117 81 34MB
English Pages 468 [456]
ASPECTUAL INQUIRIES
Studies in Natural Language and Linguistic Theory VOLUME 62 Managing Editors Marcel den Dikken, City University of New York Liliane Haegeman, University of Lille Joan Maling, Brandeis University Editorial Board Guglielmo Cinque, University of Venice Carol Georgopoulos, University of Utah Jane Grimshaw, Rutgers University Michael Kenstowicz, Massachusetts Institute of Technology Hilda Koopman, University of California, Los Angeles Howard Lasnik, University of Connecticut at Storrs Alec Marantz, Massachusetts Institute of Technology John J. McCarthy, University of Massachusetts, Amherst Ian Roberts, University of Cambridge
The titles published in this series are listed at the end of this volume.
ASPECTUAL INQUIRIES Edited by
PAULA KEMPCHINSKY University of Iowa, Iowa City, IA, USA
and ROUMYANA SLABAKOVA University of Iowa, Iowa City, IA, USA
4y Springer
A C.I.P. Catalogue record for this book is available from the Library of Congress.
ISBN 1-4020-3035-5 (PB) ISBN 1-4020-3032-0 (HB) ISBN 1-4020-3033-9 (e-book)
Published by Springer, P.O. Box 17, 3300 AA Dordrecht, The Netherlands. Sold and distributed in North, Central and South America by Springer, 101 Philip Drive, Norwell, MA 02061, U.S.A. In all other countries, sold and distributed by Springer, P.O. Box 322, 3300 AH Dordrecht, The Netherlands.
Printed on acid-free paper
All Rights Reserved © 2005 Springer No part of this work may be reproduced, stored in a retrieval system, or transmitted in any form or by any means, electronic, mechanical, photocopying, microfilming, recording or otherwise, without written permission from the Publisher, with the exception of any material supplied specifically for the purpose of being entered and executed on a computer system, for exclusive use by the purchaser of the work. Printed in the Netherlands.
To our colleague, fellow aspectologist and friend, Alice Davison
TABLE OF CONTENTS
PAULA KEMPCHINSKY and ROUMYANA SLABAKOVA / Issues and Interfaces in the Study of Aspect
1
Part I ASPECT AND THE INTERNAL STRUCTURE OF THE CLAUSE ELIZABETH RITTER and SARA THOMAS ROSEN / Topic or Aspect 21 HAGIT BORER / Some Notes on the Syntax of Quantity LISA DEMENA TRAVIS / Articulated vPS and the Computation of Aspectual Classes RAFFAELLA FOLLI and HEIDI HARLEY / Flavors of v CHRISTINA SCHMITT / Semi-Copulas MA1 TUNGSETH / PP, FP and the TelidAtelic Distinction in Norwegian Motion Constructions Part II ASPECT, TENSE AND DISCOURSE KAREN ZAGONA / The Spatio-temporal Path and Aspectual Composition HAMIDA DEMIRDACHE and MYRIAM URIBE-ETXEBARRIA / Aspect and Temporal Modification CARLOTA S. SMITH / Aspectual Entities and Tense in Discourse PATRICK CAUDAL / Stage Structure And Stage Salience For Event Semantics
239
v iii
TABLE OF CONTENTS
PATRICK CAUDAL and LAURENT ROUSSARIE / Aspectual Viewpoints, Speech Act Functions and Discourse Structure
265
Part III ASPECT AND ACQUISITION
NINA HYAMS / Child Non-Finite Clauses and the Mood-Aspect Connection
293
ANGELIEK VAN HOUT / Imperfect Imperfectives
3 17
MARTYNA KOZLOWSKA-MACGREGOR / L2 Acquisition of Aspectual Distinctions in Polish
345
ASYA PERELTSVAIG / Aspect Lost, Aspect Regained
369
KATHLEEN BARDOVI HARLIG / Tracking the Elusive Imperfect in Adult L2 Acquisition
397
REFERENCES
42 1
INDEX
45 1
PAULA KEMPCHINSKY AND ROUMYANA SLABAKOVA
ISSUES AND INTERFACES IN THE STUDY OF ASPECT
The study of aspect—the internal temporal structure of an event—has a long philosophical tradition and a comparatively short but extremely fruitful history of linguistically-based research. The philosophical study of aspect can be dated to the event typology in Aristotle's Metaphysics; the birth of the linguistic study of aspect is usually dated to Vendler's extremely influential 1957 article on "Verbs and times" (reprinted in Vendler, 1967), which analyzed the distributional properties of different event types. One could argue, therefore, that the study of aspect represents one of the first attempts to examine properties of the syntax-conceptual/intentional interface. We believe that the study of aspect is in fact fundamentally concerned with interfaces, both in the sense of interfaces formally recognized as such in the architecture of the grammar and in the sense of interfaces between subdisciplines of linguistics. The question of the relative contribution of lexical properties of a verb vs. constructionalist representations of lexical aspect (situation aspect, or aktionsart) necessarily entails focusing on the interaction between lexical semantics and syntactic projection. In turn, the interplay between lexical aspect and sentential aspect (viewpoint aspect) in determining whether the event is interpreted as bounded in time forces us to look more closely at the relationship between the lexical layer of the clause and the functional layer, as well as examining the interface between the syntactic representation and the semantic representation of the clause. Many researchers working on the acquisition of the linguistic parameters of aspect have identified this interplay between situation aspect and viewpoint aspect as a fundamental factor in the acquisition process, both of first and of second languages, highlighting the relevance of acquisition research to semantic-syntactic research on aspect. The ways in which aspectual interpretation enters into the discourse, and how discourse factors affect aspectual interpretation, makes the study of aspect a particularly useful tool for elucidating the interface between the sentential level of clausal semantics and the broader, multisentential level of the passage. It was from this perspective that we organized the "Workshop on the Syntax, Semantics and Acquisition of Aspect", held at the University of Iowa in May of 2002 and funded by the National Science Foundation (Grant 0111290). The purpose of the workshop was precisely to bring together scholars in different areas of specialization in linguistics to discuss the syntactic and semantic characterization of aspect from the perspectives of generative syntax, discourse representation theory,
P. Kempchinsky and R. Slabakova (eds.), Aspectual Inquiriess, 1-18. © 2005 Springer. Printed in the Netherlands.
2
P. KEMPCHINSKY AND R. SLABAKOVA
and psycholinguistics. Although researchers within any one of these fields are familiar with the work of others in the same field, there has been less communication across fields. Our goal in organizing this workshop was to purposefully set out to bring these different threads of research together, and thus create a context in which these various researchers could engage in active discussion and as a result further our understanding of this linguistic phenomenon. The papers gathered in this volume represent the results of this workshop. Prior dissemination of initial drafts of the workshop papers on the workshop website, plus a schedule which provided extensive time for discussion at the end of each presentation and a roundtable discussion to wrap up the workshop, encouraged active and fertile interaction among the participants. To further this crossdisciplinary communication, each of the authors received comments on the first draft of the written version of the papers from another workshop participant, from a different subdiscipline. In this introduction we provide a brief review of the issues within the field of aspectology that arose in the workshop presentations, followed by a summary of the main contributions of each paper. We believe that fellow aspectologists-to borrow the term from the Slavic tradition-will find useful and insightful additions here to the growing abundance of linguistic literature on the topic. 1. GENERAL OVERVIEW
1.1. The determination of lexical aspect The aspectual properties of verbal predicates-often designated Aktionsart-have been investigated from various perspectives. The typology of Vender (195711967) as refined by Dowty (1979), the Vendler-Dowty classification, is the point of departure for many subsequent studies. This classification divides predicates into four classes, based on the properties of definiteness (or telicity) and process (or duration): states, activities, accomplishments and achievements. As is well-known, each of the aspectual classes displays characteristic linguistic behavior. Thus, the atelic classes of states and activities are compatible with durative adverbials, while the telic classes of accomplishmentsand achievements are compatible with frame adverbials: (1) a. b. c. d.
That box contained a letter for a weeW*in a week. Mary drew circles for an hourl*in an hour. Mary drew a circle in an hourl*for an hour. Mary discovered the answer in an hour/*for an hour.
An ongoing issue in the study of aspect has been whether the four Vendlerian aspectual classes (or alternative classifications) are lexical primitives, e.g. a feature of the verb, or whether event types are compositional: the "lexicalist" vs. "constructionalist" debate. Verkuyl (1972, 1993), in a model-theoretic approach, proposed that situation aspect ('inner aspect', in his terms) is a property of sentence
ISSUESAND INTERFACESIN THE STUDYOF ASPECT
3
structure, derived on the basis of three parameters: (i) a verbal feature expressed informally as [ADDTO], whch in effect means that the verb can combine with its direct object in order to use the object as a counting device, (ii) a nominal feature, [SQA] (Specified Quantity of A, following Gruber 1976), which determines whether the object has specified cardinality or not, and (iii) the nature of the thematic relation between the verb and the object, totally affected or not. Durative aspect (e.g. activities and states) is considered to be unmarked, and will obtain when any one (or all) of the three parameters has a negative value. Terminative aspect, the marked form, obtains only when all three parameters have a positive value, as illustrated in (2), focusing on the contribution of [fSQA]: (2) a. b.
Mary drew circles (for an hour). : durative [+ADDTO],[-SQA], [+TOTALLYAFFECTED] Mary drew twenty circles (*for an hour). [+ADDTO],[+sQA], [+TOTALLY AFFECTED] : terminative
The essential semantic notion underlying this system is that of path: a verb which is [+ADDTO]combines with its object to form a PATH,which is then a bounded path if AFFECTED]. the object is [+SQA]and [+TOTALLY Verkuyl's work served to highlight the importance of the relationship between the verb and its direct object for the 'determination of aspect; hence the interest in aspectual or event classification for syntacticians. Tenny (1987), much refined in Tenny (1994), was the first significant work within generative syntax to examine the connection between the verb-direct object relation and event type. She proposed that lexicon to syntax mapping of verbal arguments is determined by aspectual roles: the Aspectual Interface Hypothesis. This hypothesis claims that universal lexicon to syntax llnking principles are sensitive only to links between (subparts of) events and the event participants. In particular, the role of the direct object in delimiting the event is highlighted by her Measuring-Out Constraint on Direct Internal Arguments, according to which the direct object of a verb undergoes no necessary internal motion or change unless said motion or change measures out the event over time. Furthermore, direct objects are the only overt arguments of the verb whch can fulfill this aspectual role. Tenny's work focuses on the initial mapping of arguments in the syntax, not on subsequent syntactic movement. In addition to the effect of the internal structure of the direct object on aspectual classification, illustrated by the activity/accomplishment alternation of (2a) vs. (2b) above, there are of course other ways in which the overall syntactic structure of the verb phrase affect event interpretation, such as the addition of a resultative phrase or a goal PP: (3) a. a.' b. b.'
Bill hammered the metal (for two minutes). Bill hammered the metal flat (in two minutes). Sisyphus pushed the rock (for days). Sisyphus pushed the rock to the top of the hill (in an hour).
4
P. KEMPCHINSKY AND R. SLABAKOVA
Much of the syntactic work on aspect has focused in particular on accomplishment/activity alternations such as those above, which prima facie appear to provide the strongest empirical evidence for the constructionalist position. Nevertheless, Rappaport Hovav and Levin (1998) argue for a lexicalist-based approach to these alternations. Their basic underlying assumption is that it is the lexical properties of the verb which determine the syntactic frame in which it may appear, and hence its aspectual interpretation. They propose that the basic elements of verb meaning are represented by primitive predicates such as ACT, CAUSE and BECOME,organized according to certain specified event structure templates, whlch define the aspectual classes. Thus, for example, an achevement has the template [ BECOME [ x I], while an accomplishment may have, as one of its possible templates, [ x CAUSE [ BECOME [ y I]]. Mapping from the lexicon to the syntax is governed by conditions such as the Subevent Identification Condition, which specifies that each subevent in the event structure must be identified by a lexical head in the syntax. Aspectual alternations occur via the process of Template Augmentation, according to which a less complex event type (e.g. an activity or a state) may be augmented to a more complex event type (an accomplishment or an achievement), but the reverse does not hold. Hence verbs whose basic event type is activity may appear in the event structure template of accomplishments. An important issue within the constructionalist side of the debate is the nature of the functional projections involved in the syntactic representation of aspect. Some researchers have proposed that there are functional projections specifically dedicated to aspect. Borer (1994) proposes that there is an Aspect Phrase headed by the feature [+EM] (Event Measure) whose specifier position both licenses the aspectual role of the direct object and is the locus of accusative Case checking; hence, in a sense she recasts the purely functional category of AgrO (Chomsky 1993) in a more interpretive mode (also see Borer, this volume). In contrast, Schrnitt (1996) and Zagona (1999) specifically propose that the formal category Agr is the locus of aspectual calculation; the purely formal need to check Case provides, as it were, the necessary syntactic configuration for the semantic features of the verb and its internal argument (e.g., Verkuyl's features of [+ADDTO] and [+sQA])to become visible to each other. Travis' work on event phrases in syntax (Travis 1994, 2000a) also examines this question. She proposes to use the split VP structure originally proposed by Larson (1988) to represent in the syntax each of the predicate operators in the semantic representations of the lexical aspectual classes. The lower VP represents the most deeply embedded,stative predicate, while the head of the higher VP corresponds to the CAUSE operator of, for example, accomplishment predicates. In her contribution to this volume she revisits this issue and provides a refinement of her earlier work. 1 2. The representation of sentential aspect
There are two main issues in the study of sentential aspect: (i) the nature of the interaction between sentential aspect and lexical aspect in deriving the interpretation
ISSUESAND INTERFACES IN THE STUDYOF ASPECT
5
of the event, and (ii) the syntactic representation of sentential aspect within a highly articulated clausal structure. Perfective vs. imperfective sentential aspect corresponds to the notions of 'boundedness' and 'unboundedness', and is sometimes conflated with the notions of 'telicity' and 'atelicity'. The two types of aspect can clearly be grammaticalized in different ways, as illustrated by the Bulgarian examples below. Preverbs (aspectual prefixes on the verb) encode telicity, while aspectual tense suffixes encode boundedness. Preverbs and aspectual suffixes are given in bold. (4) a.
b. c.
d.
Ivan gotv-i jadene. (atelic bounded) Ivan cook-AORISTfood 'Ivan cooked (an unspecified quantity of) food.' Ivan z-gotv-i jadene. (telic bounded) Ivan PREVERB-cook-AORIST food 'Ivan cooked (some specified quantity of) food.' jadene kogato ti dojde. (atelic unbounded) Ivan gotv-eie Ivan cook-IMPERFECT food when you came 'Ivan was cooking (an unspecified quantity of) food when you came' Ivan z-gotv-jaSe jadene :tom beSe svoboden. Ivan PREVERB-cook-IMPERFECT food when was free (telic unbounded) 'Ivan cooked (some specified quantity of) food whenever he was free.'
Given the general lack of grammaticalization of lexical aspect in the Germanic and Romance languages, linguists working on these languages have been somewhat divided on the question of the distinction of the two aspects. Verkuyl(1993) argues that viewpoint aspect ('outer aspect') may modify situation aspect, but cannot undo it. For him, there is no essential difference between the two with respect to their interpretive effects, although he recognizes a structural difference, in terms of their place in the syntactic representation of the sentence. The essence of both is that the event (or "temporal entity") is conceived of as bounded; beyond that, he finds the distinction between the two "distracting" (op. cit., p. 11). Smith (1991) points out a clear semantic distinction between lexical telicity and temporal boundedness; as she puts it, telic events finish or are completed, while atelic events, even when temporally bounded, only stop or are terminated. It is precisely the notion of boundedness that underlies Giorgi and Pianesi's (1997) analysis of perfective vs. imperfective aspect. To dstinguish the semantic interpretation of a past imperfective sentence from a past perfective one (as in Italian Gianni rnangi6/mangiava una mela 'Gianni ~ ~ ~ . P E R F / ~ ~ ~an. Iapple'), MPEW they propose the semantic primitive of TOPOLOGICAL CLOSURE.A perfective predicate may denote only topologically closed events; therefore, the temporal variable associated with the sentence must be existentially bound. Conversely, an imperfective predicate denotes a topologically non-closed event, and the temporal variable may be bound by, for example, a universal quantifier (cf. Bonomi (1997),
P. KEMPCHTNSKY AND R. SLABAKOVA
6
who proposes that the imperfect always introduces a quantificational structure of universal quantification over circumstances.) Since a topologically non-closed event is one which does not contain a boundary, this means that a predicate expressed with imperfective morphology must have a process subevent in its event structure. The issue then becomes how to represent the interaction of sentential aspect and lexical aspect in the syntax, in terms of the functional categories involved. That is, how is the necessary relation between sentential aspect and the relevant subevent of the event structure of the predicate established? One possible approach to this question is provided by Demirdache and Uribe-Etxebarria (1997, 2000, 2002), who set out to provide a syntactic structure for tense and sentential aspect; their general analysis will be discussed in greater detail in the next section. Nevertheless, an open question is whether such aspectual interactions are represented in the syntax or are more properly treated as post-syntactic, as in Discourse Representation theory. This question is particularly acute for apparent aspectual class alternations which seem to be due not to differences in verb phrase syntax, but rather strictly to morphological distinctions of perfectivity. This is the case in, for example, the Romance languages. Normally stative predicates, when appearing in a perfective tense, denote the inception of a resulting state rather than the state itself, and hence show the linguistic characteristics of achievement predicates, as illustrated by the fact that they may be modified by point of time adverbials (and see Zagona, this volume):
(5) a. b.
Maria conocia a Juan por muchos aiiosl*ayer a1 mediodia. 'Maria knew (IMPERFECT) Juan for many yearslyesterday at noon.' Maria conocio a Juan ayer a1 mediodia. 'Maria knew (PRETERITE) Juan yesterday at noon.' = Maria met Juan yesterday at noon.
The alternation in (5) is an example of aspectual shift. De Swart (1998), worlung within DR theory, proposes that such examples of aspect shift are coerced by the discourse context, and as such are distinct from aspectual alternations which are overtly marked by explicit grammatical markers. In this sense, sentential aspect works as an "eventuality description modifier", but the coercion itself is due to the interaction of sentential aspect with the immediate linguistic context and external real world knowledge. The interaction of the two types of aspect thus also bears on the broader question of the nature of the syntax-semantics interface, as well as on the nature of the interaction between the grammar per se and extragrammatical factors such as discourse context. These two broad issues have also surfaced in research on f i s t and second language acquisition of aspect,
ISSUESAND INTERFACES IN THE STUDYOF ASPECT 1.3. The role of aspect in language acquisition The first and second language acquisition of tense and aspect has probably been the most prolific topic of research in the field of applied linguistics. The body of literature on, say, acquisition of questions, negation, null subjects, even of inflectional morphology, pales in comparison. Furthermore, tense and aspect have been approached from a number of research perspectives with different epistemological affiliations. Although we cannot do justice here to this enormous body of work, we shall briefly mention the main findings, some prominent explanations, and some trends in recent research. Since the 1970s, the following four associations have been observed when children and adults are learning their first and second languages: Learners first use (perfective) past marking on achievements and accomplishments, eventually extending use to activities and statives. (b) In languages that encode the perfectivelimperfective distinction, imperfective past appears later than perfective past, and imperfective past marlung begins with statives, extending to activities, accomplishments, and achievements. In languages that have progressive aspect, progressive markmg (c) begins with activities, then extends to accomplishments and achevements. (d) Progressive marking is not incorrectly overgeneralized to statives.
(6) (a)
Based on those observations, the claim of the Primacy of Aspect (POA) Hypothesis (Antinucci & Miller, 1976; Bloom, Lifter & Hafitz, 1980; Bronckart & Sinclair, 1973) is that the language acquirer initially marks aspect, rather than tense. Many explanations for the observed'associations have been proposed so far. The Prototype Theoly explanation argues that these associations are due to some mapping of prototypical features. For example, the prototype for the category "progressive" can be characterized as "action in progress". The lexical classes that exhibit this meaning are activities, having the semantic features [dynamic] and [atelic]. Within L2 acquisition research, Bardovi-Harlig (1998) has argued for a role for discourse factors. The discourse explanation argues that narrative structure influences the distribution of tense-aspect morphology in interlanguage: predominantly perfective in the foreground, predominantly progressive in the background. The hypothesis is that ultimately the acquisition of the appropriate verbal morphology is tied to the discursive ends for which the perfective1 imperfective distinction is employed. Early generative approaches to the issue include Bickerton's (1981, 1984) Language Bioprogram explanation that two semantic contrasts are pre-programmed, so they emerge early in child language: the state-process distinction, and the punctual-nonpunctual distinction. A more recent offering is Olsen and Weinberg's (1999) Subset Principle (Benvick, 1985; Wexler & Manzini, 1987) explanation. Children begin the acquisition process constrained by parametric options.
8
AND R. SLABAKOVA P. KEMPCHINSKY
Furthermore, due to learnability considerations, they initially hypothesize the smallest possible grammar, which they can later abandon in favor of a more inclusive grammar only on the basis of positive evidence. In the acquisition of aspect, the most restricted association that children can posit is one-to-one. Since there exist natural languages (for example, Mandarin and Korean) that make use of straightforward associations-'one lexical aspectual feature is always linked to one grammatical aspect morpheme'-it is argued that children assume this most restrictive mapping as their initial hypothesis. At the onset of the acquisition process, imperfective is mapped onto dynamic, durative predicates, and perfective is mapped onto telic predicates. If these initial values are incorrect for a specific language, upon hearing child-directed speech that disconfirm them, children will relax their initial undergeneralization. What has not been clear from this body of research, however, is the extent to which L1 and L2 acquirers actually know the relevant semantic interpretations of the aspectual distinctions. In her pioneering work on child acquisition of lexical aspect, van Hout (1998) defines the theory of telicity at the interface of the lexicon, syntax, and semantics. She is interested in how Dutch children acquire the eventsemantic knowledge of verb-frame alternations (e.g., activity versus accomplishment, signaled by presence of an object of [+SQA]). van Hout finds that children up to the age of five do not know the exact aspectual implications of the transitive-intransitive alternation, nor do they aspectually distinguish predicates with [+SQA] objects from those with [-SQA] objects. These findings are compatible with similar experimental results from English child language acquisition (Wagner 1997). The question arises, then, of how to integrate the claims of the POA Hypothesis, according to.which children use lexical aspect to bootstrap themselves into viewpoint aspect, and the findings of van Hout and Wagner, according to which children are not aware of grammatical markers of lexical aspect. Furthermore, as van Hout and Hollebrandse's (2001) and Wagner's (1998, 2001) research findings suggest, even children who produce adult-like aspectual viewpoint morphology may have non-adult aspectual interpretations. Explaining why this is so is one of the most promising areas of future research on aspect acquisition. Within second language acquisition, work by Montrul and Slabakova (2002, 2003), Slabakova and Montrul (2002, 2003), Slabakova (2001) as well as Kozlowska-Macgregor (this volume) follows this intriguing avenue of further research into interpretive properties of language development (see Slabakova (2002) for an overview of recent acquisition research on aspect). 2. THE ARTICLES
The papers from the workshop clustered within three main areas of research, and are so organized in this volume. The papers in Part I are primarily concerned with the internal structure of the clause and its relationship to lexical aspect. The papers in Part I1 examine, from a variety of perspectives, the interaction of aspect, tense and discourse, while the papers in Part I11 present research results from studies on aspect in f i s t and second language acquisition and in language attrition. As we had
ISSUESAND INTERFACESIN THE STUDYOF ASPECT
9
hoped would be a result of the workshop, several common themes arose which cut across these divisions, and we have chosen to summarize the contributions to this volume in an order which highlights these links. A number of papers address the issue of the relationship of event structure and the syntactic projection of arguments, and the determination of the true locus of situation aspect: as a lexical property of verbs, or as a result of the syntactic structure into which the verb and its arguments merge. The paper by Elizabeth Ritter and Sara Rosen sets out a programmatic approach to the investigation of how event information is encoded in the syntax. Noting that a considerable body of research converges on the view that certain functional categories are the grammatical expression of event information, they propose a tripartite classification in terms of how languages organize arguments in the syntax: object split languages, subject split languages, and topic-comment languages. The first two types grammaticize the "contours of the event", either by encoding event delimitation (object split languages) or event initiation (subject split languages). The third type, by their proposal, does not encode event structure at all in syntactic structure, but rather organizes arguments on the basis of their discourse function. Object split languages-that is, languages which do not mark all objects alikegenerally distinguish between objects on the basis of definiteness. The definiteness distinction may in turn be related or not to event delimitation, depending on whether the [Quant] feature on the head of Aspect Phrase is uninterpretable or interpretable. Subject split languages, on the other hand, show Case distinctions on subjects based on either animacylperson restrictions'or on agentivity vs. non-agentivity. Ritter and Rosen propose that such languages grammaticize event initiation, with the relevant functor category being TP, carrying a [person] feature. Finally, in the third type of language, topic-comment languages, there are no event features to be checked in the syntax, nor, as a consequence, phi-features. Such languages, therefore, are predicted to have no subject requirement. Clearly within the realm of aspect studies, the most extensively studied type of language in Ritter and Rosen's classification are languages which grammaticize event delimitation, where, following in particular work by Krifka (1989, 1992), the relevant semantic property of the object (Verkuyl's +SQA) is usually assumed to be quantization. This is precisely the issue taken up in her paper by Hagit Borer, who argues that the relevant notion is not quantization, but rather quantity. She starts by examining the syntactic and semantic characterization of quantity in nominals. She notes that if in DP there is assumed to be a functional projection of Quantity Phrase (#P), in which prenominal quantifiers such as some, too much or three are located, then it is desirable to assume that the nominal plural affix -s is not a head of #P, but rather is a classifier. Syntactically, bare plurals are llke mass nouns in lacking #P. Semantically, bare plurals do not presuppose the existence of singulars, in contrast with vague quantifiers such as more than three (apples) which do, at least, imply the existence of (in this case) at least 'three singulars. Borer proposes that a nominal (specifically, a direct object) may license telicity whenever #P within DP has some value, thus deriving the lack of telicity with bare plurals.
10
P. KEMPCHINSKY AND R. SLABAKOVA
The crucial distinction, Borer argues, between quantization (as in Krifka's work) and quantity is that quantization requires that every subpart of a quantized P be not P, while P may be quantity iff P is not homogeneous, which does allow for the possibility that some parts of P may be P. She proposes that semantic quantity is mediated through a specific functional structure, AspQ. In Spec of this projection, a nominal with quantity transfers the quantity value to the head ASPQ, rendering the event represented by the V as a quantity event. In languages such as English, the quantity nominal is necessary for telic interpretation precisely because the nominal is the source of the feature [quantity], while in Slavic-type languages the preverbal prefix which typically appears on accomplishment predicates assigns quantity value directly to the head of AspQ. In the .final part of her paper Borer shows how the weaker notion of quantity applies profitably to the computation of telic events: a predicate may be non-homogenous, and hence telic, even if some subparts of the event are proper subparts (the case of run to the house), or proceed past the 'endpoint' ($11 the room with smoke) or involve actions which cannot be measured by changes in the direct object (read a book). Borer argues that "co-finality" is in fact just a special case of telicity, rather than the defining case. Lisa Travis also takes as her point of departure the link between properties of the object (Case-marking, movement) and the calculation of telicity, questioning the view that it is (functional) specifier positions outside of the verb phrase which are crucial for the computation of aspectual classes. Her goal is to explore an alternative approach, according to which the syntactic structure and operations relevant to the computation of aspectual classes occur entirely within the verb phrase, specifically, within vP. The structure which she assumes for vP includes a Aspect projection between vP and VP, as in Ritter & Rosen's discussion of event delimitation languages, and is the location of one of the three positions w i t h vP relevant for the calculation of telicity: the head v of vP, the head Asp of AspP, and the head X of a lexical complement category within VP, generally PP or AP. Any of these three positions-v, Asp or X-may express the endpoint of an event, according to specific options realized by individual languages. In addition, the heads Asp and v may also encode a beginning point, and v may encode an arbitrary point in the event. Thus in Malagasy a morpheme merged into the head Asp encodes either the endpoint of the event or the beginning point, depending on the basic lexical class of the main verb. The Slavic preverbs, which for Travis are in v (rather than in some Asp(ect) position, as for Borer) may express the natural endpoint, beginning point, or arbitrary endpoint in the event, depending on the particular preverb. A given language may in fact employ more than one of these three telicity markers, as she shows with her analysis of complex verbs in Navajo and Slave. The overall picture which emerges from her analysis is that situation aspect is syntactically encoded in terins of positions within vP, but is quasi-lexical in nature in that vP is assumed to be the domain of idiosyncratic lexicalization; that is, it is the 1-syntax domain (Hale & Keyser, 1993). The acquisition study by Martyna Kozlowska-Macgregor addresses precisely the theoretical issue of the aspectual contribution of the Slavic perfective preverbs to VP interpretation, using data from Polish. Exploiting the distinction
ISSUESAND INTERFACES IN THE STUDYOF ASPECT
11
between syntactic (s-) and lexical (I-) syntactic features involved in morphosyntactic derivations as proposed by Travis, Kozlowska-MacGregor describes two homophonous prefixes po-. One instantiation ofpo- conveys that the event or state continued for a while; the other po- morpheme is used with plural objects and describes a bounded series of events completed one after another. She argues that these two morphemes have semantic features, and consequently different syntactic positions: one in 1-syntax and the other in s-syntax. The empirical base of the paper is an L2 study which tested whether'English learners of Polish are aware of these subtle interpretive properties. The theoretical account is validated by the strong performance of the Polish native speakers, while the near-native speakers' performance reveals an unstable, complex system, in which knowledge of 1syntactic features is present but knowledge of s-syntactic features is not. The paper by Raffaella Polli and Heidi Harley also focuses on one of the three telicity markers in Travis' schema: the nature of v. They set out to show that the main empirical problem for constructionalist approaches to aspectual calculationnamely, the lack of alternation patterns of the activity to accomplishment variety across all transitive verbs-can be .satisfactorily accounted for if the locus of variability is taken to be not the lexical verb, as in lexicalist approaches, but rather the light verb v. Specifically, they propose that v is available in three "flavors": vDo, VCAUSE and VSE,with each requiring a particular type of complement structure and imposing specific requirements on their external arguments. Thus vDo may appear with an incremental theme and must have an animate subject, as in John ate the apple, but V C A U ~ Emust appear with a resultative complement of a change of state and need not have an animate subject, as in The sea ate the beach away. A fundamental assumption of their analysis is that the event structure of predicates is decomposable both syntactically and semantically. In structures with VCAUSE, there must be a syntactic projection which encodes the resultant state, because what CAUSES do is to initiate a change of state. For Folli & Harley, this syntactic projection is a small clause structure. The alternation between event types and hence between "flavors" of v has language-specific reflexes; in this paper they concentrate on the role of the verbal particle in English and on the role of the inchoative reflexive morpheme si in Italian, Folli & Harley go on to briefly discuss some possible extensions of their approach, noting various verb classes in which animacy restrictions-for them, the result of inserting one type of v versus another-surface systematically with changes in event structure. Thls recalls Ritter & Rosen's discussion of subject split languages which encode such animacy distinctions grammatically. A different case of aspectual alternation is analyzed by Mai Tungseth: the variable behavior of combinations of verbs of motion and certain stative prepositions in Norwegian, shown in (7): (7)
Jon syklet i grerfta i en timelpg to sekunder. Jon biked in ditch.^^^ for an hourlin two seconds 'Jon biked in the ditch for an hourlinto the ditch in two seconds.'
12
P. KEMPCHINSKY AND R. SLABAKOVA
Also adopting an essentially constructionalist approach to aspectual classification, she argues that the two different interpretations of sentences such as (7) are a function of the location of the PP i grq7a in the syntactic structure. In both cases, the PP is assumed to be the complement of a generally phonologically empty F head. In the telic, directed motion interpretation, this FP appears in a complement position to the verb, while in the atelic, located motion interpretation, the FP is outside the VP in an adjunct position. VP constituency tests such as VPtopicalization and do-substitution support this analysis: only when the PP is interpreted directionally, as shown by the co-occurrence of interval adverbials, does it behave syntactically as part of the VP. These constituency tests are fwther bolstered by data on accent placement and on binding; for example, a direct object may bind a possessive anaphor within the PP only when the PP has a directional interpretation. The basic analysis is cast in terms of feature-checking: the head F of the FP dominating the PP carries an uninterpretable [dir(ectional)] feature which must be checked by the corresponding interpretable feature on the verb. On the assumption that t h s checking must take place within the verb's complement domain, the FP will have the [dir] feature checked only when it is in a complement position to the verb. Locative PPs, on the other hand, are assumed to represent the default reading; hence, the F head of the dominating functional projection has no feature to be checked, and so may appear in adjunct position. Tungseth extends the analysis to case of complex directional PPs, in which a directional particle is proposed to be the overt realization of the [dir] feature. Although Tungseth's analysis is.cast withn a constructionalist stance, it does appeal as well to a minimal notion of lexically-based variation, in that it proposes that it is the interpretable [dir] feature carried by the verb which satisfactorily checks and deletes the uninterpretable feature on the FP. Cristina Schmitt initiates her paper by highlighting the difficulty of separating non-compositional meaning from compositional meaning, particularly within Distributed Morphology (DM), which essentially eliminates the division between lexical rules and syntactic processes. If "lexical" properties of a verb include both the idiosyncratic (encyclopedic) meaning of the root, plus its abstract aspectual features, and if both of these contribute to licensing of the verb's arguments, then the task of separating the one from the other becomes a tricky one. Her aim in this paper is to tackle this problem by concentrating on the aspectual properties of copulas and copula-like verbs, since these are verbs which as lexical items are severely underspecified. To do this, she focuses on the Portuguese copulas ser, estar andficar, and she employs tools from both DM and the Generative Lexicon (GL) as set forth in Pustejovsky (1995). Since for DM copula-like verbs are essentially the analytic spell-out of the processes which create synthetic words, isolating their properties is a means of precisely identifying these processes. At the same time, she proposes to show that the basic aspectual 'ingredients' of a verb's lexical representation such as STATE, PROCESS or TRANSITION, as proposed by Pustejovsky, can characterize the basic meanings of these three copulas. Copulas are v elements; ser is a 'pure v', estar denotes a state and is hence v + STATE], and ficar denotes a transition and is
ISSUESAND INTERFACES IN THE STUDYOF ASPECT
13
therefore v + TRANSITION]. These aspectual features can also account for predictable meaning shifts, according to the nature of the complement-for example, 'act be' readings of ser and 'become' vs. 'stay' readings of ficar, depending on whether the complement is an AP or a PP. That a PP within VP plays a crucial role in determining telicity is well-known from examples of the type John pushed the cart to the river, where the PP is a Path (Tenny, 1994; Jackendoff, 1996) establishing a spatial axis which measures the progress of an event participant. Karen Zagona in her paper provides an analysis of the temporal interpretation of double access readings, as in (8), based on a temporal notion of Path: (8)
John said/announced that Mary is in Seattle.
The term 'double access reading' (DAR) reflects the fact that the time of the embedded (stative) predicates includes both the time of the matrix event and speech-time. Zagona notes that the availability of the DAR is aspectually constrained in that the matrix verb must be non-stative, and proposes that if a constituent bears features which allow it to be construed as a (temporal or spatial) Path, then that constituent can be interpreted as a (temporal or spatial) location. In (8) the embedded clause provides a temporal location for the matrix event. More precisely, she adopts Jackendoff s principle of Structure Preserving Binding, by which an event is interpreted as bounded if there is an identity relationship [Event' pathi ~ime']. Zagona gives evidence showing that for verbs of communication, the (temporal) Path is the communicative activity itself. Since the temporal relationship of the embedded CP in sentences such as (8) to the matrix event is one of inclusion but not identity, it must therefore be the case that the CP is syntactically outside the scope of the operation of Structure Preserving Bounding. This suggests that it is structurally akin to a depictive adjunct, which she proposes is adjoined to AspP, where AspP is the functional projection headed by sentential (grammatical) aspect, dominating VP. As she notes, this makes these CPs structurally parallel to the locative PPs in the aspectually ambiguous Norwegian sentences analyzed by Tungseth. In the final section of the paper she considers an alternative syntactic analysis, based on a structural analysis of Path constituents along the lines of Hale and Keyser's (1993) analysis of verbs such as put. The paper by Hamida Demirdache and Myriam Uribe-Etxebarria also considers temporal relations in spatial terms, focusing on the syntax and semantics of sentential Aspect and Tense. They have argued in previous papers that there is a single, uniform grammar for temporal and aspectual relations, based on the notion that tenses, aspects and time adverbials are dyadic predicates of spatiotemporal ordering. Tense has an external argument of Utterance-Time (UT-T) and an internal argument of Aspect Phrase, headed by Aspect, which in turn has an external argument of Assertion-Time (AST-T) and an internal argument of VP, whose external argument is Event-Time (EV-T). Like purely spatial entities, these predicates locate some Figure with respect to the Ground, and thus establish relations of inclusion, subsequence or precedence. In this paper the authors focus on
14
P. KEMPCHINSKY AND R. SLABAKOVA
temporal modification by time adverbials, proposing that temporal modification is semantically and syntactically parallel to nominal modification. A time adverbial is base-generated adjoined to the temporal phrase whose reference it restricts. For example, in the sentence Abdel had left the house at 3 pm, the PP may either be adjoined to EV-T, yielding the reading that the leaving occurred at 3 pm, or it may be adjoined to AST-T, yielding the reading that the leaving had occurred prior to 3 pm. They extend this analysis to bare locating time adverbs such as yesterday, which they claim are headed by empty Ps of central coincidence, hence establishmg an inclusion relation, and also discuss various syntactic restrictions on the cooccurrence of multiple adverbs. In the final section of their paper, Dernirdache & Uribe-Etxebarria turn to the analysis of the temporal relations between subordinate and matrix clauses, which they argue are governed by two economy principles. A temporal derivation is optimal if this derivation yields some ordering of the Assertion Time of one clause relative to that of another clause, and if no step in the derivation is semantically vacuous. They demonstrate how the interaction of these two principles accounts for both possible and impossible temporal relations between clauses, both for complement subordinate clauses and adverbial subordinate clauses. Their model explicitly assumes that temporal interpretation happens at various steps in the derivation, and is therefore couched within a view of the grammar that conceives of a multiple interface between syntax and semantic interpretation. The interaction of lexical aspect and grammatical aspect is addressed by the intriguing study by Asya Pereltsvaig, focusing again, as in KozlowskaMacgregor's study, on the issue of Slavic perfective preverbslprefixes. Her objective is to determine what exactly these prefixes encode in the grammar of attrited Russian speakers (usually second-generation immigrant speakers, whose native Russian system has changed since they have become English-dominant). Assuming that Russian prefixes are grammatical aspect markers, Pereltsvaig argues that the attrited Russian grammar, unlike standard Russian, does not include a functional projection for encoding grammatical aspect, which she labels OuterAspP. This reduction of standard Russian phrase structure is possible if one assumes that grammatical aspect locates the event in time, as in the analysis by Dernirdache & Uribe-Etxebarria, and that events and times are of the same ontological type. Accordingly, standard Russian encodes reference time in OuterAspP, speech time in T, and event time in vP, while in attrited Russian TP selects vP directly as a complement. Therefore T relates the interval E with respect to speech time. Essentially, attrited Russian uses the perfective morphology to encode lexical aspect. Fewer temporal relations are encoded, but the semantic computation still goes through because of the above assumptions. But what are the temporal relations that are lost in this simplified system? Pereltsvaig shows that the semantic relations of precedence and simultaneity, as expressed by the progressive (John is eating right now) and perfect morphology (John had left by 5 o'clock) cannot be expressed in American Russian. Thus a syntactic analysis of tense and aspect along the lines of Demirdache & Uribe-Etxebarria's proposal is put to work in accounting for language breakdown phenomena.
ISSUESAND INTERFACES IN THE STUDYOF ASPECT
15
The idea that temporal and aspectual interpretation is based on the same set of semantic and syntactic primitives is further supported by Nina Hyams' study, in which she accounts for a curious pattern in child Greek: the use of a bare perfective form with third person singular frozen agreement. She argues that this form is a root infinitive analogue in Greek, a language without infinitives. The bare perfective form has a number of identifiable properties: it is non-finite, since the agreement morphology is not productive; it has a modal or irrealis meaning; it is restricted to eventive predicates; and it co-occurs with finite clauses. These properties bring the bare perfective in line with root infinitives in Germanic languages, French, and Russian. In the early grammar, Hyams argues, there is an opposition between modal and non-modal, or temporal, meanings, which children map onto non-finite and finite verbal forms, respectively. Hyams analyzes the construction as follows: in the adult grammar, the hierarchy of functional projections is MoodP - TPIAgrP -AspP - VP. In the child grammar, on the other hand, TPIAgrP and AspP are underspecified. In other words, the time-denoting heads are eliminated; the perfective feature merges under V and is checked against Mood. In this way, the analysis derives the lack of productive agreement and the modal reference effect. But what is the connection between modality and perfectivity? Deontic, or volitional modality encodes a polarity transition: to require, want, or intend P means that at the present stage, no P is true, while P will be true in some future interval. The same transition feature (7 P, then P) is involved in perfective aspect. In line with the continuity hypothesis (child grammar falls within the hypothesis space constrained by UG), Hyams identifies cases in Romance dialects in which features originally lower down in the tree, e.g., the past feature, can also license Mood. The interaction between lexical (situation) aspect, grammatical (viewpoint) aspect and tense is also a concern in the paper by Carlota Smith, but at the discourse rather than sentential level. Her unit of analysis is the passage, and she posits the existence of five "discourse modes", each of which can be characterized by a certain cluster of linguistic features. She sets out to demonstrate that aspectual situation categories identify specific types of discourse units, thus providing a framework for the linguistic study of discourse. In setting up this framework, she proposes to expand the classes of situation entities to include the classes of abstract entities and general statives, in addition to the more familiar eventualities of specific events and states. Each of the discourse modes has a predominant type of situation aspect; thus, for example, the Narrative mode is characterized by eventualities while the Report mode is characterized by a mixture of eventualities and general statives. A major section of her paper is devoted to the analysis of text progression in each of the five discourse modes. According to the discourse mode and hence the type of predominant situation entities, this advancement will either be temporal or atemporal. Atemporal text progression may be reflected by changes in spatial location-continuing the analogy of spatial and temporal progression-where this spatial location may be actual (Description, e.g. of a scene) or metaphorical, with changes of location through the information space of the text (Argument and Information modes). Smith illustrates text progression for the five discourse modes
16
AND R. SLABAKOVA P. KEMPCHINSKY
with examples of written text extracted primarily form non-fictional sources such as magazine and newspaper articles. She argues that couching the study of discourse in terms of these discourse modes is potentially more useful than focusing on genres, because the latter is very context-dependent. Two papers in the volume address the discourse functions of aspect within the framework of Segmented Discourse Representation Theory. Patrick Caudal's contribution, in his own words, is an attempt at a formal semantic treatment of viewpoint aspect, but in actual fact it formalizes the way aspectual meanings are calculated right up to the level of integrating discourse information. He distinguishes his theory from recently proposed approaches to viewpoint aspect in the literature, namely, de Swart's (1998) proposal that the aspectual contribution of the tenses is to provide aspectual type shifts; Kamp and Reyle's (1993) stage decomposition approach within Discourse Representation Theory, and Pustejovsky's (1995) mereological approach, treating stage relations as part-of relations. Caudal's model actually incorporates features from some of the models mentioned above. He discriminates among at least three canonical types of eventuality stages: 1) the inner stages or core stages, ascribed to all eventualities, they are "picked out7' by the neutral past simple tense (for example, for telic predicates, the inner stages include the terminus); 2) preparatory stages, or causal stages selected under some prospective readings of the progressive tenses, e.g., John was reaching the summit; 3) resultative stages, also ascribed to all eventualities and focused on by the perfect tenses. The defining property of stages is their ability to come into focus as a result of the application of viewpoint operators. He argues that much aspectual information, in the form of eventuality descriptors, is encoded in the lexicon; a verbal lexical item contains at least some information about stage salience. At the next level, he views aspectual tenses as "camera lenses" capable of focusing on a particular stage as per the viewpoint of the speaker (recalling Smith's (199111997) term of viewpoint aspect). Furthermore, aspectual VP modifiers (adverbials) also have the ability to bring stages into focus. Finally, context is the ultimate "focuser" of stages. Caudal's theory is inherently modular, since all these operations can be viewed as presupposing one another hierarchcally and extending over time. Thus, his model is compatible with the findings of the acquisition studies by Pereltsvaig and Hyams, discussed above, as well as the study by van Hout, to be discussed below. Building on the viewpoint aspect treatment proposed in Caudal's article, Patrick Caudal and Laurent Roussarie treat tenses not simply as viewpoint operators, but as illocutionary viewpoint functions which constrain rhetorical relations. They start out from the classic observation that not all aspectual tenses can appear in all speech acts. For example, the Frenchpassd simple cannot occur in hypothetical speech acts while the imparfait can. Following work by Asher and Lascarides (1994, 1998, 2001), the authors assert that the viewpoints themselves are a specific sort of speech act infomiation, capable of interacting with discourse interpretation via discourse relations. One of the applications of this theory is demonstrated by describing and capturing formally the illocutionary force of the perfective and imperfective aspect in French. The perfective viewpoint is simply
ISSUESAND INTERFACES IN THE STUDYOF ASPECT
17
assertive, while the imperfective viewpoint is discursively underspecified. This means that it can be associated with a variety of speech acts. Although its main contribution out of context is to provide the background discourse relation, this function is defeasible by context. All of the rest of the conversational implicatures of the imparfait (e.g., free indirect speech (9), politeness imparfait (lo), and so on) are compositional interpretive effects and not part of its semantics. (9)
I1 partait, sa decision ttait prise. 'He was leaving, he had made his decision.'
(10)
Je voulais voir mon fils. 'I wanted to see my son.'
Similarly, the narrative use of the imparfait (used to describe subsequent finished events) is due to its use within the context of a literary narration, a fact which remains unexplained by existing viewpoint theories. It is conceivable that the added complexity of the imperfective as opposed to the perfective is reflected in longer processing time, later acquisition, and faster attrition. In her experimental study of children's knowledge of viewpoint aspect in Polish, Angeliek van Mout specifically sets out to determine whether children display "imperfect" or "perfect" imperfectives in their grammar. To do this, she improves on the methodologies of previous studies on this topic. Whereas previous comprehension experiments gave children a choice of only two different situations, she presents children with three kinds of situations: complete, incomplete, and ongoing. In addition, the experimental setup requires the children to integrate an event variable into an existing discourse structure. Adults chose only completed situations when given perfective verb sentences and only ongoing situations for imperfective verb sentences, as the discourse structure integration for the specific situation posits. On perfective verbs,children behaved like adults; in contrast, they mapped all three types of situations onto imperfective verb sentences. Van Hout argues that the most plausible explanation for this non target-like pattern is that children are aware of viewpoint aspect semantics, but somehow fail to integrate this knowledge at the semantics-discourse interface. More specifically, children fail to relate the event of the test sentence to the relevant one in a sequence of events specified by the particular discourse structure of the test situation. Apparent difficulties in the acquisition of imperfective grammatical aspect in second language acquisition are the concern of Kathleen Bardovi-Harlig's contribution. Her point of departure is the Aspect Hypothesis (Andersen and Shirai 1994, see section 1.3 above), and she addresses the second part of this hypothesis: in languages that encode the perfectivelimperfective distinction, imperfective past appears later than perfective past, and imperfective past marking begins with statives, extending next to activities, then to accomplishments, and finally to achievements (Shirai, 1991). This particular claim has been little investigated, perhaps because the bulk of the L2 studies in this area have been based on L2 English, and English does not have (simple) imperfective aspect. Bardovi-Harlig
18
P. KEMPCHINSKY AND R. SLABAKOVA
surveys what has been established in the literature on second language acquisition of the imperfect. Several factors have been documented as influencing the distribution of imperfective morphology. Narrative structure (text structure) is one factor, in that the imperfective is typically found in the background part of the learner narrative (where native speakers would also use it). Another factor is the type of narrative: learners are more accurate with the use of imperfect in personal narratives than in impersonal ones. Many studies have found that only a very limited number of stative verbs appear with the imperfect, contrary to the Aspect Hypothesis claim; t h s is attributed to limited lexical knowledge. Bardovi-Harlig discusses the well-documented late acquisition of the imperfect and the possible sources of t h s delay, and suggests new avenues for research on the imperfect in second language acquisition, talung into account its poly-semantic nature and its discourse functions. That is, learners seem not to have acquired the discursive underspecification of the imperfect discussed by Caudal and Roussarie. As we have shown in this introduction and review of the articles, approachmg the study of aspect from the perspective of different subdisciplines broadens our understanding of the general phenomenon. We believe that the different angles brought together in t h s volume have resulted in a more comprehensive picture of the representation of aspect in the mindhain of the speaker. University of Iowa
PART ONE
ASPECT AND THE INTERNAL STRUCTURE OF THE CLAUSE
ELIZABETH RITTER AND SARA THOMAS ROSEN
TOPIC OR ASPECT Functional heads, features and the grammaticization of events*
1. INTRODUCTION In recent years, there have been a number of attempts to explain how and to what extent event information is encoded in syntactic structure. Proponents of the view that event information is encoded in the syntax (including Borer, 1994, 1998; van Hout, 2000; Kratzer, 1989; Manzini & Savoia, 1998; Ramchand, 1997; Ritter & Rosen, 1998, 2000, 2001; Travis, 2000a) have all assumed that it is the functional categories within the extended projection of the predicate that encode such information. Building on our own previous work, we provide evidence that if event information is grammaticized, it is encoded in the functional head responsible for licensing direct objects. Specifically, a language that grammaticizes event delimitation has Aspect in its inventory of functional categories. We assume that in such a language Aspect carries this information, and only objects that delimit the event check their phi/Case features in Spec, AspP (Travis, 2000a). Objects of nondelimited predicates either fail to check their features or check them in situ. This leads to object splits not only on the basis of delimitation, but also on the basis of properties of the object, including specificity or definiteness. Languages with object splits based upon aspect are languages that grammaticize event delimitation. In contrast, a language that has no aspectual feature associated with the functional category that mediates object agreement and Case has no grammatical reflexes of delimitation (or telicity). We argue here that there are two alternative organizing principles for such a language. One alternative is to organize the arguments around agentivity. We assume that the phi and Case features of the subject are checked in Spec, TP, following Chomsky (2001). In languages that grammaticize agentivity, Tense carries information about agents, i.e., only agentive subjects check their phi/Case features in Spec, TP. Non-agentive subjects either have no phi features to be checked, or they check their features elsewhere. The result is a split in the behavior of subjects, based upon person or animacy, since prototypical agentive subjects are animate or human (Dowty, 1991). The claim, then, is that languages that have subject splits along the lines of person or animacy are languages that grammaticize agentivity. The other alternative is to organize the arguments based on their discourse function, e.g. whether they serve as topic or comment. Reasonably there needs to be 21 P. Kempchinsky and R. Slabakova (eds.), Aspectual Inquiriess, 21-39. © 2005 Springer. Printed in the Netherlands.
some basic organizing principle that determines how the pieces of a sentence are put together. Observationally, there are two possibilities: the discourse-what it is that the speaker is talking about, and what he or she has to say on the topic; and the event being described-who is doing what to whom. In the first case, the relationship between the topic, the most prominent DP in the clause, and what follows is unpredictable; the topic may bear any thematic role in the event denoted by the verb. In the second case, the clause is structured around the verb and its arguments, and the interpretation of the most prominent DP, the subject, is predictable, given the meaning of the verb. We follow Rizzi (1997) in assuming the existence of a Topic Phrase whose head (Top) lacks features; this lack of features permits an explanation of the following facts: (i) TopP is completely optional (there are no features to check); (ii) there is no topic agreement (again, there are no features); and (iii) whenever TopP is present, an XP moves into its Spec (if the head is empty the Spec must be filled). As a discourse feature, Topic appears at a syntactic edge. Edge-related processes are post-syntactic and do not have the capacity to look inside syntactic constituents. The functional category Topic is fundamentally different from Tense and Aspect in two important respects (i) it is optional and (ii) it lacks phi and Case features. These differences are due to its distinct function: Topic provides information about the structure of the discourse, not the event. Cross-linguistic variation in the content of Tense and Aspect is attributed to variation in the feature content of these heads. Cross-linguistic variation in the frequency of topic sentences is related to the degree to which event information is grammaticized: Topic sentences will be most common in a language that fails to grammaticize either agentivity or delimitation simply because such a language will rely more heavily on discourse determined sentence structure when event information is unavailable as an alternative. Thus, we will argue for a tripartite classification of languages. Some languages organize their arguments according to their discourse function while others organize them according to properties of the subject or the object. For those languages that use the arguments as the primary organizational component, some languages do so by means of delimitation or telicity while others do so by means of person or animacy of the subject. Our approach is minimalist in that we assume only those functional projections that have independent semantic motivation, i.e., projections of temporal functional categories, Tense and Aspect, and a discourse-determined functional category, Topic. Additionally, we assume that the feature content of these heads is minimalspatio-temporal features as required for Tense and Aspect (perhaps along the lines of proposals by Dernirdache & Uribe-Etxebarria, this volume), and in some languages, these functional categories may or must also bear phi and Case features. An analysis of the content of Tense and Aspect will determine whether a given language grammaticizes event initiation and/or delimitation.
TOPICOR ASPECT 2. GRAMMATICIZING DELIMITATION 2.1. Object Splits
An object split language is one that does not mark all objects alike. For example, some objects receive accusative Case.while others receive some other kind of Case; or some objects trigger verb agreement while others do not; or some objects undergo object shift while others do not. We have observed two types of object splits: (i) those that are conditioned by the definiteness or specificity of the direct object, and (ii) those that are conditioned by aspect-the delimitation of the eventin addition to the definiteness or specificity of the object. Languages that grarnrnaticize delimitation do so via object splits. Delimited events are quantized in the sense of Krifka (1989, 1992)-in other words, they are discrete and countable. The event participant that bounds, delimits or terminates the event is the delimiter, and the prototypical delimiter is an affected object (Tenny, 1994). The general claim is that objects that delimit tend to be quantized as well; they are definite or specific, discrete and countable. Borer (2004, this volume) discusses the notion of quantization in detail, and argues that quantity in the event cannot be equated with delimitation. The structural manifestations of telicity, she argues, are manifested of events with quantity, and not necessarily of events with perceived ends. We argue that the quantization of events and the quantization of direct objects are different manifestations of the same object checking relation. Icelandic and Hebrew both have object splits based upon the definiteness of the object. Mandarin Chmese, Finnish and Russian all have object splits based upon definiteness of the object and delimitation of the event. In Icelandic, object shift is not conditioned aspectually, as the examples in (1) and (2) show. Only definite direct objects may undergo object shift, and in particular, object shift in Icelandic is not sensitive to the delimitation of the event, as shown in (3) and (4) (examples from Collins & Thrainsson, 1996, p. 392; Diesing, 1997, p. 412). (1) a.
b.
(2) a.
b.
Jon las ekki [bzkurnar]. John read not the books 'John did not read the books.' Jon John
las [bakurnar] ekki. read the books not
E~ peklu ekki [Jon]. I know not John 'I don't know John' E~ pekki [Jon] ekki.
I
know John not
(3) a.
H a m las ekki [bzkur]. he read not books 'He didn't read books.'
b. * H a m las he read (4) a.
Eg
b.
* Eg
[bzkur] ekki. books not
peklu ekki [morg born]. I know not many children 'I don't know many children'
I
p e a [mijrg born] ekki. know many children not
The same basic facts obtain with overt accusative Case marking in Hebrew. In this language only definite direct objects bear accusative Case. The accusative Case marking is, again, not sensitive to the delimitation of the event.
(5) a.
b.
(6) a.
b.
(7) a.
b.
ani makir et dani. I know ACC Dani 'I know Dani. ' ani makir (*et) harbe yeladim xaxarnim. I know ACC many children smart 'I know many smart children.' ani raiti et dani. I saw ACC Dani 'I saw Dani.' ani raiti (*et) harbe yeladim xaxarnim. I saw ACC many children smart 'I saw many smart children.' ani karati et ha-sefer. I read ACC the-book 'I read the book.' ani karati (*et) sefer. I read ACC book 'I read a book.'
In contrast to the languages just mentioned, there are languages in which the
position of the object or the Case on the object is aspectually conditioned. In Mandarin Chnese, object splits are conditioned by both the specificity of the object and the delimitation of the event. The Mandarin BA construction involves object shift to the preverbal position, and is only possible when the event denoted by the predicate is delimited, as the examples in (8) and (9) show. Since delimited events require definite or specific objects, the shifted objects in the BA construction are also definite (examples from Cheng, 1988, pp. 74-75).' (8) a.
b.
(9) a.
b.
Ta sha-le [Zhangsan]. He kill-ASP Zhangsan 'He killed Zhangsan. ' Ta [ba Zhangsan] sha-le. He BA Zhangsan kill-ASP 'He killed Zhangsan.' Lisi kanjian-le [Zhangsan]. Lisi see-ASP Zhangsan 'Lisi saw Zhangsan.'
* Lisi
[ba Zhangsan] kanjian-le. Lisi BA Zhangsan see-ASP 'Lisi saw Zhangsan.'
Finnish is another example of a language that has event-based object splits. When the event is delimited, the object is marked with accusative Case; when the event is non-delimited or unbounded, the object is marked with partitive Case (Kiparsky, 1998, pp. 2-3, 5).' (10) a.
b.
(1 1) a.
b.
Anne rakensi talon vuode-ssa/*vuode-n. Anne built house-Acc year-INEssrvE/year-Acc 'Anne built a/the house in a year/*for a year.' Anne rakensi taloa tunni-n /*tunii-ssa. Anne build house-PARThour-ACC~~OU~-INESSIVE 'Anne was building dthe house for an hour/*in an hour.' Tiina beitti keihaan metsaan sekunni-ssd*sekunni-n. Tiina threw javelin-Acc into-the-forest s e c o n d - ~ ~ ~ s s ~ v ~ / s e c o n d - ~ c c 'Tiina threw the javelin into the forest in a second/*for a second.' Tiina heitti keihhta tunni-n/*tunii-ssa. Tiina threw javelin-PART ~OU~-ACC/~OU~-INESSIVE 'Tiina threw the javelin for an hour/*in an hour.'
Thus, Icelandic, Hebrew, Mandarin Chinese and Finnish all display object splits, but the triggering conditions and manifestations are different for each language. In Icelandic, the position of the object is determined by the definiteness of the object alone; in Hebrew, morphological Case is determined on the same basis. In Mandarin, the position of the object is determined by the specificity of the object and by the delimitation of the event; in Finnish the Case or aspectual marking of the object is determined by the definiteness of the object and by the delimitation of the event. The findings are summarized in Table 1.
Tablel. Object Split Languages Language
Dejniteness
Icelandic Hebrew Mandarin Finnish
Event
Manifestation object shift object shift Case
2.2. Analysis of Delimitation Languages We argue that object splits derive from the checking of the direct object's Case and agreement features in the relevant functional projection (FP). There seems to be a consensus in the literature that at least one FP is situated between vP and VP, but some disagreement as to its category, specifically whether it is AspP or Agr-0. We suggest that when the head of this FP has event-related featural content it is of the category Aspect, and otherwise it is Agreement. In Ritter & Rosen (2001) we proposed that in object split languages, definite direct objects are specified for a formal quantization feature ([Quant]) which must be checked against the [Quant] specification of the head of the FP. Here we propose that Asp differs from Agr-0 in that only the former may have an interpretable [Quant], indicating that the predicate denotes a delimited event. If AspIAgr-0 includes [Quant], only a definite direct object will be able to check its features by raising to Spec, AspIAgr-0. In a language with event-sensitive.object splits, the [Quant] feature on Asp is not deleted but rather is interpreted at LF as indicating that the predicate denotes a delimited event, and the definite direct object in Spec, AspP serves to delimit that event. In a language with an object split that is not sensitive to event type, on the other hand, [Quant] is an uninterpretable feature of Agr-0, which will delete before LF. In this case, the object split is conditioned by the definiteness or specificity of the direct object. The basic checking procedure is schematized in (12) below:
(12) a.
Grammaticization of both the event and the object
agent
v'
A
b.
Grammaticization of the object
agent
v'
A Spec
T
Agr-0'
I Agr-0 II-[
VP
A V
OBJ [QUANTI
In a language like Icelandic or Hebrew, in which the functional category (Agr0 ) is not specified for the event feature, the [Quant] feature of Agr-0 only checks definiteness or specificity of the object DP. In contrast, Mandarin and Finnish have object splits conditioned by both delimitation and definiteness or specificity. When Asp is specified for [Quant] this feature characterizes the event and is not deleted. By hypothesis, [Quant] is a unary valued feature. Hence, the Asp head of a nondelimited predicate has no [Quant] feature. Objects without this feature remain in situ and receive inherent (partitive or other) To summarize, we are claiming that all languages have a functional projection intermediate between vP and VP that checks the Case and agreement of direct
objects, but that the category and feature specification may vary across languages, depending on the content of the functional category that heads t h s projection. If this functional category bears event related information, i.e. if it is specified for an interpretable quantization feature, then the projection is AspP, and we find clear evidence that delimitation is grammaticized. For example, object Case in Finnish and object position in Mandarin Chinese are both sensitive to whether the event is delimited. 3. GRAMMATICIZING AGENTIVITY 3.1. Subject Splits
A subject split language is one that does not mark all subjects alike. For example, some subjects receive nominative Case, while others do not; or some subjects trigger agreement, while others do not; or some subjects are in a designated position, while others are not. We have observed two kinds of subject splits: (i) subject splits based on animacy, or more specifically person and (ii) subject splits based on thematic role, specifically agent vs. non-agent. Though on the surface these appear to be two distinct phenomena, we suggest that in fact both hnds of subject split grammaticize event initiation by identifying the external argument, i.e. the agent. This follows from the definition of an agent as the event participant that launches, initiates, or acts to perform the event, and the observation that prototypical agents are highly animate. Consider the animacy hierarchy proposed by Dixon (1994, p. 85), and adapted by Croft (1990, p. 112):~ (13)
lst/2nd person pronoun > 3rd person pronoun > proper name > human common noun > non-human animate common noun > inanimate common noun
The highest items on this hierarchy are 1" and 2ndperson pronouns, which identify the discourse participants, i.e. the speaker and addressee, and which normally have a human referent5 At the bottom of the hierarchy are common nouns that refer to inanimate entities. If volitional involvement and sentience are necessary characteristics of agents (Dowty, 1991), then the items high on the hierarchy are always potential agents, while those low on the hierarchy are not6 Thus, we suggest that animacy restrictions are a way of encoding agentivity in the agreement system. Dyirbal (Australian) and Euchee (Amerindian isolate) illustrate different , marhng of the subject is sensitive manifestations of subject splits. In ~ y i r b a l Case to animacy, or more specifically person. Only 1" and 2ndperson subjects receive nominative Case; others receive ergative Case. The examples in (14) and (15) illustrate the nominative Case on lStand 2ndperson subjects, and the ergative Case on 3'd person subjects, respectively (Dixon, 1994, p. 161).
(14) a.
b.
(15)
gana miyanda-nYu we.all-NOM laugh-NONFUT 'We laughed.' bura-n nYurra qana-na y o u . a l l . ~ o ~we.al1-ACC See-NONFUT 'You all saw us.' quma yabu-rjgu bura-n Father.~ss mother.^^^ see-NONFUT 'Mother saw father.'
Euchee is another example of a language with a person-based subject split. In this language, the position of a subject agreement prefix on the verb varies with the person specification of the subject (Linn, 2001; Linn & Rosen, 2001). When the subject is 1" or 2ndperson, the subject agreement prefix appears next to the verb root, as in (16). But when the subject is 31d person, the agreement prefix precedes the object agreement marker, as in (17).'
'
h8-di-'n& hlm(0BJ)- I(SUBJ)-see 'I see him.'
3 lStperson subject
= 0-S1-V
nedze-di-'n& ~OU(OBJ)-I(SUBJ)-see 'I see you.' h8-ne-'n& 3 2ndperson subject him(0~~)-~OU(SUBJ)-see 'You see him.'
= 0-S2-V
dze-ne- 'n& ~~(oBJ)-~OU(SUBJ)-see 'YOUsee me.' h6-dze-'n& ~~(SUBJ)-I(0B.T)-see 'He sees me.' h6-nedze-'n& he(suBJ)-you(0~~)-see 'He sees you.'
+ 3rdperson subject
= S3-0-V
Our claim is that languages like Dyirbal and Euchee grammaticize event initiation, as evidenced by their differential treatment of subjects that are proto-typical agents and those that are not. The distinction in Dyirbal and Euchee is between speech act participants and others. The former are highly animate, typically human, which is one defining characteristic of agentive subjects (Dowty, 1991). Agents are those arguments that typically initiate, cause, or control the event. We speculate that languages that grammaticize event initiation are sensitive to one of these notions. This is reminiscent of the range of properties manifested by languages grammaticizing terminal aspect. Among these properties are delimitation, telicity, boundedness, quantization and quantity. We leave to future research the question of identifying the concepts that unify different types of initiation and termination.
3.2. Analysis of Person-based Subject Splits We suggest that subject splits arise as a consequence of feature checking in TP. T(ense) has a person feature among its set of agreement features, and in languages llke Dyirbal and Euchee, person is the only agreement feature that Tense has. We assume that only lS' and 2ndperson DPs are in fact specified for person; so-called 3rd person DPs lack a person feature specification, though they may be specified for number andlor gender (Ritter, 1995; Harley & Ritter, 2002). Thus, in a language that grarnrnaticizes the event via agentivity, person is the mechanism for the grammatical marking of the event. Only lS' and 2ndperson subjects have the requisite feature to raise to Spec, TP to check their Case and phi features.
T'
T
I I
/", VP [person]
A
agent [person] t - - - - - J v
v'
/\ AspP
A Also in such a language, if a subject has no person feature (i.e., if the subject is 3'* person or inanimate), then it cannot raise to Spec, TP. We argue that when this happens, one of two alternative strategies to identify this subject argument must be used: (i) the argument is checked in a different functional projection such as a Gender Phrase (see Linn & Rosen (2001) for Athapaskan and Euchee), or (ii) the subject remains inside vP and receives ergative Case (see Bittner & Hale (1996) for Walpiri and Inuit, here exemplified by Dyirbal).
We begin with an account of the Dyirbal facts. In this language 1" and 2nd person subjects receive nominative Case, but 3 1 person ~ subjects receive ergative Case because 1" and 2ndperson subjects, bearing person features, raise to Spec, TP where they check phi and Case features. 31dperson subjects, lacking person features, cannot raise to Spec, TP. It has been suggested by Bittner and Hale (1996) that ergative Case is assigned to the subject inherently in Spec, vP. (19)
Dyirbal subject Case
TP
/'-.',
(DP) T' [1"12"~] , " , NOM T VP
A (DP) [3rd1
v
A v
ERG
AspP
& In Euchee the subject agreement marker appears in a different position in the morphosyntax depending upon the person features: (20)
Euchee subject agreement TP
A (DP) [l"12"~pers] T
Tf GenP
A (DP) [Ydpers]
Genf
A GEN
VP
/'-'.
agent
v'
/'-', v AspP A
Adapting Linn & Rosen's (2001) proposal, we suggest that 1" and 2nd person
subjects raise to Spec, TP to check the person features in T. Because 31dperson does not bear person features in this language, a 31d person subject fails to raise to Spec, TP. Linn & Rosen (2001) argue that 3rdperson subjects check gender features in a distinct functional projection, Gender Phrase (GenP).
3.3. ActiveStative Marking: A lexical property of verbs A related phenomenon observed in many languages is activelstative marlung or agentlpatient marking (Mithun, 1991). Euchee displays activelstative marking in 1'' and 2ndperson subjects. As the examples in (21) show, a lS'or 2ndperson subject of an active predicate has mono-morphemic agreement; a lStor 2ndperson subject of a stative predicate has bi-morphemic agreement, but no such distinction is made with 3rd person subjects. Note, however, that there can be no inanimate subject of an active predicate as shown in (22). (2 1) a.
b.
c.
(22)
Active marking in 1" and 2nd dithe 'I ran' 'tithe 'We (inclusive) ran' ntithe 'We (exclusive) ran' nethe 'You ran' 'anethel'zthe 'You (plural) ran' Stative marking in the 1" and 2nd(bi-morphemic) 'I am short' dzes'i'C '6-s'i'C 'We (inclusive) are short' nti-s'i'C 'We (exclusive) are short' ne&s7i'& 'You are short' '5-s'i'C 'You (plural) short'
No distinction in agreement for active and stative predicates; gender in 31dperson htithe 'He ran' htis'i'C 'He is short' * hti-s'i'C 'He is short' sethe 'She ran' ses'i'C 'She is short' No inanimate subject of an active predicate. 'It ran' hish'o 'It is withered'
* hithe
Similarly, Lakhota makes a morphosyntactic distinction between subjects that "perform, effect or instigate" the action and those that do not (Mithun, 1991, pp.
515-516). The examples in (23) and (24) illustrate that in Lakhota, as in Euchee, the form of subject agreement on the verb is determined by situation type of the event denoted. However, Lakhota makes a morphosyntactic distinction between agentive and non-agentive subjects rather than between subjects of active and stative predicates.g (23) a. b. c. d. e. f. g.
mawani wak'e wanuwe wathi waxphye blowakaska wagltpa
'I walk' 'I dug' 'I swam, bathed' 'I live, dwell' 'I'm lying' 'I hiccough' 'I vomit'
(24) a. b. c. d. e.
mahixpaye mat'& amikisni imaphi malhkhota
'I fell' 'I fainted, died' 'I got 'I'm tired' 'I'm Sioux'
It is tempting to conclude that activelstative marking is another manifestation of subject splits. However, it is entirely possible that the activelstative marking found in languages like Euchee and Lakhota is lexically-based, determined by verb class. We related the Dyirbal and Euchee facts in (14)-(17) to feature checking, proposing specifically that V raises to T to check its person feature in these languages, assuming that only highly animate subjects have the necessary person feature to enter into this checking relation in Spec, TP. The facts in (21)-(24) have to do with the form of agreement selected, which is reasonably treated as a lexical property of the verb, like situation-based classification. It is perhaps worth exploring whether verb class-based paradigms of this sort are only found in subject split languages, i.e., whether these different paradigms provide learners with the necessary cues to determine that they are acquiring a language of this type. 4. TOPIC-COMMENT ORIENTATION 4.1. Topic Languages
In sections 2 and 3 of the paper, we argued that information about the internal structure of the event is grammaticized in Tense and Aspect. That is, the fknctional categories that check the syntactic features of the subject and the direct object (i.e., Case and agreement features) grammaticize the contours of the event, including delimitation (through the object) and initiation (through animacy). In this section, we suggest that not all languages organize the clause around the internal structure of the event or even around the notion of subject and object. Rather, some languages organize the arguments around the discourse requirements of topic and comment.
Such discourse-oriented languages will highlight or mark arguments on the basis of the role each plays in the overall discourse rather than the role each plays in the event denoted by the predicate. Li & Thompson (1976) have developed a typology which distinguishes, inter alia, topic prominent languages in which the topic, rather than the subject (or direct object), plays a significant role in the organization of the clause. Their typology is reproduced as Table 2, and indicates that languages may be subject prominent, topic prominent, both, or neither. We would add to this typology the object-prominent languages discussed in section 2. Our purpose here, however, is to explore the properties of topic prominence.
Table 2. Li & Thompson's subject/topic based typology Properties subject-predicate plays a major role
Classification Subject Prominent
I
Topic Prominent
I
basic sentences have topiccomment structure
Both Subject prominent and Topic Prominent
I
in some basic sentences, subject-predicate plays a major role; other basic sentences have topiccomment shucture subject and topic have merged and are no longer distinguishable in all sentence types
Neither Subject Prominent nor Topic Prominent
I
I 1 I I
Languages Indo-European, Niger-Congo, FinnoUgric, Simitic, Dyirbal, ~ndonesian,Malagasy Mandarin, Lahu (Lolo-Burmese), Lisu (Lolo-Burmese) Japanese, Korean
Tagalog, Illocano
Li and Thompson point out that highly topic prominent languages have a distinct set of linguistic characteristics; those that they discuss are listed in (25). (25) a. b. c. d. e. f.
Characteristics of a Topic-Prominent Language Surface coding for topic in some topic prominent languages, but not necessarily for subject. No expletive subjects in topic prominent languages. Passivization is not common/not permitted in topic prominent languages. Topic prominent languages tend to be verb-final languages. No constraints on what can serve as a topic. "Double subjects" (i.e., sentences with a subject and a base-generated topic) are pervasive 'in topic-prominent languages, but rare or unattested in pure subject-prominent languages.
As noted in (25f), one salient characteristic of topic prominent languages is the existence of so-called "double subject" sentences containing both a topic and a
subject. It is important to point out that the topic is not a theta-marked argument, but is rather an extra element that indicates what the sentence is about. Given that it is not an argument of the verb, the topic is presumably merged into its spell-out position, rather than raised from some argument position. Examples of "double subject" sentences are given in (26) from Japanese and (27) from Mandarin Chinese (Li & Thompson, 1976, p. 468). (26)
Sakana wa tai ga oisii. TOP red snapper SUBJ delicious fish ' Fish (topic), red snapper is delicious.'
(27)
Nbike she ybzi da. that tree leaves big 'That tree (topic), the leaves are big.'
Topics are distinct from subjects in several ways. First, topics must appear at the left edge of the clause, whereas subjects may appear in a variety of positions (Rizzi, 1997). Notice that in the "double subject" constructions above, the topic always appears to the left of the subject. Second, topics rarely trigger verb agreement, whereas subjects often do. Many highly topic prominent languages, including Mandarin, Japanese, and Lisu (discussed below), appear to have no overt agreement. Third, topics must be definite (as illustrated in (28) for Japanese), but subjects need not be (Rizzi, 1997). Finally, topics have a constant discourse rolethey determine what the sentence is about, whereas subjects have different roles, depending upon the thematic requirements of the verb, ranging from expletives to themes to agents (Portner & Yabushita, 1998, p. 120). (28)
* Am
onna-no-hito-wa sono neko-o uchi-ni tsurete-kaerimashita. some woman-TOPIC that cat-ACChome-~octaking-retum.home 'A woman took the cat home.'
4.2. Lisu: the limiting case? Lisu, a Lolo-Burmese language, is considered a highly topic prominent language because almost every clause contains a topic (Li & Thompson, 1976). A striking property of this language is that while every clause has an identifiable topic, it is often impossible to distinguish subject from direct object or alternatively, agent from patient. There are no diagnostics that reliably identify subjects (or objects) in Lisu. Consider the examples in (29) and (30), from Li and Thompson (1976, p. 475). In these sentences, as in almost every Lisu sentence, the topic appears in initial position, but it is often impossible to distinguish subject (agent) from direct object (patient). Lisu has no morphological marking (Case or agreement) to distinguish subjects from objects, and no ordering restrictions to distinguish subjects from objects.
(29)
lathyu nya ana khu - a. people TOPIC dog bite - DECLARATIVE 'People, they bite dogs.' OR 'People, dogs bite them.'
(30)
ana nya lathyu khu - a. dog TOPIC people bite - DECLARATIVE 'Dogs, they bite people:' OR 'Dogs, people bite them.'
In (29) the topic people is interpreted as either the agent or patient of biting, with the non-topicalized argument dogs as the other argument of the verb. In other words, it is unclear whether the non-topicalized argument is the subject or the direct object of the verb. This is because Lisu is an SOV language lacking verb agreement. This makes it impossible to tell whether the non-topicalized constituent is a subject or an object. Lisu "reflexive" DPs can have a topic as their antecedent, as in (31a), and the so-called reflexive can consist of a copy of the antecedent, as in (3 lb). (3 1) a.
b.
lama nya yi kudwi: khu - a. tiger TOPIC he body bite - DECLARATIVE 'The tiger (topic), he bit his body (=hrnself).' lama nya lama kudwe khu - a tiger TOPIC tiger body bite - DECLARATIVE 'The tiger (topic), he bit his body (=himself).'
Even Lisu reflexive DPs can be topicalized, as the example in (32b) illustrates. (32) a.
b.
lama kudwt: nya lama khu - a. tiger body TOPIC tiger bite - DECLARATIVE 'His body (topic), the tiger bit it.' yi kudwk nya lama khu - a. he body TOPIC tiger bite - DECLARATIVE 'His body (topic), the tiger bit it.'
We take from these facts about anaphora that reflexive DPs in Lisu are not locally A-bound. In general, anaphors are bound by subjects (or objects), but this does not appear to be the case for Lisu. We suggest that the absence of A-bound anaphora is related to the insignificance of A-positions (subjects and objects) in Lisu. Next, sentences without topics are extremely rare in Lisu; Hope (1974) gives but one example of a sentence that appears to have no topic, in that the argument marked with the topic marker is non-specific. But when such sentences do occur, some DP must be marked with nya (the topic marker). The choice of which
argument to nya-mark is based on the thematic hierarchy, where agent < dative < object < instrument. The example in (33) is the one such example that Hope provides. It appears that nya identifies the subject in these rare sentences, but the notion of subject seems to play no other role in the grammar of Lisu (Li & Thompson, 1976, p. 477; example from Hope, 1974, p. 60). (33)
swu nya atha d3 - a. one NYA knife forge - DECLARATIVE 'Someone is forging a knife.'
Finally, in Lisu thematic and aspectual information is encoded in the lexicon, giving rise to fine-grained distinctions in Lisu verbs. As a result, Lisu verbs include more specific information about selectional properties, notably animacy, than their English counterparts. For example, the verb thywu means 'burn an inanimate object' and the verb sye 'kill' obligatorily co-occurs with the noun yi-p ' 'an end', but need not occur with a patient argument. As the example in (34) shows, although the meaning of the verb is highly specific and likely delimited, the verb does not require a patient argument as its direct object. Thus, it appears that event information is encoded in the lexicon, but not in the syntax. (34)
ha nya yi-p3 syk - a asa TOPIC end kill - DECLARATIVE 'Asa killed and an end resulted.'
4.3. The Syntax of Topic-Prominent Languages Rizzi (1997) proposes that the complementizer layer of a clause consists of several distinct projections, including a Topic Phrase (TopP). TopP is optional, occurring only when needed, and its head, Top, checks no phi or Case features.
According to Rizzi, topic-comment structure is simply a kind of "higher predication," a predication within the Comp system; Top lacks inherent semantic content. Some consequences of Rizzi's theory of topic are that not all clauses contain TopP, topics must mergelmove to Spec, TopP before spell-out, and there are no topic splits parallel to subject and object splits. Because topic-prominent languages are discourse oriented, we suggest that they have no event feature to be checked and no phi features to be checked. If this is correct, then a topic-prominent language should have no subject requirement. We should find topic prominent languages with no head marking (due to the lack of phi features), and with no grarnmaticization of the event through either the subject or direct object. In particular, if topic involves movement or merger into the Spec of either CP or TopP, then its head, C or Top, has no content. The relevant Spec must
be filled in order to identify the functional projection (potentially an EPP requirement on this position). This results in (i) the requirement that every sentence have a topic, (ii) the lack of thematic restriction on the topic, (iii) the lack of restriction on the number of topics a sentence can have, and (iv) the lack of any agreement-type features on the topic. Notice that in Lisu no event information is encoded in the syntax, i.e., Tense and Asp have neither spatio-temporal nor Caselagreement features. The language has no tense or aspect marking on the verb, no Case marking on DP or agreement marking on the verb, and only lexical aspectual distinctions. In the absence of any event information, the organization of the clause is determined almost exclusively by discourse considerations (note that the thematic hierarchy plays a marginal role). A topic is obligatory in Lisu because tense and aspect are completely laclung in inherent content, and Top (not Tense or Asp) has an EPP feature.
5. DISCOURSE ORIENTATION VS. EVENT ORIENTATION To conclude, we have argued that some languages organize their arguments according to the type of event denoted by the clause. Some of these languages may grarnmaticize delimitation, while others grammaticize agentivity. We hypothesize here that V-related projections are event related. That is, the inflectional projections that license subjects and objects are event-related. AspIAgr-OP licenses objects by checking Case and agreement; we argue that it also contains a quantization feature that checks quantized objects. In some languages the quantization feature is an aspectual feature and will only check quantized objects of delimited events. TP licenses subjects by checking Case and agreement; we argue that in some languages only animate or agentive subjects have agreement features that are checked in Spec, TP. Other subjects must check their features (e.g. gender features) elsewhere. Asp and T are V-related projections in thesense of Grimshaw (1991) and Rizzi (1997). On the other hand, complementizer projections that license topics are discourserelated, not event-related. As Rizzi (1997) points out, elements in the complementizer layer do not bear agreement or Case marking. The complementizer layer is not part of the extended projection of VP (in the sense of Grimshaw (1991)). Thus, the complementizer layer is not V-related (Rizzi, 1997). Highly topic prominent languages organize their arguments around the discourse notions of topic and comment. In such languages there are no argument splits, and the arguments bear no Case or agreement marking. Any similarities between topics in topicprominent languages and subjects in event-prominent languages can be attributed to the fact that both enter into a kind of predication relation (cf. Chao, 1968; Rizzi, 1997). Ritter, University of Calgary Rosen, University of Kansas
NOTES *
The order of the authors' names is alphabetical; this work represents equal participation of both authors. We thank Paula Kempchinsky and Roumyana Slabavoka and the participants in the Workshop on Events at the University of Iowa. We also thank Mary Linn for her Euchee data and theoretical insights.
'
Interestingly, the BA objects in Mandarin need not be direct objects. They may include part whole relations and even adjuncts of various sorts (Yan Ling, personal communication). Ta ba ge baba si-le. he BA CL dad die-ASP 'His father died.' (implies that he has some responsibility)
(i)
Further, specific objects of delimited events need not enter into the BA construction, as the grammaticality of (8a) indicates It should be noted, however, that (8a) need not be interpreted as delimited. This sentence is compatible with an interpretation in which Zhangsan does not in fact die. Thanks to Liina Pylkkanen (personal communication) for the data in (10) and (1 1). Even object shift in English appears to be weakly conditioned by definiteness. Lasnik (2001) argues, following Johnson (1991), that verb particle constructions are a reflex of object shift. He argues that when the object appears between the verb and the particle, this is due to raising of the object to Spec, AspP and the verb to v. (Lasnik argues that Asp is actually Agr-0.) An interesting fact about particle shift is that it is required of pronouns. Pronouns are highest on the continuum of definiteness. (i) (ii)
a. b. a. b.
*
John looked up the number. John looked the number up. John looked up it. John looked it up.
DP inside VP DP raised to AspP * Pronoun inside VP Pronoun raised to AspP
In addition, it is a well-known fact that extraction from an indefinite DP is better than from a definite DP. One might conclude that extraction out of a definite DP in Spec, AspP creates a CED violation (Lasnik, 2001), while the indefinite DP remains in situ where extraction is allowed. (iii)
a. b. c.
What did you see pictures of? ? What did you see the pictures, of!
* What did you see Bill's
pictures of!
See also the description in Silverstein (1976). The idea that 3'* person DPs lack a person feature is not a new one. Such a proposal was originally made by Benveniste (1971) and subsequently adopted by Harley & Ritter (2002), Uriagereka (1995), and others. Dowty also considers the following properties as contributing to agent proto-roles: (i) movement relative to the position of another participant, (ii) having an existence independent of the event (i.e. not created by the event), and (iii) causing the event. These are also characteristic of animate agents, although they may not pick out animates to the exclusion of all inanimates. According to Rice (2000) Slave (Athapaskan) shows the same person split as Euchee. Although our data do not show this, the verb agrees with a 3rdperson subject in gender (Linn, 2001). There exists a basic distinction between stative and non-stative or active situation types, and among the non-stative situation types there is a subdivision between those that involve a process (and thereby require an agent to perform, effect or instigate the process), and those that do not. The former class consists of activities and accomplishments. Thus, both Euchee and Lakhota have event-based subject splits, although the determining factors in each language are slightly different, and the splits are subtly different: In Euchee a non-agentive subject of an active predicate patterns with an agentive one, whereas in Lakhota the two kinds of subjects pattern differently.
'
HAGIT BORER
SOME NOTES ON THE SYNTAX OF QUANTITY
1. INTRODUCTION It has been suggested by Verkuyl (1972, 1989, 1999) that the presence of a direct argument with some specific properties (for Verkuyl, op. cit, [+SQA] 'specified quantity of A') is essential for the derivation of a telic interpretation, an assumption that has proven very influential in subsequent treatment of aktionsart (see Platzack, 1979; Tenny, 1987; Dowty, 1991;. Krifka, 1992, 1998; among many others). Wishing to abstract away, for the moment, from the debate concerning the properties of the argument in question (and the accuracy of the generalization in general), suppose we put together the broad description in (1), referring to it, without prejudice, as Verkuyl's generalization: Verkuyl's Generalization: Telic interpretation can only emerge in the context of a direct argument with property α
(1)
To illustrate, the direct arguments in (2) are usually assumed to have property a, thereby giving rise to a possible telic interpretation, but not so the direct arguments in (3): (2)
a. b. c. d.
Kim ate some apples. Pat drank too much beer. Robin read two books. Marcia built a house.
(3)
a. b.
Kim ate apples. Pat drank beer.
Although the nature of property a is generally sought within the domain of quantification, its precise characterization is not agreed upon, a matter to which I turn shortly. Suppose we adopt in this context the prevailing view that the telicityatelicity distinction is to be viewed as equivalent to the semantic distinctions which are attested within the nominal domain, and specifically, to the distinction between quantity nominals and non-quantity nominals (cf. Bach, 1986; Krifka, 1989, 1992). Thus in some sense to be made precise, telic events are quantities, while atelic events are not. In turn, if the property quantity, however refined, is to be represented
41 P. Kempchinsky and R. Slabakova (eds.), Aspectual Inquiriess, 41-68. © 2005 Springer. Printed in the Netherlands.
within the nominal domain as a specific syntactic node, then a principled approach to the syntax-semantics interface would require that within the domain of events as well, the property quantity must be represented as a specific syntactic node. The purpose of this article is to elaborate on this parallelism between the nominal domain and the event domain and to pursue its ramifications, with the aim of shedding light on the constituent structure of both nominals and events. To do so, we must commence with a discussion of the structure of quantity nominals, where the existence of a specific syntactic node devoted to quantity is less controversial, although, as we shall see, its properties and its interaction with nominal interpretation is not self evident. 2. QUANTITY IN NOMINALS Largely due to work by Ritter (1991, 1995) and others, it is now quite accepted that there is, within nominals, a functional projection dedicated specifically to quantity or number interpretation and which we shall call Quantity Phrase (#P). For the nominals in (2), then, a preliminary representation would be as in (4) (the nature and the status of the DP projection in nominals such as those in (4) is set aside for expository purposes):
An immediate question which arises with respect to (4) concerns the plural marker -s in (4a,c). Usually, the plurality marker is considered a species of Number. Under such an analysis, -s (or PL) would be the head of #P, triggering head movement (overtly or covertly) while the quantity expressions in (4a,c) would, presumably, be specifiers. The resulting (hypothesized) structure would be as in (5) (checking derivation assumed for expository purposes):
However, if the -s is a Number head, this means that bare plurals, as in (3a), llkewise must include a Quantity Phrase, yielding the structure in (6) (presumably, in English, with covert N movement):
In turn, for the bare mass noun in (3b), there is no prima facie reason to postulate an #P, and the simplest representation for it would be as in (7):
SOMENOTESON THE SYNTAX OF QUANTITY
43
A clear difficulty now emerges concerning the syntactic characterization of the notion quantity, such that it renders the nominal in (2) quantity, but not so the nominal in (3a). Nor can the notion of head vs. specifier be appealed to successfully to give a common syntactic description to (2) and (3a) (i.e., postulating that a nominal is quantity in the relevant syntactic sense only if it has a quantity specifier), as at least a in a house is clearly a head, and not a specifier, and yet a house is a quantity expression in the required sense. It would thus appear that if structures (6)-(7) are adopted, then the distinction between quantity nominals and nonquantity nominals must be exclusively semantic, rather than syntactic in nature, and specifically, one would have to claim that two houses, or a house are in some semantically relevant sense quantity but houses is not, although their syntactic structure is, for all intents and purposes, identical. The undesirability of lifting the relevant distinction from the realm of the syntax to the realm of the semantics, we note, is couched within a particular view of the syntax-semantics interface, and may not be viewed by some as problematic. However, as we shall see shortly, shifting the distinction to the semantics likewise fails to return correct results. In turn, once the semantic analysis is modified, it becomes possible to return to the syntax, and give a distinct syntactic structure to quantity and non-quantity nominals, thereby rescuing, within this domain, a stricter view of the syntax-semantics interface. Within join, semi-lattice type of approaches to plurality and mass interpretation (cf. Link, 1983; Bach, 1986; Krifka, 1989; among others), the affinity between bare plurals and mass nouns is blamed on the fact that the interpretation of plurals is vague in the following sense. Suppose the denotation of plurals consists of all or any subsets of singularities, as the diagram in (8) shows:
{a, b, c ) {a, b, d ) {b, c, d ) {a, c, d ) ... {a, b } {a, c } {a, d } a b c d ....
{b, C )
ib. d }
ic, d ) ... = Atoms
Suppose now that we talk about circles. The number of atoms implicated in circles remains entirely vague, as it can consist of any of the sets above the atom line in (8). Similarly, for mass nouns, any reference to, e.g., sand suffers from similar vagueness, as the actual amount of sand involved is under-determined by the utterance. It is on that level, then, that the commonality of bare mass nouns and bare plurals can be defined so as to account for their similar properties. There are, however, a number of problems with the join semi-lattice approach to plurality, based as it is on the assumption that plurals are a function from singulars. First, note, the interpretation of the bare plural does not actually consist of any subset of (well-defined) singulars. Specifically, if Kim ate a portion (potentially of a different size) from a number of apples, although we are justified in saying that she ate apples, the set of elements that she ate does not consist of singular apples, but
rather, of apple portions. More devastatingly, consider the following paradigm:
(10)
zero apples/*apple
Zero apples or 0.5 apples do not presuppose the existence of singulars although they occur with plural marking, and hence are not well-defined sets within the join semilattice representation in (8). An additional set of problems, often pointed out in conjunction with telicity, is associated specifically with the quantification in (1 lb-c). (1 1) a. b. c.
Kim ate three apples (in an hour). Pat built more than three houses (in two months). My lud sister drew some circles (in half an hour).
Although an expression such as more than three narrows somewhat the possible interpretation of plurals (e.g., more than three cuts off the bottom two, non atomic, lines in (8)) one would be somewhat hard pressed to suggest that this narrowing down suffices to make the vagueness of the plural interpretation disappear, so as to allow for the appropriate contrast to be drawn between the bare plural apples and more than three apples. Likewise some, which potentially refers to any of the nonatomic sets in (8), nevertheless triggers a telic interpretation, unlike a bare plural. The problem extends to the domain of mass nouns as well. Specifically, how much is too much? How much is more than enough (assuming that much and enough are well-defined quantities in a given context)?' (12) a. b. c.
Kim ate more than enough meat (in an hour). Pat built most furniture (in two months). Robin sifted (too) much sand (in half an hour).
Viewed differently, however, note that although expressions such as more than three circles or some circles do not resolve the vagueness associated with the interpretation of bare plurals, they do have an interpretation which is radically different from that of bare plurals. If Kim drew more than three circles, then there are at least three individual circles such that Kim drew them. The truth conditions of more than three circles could not, in fact, be computed without assuming individual circles. Likewise, if Kim drew some circles, then there are at least two individual circles such that Kim drew them, and the sentence cannot be true if there is only an assortment of incomplete arches on the page, a situation which would render drew circles Cfor an hour) true. In each case, the computation of the meaning does appear to be presupposing individuals, in a way which the interpretation of bare plurals
SOMENOTESON THE SYNTAX OF QUANTITY
45
does not. We submit, then, that individuals cannot be created by the plural inflection and the plurality marker does not imply the existence of a set of singulars. Thus, the plural inflection can occur without such individuals. In turn, individuals within a plural set emerge as a result of the presence of a quantity expression distinct from the plurality marker. But if this is the case, then perhaps there is much to be gained from assuming that plurality markers are not quantity heads, but rather, fulfill a different function in the grammar. I have argued elsewhere that plural markers are, in actuality, classifiers, and that like other classifiers, their main function is divisional. I will set aside the detailed argumentation for that position, noting only the following paradigm, from Armenian (pointed out to me by M. Siegler): (13) a.
Numeral, no classifier, no plural yergu hovanoc uni-m two umbrella have- 1s 'I have two umbrellas.'
b.
Numeral, classifier, no plural yergu had hovanoc uni-m two CL umbrella have-1s 'I have two umbrellas.'
c.
Numeral, no classifier, plural yergu hovanoc-ner uni-m two umbrella-PL have- 1s 'I have two umbrellas. '
d.
* Numeral, classifier, plural * yergu
had hovanoc-ner uni-m two CL umbrella-PL have-1s 'I have two umbrellas. '
Setting aside here the case in (13a), in whlch a numeral can occur without either a classifier or plural marking (and likewise, the complementary distribution of numerals and plural markers in Hungarian and in Turkish), and focusing on (l3b-d), we note the complementary distribution of numerals and classifiers, which may never occur together. A similar situation is attested in Chinese, if we follow Li (1998) in assuming that -men in Chinese is a plurality marker (and see also Li and Thompson, 1981), noting specifically that it never co-occurs with a classifier. An explanation for this complementary distribution as well as the non-quantity properties of the plural marker in a language such as English follow directly if we assume that so-called plural inflection in actuality reflects the presence of a classifier phrase, and not that of an #P. That classifier fulfills a divisional function, rather than a quantity function. Consider now a specific execution of the creation of individuals by the counting
function. According to this execution, all noun denotations are mass. More specifically, we may assume that a mass denotation emerges whenever a noun denotation is not grammatically otherwise specified (which is to say, whenever it fails to be associated with some non-mass functional structure). The divisional function (associated with the classifier, or more accurately, a divisional head) involves the superimposition, on a mass denotation, of an infinite set of webs, or reticules (including, potentially, a reticule without any divisions, reticules without complete cells, or reticules which create cells that do not correspond to a canonical singular). The function of #P, on the other hand, is that of assigning quantity, or in the presence of a divisional structure, that of counting. More specifically, it involves the selection, among the reticules, of one which matches the properties of the specific For example, for a cardinal such as three it involves the selection of a reticule with three cells, to which a uniform extension, presumably that associated with the conceptual meaning of the relevant N, can be applied. For zero, on the other hand, it will involve the selection of a reticule without any (completed) cells. For more than three, all reticules which include at least three complete cells will be selected, etc. For 0.5, a reticule will be selected which involves a portion of the mass which does not correspond to a canonical singular. Of special interest is some when it occurs with a divisional structure (i.e., with plural inflection), where, I suggest, the # function will be equivalent to that of any unspecified number bigger than one. Some, then, could choose any (or all) reticules in which there are at least two cell divisions, thereby giving rise to some unspecified number of individuals, but individuals nevertheless. Interestingly, in Romance languages, in which bare plurals exist alongside plurals with indefinite (plural) articles, we find the contrast in (14): (14) a. b.
Juana comiomanzanas (*en media horaldurante media hora). 'Juana ate apples (*in half an hourlfor half an hour).' Juana comi6 unas manzanas (en media hora/?durante media hora). 'Juana ate ART.PL apples (in half an hour/?for half an hour).'
We note that there is no sense in which the quantity of manzanas 'apples' in (14a) could possibly be assumed to be less well defined from that in (14b). The difference between (14a) and (14b), we submit, is in the function of the plural indefinite article. It is, we propose, a counter, although its cardinality remains undetermined, and as such, it performs a similar function to that of some or several. Syntactically, it licenses #P. Semantically, it selects those reticules in which there are at least two cell divisions, thereby giving rise, just like some, to an unspecified number of individuals, but individuals nevertheless. We conclude that from a purely semantic perspective, plurality must be treated as different from quantity, thereby enabling us not only to resolve the problems for the semantic description of quantity, but also to resurrect the syntactic distinction between bare nouns, be they mass or plural, and quantity norninals. Specifically, the syntactic generalization is that whenever #P is assigned some value, telicity may be licensed by the emerging nominal. The structures in (6)-(7) must be rejected, then,
SOMENOTESON THE SYNTAX OF QUANTITY
47
and an alternative proposed structure for the nominals in (2)-(3) should thus be as in (15). We may now state the property quantity in syntactic (as well as semantic) terms, as that which involves the projection of #P:
3. THE INTERPRETATION OF QUANTITY We must now determine why the existence of individuals within, e.g. some apples, as formed by the #P function, should give rise to telicity, while apples, which lacks such individuals within its denotation, should fail to do so. To see what effect the 'formation of individuals' may have, it might be worth while to consider some of the formal treatments of nominals and telicity, reviewing some of their merits and demerits. Such a review in fact indicates that the claim that apples does not consist of singular apples, but more than three apples does, can resolve a number of rather tenacious problems for the interaction between telicity and nominal structure. Consider, again, the determiners in ( i l ) , which give rise to (optional) telicity, when compared with the absence of telicity in (3). Krifka (1989, 1992, 1998), in attempting to characterize the class of nominals which give rise to a telic interpretation, proposes that such nominals must be quantized. Quantized, in turn, contrasts with cumulative, a property which is attributed to bare plurals and bare mass nouns, and which gives rise to atelicity. The definitions are given in (16): (16) a.
b.
Cumulative (Krifka, 1998) Xis cumulative iff there exist y, x with the property X (and x distinct from y) such that for all x and y, if x,y have the property X, then X is a property of the sum of x and y (bX _c UP[CUMP(X) *3,y[X(x)fi(V)~ ~ = y Al E7y[X(x)4.i) + X(x WII) Quantized (Krifka, 1998) Xis quantized iff for all x,y with the property X, y is not a proper part of x &Y[X(X)~(V) + TY< ~~11) (bX _c Up[eUAP(X)
*
Bare plurals and bare mass nouns are clearly cumulative in the required sense (apples plus apples gives apples; meat plus meat gives meat), as well as nonquantized, given the fact that it is certainly not the case that all subparts of meat are not meat, or that all subparts of apples are not apples. Likewise, cardinals are clearly quantized, in that no part of three apples is three apples, etc. As often observed, however, some as well as more than three both trigger telicity, although by the
definitions in (16 ) they are cumulative, rather than quantized. A somewhat different approach is put forth in Kiparsky (1998). Kiparsky suggests that the key notion for the emergence of quantity interpretation (boundedness, in his terms) involves the notions divisiveness, cumulativity and diversity, as defined in (17):~ (17) a.
b.
c.
(1 8)
P is divisive if and only if for all x with property P,where x is nonatomic, there is a y, proper subset of x with the property P. (Pis divisive i f f Yx [P(x) A latom(x) +i$[y c x A P(y)]]) P is cumulative if and only iff for all x with property P, where x is not the maximal element with property P, there is a y, proper superset of x with the property P. (P is cumulative iff bk [P(x) A +up(x,P) +3(y)(x cynP(y)]) P is diverse if and only iff for all x with the property P and ally with the property P, and x distinct from y, x is not a proper subset of y and y is not a proper subset of x. (Pis diverse i f f Vx Yy[P(x) A P(y) A x # y + 3 c y A ~y cx]) A predicate P is unbounded (-B) iff it is divisive and cumulative and not diverse.
Cumulative reference requires, in essence, that every proper superset of x within the denotation of P be within the denotation of P. Divisive reference requires, in essence, the existence of some subset of x within the denotation of P to be within the denotation of P. In turn, if bounded is true of any predicate which is not unbounded, then in order for a bounded reading to emerge either divisive reference or cumulative reference must fail (or alternatively, the predicate must be diverse). The crucial contribution of Kiparsky's approach, which I will adopt below, is to set up conditions for the failure of non-quantity, so to speak, rather than setting up conditions for the failure of quantity. As such, it provides a solution for at least one class of cases which are a problem for Krifka's quantization approach, namely, cases such as at least three apples. Specifically, at least three apples is cumulative, but it is not divisive (i.e., there need not be a proper subset of at least three which is at least three), and hence is correctly predicted to give rise to telicity (or boundedness). We note however that for the bulk of the cases which are problematic for Krifka, the definitions in (17)-(18) do not provide a solution. Specifically, some apples, several apples, many apples etc. continue to be unbounded, as they are both cumulative and divisive (and non-diverse). A few more comments are in order here concerning cases such as more than three apples, several apples, some apples and similar examples. For Krifka, they were cumulative, and in fact, non-quantized, as it is certainly not the case that every proper subset of more than three apples must be distinct from more than three apples. For Kiparsky's definition they continue to be cumulative, of course. Are they divisive, however? Let us consider more carellly the possible divisive properties of, e.g., more than three apples. If the set under consideration involves
SOMENOTESON THE SYNTAX OF QUANTITY
49
five or more apples, more than three is clearly divisive, as by definition, any set which is five or more has a subset, i.e., four apples, which is more than three, thereby passing the divisiveness test. What, however, of a set of four apples? Here, divisiveness fails (thereby making the entire more than three predicate non-divisive, as required), but only if we consider whole apples, exclusively. As a set of four whole apples does not have a subset which is more than three whole apples, it is non-divisive. If portions of apples are considered, on the other hand, than a set of four apples is divisive, with respect to more than three apples, as is any quantity of apples between three and four, under infinite division. A similar rationale applies to several apples and some apples, where, arguably, a set of two apples is not divisive as it has no subparts which are several or some whole apples. One could propose, at this point, that only whole singular apples count, referring specifically to the denotation of plurals as sets of singulars. Under such an interpretation, note, more than three apples as well as several apples and some apples would become non-divisive, and hence bounded, as required, thereby solving the problem. However, the solution here is only apparent. Thus consider apples, under a scenario that includes two apples. If plural refers exclusively to sets of whole singulars, then apples, when referring to two apples, does not have any subparts which are apples, and we predict, erroneously, that apples is non-divisive. If, on the other hand, non-whole parts of apples do count, apples will be divisive, correctly, but so would more than three apples, some apples and several apples erroneously. It thus appears that even if something like the definitions in (17) could be made to work, there must be a distinction drawn, in principle, between the sort of plurality that is referred to by expressions such as more than three apples, and which seems to apply to sets of singulars, and the sort of plurality that is referred to by expressions such as apples, in short, precisely the distinction we already suggested must exist, making more than three apples a true case of sets of singulars, but not so apples. Seeking, however, to address specifically the cases which are resolved neither by Krifka's definition nor by that put forth by Kiparsky (1998), suppose we adopt a slightly modified definition of homogeneity as proposed in Kiparsky (1996) but abandoned in Kiparsky (1998), and which is based, specifically, on (a slightly modified) definition of divisiveness originally proposed in Krifka (1992): (19) a. b.
quantity: P is quantity iff P is not homogenous. P is homogeneous iff P is cumulative and divisive. i. P is divisive iff for all x with property P there is a y, proper subset of x, with property P,such that subtracting y from x yields a set with the property P. P is divisive fz bk [P(x) +3, (Ph)A y Perfect Aspect: ]-[AST-T/F]-> AST-T AFTER EV-T
The proposal in (1) allows us to explain the diversity of predicates that surface crosslinguistically to express temporal and aspectual relations. In particular, (la) immediately explains why predicates expressing central coincidence between the Figure and the Ground-e.g. pre/postpositions such as in, on, at, as well as verbs of location, stance or posture such as stay, sit, lie down, live, or even verbs of nondirected motion such as walk-are systematically used to form progressive sentences across languages. (lb) straightforwardly explains why predicates expressing [-central, +centripetal coincidence] between F and *typically, verbs of cross-linguistically used to express the motion towards G such as go-are Prospective/Future. Finally, (lc) explains why predicates of [-central +centrifugal coincidence] between F and *that is, prepositions such as after or verbs of motion from G such as come from or throw away-are used cross-linguistically to express the PerfectTast (see D&UE, 2000,2004b, and references therein). We capture the semantic parallel between Tense and Aspect syntactically by proposing a uniform phrase structure for temporal and aspectual relations. The idea that Tense relates two times has been captured in syntactic terms by breaking down Tense syntactically into its semantic components. For Zagona (1990), Tense is a head that projects a maximal projection taking two time-denoting arguments. Its external argument is a Reference Time (REF-T)-typically, the UT-T. Its internal argument is the EV-T. Stowell (1993)extends this proposal and argues that Tense is a two-place predicate of temporal ordering. D&UE propose that Aspect, just like Tense, is a dyadic predicate taking time-denoting phrases as arguments, and projecting its temporal structure in the syntax. The external argument of ASP0 is a reference time (the AST-T),its internal argument is the time of the event denoted by the VP (the EV-T). Likewise, the external argument of T" is a reference time (typically, the UT-T),its internal argument is the AST-T. Our proposal is given in (2). The Phrase Structure of Tense and Aspect (D&UE, 1997, 2000) TP
UT-T, AST-T, EV-T are time-denotinglzeit phrases
/", UT-T
T'
A
To
ASP-P
A AST-T
ASP'
A
ASPo
VP
A EV-T
. VP
ASPECTAND TEMPORAL MODIFICATION
193
In the following subsections, we briefly illustrate our proposal with the analysis of two aspects: the Progressive and the Perfect. We then extend this proposal to time adverbs by arguing that time adverbs are PPs headed by a spatiotemporal predicate establishing an ordering relation between two time arguments (section 2). With this proposal, we seek to reduce the syntax and semantics of tenses, aspects and time adverbs to the same set of substantive and structural primitives.
1. I . Progressive Aspect Progressive Aspect is the spatiotemporal predicate WITHIN. Progressive Aspect thus acts like a Present Tense: both are predicates with the meaning of WITHIN. We illustrate the phrase structure of present and past progressive sentences in (3).
The present progressive
(3) a. a'.
Maddi is drawing a house.
The past progressive b.
Maddi was drawing a house.
b.'.
TP
TP
A
A UT-T
UT-T
T'
To ASP-P AFTER
ASP-P
To
WITHIN
T'
A
A
A
A
ASP'
AST-T
AST-T
ASP'
ASP'
VP
VP
WITHIN
WITHIN EV-T
ASP'
A
A
EV-T
VP
VP
In (3a3, Progressive Aspect orders the AST-T WITHIN the EV-T. It thus picks out a time within the interval defined by the EV-T,as illustrated by the temporal schema in (4a). Present Tense then orders the UT-T within the AST-T, as shown in (4b). (3a) thus focuses a subinterval of the EV-T which itself overlaps with the UT-T. UT-T
EV-T
(4) a.
b.
-[----[----I-]->
--[-I-]->
AST-T
AST-T
EV-T
(5) a.
UT-T
b.
-[-[----]----I-> AST-T
-[]-I> AST-T
The past progressive sentence (3b) has exactly the same phrase structure except that Tense is the spatiotemporal predicate AFTER. Progressive orders the AST-T WITHIN the EV-T, as in (5a). Past then orders the UT-T AFTER the AST-T. Note that since the AST-T denotes a proper subpart of the EV-T, whch includes neither its initial nor its final bound, T" orders the uT-T after a subinterval of the EV-T (=AST-T),as shown
in (5b). The UT-T is therefore ordered with respect to the AST-T, but remains unordered with respect to the final bound of the event. Thus, the described event might or might not have culminated before UT-T, even when the sentence is in the past. To recapitulate, the role of grammatical Aspect is to convey a viewpoint on the situation described by a sentence-that is, a temporal perspective that focuses all or parts of the described situation (Smith, 1991). Why does aspect focus (pick out) a time interval in the temporal contour of the event described by a sentence? Aspect focuses a time interval because Aspect (just like Tense) is a spatiotemporal predicate establishing a topological relation between two time arguments. Progressive aspect focuses a subinterval of the EV-T because it is a predicate of central coincidence ordering the ~ i g u r e(the AST-T, the time for which the speaker makes a statement) WITHIN the cround (the EV-T).' 1.2. Perfect Aspect
We analyze the Perfect (Have V-ed) as the spatiotemporal predicate AFTER. The Perfect thus acts like a Past Tense: both Past and Perfect are spatiotemporal predicates with the meaning of AFTER. We illustrate the phrase structure of present and past perfect sentences in (6).
The present perfect
(6) a.
Maddi has drawn a house.
a'.
b.
Maddi had drawn a house.
b'.
TP
A UT-T
The past perfect
TP
A
T'
UT-T
A To
TO
ASP-P
WITHIN
T'
A A
AST-T
ASP-P
AFTER
ASP'
AST-T
A
A VP
ASPO
AFTER
ASP'
A
EV-T
VP
ASP'
VP
AFTER
A
EV-T
VP
In (6a'), Perfect orders the AST-T AFTER the EV-T. It thus picks out a time after the interval defined by the EV-T,as illustrated by the temporal schema in (7a). Present is the spatiotemporal predicate WITHIN. It orders the UT-T WITHIN the AST-T, as shown in (7b). The event of drawing a house is viewed as completed prior to a present reference time because the Perfect orders a reference time (our AST-T), itself overlapping with UT-T,after the time defining the final bound of the event. EV-T
(7) a.
-[---I-
EV-T
[-I-> AST-T
b.
UT-T
-[-]-[-I-]-> AST-T
ASPECTAND TEMPORAL MODIFICATION
195
The Past Perfect in (6b) has the same phrase structure, except that Tense is the spatiotemporal predicate AFTER. Perfect Aspect orders the AST-T AFTER the EV-T, as represented in (8a). Past Tense then orders the UT-T AFTER the AST-T, as in (8b). Since the UT-T follows the AST-T, which itself follows the EV-T, the event of drawing a house is viewed as completed prior to a past reference time (REF-T= ASTT).* EV-T
(8) a.
UT-T
EV-T
b.
-[-]-[----I->
-[---]---[---]-I->
AST-T
AST-T
We now turn to the grammar of simple tenses. 1.3. Tense without Morphological Aspect To derive the grammar of temporal relations in systems without either morphological Tense or morphological Aspect, we make the following two assumptions: i) TP and ASP-P are always projected; ii) when either To or ASP0 is not morphologically overt, its external argument binds its internal argument. This proposal yields the phrase structure in (9) for simple past and simple f ~ t u r e . ~ (9) a.
Future Tense without Aspect
b.
T'
UT-T
Past Tense without Aspect
T'
UT-T
A
A To
To ASP-P BEFORE
ASP-P
AFTER
In (!la), T" orders the UT-T BEFORE the AST-T. Since ASP0 has no morphological content, the AST-T binds the EV-T. The EV-T is thus indirectly ordered in the future with respect to the UT-T, yielding the simple Future Tense interpretation in (10). UT-T
( 10)
AST-T
AST-T
(1 1)
-[---]---[----I-> EV-T
UT-T
-[-]-[---I-> EV-T
Turning to (9b), TO orders the UT-T AFTER the AST-T. ASP0 having no morphological content, the AST-T binds the EV-T. The EV-T is thus indirectly ordered in the past with respect to the UT-T, yielding the simple Past Tense interpretation in (1 1). This explains why the event described by a simple past tense is portrayed in its entirety-
so called Perfective or Aoristic Aspect. The event is viewed as including both its initial and final bounds because the AST-T coincides temporally with the EV-T. In sum, the role of ASP' is to overtly specify the relation holding between the EV-T and the time the for which an assertion is made. If there is no morphological encoding of ASP', the relation between the AST-T and the EV-T is established anaphorically, via binding. Binding yields an ordering of exhaustive coincidence: the two intervals are cotemporal. Within a model where time spans are discourse referents projected in the syntax as temporal DPsIarguments, we expect that they can enter into relations of coreference or disjoint reference-just as any other DP can. 2. TIME ADVERBS AS MODIFIERS OF THE SPATIOTEMPORAL ARGUMENTS PROJECTED BY ASP0 We have assumed, following Zagona (1990) and Stowell (1993), that time arguments are represented in the syntax as temporal DPs or Zeit phrases. The null assumption is then that they can be modified, just like any DP can. This, we argue, is precisely the role of time adverbs. Time adverbs are the semantic and syntactic modifiers of Zeit phrases projected in the syntax as arguments of ASP0 (see D&UE, 2002, 2004a). We first incorporate prepositional time adverbs into our model. We then extend our analysis to bare adverbs and subordinate adverbial clauses.
2.1. Locating PP Time Adverbs Consider the following examples: (12) a. b.
Layla moved to Ethiopia in 1999. Layla moved to Ethiopia afterhefore Easter.
The simple past sentence Laila moved to Ethiopia describes an eventuality that occurred at a past time. The PPs in 1999 or aftedbefore Easter serve to restrict the reference of this past event time by ordering it (WITH)IN or AFTER~BEFOREthe time denoted by their internal argument (1999/Easter). Locating adverbs are thus semantically restrictive modifiers of either the EV-T or the AST-T (cotemporal in (12) since the AST-T binds the EV-T when there is no morphological aspect-cf. (9bD The null assumption then is that locating time adverbs are also syntactically restrictive modifiers of the time spans projected in the syntax as arguments of ASP0 (that is, AST-T/EV-T).Modifiers of nominal arguments are analyzed as adjuncts base-generated adjoined to the NP whose reference they restrict, as shown in (13a). Modification is established via predication: the modifier is predicated of its sister NP. The latter is formally its external argument.4
ASPECT A N D TEMPORAL MODIFICATION
The syntax of nominal and temporal modzfication
(13) a.
b.
NP
A
AST-TIEV-T
A
NP
MODIFIER
book
on the table
MODIFIER
AST-T/EV-T
idbeforelafter 1999
We establish a semantic and syntactic parallel between nominal and temporal modification: modifiers of time arguments are base-generated adjoined to the Zeit phrase whose reference they restrict, as in (l3b). Temporal modification is thus, just like nominal modification, established via predication. The spatiotemporal predicate in/before/after 1999 in (13b) is predicated of the AST-T/EV-T, its external argument.' Turning to the temporal syntax of (12a,b), we note that the prepositions in, before, after are spatiotemporal predicates, modifying a time argument of ASP0 (either the AST-T or the EV-T, cotemporal in (12) since we have simple Past Tense sentences). The syntax of temporal modification in (12a) is given in (14a): the preposition in restricts the reference of the AST-T/EV-T, its external argument, by ordering it WITH)^ the time denoted by its internal argument 1999. The syntax of temporal modification in (12b) is given in (14b): the Po beforelafter restricts the reference of the AST-T/EV-Tby ordering it BEFORE/AFTER the time denoted by Easter. (14) a.
AST-TIEV-T
A AST-T/EV-T
A PO
AST-TIEV-T
A
PP
IN
b.
1999
AST-T/EV-T
PP
A Po EASTER BEFORE/AFTER
The full temporal structure proposed for (12a) is given in (15). The PP in 1999 is predicated of the AST-T. It restricts the reference of the AST-T (its external argument) by ordering this time span WITHIN the time denoted by its internal argument 1999. Since the AST-T is itself cotemporal with the EV-T (perfective aspect), the adverb indirectly serves to locate the EV-T, as illustrated by the temporal schema in (15b-c).
AST-T=EV-T UT-T
UT-T
I---t->
4 T'
A ASP-P
TO
AFTER
AST-Ti
ASP'
A
A
AST-T'EV-T
UT-T
Sentence (12b) will have the same phrase structure except that the PP predicated of the AST-T is before/after Easter. The preposition before orders the AST-T (its external argument) BEFORE the time denoted by its internal argument Easter, whereas the preposition after orders the AST-T AFTER the time denoted by Easter. Since the AST-T is itself cotemporal with the EV-T (perfective aspect), the time adverb indirectly serves to locate the EV-T,as illustrated by the temporal schema in (16a-b). (16) a.
before Easter AST-T=EV-T
-[---]-[-]-I->
b. EASTER
UT-T
after Easter
-[I
EASTER
AST-T=EV-T
]-I>
UT-T
We now turn to cases where AST-T modification and EV-T modification yield distinct construals. We illustrate this case with the analysis of the past perfect sentence in (17a), which has the phrase structure in (l7b). (17b) yields the temporal schema in (17c). Recall (from section 1.2) that the Perfect orders the AST-T AFTER the EV-T, as illustrated in (l7b-c). (17) a.
Abdel had left the house. EV-T
b.
AST-T
UT-T
C.>- I - - ] - - [ - ] - [ -
TP
A UT-T
T'
A To AFTER
ASP-P
ASP'
AST-T
A ASPO
VP
AFTER EV-T
VP
Since the AST-T and the EV-T here are disjoint in reference, adding a PP to (17) will
ASPECTAND TEMPORAL MODIFICATION
199
yield two distinct readings depending on whether the time adverb modifies the EV-T or the AST-T, as illustrated in (18). (18) a. b.
Abdel had left the house at 3 PM. the leaving occurs at 3 PM
c.
the leaving is prior to 3 PM
EV-T
AST-T
A EV-T
A AST-T
PP
PP
A Po
A 3PM
Po AT
AT
3 PM
3PM
3 PM
b'.>- I - - ] - [ - - ] - - [ -
c'. EV-T
AST-T
>- I-]- -[
-[-I--
UT-T
EV-T
AST-T
UT-T
When the PP at 3 PM is predicated of the EV-T, as in (18b,b1), the preposition AT establishes a relation of central coincidence between the EV-T and its internal time argument, thus restricting the reference of the EV-T to the time span designated by 3 PM. This yields the so-called EV-T reading, where Abdel's leaving occurs at 3 PM. When the PP at 3 PM is predicated of the AST-T, as in (18c,c1), the preposition AT establishes a relation of central coincidence between the AST-T and the span designated by 3 PM. This yields the so-called Reference Time reading, where Abdel's leaving occurs prior to a reference time-that is, 3 PM. Our proposal straightforwardly integrates recursive temporal modification. Recursive modification is illustrated in (19) where two PPs have been predicated of the AST-T. The preposition at restricts the reference of the AST-T by relating this time to the time denoted by 5 PM. This ordering relation is one of central coincidence. The preposition after further restricts the reference of the AST-T by establishing a relation of noncentral coincidence between the AST-T and the span designated by dawn. Since the AST-T and the EV-T are cotemporal (perfective aspect), the PPs in (l9a-b) serve to indirectly locate the EV-T, as shown in (19c).~ (19) a.
Maddi was born at 5 AM, after dawn.
b.
AST-T
A AST-T
A AST-T
AST-FEV-T
c.
PP
A PP
A
AFTER
>-I - ]
-[-I-[ DAWN
DAWN
5AM
UT-T
2.2. Bare Locating Time Adverbs We now turn to the temporal syntax of bare time adverbs; that is, time adverbs which do not have the syntax of a PP; but merely of a DP. The time adverbs in (20) are locating adverbs: they specify that the past time of Oihana's arrival is contained WITHIN the span denoted by June 1, 2000/Sunday/yesterday, as illustrated by the temporal schema below: AST-T=EV-T
(20) a. b.
Oihana left June 1, 2000. Olhana left Sundaylyesterday.
c.
-[-[JUNE
UT-T
I-I---P ~,~OOO/YESTERDAY
Past orders the UT-T AFTER the AST-T. The AST-T itself binds the EV-T (perfective aspect). The time adverb then locates' the AST-T (cotemporal with the EV-T)WITHIN the time denoted by June 1, ZOOOIyesterday, as shown in (2 lc). Bare NP adverbs are thus concealed PPs-PPs headed by a silent preposition of central coincidence (see Jespersen, 1929; Kamp & Reyle, 1993). Their phrase structure is given in (21). (2 1)
Bare NP time adverbs
Time adverbs are thus uniformly analyzed as PPs headed by a dyadic predicate of spatiotemporal ordering. Thls predicate can be either overt or null. Silent spatiotemporal predicates express central c~incidence.~ Our proposal extends to all locating adverbs whether they have the syntax of PPs, DPs, or CPS.' The syntax of clausal time adjuncts will be discussed in section 2.5 below. (22)
Locating Time Adverbs PP: at Christmas, in 2000, before Sunday, on Easter Monday DP: Sunday, the lothof June 2001, t h s morning, tomorrow CP: afterhefore Zazy left, when Zoey arrived
2.3. Durational Adverbs We extend our analysis of locational adverbs to durational adverbs. As shown in (23), these adverbs also have the syntax of PPs, bare NPs or CPs. We analyze durational adverbs, just like locational adverbs: they are PPs predicated of Zeitphrases projected in the syntax as arguments of ASP0--either the AST-T or the EV-T, as illustrated in (24). The PPs in (24) specify either the durationhemporal size or the
ASPECTAND TEMPORAL MODIFICATION
20 1
boundaries of the AST-T~EV-Tby relating this time span to the time span denoted by their internal argument. (23)
Durational adverbs PP: for three weeks, in an hour, from 3 to 4, until 2001, from 1924 NP: six months (e.g. Lai'la traveled six months.) CP: since Zazy left, until Zoey arrived
In (24a), the spatiotemporal predicate FOR~INmeasures the duration, temporal size by relating it to the time span designated of its external argument, the AST-T/EV-T, by its internal argument 3 hours; this ordering relation is one of central coincidence. In (24b), UNTIL specifies the final bound of the AST-T/EV-T by establishing a relation of noncentral centripetal coincidence between this time (its external argument) and the time denoted by 1924. UNTIL specifies that the locatiodtrajectory of F (ASTT/EV-T)ends at G (1924). UNTIL thus makes visible the final bound of the AST-T/EVT by indicating that it ends at 1924. In contrast, FROM delimits its initial bound by establishing a relation of noncentral centrifugal coincidence between the AST-T/EV-T and the time denoted by 1924. That is, FROM specifies that the locatiodtrajectory of F (AST-T/EV-T) begins at G (1924). FROM thus makes visible the initial bound of the AST-T/EV-T by indicating that it begins at 1924. Finally, in (25), we illustrate the phrase structure for recursive modification by durational time adverbs. (25)
Recursion of Durational Time Adverbs
AST-T/EVT
PP
TO 2001
A FROM
1924
We next turn to certain co-occurrence restrictions with multiple adverbs. 2.4. Syntactic Restrictions on the Co-occurrence of Multiple Adverbs The examples in (26) illustrate distributional restrictions on multiple adverbs discussed by Hornstein (1990, p. 25).
(26) a. Sue left a week ago yesterday. b. Sue left a week ago. c. Yesterday, Sue left a week ago. d. * A week ago, Sue left yesterday. Since ago is etymologically derived from [(intensifier) + gan 'to go'] (Morris, 1978), we analyze ago as a predicate of noncentral centripetal coincidence, indicating that the location/trajectory of F is TOWARDS/BEFORE G . This proposal yields the following phrase structures for the complex adverbs a week ago yesterday and a week ago: (27) a. b'
yesterday]] [zE,T-pa week [,, GO/BEFORE [ZEIT-P [ZEIT-p a week [pp iZE1T-p PROa~bll
PRO,,^ = UT-T)
In (27b), the internal argument of GO~BEFOREis an implicithull REF-T.We extend Stowell (1993) and assume that this null REF-T argument is a Zeit-P analog of PRO. When PRO is not controlled (has no c-commanding antecedent), as is the case in (26b/27b), it denotes the uT-T. (27b) is thus interpreted as 'John left a week before the UT-T. ' When PRO has a c-commanding antecedent, as is the case in (26c), it gets controlled (bound) by the time argument c-commanding it, as shown in (28). (28) yields the interpretation 'John left a week before yesterday'.
Now, in the ungrammatical (26d), PRO is not c-commanded by yesterday. Hence, it cannot be controlled by this adverb and can only denote the UT-T,as shown in (29).
(26d) is thus assigned the uninterpretable structure in (29) (that is, *"A week before now, John left yesterday") and, as such, is ungrammaticallruled out. In contrast, (26c) is licit as it is assigned the structure in (28), which yields a well-formed interpretation ("John left a week before yesterday"). Under this analysis, the contrast between (26a) and (26d) is ultimately structural: control of an implicit reference time requires c-command, satisfied in (26c) but not in (26d). 2.5. Temporal Adverbial Clauses We now turn to the temporal syntax of clausal time adverbs, as illustrated in (30a). (30) a. b.
Josu left [befordafter Marivi arrived]. Josu left [before/aftr Easter].
ASPECTAND TEMPORAL MODIFICATION
203
Syntactically, the adjunct clauses in (30a) are PPs headed by the spatiotemporal predicate: before/after. Moreover, the spatiotemporal predicates in (30a) have exactly the same semantic function as those in (30b). That is, these PPs serve to restrict the past temporal reference of the event described by the clause Josu left. In (30a), the prepositions before/after serve to establish an ordering relation between the event time of the matrix and the event time of the subordinate clause. The past time of Josu's departure is ordered BEFORE/AFTER the past time of Marivi's arrival. Note, however, that whereas the prepositions in (30b) take a time denoting NP as internal argument, those in (30a) take a clause as internal argument (Marivi arrived). Since a clause cannot itself denote a time interval, we analyze the subordinate clauses in (30a) as subordinate relative clauses, restricting the reference of an implicit Zeit-P, as roughly illustrated in (3 1). (3 1) ,[ beforelafter [,,-,
the timei [CpOi ,[ Marivi arrived ti I]]
We conclude that all time adverbs, whether they have the overt syntax of PPs, bare NPs or CPs (see note 8), are PPs headed by a spatiotemporal predicate. They modifylare predicated of a time argument projected by ASP0, as illustrated in (32).
AST-T/EV-T
A
[ZEIT-P EASTER] BEFORE/AFTER [ZEIT-P the time ,,[ at which ,[ Marivi arrived]]] pa
Let's now see how the temporal adjunct clause in (31) is integrated in the temporal structure of the matrix clause Josu left. Recall, first, that in sentences with a simple past, such as (30), we cannot tell whether a temporal adverb modifies the AST-T or the EV-T of the matrix-since these two times are cotemporal (perfective aspect). For concreteness, we assume here that the temporal adjunct clauses in (30a) modify the AST-T of the matrix clause (see D&UE (2004a, 2002) for discussion). Now, the matrix clause Josu left in (30a) has the temporal schema in (33). The AST-T of the matrix (AST-TI)binds the EV-T of the matrix (EV-TI),since we have perfective aspect. Further, Josu's departure culminates at a past time, since the Past Tense of the matrix orders the UT-T AFTER AST-TI, itself cotemporal with EV-TI: (33)
AST-TI= EV-TI ---[
UT-T
]-I-'
The reference of A S T - ~is1 itself fiu-ther restricted by the PP before Marivi arrived in (3 1). The phrase structure for temporal modification of A S T - ~is1 given in (34).
(34) a.
,[ BEFORE (the time at which) Marivi arrived]
ZEIT-Pi
Po
ZEIT-Pi 'AST-T2
BEFORE
c.
UT-T
]--I'
-[
EV-TZ
Marivi's arrival
A UT-T
A
The PP in (34a) is a modifier-that is, it is predicated-of the matrix assertion time AST-T 1, as shown in (34b). BEFORE thus establishes an ordering relation between two times: AST-TIand an implicit time interval, Zeit-P. The CP [Marivi arrive4 has been analyzed as a relative clause restricting the reference time of this Zeit-P. Restrictive temporal modification is established via predication. Predication requires that the head of the relative clause (that is, Zeit-P) be coindexed with the null operator which moves to Spec,CP from a temporal argument position (that is, AST-~2), in order to create a predicate variable. Predication via movement of a null temporal operator in (34b) thus ensures that the internal temporal argument of BEFORE denotes the AST-T of the subordinate clause (Zeit-P = AST-T2),as shown in (34c). Now, A S T - ~is2 itself in the past, since T" in the subordinate clause orders the UT-T after AST-~2. Since, moreover, A S T - ~binds 2 E V - ~ 2(perfective aspect), the internal argument of BEFORE indirectly denotes the time of the event described by the VP MARIVI ARRIVE (Zeit-P = AST-~2 = EV-T2), as illustrated in (34c). That is, BEFORE indirectly orders AST-T1 BEFORE EV-T2. The past event described by the matrix (Josu's departure) is thus ordered before the past event described by the adjunct clause (Marivi's arrival): (35)
AST-TI
1--
-[-I
EV-TI
J's departure
AST-T2
UT-T
'-I] EV-T2
M's arrival
ASPECTAND TEMPORAL MODIFICATION
205
Summarizing, we have proposed that temporal adjunct clauses are PPs headed by a spatiotemporal predicate relating two times: the AST-T of the matrix and another time which itself indirectly denotes (via predication of the adjunct clause mediated by movement of a null temporal operator) the AST-T of the adjunct c l a ~ s e . ~ Note, however, that adjunct clauses expressing central coincidence do not always have the overt syntax of a PP: the English adjunct clause in (36a) is not headed by an overt spatiotemporal preposition, unlike the before/after clauses in (30). (36) a. b.
Josu was leaving when Marivi arrived. Josu was leaving *when Easter.
We analyze the wh-phrase when as a relative temporal operator, that is, as the overt realization of the null temporal operator posited in (34): (37) a. b.
,[ beforelafter [,-, ,[ 0 [,-,
,[ Zoey arrived ti I]]] the timei [aai the timei ,[ wheni ,,[ Zoey arrived ti I]]]
Under this proposal, when clauses are PPs headed by a silent spatiotemporal predicate. That when clauses express central coincidence then follows automatically from the generalization defended in section 2.2: silent spatiotemporal predicates always express central coincidence, whether they take NP arguments as in (20), or CP arguments as in (36a). Now, in contrast to their English counterparts, French temporal adjunct clauses can be overtly headed by a predicate of central coincidence, as shown in (38a). French thus instantiates three distinct but-given the parameters of variation defined here-logically possible structures for adjunct clauses of central coincidence: (38) a.
b.
Overt preposition of central coincidence - null temporal operator 0 [que Franny travaille I]] that Franny works
Zok Zoey
,,[ pendant ,,[ dort sleeps Po
ZoC Zoey
,,[ pendant ,,[ le jour I] dort sleeps Po the day
Nullpreposition of central coincidence - overt temporal operator ZoC Zoe
travaille ,[ 0 ,,[ works
quand when
[Franny dort I]] Franny sleeps
c.
Null preposition of central coincidence - null temporal operator La mort nous prend ,,[ 0,,[ 0 [,o que [nous sommes encore tout pleins de nos miseres et de nos bonnes intentions]]] Literally: 'Death takes. us that we are still full of our miseries and good intentions.' 'Death takes us while we are still full of our miseries and good (Grtvisse, 1980, p. 1333) intentions.'
The proposal defended here establishes a strict parallel between the syntaxsemantics of Tense, Aspect and Time Adverbs. Tenses, aspects and time adverbs are defined in terms a uniform set of substantive primitives: they are spatiotemporal predicates establishing a relation of (non)central coincidence between two timedenoting arguments. This strict semantic parallel is recapitulated in (39).
Tenses, Aspects and Time Adverbs as Spatiotemporal Predicates
(39) a.
[+central coincidence]: F is WITHIN G Tense: Present Aspect: Progressive
-[-[--]-I->
UT-T
AST-T
-
[
AST-T
-[-]I->
1924
[-central +centrifugal coincidence]: F is AFTER G Tense: Past Aspect: Perfect Adverb: afer 1924 AST-T
-[--I-[--]' c.
>
EV-T
AST-T
b.
-
Adverb: in 1924
UT-T
EV-T
-[--]-[-I->
AST-T
1924
-[-]-[--I->
AST-T
[-central +centripetal coincidence]: F is BEFORE G Tense:future Aspect: Prospective Adverb: before 1924 UT-T
-[---I-[----]->
AST-T
AST-T
-[-I-[-]->
EV-T
AST-T
-[-]-[-I->
1924
By proposing a uniform phrase structure for spatiotemporal predicates, we have further reduced the grammar of tenses, aspects and time adverbs to the same set of structural primitives. 3. ANCHORING SUBORDINATE CLAUSES We now extend our analysis of the temporal structure and interpretation of subordinate clauses to incorporate complement clauses. We examine both adverbial temporal clauses and complement clauses, and account for the morphosyntactic and interpretative restrictions that constrain their distribution. We argue that the set of available readings-dependent vs. independent construal of the subordinate (be it complement or adjunct) clause-as well as the morphosyntactic restrictions
ASPECTAND TEMPORAL MODIFICATION
207
imposed on the combinations of tenses and aspects in subordinate contexts can be uniformly derived from two general economy principles governing temporal derivations. 3.1. Anchoring Temporal Adjunct Clauses The paradigms in (40)-(43) exemplify restrictions that obtain between tense-aspect combinations in matrix and adjunct clauses. There are two types of restrictions: (i) restrictions on the occurrence of tenses and aspects in adjunct clauses (compare (40) with (41)-(42)), and (ii) restrictions on their interpretation: examples (43) are grammatical but only under a future shifted construal of the present (perfect) in the adjunct. (40) a. b. c.
Max left after Sue arrived. Max left before Sue arrived. Max left when Sue arrived.
(41) a. b.
* Max will leave after Sue arrived. * Max left before Sue will arrive.
(42) a. b. c. d.
* Max left when Sue will arrive. * Max will leave before Sue arrived.
(43) a. b. c.
* Max left after Sue will arrive.
* Max will leave when Sue arrived.
Max will leave after Sue arriveshas finished. Max will leave before Sue arriveslhas finished. Max will leave when Sue arriveslhas finished.
We argue that theses restrictions on the possible combinations and interpretations of tenses and aspects-when a temporal adjunct is merged into a matrix clause-fall out uniformly from the principles given in (44)-(45): (44)
Anchoring a subordinate clause into a matrix clause is optimal iff it yields an ordering of their Assertion Times relative to each other.
(45)
No step in a given temporal derivation can be semantically vacuous: every step in the derivation must be temporally informative (that is, must yield a temporally distinct interpretation).
The principles in (44) and (45) are economy principles constraining the representations and steps in a derivation at the interface with the semantic component. Principle (44) ensures that anchoring a subordinate clause into a matrix yields an optimal output-that is, an output in which the two times for which the
speaker makes a statement, the matrix and subordinate AST-~imes,end up ordered with respect to each other. Principle (45) is designed to ensure that a given temporal interpretation is achieved in an optimal manner, in the spirit of Fox's (2000, pp. 2-6) 'Scope and Variable Binding Economy Principles'-that is, "with no more effort than necessary." Principle (45) will rule out the illicit combinations in (41), and explain why the present tense in the adjuncts in (43) must shift to the future." To illustrate how these principles operate, consider the derivation of the ungrammatical (41a) *Max will leave after Sue arrived, given in (46).11 (46)
Derivation of (4la): *Max will leave after Sue arrived Step #I. Compute the temporal interpretation of the matrix Max FUT leave i. UT-T -AST-T1 Step #2. Compute the temporal interpretation of the adjunct Sue PAST arrive ii. AST-TZ UT-T Output of Steps #1 & 2 ...
111.
AST-T2 -UT-T -AST-TI
Step #3. Compute the temporal contribution of the connective afer iv. AST-T2 -UT-T -AST-TI The derivation in (46) is ruled out by the economy principle in (45). Step #3 of the derivation is illicit as it yields an output (iv) that is not temporally distinct from (iii): (47)
iii.
Output of Steps #I & 2 AST-TI (That is, AST-TIafter AST-T2)
AST-T~UT-T-
iv.
Output of Step #3 AST-TIAFTER,-ONNECTlVE AST-T2
(That is, AST-TI after AST-T2)
Step #3 in the derivation of (41a) is thus semantically vacuous, temporally uninformative: it yields exactly the same output as steps #1 & 2 combined. The generalization underlying our analysis is simple and intuitive. When a temporal adjunct is merged into a matrix clause, the ordering relation between the Assertion Times of each clause must be established by the temporal connective itself-since the semantic function of the connective, as a predicate of spatiotemporal ordering, is precisely to order the matrix and adjunct AST-Ts (as argued in section 2.5). The combinations in (41) are illicit because the order between the matrix and adjunct AST-TS is already established by the respective tense in each clause (that is, in (41a), the future in the matrix is after the past in the adjunct). The economy principle (45) will likewise rule out (4lb),*Max left before Sue will arrive, as temporally uninformative: the connective is not contributing
ASPECTAND TEMPORAL MODIFICATION
209
semantically to the temporal ordering of the AST-TS of the matrix and adjunct clauses. The ordering is itself already determined by the tenses of the clauses (the past in the matrix is before the future in the adjunct). We now turn to the ungrammatical sentences in (42). They will be ruled out because their (respective) derivations yield temporally uninterpretable-that is, contradictory-orderings of the AST-TS. We illustrate with the derivation of the ungrammatical (42c) *Max will 1eave.before Sue arrived, given in (48). (48)
Derivation of (42c): *Max will leave before Sue arrived
Step #I. Compute the temporal interpretation of the matrix Max F(IT leave i.
UT-T -AST-T1
Step #2. Compute the temporal interpretation of the adjunct Sue PAST arrive ii. AST-T2 UT-T Output of Steps #I & 2 iii. AST-T2-UT-T -AST-TI Step #3. Compute the temporal contribution of the connective before iv. AST-T1 AST-T2 The derivation of (42c) in (48) crashes since it yields contradictory orderings of the AST-Ts in (iii-iv). Since the derivation fails to order the AST-Ts, (44) is violated. In contrast, the derivations of the sentences in (40) converge as they yield semantically well-formed outputs, satisfying both (44) and (45). In particular, the connective contributes semantically to the ordering of the AST-TS, as the derivation of the grammatical (40b) Max left before Sue arrived illustrates: (49)
Derivation of (40b): Max left before Sue arrived
Step #I. Compute the temporal interpretation of the matrix Max PAST leave i. AST-T1 -UT-T Step #2. Compute the temporal interpretation of the adjunct Sue PAST arrive ii. AST-T2-UT-T Step #3. Compute the temporal contribution of the connective before iii. AST-TI -AST-~2 UT-T This time, step #3 in the temporal derivation is not temporally vacuous: the predicate before contributes to the temporal computation of the sentence by establishing the relative ordering of the AST-Ts of the matrix and adjunct clauses.
Summarizing, the combinations in (40) are licit because their derivation yields an interpretable, optimal output, satisfying (44). Moreover, the resulting temporal interpretations are achieved in an optimal manner (no vacuous step in the derivation), satisfymg (45). In contrast, the combinations in (41) are illicit because their derivation is temporally uninforinative. That is, at the step in the derivation at which the temporal contribution of the connective is computed, it no longer contributes to the temporal ordering of the two times for which the speaker makes a statement (the AST-TS).It is thus semantically vacuous, in violation of (45). Finally, the combinations in (42) are illicit because their derivation yields an uninterpretable/contradictoryordering of the AST-TS,in violation of (44).12 Turning to (43), these examples raise an intriguing problem. They are licit but only under a future shifted construal of the present in the adjunct clause. Why is a present tense interpretation of the adjunct clause unavailable, or ruled out? It is ruled out because it yields either a temporally uninterpretable output (for (43b,c)), or a temporally uninformative derivation (for (43a)). We illustrate with the derivation of (43a): (5 1)
Derivation of (43a): Max will leave after Sue arrives
Step # I . Compute the temporal interpretation of the matrix Max FUT leave i. UT-T -AST-TI Step #2. Compute the temporal interpretation of the adjunct Sue PRES arrive 11.
UT-T
I AST-T2
Output of Steps #I & 2: ... UT-T -AST-TI
111.
I Step #3. Compute the temporal contribution of the connective afer iv. UT-T AST-T1
I
Once again, the derivation in (51) crashes because the last step in the derivation does not yield a construal temporally distinct from the one generated at the previous step. (51) is thus ruled out by the economy principle in (45). However, shifting the present in the adjunct to the future rescues the derivation of (43a), as shown below.
ASPECTAND TEMPORAL MODIFICATION Derivation of (43a) Max will leave afer Sue arrives
(52)
Step # I . Compute the temporal interpretation of the matrix i. UT-T -AST-T1 Step #2. Compute the temporal interpretation of the adjunct [Future shifted] ii. UT-T -AST-T2 Step #3. Compute the temporal contribution of the connective after ... 111.
UT-T -AST-T2 -AST-T1
In sum, shifting to the fiture rescues the derivation as it yields a temporally informative output: the connective is no longer semantically vacuous, it establishes the temporal ordering of the AST-TS,relative to each other.I3 3.2. Anchoring Temporal Adjunct Clauses vs. Anchoring Complement Clauses
We now turn to a major difference in the anchoring of temporal adjunct clauses, as opposed to complement clauses. As is well known, complement clauses can yield dependent construals. The dependent construal of (53a), with a Past under a Past, is given in (53b). Under t h s reading, the external argument of To in the complement clause is anaphorically anchored to (bound by) the matrix AST-T (AST-TI),itself cotemporal with EV-TI (since we have perfective aspect in the matrix). The embedded Past Tense will then order its external argument (now identified with AST-TI)after its internal argument AST-T2, itself cotemporal with EV-T2(since we also have perfective aspect in the complement clause). The embedded Past Tense thus indirectly orders the past time of saying after the time of kissing, as (53b) illustrates. (53) a. b.
Max said that Lou kissed Billie
Dependent (past shiftedanaphoric) construal: EV-T2 AST-TI=EV-T1 UT-T AST-T~= -[ I-[ '-I KISSING
c.
SAYING
*Independent (non shifteddeictic) construal: AST-TI= EV-T1 AST-T2= EV-T2 UT-T -[ [-I '-I SAYING
KISSING
Note that the independent construal of the complement clause in (53a) is not available. Under this reading, the external argument of To in the complement clause is deictically anchored to the UT-T. The embedded Past Tense then orders its external argument (now identified with the UT-T)after its internal argument AST-~2,
itself cotemporal with EV-T2. The time of kissing (EV-~2) is thus (indirectly) ordered in the past with respect to the UT-T.Crucially, however, it remains unordered with respect to the past matrix AST-T (AST-T~=EV-TI). This incorrectly predicts that (53a) allows the reading in (53c) where the past time of kissing follows the time of saying. Finally, in (54), we see that in a future complement clause embedded under a future matrix clause, the dependent construal illustrated in (54b) is likewise enforced. That is, the independent construal in (54c) is not available. (54) a. b.
Max will say that Lou will kiss Billie
Dependent (future shiftedJanaphoric) construal: UT-T AST-TI' EV-T1 AST-TZ= EV-T2
-[I
[-I
c.
I-' KISSING
SAYING
* Independent (non shiftedldeictic) construal: UT-T
AST-T2= EV-T2
-I[
. AST-TI= EV-T1
[-I KISSING
I-' SAYING
Now, in contrast to complement clauses, temporal adjunct clauses always yield independent construals. That is, the external argument of To in an adjunct must be deictically anchored to the UT-T;it cannot be anaphorically anchored to the matrix AST-T.This difference between complement and adjunct clauses is not stipulated: it falls out automatically from the economy principles in (44)-(45). That is, anaphorically anchoring the external argument of To in an adjunct clause will always be ruled out for either of two reasons. (i) At the point in the derivation where the temporal contribution of the connective is computed, it no longer contributes semantically to the temporal ordering of the AST-Ts. As this step in the derivation is semantically vacuous, (45) is violated. (ii) Alternatively, the derivation yields a temporally uninterpretable (contradictory) ordering of time spans, violating (44). We illustrate each of these subcases with the derivations in (55) and (56). Note the anaphoric anchoring of the adjunct clause at step #2 of the derivation: the embedded Past Tense orders the AST-T of the adjunct clause (AST-T2)AFTER the matrix AST-T (AST-TI),and not after the UT-T as was the case in the derivation (49). (55)
Anchoring the external argument of subordinate To to matrk AST-T Max left before Sue arrived
Step # I . Compute the temporal interpretation of the matrix Max PAST leave i. AST-T1 -UT-T
ASPECTAND TEMPORAL MODIFICATION Step #2. Compute the temporal interpretation of the adjunct Sue PAST arrive ii.AST-T2 -AST-TI a Anchor time is the matrix AST-T Output of Steps #I & 2 iii. AST-T2 -AST-TI
UT-T
Step #3. Compute the temporal contribution of the connective before iv. AST-TI AST-T2 a Contradictory orderings of the AST-Ts The derivation crashes since it yields the uninterpretable, contradictory, ordering of AST-Ts in (iii)-(iv). Since the derivation fails to order the AST-Ts, (44) is violated. Consider now the other subcase, where anchoring the external argument of To in the embedded adjunct to the matrix AST-T generates a semantically vacuous output: (56)
Anchoring the external argument of subordinate Toto matrk AST-T Max left after Sue arrived
Step # I . Compute the temporal interpretation of the matrix Max PAST leave i. AST-TI UT-T Step #2. Compute the temporal interpretation of the adjunct Sue PAST arrive ii AST-T2 AST-TI Anchor time is the matrix AST-T Output of Steps #1 & 2 ...
111.
AST-T2 -AST-TI
-UT-T
Step #3. Compute the temporal contribution of the connective after iv. AST-T2 AST-TI -UT-T a Temporally vacuous output Once again the derivation crashes: the last step in the derivation is illicit since it does not yield a temporally distinct construal-(iii) and (iv) are non-distinct outputs. Adjunct clauses thus contrast sharply with complement clauses. Temporal anchoring of before and after adjunct clauses into the matrix can only be established via the temporal connective, as predicted by our economy principles. In contrast, temporal anchoring of the complement clauses in (53)-(54) into the matrix can only be established anaphorically, via control (/binding) of the subordinate external argument of T by the matrix AST-T.The enforced dependent construal in (53)-(54) likewise follows automatically from the principles in (44)-(45), as we now show.14 3.3. The Typology of Construals in Complement Clauses
We have argued that restrictions on the distribution and interpretation of tenses in temporal adjunct clauses follow uniformly from two economy principles. We now show that these principles also explain the typology of construals in complement
clauses-that is, the distribution of (in)dependent construals (enforced, optional, or proscribed) in complement clauses. We consider here only eventive verbs.15 We take the null hypothesis to be that the external argument of To in a subordinate clause-henceforth, the anchor (time)--can be freely set either to the UT-T, or to the matrix AST-T (AST-TI). When the external argument of To is identified with the UT-T,the temporal anchoring of the subordinate clause is deictic. When the external argument of To is controlled by AST-TI,the temporal anchoring of the subordinate clause is anaphoric. For concreteness, we assume here the default setting for the anchor time to be the UT-T.'~Resetting to the matrix AST-T is free however-as long as the resulting derivatiodoutput satisfies the economy principles in (44)-(45). This proposal yields three different scenarios for complement clauses: (i) obligatory resetting of the anchor to the matrix AST-T, (ii) optional resetting of the anchor to the matrix AST-T, (iii) vacuous resetting of the anchor to the matrix AST-T. 3.3.1. Obligatory Resetting of the Embedded Anchor to the Matrix AST-T. Let us start with the first scenario: obligatory resetting of the anchor to the matrix AST-T. This scenario is illustrated with the examples in (53)-(54), repeated as (57). (57) a. b.
Max said that Lou kissed Billie Max will say that Lou will kiss Billie
Recall that these examples only allow a dependentlshifted interpretation (see the discussion and schema in section 3.2 above). Why is the dependent construal enforced in (57)? We argue that pastlfuture shifted readings obligatorily arise when merging a complement clause into a matrix clause requires resetting the anchor time from its default value (UT-T)to the matrix AST-T-otherwise, the economy principle in (44) is violated. We illustrate with the derivation of (57a), given in (58). (58)
Max said that Lou kissed Billie Step #I. Compute the temporal interpretation of the matrix Max PAST say i. AST-TI-UT-T Step #2. Compute the interpretation of the complement Lou PAST kiss Billie ii. AST-T2-UT-T a Anchor time is the UT-T 3 Output
of Steps #1 & 2 is not optimal. (44) is violated.
The temporal derivation in (58) does not yield an optimal output. The AST-TS of the matrix and complement clauses-that is, the two times for which the speaker makes an assertion-are both ordered in the past with respect to the UT-T, but remain
ASPECTAND TEMPORAL MODIFICATION
215
unordered with respect to each other;violating (44). Resetting the anchor time from its default value to the matrix AST-T is consequently required in order to achieve a relative ordering of the AST-Ts, as (59) illustrates. (59)
Max said that Lou kissed Billie Step #I. Compute the temporal interpretation of the matrix Max PAST say i. AST-TI -UT-T Step #2. Compute the interpretation of the complement Lou PAST kiss Billie ii. AST-T2 -AST-T1 3 Anchor time reset to AST-TI Output of Steps #I & 2 ...
111.
AST-T2 -AST-T1
UT-T
=. Output of Steps #I & 2 optimal This time the derivation converges: it yields the optimal output in (59iii) where the AST-Ts are ordered relative to each other.
3.3.2. Optional Resetting of the Embedded Anchor to the Matrix AST-T. We have explained why the examples in (57), with past under past and future under future, require obligatory resetting of the anchor time of the complement clause to the matrix AST-T. However, resetting the anchor time is not always obligatory; it can be optional. This scenario is illustrated with the following tense-aspect combinations in the matrix and adjunct clause. (60) a. b.
Max will say that Lou is kissing Billie Max will say that Lou is lussing Billie
(61) a. b.
Max said that Lou will kiss Billie Max said that Lou would kiss Billie
As we shall see, resetting the anchor time in (60)-(61) from its default value (UT-T) to the matrix AST-T is a legitimate (that is, temporally informative) step. However, t h s time, resetting will be optional, and not enforced as was the case for (57) above. To illustrate, let's go through the possible derivations of (6Oa) given in (63)-(64). (63)
Max will say that Lou is kissing Billie Step #I. Compute the temporal interpretation of the matrix Max FUT say i. UT-T -AST-T1
Step #2. Compute the interpretation of the complement Lou PRES kiss Billie 11. UT-T 3Anchor time is the UT-T
I
Output of Steps #I & 2 ...
111.
UT-T
AST-TI
I 3Output
of steps #1 & 2 is optimal
a Steps #1 & 2 are temporally informative The derivation in (63) converges: each step is temporally informative and the resulting output is optimal. Note that under the reading illustrated by the schema in (63iii), the time of saying follows the time of kissing, which itself overlaps with the UT-T. Now resetting the anchor time to AST-TIyields the temporally distinct output in (64iii'): the W r e time of saying is now simultaneous with (and not subsequent to) the time of kissing. In sum, resetting the anchor time in (64) is temporally informative and, thus, legitimate-but it is not enforced (as was the case for (57)). (64)
Resetting the anchor time to AST-Tl( 3 AST-TIWITHIN AST-T2) iii'.
UT-T -AST-TI
I
By the same reasoning, (60b) allows the two possible derivations given in (65)-(66). Setting the anchor time to the UT-T yields the well-formed output in (65iii), where the time of kissing is ordered in the past with respect to both the UT-T and the time of saying. Resetting the anchor is freely allowed as it yields the temporally distinct interpretation in (66iii') where the time of kissing is ordered in the past with respect to the time of saying, but unordered with respect to the UT-T. That is, the time of kissing can either follow or precede speech time. The alternative derivation in (64) is thus temporally informative, and (45) is satisfied. Notice also that the output in (66iii') is optimal since the matrix and embedded AST-TS are ordered relative to each other (AST-T1AFTER AST-T2),thus satisfying (44). (65)
Max will say that Lou kissed Billie Step #1. Compute the temporal interpretation of the matrix Max FUT say i. UT-T -AST-TI
ASPECTAND TEMPORAL MODIFICATION
Step #2. Compute the interpretation of the complement Lou PAST kiss Billie ii. AST-T2 UT-T 3 Anchor time is the UT-T Output of Steps #I & 2 ...
111.
(66)
A S T - ~ 2-UT-T
AST-T1
Resetting the anchor time to AST-TI (=AST-T1AFTER AST-2) iii' UT-T
Or
UT-T
Optional resetting is thus allowed in all these configurations since it is temporally informative and yields a relative ordering of the AST-TS. Finally, we argue that resetting the anchor time can have an overt morphological reflex, e.g., the wilNwould alternation in (61). Thus, consider the derivation of (61a) given in (67), where the anchor time of the embedded clause is set to the UT-T. (67)
Max said that Lou will kiss Billie
Step #I. Compute the temporal interpretation of the matrix Max PAST say i. AST-TI UT-T Step #2. Compute the interpretation of the complement Lou FUT kiss Billie ii. UT-T AST-T2 3 Anchor time is the UT-T Output of Steps #I & 2 .. .
111.
AST-T1 -UT-T
AST-T2
The derivation in (67) converges since it yields the relative ordering of the AST-TS in (67iii). Note that the time of saying is ordered in the past with respect to both the UT-T and the time of kissing. Consider now the alternative temporal derivation in (68), where the anchor of the embedded clause has been reset to the matrix AST-T. (68)
Resetting the anchor time to AST-TI iii'. AST-TI
t UT-T
( 3AST-T2AFTER AST-TI)
AST-~2
? or
UT-T
Note that resetting the anchor time to AST-TI is temporally informative (thus satisfying (45)) since it yields the temporally distinct construal in (68iii'j where the time of kissing is ordered in the future with respect to the time of saying, but
unordered with respect to the UT-T.That is, the time of kissing can either precede or follow speech time. Note further that the output in (68iii7 is optimal since the s ordered relative to each other (AST-TIBEFORE matrix and embedded A S T - ~ lare AST-T2),thus satisfying (44). Now, if we assume that will and would are derived from the common abstract predicate WOLL (see Ogihara (1989) and references therein), then resetting the anchor from the UT-T to the matrix AST-T will have an overt morphological reflex. When the anchor time of the embedded clause is the UT-T,as is the case in the derivation (67), the future in the embedded clause is spelled out as will. In contrast, when the anchor of the embedded clause is the past matrix AST-T,as is the case in the derivation (68), the future in the einbedded clause is spelled out as would.
3.3.3. Vacuous Resetting of the EmbeddedAnchor to the Matrix AST-T We now illustrate the last possible scenario for the anchoring of subordinate clauses. Setting the anchor time of the complement clauses in (69) to either the UT-T or the matrix AST-T will yield exactly the same temporal ordering of time spans. We exemplify this scenario with the derivation of (69a) in (70)-(71). (69) a. b. c.
(70)
Max says that Lou will kiss Billie Max says that Lou kissed Billie Max says that Lou is kissing Billie
Max says that Lou will kiss Billie Step #l. Compute the temporal interpretation of the matrix Max PRES say i. UT-T
I
AST-TI
Step #2. Compute the interpretation of the complement Lou FUT kiss Billie ii UT-T_ AST-T2 3 Anchor time is the UT-T Output of Steps #I & 2 ...
111.
UT-T -AST-T2
I (71)
Resetting the anchor time to AST-TI( 2 UT-T WITHIN AST-TI) iii'.
U T - T
I
AST-~2
2 Temporally
uninformative/vacuous
ASPECTAND TEMPORAL MODIFICATION
219
The two settings of the anchor time (deictic in (70) vs. anaphoric in (71)) yield identical temporal outputslconstruals. Resetting the anchor time from its default setting-whether we take the default to be the UT-T or the matrix AST-T (see note 16)-is thus semantically vacuous, temporally uninformative: it yields an output identical to the output generated by the alternative derivation. Resetting the anchor time in the temporal derivation of examples (69) is thus not a legitimate step: it violates the economy principle in (45). 4. CONCLUSION The anchor time in a subordinate clause-be it a temporal adverbial or a complement clause-can either denote the UT-T, or be referentially dependent on the matrix AST-T.Anchoring a subordinate clause is subject to two economy constraints. (i) It must yield an optimal output. The output is optimal if the AST-Ts of the matrix and subordinate clauses end up ordered relative to each other. (ii) A given temporal construal must be achieved in an optimal manner-that is, no step in the derivation may be semantically vacuous, temporally uninformative. The proposal that economy principles govern the steps in-and the representations generated by-the temporal computation of a given sentence accounts for three different sets of facts. First, it can derive the restrictions on tense-aspect combinations in matrix and adjunct clauses (section 3.1). Second, it explains a major difference in the anchoring of temporal adjunct vs. complement clauses. In particular, a pastlfuture shifted reading in a complement clause can only be achieved via a referential dependency of the anchor time (in the complement clause) on the matrix AST-T.In contrast, a pastlfuture shifted reading in an adjunct clause can only be acheved via the temporal contribution of the connective beforelafter (section 3.2). Third, t h s proposal .accounts for the typology of construals in complement clauses with eventive verbs (section 3.3). Before closing, we highlight a novel aspect of the model presented here. As the derivations throughout this last section illustrate,17 we are implicitly assuming that computing the temporal reference of a given sentence does not take place after syntax-that is, once phase structure units have been merged togetherlassembledbut rather through out the derivation, in parallel to the assemblinglmerger of syntactic units. If this proposal is on the right track, then it questions the traditional assumption that there is a unique interface level between syntax and the interpretative component (LF). That is, the proposal that temporal interpretation does not take place at the end of the'syntactic derivation-but at different steps in the derivation, in the construction of the (temporal) phrase-structure of a sentencesuggests that the interface between syntax and meaning is multiple-as has been proposed elsewhere for the syntax-PF interface (Nunes & Uriagereka, 2000).18 Demirdache, Universitd de Nantes Uribe-Etxebarria, Euskal Herriko Unibertsitatea
NOTES *
We thank Brenda Laca, Paula Kempchinsky, Javier Ormazabal and the participants of the international workshop on the syntax, semantics and acquisition of aspect (University of Iowa, May 2002) for insightful comments, questions and discussion. This project has been partly funded by the University of the Basque Country [9/UPV 00033.130-13888/2001] and by a grant from the Basque Government (PI-198-127).
'
Note that D&UE (1997b, 2002) assume a finer-grained representation of the EV-T. In particular, a causative predicate is assigned a complex event structure consisting of two subevents-a process and a result state-represented as a layered VP structure (cf. Zagona, 1990, and others). Each subeventNP is associated with its own time argument: EV-TI for the process subevent, EV-T2 for the result state. This decomposition of the EV-T entails that the Progressive orders the AST-T WITHIN EV-TI-that is, within the time which defines the process part of the described event. This analysis correctly predicts that Zoe idwas closing the door means that Zoe islwas in the process (and not the result state) of closing the door. D&UE (1997b, 2002) argue that the proposal that the Perfect is the spatiotemporal predicate AFTERtogether with a finer-grained representation of the EV-T(see note (I))--uniformly derive the existential, resultative and continuative readings of the English present perfect and, more generally, the construals of the Perfect with telic vs. atelic predicates. By Tense without overt/morphological Aspect, we do not mean only non-compound tenses. In particular, D&UE (2002) anlyse the French impafait, which is not a compound tense, as a past Tense morphologically marked for imperfective Aspect (itself analysed as the spatiotemporal predicate WITHIN). For discussion of the differences between the progressive and the imperfective, see Demirdache (2003). We do not distinguish formally between the specifier of XP and adjunction to XP (cf. Kayne, 1994; Chomsky, 1995). Notice that analyzing the time adverbs in (13) as either (i) modifiers restricting the reference of the EV-T, or (ii) modifiers restricting the reference of the AST-T, will yield non-distinct interpretations for ( 1 3 b s i n c e the AST-T binds the EV-T (perfective aspect). In (17)-(IS), we discuss perfect sentences where modification of the EV-T and modificatiqn of the AST-T by a time adverb yield distinct construals. Note that the relative ordering of the two temporal PPs in (19) follows automatically. Since the AST-T in (19b) is required to denote a time span which both falls ATIWITHIN 5AM and is AFTERhbsequent to the span designated by 'dawn', then 'dawn' in (19) automatically denotes a time which precede 5 AM, as the diagram in (i) illustrates.
' '
(i)
-[DAWN] -[SAM] -[DAWN] ->
I
[AST-T] Paula Kempchinsky (pc) suggests the following indirect argument to support the idea that bare time adverbs are concealed PPs (based on an observation by Stroik (1966, p. 119)). Bare time adverbs have a freer distribution than bare locative adverbs. As (ia-b) illustrate, locative adverbs generally need Case unless they are predicates: (i)
a. This is Madison The show is *Madisonfin Madison b. (ii) a. This is Monday b. The show is Mondaylon Monday The grammaticality contrast between (ib) and (iib) would follow if the bare temporal adverb Monday is in fact a PP, since PPs don't need Case, and typically show Case resistance effects. D&UE (2002) identify two distinct types of bare time adverbs: bare NP adverbs and bare CP adverbs -that is, temporal adjunct clauses which are not introduced by an overt preposition, but merely by a complementizer, as the French example in (38c) in the text illustrates. Bare CP time adverbs are analyzed, just like bare NP time adverbs, as PPs headed by a silent dyadic predicate of spatiotemporal ordering expressing central coincidence, as the structure assigned to (38c) illustrates.
ASPECTAND TEMPORAL MODIFICATION
22 1
D&UE (2004a) argue that the proposal that predication takes place via movement of a temporal operator automatically explains the puzzling unavailability of certain readings in time adjunct clauses with perfect aspect, as reported by Homstein (1'990). loIn D&UE (2000), the economy principle in (45) rules out illicit combinations of recursive aspect, namely the Progressive of the Perfect and the Progressive of the Progressive, while at the same time allowing the Perfect of a Progressive which, in contrast to the former combinations of aspects, is licit. I' For expository reasons, we henceforth simplify the schemas representing the temporal construal of a given sentence: we do not represent the EV-T if it coincides (/is cotemporal) with the AST-T (as is the case with simple tenses). (46i-ii) for example, should thus be understood as shorthand for (i)-(ii), respectively: UT-T-AST-TI= EV-TI (9 AST-2 = EV-T2-UT-T (ii) lZ Note that our system allows for a future adjunct clause embedded under a future matrix clause. While such patterns are not very felicitous in languages like in English, they are fine in languages like Greek (see Shaer (1996)) or French (compare (ia) with (iia), or (iiia) with (iiib)). Indeed, even the preference for a future shifted present over a morphological future in English adjunct clauses does not hold in French (compare (iia) with (iib)). Note finally that future under future is allowed in before-clauses such as (iv). a. ? Max will leave when Rosa will leave. b. Max will leave when Rosa leaves. (ii) a. Max partira quand Rosa partira. b.*/?? Max partira quand Rosa part. (iii) a. ? Max will leave after Rosa will have finished her thesis. b. Max partira apr& que Rosa aura fini sa thkse. Pigs will fly before he'll become a mathematician. (Shaer, 1996, p. 246) ( 4 l 3 We do not address here the issue of why present can shift to the future in either: (i) an adjunct clause under a future matrix clause (e.g. (43)); or (ii) in a simple clause under a future adverb (e.g. Max leaves tomorrow). We only explain why a semantic present in an adjunct under either past in the matrix (e.g. *Max left afterlbefore Sue arrives), or future in the matrix (e.g. (43)) is illicit, and further, why shifting the present in the adjunct in (43) to the future rescues the sentenceas the derivation in (52) illustrates. l4 We assume relative clauses involve coordination, and not subordination (see Demirdache (1991), Kayne (1994), and references therein). As a result, merging a relative clause with a matrix clause will not be governed by the economy principle in (44), since the latter is a constraint on subordination. l5 For discussion of the temporal construals of complement clauses with stative verbs and, in particular, for an analyis of the so called simultaneous and double access readings, see D&UE (in prep). l6 We have assumed here that deictic anchoring (that is, setting the anchor time of the subordinate clause to the UT-T)is the default setting of the anchor time. We stress, however, that the reasoning would go through equally well if we had chosen anaphoric anchoring (that is, setting the anchor time of the subordinate clause to the matrix AST-T)as the default value, with resetting taking place to the UT-T.See D&UE (in prep) for discussion of how the anchor time is set throughout a temporal derivation. l7 Notice in particular that in the derivations illustrated in (46) through (71), the temporal interpretations of the matrix and subordinate clause are computed independently, in parallel (Steps 1 & 2). After merger of the subordinate clause into the matrix (Step 3), the computation continues (that is, the temporal contribution of the connective is evaluated after merger; likewise resetting the anchor time takes place after merger). See D&UE (in prep) for further elaboration. For a detailed discussion of the implications of our model for the architecture of the grammar, see D&UE (in prep). (i)
CARLOTA S. SMITH
ASPECTUAL ENTITIES AND TENSE IN DISCOURSE
1. INTRODUCTION: LINGUISTIC STUDY OF DISCOURSE Linguistic expressions form patterns in discourse. In the domain of temporality, the morphemes of tense and aspectual viewpoint pattern significantly. Shifts of viewpoint and tense are often associated with shifts of direction, and with the distinction between foreground and background.1 One important reason for this effect is that aspectual viewpoint and tense form closed sub-systems in language. In a closed system, choice of one term implies contrast with the other possibilities. The element of choice and its contrastive significance allows for more than one level of discourse meaning. This article is concerned with the patterning of covert linguistic categories in discourse. I will show that aspectual situation categories such as event and state pattern together to establish discourse units at the relatively local level of the passage. The situation categories also form a closed system. As an introduction to the research, I would like to address a very general issue about the linguistic study of discourse. Linguistic study of discourse poses the problem of finding a level that is fruitful for close analysis. Typically researchers have focused on genre, e.g. conversation, personal narrative, highly structured written materials. This is a reasonable strategy, because each genre represents an activity with its own purpose, structure and conventions. The various genres are different 'activity types' in which language plays a key role, as Levinson (1979) emphasized. As an activity embedded in context, discourse is understood mainly on the basis of convention and expectation, rather than on particular linguistic form. Indeed, the global structure of a discourse is often not presented directly in a text. What this suggests is that genre is the wrong level for close linguistic study of discourse, due to its strong pragmatic basis. I will work at the more local level of the passage. Within a text one recognizes stretches that are intuitively different, e.g. narrative, description, argument. Each has a particular force and a characteristic cluster of linguistic features and interpretations. I shall say that they realize different "discourse modes". I posit five modes: Narrative, Report, Description, Information, and Argument. Each makes a different contribution to the text. The list is not exhaustive—it omits conversation and procedural discourse, for instance—but it includes the modes that commonly appear in written texts. Persuasive discourse is not a distinct mode or genre: persuasion appears in texts of all kinds. Many factors are the same in the spoken and written modalities; for the work reported on here I consulted written material only. The generalizations and examples
223
P. Kempchinsky andR. Slabakova (eds.), Aspectual Inquiriess, 223-237. © 2005 Springer. Printed in the Netherlands.
224
CARLOTA S. SMITH
are based on a group of about thirty texts that I examined carefully for this work. They vary in genre, length, and venue, including newspaper articles, essays in journals, short stories, and novels. I use the term 'discourse' for both spoken and written material, and 'text' for written. The discussion is confined to English. I think that the analysis will hold across languages, although some details will differ for different languages, Section 2 gives a general characterization of the modes. Section 3 discusses situation entities; section 4 discusses principles of text progression, with examples from each mode; section 5 concludes. 2. DISCOURSE MODES To introduce the modes and the intuitive distinctions between them, I present two text passages. In each there is a shift from one mode to another. Consider first the passage in (I), from a newspaper editorial. The original paragraphing is preserved. The provenance of the natural examples is given at the end of the article. Argument to Narrative 1 I feel reasonably certain of the final verdict on the current impeachment affair because I think history will see it as the climax of a six-year period marred by a troubling and deepening failure of the Republican party to play within the established constitutional rules. 2 It was on Election Night 1992, not very far into the evening, that the Senate minority leader, Bob Dole, hinted at the way his party planned to conduct itself in the months ahead: it would filibuster any significant legislation the new Democratic President proposed. This passage begins quite abstractly in the mode of Argument, and shifts at the second sentence to Narrative. The shift occurs with introduction of a particular time, place, and individual; and a dynamic event (Bob Dole hinted ...) in S2. The cleft is a typical scene-setting device in the fairy tale genre, which gives it an ironic flavor in this c o n t e ~ t . ~ The second example also involves a shift to narrative. The first two sentences express generalizations, typical of the Informative mode. Then with the specifics of S3, the mode of the passage shifts to Narrative. (2)
Informative to Narrative 1 When a big whale dives, currents set in motion by the passage of so many tons of flesh come eddying back up in a column that smoothes the restless surface of the sea. 2 Naturalists call this lingering spool of glassy water the whale's footprint. 3 Out between the Hawaiian islands of Maui and Lanai, Jim Darling nosed his small boat into a fresh swirl. 4 The whale that had left it was visible 40 feet below, suspended head down in pure blueness with its 15-foot-long arms, or flippers, flared out to either side like wings.
ASPECTUAL ENTITIES AND TENSE IN DISCOURSE
225
Such shifts are common in actual texts. There is normally some variation within texts of a given genre. Narrative fiction, for instance, expresses sequences of event, bound together more or less closely by a unifying theme. But narratives rarely consist entirely of such sequences. They also have descriptive passages, and sometimes commentary. In expository texts one often finds narrative sequences which depart from the argument line. The point here is that texts of almost all genre categories are not monolithic, but rather have passages of different modes. This may be the reason that genre-based searches for linguistic regularities have not been particularly successful. Two linguistic features characterize the modes, both involving temporality. The first is the type of situation-events, states, and others discussed in the next section-that a text passage introduces into the universe of discourse. The second is the principle of text progression. Three of the modes are temporal: Narrative, Description, and Report. Text passages in these modes progress by temporal or spatial change. In the atemporal modes of Information and Argument, the situation entities are primarily static. Atemporal texts progress by metaphorical change of location through the information space of the text. 3. SITUATION ENTITIES AND THE DISCOURSE MODES Texts introduce individuals, concepts and times into the universe of discourse. They also introduce situations such as events and states, which are familiar from aspectual studies. In formal accounts, individuals, times and situations are represented as different types of entity in the representation. Each entity is licensed by information in a text. The analysis that I sketch in t h s article is developed more formally in the framework of Discourse Representation Theory in Smith (2003). I propose a new classification of situation entities that recognizes differences among non-stative situations. There are three main types of situation entity: Eventualities, or specific events and states; General Statives, or generics and states that involve a pattern or regularity, and Abstract Entities, facts and propositions. I will argue that the three types are distinct on conceptual grounds, and that they have linguistic correlates. T h s extended notion of situation entity is the most important feature of the five discourse modes. Situation entities are conceptual categories, expressed linguistically at the level of the clause. They are realized by verb constellations and nominal forms (Smith 199111997).The sentences in (3) exemplify members of the category of Eventuality: (3) a. b.
Eventualities (particular) The lobster won the quadrille. Lee rehearsed. (Events) The cat is on the mat. The Colonel owns the farm. (States)
The semantic property of dynamism distinguishes events from states. Clauses expressing events have distributional and interpretive properties associated with dynamism, whereas those of states do not. The properties are familiar, and I shall not discuss them here. Nor will I discuss finer distinctions among events and states, also familiar and not directly relevant to the discourse modes. The verb
226
CARLOTA S. SMITH
constellations that express events and states have distinct distributional properties, as shown by Vendler (1957), Verkuyl(1972) and Dowty (1979). They are therefore covert categories of grammar in the sense of Whorf (1956), and can be considered part of a speaker's knowledge of the language. There are many types of states. Particular states such as those expressed in (3b) hold at a particular time and place; they may last for a long or short period. I now consider some other types of non-dynamic situation categories. The class of General Statives includes Generalizing and Generic sentences, as exemplified in (4): (4) a. b. c.
Statives (general) The lion has a bushy tail. (Generic: kinds) John speaks French. (Generalizing: patterns of situations) I1 Mary often fed the cats last year.
Generalizing sentences express regularities, patterns of situations rather than particular events or states. They are also known as gnomic, dispositional, general, and habitual. The latter two labels reflect the fact such sentences often have a frequency adverbial (sometimes, always, never). One test for whether a sentence is of the Generalizing type is whether it allows a frequency adverbial without disturbing the syntax or interpretation. If it does allow a frequency adverbial, it is almost certainly a generalizing sentence. Krifka et al. (1995) use the term 'characterizing sentence' for this class. The generalizing interpretation of sentences like (4c) is due to a constraint that prevents bounded events from being located in the Present. I will call it the Bounded In the Present, events must be presented as ongoing, e.g. John is Event ~onstraint.~ speaking French, Mary is drawing a circle. The Bounded Event Constraint is due to a pragmatic convention of communication that has semantic consequences. In taking the temporal perspective of the present, speakers follow the tacit convention that communication is instantaneous. The perspective of the Present is thus incompatible with a bounded event, because the event-even an instantaneous event-would go beyond that perspective (Kamp & Reyle, 1993). In English, a present tense clause with an event verb constellation and the simple, verb form is semantically stative, conveying a general pattern rather than a particular event, as in (4c) above. (I take it that the simple verb form conveys the perfective viewpoint. For English statives, the perfective viewpoint focuses the entire temporal schema of the state, which does not include endpoints; see Smith (199111997) for discussion. Thus the interpretation of sentences like (4c) is consistent with other statives in English.) The shift to a stative situation type arises by coercion, triggered by the Bounded Event Constraint. There are exceptions to the Constraint, all of which telescope time. Most familiar are perfonnatives (I hereby christen this ship the 'Queen Elizabeth ') and sports-announcer reports (Now Jones throws the ball to third base). Also exceptional are literary commentary (Here the author invites us to notice...) and fictional narrative entirely set in the present, which has its own conventions. Languages have different ways of realizing the Bounded Event Constraint. In Russian, for instance, bounded event sentences in the present
ASPECTUAL ENTITIES AND TENSE IN DISCOURSE
227
tense are systematically taken as Future; while in some languages the present tense has only an imperfective viewpoint value. Although generalizing sentences lack the dynamism of particular event sentences, they have some distributional properties of dynamism (cf Smith, (199111997). They can appear with forms associated with agency and control, and with pseudo-cleft do: (5) a. b. c.
John deliberately plays tennis every Friday. I persuaded John to plays tennis every Friday. What John does is to play tennis every Friday.
These distributional facts reflect the hybrid nature of generalizing sentences. Although stative, they often have dynamic verb constellations (e.g., play tennis) and they involve a pattern of dynamic events. Of course, stative generalizing sentences exist as well: John is often in love, Maiy usually knows the answer. Generic sentences refer to kinds rather than individuals. In a sentence like The lion has a bushy tail the subject NP denotes the entire class of lions, not a particular lion or lions. Definite NPs (the lion) and bare plurals (lions) are the main types of NP that are taken as kind-referring. Characterizing generic sentences by syntactic means alone is difficult, but there are typical forms and interpretations; see Carlson & Pelletier (1995). Generic and Generalizing sentences are derived by coercion from verb constellations that express specific eventualities at the basic level of categorization. For instance, the verb, object argument and adverbial of examples (4b,c) above (speak French, feed the cat) express specific events. In context with a definite NP subject, simple viewpoint and present tense, however, they have the generalizing interpretation. Similarly, generic sentences may have event verb constellations, as in The lion eats meat. Verb constellations that express states at the basic level also appear in general statives, as in (4a) above, have a bushy tail. I will assume that General Statives-generic and generalizing sentences-can be recognized by their distributional and semantic properties, and are also covert categories in the grammar. This is somewhat optimistic, since these sentences are notoriously difficult to characterize' in terms of their linguistic properties. For generalizing sentences the strongest properties are the presence andfor possibility of frequency adverbs, and the combination of present tense, perfective viewpoint and event verb constellation. Past generalizing sentences are more difficult to characterize linguistically. Now consider the abstract entities of Facts and Propositions. They are situation entities introduced by verb constellations in clausal complements of certain predicates; I use the term 'clausal complement' for arguments with the internal structure of a clause (predicate and argument) that appear in both subject and nonsubject position.
228
CARLOTA S. SMITH
(6) a. b.
Abstract entities Facts: obiects of knowledge I know that Mary refused the offer. Mary's refusal of the offer was significant.
c. d.
Propositions: obiects of belief I believe that Mary refused the offer. Mary's refusing the offer was unlikely.
Clausal complements referring to facts and propositions have characteristic distributional and other linguistic features, and thus function as covert linguistic categories, as Vendler showed in his article "Facts and Events" (1967). Abstract entities differ from the other types of situation in how they relate to the world. Eventualities and general statives are located spatially and temporally in the world; abstract entities are not. Facts are the objects of knowledge, while Propositions are the objects of belief. Conceptually and truth-conditionally, Facts and Propositions differ from Eventualities. The latter are spatiotemporally located and have causal powers. Facts themselves are not located, yet they are contingent for truth on situations being a certain way; arguably, they also have causal powers (Asher, 1993). The class of Factive predicates includes emotives (regret, resent, deplore), verbs of communication (say, tell, show, indicate), and verbs of conjecture (guess, predict). Propositions are the contents of mental states like beliefs, expectations, decisions, and intentions. Propositions are not located, are not contingent, and do not have causal powers. They are typically expressed by clausal arguments of propositional attitude verbs and other predicates, e.g believe, doubt, fear, hope, want, think, affirm, deny; be unlikely, consistent. This notion of 'proposition' should not be confused with the more general use of the term, in which a proposition is the content that a sentence expresses. Abstract entities may seem like other states in their strictly aspectual properties. Traditionally, aspect involves the internal temporal structure of situations; thus, the standard situation types differ in the factors of dynamism, telicity and duration. However, if we extend the notion of internal structure to include truth-conditional verification of situations, then Facts and Propositions are distinct from other classes of states. Spatiotemporal location and contingency are both relevant to the truthconditions of a sentence. These factors differentiate Facts and Propositions from specific states and from general statives. Linguistically, fact complements can be distinguished from Propositions and clauses referring to Eventualities. Peterson (1997) offers two keys to identifying them: (i) grammaticality of that-S clausal complements and (ii) grammaticality of wh- clausal complements. To identify the complement clause of a given sentence, we test by substituting complements of the key forms. To begin, consider the pattern for Eventualities: a clausal complement refers to an Eventuality if both substitution tests destroy grammaticality. (7) illustrates with an initial gerundive clausal complement:
ASPECTUAL ENTITIES AND TENSE IN DISCOURSE (7) a. b. c.
229
Mary's refusing the offer was followed by silence.
* That Mary refused the offer was followed by silence. * What Mary refused was followed by silence.
While (7a) is grammatical, neither type of substitution preserves grammaticality. This shows that the complement of (7a) refers to an Eventuality. In contrast, a clausal complement refers to a Fact if both substitutions preserve grammaticality: (8) a. b. c.
Mary's refusing the offer was significant. That Mary refused the offer was significant. What Mary refused was significant.
All three versions are grammatical, showing that the complement of (8a) refers to a Fact. And finally, a clausal complement refers to a Proposition if the first substitution test preserves grammaticality and the second does not: (9) a. b. c.
Mary's refusing the offer was inconsistent. That Mary refused the offer was inconsistent. * What Mary refused was inconsistent.
Whle (9a) and (9b) are grammatical, the substitution of a wh-complement in (9c) destroys grammaticality. This shows that the complement of (9a) refers to a ~ro~osition.~ These substitution tests are sometime difficult to apply. They may require the analyst to make changes in a sentence to obtain the appropriate base form. To do this, Peterson (1997, p. 75) proposes 'vendlerization', which involves 'pseudotransformations' that derive a nominal complement from an underlying full-sentence structure; and its opposite, a way of obtaining the underlying full-sentence structure. The changes may go beyond the standard notion of transformation. A quantificational test that distinguishes complements referring abstract entities is proposed by Asher (1993).' These tests and distributional patterns are quite effective in identifying complements referring to abstract entities. It is important to note that the characterization presented here is limited to clausal complements that refer to abstract entities. Sentences directly expressing facts and propositions are not distinct linguistically. In spite of this very real limitation, the category is useful in distinguishing the atemporal discourse modes. This completes the list of types of situation entities. Different types of entities predominate in passages of different modes, as summarized below:
CARLOTA S. SMITH
Predominant situation entities and discourse modes Narrative : eventualities Report: eventualities, general statives Description: states, ongoing events Information: general statives Argument : abstract entities, general statives 4. TEXT PROGRESSION IN THE DISCOURSE MODES 4.1. Temporal and atemporal text advancement All texts advance through a structure. We advance through the episodes of a story, the stages of an argument, the classifications of an informative text. Passages of the temporal modes advance as location-temporal or spatial-changes. The text modes of Argument and Information are not temporally organized, though they may include eventualities that have temporal locations. Atemporal texts progress with metaphorical changes of location through the information space of the text. (1 1) gives a classification of the modes as temporal or atemporal, according to the predominant types of entities: (1 1)
Temporality of the modes Temporally located, dynamic: Narrative, Report Temporally located, static: Description Atemporal: Information, Argument
Narrative advances through narrative time, with situations related to each other. Description is static temporally; it advances by changes in spatial location, within the scene described. In Reports, situations are related to the time of report, often the present (Speech Time) and advancement involves a change of time. There are three principles of tense interpretation, according to the discourse mode of a passage: Continuity, Anaphora, and Deixis. In Narrative, situations are related to each other, and tense conveys continuity. In Report, situations are related to Speech Time and tense is deictic. In Description, situations are related to an already-established time and tense is anaphoric. Tense interpretation in the atemporal modes is deictic. The principles of interpretation are modeled using an extended Reichenbach framework. In texts with the principle of tense continuity, Reference Time advances according to bounded events or temporal adverbials; see 4.2.1 below. In texts with the principle of deictic tense, Reference Time changes with different relations to Speech Time; see 4.2.2. In texts with the principle of tense anaphora, Reference Time is simultaneous with a previously established time; see 4.2.3. To calculate temporal location in a discourse passage one must have access either directly or indirectly to the discourse mode of the passage. I do not have space to develop the formal analysis here; see Smith (2003) for a detailed account. To understand text advancement in the atemporal modes we need something
ASPECTUAL ENTITIES AND TENSE IN DISCOURSE
23 1
other than dynamism. We can find such a principle in the notion of a spatial domain that underlies all others. I will use the notions of metaphorical location and metaphorical motion. The semantic domain of an atemporal discourse can be seen as terrain to be traversed: a metaphorical space. he discourse advances as key reference moves metaphorically from one part of the domain to another. We need the complexity of space to model metaphorical motion. Space is not unidimensional, like time; rather, it allows directions of various kinds. Similarly direction in a text domain can be hierarchically up or down, lateral, etc. Texts are organized according to the domain and the particular focus of a given text. There are conventional organizing principles such as hierarchy, geography, chronology, cause. Understanding the way domains are organized is itself an interesting and difficult problem, beyond the scope of this discussion. I will identify for each clause a key referent that is semantically central; I then look for the metaphorical location of.that referent, called the Primary Referent. For events the Primary Referent is that referent that moves or changes; for states, it is the referent to which a property is ascribed or location maintained. This approach owes a great deal to the work of Talmy (1985, 2000). Motion and location may be literal or metaphorical. States are maintained rather than changed, focusing on what is predicated of the primary figure and how the components of the situation pertain to the figure. Similar ideas were put forth by Gruber (1965), and developed more fully in localist theory. The basic insight is that "...the formalism for encoding concepts of spatial location and motion, suitably abstracted, can be generalized to many other semantic fields" (Jackendoff, 1990, p. 25). Examples of the parallelism between the spatial and other semantic fields include possession (The inheritance went to Philip), ascription of properties (The light changed from green to red), and scheduling (The meeting was changed from Monday to Tuesday); the examples are Jackendoff s. The notion of primary referent is not dependent on surface structure. Rather, it depends on the type of situation expressed in a text, factoring out the linguistic features that depend on presentation or perspective. Thus the primary referent of a clause is the same whether the clause is active or passive. In Mary opened the door and The door was opened by Mary, 'the door' is primary referent in both cases. I list some criteria for identifymg the primary referent of a clause, based on the intuitive notions given above. ~ x a m ~ lare e s taken from the texts of this study. (12)
Criteria for Primary Referent of Events The primary referent is that entity in an event which (i) Undergoes a change of state The high school outsider becomes the more successful adult. (ii) Is causally affected by another participant The national outpouring has forced us to confront the situation. (iii) Doesn't exist independently of the event High school students present and past have come forward with stories about cliques and an artificial world. (iv) Moves or otherwise changes
CARLOTA S. SMITH
Criteriafor Primary Referent of States The primary referent is.that entity in a state which is (i) Literally or metaphorically located Dragons are usually arranged almost heraldically round a conceptual center point. (ii) Dependent on the situation for existence The predominant output was the white ware with transparent ivory-toned glaze which made the kilns famous. (iii) Figure relative to a Ground A group of kilns is northeast of Ch'ang-an, the capital city of the T'ang Dynasty. (iv) Has a property ascribed to it The most important kilns are those at Tao-chu in Shensi. The more specific criteria are applied before the general. Criterion (13iv) for instance, holds of the subject in any state sentence. Primary referents can usually be identified in clauses according to the sernanticsyntactic notion of thematic role. The primary referent usually coincides with the Theme, or Patient, argument. I assume that information about the Theme argument is available in the analyzed surface structure of a sentence, along with information about verb class and other features. Analyzed surface structures are the input to the construction rules of Discourse Representation Theory. 4.2. Passages of the discourse modes In this section I demonstrate the analysis outlined above for each of the modes. The examples have situation entities and text progression characteristic of the mode in question. 4.2.1. Narrative The main types of entity introduced in a narrative are eventualities, specific events and states. They are related to each other. Narrative presents a sequence of consequentially related events and states, and the order in which they occur is crucial for understanding (Moens, 1987). Dynamism is essential to this sequence: we understand a narrative to consist of events that occur one after the other in time. The interpretation of sequence is due to linguistic forms conveying that the beginning of one situation follows the end of another. The forms are mostly aspectual: narrative time advances with perfective event sentences (and with explicit temporal adverbials), and fails to advance otherwise. This is the basic finding of discourse dynamics (Hinrichs, 1986; Kamp & Rohrer, 1983; Partee, 1984). I take it that the default is sequence; for simplicity I ignore flashbacks, changes of scale and so forth. Bounded events are conveyed by event verb constellations and the simple verb form in English. Unbounded eventualities are ongoing events-in the progressiveand states. For discussion see Smith (199111997). (14) is an example of a narrative with the relevant information displayed. Events and states are indicated by
ASPECTUAL ENTITIES AND TENSE IN DISCOURSE
233
subscripts for each clause. The tensed clauses are numbered; in sentences with more than one clause, the clauses are lettered. Clauses that advance narrative time are marked with an arrow +: (14)
l E --+ A few days later I called on Dr P and his wife at home, with the score of the Dichterliebe in my briefcase and a variety of odd objects for the testing of perception. 2aE --+ Mrs. P showed me into a lofty apartment, bs which recalled fin-de-sibcle Berlin. 3a A magnificent old Bosendorfer stood in state in the centre of the room, and bs all around it were music stands, instruments, scores. 4aE --+ Dr. P came in, a little bowed, bE and --+ advanced with outstretched hand to the grandfather clock, cE but, hearing my voice, --+ corrected hlmself, dE and + shook hands with me. 5aE --+ We exchanged greetings bE and + chatted a little of current concerts and performances. 7 Diffidently, aE --+ I asked him bs if he would sing.
The continuity interpretation of tense is indicated below, considering only events that advance the narrative. Tense is conventionally past in narrative, with changes only to past perfect or future-in-past. The first clause of a narrative indicates a Reference Time (RT) that precedes Speech Time (< SpT). After that, a bounded event advances the narrative, so that one RT follows another (RT2 > RT1): (15)
Continuity interpretation of tense E 1.................E2....... . ....................E3........... RT 1< SpT RT2 >RT 1 RT3 > RT2
States and ongoing events do not advance RT and are not treated here. They are interpreted by a limited version of the principle of tense anaphora, presented below. 4.2.2. Report In Reports, eventualities or general statives are related to Speech Time rather than to each other. Time and space adverbials are common in reports, as are changes in tense. Significantly, the events of a report could appear in different order without making a real difference, unlike narrative (cf. Caenepeel, 1995). Tense and adverbials are Deictic, oriented to Speech Time. Below is a time line giving the deictic interpretation of tense.
234
(16)
(17)
CARLOTA S. SMITH l E+ A week that began in violence ended violently here, with bloody clashes in the West Bank and Gaza and intensified fighting in Southern Lebanon. 2E + Despite the violence, back-channel talks continued in Sweden. 3s -+ Israeli, Palestinian and American officials have characterized them as a serious and constructive effort. 4s + Israel is offering'as much as 90 percent of the West Bank to Palestinians. Deictic interpretation of tense E l ..........E2 .........S1.......S2........... RT