accommodation – proper names – new information vs. given information – Nieuwland, M. S. –,  Nieuwland, M. S. and J. Van Berkum , ,  Nieuwland, M. S. et al. , , n NIST metric ,  No Such Proposition view (Kripke) – non-linguistic factors effect on choice of referring expressions , – and salience – non-linguistic reference  non-specific de re – Norwegian bare nominals  definiteness – notional readings vs. relational readings  Noun Phrases (NPs) de re/de dicto distinction – vs. Determiner Phrases (DPs) n Nref (referentially induced frontal negativity) effect , –, ,  null cataphors in East Asian languages –

OUP CORRECTED PROOF – FINAL, 21/1/2019, SPi





null forms, in first language acquisition ,  null pronouns vs. overt pronouns – O’Neill, D. K.  object domain vs. spatial domain – object language, intensional variables – Odawa language, word order  Olive Oil scenario  open-world algorithm  open-world approaches vs. closed-world approaches  open world resolution  optimality of referring expressions – Osterhout, L. and P. Holcomb n Overspecification, Judicious () – Overstreet, M. and G. Yule  overt pronouns vs. null pronouns – Öztürk, B.  P effects , – P effects –, , , , n Pablos, L., J. Doetjes, B. Ruijgrok, and L. Cheng – Paraboni, I. and K. van Deemter ,  Paraboni, I. et al.  paradigms of reference – Partee, B. H. , , –,  Partitive Constraint , – partitive contrasts in specificity  partitive specificity  partitives –,  partitivity, relationship with specificity  Passonneau, R. J. ,  pathological utterances – Pechmann, T. ,  Percus, O. , –,  Perfect Recall Percentage (PRP) ,  Perry, J. –, n, n, n, , n, , –, n, n Persian bare nominals – referential objects ,  personal pronouns – perspective-taking – Pesetsky, D.  Peters, S. and D. Westerståhl , 

piano analogy in joint action –, ,  Pica, T. et al. – pilotage joint action task (Hutchins) –, – plan-based approach to   planning theory –, ,  Plato  Poesio, M. and R. Viera , –,  Pointwise Mutual Information (PMI)  Politzer-Ahles, S. et al.  Pollack, M. E. n Portner, P.  Portner, P. and K. Yabushita – Position of Antecedent Hypothesis (PAH)  positioning – Possibility Constraint n possible world pronouns – Power, R.  pragmatic definites ,  pragmatic reference – vs. semantic reference ,  pragmatic set (P-set) ,  Prat-Sala, M. and H. P. Branigan –,  predication  predicative statements n, – predicative use of demonstratives – predictability – Preference Order , ,  presupposition –,  presupposition accommodation, definite descriptions – presupposition failure  primary intentions  Prince, E. –, ,  Prior, A. N.  private languages n private shared knowledge, uses of demonstratives – pro-drop languages n, n, – procedural attachments  procedural semantics view  processing depth – processing limitations – pronoun generation  effect of cognitive load  pronoun interpretation –

OUP CORRECTED PROOF – FINAL, 21/1/2019, SPi

 and coherence relations  in free indirect discourse (FID) – processing depth – pronouns n,  anterior negativities – in negative existentials  uniqueness  proper names ,  cataphoric relationships  familiarity presuppositions  generic uses of demonstratives – neuroscience approach – uniqueness  properties, predication  propositional attitudes – propositional functions n proximal demonstrative forms – psycholinguistic perspective – anaphoric forms – null vs. overt pronouns – pure indexicals  Pure Metalinguistic Description theory (PMT, Katz) – purposive actions ,  Pylkkänen, L. et al. – Q-based implicatures –,  Q-based implicatures , – QR-theory (Quantifier Raising) ,  quantificational operators – quantificational phrases  quantificational status of definiteness  quantificational uses of demonstratives – quantificationally bound definite NPs – quantifiers n Quantity Maxim (Grice)  queer connections n questions, in language acquisition – questions under discussion, or discourse goals  Quine, W. v O. –, , –, ,  recency, effect on choice of referring expression  Reference Failure , , ,  reference form, effect on retrieval operations –



reference resolution –,  cross-linguistic variation – open-world vs. closed-world approaches  in robotics – topicality – referential anaphora – referential anchoring approach to specificity (von Heusinger) –,  referential cognitive status –, , , , ,  referential communication tasks , –, – referential communication, models – Referential Compositionality –, , –, –,  referential contrasts in specificity – referential determiners – referential discourse, model  referential expressions  referential intentions –,  content – determining reference n explanatory failure – form – function – hyperintentionalism – identifying – motivation for theories – standard form  referential presupposition  referential stability  referential theories of specificity  referential uniqueness vs. semantic uniqueness – referential use vs. attributive use –, , n, – referentiality and definiteness – overt markers – theories of specificity – referring expressions (s) – children – choice of – accessibility – dissociation from choice of referent –

OUP CORRECTED PROOF – FINAL, 21/1/2019, SPi





referring expressions (s) (cont.) factors influencing – non-linguistic factors , – complex – functions of – hierarchies – interactive settings – limitations in the types  optimality – understanding by robots – without speech acts of reference  Referring Expressions Generation () , , –, –, – classic  task  Constraint satisfaction – Content Determination – context –,  evaluation – Graph search – Knowledge Representation (KR) , – metrics for – plan-based approach  pre- history – reference to sets – relational descriptions – salience – vague references  variations and extensions –, – widening the scope  Reimer, M. ,  Reiter, E. and R. Dale – see also Dale, R. and E. Reiter relational conditions  relational descriptions – relational readings vs. notional readings  relative clauses – Relevance Maxim (Grice)  Relevance Theory (Sperber and Wilson)  reminder that  Ren, Y., K. van Deemter, and J. Pan ,  repetition, in language acquisition – requests, fulfillment conditions  Restriction vs. Saturation  restrictive that – result relations vs. narrative relations  retrieval operations, effect of reference form –

Richard, M. n Road Gritting scenario  Roberts, C. , , , , ,  robotics importance of natural language – natural language  reference resolution – Romanian, indefinites  Rosch, E.  Rothschild, D.  rouge- and rouge-su measures  Roy, D.  Rozendaal, M. and A. Baker – Runner, J.  Runner, J. and A. Ibarra  Russell, B. , –, –, , –, –, , , , , , – Russian, cognitive statuses , , , , –, , ,  Sæbø, K. J. – Sag, I. n Sag, T. and J. Hankamer  Salazar Orvig, A., H. Marcos, et al. , ,  salience , – discourse ,  linguistic – visual – and word order – salience weights (sws)  salient entities, and starting point of utterances – Salish languages – Salmon, N. n, , –,  Salomo, D. et al. – satisfaction conditions , , – Saturation vs. Restriction  Schaeffer, J. and L. Matthewson ,  Schober, M. n Schumacher, P., L. Roberts, and J. Järvikivi  Schwarz, F. , ,  Schwarzschild, R. ,  scopal contrasts in specificity –,  scope islands –, ,  scope restraint on anaphoric relations 

OUP CORRECTED PROOF – FINAL, 21/1/2019, SPi

 scope taking – scope theory –, – and the de re/de dicto distinction – split intensionality theory –, – Scott, K. J.  Searle, J. R.  Sechelt language, referential determiners – second language acquisition, referential communication tasks  ‘second-order’ referential intentions  secondary intentions  semantic argument – semantic compositionality  semantic contribution of a name  semantic definites ,  Semantic Fields approach – semantic presupposition n semantic reference n vs. pragmatic reference ,  vs. speaker reference – semantic underdetermination  semantic uniqueness vs. referential uniqueness – Serratrice, L.  Set of Activated Entities (ACT) –, – Set of Familiar Entities (FAM) –,  shared cultural narratives – Sharvy, R.  shifted perspective – SHRDLU system – Siddharthan, A. and A. Copestake , – Simchen, O. , n, , n Simplified English ,  singular negative existentials – singular propositions  singular terms  Skarabela, B. et al.  Skolem function – Soames, S. n Song, H. and C. Fisher  Sowa, J.  Spanish bare nominals  cognitive statuses , –, , , –, –,  null vs. overt pronouns 



Position of Antecedent Hypothesis (PAH)  proximal demonstrative determiners  specificity  spatial domain vs. object domain – Speaker Control Principle –, ,  speaker intentions – identifying – see also referential intentions speaker reference – vs. semantic reference – specific indefinites in speech reports (Sæbø) – specific intentions , – specificity n, n, – contrasts – of de re – expressing – links between notions of  theories of – speech acts of expression  speech acts of predication  speech acts of reference – absent from referring expressions  vs. linguistic reference – split intensionality theory –, – spoken language/discourse deictic demonstratives  pathological utterances – spontaneity  vs. written language ,  Squamish language  Stainton, R. n Stalnaker, R. , , , –, n starting point of utterances – Stevenson, R. J. et al.  Stewart, A. et al. – Stoia, L. et al.  Stone, M. and B. Webber ,  Stone, M. et al.  story completion experiments –, , – strategic conversations  Strawson, P. F. –, , , , n, , n, –, ,  Streb, J. et al.  strength –

OUP CORRECTED PROOF – FINAL, 21/1/2019, SPi





string distance metrics –,  strong pronouns vs. weak pronouns  structural ambiguity – sub-sentential intentions  subjecthood –,  subjectivist semantics n success conditions – summarization  supplementives n, n Surface Realization – symbol grounding problem  syntactic anaphora – syntactic positive shifts (SPS) – Szabo, Z. G. – target identification  Tellex, S. et al.  temporal operators  tense, de re/de dicto distinction – than-clauses  the vs. a/an –,  theoretical modeling  theory of descriptions (Russell) – theory of mind (ToM) n, ,  thinking, reference in – this-indefinite , –, , – Thomasson, A. n topic-drop languages – topic shift  topical contrasts – topicality –,  topicality-based approach , – topicalization n topics – traditional formal semantics see formal semantics transactional contingent patterns  transference  true demonstratives  true propositions  truth-conditional referential intentions n truth conditions – in formal semantics –,  negative existentials – truth value, reference as  TUNA corpus –, –

Turkish bare nominals – referential objects ,  Turner, R. et al.  type identifiable cognitive status , , , , , , , , , ,  type-shifting operations  Ueno, M. and A. Kehler  uncertain properties  understanding, of referring expressions by robots – unexpectedness – unique satisfaction proposition  uniquely identifiable cognitive status , , , n, , , , –, , – uniqueness – contradiction to – Universal Grammar  unknowing and knowing readers – Unnamed Company scenario  UserKnows function  utterance interpretation n starting point – vagueness  Van Berkum, J. J. A. – Van Berkum, J. J. A., C. M. Brown and P. Hagoort , –,  Van Berkum, J. J. A., C. M. Brown, P. Hagoort, and P. Zwitserlood ,  Van Berkum, J. J. A., A. W. Koornneef, M. Otten, and M. S. Nieuwland – Van Berkum, J. J. A., P. Zwitserlood, M. C. M. Bastiaansen, C. M. Brown, and P. Hagoort – van Deemter, K. , ,  van Deemter, K. et al.  van der Sluis, I. and E. Krahmer  van Gompel, R. and S. Liversedge – van Hout, A. et al.  Van Nice, K. Y. and R. Dietrich – Van Rij, J., H. Van Rijn, and P. Hendriks ,  Van Rooij, R. n

OUP CORRECTED PROOF – FINAL, 21/1/2019, SPi

 Vendler, Z. n verbs, de re readings  Viethen, J. and R. Dale ,  visual information – visual perception – visual salience – Visual World paradigm – Vogels, J., E. J. Krahmer, and A. Maes , –, –, , ,  von Fintel, K. and I. Heim ,  von Heusinger, K. n, , , , – von Stechow, A. ,  von Wright, G. H.  Walton, K. n Ward, G. and B. J. Birner – weak pronouns vs. strong pronouns  Webber, B. L.  Wechsler, S.  Wettstein, H.  ‘Who is?’ scenario – wide scope 

Wilder, C. n Wilson, F.  Winograd, T. , – within-structure processes  Wittek, A. and M. Tomasello  Wittgenstein, L. n, n Wolter, L. n Woodbury, R.  Wooding, D. et al.  word order and salience –, – and subjecthood – words as tools  working memory , , – and accessibility  Wright, S. and T. Givón ,  written language referential intentions  vs. spoken language ,  Yang, C. L. et al.  Yule, G. ,  Zender, H. et al. –



OUP CORRECTED PROOF – FINAL, 17/1/2019, SPi

OXFORD HANDBOOKS IN LINGUISTICS THE OXFORD HANDBOOK OF AFRICAN AMERICAN LANGUAGE Edited by Sonja Lanehart

THE OXFORD HANDBOOK OF APPLIED LINGUISTICS Second edition Edited by Robert B. Kaplan

THE OXFORD HANDBOOK OF ARABIC LINGUISTICS Edited by Jonathan Owens

THE OXFORD HANDBOOK OF CASE Edited by Andrej Malchukov and Andrew Spencer

THE OXFORD HANDBOOK OF CHINESE LINGUISTICS Edited by William S-Y Wang and Chaofen Sun

THE OXFORD HANDBOOK OF COGNITIVE LINGUISTICS Edited by Dirk Geeraerts and Hubert Cuyckens

THE OXFORD HANDBOOK OF COMPARATIVE SYNTAX Edited by Gugliemo Cinque and Richard S. Kayne

THE OXFORD HANDBOOK OF COMPOSITIONALITY Edited by Markus Werning, Wolfram Hinzen, and Edouard Machery

THE OXFORD HANDBOOK OF COMPOUNDING Edited by Rochelle Lieber and Pavol Štekauer

THE OXFORD HANDBOOK OF COMPUTATIONAL LINGUISTICS Edited by Ruslan Mitkov

THE OXFORD HANDBOOK OF CONSTRUCTION GRAMMAR Edited by Thomas Hoffman and Graeme Trousdale

THE OXFORD HANDBOOK OF CORPUS PHONOLOGY Edited by Jacques Durand, Ulrike Gut, and Gjert Kristoffersen

THE OXFORD HANDBOOK OF DERIVATIONAL MORPHOLOGY Rochelle Lieber and Pavol Štekauer

THE OXFORD HANDBOOK OF DEVELOPMENTAL LINGUISTICS Edited by Jeffrey Lidz, William Snyder, and Joe Pater

THE OXFORD HANDBOOK OF ELLIPSIS Edited by Jeroen van Craenenbroeck and Tanja Temmerman

THE OXFORD HANDBOOK OF ERGATIVITY Edited by Jessica Coon, Diane Massam, and Lisa deMena Travis

THE OXFORD HANDBOOK OF EVIDENTIALITY Edited by Alexandra Y. Aikhenvald

THE OXFORD HANDBOOK OF GRAMMATICALIZATION Edited by Heiko Narrog and Bernd Heine

OUP CORRECTED PROOF – FINAL, 17/1/2019, SPi

THE OXFORD HANDBOOK OF HISTORICAL PHONOLOGY Edited by Patrick Honeybone and Joseph Salmons

THE OXFORD HANDBOOK OF THE HISTORY OF ENGLISH Edited by Terttu Nevalainen and Elizabeth Closs Traugott

THE OXFORD HANDBOOK OF THE HISTORY OF LINGUISTICS Edited by Keith Allan

THE OXFORD HANDBOOK OF INFLECTION Edited by Matthew Baerman

THE OXFORD HANDBOOK OF INFORMATION STRUCTURE Edited by Caroline Féry and Shinichiro Ishihara

THE OXFORD HANDBOOK OF JAPANESE LINGUISTICS Edited by Shigeru Miyagawa and Mamoru Saito

THE OXFORD HANDBOOK OF LABORATORY PHONOLOGY Edited by Abigail C. Cohn, Cécile Fougeron, and Marie Hoffman

THE OXFORD HANDBOOK OF LANGUAGE AND LAW Edited by Peter Tiersma and Lawrence M. Solan

THE OXFORD HANDBOOK OF LANGUAGE AND SOCIETY Edited by Ofelia García, Nelson Flores, and Massimiliano Spotti

THE OXFORD HANDBOOK OF LANGUAGE EVOLUTION Edited by Maggie Tallerman and Kathleen Gibson

THE OXFORD HANDBOOK OF LANGUAGE POLICY AND PLANNING Edited by James W. Tollefson and Miguel Pérez-Milans

THE OXFORD HANDBOOK OF LEXICOGRAPHY Edited by Philip Durkin

THE OXFORD HANDBOOK OF LINGUISTIC ANALYSIS Second edition Edited by Bernd Heine and Heiko Narrog

THE OXFORD HANDBOOK OF LINGUISTIC FIELDWORK Edited by Nicholas Thieberger

THE OXFORD HANDBOOK OF LINGUISTIC INTERFACES Edited by Gillian Ramchand and Charles Reiss

THE OXFORD HANDBOOK OF LINGUISTIC MINIMALISM Edited by Cedric Boeckx

THE OXFORD HANDBOOK OF LINGUISTIC TYPOLOGY Edited by Jae Jung Song

THE OXFORD HANDBOOK OF LYING Edited by Jörg Meibauer

THE OXFORD HANDBOOK OF MODALITY AND MOOD Edited by Jan Nuyts and Johan van der Auwera

OUP CORRECTED PROOF – FINAL, 17/1/2019, SPi

THE OXFORD HANDBOOK OF MORPHOLOGICAL THEORY Edited by Jenny Audring and Francesca Masini

THE OXFORD HANDBOOK OF NAMES AND NAMING Edited by Carole Hough

THE OXFORD HANDBOOK OF PERSIAN LINGUISTICS Edited by Anousha Sedighi and Pouneh Shabani-Jadidi

THE OXFORD HANDBOOK OF POLYSYNTHESIS Edited by Michael Fortescue, Marianne Mithun, and Nicholas Evans

THE OXFORD HANDBOOK OF PRAGMATICS Edited by Yan Huang

THE OXFORD HANDBOOK OF REFERENCE Edited by Jeanette Gundel and Barbara Abbott

THE OXFORD HANDBOOK OF SOCIOLINGUISTICS Second Edition Edited by Robert Bayley, Richard Cameron, and Ceil Lucas

THE OXFORD HANDBOOK OF TABOO WORDS AND LANGUAGE Edited by Keith Allan

THE OXFORD HANDBOOK OF TENSE AND ASPECT Edited by Robert I. Binnick

THE OXFORD HANDBOOK OF TRANSLATION STUDIES Edited by Kirsten Malmkjaer and Kevin Windle

THE OXFORD HANDBOOK OF UNIVERSAL GRAMMAR Edited by Ian Roberts

THE OXFORD HANDBOOK OF THE WORD Edited by John R. Taylor

THE OXFORD HANDBOOK OF WORLD ENGLISHES Edited by Markku Filppula, Juhani Klemola, and Devyani Sharma