The Theory and Practice of Language Faculty Science 9783110724790, 9783110724677

The book demonstrates that it is possible to study the language faculty with the core scientific method, i.e., by deduci

220 14 33MB

English Pages 472 Year 2022

Report DMCA / Copyright

DOWNLOAD PDF FILE

Table of contents :
Series Preface
Preface
Contents
Part 1: The Past History of Our Attempts to Detect C-command
From Compatibility to Testability – Some Historical Background
On Non-individual-denoting So-words
Research Heuristics in Language Faculty Science
Part 2: The Correlational Approach
The Key Tenets of Language Faculty Science
Detection of C-command Effects
Replication: Predicted Correlations of Judgments in Japanese
Predicted Correlations of Judgments in English
Implementing Experiments on the Language Faculty
Part 3: LFS as an Exact Science
Language Faculty Science and Physics
Conclusion: Why Language Faculty Science?
Index
Recommend Papers

The Theory and Practice of Language Faculty Science
 9783110724790, 9783110724677

  • 0 0 0
  • Like this paper and download? You can publish your own PDF file online for free in a few minutes! Sign Up
File loading please wait...
Citation preview

The Theory and Practice of Language Faculty Science

The Mouton-NINJAL Library of Linguistics

Edited by Yukinori Takubo Haruo Kubozono

Volume 3

The Theory and Practice of Language Faculty Science Edited by Hajime Hoji, Daniel Plesniak and Yukinori Takubo

ISBN 978-3-11-072467-7 e-ISBN (PDF) 978-3-11-072479-0 e-ISBN (EPUB) 978-3-11-072489-9 ISSN 2626-9201 Library of Congress Control Number: 2022941713 Bibliographic information published by the Deutsche Nationalbibliothek The Deutsche Nationalbibliothek lists this publication in the Deutsche Nationalbibliografie; detailed bibliographic data are available on the internet at http://dnb.dnb.de. © 2023 Walter de Gruyter GmbH, Berlin/Boston Cover image: piranka/E+/Getty Images Typesetting: Integra Software Services Pvt. Ltd. Printing and binding: CPI books GmbH, Leck www.degruyter.com

Series Preface The Mouton-NINJAL Library of Linguistics (MNLL) series is a new collaboration between De Gruyter Mouton and NINJAL (National Institute for Japanese Language and Linguistics), following the successful twelve-volume series Mouton Handbooks of Japanese Language and Linguistics. This new series publishes research monographs as well as edited volumes from symposia organized by scholars affiliated with NINJAL. Every symposium is organized around a pressing issue in linguistics. Each volume presents cutting-edge perspectives on topics of central interest in the field. This is the first series of scholarly monographs to publish in English on Japanese and Ryukyuan linguistics and related fields. NINJAL was first established in 1948 as a comprehensive research organization for Japanese. After a period as an independent administrative agency, it was re-established in 2010 as the sixth organization of the Inter-University Research Institute Corporation “National Institutes for the Humanities”. As an international hub for research on Japanese language, linguistics, and Japanese language education, NINJAL aims to illuminate all aspects of the Japanese and Ryukyuan languages by conducting large-scale collaborative research projects with scholars in Japan and abroad. Moreover, NINJAL also aims to make the outcome of the collaborative research widely accessible to scholars around the world. The MNLL series has been launched to achieve this second goal. The authors and editors of the volumes in the series are not limited to the scholars who work at NINJAL but include invited professors and other scholars involved in the collaborative research projects. Their common goal is to disseminate their research results widely to scholars around the world. This volume represents the cumulative efforts of Hajime Hoji, his students, and his colleagues over several decades to establish how the language faculty can be studied in a definite and reproducible manner consistent with the basic scientific method. A picture of how this can be done, and indeed is being done, convincingly emerges through the combined endeavors of the book’s contributors. The core concepts articulated in this volume will thus surely form the basis for future investigations into the language faculty for decades to come. Yukinori Takubo Haruo Kubozono

https://doi.org/10.1515/9783110724790-202

Preface This book aims to show that it is possible to accumulate knowledge about the language faculty by the basic scientific method, i.e., by deducing definite predictions from our hypotheses and obtaining and replicating experimental results precisely in line with such predictions, as is done in other scientific disciplines such as physics (what we may call “the method of exact science”). We call this endeavor Language Faculty Science, LFS for short. Modern linguistics has often been said to be a scientific study of language, and indeed essentially every introductory linguistics textbook makes some version of this claim. It is rather rare, however, that we find serious or in-depth discussion in such books about what is meant by “language” and what kind of activities are involved in “science”. It is doubtful that every phenomenon having to do with language can be studied by a method of exact science analogous to that of physics. In order to determine what linguistic phenomena can be studied by such a method, it is necessary first to understand what kinds of activities are involved in “science” and what kinds of problems will have to be dealt with in order to apply the scientific method to linguistic phenomena and obtain meaningful results. Since the earliest days of generative grammar, Chomsky has maintained that we must study linguistic competence as natural scientists study their subject matters, claiming that it is possible to do so.1 What do we mean by linguistic competence? Barring any serious impairment, every member of the human species is able to produce and comprehend the language(s) to which they are exposed. Underlying this ability of ours to relate linguistic form, i.e., sounds/signs, and meaning is the language faculty. It is hypothesized that the language faculty in its initial state (sometimes called Universal Grammar) is uniform across the members of the species and, in its steady state (sometimes called a person’s I-language), where its “maturational” growth has stopped, it varies in accordance with one’s linguistic experience, within the limit imposed by the genetic endowment. I in I-language stands for “internal” and “individual” (Chomsky 1995: 13 and elsewhere). The basic property common to all I-languages that has so far been studied by the method of exact science is the human computational ability to handle the “discrete infinity” of language, that is, to relate finite sounds and signs to infinite number of sentences with forms and meaning. Other aspects of the linguistic competence seem to involve too many variables for this purpose, and as such, none of them seem to have been

1 See Section 3 of the introduction to Chomsky (1975), drawn from an unpublished 1955–56 manuscript. https://doi.org/10.1515/9783110724790-203

VIII 

 Preface

isolated in ways that would make them suitable objects of inquiry for research that pursues the method of exact science. The concept Merge, which has been proposed for the purpose of accounting for discrete infinity (via the imposition of hierarchical structures) is an operation that combines two elements, each having form and meaning, and forms one; its recursive application makes it possible to generate recursively enumerable hierarchical structures. This conceptualization can be understood as an accomplishment of the generative enterprise, which Chomsky founded as an attempt to pursue linguistic research as a scientific discipline. One might, however, reasonably suggest that the existence of recursive Merge and the hierarchical structures that arise due to its application, though they have been assumed, have never been (thought possible to be) subjected to the empirical testing via the method of exact science This volume is concerned with the demonstration of the existence of c-command, i.e., the detection of c-command effects, so as to demonstrate the existence of recursive Merge. The concept of c-command itself was proposed in the 1970s, and its critical relevance/significance has been recognized in relation to the phenomenon of bound variable anaphora (BVA), among other phenomena. With the understanding that it is defined in terms of Merge, c-command is now redefined as a more restricted concept. “x c-commanding y” is defined as in (1) by using Merge. (1) x c-commands y iff x is merged with z that contains y, z contains y iff (i) X and Y are the daughters of Z iff Z={X,Y} (ii) X contains Y iff: a. Y is X’s daughter or b. Y is the daughter of an element Z that X contains. x c-commands y iff y is a member of a set Z that is merged with x or y is a member of a subset of Z (or a subset of a subset of Z, etc.). The c-command relation is thus defined to hold between x and y only when Merge is recursively applied. In short, if we can show the existence of a phenomenon, such as a meaning relation between two elements that can arise only if there is a c-command relation between them, that constitutes the demonstration of recursive Merge. Identification/determination of properties of the computational system of natural language, which maps the form of sounds and signs to that of meaning, follows from such a demonstration. In other words, if we can identify phenomena that require c-command, that opens up a path for empirically testing the hypothesis about the existence of recursive Merge. In this volume, we refer to the theoretical and experimental identification of a c-command (accompanied by rigorous testability) as “c-command detection”.

Preface 

 IX

C-command detection can be a tool for empirically testing the thesis that recursive Merge exists at the core of the computational system of the human language. It has in fact been argued that for the variable-binding relation to obtain between X and Y, it is necessary for X to c-command Y. The existence of similar formal relations has been entertained in relation to distributive readings and coreference. Let us notate such meaning relations between X and Y as MR(X, Y). The attempt to establish the link between meaning relations and formal/structural constraints can be understood, from the current perspective, as an attempt at c-command detection. Since the identification of the acceptability of relevant meaning relations is based on native speaker judgments, however, the inherent variability of such judgments poses a perpetual problem. This includes the question of what the sources of such variation might be. If X c-commanding Y is a necessary condition for a given MR(X, Y), a sentence in which X does not c-command Y must necessarily be judged unacceptable under the MR in question; a structural relation like c-command either holds or does not hold, leading to a categorical distinction. In reality, however, there are cases where the aggregate judgments on the availability of a given MR in a given sentence are not categorical, pointing to the possibility that there are sources for MR(X, Y) other than c-command (with the variance in judgments attributable to variance in these other, non-command sources). Unless we successfully exclude the possibility of such non-c-command sources for MR(X, Y), c-command detection cannot be attained. The aim of this volume is to articulate how this is possible and thereby to argue that LFS as an exact science is possible. The volume consists of ten chapters, organized in the following three parts. Part 1: The Past History of our Attempts to Detect C-command Part 2: The Correlational Approach Part 3: LFS as an Exact Science Part 1, “The Past History of our Attempts to Detect C-command”, addresses how c-command detection was attempted in past works, pointing out its shortcomings, and suggesting solutions to the problems. Chapter 1 discusses works from 1985 to 2015 by Hajime Hoji, who has proposed the methodology pursued in this volume and is one of the three co-editors. The chapter explains, based on concrete illustration, how and why his attempts for c-command detection in those works fell short of being an instance of exact science. Though he does not go so far as to say this, I believe that problems pointed out and suggested solutions given in the chapter are equally applicable to works by researchers other than Hoji. Chapter 2 can be understood as our initial attempts to identify non-formal sources of the BVA(X, Y) interpretation, that is, BVA(X, Y) interpretations that are possible despite X not c-commanding Y. This phenomenon has been called

X 

 Preface

“Quirky binding” and it has been recognized as posing a problem when we try to use BVA for c-command detection. This chapter is written by Ayumi Ueyama, who gave the first systematic account of this phenomenon in her dissertation. The chapter is based on Appendix D of her 1998 dissertation (Ueyama 1998). A close analysis of this phenomenon is a necessary step for understanding what must be controlled for in order to use BVA for c-command detection. Chapter 3 presents, based on Chapters 1 and 2, a detailed illustration of the experimental method pursued in the Hoji research group up to 2015. The chapter does not (directly) address predicted correlations of judgments, which figure prominently in the approach pursued in the remainder of the volume, but it offers the reader basic knowledge for understanding how experiments can be conducted without the use of such correlations and what their limitations are. Part 2, “The Correlational Approach”, based on Part 1, provides a methodology for carrying out LFS research and presents actual research results. Chapter 4 addresses the basic tenets of the correlational methodology and how it can be put to practice; Chapter 5 discusses self-experiments in Japanese. Chapters 6 and 7 go over non-self-experiments in Japanese and English, respectively. Finally, Chapter 8 is intended to be a preliminary form of a manual for non-self-experiments in LFS. Part 3, “LFS as an Exact Science”, situates LFS in a broader context. Chapter 9 does so by comparing LFS with physics, addressing how categorical predictions and their experimental testing are possible in LFS, despite the fact that measurement in LFS is qualitative rather than quantitative. Chapter 10 concludes the volume by summarizing the preceding chapters; it addresses how compatibility-seeking approaches can fail to make definite predictions and how the correlational methodology in LFS makes it possible to do so. I would now like to briefly talk about the evolution of the project of this volume. I was acquainted with Hajime’s research since the early 1990s when he was struggling with ideas that eventually coalesced into LFS; I have witnessed his pursuit of LFS as an exact science and his enthusiasm and effort over the years, and I have also done joint research with him, including some published works. During a meeting in my office in December of 2019, he explained the correlational methodology, writing down its core idea on the whiteboard; that idea came to be the basis of this volume. I was convinced then, based on my own self-experiments, that it is possible to obtain categorical judgments by using correlation. I in fact came to understand that this method was an explicit statement of the method he had been using since the 1990s as understood in the terms of the current approach. Since that point on, I had been hoping that his theory and concrete method of experiments for testing hypotheses be put in one place in some way and be published in the form of a book. Fortunately, Ayumi Ueyama, Emi Mukai, and Daniel Plesniak, (the latter of

Preface 

 XI

whom subsequently became one of the editors), have agreed to contribute chapters to this volume; they have written dissertations under Hajime’s supervision, having made their own contributions at different stages of LFS’s development. The book proposal was then approved as a volume in Mouton-NINJAL Library of Linguistics series by NINJAL and De Gruyter Mouton. Every chapter of the volume underwent internal review by contributors to the volume, and the three editors commented on every chapter regarding its content and style, and the submitted version thus created was reviewed by an external reviewer. The chapters were revised based on the external review, resulting in the final versions. I would like to thank the external reviewer for taking on the difficult task of reviewing an earlier draft of the volume, and Haruo Kubozono of NINJAL, co-editor of the series, Michaela Göbels, Birgit Sievert, and Kirstin Boergen of De Gruyter Mouton, for their generous help at various stages of the production of the volume. The editors and the authors would like to thank Kiyoko Kataoka, Teruhiko Fukaya, Audrey Bongalon, Junichi Iida, Asako Miyachi, Felix Qin, Carolin Scherzer, Shun Shiranita, and Yoona Yee for their contributions at various stages in the development of LFS and to the completion of this volume. We would also like to thank Yasuo Deguchi for giving us insightful comments regarding how LFS as an exact science may be placed in the context of the philosophy of science, and Jiro Gyoba for his suggestion that has led to the visual presentation in this volume of core ideas of LFS, including the Venn-diagram-based presentation of experimental results. Special thanks are due additionally to Emi Mukai, who did extra work in checking references and consistencies in formatting. This volume is the first book that presents a concrete illustration of how research that deals with a mental phenomenon, such as the computational system that is responsible for mapping between linguistic forms and meaning, can be pursued as an exact science. I hope many young researchers in the coming generations will join this enterprise. Yukinori Takubo, April 2022

References Chomsky, Noam. 1975. Logical structure of linguistic theory. New York: Springer. Chomsky, Noam. 1995. Language and nature. Mind 104(413). 1–61. Ueyama, Ayumi. 1998. Two types of dependency. Los Angeles, CA: University of Southern California dissertation.

Contents Series Preface  Preface 

 V

 VII

Part 1: The Past History of Our Attempts to Detect C-command Hajime Hoji From Compatibility to Testability – Some Historical Background  Ayumi Ueyama On Non-individual-denoting So-words 

 3

 33

Emi Mukai Research Heuristics in Language Faculty Science 

 57

Part 2: The Correlational Approach Hajime Hoji The Key Tenets of Language Faculty Science  Hajime Hoji Detection of C-command Effects 

 117

 163

Hajime Hoji Replication: Predicted Correlations of Judgments in Japanese  Daniel Plesniak Predicted Correlations of Judgments in English 

 329

Daniel Plesniak Implementing Experiments on the Language Faculty 

 357

 223

XIV 

 Contents

Part 3: LFS as an Exact Science Hajime Hoji and Daniel Plesniak Language Faculty Science and Physics 

 387

Hajime Hoji and Daniel Plesniak Conclusion: Why Language Faculty Science?  Index 

 449

 429

Part 1: The Past History of Our Attempts to Detect C-command

Hajime Hoji

From Compatibility to Testability – Some Historical Background 1 Introduction The main thesis of this volume is that we can accumulate knowledge about the language faculty by the basic scientific method. The language faculty is the component of the mind that underlies our ability to relate linguistic sounds/signs (henceforth, simply “sounds”) and meaning. One important feature of the basic scientific method is that we seek to deduce definite/categorical predictions from hypotheses and obtain and replicate experimental results precisely in line with such predictions. Because the language faculty is internal to an individual, the relevant investigation must, therefore, be concerned, at least at the most fundamental level, with an individual’s linguistic intuitions about the relation between sounds and meaning. As is discussed in some depth in Chapter 4, adopting the basic scientific method leads us to seek to state our hypotheses in terms of the basic structural concept derivable from the most basic hypothesis about the language faculty, due to Chomsky (1995 and 2017 among many other works), that there is a computational system of the language faculty whose sole structure-building operation is one that combines two discrete items to form a larger one. The operation is called Merge, and it is illustrated in (1). (1) {D,{C,{A,B}}} D

{C,{A,B}} C

{A,B} A

B

If A and B are merged, we get {A,B}, if C is merged with {A,B}, we get {C,{A,B}}, if D is merged with {C,{A,B}}, we get {D,{C,{A,B}}}, etc. {D,{C,{A,B}}} can be represented in terms of a “tree representation” as in (1).1

1 (1) and the exposition here are based on Plesniak 2022. https://doi.org/10.1515/9783110724790-001

4 

 Hajime Hoji

Language Faculty Science (LFS), by which we mean the research program that aims at the accumulation of knowledge about the language faculty by the basic scientific method, thus crucially relies on hypotheses formulated in terms of a basic structural concept derivable from Merge. We use “c-command”, as defined in (2), as such a concept.23 (2) x c-commands y if and only if x is merged with something that contains y.3 What is minimally required for the success of LFS research is, therefore, the reliable detection of effects of the c-command relation on the possibility of a certain meaning-sound pairing. Much of this volume is in fact concerned with what counts as reliable detection of c-command effects on the possibility of a given meaning-sound pairing. Roughly speaking, detection of c-command effects (henceforth simply “c-command detection”) takes the form of observing (i) effects of the absence of c-command and (ii) effects of its presence. “C-command” is a theoretical concept, and the language faculty is internal to an individual. “C-command detection” thus requires articulation of what (universal, and language-particular) hypotheses lead to rigorously testable predictions about c-command effects with respect to an individual. It further requires articulation of how the observed effects of c-command with a given individual can be replicated with other speakers, not only within the same linguistic community as the individual but also beyond the linguistic community in question. Chapter 4 of this volume provides relevant articulation and the general methodology proposed there is illustrated in subsequent chapters of this volume. For the purpose of our initial discussion in this chapter, we adopt (3) as our “definition” of “c-command detection”.4

2 See Chapter 4: Section 1 for the motivation of adopting Merge and the critical importance of the concept of c-command in a study of the language faculty by the basic scientific method. 3 For the definition of contain, I adopt (i) and (ii), based on Plesniak’s (2022: 2.2.2) formulation, with slight adaptation. (i) X and Y are the daughters of Z iff Z={X,Y} (ii) X contains Y iff: a. Y is X’s daughter or b. Y is the daughter of an element Z that X contains. 4 If an individual speaker’s judgments obtain in line with (32) below, for example, that constitutes c-command detection.

From Compatibility to Testability 

 5

(3) C-command effects are detected iff: A certain interpretation pertaining to two linguistic expressions X and Y is judged to be possible only if X c-commands Y.5 As discussed in some depth in Hoji 2015: Chapters 3 and 4, and also in Chapter 4 of this volume, checking whether or not a certain interpretation pertaining two linguistic expressions X and Y is possible in a given sentence containing X and Y is considered an instance of an experiment, where we test the validity of the hypotheses that predict the presence of c-command. Given that the language faculty is internal to an individual, experiments to detect c-command effects start with an individual. As will be discussed in some depth in Chapters 4–6 of this volume, conducting a self-experiment (an experiment where the researcher checks his/her own judgments) seems to be the most effective way to find out about properties of the language faculty. Hoji 1985, Hoji 2015, and works written in the time between the publication of these two constitute sustained (yet incomplete) efforts to detect c-command effects in Japanese;56 these efforts were, in fact more extensive and intensive than those in any other published works, as far as I am aware. I thus consider Hoji 1985 and Hoji 2015 (and intervening works), as the primary “historical background” for the discussion in this volume. The main points of this chapter are as follows: Hoji 1985 can be understood as an attempt to obtain c-command detection with certain meaning relations in Japanese, and its claims are based primarily on results of my self-experiment. As remarked in Hoji 2016: Section 2, my research subsequent to Hoji 1985 started out as an attempt to overcome a major shortcoming of Hoji 1985, namely that the empirical generalizations put forth (or adopted) there are often far from being robust, despite the fact that in much of the subsequent generative research, they have often been accepted and comprise one of the basic sets of generalizations in Japanese syntax, along with the proposed/assumed structural analyses for

5 This differs from the “definition” of c-command detection adopted in subsequent chapters, the presentation of which requires concepts that have not been introduced. 6 C-command detection, as discussed in this volume requires a “detector” of c-command effects, and the “detector” in question, which is not a physical detector, involves articulation of prediction-deduction, including the explicit statement of the hypotheses in question, the deduction of the prediction, the design of experiment, etc., as discussed in subsequence chapters of this volume. Hoji 1985 does not have such articulation. In this sense, what is attempted in Hoji 1985 does not count as an attempt at c-command detection, and one might call it an attempt to “identify” c-command effects. The use of ‘detect’ here with regard to Hoji 1985 is for the ease of exposition.

6 

 Hajime Hoji

the sentence patterns in question.7 Although subsequent research expanded the empirical coverage considerably (see below), including the empirical considerations that would make the experiment more reliable (see below and also Chapters  3–7 of this volume), it continued to rely on my own judgment as the main source of data.8 With my gradual realization that reliance on results of selfexperiments might not be effective in convincing other practitioners in the relevant field, I began to expend efforts in articulating methodology for non-self-experiments (experiments conducted on speakers other than the researcher him/ herself). The efforts resulted in Hoji 2015, which attempts to articulate: (i) how we deduce definite predictions about the judgments of an individual speaker on the basis of universal and language-particular hypotheses, (ii) how we obtain experimental results precisely in accordance with such predictions, and (iii) the actual “large-scale” “non-researcher-informant” experiments that had been performed in line with (i) and (ii). As noted in Chapter 4: Section 6.2, the results of “multiple-non-researcher-informant experiments” reported in Hoji 2015 in fact came quite close to obtaining experimental results in line with our categorical predictions, with the level of experimental success that, as far as I am aware, had not been attained previously. Problems remained, however, most notably because these “close” results still deviated from the predicted results in non-trivial ways. Recognizing that the problem persists for principled reasons and that we could not, generally, obtain experimental results in line with our categorical predictions in self-experiments, let alone non-self-experiments, if we focused on judgments on the availability of a particular meaning relation in isolation, has led to the correlational methodology to be introduced in subsequent chapters. We return to the approach taken in Hoji 1985 with regard to focus on self-experiments, but under the correlational methodology just alluded to, we develop methodology for non-self-experiments based on the methodology for self-experiments.9

7 The relevant issues have been extensively discussed in a series of works by A. Ueyama (including Ueyama 1998) and by J.-R. Hayashishita (including Hayashishita 2004), and Hoji 2003a provides a review. 8 As stated at the outset of this chapter, because the language faculty underlies our ability to relate sounds and meaning, and it is internal to an individual, investigation about the language faculty must be concerned, at least at the most fundamental level, with an individual’s linguistic intuitions about the relation between sounds and meaning. Experiments thus take the form of checking an individual’s linguistic intuitions, as discussed in some depth in Chapter 4 of this volume as well as in Hoji 2015. 9 This is discussed in some depth in Chapter 4 of this volume (particularly Section 6) and illustrated in Chapters 5 and 6 of this volume.

From Compatibility to Testability 

 7

2 Compatibility-seeking Research Hoji’s (1985) attempt to obtain c-command detection was in the context of arguing for (4b) over (4a), with regard to sentences of the form “NP-ga NP-o V” (roughly, Subject Object Verb), leaving aside the details.10 (4) a. Hypothesis 1: “NP-ga NP-o V” is represented as in (5). b. Hypothesis 2: “NP-ga NP-o V” is represented as in (6). IP

(5) VP NP1 -ga NP2 -o (6)

-ta V

IP VP NP1 -ga

-ta V'

NP2 -o

V

A predominant view in the relevant field at the time of Hoji 1985 adopted the hypothesis in (4a) and hence assumed that the structural representation of the surface string that corresponds to (7) in Japanese, as exemplified in (8), for example, was something like (5), rather than (6).11 (7) NP1-ga NP2-o V(-ta/-ru) (8) [NP1 Mary]-ga [NP2 susi]-o Mary-nom sushi-acc ‘Mary ate sushi.’

tabeta ate

10 Arguments for (4b) (and its more general version) are reviewed in Hoji 2003a: Section 2. 11 The point also applies to a surface string that corresponds to NP1-ga NP2-ni V(-ta/-ru), such as (i). (i)

[NP1

Mary]-ga [NP2 John]-ni Mary-nom John-dat ‘Mary tried to seduce John’

iiyotta approached

8 

 Hajime Hoji

This view was coupled with another prevailing conception at the time that a different “word order” such as “Object Subject Verb” (i.e., NP-o NP-ga V) was “freely generated” as a variant of (7), “reflecting” the absence of the “structural primacy” of the Subject over the Object, i.e., the absence of the asymmsetrical c-command relation between the Subject and the Object, as indicated in (5).12 Consider the sentence patterns in (9). (9)

a. [Y-no N]-ga X-o V b. Y-ga X-o V

The structural relations between X and Y in (9a) and (9b) are as indicated in (10) and (11), respectively, according to the two hypotheses in (4), suppressing the “case markers” such as -ga, -o, and -no, and with “α” added for the purpose of easy reference to a particular “node”. (10) Structures of (9a) a. According to (4a) α

Y

N

X

V

b. According to (4b) α Y

N

X

V

12 Similar remarks apply when we consider “word orders” among the “direct object” and the “indirect object”, along with the “subject”. Under the predominant view in question at the time, sentences that contain “subject”, “direct object” and “indirect object” can be “generated” “freely” with regard to the linear order among them, again reflecting the absence of asymmetrical c-command relation among them. One of the claims in Hoji 1985 is that “in the base order”, the “subject” asymmetrically c-commands the “indirect object”, which in turn asymmetrically c-commands the “direct object”.

From Compatibility to Testability 

 9

(11) Structures of (9b) a. According to (4a) α

Y

V

X

b. According to (4b) α Y

X

V

According to the Merge-based conception of structure-building in the language faculty, structures such as (10a) and (11a), which are based on the hypothesis in (4a), are not possible. Hoji 1985, which did not “have” Merge, tried to argue for (10b) and (11b), in line with the hypothesis in (4b), on an empirical basis. The definition of c-command used in Hoji 1985 is distinct from (2), as given in (12), which was one of the most widely accepted definitions of c-command while Hoji 1985 was prepared.13 (12) (From Hoji 1985: Chapter 1) X c-commands Y if neither dominates the other and the first branching node that dominates X dominates Y. A node X dominates another node Y iff X contains Y; see note 3 for “containment”. The first branching node that dominates X is marked as “α” in each of the trees in (10) and (11). Since α also dominates Y in (10a) and (11a), X c-commands Y in (10a) and (11a), hence in (9a) and (9b), according to (12). In (10b) and (11b), by contrast, α does not dominate Y; therefore, X does not c-command Y. That is according to (12), and also according to the definition of c-command in (2).14

13 As will be observed below, research based on the definition of c-command in (2) and that based on (12) can be compared meaningfully as long as we focus on what predictions they make in terms of schemata. This illustrates the point made in Hoji 2015: Chapter 3: Section 6: note 33 that schema-based predictions transcend differences of frameworks. 14 It may be interesting to observe that while Merge-based ‘c-command’, as in (2), is (very close to being) part of “the hardcore” in the terms of Lakatos 1978 (see Chapter 9 of this volume), the “first-branching-node-based” ‘c-command’, as in (12), is not. As noted, the “flat-structure” hypothesis, as in (5), is not allowed to begin with under the Merge-based conception of the language faculty; it may, however, provide us with an interesting exercise as to what testable predictions it can give rise to, with what kinds of hypotheses.

10 

 Hajime Hoji

If we have a meaning relation holding between X and Y that must require X asymmetrically c-commanding Y, we can use its (un)availability to distinguish between the two hypotheses in (4). For such meaning relations, Hoji discusses BVA(X, Y) and DR(X, Y). Consider the English sentences in (13) and (14).15 (13) BVA(every engineer, his): a. every engineer praised his robot b. his robot praised every engineer c. his robot, every engineer praised (14) DR(every engineer, three robots): a. every engineer praised three robots b. three robots praised every engineer c. three robots, every engineer praised What is intended by BVA(every engineer, his) in (13a, c) and in (13b) is an interpretation like (15a) and (15b), respectively. (15) a. Each of the engineers in question praised his own robot(s). b. Each of the engineers in question was praised by his own robot(s). What is intended by DR(every engineer, three robots) in (14a, c) and in (14b) is an interpretation like (16a) and (16b), respectively. (16) a. Each of the engineers in question praised a distinct set of three robots. b. Each of the engineers in question was praised by a distinct set of three robots. If we use John-sae ‘even John’ and John to Bill ‘John and Bill’ in place of what corresponds to every engineer in (13) and (14), as is done in Hoji 1985, the intended BVA and the intended DR would be roughly like (17) and (18), respectively.

15 Another “meaning relation” discussed in Hoji 1985 (and also in Hoji 1990) is a “coreference” relation, hypothesized to be “regulated” by a “negative condition” (such that the “coreference” relation is not possible under certain conditions), often called Binding Condition C/D (see Hoji 2003a: note 7); see Ueyama 1998: Appendix C and Chapters 5 and 6 of this volume for how the “coreference” relation discussed in Hoji 1985 and 1990 is different from Coref(alpha, Y) as discussed in Chapters 5 and 6 of this volume.

From Compatibility to Testability 

(17)

 11

a. BVA(even John, his) Each of the individuals in question praised his or her own robot(s), including John who was least expected (among the individuals in question) of having the property of praising one’s own robot(s). b. BVA(John and Bill, his) Each of John and Bill praised his own robot(s).

(18)

a. DR(even John, three robots) Each of the individuals in question praised a distinct set of three robots, including John who was least expected (among the individuals in question) of praising three robots. b. DR(John and Bill, three robots) Each of John and Bill praised a distinct set of three robots.

Now consider the hypotheses in (19), whose validity Hoji 1985 assumes, rather than tries to argue for, with the definition of c-command as given in (12). (19)

a. BVA(X, Y) is possible only if X c-commands Y. b. DR(X, Y) is possible only if X c-commands Y.

With the hypothesis in (4b), combined with (19), we make the predictions as indicated in (20) because, according to (4b), X does not c-command Y in (9a) and (9b), as discussed above (20) a. Prediction based on the combination of (4b) and (19a): BVA(X, Y) is not possible in (9a): [Y-no N]-ga X-o V b. Prediction based on the combination of (4b) and (19b): DR(X, Y) is not possible in in (9b): Y-ga X-o V (4a), when combined with (19a) and (19b), on the other hand, do not lead to the predictions in (20) because, according to (4a), X does c-command Y in (9a) and (9b), as also discussed above. It is claimed in Hoji 1985 that the predictions in (20) are borne out, by and large, with various combinations of X and Y, which is taken to constitute evidence

12 

 Hajime Hoji

for (4b). For example, it is claimed in Hoji 1985 that BVA(John to Bill, ei) and BVA(John mo Bill mo, ej) are impossible in (21) while they are possible in (22).16 (21) a.

(=Hoji 1985: Chapter 4 (28a)) ✶ [NP[S ei ej kakumatte ita] otokoj]-ga [VP John to Billi -o uragitta] was protecting man-nom and -acc betrayed (✶The man that hei was protecting betrayed John and Billi.)

b. (=Hoji 1985: Chapter 4 (28e)) ✶ [NP[S ei mukasi ej osieta] senseii]-ga before taught teacher-nom [VP imademo John mo Bill-moj oboete iru/kiratte iru] (koto) even now also also remember/hates (✶The teacher who taught himj years ago still remembers/hates [John as well as Bill]i.) (22) a.

(=Hoji 1985: Chapter 4 (29a)) John to Billi -ga [VP [NP[S ei ej kakumatte ita] otokoj]-o uragitta](koto) and    -nom was protecting man-acc betrayed (John and Billi betrayed the man that hei was protecting.)

b. (=Hoji 1985: Chapter 4 (29e)) John mo Bill-moi [VP imademo [NP[S ei mukasi ej osieta] gakuseii]-o even now before taught student-acc oboete iru/kiratte iru] (koto) remember/hate ([Both John and Bill]i still remember/hate the student that hei taught years ago.) It is further claimed in Hoji 1985 that the impossibility of BVA(X, Y) in (9a) and that of DR(X, Y) in (9b) cannot be due to Y preceding X because BVA(X, Y) is possible in (23a), in contrast to (9a), and DR(X, Y) is possible in (23b), in contrast to (9b).17 (23) a. [Y-no N]-o X-ga V b. Y-o X-ga V 16 “e” (standing for “empty element/category”) is used to represent the “empty nominal” (assumed to be present) in the embedded subject and object positions in (21) and (22), where an overt nominal is missing. The subscript “i” is used in “John to Billi” and “ej” to indicate the relevant meaning relation is intended to hold between these two. 17 The following discussion focuses on BVA.

From Compatibility to Testability 

 13

Sentences such as (24) are provided as an illustration. (24) a. (=Hoji 1985: Chapter 4 (38e), simplified) [NP[S ei [VP Ginza-de ej katta]] yubiwaj]-o John to Billi-ga suteta (Lit. [The ring that hei bought at Ginza], [John and Bill]i threw away.) b. (=Hoji 1985: Chapter 4 (38b), simplified) [NP[S ei [VP Ginza-de ej katta]] yubiwaj]-o John-mo Bill-moi suteta ([The ring that hei bought at Ginza]k, [both John and Bill]i threw away.) In short, the main argument in Hoji 1985 is that if we assume (19), the judgments reported on the relevant sentences instantiating the patterns in (9) and (23) suggest the correctness of (4b), not (4a).18, 19 Some instances of X for BVA(X, Y) and DR(X, Y) considered in Hoji 1985 are listed in (25), including the first two mentioned above. (25) a. b. c. d. e. f. g.

John-sae ‘even John” John to Bill ‘John and Bill’ John ya Bill ‘John and Bill and so on’. John-mo ‘John also’ John-mo Bill-mo ‘both John and Bill’ daremo ‘everyone’ dono hito-mo ‘every person (Lit. whichever person)’

The three choices for Y of BVA(X, Y) discussed in Hoji 1985 are listed in (26). (26) a. kare ‘he’ b. zibun ‘self’

18 Extending the empirical coverage to the “ni-marking” verb (the verb that “marks” its object with -ni instead of -o), the sentence patterns with the “direct object” and the “indirect object”, and sentences involving adverbials/adjuncts (indicating place/time/etc.), Hoji 1985 makes a more general proposal that the Japanese phrase structure is strictly binary; see Hoji 1985: Chapter 1, note 15 and Hoji 2016: note 2 for how this general proposal in Hoji 1985 may be related to Kayne’s (1981, 1984; Introduction) hypothesis that binary branching is the only permissible branching in any language. 19 While Hoji 1985 is about “c-command”, “c-command” in Hoji 1985 was not a concept directly derived from Merge (as the only structure-building operation in the Computational System of the language faculty), which was introduced in Chomsky 1995 and further motivated conceptually in Chomsky 2017.

14 

 Hajime Hoji

c. pro (the so-called “empty pronoun”, i.e., the “covert” nominal that is hypothesized to “appear” in place of an overt nominal, which we represent here as pro)20 As discussed in some depth in Hoji 2003a, there is a great deal of judgmental variation (and instability) among speakers regarding the possibility of BVA(X, Y) in sentences of the forms in (9). One of the reasons for this is crucially that what is used as Y of BVA(X, Y) in Hoji 1985 is (26c).21 With the use of pro as Y of BVA(X, Y), we cannot rule out the possibility that pro is taken as “plural-denoting”, as it is possible to understand John-ga pro hihansita ‘John criticized pro’ as meaning “John criticized a group of individuals”. The use of pro as Y of BVA(X, Y) then gives rise to an issue similar to how to differentiate (28a) from an interpretation we might be able to “get to” through (28b), as an interpretation for (27), because the situation depicted by (28b) is compatible with the situation depicted by (28a). (27)

Every boy praised their robots.

(28) a. Each of the boys in question praised his own robots. b. The boys in question praised their (=the boys’) robots. This makes the availability of BVA(every boy, their) in (29a) much more difficult to “assess” than that of BVA(every boy, his) in (30a). (29) a. Their robots praised every boy. b. Their robots, every boy praised. (30) a. His robots praised every boy. b. His robots, every boy praised. 20 This corresponds to “e” with ‘e’ standing for ‘empty (category)’ in (21), (22), and (24). 21 It was widely accepted in the field, although without having been tested “experimentally”, that kare cannot be Y of BVA(X, Y), and that makes kare not usable for the purpose at hand. Zibun, on the other hand, seems to have different issues, due to its “subject-orientation” (it “prefers” to have its antecedent in a “subject” position), its “sensitivity” to non-formal, i.e., non-structural, factors, such as “points of view,” “empathy”, and the like, as is well known. Kare and zibun were deemed in Hoji 1985 not suitable for investigation concerned with c-command detection. Recent research has shown, however, that it is possible to use kare as possible Y of BVA(X, Y) once we adopt the correlational methodology to be discussed in subsequent chapters, including Chapters 4–8; see Chapter 5: 5, for example. Something similar might also happen to zibun in the future; see Plesniak’s (2022) discussion of ziji in Mandarin Chinese and casin in Korean for some initial steps towards using such elements for the purposes of c-command detection.

From Compatibility to Testability 

 15

Contrary to what is reported in Hoji 1985, I (now) find BVA(X, pro) to be possible in sentences in Hoji 1985 where X, according to hypotheses such as (4b), fails to c-command pro. For example, I find the BVA in (21) available. While that is not particularly surprising in light of the above considerations, it makes one wonder how the alleged generalizations reported in Hoji 1985 were made. One may further wonder how it came to pass that they were not critically examined before their validity was widely accepted (in the relevant field) as empirical support for hypotheses such as (4b), and as a basis for further empirical (and theoretical) discussion. One might suggest that the judgments reported in Hoji 1985, which seem to have been accepted by other speakers of Japanese in the field, at least to the extent that they had become part of the “standard generalizations”, are “expert judgments” such that “experts” can overcome the problems noted above, such as the possibility of pro being taken as “plural-denoting”. Be that as it may, the fact that I can now readily accept what appears to be the BVA reading in (21) would, under this suggestion, have to indicate that I have ceased to be an expert in the relevant respect (which I would humbly like to deny). Furthermore, my own case aside, if we do not have a way to replicate such “expert judgments” in some (testable) way in other speakers, the strength of such “empirical evidence” remains rather unclear; see Hoji 2003a: 2.2.2, especially note 14. A more reasonable explanation for the creation and longevity of these unsupported generalizations seems to be related to the general research orientation adopted and pursued in Hoji 1985, which we can broadly characterize as compatibility-seeking research, as opposed to testability-seeking research. The difference between testability-seeking research and compatibility-seeking research can be understood in relation to what is typically considered as supporting evidence for hypotheses in each type of research. Testability-seeking research tries very hard to look for ways in which its hypotheses can be shown to be incorrect. What constitutes evidence in support of a set of hypotheses under the testability-seeking research is the definite prediction(s) derived from the hypotheses having survived a rigorous attempt at disconfirmation.22 Recall the hypothesis in (4b) and (19a), repeated here. (4)

b. Hypothesis 2: “NP-ga NP-o V” is represented as in (6).

22 See Hoji 2016 for more discussion about the two types of research orientation in question.

16 

 Hajime Hoji

(6)

IP VP NP1 -ga

-ta V'

NP2 -o (19)

a.

V

BVA(X, Y) is possible only if X c-commands Y.

The hypothesis in (4b) is compatible with the general, compositional-semanticsbased, approach to meaning, where binary composition is standard. Likewise, it is also compatible with the general structural analyses of languages like English, where this rough structure for subjects, objects, and verbs has long been adopted. Likewise, the assumption that (19a) is (correct and) applicable to Japanese would make Japanese more “compatible” with other languages (such as English), where (19a) has been “successfully applied”. Consider now the alleged generalization that BVA(X, Y) is not available in (31a) but it is in (31b). (31) The alleged generalization defended in Hoji 1985: BVA(X, Y) is not possible in (a) but it is in (b). a. (=(9a)) [Y-no N]-ga X-o V b. (=(23a) [Y-no N]-o X-ga V We can ask what set of judgments would be taken as a basis for the generalization. According to the hypothesis in (19a), we predict speaker judgments that any sentence instantiating (31a) (and more generally (32a)) disallows BVA(X, Y) (while sentences instantiating (31b) (and more generally (32b)) allow it23) with any combination of X and Y for BVA(X, Y) (that would qualify as “BVA(X, Y)”) for any native speaker of Japanese.

23 The difference between the nature of the prediction about (32a) and the nature of the prediction about (32b) is addressed in some depth in Hoji 2015: Chapter 2: 2.4, and also in Chapters 5 and 6 of this volume.

From Compatibility to Testability 

(32)

 17

Generalized version of (31): BVA(X, Y) is not possible in (a) but it is in (b). a. [[. . . Y . . . ]N]-ga X-o V b. [[. . . Y . . . ]N]-o X-ga V

Research that seeks rigorous testability would test the predictions in this way, with any combination of X and Y, with any speaker (who is willing to judge the relevant sentences), with various types of instantiations of (32); it would even consider effects of the sentences being embedded in a larger structural/syntactic or pragmatic context. Any judgment from any speaker that a sentence instantiating (32a) is acceptable with BVA(X, Y) would be considered a disconfirmation of the prediction and would force us to look into what has led to such a judgment.24 Compatibility-seeking research, which we understand Hoji 1985 to be an instance of, takes a radically different approach. To begin with, the testing of the prediction is not systematic or “exhaustive”, it is done “selectively”. When some combinations of X and Y lead to a judgment not as predicted, other combinations are checked. When the prediction is “confirmed” with some other combination of X and Y, for example, it is taken as evidence for the hypotheses that have led to the prediction. Compatibility-seeking research thus seeks confirming evidence for its hypotheses instead of trying to make (rigorous) attempts at disconfirmation of the predictions derived from those hypotheses.25 That is how compatibilityseeking research accumulates research results.26 To the extent that the predictions of the form in (32) are due to (19a), we predict a fundamental asymmetry between (32a) and (32b). With the necessary condition for BVA as given in (19a), the disconfirmation of (32a) would only take one clear “instance” where BVA(X, Y) is possible without X c-commanding Y, but

24 As remarked in Chapter 6: note 77, that would in fact provide us with an opportunity for discovery. 25 This is like “fooling oneself” and against “a kind of leaning over backwards” as Feynman (1985: 340–343) put it in his “Cargo Cult Science”. 26 The parasitic-gap analysis in Hoji 1985: Ch. 2 of what would later be called the “A-Scrambling” construction in Japanese seems to be a good example of another manifestation of compatibility-seeking research, where an analogy is made of some loosely understood “phenomenon” in a non-English language to a phenomenon in English that is analyzed in highly theoretical terms. If the descriptive generalization in English is not as robust as one wishes to begin with, and if aspects of the theoretical characterization of the phenomenon are not independently motivated on empirical grounds, the “theoretical characterization” of the loosely understood phenomenon in Japanese “analogous to English” is bound not to survive the test of time; see Hoji 2016: 6.2. The fact that virtually no substantive reference was made to the term “parasitic gap” in Japanese in subsequent years, even though the relevant paradigms continued to be discussed in the field fairly extensively, seems quite suggestive of the current state of affairs.

18 

 Hajime Hoji

(32b) cannot be disconfirmed. Testability-seeking research thus focuses on (32a) when attempting to achieve testability.27 Compatibility-seeking research does not focus on (32a). It usually focuses on instances of “contrast” between sentences of the pattern in (32a) and those of the pattern in (32b), and it sometimes even takes isolated “confirmation” of the prediction of the form in (32b) as its primary supporting evidence. Since BVA(X, Y) has played such a central role in the discussion of the so-called “scrambling” in Japanese (as has DR(X, Y)), works on this topic provide us with nice illustration of the distinction between testability-seeking research and compatibility-seeking research; see Hoji 2006a. In order for a given hypothesis to have the chance to receive empirical support, it must first be possible for the hypothesis to give rise to a definite prediction in conjunction with other hypotheses, so that it is conceptually possible for the definite prediction to turn out to be wrong. Therefore, when a hypothesis is put forth in testability-seeking research, one of the first questions to be considered is how it can be put to rigorous empirical/experimental test, i.e., how its validity can be tested experimentally, and how the hypothesis can be invalidated. Under this approach, the formulation of hypotheses and even the choice of the specific research topic are severely limited by the desire to seek rigorous testability. Compatibility-seeking research, on the other hand, does not make (rigorous) attempts at disconfirmation of the predictions made under its hypotheses. Instead, it typically seeks confirming evidence for its hypotheses. What constitutes confirming evidence depends in part upon how rigorously one carries out one’s research. It may be the identification of some pairs of sentences that seem to exhibit a particular contrast in the direction of what is suggested by the hypotheses in question, often despite the fact that it can be easily shown that the prediction of the form in (32a) is disconfirmed; see Hoji 2016: Section 6 for related discussion. If one pursues compatibility-seeking research when dealing with a particular language, one may try to express/describe, as Hoji 1985 does, some “phenomena” in the language in the terms of the “theory” one adopts and considers what could be said about the “theory” on the basis of one’s “findings” (as pointed out in Hoji 2016: Section 2). For example, (as I understand my past way of thinking) Hoji 1985 assumed the existence of pro in Japanese because that would make Japanese compatible with the standard analysis of sentence structure accepted in the field. How the postulation of the existence of “empty pronoun” leads to definite and testable predictions is rarely addressed in work following similar approaches

27 Predictions of this type is called “negative predictions” in Hoji 2009 and it is called ✶Schema-based prediction in Hoji 2015.

From Compatibility to Testability 

 19

to pro.28 In relation to (21), we addressed issues with regard to the possibility of “pro” being “plural-denoting”, as a possible source of speaker judgments contrary to the prediction in (32a). Another possible source is that what is missing overtly is indeed missing and there is no “pro”. As will be seen in Chapter 5 of this volume (Section 7.2) this issue is in fact quite centrally important when we consider “anti-locality” effects. The compatibility-seeking research orientation as addressed above seems to have remained in “syntax research”, and in my view it is actually even more conspicuous now than it was when Hoji 1985 was written. This seems to reflect the general absence of a well-articulated methodology for pursuing rigorous testability. I believe this is compounded by the claim elsewhere in the field that the absence of such a methodology is due to a principled reason, namely the nature of the subject matter with the claim being that it is not possible to obtain and replicate an individual speaker’s linguistic judgments that are categorical in nature. This, I reject; that it is in fact possible to do so is the thesis pursued by the methodology articulated in Chapter 4, and the viability of such a methodology is further illustrated in other chapters of this volume.

3 Towards Testability-seeking Research As remarked in Hoji 2016: 6.2, a reasonable interpretation of Hoji 1985 is that it tried to identify as good a probe as possible for discovering the universal properties of the language faculty through the investigation of the availability of BVA(X, Y) and DR(X, Y) in Japanese. The probes thus identified were used to argue for the thesis that the Japanese phrase structure is strictly binary branching. My own current judgments and works such as Ueyama 1998 and Hayashishita 2013, among other places, however, indicate that the alleged generalizations put forth in Hoji 1985 do not survive (minimally) rigorous empirical scrutiny. For example, many, if not most of, the sentence patterns that are predicted in Hoji 1985 to be unacceptable under BVA(X, Y) are found acceptable by many speakers.29 We thus lose the empirical generalizations (at least in the sense of them being empirically testable) that would have constituted compelling support of (4b) (and its “generalized” version).

28 Hoji 1987 is one of the few exceptions. 29 In fact, most, if not every, sentences reported in Hoji 1985 as unacceptable, under the “intended BVA” or under the “intended DR”, are acceptable to me now.

20 

 Hajime Hoji

Subsequent research (such as Hoji 1990, 1995, 1998, etc.) tried to remedy the situation. As noted, the problem with sentences like (21) is likely related to the fact that pro can be “plural-denoting”. If we can identify an overt nominal in Japanese that cannot be “plural-denoting”, the use of such a nominal as Y of BVA(X, Y) might thus lead to clear(er) judgments in support of (31), and hence for c-command detection. In the meantime, it was claimed in works such as Nishigauchi 1986 and Yoshimura 1989, 1992, that overt nominals such as sore ‘it’ and soko ‘that place’ (referring to “that company”, for example) can be Y of BVA(X, Y). To avoid issues arising from the possibility of the nominal intended as Y of BVA(X, Y) being “plural-denoting” (see above), Hoji 1990: Chapter 4 used the “split antecedence” test on overt nominals such as soko ‘it’ and soitu ‘that guy’.30 For example, the test checks whether sentences like (33) clearly disallow “split antecedence” as indicated below while the use of a clearly “plural-denoting” nominal, such as sono 2-sya ‘those two companies”, in the position of soko allows “split antecedence, for a given speaker (at a given time). (33) (=Hoji 1990: Chapter 4: (60), slightly adapted) a. ✶Toyota(1)-ga Nissan(2)-ni [S’ soko(1, 2)-no zyuugyooin-ga Toyota-nom Nissan-dat it-gen employee-nom issyoni pikunikku-o subekida to] teian-sita together picnic-acc should:do that suggested ‘Toyota(1) has suggested to Nissan(2) that their(1,2) employees should have a picnic together’ b.

Toyota(1)-ga Nissan(2)-ni [S’ Amerika-no oote kigyoo-ga Toyota-nom Nissan-dat America-gen major company-nom soko(1, 2)-to zyointo ventyaa-o sitagatteiru to] tugeta it-with joint venture-acc want:to:do that told ‘Toyota(1) has told Nissan(2) that a big American corporation wants to do joint venture with them(1,2)’



30 The test would be used as part of the “Sub-Experiments” in the experiments discussed in Hoji 2015 for checking whether nominals such as soko ‘it/that place’ and soitu ‘that guy’ and their a-NP counterparts (asoko ‘it’ and aitu ‘that guy’) cannot be “plural-denoting” for a given speaker (at a given time). The test is employed also for nominals such as sono N ‘that N’ and ano N ‘that N’, as will be discussed in Chapters 5 and 6 of this volume. “Sub-Experiments” as discussed in Hoji 2015: Sections 4.3 and 4.4, and “sub-preliminary experiments” in Chapter 6: Section 2.4 of this volume also test (i) and (ii). (i)

whether the participant is paying close attention to the instructions

(ii)

whether the participant understands the intended meaning relation as specified

From Compatibility to Testability 

 21

For a speaker who disallows “split antecedence” as indicated in (33), Hoji 1990: Chapter 4 argues, the meaning relation holding between Toyota to Nissan ‘Toyota and Nissan’ and soko ‘it’ cannot obtain as an instance of “coreference” between Toyota to Nissan ‘Toyota and Nissan’ and some plural-denoting expression, as soko is not plural denoting; the meaning relation in question must, therefore, be an instance of BVA(Toyota to Nissan, soko), and for such a speaker, we predict judgments as indicated in (35), in line with (31). (34) (=Hoji 1990: Chapter 4: (61a), slightly adapted) [Toyota to Nissan]i-ga sokoi-no zyuugyooin-ni Toyota and Nissan-nom it-gen employee-dat kirokutekina boonasu-o dasita (koto) record-breaking bonus-acc gave ‘[Toyota and Nissan]i gave a record-breaking amount of bonus to sokoi’s employees.’ (35)

(With BVA(Toyota to Nissan, soko)) a. ✶Soko-no robotto-ga Toyota to Nissan-o hihansita. it-gen robot-nom Toyota and Nissan-acc criticized ‘Its robot(s) criticized each of Toyota and Nissan.’ b. Soko-no robotto-o Toyota to Nissan-ga hihansita. it-gen robot-acc Toyota and Nissan-nom criticized ‘Each of Toyota and Nissan criticized its robot(s).’

With the use of soko as Y of BVA(X, Y), judgments indeed obtain more clearly in line with (31) (and its more general version in (32)) than with the use of pro as Y of BVA(X, Y). As I will discuss below, however, the use of soko alone is still insufficient to obtain such judgments in a definitely reliable and reproducible manner. Attempts to obtain clear judgments not only involved expansion of empirical coverage but it also led to theoretical articulation of what formal properties might be responsible for the relevant linguistic judgments and even to conceptual articulation of how rigorous testability can be pursued in dealing with language and the language faculty. For example, we can consider the efforts to expand the empirical coverage of the relevant generalizations by considering a wider range of choices for Y of BVA(X, Y). In particular, this involved the examination of certain ellipsis-related meaning relations that seem to be sensitive to c-command relations, namely the so-called “sloppy identity” reading. Such investigation led to the observation reported in Hoji 1997, 2003b, among other places, that kare ‘he’ can serve as a “sloppy pronoun”. This was somewhat unexpected in light of the observation, made

22 

 Hajime Hoji

in Hoji 1991 and earlier works, that kare could not be Y of BVA(X, Y), but it in fact seems to have anticipated judgments by speakers that kare can be Y of BVA(X, Y), as discussed in Hoji et al. 1999; see Chapter 5 of this volume: note 24.31, 32 Further, the “anti-locality condition” that seems to be imposed on BVA(X, Y) and also the sloppy identity reading was also investigated intensively; this was done in relation to the attempt to obtain clearer judgments in line with (31) (and its general form in (32)), with regard to both “non-elliptical” and “elliptical” sentences.33 This research, particularly on the sloppy identity reading, led to the

31 What is meant by “BVA(X, Y)” here is a c-command-based instance of BVA(X, Y). As will be discussed extensively in subsequent chapters of this volume, especially, Chapters 2 and 5, it is possible for BVA(X, Y) to arise without the c-command relation pertaining to X and Y. The correlational methodology to be proposed in subsequent chapters has been motivated by the recognition that meaning relations such as BVA(X, Y) can arise without the c-command relation in question. 32 What is meant by “sloppy identity” reading is the interpretation where the second conjunct of (i) is understood as in the second conjunct of (ii). (i) (ii)

John washed his car and so did Bill. John washed John’s car and Bill washed Bill’s car

Reinhart (1983: Ch. 7) argues that “sloppy identity” reading is possible only if the requisite c-command condition is satisfied (in the case of (i), John c-commanding his). The assessment of the validity of this claim is rather involved. Among the relevant issues is how to determine what “elliptical” construction in a language in question can effectively serve for the relevant assessment. Considerations in Hoji 1997, 1998, and 2003b, Fukaya and Hoji 1999 and Fukaya 2007, among other places, suggest that “sloppy identity” reading” can arise in two distinct ways, with the distinction being reminiscent of Hankamer and Sag’s (1976) deep and surface anaphora distinction and that only one of the two types of “sloppy identity” reading seems to obey various structural conditions, including the c-command requirement, and the lexical requirement on the “sloppy pronoun”. The lexical requirement on the “sloppy pronoun” is related to the observation that if his in (i) is replaced by John, as in (iii), the interpretation in (ii) is no longer possible for most speakers even when the first conjunct of (iii) is acceptable for them, with the two instances of John understood as referring to the same individual named ‘John’. (iii)

John washed John’s car and so did Bill.

Sloppy-identity-based research did not play a central role in the development of the correlational methodology for LFS to be presented in subsequent chapters because of additional complication/ difficulty with regard to which “elliptical structure” in a given language provides us with a good “testing ground” for c-command detection, how to identify the presence and the absence of the requisite c-command relation in the absence of precedence, among other (more involved) issues, as discussed in Hoji 2003b. 33 What is meant by “anti-locality condition” is related to the observation that BVA(every boy, his) is not possible in (i) but BVA(every boy, him/his) is in (ii), for a vast majority of speakers. (i) (ii)

a. b.

Every boy praised him. Every boy praised his teacher. Every boy thought that John would praise him.

From Compatibility to Testability 

 23

recognition of two distinct sources for BVA(X, Y), one formal and c-commandbased and the other non-formal (and thus not c-command-based); this insight is one of the most crucial realizations that led to the creation of this volume, and will be discussed in Chapters 2, 5 and 6. See Hoji 2016 for a review of many of the works just alluded to. Despite the efforts, and with various choices of X and Y, it had proved to be difficult to obtain clear judgments precisely in line with (31). When I had clear judgments and my colleagues agreed with me, judgments reported in published works were often quite different. For example, when judgments reported in published works are in line with the prediction of the form in (31a), many speakers, including myself, disagree and find the relevant sentences (readily) acceptable.34 It has in fact been widely observed and addressed that there is frequent, if not constant, disagreement among researchers about judgments; most crucially for our concern, this includes cases where judgments on sentences are predicted to be impossible under a particular meaning relation. The empirical bases of the claims made in Hoji 1985 and subsequent works of mine relied mostly on my own judgments, collected not in nearly as systematic a way as one would have hoped, in line with the basic research practice of Hoji 1985 (which I take to be a reflection of the general research practice in the field of syntax). I thus recognized the need to make use of the collected judgments of a large group of speakers, perhaps preferably mostly non-researchers, to substantiate the claims that used to be made based solely (or at least mainly) on my own judgments. Hoji 2015 was an attempt to do that in light of the considerations noted above. As summarized in Hoji 2016: 5.3, Hoji 2015 argues that we can consider the result of our “Main-Experiment” (in the context of our current discussion, experiments checking speaker judgments on sentences that instantiate (32a) and (32b)) to be revealing about the validity of the hypotheses the lead to the

According to Reinhart’s (1983: Ch. 7) proposal that what underlies BVA(X, Y) and “sloppy identity” reading must satisfy not only the c-command condition but also the “anti-locality condition”. Assessment of the validity of the proposal is rather involved and it cannot be provided here. Hoji 1995, along with Hoji 1990, provide initial discussion of the relevant complications; see also the previous footnote, especially its last sentence. 34 Reported judgments in question are often accompanied by a disclaimer that the judgments are “subtle”, and the relevant judgments are given as “??” or “?✶, instead of “✶”, yet taken as providing empirical support for the hypotheses that lead to “impossibility”. (“✶” indicates “unacceptable” in commonly used notation.)

24 

 Hajime Hoji

predictions in question, only if we focus on the speakers for whom, according to their reported judgments in the “Sub-Experiments”, the instructions in the “Main-Experiments” are clear and effective. Further, the relevant “Sub-Hypotheses” (again in the context of the present discussion, hypotheses that a particular choice of Y of BVA(X, Y) is “singular-denoting”) must be deemed valid for the speaker(s) in question, according to some pre-established test(s). Hoji 2015 was thus an attempt to obtain judgments in line with the categorical prediction indicated in (32), repeated here, from non-researcher participants in a large-scale experiment, focusing on the subset of the participants who meet the criteria defined above. (32)

Generalized version of (31): BVA(X, Y) is not possible in (a) but it is in (b). a. [[. . . Y . . . ]N]-ga X-o V b. [[. . . Y . . . ]N]-o X-ga V

Choices of X and choices of Y considered in Hoji 2015 are listed below (although not all of them are actually used in the experiments discussed in Hoji 2015). (36)

(37)

(See Hoji 2015: Ch. 7: (74).) Choices of X of BVA(X, Y): a. subete-no N b. 3-tu-no N d. kanarino kazu-no N e. N-cm35 3-tu f. N-cm 3-tu izyoo g. N-cm sukunakutomo 3-tu izyoo h. NP(-cm)-dake

‘every N’ ‘three N’s’ ‘a good number of Ns’ ‘three Ns’36 ‘three or more Ns’ ‘at least three or more Ns’ ‘only NP’

(See Hoji 2015: Ch. 7: (75).) More choices of X of BVA(X, Y): a. nan % izyoo-no N ‘what % or more Ns’ b. sukunakutomo 3-tu-no N ‘at least three Ns’

35 “cm” standing for “case marker”, such as -ga, -o, etc. 36 There are two versions of each of (36e), (36f), and (36g). One is where the “#-cl” (=a number followed by a classifier (e.g., 3-tu in (36b)) is adjacent to “its host NP” (the N-cm” in (36e), (36f), and (36g)) and the other is where the two were separated, for example, by an “adverbial” phrase. 37 XP-sika requires the presence of Negation.

From Compatibility to Testability 

c. d. e. f. g.

do-no N(-cm)-mo do-no N-cm NP(-cm)-sika37 NP(-cm)-sae 2-wari izyoo-no N

 25

‘whichever N, every N’ ‘which N’ ‘{no one/no place/nothing} but NP’ ‘even NP’ ‘20% or more N(s)’

(38) (See Hoji 2015: Ch. 7: (76).) Choices of Y of BVA(X, Y) a. soko ‘it, the place, that place’ b. soitu ‘the guy, that guy’ c. so-no N ‘the N, that N’ As noted in Chapter 4: Section 6.2 of this volume, the results of the “multiple-nonresearcher-informant experiments” reported in Hoji 2015 came quite close to obtaining experimental results in support of the categorical prediction in question, with the level of experimental success that had not been attained previously (as far as I am aware); see the results reported in Hoji 2015: 259–262 (cf. also http://www.gges. xyz/hojiCUP/index.shtml). For example, if we focus on speakers for whom soko ‘it’ and soitu ‘that guy’ are necessarily “singular-denoting”, based on “split antecedence tests”, with X being each of (39),38 a vast majority of the speakers reject (40b), while accepting (40a).39 (39)

a. b. c. d.

3-tu-no kyuuudan subete-no kyuudai 55% izyoo-no tihoo zititai karari-no kazu-no seizika

‘three ballclubs’ ‘every ballclub’ ‘55% or more local governments’ ‘a considerable number of politicians’

38 For the ease of exposition, let us refer to such speakers as “qualified” speakers, with a clear understanding that what is intended by the term is what is noted just above. 39 Whether not a given speaker clearly reports that asoko ‘it’ and aitu ‘that guy’ cannot be Y of BVA(X, Y) while soko ‘it’ and ‘soitu ‘that guy’ can (in line with the generalization that so-NP can, but a-NP cannot, be Y of BVA(X, Y)) was also considered. The result of one of the “Experiments” testing this (EPSA [10]-5; see http://www.gges.xyz/hojiCUP/index.shtml) is that among the speakers for whom soko and soitu are necessarily “singular-denoting”, only one out of 63 such speakers allowed BVA(X, asoko); two other speakers accepted it but did so only one out of eight times. The remaining speakers in questions never allowed BVA(X, asoko).

26 

 Hajime Hoji

(40) (=Hoji 2015: Chapter 7: (99a, b), slightly adapted) (With BVA(tihoozititai-ga/o 3-tu izyoo, soko) ‘BVA(three or more local governments, it)’) a. Sentence instantiating (31b): (zibun-no tokoro-no syokuin-o hihansita tihoozititai-ga own-gen place-gen employee-acc criticized local:government-nom 3-tu izyoo aru toyuu imi-de) 3-CL more exist that interpretation-with ‘under the interpretation that there are three or more local governments which criticized their own employees’ So-ko-no syokuin-o tihoozititai-ga 3-tu izyoo that-place-gen employee-acc local:government-nom 3-CL more hihansita. criticized ‘Approx: Each of three or more local governments criticized its employees.’ b. Sentence instantiating (31a): (zibun-no tokoro-no syokuin-ni hihansareta own-gen place-gen employee-dat was:criticized tihoozititai-ga 3-tu izyoo aru local:government-nom 3-CL more exist toyuu imi-de) that interpretation-with ‘under the interpretation that there are three or more local governments which were criticized by their own employees’ So-ko-no syokuin-ga tihoozititai-o 3-tu izyoo That-place-gen employee-nom local:government-acc 3-CL more hihansita. criticized ‘Approx: Its employees criticized each of three or more local governments.’ Only 4 out of 23 such speakers (i.e., “qualified” speakers; see footnote 38) ever accepted (40b). If we consider a version of (40b), with a verb that “marks” its object with -ni, instead of -o, none of the 23 speakers in question accepted this modified version of (40b). Similar results obtained with X being one of (41). (41)

a. subete-no tihoo zititai ‘every local government’ b. 3-tu-no tihoo zititai ‘three local governments’ c. tihoozititai-ga/o sukunakutomo 3-tu izyoo ‘at least 3 or more local governments’

From Compatibility to Testability 

 27

There are, however, still clear problems. First, although it came quite close to obtaining experimental results in support of the categorical prediction in question (at least with some combination of X and Y, and with a particular property of the sentences under consideration, such as having to do with the “case marking” on the phrase that contains Y), it did not consistently attain results in line with the definite and categorical predictions. Cases like the one mentioned above, where none of the “qualified” speakers ever accepted sentences of the form in (32a), were exceptional rather than the “normal” cases. The prediction was that no “qualified” speakers would accept sentences instantiating (31a) under the specified BVA(X, Y), but in many cases, a few such speakers did accept them. Second, if I participated in the experiments myself, my judgments would disconfirm the definite predictions in most, if not all, experiments; that is, I would be a “qualified” speaker who nevertheless accepted sentences instantiating (31a) under the specified BVA(X, Y). Even more problematic for this endeavor, I have observed that my own judgments can change; it is usually the case that the more extensively I check the relevant sentences, the more likely it becomes that I can accept the sentences that instantiate (31a) with the relevant BVA(X, Y) readings. This judgmental instability is related to the fact that, although it does help us avoid the “complication” caused by the use of pro as Y of BVA(X, Y) and gives us results closer to our definite predictions, the use of overt nominals for Y of BVA(X, Y) (such as soko ‘it’ and soitu ‘that guy’) does not always result in speaker judgments in line with the prediction given in (31). This is extensively discussed in Ueyama 1998, especially in its Appendix D (see also Chapter 2 of this volume). These nominals seem “singular-denoting” for most speakers, based on “split antecedence” test, so such results suggest the presence of further confounding factors at play, to say nothing of the fact that such factors can apparently vary in their effects over time. In light of the fact that the language faculty is internal to an individual, a serious shortcoming of Hoji 2015 is its failure to focus on self-experiments.40 Subsequent chapters in this volume will present a correlational methodology that makes it possible to account for judgmental changes, alluded to above. As

40 The way the charts for experimental results are presented in Hoji 2015 in fact gives one the (false) impression that Hoji 2015 was concerned with the aggregate of judgments by a group of speakers, rather than judgments by an individual speaker. “False” in the sense that Hoji 2015 does note, explicitly and in more than a few places, that it is concerned with individual speakers, but what is “practiced” in that work seems to suggest something different.

28 

 Hajime Hoji

a result, this correlational methodology in turn makes it possible to deduce definite and categorical prediction about an individual, including myself, and obtain and replicate experimental results precisely in line with such predictions, despite the judgmental changes. This methodology, to be introduced in Chapter 4, allows us to deduce predictions that can survive rigorous attempts at disconfirmation, finally enabling us to successfully pursue true testability-seeking research.

4 Concluding Remarks Consider the predictions in (32), repeated here. (32)

Generalized version of (31): BVA(X, Y) is not possible in (a) but it is in (b). a. [[. . . Y . . . ]N]-ga X-o V b. [[. . . Y . . . ]N]-o X-ga V

As noted above, the prediction regarding (32a) is the most crucial in (32) because it is this prediction that is disconfirmable. The disconfirmation of the prediction of the form in (32a) would therefore be a sufficient basis for rejecting a hypothesis that leads to the prediction (in conjunction with other hypotheses). Works such as Hoji 2006a, 2006b, 2009, 2010a, and 2010b evaluate predictions of the general form in (32a), made under specific “lexical” hypotheses that state that certain choices of Y, due to its hypothesized lexical properties, must be c-commanded by its intended antecedent (X).41 It is interesting to note that many of those works (the first three cited above) give result charts that include standard deviations, in addition to the “average” of responses by the group of participants in question, in addition to information about the number of participants who “accepted” the sentences in question (with

41 See the works cited above for the details; the lexical hypotheses in question are one stating that otagai in Japanese is a local anaphor (analogous to each other in English) and one stating that zibun-zisin in Japanese is a local anaphor (analogous to himself in English). Results of non-self-experiments, reported in those works, indicate that the predictions of the general form in (32a) made under the relevant lexical hypotheses are disconfirmed in the sense that the majority of the participants in the relevant experiments accept the sentences of the form in (32a), clearly indicating that there are speakers who accept the sentences of the form in (32a). What results we might obtain once we apply the correlational methodology to those hypotheses is yet to be checked.

From Compatibility to Testability 

 29

the relevant meaning relation in question), along with the total number of the participants in question; see, for example, Hoji 2009: Chapter 2: Sections 3.2.1 and 3.2.2. The latter two works cited above present result charts in a similar way, but do not mention standard deviations. Result charts in Hoji 2015 do not indicate the “average” of responses by the group of participants (nor the standard deviation). Different ways of presenting result charts in these works, just mentioned, suggest a progressively clearer understanding (on my part) of the nature of experiment in LFS (including non-self-experiments) that it deals with individuals, not averages across a group. Result charts in Hoji 2015 indicate what percentage of judgments reported on a particular sentence pattern (such as (32a) and (32b)) is “Acceptable at least to some extent”. This can give the reader the wrong conception that the work is concerned with responses of the group of speakers, analyzed in some way, rather than responses from an individual, as pointed out in note 40. It is clearly stated in Hoji 2015 that LFS makes definite predictions about an individual’s judgments and tries to obtain experimental results in line with such predictions. Since the prediction about sentences of the form in (32a) is that no “qualified” speakers accept them under the intended meaning relation, the predicted percentage of “acceptable at least to some extent” answers on such sentences should be zero. The “zero” result, however, obtained only very occasionally in Hoji 2015, as briefly mentioned above. As will be discussed in some depth in subsequent chapters, this is not accidental and is actually for a principled reason; this reason is what leads to the correlational methodology for LFS to be proposed. We could not, for this same principled reason, expect to obtain experimental results precisely in line with our categorical predictions in self-experiments, and hence in non-self-experiments, if we focused only on judgments on the availability of a particular meaning relation in isolation. The approach pursued in this volume in some ways returns to the one taken in Hoji 1985 with regard to its focus on self-experiments, but with its additional focus on the correlational methodology just alluded to, which makes it possible for our research to be testability-seeking. Articulation of the methodology for non-self-experiment should be based on the methodology for self-experiments, and this is indeed the case in this volume, as will be addressed in some depth in the context of “replication” in Chapters 5 and 6. Reliance on self-experiments raises an interesting challenge regarding how to replicate results of a given self-experiment with other speakers and in fact with any speaker of any language, the topic that will be addressed in the bulk of this volume.

30 

 Hajime Hoji

References Most of the works by Hoji listed below are available at ResearchGate. https:// www.researchgate.net/profile/Hajime-Hoji. Chomsky, Noam. 1995. The minimalist program. Cambridge, MA: MIT Press. Chomsky, Noam. 2017. The Galilean challenge. Inference: International review of science 3(1). https://inference-review.com/article/the-galilean-challenge (accessed 14 February 2022) Fukaya, Teruhiko and Hajime Hoji. 1999. Stripping and sluicing in Japanese and some implications. In Sonya Bird, Andrew Carnie, Jason D. Haugen and Peter Norquest (eds.), Proceedings of WCCFL 18, 145–158. Somerville: Cascadilla Press. Fukaya, Teruhiko. 2007. Sluicing and stripping in Japanese and some implications. Los Angeles, CA: University of Southern California dissertation. Feynman, Richard. 1985. Surely you’re joking, Mr. Feynman!: Adventures of a curious character. New York, NY: W.W. Norton & Company. Hankamer, Jorge and Ivan Sag. 1976. Deep and surface anaphora. Linguistic Inquiry 7. 391–428. Hayashishita, J.-R. 2004. Syntactic and non-syntactic scope. Los Angeles, CA: University of Southern California dissertation. Hayashishita, J.-R. 2013. On the nature of inverse scope readings. Gengo Kenkyu 143. 29–68. Hoji, Hajime. 1985. Logical form constraints and configurational structures in Japanese. Seattle, WA: University of Washington dissertation. Hoji, Hajime. 1987. Empty pronominals in Japanese and subject of NP. In Joyce McDonough and Bernadette Plunkett (eds.), NELS 17, 290–310. Amherst, MA: University of Massachusetts, Amherst, GLSA Publications. Hoji, Hajime. 1990. Theories of anaphora and aspects of Japanese syntax. Unpublished ms., University of Southern California. Hoji, Hajime. 1991. KARE. In Carol Georgopoulos and Roberta Ishihara (eds.), Interdisciplinary approaches to language: Essays in honor of S.-Y. Kuroda, 287–304. Dordrecht: Kluwer Academic Publishers. Hoji, Hajime. 1995. Demonstrative binding and principle B. In Jill N. Beckman (ed.), NELS 25, 255–271. Amherst, MA: University of Massachusetts, Amherst, GLSA Publications. Reprinted in: Hoji 2013. Hoji, Hajime. 1997. Sloppy identity and formal dependency. In Brian Agbayani and Sze-Wing Tang (eds.), Proceedings of the 15th West Coast Conference on Formal Linguistics, 209–223. Stanford, CA: CSLI Publications. Hoji, Hajime. 1998. Null object and sloppy identity in Japanese. Linguistic Inquiry 29. 127–152. Reprinted in: Hoji 2013. Hoji, Hajime. 2003a. Falsifiability and repeatability in generative grammar: A case study of anaphora and scope dependency in Japanese. Lingua 113. 377–446. Reprinted in: Hoji 2013. Hoji, Hajime. 2003b. Surface and deep anaphora, sloppy identity, and experiments in syntax. In Andrew Barss (ed.), Anaphora: A reference guide, 172–236. Cambridge: Blackwell. Reprinted in: Hoji 2013.

From Compatibility to Testability 

 31

Hoji, Hajime. 2006a. Assessing competing analyses: Two hypotheses about “scrambling” in Japanese. In Ayumi Ueyama (ed.), Theoretical and empirical studies of reference and anaphora: Toward the establishment of generative grammar as an empirical science (A report of the Japan Society for the Promotion of Science Grant-in-Aid for Scientific Research (B), Project No. 15320052), 139–185. Fukuoka: Kyushu University. Hoji, Hajime. 2006b. Otagai. In Ayumi Ueyama (ed.), Theoretical and empirical studies of reference and anaphora: Toward the establishment of generative grammar as an empirical science (A report of the Japan Society for the Promotion of Science Grant-in-Aid for Scientific Research (B), Project No. 15320052), 126–138. Fukuoka: Kyushu University. Hoji, Hajime. 2009. A foundation of generative grammar as an empirical science. Unpublished ms., University of Southern California. Hoji, Hajime. 2010a. Hypothesis testing in generative grammar: Evaluations of predicted schematic asymmetries. Journal of Japanese Linguistics 16: Special issue: In Memory of S.-Y. Kuroda, 25–52. Hoji, Hajime. 2010b. Evaluating the lexical hypothesis about otagai. Linguistic Research 27(1). 65–119. Hoji, Hajime. 2013. Gengo kagaku-o mezashite [Towards linguistic science]: Issues on anaphora in Japanese. (Ayumi Ueyama and Yukinori Takubo (eds.).) Shiga: Ohsumi Publisher. Hoji, Hajime. 2015. Language faculty science. Cambridge: Cambridge University Press. Hoji, Hajime. 2016. Towards language faculty science: Remarks on the papers collected in Hoji 2013. Preface to the e-edition of Hoji 2013. Hoji, Hajime, Satoshi Kinsui, Yukinori Takubo and Ayumi Ueyama. 1999. Demonstratives, bound variables, and reconstruction effects. Proceedings of the Nanzan GLOW: The Second GLOW Meeting in Asia, September 19–22. 141–158. Kayne, Richard S. 1981. Unambiguous paths. In Robert May and Jan Koster (eds.), Levels of syntactic representation, 143–183. Dordrecht: Foris Publications. Kayne, Richard S. 1984. Connectedness and binary branching (Studies in generative grammar). Dordrecht: Foris Publications. Lakatos, Imre. 1978. The methodology of scientific research programmes (Philosophical papers Volume 1). (John Worrall and Gregory Currie (eds.).) Cambridge: Cambridge University Press. Nishigauchi, Taisuke. 1986. Quantification in syntax. Amherst, MA: University of Massachusetts, Amherst dissertation. Plesniak, Daniel. 2022. Towards a correlational law of language: Three factors constraining judgment variation. Los Angeles, CA: University of Southern California dissertation. Reinhart, Tanya. 1983. Anaphora and semantic interpretation. Chicago: University of Chicago Press. Ueyama, Ayumi. 1998. Two types of dependency. Los Angeles, CA: University of Southern California dissertation. Yoshimura, Noriko. 1989. Comments on Stowell’s paper. Paper presented at UCSD Workshop on Japanese and Logical Form, San Diego. Yoshimura, Noriko. 1992. Scrambling and anaphora in Japanese. Los Angeles, CA: University of Southern California dissertation.

Ayumi Ueyama

On Non-individual-denoting So-words 1 Bound Variable Anaphora and Dependent Terms Since Reinhart 1983ab, among others, it has been assumed in many works that a syntactic bound variable anaphora interpretation can be established only if the argument position of the quantifier c-commands the dependent term. In other words, when the former does not c-command the latter, the bound variable anaphora interpretation is not expected to obtain. In the case of Japanese, since a nominative (i.e., ga-marked) NP is supposed to c-command an accusative (i.e., o-marked) or a dative (i.e., ni-marked) NP, this assumption predicts that a bound variable anaphora is possible when the quantifier is nominative and the dependent term is included in an accusative (or dative) NP. (1) SO-type construction: Toyota-sae-ga [so-ko-o tekitaisisiteiru kaisya]-o uttaeta Toyota-even-nom that-place-acc be:hostile company-acc sued ‘Even Toyota sued the company which is hostile to it.’ (Ueyama 1998: ch.2 (36)) It is also expected that the bound variable anaphora interpretation does not obtain when the quantifier is an accusative or dative and the dependent term is included in the nominative NP. Quite unexpectedly, however, there are cases when the bound variable anaphora reading appears to obtain even when the quantifier does not c-command the dependent term. In this chapter, I call this phenomenon as quirky binding, and address when it is possible and when it is not, mainly based on the description given in Ueyama 1997, 1998. Before going into the discussion, I briefly introduce what kind of expressions can play the role of a dependent term in Japanese.

https://doi.org/10.1515/9783110724790-002

34 

 Ayumi Ueyama

2 Demonstratives in Japanese 2.1 Three Demonstrative Prefixes Since most of the anaphoric relations in Japanese involve a demonstrative NP, I introduce the descriptive properties of demonstrative NPs in Japanese in this section.1 A demonstrative NP in Japanese is formed with one of the demonstrative prefixes (listed in (2)) followed by (i) a bound morpheme of a certain type or (ii) -no (genitive) and an NP.2 Some of the bound morphemes which can follow them are given in (3).3 (2) a. kob. soc. a(3) a. -ko ‘place/institution’4 b. -re ‘thing’ c. -itu ‘guy’ The4terms such as deictic use and anaphoric use are often referred to in describing the function/meaning of demonstrative NPs.5 A demonstrative NP is said to be used deictically use when the target individual is visible to the speaker and the addressee, sometimes accompanied by ostension/pointing. The deictic use is 1 ‘He’ in English is also often translated as kare in Japanese. But as often pointed out, kare does not share many properties with ‘he’ in English. One of the properties of kare that is crucial to the present discussion is that kare cannot be construed as a bound variable at least as easily as ‘he’ in English or so-ko ‘that-place/it’ and so-re ‘that-thing/it’ in Japanese. (Cf. Saito and Hoji 1983, Hoji 1990, 1991 among others. In addition, Nakai 1976: 34a is referred to in Kitagawa 1981: 71.) It is also pointed out in Hoji 1991 and Takubo 1996 that it is not totally impossible for kare to be used as a bound variable. I will not provide further discussion on kare in this work. More discussion is found in Hoji’s Chapter 5 in this volume, and Takubo 1996. 2 Do- ‘which’ is often listed together with these demonstrative prefixes in reference grammar books, since they share the morphological distribution. 3 The demonstrative prefixes also appear in non-nominal words such as so-o ‘in that way/so’, ko-nna ‘like this’, and so on. I do not discuss these cases in this work, since I focus on the anaphoric relation with respect to an individual. 4 When this morpheme follows the demonstrative a-, the resulting word becomes a-soko instead of ✶a-ko. According to Satoshi Kinsui (p.c., summer 1997), a-soko comes from the form a-si-ko. While the etymology of this si is not clear, it seems unrelated to the demonstrative so- in so-ko, so-re, or so-itu. 5 See Kuno 1973: ch.24, for example.

On Non-individual-denoting So-words 

 35

not much of a concern here, but generally speaking, when used deictically, ko- is used for something close to the speaker, so- for something close to the addressee, and a- for something away from both the speaker and the addressee. A demonstrative NP is said to be in an anaphoric use when the target individual is not visible to the speaker but is mentioned in the discourse. According to these characterizations of deictic and anaphoric use, however, a demonstrative NP would be neither deictic nor anaphoric if (i) the target individual is not visible to the speaker and (ii) it is not mentioned in the discourse. Instead of using the term ‘anaphoric use’, therefore, I will use the term non-deictic use in the following discussion to refer to a case in which the target individual is not visible to the speaker. As exemplified in (4), all of the three demonstrative forms have a non-deictic use, since the sentence in (4) can be felicitously uttered, with each of the demonstrative forms, in the absence of the person under discussion. (4) So-no toki kyuuni {a-itu / so-itu / ko-itu}-ga sakenda that-gen time suddenly {that-guy/that-guy/this-guy}-nom screamed n da. comp copula ‘And then suddenly he screamed.’ The following subsections illustrate the properties of the non-deictic use of a-words and so-words along the lines of Takubo and Kinsui 1997 and Kuroda 1979.6 It will be shown that a-words are always used to refer to an individual independently of other linguistic expression, while so-words must have some linguistic antecedent.

2.2 A-words must be “Referential” As shown in (5)–(7), a-words can be used non-deictically, even at the very beginning of a discourse.7

6 I put aside ko-words in this work, which roughly correspond to ‘this NP’ in English. Discussion on non-deictic use(s) of ko-words is found in Kinsui and Takubo 1990, Kinsui 1998 and the references therein. 7 As noted just above, the description in this subsection is largely based on Takubo and Kinsui 1997 and Kuroda 1979.

36 

 Ayumi Ueyama

(5)

(Situation: The detective is looking for a man. He somehow believes that the man should be hiding in a certain room. He breaks into the room and asks the people there.) [A-itu]-wa do-ko-da? that-guy-top which-place-copula ‘Where is [he]?’

(6)

(Situation: The speaker (Mr. A) gave some cookies to the addressee (Mr. B). Mr. A really wants to know how Mr. B finds it and asks him at the very beginning of their conversation.) [A-re] tabeta? that-thing ate ‘Have you tried [that]?’

(7)

(Situation: The speaker tries to recall who he met the day before. He calls his secretary and asks:) [Kinoo kita a-no gakusei], namae nan datta? yesterday came that-gen student name what was ‘What is the name of [that student who came yesterday]?’

It appears that a-words can also refer to an individual that has been mentioned in the preceding discourse. (8)

A: Kinoo Yamada-ni atta yo. yesterday Yamada-dat met particle ‘I met Yamada yesterday.’ B: Soo. A-itu genki datta? yes that-guy fine was ‘Really? Was he fine?’

(9)

A: Kondo banana keeki yaiteageru ne. next:time banana cake bake:for:you particle ‘I will bake some banana cake for you next time.’ B: A-re daikoobutuna n da that-thing favorite comp copula ‘That is my favorite.’

On Non-individual-denoting So-words 

 37

(10) A: Kinoo-mo gakusei-san-ga mattemasita yo. yesterday-also student-Mr-nom waiting particle ‘The student was waiting for you yesterday again.’ B: A-no gakusei kyoo-mo kuru to omou? that-gen student today-also come comp think ‘Do you think he will come again today?’ However, even if the target individual has been introduced in the discourse, a-words cannot be used if the speaker does not know him/her/it through direct experience. (11)

(Situation: A wife told her husband on the phone that someone had called him. He has no idea who the person is. He asks her:) #A-itu-wa nante itteta? that-guy-top what said ‘What did he say?

(12) (Situation: Mary told John about the movie she just saw. John did not know about it, but he got interested.) #A-re-wa omosiro-sooda ne. that-thing-top interesting-sound particle ‘That sounds interesting.’ (13) (Situation: The secretary told the professor that a student waited for him for an hour at the door the day before. The professor has no idea who the student is, but he feels sorry for him and tells the secretary:) #Kinoo kita a-no gakusei-ga mooitido ki-tara, Yesterday came that-gen student-nom again come-if sugu osietekure. soon tell:me ‘Please tell me immediately if that student who came yesterday comes again.’ The observation in (11)–(13) indicates that a-words cannot be really “anaphoric” to another linguistic expression in the discourse: namely, although a-words appear to have an anaphoric relation with another linguistic expression in (8)–(10), they are in fact “referential” themselves, just as in the cases in (5)–(7). As has been pointed out in Hoji 1991 among others, an a-word can never be a bound variable. For example, (14a) can never be interpreted as in (14b), unlike (14c):

38 

 Ayumi Ueyama

(14) a.

do-no kaisya-mo a-soko-no bengosi-o uttaeta which-gen company-also that-place-gen attorney-acc sued ‘every company sued its attorney’ ✶

b. ∀x[ x sued x’s attorney ] c. do-no kaisya-mo so-ko-no bengosi-o which-gen company-also that-place-gen attorney-acc ‘every company sued its attorney’

uttaeta sued

This fact naturally follows from the generalization that an a-word is always “referential” just as are names.

2.3 Non-deictic So-words Cannot be “Referential” Unlike a-words, non-deictic so-words cannot be used at the very beginning of a discourse, even if the speaker knows the target individual well. Compare (15)–(17) with (5)–(7) above, respectively, where the identical situations are postulated. (15) (Situation: The detective is looking for a man. He somehow believes that the man should be hiding in a certain room. He breaks into the room and asks the people there.) #[So-itu]-wa do-ko-da? that-guy-top which-place-copula ‘Where is [he]?’ (16) (Situation: The speaker (Mr. A) gave some cookies to the addressee (Mr. B). Mr. A really wants to know how Mr. B finds it and asks him at the very beginning of their conversation.) #[So-re] tabeta? that-thing ate ‘Have you tried [that]?’ (17)

(Situation: The speaker tries to recall who he met the day before. He calls his secretary and asks:) #[Kinoo kita so-no gakusei], namae nan datta? yesterday came that-gen student name what was ‘What is the name of [that student who came yesterday]?’

On Non-individual-denoting So-words 

 39

These examples are infelicitous whether or not the addressee understands who/ what the target individual is. On the other hand, so-words can “anaphorically” refer to an individual introduced in the discourse, whether or not the speaker knows the target individual. The examples in (18)–(20) are to be compared with those in (11)–(13), respectively. (18)

(Situation: A wife told her husband on the phone that someone had called him. He has no idea who the person is. He asks her:) So-itu-wa nante itteta? that-guy-top what said ‘What did he say?

(19)

(Situation: Mary told John about the movie she just saw. John did not know about it, but he got interested.) So-re-wa omosiro-sooda ne. that-thing-top interesting-sound particle ‘That sounds interesting.’

(20) (Situation: The secretary told the professor that a student waited for him for an hour at the door the day before. The professor has no idea who the student is, but he feels sorry for him and tells the secretary:) Kinoo kita so-no gakusei-ga mooitido ki-tara, Yesterday came that-gen student-nom again come-if sugu osietekure. soon tell:me ‘Please tell me immediately if that student who came yesterday comes again.’ The observation given in this subsection indicates that non-deictic so-words must have some linguistic antecedent: namely, they are always dependent on some other linguistic expression.8

8 As Hajime Hoji (p.c., spring 1997) points out to me, there can be some cases in which it is not easy to determine whether the so-word is deictic or non-deictic. Consider the example in (i), by which he points out that the so-word can refer to something which is not visible to the speaker or the hearer, without the “antecedent” in the preceding discourse. (i)

A, [so-re] watasi oh that-thing I ‘Oh, I have eaten [it].’

tabetyatta. have:eaten

40 

 Ayumi Ueyama

As is expected from this property, a so-word can be a dependent term in a bound reading. (14) a.

do-no kaisya-mo a-soko-no bengosi-o uttaeta which-gen company-also that-place-gen attorney-acc sued ‘every company sued its attorney’ ✶

b. ∀x[ x sued x’s attorney ] c.

do-no kaisya-mo so-ko-no bengosi-o uttaeta which-gen company-also that-place-gen attorney-acc sued ‘every company sued its attorney’

One may suspect that so-ko in (14c) might simply stand in a coreferential relation with the companies under discussion considered as a “group entity”. However, as repeatedly shown in a series of works by Hoji (cf. Hoji 1995, 1998, 2003, 2015), so-ko is strictly singular-denoting and hence cannot be coreferential with a plural-denoting NP. This generalization is demonstrated by the fact that so-ko cannot have ‘split antecedents’ unlike plural-denoting elements such as karera ‘they’. (21) a. Tom1-ga Nick2-ni [CP CIA-ga karera1+2-o sirabeteiru Tom-nom Nick-dat CIA-nom them-acc is:investigating to] tugeta. comp told ‘Tom1 told Nick2 [that the CIA was investigating them1+2].’

I agree with him that (i) is more or less acceptable under the situation given in (ii): (ii)

Situation: John has hurried back home, desiring to eat a piece of cake he has kept for this moment. But when he opens the refrigerator door, he finds nothing there. He is shocked and cannot say a word, staring at the place it used to be. Looking at him, Mary says:

I maintain that a non-deictic so-word cannot be “referential” and that such examples as in (i) (under the situation in (ii)) should be regarded as an extended case of deictic use, although the target individual is not visible anymore. For me, it is crucial that John is staring at the place it used to be, in a sense “pointing” at something which does not exist anymore. Hence, (i) seems to me to become quite unacceptable under the scenario described in (iii), which minimally differs from (ii) but involves no “pointing” at the target individual: (iii) Situation: John has hurried back home, desiring to eat a piece of cake he has kept for this moment. But when he opens the refrigerator door, he finds nothing there. He is shocked and cannot say a word, staring at Mary, who has been at home all day. Looking at him, Mary says:

On Non-individual-denoting So-words 

b.

 41

Toyota1-ga Nissan2-ni [CP CIA-ga soko1+2-o Toyota-nom Nissan-dat CIA-nom it-acc sirabeteiru to] tugeta. is:investigating comp told ‘Toyota1 told Nissan2 [that the CIA was investigating it1+2].’



3 Non-individual-denoting So-words If a so-word always requires a linguistic antecedent and the relevant syntactic relation is contingent upon the c-command condition, it is expected that an anaphoric relation should not be available if the antecedent does not c-command the so-word at LF. However, one might come up with examples which appear to go against this generalization, such as (22) and (23). (22) a.

b.

(?)So-ko-no bengosi-ga Toyota-o uttaeta. that-place-gen attorney-nom Toyota-acc sued ‘Its attorney sued Toyota.’ (?)So-ko syussin-no hito-ga so-no daigaku-no that-place source-gen person-nom that-gen university-gen naizyoo-o bakurosita. inside:story-acc exposed ‘Its {graduate/former employee} exposed the inside story of that university.’

(23) [So-ko-no ko-gaisya-to torihiki-o siteiru kaisya]-ga that-place-gen child-company-with business-acc do company-nom Toyota-o uttaeta. Toyota-acc sued ‘[A company which is doing business with its subsidiary] sued Toyota.’ I maintain that these are not really counterexamples to the claim: the relevant relation should be c-command if the antecedent relationship were established syntactically, but the relation between so-ko-no bengosi and Toyota that we understand in (22a), for example, does not seem to me to be yielded in terms of a syntactic relation between so-ko and Toyota. In other words, I claim that the sentences in (22) and (23) are more or less comparable with those in (24), rather than those in (25).

42 

 Ayumi Ueyama

(24) a. A/The (retained) attorney sued Toyota. b. A {graduate/former employee} exposed the inside story of that university. c. [A company which is doing business with a subsidiary] sued Toyota. (25) a. Its attorney sued Toyota. b. Its {graduate/former employee} exposed the inside story of that university. c. [A company which is doing business with its subsidiary] sued Toyota. When the expression so-ko-no bengosi is interpreted as something like ‘a retained attorney’ (i.e., when so-ko-no is interpreted as a modifier meaning ‘attached’, roughly speaking) as in (22a), the affiliation of the attorney is simply left vague as far as the linguistic expression is concerned, and hence, so-ko does not require an “antecedent” because there is no need to specify the value of so-ko in this case. Similarly, so-ko syussin-no hito (Lit. ‘people from that place’) in (22b) can mean either ‘local people’, ‘a former employee’ or ‘a graduate’, depending on the context. Exactly because these so-words need not be individual-denoting, they can be used in a generic statement, as in (26). (26) a. ?[So-ko-no bengosi] toyuu mono-wa sin’yoosi-nai hoo-ga that-place-gen attorney comp thing-top trust-not way-nom ii. good ‘It is better not to trust (the word of) [a retained attorney] in general.’ b. ?[So-ko syussin-no hito] toyuu mono-wa koohyoo that-place source-gen person comp thing-top disclose deki-nai hanasi-o sitteiru mono da. can-not story-acc know thing copula ‘It is generally the case that [people who once were inside] know some story which cannot be made public.’ Compare (26) with unacceptable sentences in (27), in which an obligatorily individual-denoting expression (i.e., ‘its’) is used in a generic statement. (27) a. b.

It is better not to trust (the word of) [its attorney] in general.’ It is generally the case that [its former employees] know some story which cannot be made public.’ ✶ ✶

On Non-individual-denoting So-words 

 43

The sentences in (26) may sound a little awkward as generic statements, but the examples in (28) show that a so-word can be clearly non-individual-denoting in some cases. (28)

a.

Watasi-no dezain-no mottoo-wa [so-no hito I-gen design-gen motto-top that-gen person rasi-sa]-o ensyutusuru koto desu. like-hood-acc direct fact copula ‘The motto of my design is to amplify [the personality/characteristics (literally: being like that person)].’

b. Kyoo-wa [so-no miti-no senmonka] to yobareru today-top that-gen way-gen specialist comp be:called kata zyuunin-ni atumatte itadakimasita. person ten-dat gather we:asked ‘Today, we have invited ten people who are regarded as [a specialist of the subject/field].’ c. Saikin-no aidoru-kasyu-wa [so-no hen-ni iru] recent-gen idol-singer-top that-gen area-at exist onnanoko-to kawaranai. girl-with indistinguishable ‘Pop-stars these days are not different from girls [in the neighborhood].’ (Takubo and Kinsui 1997: (46)) d. John-wa itumo [so-no ba kagiri]-no iiwake-o suru. John-top always that-gen situation only-gen excuse-acc do ‘John always makes [glib (Lit. only for that situation)] excuses.’ e.

Hito-o yatou karaniwa [so-re soooo]-no person-acc hire in:case that-thing suitable-gen kyuuryoo-o youisi-nakerebanaranai. salary-acc prepare-must ‘If you are going to hire a person, you must prepare for [reasonable (literally: suitable for it)] salary.’

Therefore I assume that in principle a so-word can be either individual-denoting or non-individual-denoting to a varying degree.9

9 I hold that the fact that a so-word can be non-individual-denoting should be ultimately attributed to some inherent property of so-, as has been (explicitly or implicitly) suggested in various

44 

 Ayumi Ueyama

So-ko-no ko-gaisya in (23) may require another account, since in this case so-ko-no barely contributes to the meaning of the whole phrase because every subsidiary is in nature ‘attached’ to its parent company. I consider that (23) is not distinct from (29) in its LF representation, roughly speaking. (29)

[Ko-gaisya-to torihiki-o siteiru kaisya]-ga child-company-with business-acc do company-nom Toyota-o uttaeta. Toyota-acc sued ‘[A company which is doing business with a/the subsidiary] sued Toyota.’

In sum, I maintain that the so-words in (22) and (23) do not enter into a (syntactic) anaphoric relation because they are not individual-denoting: that is to say that a linguistic antecedent is necessary for a non-deictic individual-denoting so-word, but not for a non-individual-denoting so-word. I consider that an expression such as so-ko is in principle ambiguous between individual-denoting and non-individualdenoting, presumably unlike an expression such as it or he that are necessarily individual-denoting. The possible ambiguity of so-words between individual-denoting and non-individual-denoting would bring about a serious problem in regard to the verifiability of the theory of anaphoric relations proposed in this work unless we can distinguish the cases of non-individual-denoting so-words from those of individual-denoting so-words on independent grounds. While more investigation is certainly necessary to fully understand under what conditions a so-word can (or must) be non-individual-denoting in general, Ueyama 1997 points out several ways to force a BVA reading to be yielded based on an anaphoric relation (rather than in terms of a non-individual-denoting so-word). In fact, the relevant examples in the foregoing discussion have been constructed with as much care as possible according to the results obtained in Ueyama 1997. In what follows, I will go over the relevant factors, basically drawing from Ueyama 1997.10

works on Japanese. For example, Kuroda 1979 states that a so-word is to be connected to “conceptual knowledge” (while an a-word is to be connected to “knowledge obtained through direct experience”). Under this characterization, one can say, as Takubo and Kinsui 1997 and Kinsui 1998 do, that the non-individual-denoting usage of a so-word is a realization of its “concept-denoting” property, assuming that ‘concept’ covers both individuals and non-individual concepts. Readers are referred to Kinsui and Takubo 1992, which is an anthology of the relevant works, including ones by traditional grammarians. 10 Ueyama 1997 claims that the type of so-word in question enters into a syntactic relation with some empty category, but I do not make this assumption any more.

On Non-individual-denoting So-words 

 45

4 Quirky Binding Ueyama 1997 discusses some murky cases where a so-word appears to be involved in a bound variable anaphora without being c-commanded by an apparent quantifier (henceforth QP) and calls such cases quirky binding. (30) provides some examples.11 (30) a. ?So-ko-no bengosi-ga Toyota to Nissan-o suisensita that-place-gen attorney-nom Toyota and Nissan-acc recommended (node, ato-wa dareka-ni Mazda-o because rest-top someone-dat Mazda-acc suisensite-moraw-eba ii dake da). recommend-ask-if good only copula ‘(Since) {its/a retained} attorney recommended Toyota and Nissan (, now we have only to ask someone to recommend Mazda).’ b. ?So-ko-no bengosi-ga subete-no zidoosya-gaisya-o that-place-gen attorney-nom every-gen automobile-company-acc uttaeteiru (node, zidoosya-gyookai-wa daikonran-ni sued because automobile-industry-top disorder-dat otiitteiru). be:thrown:into ‘(Since) {its/a retained} attorney has sued every automobile company (, the automobile industry has been thrown into a state of disorder).’ Informally speaking, quirky binding is possible only if the apparent QP can be understood as some kind of a ‘topic’ and the so-word is non-individual-denoting:12 a syntactic relation is not established precisely because the so-word in this case is not a dependent term. The conditions of the availability of quirky binding can be described as follows: (31) a. The apparent QP must “refer” to a specific group of individuals. b. Quirky binding is not observed in certain constructions.

11 As noted in Ueyama 1997, the availability of quirky binding tends to vary a great deal among speakers: some speakers tend to accept it rather easily, while others seem to firmly reject it. 12 Chierchia 1992: Section 3.5 mentions that some kind of ‘topicalization’ should be involved in the asymmetrical reading of an adverb of quantification in conditional sentences. I suppose that this phenomenon shares some properties with quirky binding, but I have to leave the concrete analysis open for the future research. Hayashishita (2004) is also relevant to this issue.

46 

 Ayumi Ueyama

c. There should not be other quirky binding in the relevant clause. d. The apparent QP must be in a position which is salient enough to be a “topic” of a sentence. e. The so-word must be non-individual-denoting. Among these conditions, (31a,b,c) can be shown relatively explicitly, but (31d,e) are subjective in nature and less clear at least at this stage. I make some remarks for each condition in (31) in the following subsections.

4.1 The Apparent QP must be ‘Referential’ First, in order for quirky binding to be available, the apparent QP must be used to refer to a specific group of individuals. For example, Toyota to Nissan ‘Toyota and Nissan’ can refer to a specific group consisting of Toyota and Nissan. One may suspect that so-ko in quirky binding, such as in (30a), simply stands in a coreferential relation with Toyota to Nissan. However, recall from (21) in Section 2.3 that so-ko cannot be plural-denoting and hence cannot be coreferential with a plural-denoting NP, such as Toyota to Nissan ‘Toyota and Nissan’. As shown in (30b), subete-no gakusei ‘every student’ also induces quirky binding relatively easily: it should be understood to refer to the whole group of students in this case. In contrast to such ‘possibly referential’ QPs, other QPs, such as A ka B ka ‘either A or B,’ 55%-no NP ‘55% of NP,’ do-no gakusei ‘which student’ or John-sae ‘even John’, cannot induce quirky binding. (32) a. ?✶So-ko-no bengosi-ga Toyota ka Nissan ka-o that-place-gen attorney-nom Toyota or Nissan or-acc suisensita (node, ato-wa dareka-ni Mazda-o recommended because rest-top someone-dat Mazda-acc suisensite-mora(w)-eba ii dake da). recommend-ask-if good only copula ‘(Since) {its/a retained} attorney recommended either Toyota or Nissan (, now we have only to ask someone to recommend Mazda).’

On Non-individual-denoting So-words 

 47

bengosi-ga 55%-no zidoosya-gaisya-o b. ?✶So-ko-no that-place-gen attorney-nom 55%-gen automobile-company-acc uttaeteiru (node, zidoosya-gyookai-wa daikonran-ni sued because automobile-industry-top disorder-dat otiitteiru). be:thrown:into ‘(Since) {its/a retained} attorney has sued 55% of the automobile companies (, the automobile industry has been thrown into a state of disorder).’

4.2 Certain Constructions Reject Quirky Binding Second, there are some constructions which prevent quirky binding from occurring. One such construction which I am going to introduce in this subsection is called ‘Deep OS-type’ and is discussed in detail in Ueyama (1998): it has a word order in which either an accusative or a dative NP precedes a nominative NP and the preceding NP is a quantifier (55%-no gakusei in the case of (33)) binding another dependent term appearing in the same sentence (so-itu in (33)). In this construction, quirky binding is not available: i.e., the so-word so-ko is not c-commanded by USC to UCLA ‘USC and UCLA’, and an anaphoric relation between them does not seem to be available at all. (33)

Deep OS-type: ?✶[55%-no gakusei]-ni [so-itu-o sitteiru so-ko-no 55%-gen student-dat that-guy-acc know that-place-gen sensei]-ga USC to UCLA-o suisen-saseta. professor-nom USC and UCLA-acc recommend-made Intended (but unavailable) meaning: ‘[A (USC) professor and a (UCLA) professor who know him] made [55% of the students] recommend (each of) USC and UCLA.’

In contrast, in what Ueyama 1998 calls an ‘SO-type construction’, a nominative NP precedes an accusative and a dative NPs as in (34), and this time, a so-word so-ko allows quirky binding with USC to UCLA ‘USC and UCLA’, in spite of the fact that the former is not c-commanded by the latter.

48 

 Ayumi Ueyama

(34) SO-type construction: a. ?[Chomsky-o sitteiru so-ko-no sensei]-ga USC to Chomsky-acc know that-place-gen professor-nom USC and UCLA-o kare-ni suisen-saseta. UCLA-acc he-dat recommend-made Intended meaning: ‘[A (USC) professor and a (UCLA) professor who know Chomsky] made him recommend (each of) USC and UCLA.’ b. ?55%-no gakusei-ga [[so-itu-ga sitteiru so-ko-no 55%-gen student-nom that-guy-nom know that-place-gen sensei]-ga USC to UCLA-o uttaeta to] omotteiru. professor-nom USC and UCLA-acc sued comp think Intended meaning: ‘55% of the students thinks [that [a (USC) professor and a (UCLA) professor who he knows] sued (each of) USC and UCLA].’ Ueyama (1998) cites a few more constructions which reject quirky binding, but I cannot introduce them here, since it requires some intricate explanation regarding the syntactic structure of sentences in which a nominative NP is preceded by another argument NP.

4.3 No Other Quirky Binding Furthermore, even if the construction is not of the Deep OS-type, two pairs of quirky binding cannot obtain at the same time. Thus, in the schematic representations in (35), the quirky binding between QP1 and NP1 is not available if QP2 and NP2 are also related in terms of quirky binding. (35)

a. [ . . . NP2 . . . NP1 . . .]-nom QP2-dat QP1-acc . . . V b. [ . . . NP2 . . . NP1 . . .]-nom QP1-acc QP2-dat . . . V

(35) is exemplified in (36), which are to be compared with (37) that demonstrates that one pair of quirky binding is available:

On Non-individual-denoting So-words 

(36)

 49

so-ko-no syokuin-ga a. ?✶[So-no hito]-tantoo-no that-gen person-in:charge-gen that-place-gen staff-nom [subete-no giin]-ni USC to UCLA-o suisen-saseta. all-gen senator-dat USC and UCLA-acc recommend-made Intended (but unavailable) meaning: ‘[A (USC) staff member and a (UCLA) staff member who are in charge of that man] made all the senators recommend (each of) USC and UCLA.’ b. ?✶[So-no hito]-tantoo-no so-ko-no syokuin-ga that-gen person-in:charge-gen that-place-gen staff-nom USC to UCLA-o [subete-no giin]-ni suisen-saseta. USC and UCLA-acc all-gen senator-dat recommend-made Intended (but unavailable) meaning: ‘[A (USC) staff and a (UCLA) staff who are in charge of that man] made all the senators recommend (each of) USC and UCLA.’

(37)

a. ?[So-no hito]-tantoo-no syokuin-ga [subete-no that-gen person-in:charge-gen lstaff-nom all-gen giin]-ni USC-o suisen-saseta. senator-dat USC-acc recommend-made Intended meaning: ‘[A staff who is in charge of that man] made all the senators recommend USC.’ b. ?So-ko-no syokuin-ga [subete-no giin]-ni that-place-gen staff-nom all-gen senator-dat USC to UCLA-o suisen-saseta. USC and UCLA-acc recommend-made Intended meaning: ‘[A (USC) staff and a (UCLA) staff] made all the senators recommend (each of) USC and UCLA.’

4.4 The Apparent QP must be ‘Salient’ Next, quirky binding is easier when the apparent QP is in a position which is salient enough to be a ‘topic’ of a sentence.13 Compare (38) with (30), repeated below.

13 Reinhart 1986: Appendix and references therein contain some discussions relevant to the notion ‘saliency’ in this sense.

50 

 Ayumi Ueyama

(38) a. ?✶So-ko-no bengosi-ga [Toyota to Nissan-to torihiki-ga that-place-gen attorney-nom Toyota and Nissan-with business-nom aru kaisya]-o suisensita (node, ato-wa dareka-ni exist company-acc recommended because rest-top someone-dat Mazda-o suisensite-mora(w)-eba ii dake da). Mazda-acc recommend-ask-if good only copula Intended (but unavailable) meaning: ‘(Since) a retained attorney recommended [the company which is doing business with Toyota and Nissan] (now we have only to ask someone to recommend Mazda).’ b. ?✶So-ko-no bengosi-ga [subete-no zidoosya-gaisya-no that-place-gen attorney-nom every-gen automobile-company-gen raibaru-gaisya]-o uttaeteiru (node, zidoosya-gyookai-wa rival-company-acc sued because automobile-industry-top daikonran-ni otiitteiru). disorder-dat be:thrown:into Intended (but unavailable) meaning: ‘(Since) a retained attorney has sued [the rival company of each automobile company] (, the automobile industry has been thrown into a state of disorder).’ (30) a. ?So-ko-no bengosi-ga Toyota to Nissan-o suisensita that-place-gen attorney-nom Toyota and Nissan-acc recommended (node, ato-wa dareka-ni Mazda-o Because rest-top someone-dat Mazda-acc suisensite-moraw-eba ii dake da). recommend-ask-if good only copula ‘(Since) {its/a retained} attorney recommended Toyota and Nissan (, now we have only to ask someone to recommend Mazda).’ b.

?So-ko-no bengosi-ga subete-no zidoosya-gaisya-o that-place-gen attorney-nom every-gen automobile-company-acc uttaeteiru (node, zidoosya-gyookai-wa daikonran-ni sued because automobile-industry-top disorder-dat otiitteiru). be:thrown:into ‘(Since) {its/a retained} attorney has sued every automobile company (, the automobile industry has been thrown into a state of disorder).’

As noted in Ueyama 1997, not only the depth of embedding but also the choice of the matrix verb may affect the ‘saliency’ of the apparent QP, although the acceptability of each sentence is expected to vary still more among the speakers.

On Non-individual-denoting So-words 

 51

(39) a. ?So-ko-no bengosi-ga subete-no kaisya-o uttaeteiru. that-place-gen attorney-nom every-gen company-acc sued ‘A retained attorney has sued every company.’ b. ?So-ko-no bengosi-ga subete-no kaisya-o suisensita. that-place-gen attorney-nom every-gen company-acc recommended ‘A retained attorney has recommended every company.’ c. ?So-ko-no bengosi-ga subete-no kaisya-o tubusita. that-place-gen attorney-nom every-gen company-acc bankrupted ‘A retained attorney has bankrupted every company.’ d. ?✶So-ko-no bengosi-ga subete-no kaisya-o ooensiteiru. that-place-gen attorney-nom every-gen company-acc support ‘A retained attorney supports every company.’ e. ?✶So-ko-no bengosi-ga subete-no kaisya-o that-place-GEN attorney-nom every-gen company-acc keibetusiteiru. despise ‘A retained attorney despises every company.’ f.

?✶So-ko-no bengosi-ga subete-no kaisya-ni ayamatta. that-place-gen attorney-nom every-gen company-dat apologized ‘A retained attorney has apologized to every company.’

g. ?✶So-ko-no bengosi-ga subete-no kaisya-to that-place-gen attorney-nom every-gen company-with arasotteiru. contend ‘A retained attorney is contending with every company.’ Since the ‘saliency’ in this sense is a property determined by pragmatics in nature, it is impossible to state a formal/syntactic condition regarding this aspect of the availability of quirky binding.

4.5 The So-word must be Non-individual-denoting Finally, among various kinds of so-words, so-ko ‘that place/institution’, so-re ‘that thing’ and so-no hito ‘that person’ allow non-individual-denoting reading

52 

 Ayumi Ueyama

relatively easily, while expressions such as so-itu ‘that guy’ and so-no booeki-gaisya ‘that trading company’ tend to be individual-denoting, for whatever reason.14 (40) a. ???So-itu-no bengosi-ga Tabata toyuu otoko-o uttaeta rasii. that-guy-gen attorney-nom Tabata comp man-acc sued they:say ‘They say that a retained attorney sued a man named Tabata.’ b. ?✶So-no booeki-gaisya-no bengosi-ga Tabata that-gen trading-company-gen attorney-nom Tabata toyuu booeki-gaisya-o uttaeta rasii. comp trading-company-acc sued they:say ‘They say that a retained attorney sued a trading company named Tabata.’

4.6 Summary Quirky binding is an instance of an apparent anaphoric relation yielding an apparent BVA reading. I have suggested in this section that there are several conditions for quirky binding to obtain, as summarized in (31). (31) a. b. c. d.

The apparent QP must “refer” to a specific group of individuals. Some constructions reject quirky binding. There should not be other quirky binding in the relevant clause. The apparent QP must be in a position which is salient enough to be a “topic” of a sentence. e. The so-word must be non-individual-denoting.

It is yet to be investigated how the apparent bound reading is yielded in quirky binding without any anaphoric relations, but for the purpose of revealing the nature of anaphoric relations, it is sufficient if we can properly eliminate the irrelevant cases, and we have argued that quirky binding can be avoided by not satisfying the conditions in (31), among which (31a) yields the clearest result. 14 Kare ‘he’, a so-called overt pronoun in Japanese, also hardly allows a non-individual-denoting reading. (i)

?✶Kare-no bengosi-ga Tabata toyuu otoko-o uttaeta rasii. he-gen attorney-nom Tabata comp man-acc sued they:say ‘They say that a retained attorney sued a man named Tabata.’

See footnote 1 for some remarks about this lexical item.

On Non-individual-denoting So-words 

 53

At this stage, we lack conclusive method to eliminate the cases which yield an apparent coreferential reading without involving an anaphoric relation. Therefore, although the conclusion in this chapter is based on the examples constructed as carefully as possible according to the current understanding of the relevant facts, there is a possibility that we will have to reconsider some part of our claim after the nature of non-individual-denoting so-words is further revealed.

5 Implicit Variable Binding in English I suspect that the ‘implicit variable binding’ discussed in Partee 1989 is likely an instance of something which yields an apparent bound reading without being based on an anaphoric relation.15 (41) a. Every sports fan in the country was at a local bar watching the playoffs. b. Every participant had to confront and defeat an enemy. c. Every traveler who stops for the night imagines that there is a more comfortable place to stay a few miles farther on. Partee 1989 reports that ‘implicit variable binding’ also obeys the c-command condition, on the basis of the observation in (42). (42) a. #?From five feet away I tried to toss a peanut to every pigeon. b. #?The leader of the local union wrote a letter to every untenured professor in the state. c. #?Only the nearest photographer got a good picture of every senator. If the sentences in (42) are excluded for the syntactic reason that the QP does not c-command the ‘implicit variable’, one should assume that the ‘implicit variable binding’ is based on a syntactic relation, and then it is expected that the ‘implicit variable binding’ is not available in any configuration which violates the LF c-command requirement. However, I have been informed that at least (43) can correspond to a situation in which every company is complained to by its own customer. (43) A customer complains to every company. 15 I thank Hajime Hoji for bringing Partee 1989 to my attention and making me aware of its relevance to this issue.

54 

 Ayumi Ueyama

If (43) is an instance of ‘implicit variable binding’, it becomes dubious that ‘implicit variable binding’ must be based on c-command. The contrast between (42) and (43) can be accounted for, on the other hand, if we regard ‘implicit variable binding’ as an instance something at least similar to quirky binding. Notice that the apparent QP is in the dative argument of a ditransitive predicate in (42a,b) and it is embedded in another NP in (42c). As mentioned in (31c), quirky binding rarely occurs in such cases. In contrast, the relevant NP in (43) is placed in a configuration in which quirky-binding is more readily available. I leave open how to analyze the ‘implicit variable binding’ in this work, especially for the reason that we need to know more about the syntactic functions of ‘the’ and ‘a(n)’ in order to discuss this issue.

References Chierchia, Gennaro. 1992. Anaphora and dynamic binding. Linguistics and Philosophy 15(2). 111–183. Hayashishita, J.-R. 2004. Syntactic and non-syntactic scope. Los Angeles, CA: University of Southern California dissertation. Hoji, Hajime. 1990. On the so-called overt pronouns in Japanese and Korean. In Eung-Jin Baek (ed.), Papers from the Seventh International Conference on Korean Linguistics, 61–78. Osaka: International Circle of Korean Linguistics & Osaka University of Economics and Law. Hoji, Hajime. 1991. KARE. In Carol Georgopoulos and Roberta Ishihara (eds.), Interdisciplinary approaches to language: Essays in honor of S.-Y. Kuroda, 287–304. Dordrecht: Kluwer Academic Publishers. Hoji, Hajime. 1995. Demonstrative binding and principle B. In Jill N. Beckman (ed.), NELS 25, 255–271. Amherst, MA: University of Massachusetts, Amherst, GLSA Publications. Hoji, Hajime. 1998. Formal dependency, organization of grammar, and Japanese demonstratives. Japanese/Korean Linguistics 7. 649–677. Hoji, Hajime. 2003. Falsifiability and repeatability in generative grammar: A case study of anaphora and scope dependency in Japanese. Lingua 113(4–6). 377–446. Hoji, Hajime. 2015. Language Faculty Science. Cambridge: Cambridge University Press. Hoji, Hajime. This volume: Chapter 5. Detection of c-command effects. In Hajime Hoji, Yukinori Takubo and Daniel Plesniak (eds.), The theory and practice of language faculty science. Berlin and Boston: De Gruyter Mouton. Kinsui, Satoshi. 1998. Nihongo-no shijishi-niokeru chokuji-yōhō to hi-chokuji-yōhō-no kankeinitsuite [On the relation between the deictic use and the non-deictic use of the Japanese demonstratives]. Unpublished ms., Osaka University. Kinsui, Satoshi and Yukinori Takubo. 1990. Danwa kanri riron-kara mita nihongo-no shijishi [A discourse management analysis of the Japanese demonstrative expressions]. Ninchi Kagaku-no Hatten [Advances in Japanese Cognitive Science] 3. 85–116. Reprinted in: Satoshi Kinsui and Yukinori Takubo (eds.) 1992. Kinsui, Satoshi and Yukinori Takubo (eds.). 1992. Shijishi [Demonstratives]. Tokyo: Hituzi Syobo.

On Non-individual-denoting So-words 

 55

Kitagawa, Chisato. 1981. Anaphora in Japanese: Kare and Zibun. In Ann K. Farmer and Chisato Kitagawa (eds.), Coyote papers 2, 61–75. Tucson, AZ: University of Arizona. Kuno, Susumu. 1973. The structure of the Japanese language. Cambridge: The MIT Press. Kuroda, S.-Y. 1979. (Ko) so, a nitsuite [On (ko), so and a]. In Hayashi Eiichi Kyōju Kanreki Kinen Ronbun-shū Kankō Iinkai (ed.), Eigo to nihongo to [English and Japanese], 41–59. Tokyo: Kurosio Publishers. Nakai, Satoru. 1976. A study of anaphoric relations in Japanese. Unpublished ms., University of Massachusetts, Amherst. Partee, Barbara. 1989. Binding implicit variables in quantified contexts. In Caroline Wiltshire, Randolph Graczyk and Bradley Music (eds.), Papers from the twenty-fifth regional meeting, 342–365. Chicago: Chicago Linguistic Society. Reinhart, Tanya. 1983a. Anaphora and semantic interpretation. Chicago: The University of Chicago. Reinhart, Tanya. 1983b. Coreference and bound anaphora: A restatement of the anaphora questions. Linguistics and Philosophy 6. 47–88. Reinhart, Tanya. 1986. Center and periphery in the grammar of anaphora. In Barbara Lust (ed.), Studies in the acquisition of anaphora 1, 123–150. Dordrecht: Springer. Saito, Mamoru and Hajime Hoji. 1983. Weak crossover and move α in Japanese. Natural Language and Linguistic Theory 1. 245–259. Takubo, Yukinori. 1996. The semantics and the semantic change of the Japanese distal demonstrative Kare. Paper presented at Japanese/Korean Conference, UCLA, Los Angeles, 9 November. Takubo, Yukinori and Satoshi Kinsui. 1997. Discourse management in terms of multiple mental spaces. Journal of Pragmatics 28(6). 741–758. Ueyama, Ayumi. 1997. Scrambling in Japanese and bound variable construal. Unpublished ms., University of Southern California. Ueyama, Ayumi. 1998. Two types of dependency, Los Angeles, CA: University of Southern California dissertation. Distributed by GSIL publications, University of Southern California, Los Angeles.

Emi Mukai

Research Heuristics in Language Faculty Science 1 Introduction As it is mentioned in Hoji’s Chapter 1 of this volume, language faculty science (LFS) aims to accumulate knowledge about the language faculty by the basic scientific method. In this chapter, I illustrate this attempt in a step-by-step manner, hoping it to serve as a guide for any interested researchers or researchers-to-be who wish to understand LFS with hands-on experience. This chapter, which is my rendition of the series of works by Hoji (at least up to 2015), addresses three important aspects, that is, (i) schematic asymmetries, (ii) the use of data involving some meaning relation of two linguistic expressions1 and (iii) the flowchart of conducting research. They are in turn all very crucial as (i) is needed to attain testability as discussed in Section 2, (ii) is to maximize testability as in Section 3, and the flowchart in (iii) will be introduced in Section 4 to maximize our chances of learning from errors (in a non-trivial manner).2, 3, 4 Section 5 deals with one question which is inevitable for LFS; what counts as facts in LFS. In Section 6 is an illustration of the experiments in line with the flowchart, and a summary of this chapter is provided in Section 7. 1 See Section 2 of Hoji’s Chapter 1 in this volume for meaning relation (MR). It will be introduced in Section 2.1 below as well. 2 The third heuristic (i.e., to maximize our chances of learning from errors) is adopted from Popper (1963). Listed in (i) are the first four of the 17 points that Popper gives that are crucial in the current discussion. (i) (Popper’s (1963: 965) (1)–(4)) a. All scientific knowledge is hypothetical or conjectural. b. The growth of knowledge, and especially of scientific knowledge, consists in learning from our mistakes. c. What may be called the method of science consists in learning from our mistakes systematically; first, by daring to make mistakes – that is, by boldly proposing new theories; and second, by searching systematically for the mistakes we have made, that is, by the critical discussion and the critical examination of our theories. d. Among the most important arguments which are used in this critical discussion are arguments from experimental tests. 3 “In a non-trivial manner” in the parentheses is due to Hoji’s (2010: footnote 36) concern about how exactly we can measure the degree of maximization in question. 4 The “attain testability” heuristic has epistemological priority over the other two heuristics; (i) is thus taken as the basis of (ii) and (iii). https://doi.org/10.1515/9783110724790-003

58 

 Emi Mukai

2 Attaining Testability 2.1 “Discrepancy” Between the Object of Inquiry and What is Observable It has been accepted in generative grammar since its inception (i) that its goal is to investigate the properties of an innate, algorithmic organ called the Computational System, which is hypothesized to lie at the center of the language faculty,5 and (ii) that speaker’s linguistic judgments are to be considered as a main source of evidence/data for constructing and evaluating the validity of hypotheses concerning the Computational System. However, these basic assumptions have invited enormous debates as to how hypotheses concerning the properties of the Computational System can be actually put to rigorous test. That is because it is not immediately clear how what is observable (performance in Chomsky 1965), that is, speaker judgments, is a direct reflection of the object of inquiry (competence in the sense of Chomsky 1965), that is, the properties of the Computational System. One of the main concerns in LFS is to identify those speaker judgments that are likely a (direct) reflection of the Computational System and hence that can be used as a tool to test hypotheses concerning the Computational System. The particular model of the Computational System put forth in Chomsky (1993) has been adopted in LFS (Hoji 2010: (2), Hoji 2015: Chap. 3, (1)). As shown in Figure 1, the input to the Computational System (given as CS in Figure 1) of this model is a set of items taken from the mental lexicon – such a set of items is called Numeration6 – and its output is a pair of mental representations, a PF (representation) and an LF (representation). The former is hypothesized to underlie what is directly observable in terms of ‘sounds’ (or their variants), and the latter ‘meanings’ in terms of how the selected items in Numeration are organized in an

5 One may object that the object of our inquiry is wider than just the Computational System and it is the language faculty. I am focusing on the Computational System at the moment because we can make testable predictions about the properties of the Computational System, as will be addressed below, but not about the language faculty outside the Computational System. 6 Note that in LFS, we adopt the idea of Numeration in one of the early stages of the Minimalist Program (e.g., Chomsky 1993/1995), not its “developed” versions in the subsequent stages that include “multiple Spell-out,” because it is not clear to us how to attain testability in the study of the properties of the Computational System with the latter while it seems possible to do so with the former, as will be discussed below. One might of course try to pursue the possibility of attaining testability with the latter, by articulating, and experimentally demonstrating, how that can be done.

Research Heuristics in Language Faculty Science 

 59

abstract structural representation.7 Under this view, the “pairing” of sounds (or signs) to meaning is expressed as the pairing of PF and LF. Numeration μ =>

CS => LF(μ) ß PF(μ)

Numeration μ: a set of items taken from the mental Lexicon LF(μ): an LF representation based on μ, hypothesized to underlie 'meanings' PF(μ): a PF representation based on μ, hypothesized to underlie 'sounds' (or their variants) Figure 1: The model of the Computational System (Chomsky 1993).

The object of our inquiry is the properties of the Computational System in Figure 1. A speaker’s answers to a question of the form in (1), for example, can be regarded as an instance of what is observable.8 (1) Question: Can you accept sentence α with meaning relation MR(X, Y)? MR(X, Y) stands for a meaning relation pertaining to two linguistic expressions X and Y (see Hoji’s Chapter 4: Section 3 and Chapter 5: Section 1.2 of this volume). For instance, Coref(John, himself), an instance of MR(X, Y), is taken as coreferential/ anaphoric dependency between ‘John’ and ‘himself’ when ‘John voted for himself’ is the sentence α in question. Dependency of so-called variable binding and so-called scope dependency are also instances of MR(X, Y) (e.g., BVA(everyone, his) for the meaning relation between ‘everyone’ and ‘his’ in ‘everyone defended his thesis’; DR(everyone, someone) for the meaning relation between ‘everyone’ and ‘someone’

7 Strictly speaking, an LF representation is hypothesized to underlie ‘meanings’ that are directly based on properties of the Computational System. That is to say, we are not concerned with ‘meanings’ that are ‘conveyable’ even with ungrammatical sentences such as I is a student or even ones like Mom with me shopping went. We also leave out for the time being until Section 5 the cases where certain interpretation is possible even if the necessary LF requirement for the interpretation is not met, which are called NFS (non-formal source) effects in Hoji’s Chapters 5 and 6 of this volume, the relevant discussion of which builds upon Ueyama’s (1998: Appendix D, Chapter 2 of this volume) discussion about quirky binding. 8 A speaker’s answers to a question of the form in (i) can also be regarded as a candidate of what is observable. (i) Question: Can you accept sentence α? In Section 2.3, I will return to the reason why I insist on working with (1) instead of (i).

60 

 Emi Mukai

in ‘everyone praised someone’). We can restate the rigorous testability issue as follows: How should we relate a speaker’s answers to (1) to the assessment of hypotheses regarding Figure 1? Or, how can a speaker’s responses to (1) be understood as revealing about properties of the Computational System in Figure 1? Hoji (2010) deals with this issue by adopting the view (first proposed by Ayumi Ueyama (e.g., Ueyama 2010)) that what speakers do when judging the acceptability of a sentence with a specified interpretation is to try to come up with a Numeration that produces (i) a PF representation that corresponds to the presented phonetic string and (ii) an LF representation that is compatible with (i.e., that satisfies the necessary condition(s) for) the relevant dependency interpretation.9 This commitment is the key in LFS. On the basis of this, Hoji maintains (2). (2) A (antecedent): A speaker accepts sentence α only if; C (consequent): he/she successfully comes up with a numeration that produces (i) a PF representation that is non-distinct from α and (ii) an LF representation (that satisfies the necessary condition(s) for the specified interpretation). Since, under Figure 1, an LF representation underlies ‘meanings’, it seems reasonable to assume that the availability of the specified interpretation is constrained by properties of an LF representation. One might wonder why this proposition is written with the form of ‘A only if C’, not of ‘A if C’ or of ‘A iff (=if and only if) C’. This is because there is no guarantee that speakers actually come up with such a numeration even if they (or their language faculty) have (or has) the ability to do so. In short, we adopt the proposition in (2), having the two points in (3) in mind; (3) a. Whenever a speaker accepts sentence α (with a specified interpretation) (that is, whenever a speaker does not say that α is completely unacceptable), that indicates that he/she (or his/her language faculty) has come up with such a numeration as given in (2). b. Even if such a numeration is in principle available to the speaker (or to his/her language faculty), there is no guarantee that he/she is actually able to come up with it.

9 Note that I am not claiming that adopting this view is the only way to deal with the rigorous testability problem; rather, the proposed view in LFS is just one attempt to overcome the testability problem.

Research Heuristics in Language Faculty Science 

 61

If a question of the form in (1) is taken into consideration, we obtain (4), which is a more concrete statement than (2). (4) A (antecedent): A speaker replies yes (to some extent) to a question of the form in (1) only if; C  (consequent): he/she successfully comes up with a numeration that produces (i) a PF representation that is non-distinct from α and (ii) an LF representation that satisfies the necessary condition(s) for MR(X, Y). Let us emphasize that the specified interpretation, MR(X,Y) in this case, is available only if some condition(s) is/are met at LF. Because of some common rules of inference, we can now relate a speaker’s judgments (i.e., what is observable) to the assessment/evaluation of hypotheses regarding Figure 1 (i.e., the object of our inquiry). In Section 2.2, I first discuss what is deduced from (4) in relation to prediction-making, having recourse to modus tollens (or denying the consequent) and one of common rules of inference, affirming the consequent. There, I also discuss what we can tell by looking at a speaker’s responses due to modus ponens (or affirming the antecedent) and some other rule of inference, denying the antecedent.10 This 10 Modus tollens (‘denying mode’) is sometimes referred to as denying the consequent and has the following argument form. (i)

If A, then C. (I.e., If A is true, then C is true.) (A: antecedent; C: consequent) Not C. (I.e., C is not true.) Therefore not A. (i.e., Therefore, A is not true.)

Modus Ponens (‘affirming mode’), on the other hand, is referred to as affirming the antecedent and has the form in (ii). (ii) If A, then C. (I.e., If A is true, then C is true.) (A: antecedent; C: consequent) A. (I.e., A is true.) Therefore, C. (i.e., Therefore, C is true.) Note that if A, then C (e.g., the first line of (i) and (ii) is equal to A only if C. Therefore, (4) can be rephrased as (iii). (iii) If: A (antecedent): a speaker replies yes to a question of the form in (1), then: C (consequent): he/she successfully comes up with a numeration that produces (i) a PF representation that corresponds to α and (ii) an LF representation that satisfies the necessary condition(s) for MR(X, Y).

62 

 Emi Mukai

discussion will lead us to the following; in order to attain testability, we need to have one type of prediction which says something is impossible, which is disconfirmable; another type of prediction that says something is possible plays a crucial role only if the former type of prediction has survived a rigorous disconfirmation attempt. In Section 2.3, I show that if one wants to deal with LF-related properties of the Computational System, it is necessary to work on predictions which involve an interpretation whose (im)possibility is necessarily constrained by (an) LFrelated condition(s), and one such instance is predictions involving MR(X, Y). As we will see, once we decide to narrow down our target of inquiry from any properties of the Computational System to LF-related properties and thus to deal with predictions of the type just mentioned, it follows that a ‘minimal unit of facts’ for such research consists of three types of schemata involving MR(X, Y), that is, one ✶Schema (read as ‘star-schema’) and two okSchemata. I will elaborate on this point further in Section 2.3.

2.2 Two Types of Predictions In the context of prediction-making, if one’s theory has a consequence that (4C) is false, we predict that (4A) is also false, which is due to modus tollens (‘denying mode’) which says “whenever ‘A only if C’ is true and C (consequent) is false, A (antecedent) must also be false” (see footnote 10). (5) is therefore deduced. (5)

Adopting (4) and due to modus tollens (‘denying mode’): a. When there is no numeration that would produce (i) a PF representation that is non-distinct from α and (ii) an LF representation that satisfies the necessary condition(s) for MR(X, Y) – and hence the speaker cannot come up with such a numeration; b. We predict that speakers always reply no to a question of the form in (1).

(1) Question: Can you accept sentence α with meaning relation MR(X, Y)? Notice that the prediction in (5b) is a prediction of the form that something is impossible. This is an instance of what is called a ✶Schema-based prediction. (6) A ✶Schema-based prediction: Speakers judge any ✶Example conforming to (or instantiating) a ✶Schema to be completely unacceptable with meaning relation MR(X, Y).

Research Heuristics in Language Faculty Science 

 63

The term ✶Schema is used to refer to a schematic representation (=a sentence pattern) that is predicted to be completely unacceptable (under a specified interpretation) according to the hypotheses one adopts/proposes. ✶Examples (‘star examples’) are sentences conforming to (or instantiating) a ✶Schema. If the consequent in (4) (i.e., (4C)) is not denied but affirmed, on the other hand, it does not entail, hence we do not always obtain, the affirmation of the antecedent in (4) (i.e., (4A)). This is due to one rule of inference, affirming the consequent, which says that “when ‘A only if C’ is true and C is true, A can be either true or false.” (7) Adopting (4) and due to affirming the consequent: a. When there is at least one numeration that will produce (i) a PF representation that is non-distinct from α and (ii) an LF representation that satisfies the necessary condition(s) for MR(X, Y); b. We predict that speakers can possibly reply either yes (to some extent) or no to a question of the form in (1). The prediction in (7b) is of the form that something is possible as in (8), which is called an okSchema-based prediction. (8) An okSchema-based prediction: Speakers judge some okExamples conforming to an okSchema to be acceptable (to varying degrees) with meaning relation MR (X, Y). An okSchema is a schema that minimally differs from the corresponding ✶Schema, based on our hypotheses, in terms of the hypothesized formal property, and sentences conforming to (or instantiating) an okSchema, which we shall call okExamples, are predicted to be not completely unacceptable. One of the main features of LFS is that the prediction is made regarding asymmetries among schemata, not regarding those among examples (i.e., actual sentences). That is because we are concerned with the properties of the Computational System, not those of individual lexical items. Therefore, the predictions should be given in the form of schemata, as given in (6) and (8) although the actual testing of the predictions is in relation to actual sentences instantiating a ✶ Schema and the corresponding okSchema.

64 

 Emi Mukai

2.3 Fundamental Asymmetry Between ✶ and ok

Notice that a ✶Schema-based prediction in (6) is a universal proposition while an okSchema-based prediction in (8) is an existential proposition. That is to say, ✶ Schema-based predictions are disconfirmable but not confirmable because, while it can be disconfirmed right away once one ✶Example conforming to ✶ Schema is judged acceptable, it is impossible to check all the possible ✶Examples conforming to ✶Schama and have them judged unacceptable. okSchema-based predictions, on the other hand, are confirmable but not disconfirmable by virtue of being an existential prediction, because it is impossible to check all the okExamples conforming to an okSchema and have them judged acceptable. Reflecting this distinction, Table 1 summarizes this fundamental asymmetry. Table 1: Confirmability and disconfirmability. Confirmation

Disconfirmation

Schema-based predictions possible

impossible

Schema-based predictions

possible

ok ✶

impossible

(Based on Hoji 2015: Chap. 2, (24))

Having addressed the fundamental issues about the context of predictionmaking, let us move on to the context of interpreting speaker judgments. Once we adopt (4), repeated here, it becomes possible to interpret a speaker’s responses as a manifestation of some properties of the Computational System. (4)

A (antecedent): A speaker replies yes (to some extent) to a question of the form in (1) only if; C (consequent): he/she successfully comes up with a numeration that produces (i) a PF representation that is non-distinct from α and (ii) an LF representation that satisfies the necessary condition(s) for MR(X, Y).

First, we obtain (9) due to modus ponens (‘affirming mode’) that says “whenever ‘A only if C’ is true and A is true, C must also be true” (see footnote 10).

Research Heuristics in Language Faculty Science 

(9)

 65

Adopting (4) and due to modus ponens (‘affirming mode’): a. When a speaker replies yes to the question in (1), that means; b. The speaker successfully comes up with at least one such numeration that produces (i) a PF representation that is non-distinct from α and (ii) an LF representation that satisfies the necessary condition(s) for MR(X, Y).

In (9a), (4A) is affirmed and thus we obtain (9b), which is the affirmation of (4C). When a speaker replies no, on the other hand, we might expect the negation of (4C), but that is not what the negation of (4A) entails. That is due to another rule of inference, denying the antecedent, which says “when ‘A only if C’ is true and A is false, C can be either true or false.”11 (10) Adopting (4) and due to denying the antecedent: a. When a speaker replies no to the question in (1), that means either (b–i) or (b–ii); b–i. There is no numeration that would produce (i) a PF representation that is non-distinct from α and (ii) an LF representation that satisfies the necessary condition(s) for MR(X, Y) – and hence the speaker cannot come up with such a numeration. b–ii. There exists at least one such numeration but he/she fails to come up with it for this time for some reason.11 Recall that even if such a numeration is in principle available to the speaker, there is no guarantee that he/she is actually able to come up with it (see (3)). Having (9) and (10) at hand, consider the following four possible scenarios. Table 2: Two types of predictions and four possible results. Prediction type Scenarios



Schema-based

ok

Schema-based

Scenario 1:

Speaker judgments: ok (disconfirmation)

Speaker judgments: ok (confirmation)

Scenario 2:

Speaker judgments: ok (disconfirmation)

Speaker judgments: ✶ (no confirmation)

11 Hoji (2010: Sections 2.3 and 2.4) suggests that there can be (at least) two possible reasons in (i-a) and (i-b) for the speaker not to come up with a possible numeration as in (10b–ii).

66 

 Emi Mukai

Table 2 (continued) Prediction type Scenarios



Schema-based

ok

Schema-based

Scenario 3: Predicted Asymmetry

Speaker judgments: ✶ (no disconfirmation)

Speaker judgments: ok (confirmation)

Scenario 4:

Speaker judgments: ✶ (no disconfirmation)

Speaker judgments: ✶ (no confirmation)

Recall that a ✶Schema-based prediction is a type of prediction that says that a speaker always replies no to a judgment question of the form in (1) (cf. (5) and (6)), while an okSchema-based prediction says that a speaker may reply either yes (=not no) or no (cf. (7) and (8)). In scenarios 1 and 2, the speaker judges ✶Examples acceptable and thus a ✶Schema-based prediction is disconfirmed. These scenarios are both very serious, since as we have seen in (9), the speaker’s ‘ok’ (=yes) should in principle necessarily mean the existence of a numeration that produces a pair of PF and LF representations, contrary to what is predicted under such a ✶Schema-based prediction. In these situations, it does not matter whether an okSchema-based prediction is confirmed or not; the disconfirmation of a ✶Schema-based prediction alone is already devastating and thus the confirmation of an okSchema-based prediction would not save these situations in any way. In scenarios 3 and 4, on the other hand, where a ✶Schema-based prediction is not disconfirmed, an okSchema-based prediction plays a crucial role. If it is not confirmed as in scenario 4, the validity of one’s hypotheses is not convincingly demonstrated, because speakers’ no can mean not only (10b–i) (as the researcher hypothesizes for the ✶Schema-based prediction in scenario 4) but also (10b–ii). After all, only when its corresponding okSchema-based prediction is confirmed as in scenario 3, it becomes plausible that the speaker replies no to ✶Examples conforming to a ✶Schema precisely because of (10b–i) as hypothesized. To sum up so far; we adopt (4) to deal with the testability problem (see footnote 9); and once we adopt (4) and due to modes tollens (‘denying mode’) and affirming the consequent (see (5) and (7)), it becomes clear that the key to attaining testability is to have a ✶Schema-based prediction because it is a ✶Schema-based

(i) a. Some extra-grammatical factors prevent the speaker from successfully coming up with such a numeration. (Failure of forming the numeration (or parsing difficulty))   b. The speaker does not like the intended interpretation to begin with because he/she finds it awkward/unnatural. (Unnaturalness of intended meaning)

Research Heuristics in Language Faculty Science 

 67

prediction that can be disconfirmed. Let us record this conclusion in (11) for the ease of reference. (11) To attain testability, one needs to have a ✶Schema-based prediction. I have also shown that, once we adopt (4) and due to modes ponens (‘affirming mode’) and denying the antecedent (see (9) and (10)), it becomes possible to interpret a speaker’s responses as a tool to evaluate hypotheses regarding the properties of the Computational System. More concretely, if a speaker replies yes to a judgment question (of the form in (1)) and the relevant example is an instance of the ✶Examples conforming to ✶Schema as in scenarios 1 and 2 in Table 2, the relevant ✶Schema prediction is considered disconfirmed (no matter how the “fate” of its corresponding okSchema prediction may turn out). On the other hand, if a speaker replies no to the same kind of examples as in scenarios 3 and 4, it implies that the hypotheses underlying the ✶Schema based prediction may be valid, and the value of the corresponding okSchema-based prediction becomes meaningful. In such a situation, the confirmation of the corresponding okSchema-based prediction enhances the plausibility that speakers reply no to ✶Examples precisely because the grammar does not allow it to correspond to any grammatical pair of PF and LF representations as hypothesized.

3 Maximizing Testability 3.1 Importance of MR(X, Y) We have dealt with predictions involving MR(X, Y) in our discussion so far. In this subsection, I argue that this choice is not arbitrary but inevitable; once we set our target of inquiry to LF-related properties of the Computational System, we should deal with predictions that involve MR(X, Y), as in (6) and (8), repeated here, instead of (12) and (13).12 (6)

A ✶Schema-based prediction: Speakers judge any ✶Example conforming to a ✶Schema to be completely unacceptable with meaning relation MR(X, Y).

12 The content of Section 3.1 is based on Hoji 2015: Chap. 2, Section 2.4.

68  (8)

 Emi Mukai

An okSchema-based prediction: Speakers judge some okExamples conforming to an okSchema to be acceptable (to varying degrees) with meaning relation MR(X, Y).

(12) A ✶Schema-based prediction with no specified interpretation (Cf. (6).): Speakers judge any ✶Example conforming to a ✶Schema to be completely unacceptable. (13) An okSchema-based prediction with no specified interpretation (Cf. (8).): Speakers judge some okExamples conforming to an okSchema to be acceptable (to varying degrees). Let us first show that by working on predictions involving MR(X, Y), we can, with some confidence, attribute the predicted unacceptability to the violation of some condition(s) at LF, not at any other level (e.g., at PF, at the level before-SpellOut, at the level of Numeration-formation etc.). To make our discussion concrete, suppose that, according to one’s proposal, at least one of the conditions for MR(X, Y) is not met at LF, making (14a) a ✶Schema when it is supposed to be interpreted with the meaning relation in (14b). (Note that “Y . . . X” in (14a) stands for a schema that contains X and Y with Y preceding X). (14) a. Schema: Y . . . X b. Specified interpretation: with MR(X, Y) c. Prediction: unacceptable This is a prediction of the form in (6); thus, any example conforming to the schema in (14a) is predicted to be completely unacceptable under the specified interpretation in (14b). Suppose, on the other hand, that (15a) is an okSchema because all the relevant conditions for MR(X, Y) are met at LF, corresponding to (15a). (15) a. Schema: X . . . Y b. Specified interpretation: with MR(X, Y) c. Prediction: acceptable This is the form of (8) and the prediction is that at least some example conforming to (15a) is judged acceptable (to varying degrees) with meaning relation MR(X, Y). Now, if the impossibility of (14a) with (14b) is crucially due to the violation of the condition(s) for MR(X, Y) at LF, not due to the sentence form itself instantiating the Schema in (14a), the researcher should be able to come up with acceptable sentence form instantiating the same Schema, but without the specified

Research Heuristics in Language Faculty Science 

 69

MR(X, Y), as indicated in (16). (16a) is meant to be identical to (14a), but there is no specified interpretation in (16b) unlike in (14b). (16) a. Schema: Y . . . X (=(14a)) b. No specified interpretation involving MR(X, Y) c. Prediction: acceptable That is to say, any example of the form in (14a)/(16a) should turn out acceptable if the interpretation MR(X, Y) is not at stake. If the prediction in (16) is borne out, it indicates that nothing is wrong with the following three processes; (i) the process of coming up with numeration by the speaker, (ii) the process of taking the numeration as the input by the Computational system, and (iii) the process of producing the PF representation that can surface as a phonetic string conforming to (14a)/ (16a). Therefore, when the predictions in (14) and (16) are both borne out, we can conclude that the failure to interpret the phonetic string in (14a) with MR(X, Y) as specified in (14b) is indeed due to the violation of the condition(s) at LF, not at any other level. Having (14) and (16) (as well as (15)) is possible only if we work on the predictions of the form in (6) and (8). I thus conclude that we can have speaker judgments revealing about LF-related properties of the Computational System by investigating predictions of the form in (6) and (8), not in the form of (12) and (13). Let us further illustrate the point, using some concrete examples.13 First consider the two Schemata in (17). (17)

a. b.

Schema: α V his β (with BVA(α, his)) Schema: His β V α (with BVA(α, his))

ok ✶

BVA(a, b), an instance of MR(X, Y), is abbreviation for the Bound Variable Anaphora construal between a and b. The okSchema-based prediction in (17a) and the ✶Schema-based prediction in (17b) are deduced from the hypotheses in (18) and (19).14

13 The concrete examples given in the rest of this subsection are purely for the sake of discussion, and we do not take into consideration until Section 5.1 the fact that some speakers find the ✶ examples acceptable. 14 Some of the hypotheses are written rather informally, but they are left so on purpose because they do not affect the content of our discussion here. For instance, I ignore the co-argument requirement for (what formally underlies) BVA for now though (i) is more precise than (19) (though still informal). See Hoji’s Section 7.2 of Chapter 5 in this volume. (i) BVA(a, b) is possible only if; a. a c-commands b at LF, and b. a and b are not co-arguments of the same verb (so-called binding principle B).

70 

 Emi Mukai

(18) Correspondence between a phonetic string and an LF representation The phonetic string a V b in English corresponds to an LF representation [ a [ V b] ] (where a asymmetrically c-commands b). (19) Condition for BVA(a, b) BVA(a, b) is possible only if a c-commands b at LF. Due to (18), α c-commands his at LF in the case of (17a) while α does not c-command his at LF in the case of (17b), leading to the predictions as indicated in (20). (20) Examples conforming to (17a) and (17b): a. okExample: Every player hates his former fans. (with BVA(every player, his)) ✶ b. Example: His former fans hate every player. (with BVA(every player, his)) DR(a, b), which stands for a’s wide-scope Distributive Reading with respect to b, is also an instance of MR(X, Y). (21) is an instance that involves this reading. (21) Three students voted for five professors. (with DR(three-students, five-professors)) The specified reading is such that each of three students voted for five professors (and the ‘referents’ of five professors vary depending on each of the students, meaning that the total number of professors voted for by the students is 15). The schemata in (22) are a typical DR-related paradigm. (22) a. b.

Schema: α V β Schema: β V α

ok ✶

(with DR(α, β)) (with DR(α, β))

The schematic asymmetry in (22) is due to (18) and the condition in (23).15 (23)

Condition for DR(a, b) DR(a, b) is possible only if a c-commands b at LF.

15 One might object to our view to take for granted the conditions in (18), (19) and (23), if he/she is familiar with some works whose claim is that weak-crossover effect is not that clear-cut (e.g., Y. Kitagawa 1990 and Kuno, Takami and Wu 1999) or inverse scope is sometimes possible (e.g., May 1977). This commitment of ours is, however, one of the key factors in maximizing our chances of learning from errors, as will be discussed in detail in Section 3.2.

Research Heuristics in Language Faculty Science 

 71

Since α (i.e., the intended wide-scope-taking element) c-commands β (i.e., the intended narrow-scope-taking element) in (22a), the DR reading is expected to be possible in examples conforming to (22a). In the case of (22b), on the other hand, the relevant c-command relation does not hold and thus the specified DR reading is predicted to be impossible in any examples conforming to (22b). Instances of (22) are such as those in (24). (24)

Examples conforming to (22): a. okExample: Three students voted for five professors. (with DR(three-students, five-professors)) b. ✶Example: Five professors voted for three students. (with DR(three-students, five-professors))

The above-mentioned two sets of schemata and examples are instances of those involving MR(X, Y). What is crucial is the fact that the schemata in (17b) and (22b) are no longer ✶Schemata if they do not have to be interpreted with the specified MR(X, Y) interpretation. For instance, consider (25a) and (25b) below; they are exactly the same as (17b) and (22b) in terms of their surface phonetic strings, but the specified interpretation is different. (25)

a.

(Cf. (17b).) ok Schema: His β V α (not with BVA(α, his); His ‘referring’ to one specific person) b. (Cf. (22b).) ok Schema: β V α    (not with DR(α, β))

With the BVA/DR condition in (19) and (23) not being at stake, Examples conforming to (25a) and (25b) are not predicted to be unacceptable, and they are indeed acceptable as indicated in (26). (26)

a.

Example conforming to (25a): His former fans hate every player.

(his ‘referring’ to one specific guy)

b. Example conforming to (25b): Five professors voted for three students. (not involving the scope interpretation between three-students and five-professors, meaning that the total number of professors voted for by the students is 5.) To sum up: the existence of the okSchema-based predictions in (25) entails that the ✶ Schema-based predictions in (17b) and (22b) are so predicted precisely because

72 

 Emi Mukai

the necessary conditions for BVA(α, his/him) and DR(α, β) are not satisfied at LF, as hypothesized. Let us now move on to predictions in (12) and (13), which do not involve MR(X, Y). Given in (27) is one set of such instances. (27)

a. b.

Schema: [NP Wh (N)] Aux NPSubj V? Schema: [NP Wh (N)] Aux NPSubj V1 [Complex NP N [that V2] ]?

ok ✶

Aux in (27) can be occupied by any auxiliary verb such as do/does, did, will, would, and should. In (27a), the NP at the initial position is interpreted as the object of V. In the case of (27b), on the other hand, the NP at the initial position is intended to be an argument of V2, the verb inside the complex NP. The examples in (28) conform to the Schemata in (27). (28)

a.

Example: Which book do you like? (which book is supposed to be the object of like) b. ✶Example: Which book do you like the author that wrote? (which book is supposed to be the object of wrote) ok

One could still claim that (27b) is a ✶Schema due to the violation of some LF-related constraint(s). However, there is no decisive empirical factor that would force us to adopt such a claim over a competing claim which would attribute the impossibility of (27b) to some other constraint(s) (e.g., to (a) PF-related constraint(s), to (a) before-Spell-Out constraint(s), or even to the difficulty in parsing.16 (29) and (30) are other instances of schemata and examples that do not involve MR(X, Y). (29) a. b.

ok

(30) a. b.

ok

Schema: Schema:



NP1 be Adj-er than NP2 NP1 be Adj than NP2

Example: John is taller than Bill. Example: John is tall than Bill.



We reach the same conclusion as we have with (27) and (28); there is no way to accept examples conforming to the schema in (29b) (e.g., (30b)) no matter how we would try to interpret them, and thus we cannot tell whether something is wrong at LF or at some other level (e.g., at PF or at some before-Spell-Out level), 16 It has been pointed out that not every native speaker finds (27b) unacceptable, to begin with, and thus the ✶Schema-based prediction is disconfirmed.

Research Heuristics in Language Faculty Science 

 73

regardless of how researchers wish to make a claim based on the unacceptable status of (29b). That in turn indicates that it is not clear how we can argue for or against hypotheses about LF-related properties of the Computational System decisively by looking at (29).17 To sum up the discussion; I maintain (31), following Hoji’s series of work; see Hoji 2015: Sections 3.7 and 5.6. (31)

To maximize the testability of hypotheses about LF-related properties, one should work on a Schema-based prediction about a sentence with meaning relation MR (X, Y).

3.2 Minimal Paradigm Requirement As a consequence of the discussion summarized in Section 3.1, Hoji suggests that the minimum paradigm in the study of LF-related properties of the Computational System must consist of schemata of the three types in (32). (32)

The Minimum Paradigm Requirement: a paradigm must minimally consist of schemata of the following three types (Based on Hoji 2015: Chap. 5, Section 2.2) a. okSchema1: δ (with MR(X, Y)) b. ✶Schema: α (with MR(X, Y)) c. okSchema2: α (not with MR(X, Y))

δ and α in (32) are schemata of phonetic strings. (32a) and (32b) share the specified interpretation but they are different in terms of their phonetic strings. (32b) and (32c), on the other hand, share the phonetic strings but the specified interpretation is different. The schema under the specified interpretation in (32b) is predicted to be unacceptable (i.e., it is a ✶Schema as given there) because at least one of the conditions (structural or lexical) for MR(X, Y) is not satisfied in any of the LF representations that could correspond to it, by hypothesis. The schema in (32a), which is an okSchema, minimally differs from (32b), and (all) the relevant structural and lexical condition(s) for MR(X, Y) is/are satisfied in at least 17 If one can provide coherent ways of taking care of concerns like PF constraints, spell-out constraints, parsing difficulty, etc., then it is in theory possible to make use of schemata and examples like those in (27)–(30), but this would not be an easy task given the degree of difficulty involved, especially when compared to the MR-based way of doing things, and these additional hypotheses introduced would thus interfere with maximizing testability.

74 

 Emi Mukai

one LF representation that can correspond to it, according to our hypotheses. We pair (32c) with the above two to make sure that the predicted unacceptability in (32b) is indeed due to some property/ies at LF. I maintain, following Hoji, that the speaker judgments can be regarded as a direct reflection of some LF-related properties of the Computational System only when we obtain schematic asymmetries (in judgment) conforming to (32). We have discussed the following two schematic asymmetries in English to illustrate the point, which satisfy the minimal paradigm requirement in (32). (33) a. b. c.

ok

(34) a. b. c.

ok



Schema1: Schema: ok Schema2:

α V his β (with BVA(α, his)) (=(17a)) His β V α (with BVA(α, his)) (=(17b)) His β V α (not with BVA(α, his)) (=(25a))

Schema1: Schema: ok Schema2:

αVβ βVα βVα



(with DR(α, β)) (with DR(α, β)) (not with DR(α, β))

(=(22a)) (=(22b)) (=(25b))

3.3 Predicted Schematic Asymmetries and Confirmed Predicted Schematic Asymmetries Before moving on, I have two more details to cover. First of all, let us introduce two terms to make the subsequent discussion easier to follow; predicted schematic asymmetries and confirmed predicted schematic asymmetries. When schematic asymmetries of the form in (32) are deduced based on one’s hypotheses, they are dubbed predicted schematic asymmetries. When a speaker’s judgments converge in line with the deduced predicted schematic asymmetries (i.e., when a ✶ Schema-based prediction survives rigorous disconfirmation attempt and its corresponding okSchema-based predictions come out as confirmed), such predicted schematic asymmetries are considered to be confirmed and we refer to such state of affairs as ‘we obtain a confirmed predicted schematic asymmetry’. For instance, (33) and (34) above are the predicted schematic asymmetries the former of which is deduced from (18) and (19) and the latter of which is from (18) and (23), respectively. If speaker judgments exhibit the patterns given in (33) and (34), then they are considered a confirmed schematic asymmetry. With those terms in mind, I record our conclusion as in (35). (35) The speaker judgments can be regarded as a direct reflection of some LF-related properties of the Computational System only when we obtain a confirmed predicted schematic asymmetry.

Research Heuristics in Language Faculty Science 

 75

In the next section, I construct, based on (35), a flowchart of how we should conduct our research so that we can learn from errors and make progress. I would now like to clarify the condition for MR(X, Y) in general. I have addressed BVA(a, b) and DR(a, b) as instances of MR(X, Y), and discussed (19) and (23) as the necessary conditions for them to arise. (19) Condition for BVA(a, b) BVA(a, b) is possible only if a c-commands b at LF. (23) Condition for DR(a, b) DR(a, b) is possible only if a c-commands b at LF. I have, however, only mentioned so far that “MR(X, Y) is a certain linguistic intuition involving two linguistic expressions X and Y that is available only if some condition(s) is/are met at LF.” I again follow Hoji’s series of works (see Chapter 1 of this volume) and maintain that the content of ‘some condition(s)’ for MR(X, Y) should also make reference to c-command as in (36). (36) Condition for MR(X, Y) MR(X, Y) is possible only if X c-commands Y at LF. Under this conception, the reference to c-command (19) and (23) is a consequence of the reference to c-command in (36). (36), in turn, is a consequence of adopting Chomsky’s (1993) model of the Computational System in LFS in (37) (see Hoji 2015: 29, for instance). (37) There is an operation Merge, internal and external, and that is the only structure-building operation in the Computational System. Given the assumption that the only structural relation ‘visible’ at LF is that of c-command, which is derivative of the characterization of the operation Merge, it follows that the structural condition on MR(X, Y) at LF must be stated in terms of c-command. This view is maintained in Reinhart 1983. In (38) and (39) are the two most general hypotheses in Reinhart 1983. (38) Reinhart 1983: 25, (19): Sentence-level semantic interpretation rules may operate on two given nodes A and B only if one of these nodes is in the domain of the other (i.e., A is in the domain of B, or B is in the domain of A, or both).

76 

 Emi Mukai

(39) Reinhart 1983: 26, (21): If a rule assigns node A some kind of prominence over node B, A must be a D[(=domain)]-head of the domain which contains B. Hoji (2009: Chapter 5) restates those two hypotheses as in (40) and (41), respectively, “taking “sentence-level semantic interpretation rules” as “CS-based rules or conditions that contribute to or regulate interpretive possibilities,” which seems to be a reasonable interpretation, given the discussion in Reinhart 1983.” (40) Hoji 2009: Chapter 5, (56) Reinhart 1983: 25, (19), restated: CS-based rules or conditions that contribute to or regulate interpretive possibilities can involve A and B only if A c-commands B, or B c-commands A, or both. (41) Hoji 2009: Chapter 5, (57) Reinhart 1983: 26, (21), restated: If B is dependent upon A in terms of how B gets interpreted, B must be c-commanded by A. Hoji then labels the heuristic in (42) the Reinhartian heuristic. (42) The Reinhartian heuristic: (Hoji 2009: chap. 5, (58); slightly adapted)18 The relation at LF that underlies MR(X, Y) must be based on a c-command relation between X and Y. By combining (36) and a hypothesis of how a phonetic string corresponds to an LF representation (e.g., (18)), we obtain predictions such as (33) and (34).

4 “Learning from Errors” in LFS We have seen so far (i) that there are two types of predictions, a ✶Schema-based and an okSchema-based prediction, (ii) that it is the ✶Schema-based prediction 18 The statement in (i) is a more precise version of (42). (i) The Reinhartian heuristic: (Hoji 2009: chap. 5, (58); slightly adapted) The relation at LF that underlies MR(X, Y) must be based on a c-command relation between what corresponds to X and what corresponds to Y at LF.

Research Heuristics in Language Faculty Science 

 77

that can lead us to attain testability in the study of the Computational System, and (iii) that the confirmation of predicted schematic asymmetries involving MR(X, Y) is the key to maximizing testability regarding LF-related properties of the Computational System. In this section, we turn our attention to another important heuristic, the “maximize our chances of learning from errors” heuristic.

4.1 When Confirmed Predicted Schematic Asymmetries are Obtained If a ✶Schema-based prediction and its corresponding okSchema-based prediction(s) are both borne out (i.e., if we have obtained a confirmed predicted schematic asymmetry), the use of the hypotheses that have given rise to those predictions (or the schematic asymmetry) in further discussions seems to be reasonably justified. If on the other hand the ✶Schema-based prediction is disconfirmed (i.e., if we fail to obtain a confirmed schematic asymmetry19), (i) (at least one of) the hypotheses underlying it or (ii) the experimental design, or (iii) both (i) and (ii), should be reconsidered. As we will see below, it is this reconsideration process that can give us a good chance of learning from errors. To clarify what is meant by learning from errors, let us first consider the chart in Figure 2.

(4) A new predicted schematic asymmetry formed

(3) (An) additional hypothesis/es introduced

(2) A set of hypotheses underlying the prediction becomes (a part of) licensed hypotheses

(1) A confirmed predicted schematic asymmetry obtained

(5) An experiment conducted Figure 2: How to proceed when confirmed predicted schematic asymmetries are obtained.

19 I am suppressing the possibility in the current discussion that the failure to obtain a confirmed schematic asymmetry can be due to the okSchema-based prediction failing to be confirmed.

78 

 Emi Mukai

How to read Figure 2 are given in (43). (43) a.

(1) and (2) in Figure 2 The arrow between (1) and (2) indicates that when we obtain a confirmed predicted schematic asymmetry (i.e., when the ✶Schema-based prediction has survived a rigorous disconfirmation attempt and the corresponding okSchema-based predictions are confirmed), and thus the hypotheses underlying the prediction have not been disconfirmed, I propose to refer to such ‘not having been disconfirmed’ hypotheses as “licensed hypotheses”. Licensed here means “provisionally licensed” because it means “not having been disconfirmed so far.”

b. (2), (3) and (4) in Figure 2 The arrow between (2) and (3) means that (a) new hypothesis/es is/are introduced and is/are combined with (some of) the licensed hypotheses. The arrow between (3) and (4) indicates that such addition in (3) should lead to a completely new predicted schematic asymmetries. c.

(4) and (5) in Figure 2 The arrow between (4) and (5) indicates that the newly formed predicted schematic asymmetry should be tested empirically.

d. (5) and (1) in Figure 2 The arrow between (5) and (1) means that the ✶Schema-based prediction has survived a rigorous disconfirmation attempt and the corresponding ok Schema-based predictions are confirmed. These processes in (43) can be repeated as many times as possible as long as we keep obtaining confirmed predicted schematic asymmetries. This is an ideal way to proceed.

4.2 When Confirmed Schematic Asymmetries are Not Obtained However, in reality, we often, if not most of the time, fail to obtain a confirmed schematic asymmetry. If nothing is wrong with the experimental design independent of the hypotheses in questions, such as the instructions to the speakers, that means that the researcher should try to modify (part of) the hypotheses responsible for the asymmetry. Although one could modify the hypotheses merely to eliminate the ‘counterexamples’ to the original predictions, we would like to suggest that modification processes should yield (a) new predicted schematic

Research Heuristics in Language Faculty Science 

 79

asymmetry/ies, in addition to eliminating the ‘counterexamples’. Otherwise, we will not be able to expect any (further) progress in our study of the Computational System insofar as we are to continue to pursue testability in our research.20 In short, we suggest that our research be conducted in the way illustrated in Figure 3 if we fail to obtain confirmed schematic asymmetries. (6) A confirmed predicted schematic asymmetry not obtained

(7) A set of hypotheses underlying the predictions becomes (a part of) unlicensed hypotheses

(8) (A) member(s) of the unlicensed hypotheses that has/have yielded the failure to obtain a confirmed schematic asymmetry get(s) modified (so as to accomplish both (i) the elimination of counterexamples and (ii) the deduction of a new predicted schematic asymmetry (i.e., a new *Schema-based prediction along with its corresponding okSchemabased predictions))

(4) A new predicted schematic asymmetry formed

(5) An experiment conducted

(1) A confirmed predicted schematic asymmetry obtained Continue to (2) in Figure 2 Figure 3: How to proceed when confirmed predicted schematic asymmetries are not obtained.

20 Modification processes that merely eliminate the ‘counterexamples’ to the original ✶Schema-based prediction without introducing a new ✶Schema-based prediction can be regarded as what Lakatos (1970/1978) calls regressive/content-reducing problemshift. Listed in (i) are the types of problemshift addressed in Lakatos (1970/1978); cf. Lakatos and Feyerabend 1999: 101, for example. (i) A problemshift is said to be: a. regressive/content-reducing if the new/revised hypothesis does not lead us to any new facts but only eliminates anomalies. b. theoretically progressive if the new/revised hypothesis eliminates anomalies and leads us to a new prediction. c. empirically progressive if at least some of the predictions are corroborated. By referring to the characteristics of problemshift in (i), we can rephrase our suggestion in the text discussion as follows; our modification process must not be regressive/content-reducing problemshift.

80 

 Emi Mukai

Given in (44) are the ways to read Figure 3. (44) a. (5) and (6) in Figure 3 This part indicates that, when conducting an experiment, the ✶Schemabased prediction has not survived a rigorous disconfirmation attempt and thus we fail to obtain a confirmed predicted schematic asymmetry, unlike (5) and (1) in Figure 2 (=(43d)) (but see footnote 19). b. (6) and (7) in Figure 3 The arrow between (6) and (7) indicates that when a confirmed predicted schematic asymmetry has not been obtained, the hypotheses underlying the prediction cannot be licensed hypotheses. I use the term “unlicensed hypotheses” to refer to the set of the hypotheses that have led to the failure to obtain confirmed predicted schematic asymmetries. c. (7), (8) and (4) in Figure 3 The arrow between (7) and (8) indicates that, if we fail to obtain a confirmed predicted schematic asymmetry, one or more members of the unlicensed hypotheses should be modified, and this modification process should lead us not only to eliminate the ‘counterexamples’ to the original predictions but also to the introduction of a new predicted schematic asymmetry as the arrow between (8) and (4) shows. d. (4) and (5) in Figure 3 ((4) and (5) in Figure 3 are the same as (4) and (5) in Figure 2 in (43c)). The arrow between (4) and (5) indicates that the newly formed predicted schematic asymmetry should be tested empirically. e. (5) and (1) in Figure 3 ((5) and (1) in Figure 3 are the same as (5) and (1) in Figure 2 in (43d)). The arrow between (5) and (1) means that the ✶Schema-based prediction has survived a rigorous disconfirmation attempt and the corresponding okSchema-based predictions are confirmed. If this newly-introduced predicted schematic asymmetry gets confirmed as shown in (44e), we can move on and have the chance to enter the loop illustrated in Figure 2. If it fails to get confirmed as in (44a), on the other hand, we need to go back and try another possibility. This is a method of trial and error. It seems reasonable to think that there are a number of new things to consider or to learn during the processes just alluded to. The “maximize our chances of learning from errors” heuristic therefore requires that we do our best to learn as much as possible during such modification processes.

Research Heuristics in Language Faculty Science 

 81

4.3 Flowchart of Conducting Research in LFS By combining Figures 2 and 3, we now have our Flowchart of conducting our research in Figure 4. Notice that the bottom part is the same as Figure 2, and the left part is the same as Figure 3. (6) A confirmed predicted schematic asymmetry not obtained

(7) A set of hypotheses underlying the predictions becomes (a part of) unlicensed hypotheses

= Figure 3 (8) (A) member(s) of the unlicensed hypotheses that has/have yielded the failure to obtain a confirmed schematic asymmetry get(s) modified (so as to accomplish both (i) the elimination of counterexamples and (ii) the deduction of a new predicted schematic asymmetry)

(4) A new predicted schematic asymmetry formed

(5) An conducted

= Figure 2

(3) (An) additional hypothesis/es introduced

(2) A set of hypotheses underlying the prediction becomes (a part of) licensed hypotheses

experiment

Figure 4: Flowchart of conducting our research (cf. Figures 2 and 3).

(1) A confirmed predicted schematic asymmetry obtained

82 

 Emi Mukai

I would like to propose Figure 4 as our attempt to maximize our chances of learning from errors and to make progress in the study of the properties of the Computational System21. Let us record this point as in (45). (45) To maximize our chances of learning from errors, we should conduct our research along the lines of the flowchart in Figure 4. Upon looking at Figure 4, one might naturally wonder where to ‘start’ our research in the flowchart. We will come back to this matter in Section 5.2.

4.4 Duhem’s Problem and the Study of the Computational System Because of the fact that a prediction is deduced not based on a single hypothesis but from a set of hypotheses (or a whole theory), researchers inevitably face the so-called Duhem’s problem, i.e., the problem that, when experiment disagrees with a prediction, it is in principle not possible to tell which one of the unlicensed hypotheses ought to be rejected (or “blamed”) because the disconfirmation only tells us that there is at least one hypothesis that contains an error.22 (Recall that the unlicensed hypotheses are the set of the hypotheses that fails to lead us to confirmed schematic asymmetries.) That is to say, when we proceed from the upper box to the lower box in Figure 5, we have no way of knowing a priori which member(s) of the unlicensed hypotheses is/are responsible for the failure and thus need(s) to be modified or eliminated.

21 When a researcher conducts a research (whether they are a practitioner of LFS or not), they often start by looking for a generalization of their own introspective judgments. But what matters the most to a practitioner of LFS is generalizations of schematic asymmetries (a variant of (4) in the Flowchart, as sometimes what we form is just a schematic asymmetry (not new or predicted)), crucially making a ✶schema a ‘central player’. That is to say, the disconfirmability/ testability is one of the key components of LFS. 22 Duhem (1954: 185) states “Duhem’s problem” as follows. (i)

The physicist can never submit an isolated hypothesis to the control of experiment, but only a whole ground of hypotheses. When experiment is in disagreement with his predictions, it teaches him that at least one of the hypotheses that constitute this group is wrong and must be modified. But experiment does not show him the one that must be changed.

Incidentally, Duhem’s problem is sometimes understood under the name of the so-called Duhem-Quine thesis. However, they are different in the way that Duhem does, but Quine does not, restrict his thesis to physics.

Research Heuristics in Language Faculty Science 

 83

(7) A set of hypotheses underlying the predictions becomes (a part of) unlicensed hypotheses

(8) (A) member(s) of the unlicensed hypotheses that has/have yielded the failure to obtain a confirmed schematic asymmetry get(s) modified (so as to accomplish both (i) the elimination of counterexamples and (ii) the deduction of a new predicted schematic asymmetry) Figure 5: (7) and (8) the chart in Figure 4, which are crucial in terms of Duhem’s problem.

Therefore, to conduct our research in line with the model in Figure 4 (which, I maintain, is necessary to maximize our chances of learning from errors), we need to try every possible candidate which might be responsible for the failure to obtain confirmed schematic asymmetries. In order to reduce the above-mentioned workload, Hoji (2009: chap. 3, sec. 6.3) makes a suggestion that can be summarized as (46). (46) Do not use a hypothesis to make a further theoretical deduction or to derive further empirical consequences if it is already regarded as a member of the unlicensed hypotheses in an independent experiment. By avoiding the use of any hypothesis that is already regarded as a member of unlicensed hypotheses in an independent experiment, we can reduce the number of (a) possible hypothesis/es that is/are responsible for the failure to obtain a confirmed schematic asymmetry.23 One concrete way to avoid using any member(s) of unlicensed hypotheses is to conduct preliminary experiments before testing the validity of (a) hypothesis/es that the researcher is actually concerned with.24 If we successfully obtain confirmed schematic asymmetries at this preliminary level, we can confidently declare that all the hypotheses underlying such asymmetries are not (a part of) the unlicensed hypotheses but (a part of) the licensed hypotheses, and we can confidently use any of them in the subsequent stage. The term preliminary should be taken rela-

23 For some concrete instance of the works against (46), the readers may wish to refer to Section 8.2 of Hoji 2010, which illustrates the point by making recourse to the well-known but invalid hypothesis (i.e., an instance of the hypothesis that is a member of the unlicensed hypothesis in our current term) that otagai and zibun-zisin in Japanese are instances of the [+A] category just as English each other and oneself are claimed to be. 24 As we will see in Section 5.2, preliminary experiments in the current discussion do not mean a pilot survey or experiment that aims to let speakers get used to experiments.

84 

 Emi Mukai

tively; what can be regarded as preliminary experiments always depends on what a researcher is actually concerned with at a given stage of their research. The Flowchart in Figure 4 leads us to assume (46). First of all, no member of the unlicensed hypotheses should be a part of theoretical deduction because the deduction of a (new) predicted schematic asymmetry should be achieved based on licensed hypotheses and (a) newly introduced hypothesis/es (see Figure 2). Secondly, when we fail to obtain a confirmed schematic asymmetry, the entire set of the hypotheses underlying the asymmetry becomes (part of) the unlicensed hypotheses and we cannot use any member(s) of them in the subsequent stage unless it/they get(s) modified (so as to accomplish both (i) the elimination of counterexamples and (ii) the deduction of (a) new predicted schematic asymmetry/ies) (see (8) in Figures 3, 4 and 5).25 Therefore, we should not use any hypothesis that is already regarded as a member of the unlicensed hypotheses in an independent experiment if we conduct our research along the lines of Figure 4.

5 Facts in LFS One might think that conducting research along the line of Figure 4 is not much different from the general method conducted in the field of natural science. It seems indeed quite similar to the “flowchart depicting the scientific method” on the https://www.britannica.com/science/scientific-method, for instance.26 However, 25 We should however understand that there are hypotheses that are not subject to refutation or modification. What I have listed in (i) are such hypotheses. (i)

(Based on Hoji 2009: chap. 2, (19), chap. 5, (7), (9), (11) & (12)) a. The Computational System (CS) exists at the center of the language faculty. b. The mental Lexicon exists. c. The CS is an algorithm. d. (i)  Input to the CS is a set of items taken from the mental Lexicon. (ii) Output of the CS is a pair of PF and LF representations. e. There is an operation Merge, and that is the only structure-building operation in the CS. (=(37))

Borrowing the notions of Lakatos’ 1970/1978 ‘scientific research programmes’, Hoji (2009: Chapter 5, 2015: Chapter 4, footnote 22) suggests that those hypotheses could be regarded as being part of the hard core in the study of the Computational System. (ii) Hard core: the hypotheses that are adopted without direct empirical evidence and not subject to refutation or modification (Hoji 2009: chap. 5, (5a)) 26 What we call licensed hypothesis in Figure 4 is labeled as “accepted hypothesis” in the flowchart on the link.

Research Heuristics in Language Faculty Science 

 85

unlike natural science, LFS does not have a “microscope” to look into facts. That is to say, LFS requires theory to determine which phenomena are reflections of the object of inquiry, that is, properties of the CS.27 In this section, I propose that, despite the claim in Section 4.4 that the term preliminary should be taken relatively, there is one type of preliminary experiments that is most fundamental and should be conducted every time (at least for the time being) at the very beginning of the research concerning the properties of the Computational System. This kind of experiments consists of schematic asymmetries the ✶Schema of which is the simplest case with the canonical/unmarked order (as opposed to the non-canonical/marked order)) of the target language. For instance, NPSubj V NPObj is the simplest case in English and NPSubj-nom NPObjacc V is in Japanese.28 Conducting the most fundamental preliminary experiments is necessary because (i) in reality, obtaining a confirmed schematic asymmetry is not an easy task even with the simplest case, and yet (ii) we need to obtain a confirmed predicted schematic asymmetry in order for us to conduct our research in line with Figure 4. This type of experiments is regarded as different from the other types and thus most fundamental because the aim to conduct such experiments is not to see if (a) certain predicted schematic asymmetry/ies get(s) confirmed, but to find linguistic expressions that induce meaning relation MR only if the necessary LF c-command relation is established29 and thus can most likely be suitable to use in subsequent stages. As we will see, this type of experiments lacks testability/falsifiability but it is nonetheless necessary if one wishes to conduct his/her research in line with the flowchart in Figure 4. Without the most fundamental experiments, our research activity would never get accomplished along the lines of Figure 4.

27 See Section 3 of Hoji and Plesniak’s Chapter 9 in this volume for the said difference between natural science (physics to be precise) and LFS. 28 That the simplest cases in English and in Japanese are NPSubj V NPObj and NPSubj-nom NPObj-acc V, respectively, is justified empirically as we will see in Section 5.1. 29 The linguistic expressions in question are those that do not allow quirky binding in Ueyama’s (Chapter 2 of this volume) terms, or those that do not induce NFS (non-formal source) effects in Hoji’s (Chapter 4 of this volume).

86 

 Emi Mukai

5.1 Inconsistent Judgments and Difficulty of Obtaining Confirmed Schematic Asymmetries For the schematic asymmetries in the simplest cases in English and in Japanese, we first maintain (47), which has been widely accepted in the field. (47) Correspondence between a phonetic string and an LF representation in the simplest case a. The phonetic string in (i) in English must correspond to an LF representation in (ii). (=(18)) (i) NP V NP (ii) [ NPSubj [ V NPObj] ] (where NPSubj asymmetrically c-commands NPObj) b. The phonetic string in (i) in Japanese must correspond to an LF representation in (ii).30, 31 (i) NP-nom NP-acc V (ii) [ NPSubj-nom [ NPObj-acc V] ] (where NPSubj asymmetrically c-commands NPObj) For the sake of expositions, we will refer to the Subject-Verb-Object order in English and the Subject-Object-Verb order in Japanese as the SVO order and the SOV order, respectively. Recall that it is necessary to investigate cases involving MR(X, Y) if one wishes to maximize the testability of hypotheses about LF-related properties and that MR(X, Y) is subject to the condition in (36). (36) Condition for MR(X, Y) MR(X, Y) is possible only if X c-commands Y at LF. 30 Strictly speaking, (47b–i) can also correspond to (i), since Japanese allows so called major subjects/multiple subjects and “empty nominals.” (i) [NPMajor Subj-nom [ecSubj [ NPObj-acc V] ] ] The subject in (i) is an empty category ec. This possibility however will not be taken into consideration for now, since that does not affect the current discussion in the sense that it is nonetheless the case that the subject asymmetrically c-commands the object in both (47b–ii) and (i). 31  (47b) says the same thing as (i) below, which is (13a) of Hoji’s Chapter 5 in this volume. (i) (=Hoji Chapter 5 in this volume: (13a))     NP2-ga NP1-o V (roughly, Subject Object Verb) in Japanese must correspond to a 3D (i.e., LF) representation where LF(NP2) asymmetrically c-commands LF(NP1).

Research Heuristics in Language Faculty Science 

 87

BVA(a, b) and DR(a, b) are instances of MR(X, Y) and we assume that they are constrained by (19) and (23), respectively, which are also repeated here. (19) Condition for BVA(a, b) BVA(a, b) is possible only if a c-commands b at LF. (23) Condition for DR(a, b) DR(a, b) is possible only if a c-commands b at LF. In the case of English, therefore, we obtain the ✶Schama-based predictions in (48) and (49), the former of which is deduced from (47a) and (19) and the latter of which is from (47a) and (23). (48)



(49)



Schema-based prediction with BVA(a, b) in the SVO order in English (=(33b)) ✶ Schema: [NP . . . b . . .]Subj V NPObj (with BVA(NPObj, b)) Schema-based prediction with DR(a, b) in the SVO in English (=(34b)) ✶ Schema: NPSubj V NPObj (with DR(NPObj, NPSubj))

Similarly in the case of Japanese, we obtain (50) and (51) because of (47b) and (19) for the former and (47b) and (23) for the latter. (50)



(51)



Schema-based prediction with BVA(a, b) in the SOV order in Japanese ✶ Schema: [NP . . . b . . .]-nom NP2-acc V (with BVA(NP2, b)) Schema-based prediction with DR(a, b) in the SOV order in Japanese ✶ Schema: NP1-nom NP2-acc V (with DR(NP2, NP1))

These are the ✶Schemata in the simplest cases in English and in Japanese that will be the core of the most fundamental preliminary experiments. Before moving on, however, we should address the fact that it has been pointed out that sentences conforming to (48), (49), (50) and (51) do not seem always unacceptable, with the specified MR(X, Y), in spite of their predicted values. For the case of (48), for instance, quite a few speakers find sentences like his roommate hates everyone acceptable with the intended BVA(everyone, his) reading, contrary to the predicted value in (48). Similarly, for the case of (49), not all the researchers have reported their judgments in accordance with the predicted value. For instance, one of the earliest and most well-known work on the distribution of DR in English is May (1977: 29–30), who claims that the sentence

88 

 Emi Mukai

some woman loves every man allows DR(every man, some woman), contrary to the prediction in (49). In the case of Japanese, also, some researchers have objected to the indicated impossibilities in (50) and (51) (e.g., Y. Kitagawa 1990 and Kuno, Takami and Wu 1999), and it is indeed not very difficult to come up with acceptable examples conforming to those schemata. Given in (52) and (53) are some instances of them.32 (52)

Cases where BVA(NPObj, b) is possible in the SOV order, contrary to (50): (Ueyama 1998: chap. 4, (80) (slightly adapted)33) a. ?[So-ko-no bengosi]Subj-ga Toyota to NissanObj-o suisensita That-place-no attorney-nom Toyota and Nissan-acc recommended (node, ato-wa dareka-ni Mazda-o suisensite-mora-eba (because, rest-top someone-dat Mazda-acc recommend-ask-if ii dake da). good only copula) ‘(Since) {[its/a retained} attorney]Subj recommended [Toyota and Nissan]Obj (, now we have only to ask someone to recommend Mazda).’ b. ?[So-ko-no bengosi]Subj-ga [subete-no zidoosya-gaisya]Obj-o That-place-no attorney-nom every-no automobile-company-acc uttaeteiru (node, zidoosya-gyookai-wa daikonran-ni sued (because, automobile-industry-top disorder-dat otiitteiru). be:thrown:into) ‘(Since) [{its/a retained} attorney]Subj has sued [every automobile company]Obj (, the automobile industry has been thrown into a state of disorder).’

32 The readings that are available contrary to those schemata are dubbed quirky binding and quirky scope by Ueyama (1998) and Hayashishita (2004), respectively. 33 The -no markers attached to the NPs in (50) and (51) are originally glossed as gen (i.e., genitive) in Ueyama 1998 and Hayashishita 2004. But I simply gloss it as -no here (as well as the other

Research Heuristics in Language Faculty Science 

(53)

 89

Cases where DR(NPObj, NPSubj) is possible in the SOV order, contrary to (51): a. (Ueyama 1998: chap. 2, (47) (slightly adapted)) Dareka Subj-ga uti-no subete-no sensyu Obj-o bikoositeiru Someone-nom our-no all-no athlete-acc shadow (toyuu koto-wa, zen’in-ga kiken-ni sarasareteiru toyuu (comp fact-top everyone-nom danger-dat exposed comp koto da.) fact  copula) ‘(The fact that) someone Subj is shadowing every athlete of ours Obj (means that everyone’s life is in danger.)’ b.

(Hayashishita 2004: chap. 2, (14b) (slightly adapted)) [Context: There are five bad-mannered students. You know the fact that several professors split up into five groups and went to visit each of the students. You describe your knowledge as follows.] Sukunakutomo dareka Subj-ga subete-no huryoo gakusei Obj-o At:least someone-nom all-no bad-mannered student-acc hoomonsita. visited ‘At least someone Subj visited every bad-mannered student Obj.’

For the sake of exposition, let us leave aside the cases of English, keeping it mind though whatever we will conclude in terms of the Japanese cases should also apply to the English ones. Upon facing those situations, one way to proceed along the lines of the flowchart of conducting research in Figure 4 would be to add (47b), (19) and (23) to the unlicensed hypotheses and not to use any of them to make a further theoretical deduction or to derive further empirical consequences until they are modified in line with Figure 4 (cf. the guideline in (46)).34 The question that will be discussed is, then; is (47b) responsible for the failure to obtain confirmed schematic asymmetries in accordance with (50) and (51), is it the c-command conditions ((19) in the case of BVA and (23) in the case of DR) that are invalid, or both (47b) and the c-command conditions (19) and (23) should be modified? instances of -no throughout this chapter), leaving it open whether -no is a genitive case-marker (e.g., C. Kitagawa and Ross 1982) or it is a pre-nominal form of the copula da (e.g., Kuno 1973: 25). 34 Note that if we decided to add all of (47), (19) and (23) to the unlicensed hypothesis, Duhem’s problem (see Section 4.4) would come into picture, however; the disconfirmation of the ✶Schema-based predictions just mentioned would not tell us which one(s) of (47), (19) and (23) should be modified. (19) and (23) should not be treated together because they are not used together for the deduction of our ✶Schema-based prediction.

90 

 Emi Mukai

There is some reason to believe that (47b) should be kept as it is. It has been noted that there is a clear tendency which indicates that the SOV order is unmarked while the Object-Subject-Verb (OSV) order is marked. For instance, it was claimed as early as in Kuroda 1969/1970 and maintained in Hoji 1985 (among many other places) that a sentence in the SOV order is unambiguous while a sentence in the OSV order is ambiguous with respect to the DR possibilities, as schematized in (54). (54)

SOV and OSV in terms of DR: (Kuroda 1969/1970, Hoji 1985) a. SOV: NPSubj-nom NPObj-acc V ok (i) (=(51a)) DR(NPSubj, NPObj) (ii) (=(51b)) ✶DR(NPObj, NPSubj) b. OSV: NPObj-acc NPSubj-nom V ok (i) DR(NPSubj, NPObj) ok (ii) DR(NPObj, NPSubj)

Similarly, in terms of the BVA possibilities, it is claimed/accepted in Saito 1985 and Hoji 1985, with respect to the schemata in (55), that, in the SOV order, BVA(NPSubj, b) is acceptable (as in (55a)) but BVA(NPObj, b) is not (as in (55b)), while, in the OSV order, both BVA(NPSubj, b) and BVA(NPObj, b) are possible (as in (55c) and (55d)). (55)

SOV and OSV in terms of BVA: (Saito 1985, Hoji 1985) a. SOV: NPSubj-nom [NP . . . b . . .]Obj-acc V okBVA(NPSubj, b) (=(50a)) b. SOV: [NP . . . b . . .]Subj-nom NPObj-acc V ✶BVA(NPObj, b) (=(50b)) c. OSV: [NP . . . b . . .]Obj-acc NPSubj-nom V okBVA(NPSubj, b) d. OSV: NPObj-acc [NP . . . b . . .]Subj-nom V okBVA(NPObj, b)

Based on those (empirical) observations, it has been widely assumed that the SOV order corresponds to an LF representation in which the subject asymmetrically c-commands the object (as recorded in (47b)). Though some researchers started to object in the 1990’s that the generalizations in (54a–ii) and (55b) are not valid as we have just seen above (see (52) and (53)), it is still the case that the readings in (54a–i) and (55a) are much more readily available than (54a–ii) and (55b). Besides, Ueyama (1998: Appendix D.2 as well as Chapter 2 of this volume) observes that non-(or extra-)syntactic factors affect the availability of the reading in (55b) while such is not the case for (55a). One such factor is ‘specificity’; she suggests that the intended binder must be easily taken as ‘referring’ to a specific group of individuals for the intended reading to be possible in (55b), though the reading in (55a) does not seem to be affected by such a factor. For instance, the

Research Heuristics in Language Faculty Science 

 91

objects in (52a) and (52b) (Toyota to Nissan ‘Toyota and Nissan’ and subete-no zidoosya-gaisya ‘every automobile company’, respectively) can be taken as referring to some specific group of individuals quite easily (in the contexts addressed in (52a) and (52b)). Likewise, it is observed in Hayashishita 2000, 2004: sec. 2.2 that the intended DR reading in (54a–ii) is, but that in (54a–i) is not, affected by non-syntactic factors, including ‘specificity’;35 while the former can be interpreted easily only if the wide-scope-taking-element in (54a) (i.e., the NPobj) is taken as referring to something/someone specific, the latter does not seem to be affected in that way. Therefore, considering those asymmetries between (54a–i) and (55a) on the one hand and (54a–ii) and (55b) on the other, we seem to be able to reasonably maintain that the fact that some instances of (54a–ii) and (55b) are acceptable despite the prediction(s) is due to some non-syntactic factors (including how easily particular lexical items (or NPs) can be taken as referring to something/ someone specific).36 We thus propose to maintain (47b) as it is and pursue the possibility that the c-command conditions on BVA and DR (i.e., (19) and (23)) are subject to modification. Now let us consider how we could modify the conditions on the BVA and DR readings in (19) and (23). What seems crucial is the fact that, while some instances of (50) (=(55a–ii)) and (51) (=(54a–ii)) are judged acceptable contrary to the predictions as we have just seen, it is also the case that the predicted values on (50) and (51) come out as expected with some lexical items (that cannot be taken easily as ‘referring to’ any specific entity), as observed in Ueyama 1998, Hoji 2003: Section 2 and Hayashishita 2004. That is to say, it seems the case that the confirmed schematic asymmetries are obtained if we use certain lexical items as the intended binder or as what is intended to take wide scope with respect to another scope bearing element. This state of affair can be summarized as in (56) and (57). (56)

BVA(a, b) is possible only if; a. a c-commands b at LF (cf. (19)); b. some non-(or extra-)syntactic factors allow it in some way; or c. both (a) and (b) are satisfied.

35 Hayashishita (2004: chap. 2, sec. 2.1) offers a review of works such as Ruys (1992), Ben-Shalom (1993), Beghelli and Stowell (1997) and Liu (1990) which have paid attention to similar ‘specificity effects’ in question on quantifier interpretation in English. 36 See Ueyama 1998: Appendix D.2, Ueyama’s Chapters 2 of this volume, Hoji’s Chapters 5 and 6 of this volume, and Plesniak 2022: 2.8 for more detailed discussions regarding these ‘non-syntactic factors’.

92  (57)

 Emi Mukai

DR(a, b) is possible only if; a. a c-commands b at LF (cf. (23)); b. some non-(or extra-)syntactic factors allow it in some way; or c. both (a) and (b) are satisfied.

If we replace (19) and (23) with (56) and (57), (50) and (51) are not ✶Schemata anymore and hence we could successfully explain why some instances of (50) and (51) are judged acceptable. It is however the case that (56) and (57) will not help deducing any new predicted schematic asymmetries (i.e., any new ✶Schema-based prediction involving MR(X, Y) and its corresponding okSchema-based predictions) because it does not mention how the effects of the ‘non-(or extra-)syntactic factors’ in question can be minimized. Recall that we have claimed in Section 3.1 that the deduction of a new predicted schematic asymmetry (along with the elimination of the anomaly) is necessary in the modification process in order to conduct our research along the lines of the flowchart in Figure 4. One may thus object to replacing (19) and (23) with (56) and (57). Moreover, it does not seem easy, if not impossible, to find a way to formalize restrictions that could fully eliminate the effects of such ‘non-(or extra-)syntactic factors’. Even if we could somehow find a way to do so, that would not, at least in any direct way, lead us to new insight into the properties of the Computational System since ‘non-(or extra-)syntactic factors’ fall outside the scope of the Computational System. To sum up the discussion so far; we have not obtained confirmed schematic asymmetries even in the simplest cases nor have we found a way to modify the hypotheses responsible for the failure along the lines of the flowchart in Figure 4.

5.2 The most Fundamental Preliminary Experiments 5.2.1 Proposal on the most Fundamental Preliminary Experiments Instead of further pursuing a way to ‘save’ the ✶Schama-based predictions (i.e., (50) and (51)) from disconfirmation by modifying the hypotheses that deduce the predictions in question (i.e., (47b) and (19)/(23)), we propose to consider them as the members of the most fundamental preliminary diagnostics. By making use of such diagnostics at the very beginning of our research, we may lay a firm foundation on top of which our theory can be built. That is to say, we propose to regard the experiments to test (50) and (51) (in the case of Japanese) along with their cor-

Research Heuristics in Language Faculty Science 

 93

responding okSchema-based prediction as a means to initiate our research along the lines of Figure 4. More concretely, while admitting that the BVA(a, b) and DR(a, b) are possible even if the c-command conditions in (19) and (23) are not met, we commit ourselves to the view in (58). (58) The view we adopt: If speaker judgments on sentences conforming to the ✶Schemata in (50) and (51) (in Japanese) come out as ‘unacceptable’, and those on sentences conforming to their corresponding okSchemata ‘acceptable’, the lexical items used as a and b of BVA(a, b) and DR(a, b) in such results most likely reflect the LF c-command relation and thus are most likely suitable to use in subsequent stages of investigation. The aim of conducting the most fundamental preliminary experiments is thus to find lexical items with which the predicted values in (50) and (51) and those in their corresponding okSchema-predictions are obtained (in the case of Japanese) for each speaker. Once we have identified such lexical items, we can use them in the subsequent stages of discussion and experiments with this speaker. At such stages, the conditions in (19) and (23) (as well as (47b)) are reasonably taken as part of the licensed hypotheses and they are ready to be combined with (a) newly introduced hypothesis/es to deduce (a) new predicted schematic asymmetry/ies. If we encounter the failure to confirm such (a) new predicted schematic asymmetry/ies at such stages, we will have a reasonably solid ground for assuming that the conditions in (19), (23) and (47b) are exempt from the modification process and that it is the newly introduced hypothesis/es that must undergo the process in question.

5.2.2 Justification of the most Fundamental Preliminary Experiments Our commitment to the view in (58) is due to (i) our goal, which is to reveal the LF-related properties of the Computational System, and (ii) our adoption of the Reinhartian heuristic in (42) (repeated below) which in turn is justified as long as we set the LF-related properties of the Computational System as our object of inquiry (see Section 2.3). (42) The Reinhartian heuristic: (Hoji 2009: chap. 5, (58); slightly adapted) The relation at LF that underlies MR(X, Y) must be based on a c-command relation between X and Y.

94 

 Emi Mukai

Therefore, as long as we aim at revealing the LF-related properties of the Computational System and as long as we wish to proceed as schematized in the flowchart in Figure 4, it is necessary to conduct the most fundamental preliminary experiments to select the most reliable lexical items to use in experiments at the subsequent stages. Let us also emphasize that completing the most fundamental preliminary experiments is not our goal but it is just preliminary; what is important is the subsequent stages. It is thanks to those preliminary experiments that we can expect to learn from errors in line with Figure 4 when we fail to obtain confirmed schematic asymmetries in the subsequent stages. In short, though the most fundamental preliminary experiments themselves lack testability/falsifiability, conducting them is required to obtain testability/falsifiability in the subsequent, more substantive stages of our research.

6 Illustrations of the most Fundamental Preliminary Experiments We illustrate in this section the most fundamental preliminary experiments, in relation to Japanese, and how to proceed from there following the flowchart in Figure 4.

6.1 Bound Variable Anaphora (BVA) Let us start with the minimum paradigm, the ✶Schema of which is the simplest case in terms of BVA in Japanese in (50). (59)

Most fundamental preliminary experiments regarding BVA(a, b) in Japanese a. okSchema1: [NP . . . b . . .]-acc NP1-nom V (with BVA(NP1, b)) b. ✶Schema: [NP . . . b . . .]-nom NP2-acc V (with BVA(NP2, b)) (=(50)) c. okSchema2: [NP . . . b . . .]-nom NP2-acc V (not with BVA(NP2, b))

Notice that the okSchema in (59a) is the Object-Subject-Verb order (henceforth the OSV order). We follow Ueyama 1998 and maintain the structural hypothesis in (60).37 37 This hypothesis is recorded as (13b) in Chapter 5 of this volume by Hoji. (i) (=Hoji Chapter 5 in this volume: (13b)) NP1-o NP2-ga V (roughly, Object Subject Verb) in Japanese can correspond to a 3D (i.e., LF) representation where LF(NP2) asymmetrically c-commands LF(NP1).

Research Heuristics in Language Faculty Science 

 95

(60) Correspondence between a phonetic string and an LF representation of the OSV in Japanese The phonetic string in (i) in Japanese can correspond to an LF representation in (ii). (i) NP-acc NP-nom V (ii) [ NPSubj-nom [ NPObj-acc V] ] (where NPSubj asymmetrically c-commands NPObj) Due to (60) and (19) (repeated below), we obtain the okSchema in (59a). (19) Condition for BVA(a, b) BVA(a, b) is possible only if a c-commands b at LF. One may naturally wonder why we do not use (61) as our okSchema, since (61) and (59b) are both the SOV order and thus seem to form a more straightforward minimal pair. (61)

Another possible candidate of okSchema for (59): NP1-nom [NP . . . b . . .]-acc V  (with BVA(NP1, b))

We avoid using (61) because the availability of the BVA in sentences conforming to (61) can be attributed to the surface precedence relation between the binder (NP1) and bindee (b), not to the hypothesized LF c-command relation between them (see Hoji 2015: 34 for more discussion). We will show below the results of five experiments. In each of the five experiments, each of the following five elements is used as the intended binder (i.e., a of BVA(a, b)). (62)

a. b. c. d. e.

NP-sae #% izyoo-no N(P) kanari-no kazu-no N(P) subete-no N(P) #-cl-no N(P)

‘even NP’ ‘more than #% of N(P)’ ‘a significantly large number of N(P)’ ‘all N(P)’ ‘# N(P)’ (e.g., san-nin-no hito ‘three people’)

For the intended bindee, that is, b of BVA(a. b), we use one type of the demonstrative NPs in Japanese that is formed with the demonstrative prefixes so (henceforth

96 

 Emi Mukai

simply a so-NP)38 though it seems that so-called zero pronouns (or empty categories; ec) or the other types of demonstrative NPs (ko-NPs and a-NPs) can be used in place of English personal pronouns as shown in (63) (because Japanese does not have the exact overt counterparts of English personal pronouns (e.g., Kuroda 1965: 105)). (63)

a.

John bought it.

b. John-ga John-nom

ec katta. bought

c.

John-ga John-nom

ko-re-o this-thing-acc

katta. bought

d. John-ga John-nom

so-re-o that-thing-acc

katta. bought

e.

a-re-o that-thing-acc

katta. bought

John-ga John-nom

First of all, however, among the demonstrative NPs, it is only so-NPs that can (clearly) function as a bound variable (e.g., Nishigauchi 1986 among many others). This point is illustrated in (64). (64) a.

(With BVA(every-automobile-company, so-ko)) ok Do-no zidoosyagaisya-mo so-ko-no bengosi-o which-no automobile:company-also that- place-no lawyer-acc hihansita. criticized (Functionally:) ‘Every automobile company criticized its lawyer.’

b. (With BVA(every-automobile-company, ko-ko)) ✶ Do-no zidoosyagaisya-mo ko-ko-no which-no automobile:company-also this-place-no hihansita. criticized

bengosi-o lawyer-acc

38 See Section 2 of Ueyama’s Chapter 2 in this volume for more details of the demonstrative NPs in Japanese.

Research Heuristics in Language Faculty Science 

c.

(With BVA(every-automobile-company, aso-ko)) ✶ Do-no zidoosyagaisya-mo aso-ko-no which-no automobile:company-also that-place-no hihansita. criticized

 97

bengosi-o lawyer-acc

Secondly, though zero-pronouns are most naturally used in our everyday conversations, and they seem to serve as a bound variable, evidence for their ‘existence’ is less than overwhelming (especially if the ec is intended to be a possessor as in (65a)). (65)

(With BVA(every-automobile-company1, ec1)) a. okDo-no zidoosyagaisya-mo1 ec1 bengosi-o hihansita. which-no automobile:company-also lawyer-acc criticized (Functionally:) ‘Every automobile company criticized its (=ec) lawyer.’ b.

Do-no zidoosyagaisya-mo1 [kyonen ec2 ec1 hinansita which-no automobile:company-also last:year criticized seezika2]-o uttaeta. politician-acc sued (Functionally:) ‘Every automobile company sued the politician who had criticized it (=ec) last year.’

ok

Furthermore, even if they do ‘exist’, it is doubtful that ec must be singular-denoting all the time. Hoji 2003 makes this point by making recourse to the possibility of ‘split antecedence’. He shows that, while the split antecedence is not possible with a so-NP as in (66a), it seems possible with ec as in (66b). (66) a.

Hoji 2003: (14a), which is the slightly adapted version of Hoji 1995: 259, (16) Toyota1-ga Nissan2-ni [CP zeimusyo-ga so-ko1+2-o Toyota-nom Nissan-dat tax:office-nom that-place-acc sirabeteiru to] tugeta (koto) is:investigating that informed (fact) ‘(that) Toyota1 informed Nissan2 that the tax office was investigating it1+2’ ✶

b. Hoji 2003: (16a) Toyota1-ga Nissan2-ni [CP zeimusyo-ga ec1+2 sirabeteiru to] Toyota-nom Nissan-dat tax:office-nom is:investigating that tugeta (koto) informed (fact) ‘(that) Toyota1 informed Nissan2 that the tax office was investigating them1+2’

98 

 Emi Mukai

If the intended bindee can be taken as a plural-denoting element, as soitu-ra in (67) is, it may not be easy, if not impossible, to tell the BVA reading from some other readings such as the co-reference (or group) reading in (67) and (68). (67) San-nin-no gakusei-ga so-itu-ra-no sidookyookan-o uttaeta three-cl-no student-nom that-guy-pl-no advisor-acc sued ‘Three students sued their advisors’ a. BVA(three students, so-itu-ra): each of three students sued his/her advisor b. Co-reference (or group) reading: three students1 sued their1 advisors (68) So-itu-ra-no sidookyookan-ga san-nin-no gakusei-o uttaeta that-guy-pl-no advisor-nom three-cl-no student-acc sued ‘Their advisors sued three students.’ a. BVA(three students, so-itu-ra): his/her advisor sued three students (=each of three students was sued by his/her advisor) b. Co-reference (or group) reading: their1 advisors sued three students1 It, therefore, seems that a singular-denoting so-demonstrative NP can, but the emptypronoun (ec), cannot be reliably used as the bindee (i.e., b of BVA(a, b)) in Japanese.39 Now, when we use one of (62) as the binder and so-ko ‘that place (=that institution)’ as the bindee, we obtain one set of example sentences conforming to (59). The sentences in (69), (70) and (71) constitute one such set, in which (62b) is used as the intended binder. (69) An example conforming to the okSchema in (59a) a. Sentence: [so-ko-no syokuin]Obj-o [55% izyoo-no tihoozititai]Subj-ga that-place-no staff-acc 55% or:more-no local:government-nom hihansita. criticized ‘Its staff member(s), each of 55% or more local governments criticized.’ b. Specified reading: With BVA(55% or more local governments, so-ko) = “55% or more local governments are such that each of them criticized its own staff member(s).” c. Prediction: acceptable

39 The interested readers may turn to Hoji et al 1999 and Hoji 2003 for more details.

Research Heuristics in Language Faculty Science 

 99

(70) An example conforming to the ✶Schema in (59b) a. Sentence: [So-ko-no syokuin]Subj-ga [55% izyoo-no tihoozititai]Obj-o that-place-no staff-nom 55% or:more-no local:government-acc hihansita. criticized ‘Its staff member(s) criticized each of 55% or more local governments.’ b. Specified reading: With BVA(55% or more local governments, so-ko) = “55% or more local governments are such that its own staff member(s) criticized each of them.” c. Prediction: unacceptable (71)

An example conforming to the okSchema in (59c) a. Sentence: [So-ko-no syokuin]Subj-ga [55% izyoo-no tihoozititai]Obj-o that-place-no staff-nom 55% or:more-no local:government-acc hihansita. criticized ‘Its staff member(s) criticized 55% or more local governments.’ b. Specified reading: Not with BVA(55% or more local governments, so-ko). So-ko refers to the Ministry of Finance. c. Prediction: acceptable

Likewise, we have constructed the other example sets conforming to the paradigm in (59), using one of (62) as the binder and a so-NP as the bindee. My own judgments on the relevant Examples are summarized in Table 3 below. The shaded cells are the results for the ✶Schema-based predictions. It seems to me that the results in (a) and (b) in Table 3 are in accordance with (59), while those in (d) and (e) in Table 3 are not. The one in (c) in Table 3 is in the gray zone. Based on these results, we conclude that it is less likely that the elements in (62d) and (62e) reflect the LF c-command relation, whereas the scores for (a) and (b) seem to suggest that they (most likely) do reflect the LF c-command relation, and are thus suitable to use in subsequent stages of discussion in the study of the Computational System, at least when I am the speaker for those experiments.

100 

 Emi Mukai

Table 3: The results of the most fundamental preliminary experiments on BVA (when the author herself is the speaker). a of BVA(a, b)

(59a)

(59b)

(59c)

a.

NP-sae ‘even NP’

(Cf. (62a).)

ok



ok

b.

#% izyoo-no N(P) ‘more than #% of N(P)’

(Cf. (62b).)

ok



ok

c.

kanari-no kazu-no N(P) ‘significantly large number of N(P)’

(Cf. (62c).)

ok

??

ok

d.

subete-no N(P) ‘all N(P)’

(Cf. (62d).)

ok

ok

ok

e.

#-cl-no N(P) ‘# N(P)’

(Cf. (62e).)

ok

ok

ok

6.2 Wide-scope Distributive Reading (DR) Let us now consider the paradigm involving DR in (72). (72) Most fundamental preliminary experiments regarding DR(a, b) in Japanese a. okSchema1: NP2-acc NP1-nom V (with DR(NP1, NP2)) b. ✶Schema: NP1-nom NP2-acc V (with DR(NP2, NP1)) c. okSchema2: NP1-nom NP2-acc V (not with DR(NP2, NP1)) Each of the elements in (62) (repeated below) is again used as a of DR(a, b), and in all the experiments, (62e) is used as the narrow-scope taking element (i.e., b of DR(a, b)). (62) a. b. c. d. e.

NP-sae #% izyoo-no N(P) kanari-no kazu-no N(P) subete-no N(P) #-cl-no N(P)

‘even NP’ ‘more than #% of N(P)’ ‘a significantly large number of N(P)’ ‘all N(P)’ ‘# N(P)’ (e.g., san-nin-no hito ‘three people’)

The reason why we use (62e) as b of DR(a, b) has to do with the issue of how we can be certain that speakers are judging a given sentence exactly under a ‘distributive reading’ as instructed. It seems the case that we are able to convince ourselves that the speakers judge a given sentence with DR(a, b) accurately as

Research Heuristics in Language Faculty Science 

 101

instructed only if the ‘referents’ of the intended narrow-scope taking element can vary depending upon the ‘referent’ of the intended wide-scope taking element, so to speak. The form #-cl-no NP in (62e) is an instance of such an element and thus it can be used as the narrow-scope taking element. On the other hand, it is not clear if the referents’ are varying in the above mentioned way when we use the other elements in (62a), (62b), (62c) and (62d), especially the ones in (62a) and (62d), as the intended narrow-scope taking element. For instance, consider the actual example in (73). (73) Go-nin-no bengosi-ga san-nin-no seizika-o uttaeta. five-cl-no lawyer-nom three-cl-no politician-acc sued ‘Five lawyers sued three politicians.’ Intended reading: (with DR(5 lawyers, 3 politicians)) The intended DR reading is such that the ‘suing-three-politicians’ property applies to each of five lawyers, and the ‘referents’ of the expression three politicians can vary depending on the ‘referents’ of the expression five lawyers. This reading is very easily distinguished from the one that says some specific three students were sued by (each of) five lawyers. Zen’in ‘all people/everybody’ in Japanese, on the other hand, is an instance of an element with which we cannot easily distinguish a ‘distributive reading’ from some other readings if they are used as a narrow-scope-taking element. Consider (74), for example. (74) San-nin-no gakusei-ga zen’in-o uttaeta three-cl-no student-nom all:people-acc sued ‘Three students sued everybody.’ Intended reading: (with DR(three students, zen’in)) The intended DR reading is such that the ‘suing-everybody-in-the-context’ property applies to each of three students. Because the referents of zen’in ‘all people/everybody’ cannot vary depending on the referents of san-nin-no gakusei ‘three students’, we cannot tell if the speakers are indeed judging the given sentence under the specified DR interpretation. For such reason, we avoid using elements in (62a) to (62d) as b of DR (a, b). Now, using one of (62) as a and (62e) as b of DR(a, b), we obtain example sets that conform to (72). The examples in (75), (76) and (77) constitute one such example set.

102 

 Emi Mukai

(75) An example conforming to the okSchema in (72a) a. Sentence: san-nin-no seizika Obj-o 55% izyoo-no tihoozititai Subj-ga three-cl-no politician-acc 55% or:more-no local:government-nom hihansita. criticized ‘Three politicians, 55% or more local governments criticized.’ b. Specified reading: With DR(55% or more local governments, three politicians) = “55% or more local governments are such that each of them criticized three politicians.” (The ‘referent’ of three politicians varies depending on each local government.) c. Prediction: acceptable (76) An example conforming to the ✶Schema in (72b) a. Sentence: San-nin-no seizika Subj-ga 55% izyoo-no tihoozititai Obj-o three-cl-no politician-nom 55% or:more-no local:government-acc hihansita. criticized ‘Three politicians criticized 55% or more local governments.’ b. Specified reading: With DR(55% or more local governments, three politicians) =“55% or more local governments are such that three politicians criticized each of them.” (The ‘referent’ of three politicians varies depending on each local government.) c. Prediction: unacceptable (77)

An example conforming to the okSchema in (72c) a. Sentence: San-nin-no seizika Subj-ga 55% izyoo-no tihoozititai Obj-o three-cl-no politician-nom 55% or:more-no local:government-acc hihansita. criticized ‘Three politicians criticized 55% or more local governments.’

Research Heuristics in Language Faculty Science 

 103

b. Specified reading: Not with DR(55% or more local governments, three politicians) but with DR(three politicians, 55% or more local governments) =‘Three politicians are such that each of them criticized 55% or more local governments.’ (The ‘referent’ of four local governments varies depending on each politician.) c. Prediction: acceptable Similarly, we have constructed other example sets conforming to the paradigm in (72), using the choices in (62) as the wide-scope taking element and an element of the form in (62e) as the narrow-scope taking element. My own judgments are given in Table 4. Table 4: The results of the most fundamental preliminary experiments on DR (when the author herself is the speaker). a of DR(a, b)

(72a)

(72b)

(72c) ok

a.

NP-sae ‘even NP’

(Cf. (62a).)

ok



b.

#% izyoo-no N(P) ‘more than #% of N(P)’

(Cf. (62b).)

ok



ok

c.

kanari-no kazu-no N(P) ‘significantly large number of N(P)’

(Cf. (62c).)

ok

??

ok

d.

subete-no N(P) ‘all N(P)’

(Cf. (62d).)

ok

ok

ok

e.

#-cl-no N(P) ‘# N(P)’

(Cf. (62e).)

ok

ok

ok

These results indicate that we have obtained the confirmed schematic asymmetry in line with (72) by using the elements in (62a) and (62b) as b of DR(x, b), but not by using (62d) and (62e). For instance, with the latter elements, I can easily accept the inverse scope, contrary to the prediction indicated in (72b). Recall that we have mentioned in Section 5.1 Hayashishita’s (2004) suggestion that the DR reading becomes easier in sentences of the form in (72b) only if the wide-scopetaking element is taken as referring to a specific group of individuals/entities. It seems indeed the case that the elements marked by ok in (72b) seem to be taken as referring to some specific group of individuals/entities fairly easily.

104 

 Emi Mukai

6.3 Summary of Section 6 In this section, we have provided some illustrations of the most fundamental preliminary experiments, using the elements in (62). (62)

a. b. c. d. e.

NP-sae #% izyoo-no N(P) kanari-no kazu-no N(P) subete-no N(P) #-cl-no N(P)

‘even NP’ ‘more than #% of N(P)’ ‘significantly large number of N(P)’ ‘all N(P)’ ‘# N(P)’ (e.g., san-nin-no hito ‘three people’)

The results seem to suggest that when I am the speaker, the elements in (62a) and (62b) are suitable for the subsequent, more substantive, stages of discussion. Now, having a speaker like me whose judgments match the predicted values in (59) and (72) with the lexical items (62a) and (62b), we can move on to our next step. At this point, (47b), (19) and (23) (repeated below) are already considered licensed as the solid circle in our flowchart in Figure 4 indicates. (47)

Correspondence between a phonetic string and an LF representation in the simplest case b. The phonetic string in (i) in Japanese must correspond to an LF representation in (ii). (i) NP-nom NP-acc V (ii) [ NPSubj-nom [ NPObj-acc V] ] (where NPSubj asymmetrically c-commands NPObj)

(19)

Condition for BVA(a, b) BVA(a, b) is possible only if a c-commands b at LF.

(23)

Condition for DR(a, b) DR(a, b) is possible only if a c-commands b at LF.

Now, we are ready to proceed and introduce a new hypothesis to test, as given in the dotted circle in Figure 2. After this point on, we conduct our research strictly in line with the flowchart. This is how we proceed in LFS. Before closing this section, let us emphasize one more time that getting the results shown in this section is not the goal of our inquiry, but rather, getting such results is just preliminary. That being said, it is still crucial and necessary to do so in order for us to conduct our research along the lines of the flowchart in Figure 4. Conducting our research in that way, in turn, is the key in LFS in maximizing our

Research Heuristics in Language Faculty Science 

 105

chances of learning from errors and thus in making progress in the study of the Computational System.

6.4 After the most Fundamental Preliminary Experiments One possible set of hypotheses that we would like to test in our future work has to do with the so-called ‘dative subject constructions’ in Japanese,40 which have been discussed extensively in the past literature (e.g., Kuno 1973, Shibatani 1978, 1999, Masuoka 2000 among many others). As Shibatani (1999: 45) puts it, “what appears to be a subject is marked by a dative or other oblique case” in these constructions, and they consist of the predicate types in (78). (78)

(Based on Shibatani 1999: 48 and Masuoka 2000: 237–242)41 a. Possession/existence b. Visual/auditory perceptions c. Modal state of necessity d. Modal state of potentiality

In (79) are some such examples containing each of the predicate types in (78). (79)

a.

(Possession/existence) Kono heya-ni-wa neko ga iru. this room-dat-top cat-nom be ‘There is a cat in this room. / This room has a cat.’

b. (Visual/auditory perceptions) Watasi-ni-wa neko-no nakigoe-ga kikoemasita. I-dat-top cat-no meow-nom was:audible ‘The cat’s meow was audible to me. / I heard the cat’s meow.’ c.

(Modal state of necessity) Ano kurasu-ni-wa tuyoi riidaa-ga that class-dat-top strong leader-nom ‘That class needs a strong leader.’

iru. be:needed

40 These are the same as what Hoji (Section 7 of Chapter 5 in this volume) calls “sentences with the “ergative” case-marking pattern.” 41 Shibatani (1999: 48) has two more predicate types, psychological states and physiological states, but I have excluded them from the list in (78) because the examples discussed in Shibatani as instances of those two cases are not Japanese.

106 

 Emi Mukai

d. (Modal state of potentiality) Tomoko-ni kankokugo-ga hanas-eru (koto) Tomoko-dat Korean-nom speak-potential (that) ‘(that) Tomoko can speak Korean.’ Given in (80) is the widespread claim regarding the dative subject constructions (e.g., Kuno 1973, Shibatani 1999).42 Vdat signifies any of the predicate types listed in (78). (80) A structural hypothesis re. the dative subject constructions The phonetic string in (i) in Japanese must correspond to an LF representation in (ii). (i) NP1-dat NP2-nom Vdat (ii) [ NP1-dat [ NP2-nom Vdat] ] (where NP1 asymmetrically c-commands NP2) With this claim, we are now able to form new predicted schematic asymmetries such as in (81). (81)

Predicted schematic asymmetries regarding subject constructions a. okSchema1: NP1-ni [NP . . . b . . .]-ga Vdat b. ✶Schema: [NP . . . b . . .]-ni NP2-ga Vdat c. okSchema2: [NP . . . b . . .]-ni NP2-ga Vdat

BVA(a, b) with the dative (with BVA(NP1, b)) (with BVA(NP2, b)) (not with BVA(NP2, b))

The dative subject constructions are often compared with the ‘canonical’ transitive GA-NI constructions such as those in (82) and we can form another set of predicted schematic asymmetries with them. In these examples, the verbs are not categorized as any of the predicate types in (78).

42 Strictly speaking, Shibatani’s (1999) claim is different from Kuno’s; While Kuno claims that the dative subject constructions are parallel to the canonical transitive constructions and thus the ni (dat)-marked one is the subject and the ga (nom)-marked one is the object, Shibtani argues against it and claims that the constructions must be assimilated with the so-called ‘double subject constructions’ (e.g., zoo-ga hana-ga nagai (koto) ‘(that) an elephant has a long trunk’) and thus both the ni-marked one and the ga-marked one are the subjects. Nonetheless, they share the assumption that the ni-marked one asymmetrically c-commands the ga-marked one.

Research Heuristics in Language Faculty Science 

 107

(82) a. Tanaka-ga ano seijika-ni toohyoosita (koto) Tanaka-nom that politician-dat voted (that) ‘(that) Tanaka voted for that politician.’ b. Tanaka-no inu-ga Tomoko-ni kamituita rasii. Tanaka-no dog-nom Tomoko-dat bit (I heard) ‘(I heard that) Tanaka’s dog bit Tomoko.’ c. Kurasu-no hotondo-no hito-ga Tomoko-no iken-ni class-no almost:all-no person-nom Tomoko-no opinion-dat sanseisita. agreed ‘Most of the people in the class agreed with Tomoko’s opinion.’ Those ‘canonical’ transitive GA-NI constructions have been analyzed as in (83), parallel to the transitive GA-O constructions in (47b). (83) A structural hypothesis re. the transitive GA-NI constructions The phonetic string in (i) in Japanese must correspond to an LF representation in (ii). (i) NP-nom NP-dat V (ii) [ NPSubj-nom [ NPObj-dat V] ] (where NPSubj asymmetrically c-commands NPObj) Now consider the examples in (84), the “scrambling” counterparts of (82), where the object shows up before the subject. (84) a. Ano seijika-ni Tanaka-ga toohyoosita (koto) that politician-dat Tanaka-nom voted (that) ‘(that) Tanaka voted for that politician.’ b. Tomoko-ni Tanaka-no inu-ga kamituita rasii. Tomoko-dat Tanaka-no dog-nom bit (I heard) ‘(I heard that) Tanaka’s dog bit Tomoko.’ c. Tomoko-no iken-ni kurasu-no hotondo-no hito-ga Tomoko-no opinion-dat class-no almost:all-no person-nom sanseisita. agreed ‘Most of the people in the class agreed with Tomoko’s opinion.’

108 

 Emi Mukai

It has been claimed that those sentences can still correspond to the LF representation where the subject asymmetrically c-commands the object, as they are a variant of the SOV order in (60). (85) A structural hypothesis re. the “scrambling” versions of the transitive GA-NI constructions The phonetic string in (i) in Japanese can correspond to an LF representation in (ii). (i) NP-dat NP-nom V (ii) [ NPSubj-nom [ NPObj-dat V] ] (where NPSubj asymmetrically c-commands NPObj) Combining this claim with the one in (80), the following new predicted schematic asymmetries in (86) can be formed as well. Note that the verb in (86a) is not Vdat, whereas the verbs in (86b) and (86c) are. (86) Predicted schematic asymmetries regarding BVA(a, b) with the dative subject constructions and transitive GA-NI constructions: a. okSchema1: [NP . . . b . . .]-ni NP1-ga V (with BVA(NP1, b)) b. ✶Schema: [NP . . . b . . .]-ni NP2-ga Vdat (with BVA(NP2, b)) c. okSchema2: [NP . . . b . . .]-ni NP2-ga Vdat (not with BVA(NP2, b)) Conducting experiments to test the newly formed predicted schematic asymmetries in (81) and (86) are two instances of the next step after the most fundamental preliminary experiments. We will leave the actual experiments for future work due to limitations of space.43 We nevertheless hope that the discussion in this subsection has given the readers a clearer understanding of LFS.

7 Summary LFS is an attempt to answer the perennial question of how hypotheses concerning the properties of the Computational System can be put to (rigorous) test. We have summarized this method by making reference to the following three research heuristics.

43 See Section 7 of Hoji’s Chapter 5 in this volume for more empirical data of these constructions.

Research Heuristics in Language Faculty Science 

 109

(87) Three research heuristics in LFS: a. Attain testability. b. Maximize testability on LF-related properties. c. Maximize our chances of learning from errors. (11), (31), and (45), repeated here, summarize the discussion in the preceding pages of the three research heuristics in (87). (11)

To attain testability, one needs to have a ✶Schema-based prediction.

(31)

To maximize the testability of hypotheses about LF-related properties, one should work on a Schema-based prediction about a sentence with a meaning relation MR (X, Y).

(45)   To maximize our chances of learning from errors, we should conduct our research along the lines of the flowchart in Figure 4. (11) is due to the claim in (2) that we commit ourselves to. (2)

A (antecedent): A speaker accepts sentence α only if; C (consequent): he/she successfully comes up with a numeration that produces (i) a PF representation that is non-distinct from α and (ii) an LF representation (that satisfies the necessary condition(s) for the specified interpretation).

Due to (2) and modus tollens (‘denying mode’), it follows that it is a ✶Schema-based prediction (i.e., a prediction of the form that something is impossible) that can be disconfirmed and hence that can make our hypotheses testable/falsifiable in the study of the Computational System. An okSchema-based prediction (i.e., a prediction of the form that something is not impossible) is not disconfirmable (though it is confirmable) because “something is in principle possible” does not always come out as “something is actually possible.” The necessity of MR(X, Y) as given in (31) comes from our desire to obtain a schematic asymmetry conforming to (32). (32)

The Minimum Paradigm Requirement: a paradigm must minimally consist of schemata of the following three types (Based on Hoji 2015: Chap. 5, Section 2.2) a. okSchema1: δ (with MR(X, Y))

110 

 Emi Mukai

b. c.

Schema: α (with MR(X, Y)) Schema2: α (not with MR(X, Y))



ok

The ✶Schema in (32b) and the okSchema in (32c) have the same phonetic string α but they differ from each other; the former is with MR(X, Y) while the latter is not. The existence of such an okSchema in (32c) indicates that sentences conforming to its corresponding ✶Schema would be acceptable if they did not have to be interpreted with MR(X, Y). If speaker judgments are obtained in line with what is indicated in (32b) and (32c), we can conclude that the unacceptable status of ✶ Schema is precisely because some of the conditions at LF is/are not satisfied. An ok Schema of this sort is available only if we work on a Schema-based prediction about sentence α with meaning relation MR (X, Y). Thus, we have reached the conclusion in (31). Introducing the term predicted schematic asymmetries (i.e., schematic asymmetries that conform to (32) and that are constructed as such based on one’s hypotheses) and confirmed predicted schematic asymmetries (i.e., the predicted schematic asymmetries that are confirmed by the checking of speaker judgments), we have constructed the Flowchart of conducting research in Figure 4 based on the view in (35). (35)

The speaker judgments can be regarded as a direct reflection of some LFrelated properties of the Computational System only when we obtain confirmed schematic asymmetries.

If confirmed predicted schematic asymmetries are obtained, we can justifiably use the hypotheses that give rise to those schematic asymmetries in further discussions. On the other hand, if they are not obtained, the hypotheses underlying them cannot be used to make a further theoretical deduction or to derive further empirical consequences unless they are modified so that not only the counterexamples are eliminated but also new predicted schematic asymmetries are deduced. All of these points have been summarized and schematized in the flowchart in Figure 4. I have also proposed that there is one type of preliminary experiments that is, for the time being, most fundamental. By conducting this type of experiments at the beginning of every research, we will be in a position to start our research of the Computational System in line with Figure 4. The most fundamental preliminary experiments in Japanese are to find lexical items with which the predicted values in (59) and (72) are obtained.

Research Heuristics in Language Faculty Science 

(59)

 111

Most fundamental preliminary experiments regarding BVA(a, b) in Japanese a. okSchema1: [NP . . . b . . .]-acc NP1-nom V (with BVA(NP1, b)) b. ✶Schema: [NP . . . b . . .]-nom NP2-acc V (with BVA(NP2, b)) (=(50)) c. okSchema2: [NP . . . b . . .]-nom NP2-acc V (not with BVA(NP2, b))

(72) Most fundamental preliminary experiments regarding DR(a, b) in Japanese a. okSchema1: NP2-acc NP1-nom V (with DR(NP1, NP2)) b. ✶Schema: NP1-nom NP2-acc V (with DR(NP2, NP1)) c. okSchema2: NP1-nom NP2-acc V (not with DR(NP2, NP1)) The asymmetry in (59) is deduced from (47b) and (19), while that in (72) is from (47b) and (23). (47) Correspondence between a phonetic string and an LF representation in the simplest case b. The phonetic string in (i) in Japanese must correspond to an LF representation in (ii). (i) NP-nom NP-acc V (ii) [ NPSubj-nom [ NPObj-acc V] ] (where NPSubj asymmetrically c-commands NPObj) (19)

Condition for BVA(a, b) BVA(a, b) is possible only if a c-commands b at LF.

(23)

Condition for DR(a, b) DR(a, b) is possible only if a c-commands b at LF.

The view in (58) is behind this idea of the most fundamental preliminary experiments, which in turn is justified if we wish to reveal about the properties of the Computational System in LFS. (58)

The view we adopt: If speaker judgments on sentences conforming to the ✶Schemata in (50) and (51) (in Japanese) come out as ‘unacceptable’, and those on sentences conforming to their corresponding okSchemata ‘acceptable’, the lexical items used as a and b of BVA(a, b) and DR(a, b) in such results most likely reflect the LF c-command relation and thus are most likely suitable to use in the subsequent stages.

112 

 Emi Mukai

Though the most fundamental preliminary experiments themselves lack testability/falsifiability, conducting them is necessary to have testability/falsifiability in the subsequent stages of discussion and experiments. Thanks to the most fundamental preliminary experiments, the c-command conditions in (19) and (23) (as well as (47b)) are regarded reasonably as being part of the licensed hypotheses at the subsequent stages of discussion and experiments, and they will be combined with (a) hypothesis/es that the researcher is directly concerned with at a given stage of research so as to deduce (a) new predicted schematic asymmetry/ies.

References Beghelli, Filippo and Tim Stowell. 1997. The syntax of distributivity and negation. In Anna Szabolcsi (ed.), Ways of scope taking, 71–108. Dordrecht: Kluwer Academic Publishers. Ben-Shalom, Dorit. 1993. Object wide scope and semantic trees. In Utpal Lahiri (ed.), Proceedings of SALT 3, 19–37. Washington, DC: Linguistic Society of America. Chomsky, Noam. 1965. Aspects of the theory of syntax. Cambridge, MA: MIT Press. Chomsky, Noam. 1993. A minimalist program for linguistic theory. In Kenneth Hale and Samuel Jay Keyser (eds.), The view from building 20: Essays in linguistics in honor of Sylvain Bromberger, 1–52. Cambridge, MA: MIT Press. Chomsky, Noam. 1995. The minimalist program. Cambridge, MA: MIT Press. Duhem, Pierre. 1954. The aim and structure of physical theory. Princeton, NJ: Princeton University Press. (Translated by Philip P. Wiener from the second edition, 1914, La théorie physique: Son objet, sa structure.) Hayashishita, J.-R. 2000. Scope ambiguity and ‘scrambling’. WCCFL 19. 204–217. Hayashishita, J.-R. 2004. Syntactic and non-syntactic scope. Los Angeles, CA: University of Southern California dissertation. Hoji, Hajime. 1985. Logical form constraints and configurational structures in Japanese. Seattle, WA: University of Washington dissertation. Hoji, Hajime. 1995. Demonstrative binding and principle B. In Jill N. Beckman (ed.), NELS 25, 255–271. Amherst, MA: University of Massachusetts, Amherst, GLSA Publications. Hoji, Hajime. 2003. Falsifiability and repeatability in generative grammar: A case study of anaphora and scope dependency in Japanese. Lingua 113. 377–446. Hoji, Hajime. 2009. A foundation of generative grammar as an empirical science. Unpublished ms., University of Southern California. Hoji, Hajime. 2010. Hypothesis testing in generative grammar: Evaluation of predicted schematic asymmetries. Journal of Japanese Linguistics 26. 25–52. Hoji, Hajime. 2015. Language faculty science. Cambridge: Cambridge University Press. Hoji, Hajime. This volume: Chapter 1. From compatibility to testability: Some historical background. In Hajime Hoji, Yukinori Takubo and Daniel Plesniak (eds.), The theory and practice of language faculty science. Berlin and Boston: De Gruyter Mouton. Hoji, Hajime. This volume: Chapter 4. The key tenets of language faculty science. In Hajime Hoji, Yukinori Takubo and Daniel Plesniak (eds.), The theory and practice of language faculty science. Berlin and Boston: De Gruyter Mouton.

Research Heuristics in Language Faculty Science 

 113

Hoji, Hajime. This volume: Chapter 5. Detection of c-command effects. In Hajime Hoji, Yukinori Takubo and Daniel Plesniak (eds.), The theory and practice of language faculty science. Berlin and Boston: De Gruyter Mouton. Hoji, Hajime. This volume: Chapter 6. Replication: Predicted correlations of judgments in Japanese. In Hajime Hoji, Yukinori Takubo and Daniel Plesniak (eds.), The theory and practice of language faculty science. Berlin and Boston: De Gruyter Mouton. Hoji, Hajime, Satoshi Kinsui, Yukinori Takubo and Ayumi Ueyama. 1999. Demonstratives, bound variables, and reconstruction effects. Proceedings of the Nanzan GLOW: The Second GLOW Meeting in Asia, September 19–22. 141–158. Hoji, Hajime and Daniel Plesniak. This volume: Chapter 9. Language faculty science and physics. In Hajime Hoji, Yukinori Takubo and Daniel Plesniak (eds.), The theory and practice of language faculty science. Berlin and Boston: De Gruyter Mouton. Kitagawa, Chisato and Claudia Ross. 1982. Prenominal modification in Chinese and Japanese. Linguistic Analysis 9(1). 19–53. Kitagawa, Yoshihisa. 1990. Anti-Scrambling. Unpublished ms., University of Rochester. https://ykling.sitehost.iu.edu/Resource%20files/Publication%20pdfs/AntiScrambling2000. pdf (accessed 21 March 2022) Kuno, Susumu. 1973. The Structure of the Japanese Language. Cambridge, MA: MIT Press. Kuno, Susumu, Ken-ichi Takami and Yuru Wu. 1999. Quantifier scope in English, Chinese, and Japanese. Language 75(1). 63–111. Kuroda, S.-Y. 1965. Generative grammatical studies in the Japanese language. Cambridge, MA: MIT dissertation. Kuroda, S.-Y. 1969/1970. Remarks on the notion of subject with reference to words like also, even or only. Annual Bulletin 3. 111–129. / Annual Bulletin 4. 127–152. Tokyo: Research Institute of Logopedics and Phoniatrics, University of Tokyo. Reprinted in S.-Y. Kuroda 1992, Japanese syntax and semantics: Collected papers, 78–113. Dordrecht: Kluwer. Lakatos, Imre. 1970. Falsification and methodology of scientific research programmes. In Imre Lakatos and Alan Musgrave (eds.), Criticism and the growth of knowledge, 91–195. Cambridge: Cambridge University Press. Lakatos, Imre and Paul Feyerabend. 1999. For and against method. Chicago: University of Chicago Press. Liu, Feng-hsi. 1990. Scope dependency in English and Chinese. Los Angeles, CA: University of California, Los Angeles dissertation. Masuoka, Takashi. 2000. Nihongo bunpō no shosō [Aspects of Japanese grammar]. Tokyo: Kurosio Publishers. May, Robert. 1977. The grammar of quantification. Cambridge, MA: MIT dissertation. Nishigauchi, Taisuke. 1986. Quantification in syntax. Amherst, MA: University of Massachusetts, Amherst dissertation. Plesniak, Daniel. 2022. Towards a correlational law of language: Three factors constraining judgment variation. Los Angeles, CA: University of Southern California dissertation. Popper, Karl. 1963. Science: Problems, aims, responsibilities. Federation Proceedings (Baltimore), Federations of American Societies of Experimental Biology Vol. 22(4). 961–972. Reinhart, Tanya. 1983. Anaphora and semantic interpretation. Chicago: University of Chicago Press. Ruys, Eddy. 1992. The scope of indefinites. Utrecht: Utrecht University dissertation. Published in the OTS Dissertation Series, Utrecht.

114 

 Emi Mukai

Saito, Mamoru. 1985. Some asymmetries in Japanese and their theoretical implications. Cambridge, MA: MIT dissertation. Shibatani, Masayoshi. 1978. Nihongo no bunseki [An analysis of Japanese]. Tokyo: Taishūkan. Shibatani, Masayoshi. 1999. Dative subject constructions twenty-two years later. Studies in the Linguistic Sciences 29. 45–76. Ueyama, Ayumi. 1998. Two types of dependency. Los Angeles, CA: University of Southern California dissertation. Ueyama, Ayumi. 2010. Model of judgment making and hypotheses in generative grammar. In Shoichi Iwasaki, Hajime Hoji, Patricia Clancy and Sung-Ock Sohn (eds.), Japanese/Korean Linguistics 17, 27–47. Stanford, CA: CSLI Publications. Ueyama, Ayumi. This volume: Chapter 2. On non-individual denoting so-words. In Hajime Hoji, Yukinori Takubo and Daniel Plesniak (eds.), The theory and practice of language faculty science. Berlin and Boston: De Gruyter Mouton.

Part 2: The Correlational Approach

Hajime Hoji

The Key Tenets of Language Faculty Science 1 Introduction Our ability to relate linguistic sounds/signs (henceforth, simply “sounds”) and meaning is rooted in our brain/mind; we refer to the component of the mind that underlies this ability as the language faculty. Following Chomsky’s general conception of the language faculty, we hold that it is uniform across the members of the human species in its initial state (barring serious impairment), and from  this  state, it undergoes changes based on its exposure to linguistic environment. (1) Linguistic Experience Initial State of Lang Faculty

Linguistic Experience Non-Initial State1 of Lang Faculty

Linguistic Experience Non-Initial State2 of Lang Faculty

...

Steady State1 of Lang Faculty

Linguistic Experience ...

Figure 1: Effects of linguistic experience on the language faculty.

A non-initial state undergoes further changes until “maturational” growth has stopped, resulting in “some relatively stable steady state . . ., which then undergoes only peripheral modification (say, acquiring new vocabulary items)”. (Chomsky 1986: 25) Following Chomsky 1986: 25–26, we refer to a steady state of the language faculty, as embodied in an individual,” as an I-language.1 An I-language consists of both what has been “acquired” based on exposure to linguistic environment and what was contributed by the initial state of the language faculty, as schematized in (2).

1 “I” in “I-language” stands for “internal(lized)”, “individual”, and “intensional”, pointing to the crucial nature of the subject matter (the language faculty). https://doi.org/10.1515/9783110724790-004

118 

 Hajime Hoji

(2)  Properties due to exposure to linguistic environment Contributions of the Initial State of Lang Faculty

Figure 2: A steady state of the language faculty (I-language).

The outer “circle” including the inner “circle” is meant to express the idea that what contributes to our linguistic behavior is a combination of (i) what is innate (hence, universal) and (ii) what is “acquired” (hence, I-language-particular). One feature of the basic scientific method is that we seek to deduce definite predictions from our hypotheses and obtain and replicate experimental results precisely in line with such predictions. Because the language faculty is internal to an individual and is understood to underlie our ability to relate sounds and meaning, the relevant investigation must be concerned, at least at the most fundamental level, with an individual’s linguistic intuitions about the relation between sounds and meaning. In order to accumulate knowledge about contributions of the initial state of the language faculty by the basic scientific method, we must therefore articulate each of (3a–c). (3) a. How we can deduce definite predictions about an individual’s linguistic intuitions regarding the relation between linguistic sounds and meaning b. How we can obtain experimental results pertaining to an individual precisely in line with such definite predictions c. How we can replicate the experimental results alluded to in (3b) This chapter presents a conceptual articulation of a research program with the aim just noted, which we refer to as language faculty science (LFS). This articulation will take the form of providing answers to each of (3a–c). The discussion in the chapter will be mostly conceptual, even when we address issues having to do with experiments in LFS (in relation to (3b, c)). A more concrete illustration of (3b) will be provided in Chapter 5, and (3c) will be illustrated more fully in

The Key Tenets of Language Faculty Science 

 119

Chapter 6 (with respect to Japanese) and in Chapter 7 (with respect to English), all based on actual experiments.2 Let us consider the features of the basic scientific method as stated in (4)–(6). (4) Guess-Compute-Compare (based on Feynman 1994 [1965]: 150): a. Guess-Compute: Deduce definite and testable predictions from hypotheses. b. Compare: Obtain definite experimental results and compare them with the definite and testable predictions. (5) Methodological Minimalism:3 Hypotheses must be stated in terms of the minimal number of theoretical concepts necessary. (6) Replication: Experimental results must be replicable. I contend that any scientific research program aspires to pursue this basic scientific method, as long as it is possible and strategically viable to do so. Chomsky in fact maintains that we should study the language faculty as natural scientists study their subject matters. Skepticism abounds, however, about the viability of such a claim, as expressed in Newmeyer 2008: Section 1 and Haider 2018, among many other places. The main source of the skepticism seems to be the absence of both (i) an articulated methodology for rigorously testing hypotheses, especially for obtaining and replicating experimental results in accordance with definite predictions deduced from hypotheses, and (ii) an empirical/experimental demonstration of the viability of such a methodology. Seeking to address these absences, this and other chapters of this volume build on and further articulate the claim made in Hoji 2015 that it is possible to accumulate knowledge about the language faculty by the basic scientific method. 3 We seek to accumulate knowledge about the language faculty by first deducing definite predictions from hypotheses about the language faculty and then obtaining and replicating experimental results in accordance with such

2 Further methodological details will be addressed in Chapter 8. 3 What is stated here is the aspect of methodological minimalism that is most relevant to our present discussion.

120 

 Hajime Hoji

predictions.4 In order to put forth our hypotheses and deduce testable predictions from them, we must have theoretical concepts in terms of which we formulate our hypotheses. For such a theoretical concept, we turn to Chomsky 2017, among many other places. Chomsky takes the ability to form an infinite number of sentences by using a finite number of basic discrete units, such as lexical items, as a defining property of the language faculty; he refers to it as “digital infinity”. Chomsky contends that this digital infinity requires the existence of what he terms the Computational System (CS) somewhere “within” the language faculty. Chomsky further claims the simplest possible computational operation whose recursive/ iterative application leads to the digital infinity of the language faculty is one that combines two items to form a set containing them.5 Chomsky thus suggests that the iterative application of this operation, called Merge, provides us with an account of the digital infinity in question. The methodological minimalism of the basic scientific method leads us to hypothesize that Merge is the only structure-building operation available in the CS of the language faculty, and that this is true for any member of the human species (until this minimal hypothesis has been shown to be inadequate by experimental results). We thus seek evidence for the Merge-based CS, in line with the basic scientific method, by making definite predictions about an individual based on hypotheses formulated by a concept based on Merge and obtaining and replicating experimental results in line with such predictions. If we focused on effects of Merge itself, i.e., the relation of two items merged, that would limit our investigation to effects of the very “local” relation, called “sisterhood” in earlier research on linguistic structure. That, however, would not allow us to account for the digital infinity in question, which involves (much) more than just the sisterhood relation. The account of digital infinity thus requires reference to recursive/iterative application of Merge. Adopting a term that has been used in research in generative grammar since the mid-1970s, let us refer to “c-command’ as defined in (7), which makes reference to recursive/ iterative application of Merge.

4 This leads to the recognition of the inseparability of facts and hypotheses in LFS: something counts as a fact only if it is both predicted by hypotheses and the prediction has been supported experimentally. LFS thus provides us with an example of a theory-laden research program even at very early stages of its development, as pointed out in Hoji 2015: 4–5. 5 Simpler computational operations, such as iteration (or iteration plus concatenation of the iterated elements) do not (obviously) lead to the type of infinity that the language faculty seems to produce.

The Key Tenets of Language Faculty Science 

(7)

 121

x c-commands y if and only if x is merged with something that contains y6

Detection of c-command effects (which we will often refer to below simply as c-command detection), once obtained in an individual and then replicated in other individuals, will thus provide support for this approach to the language faculty, and hence, for the existence of the Merge-based Computational System of the language faculty. Furthermore, an effective means for c-command detection will serve as a basis for further empirical and theoretical exploration of the properties of the language faculty, as will be illustrated in Chapters 5–7 of this volume. In light of the utmost importance of c-command detection in LFS, the following discussion focuses on what counts as c-command detection and what counts as its replication.

2 Merge and Abstract Representation If A and B are Merged, we get {A,B}, if C is merged with {A,B}, we get {C,{A,B}}, if D is merged with {C,{A,B}}, and we get {D,{C,{A,B}}}, etc. We can represent {D,{C,{A,B}}} in terms of a “tree representation as in (8).7 (8) {D,{C,{A,B}}} D

{C,{A,B}} C

{A,B} A

B

Recall that Merge combines two items and forms one. The linear order of the merged items as they appear in the externalized sentence is not part of what is

6 For the definition of contain, we adopt Plesniak’s 2022: 2.2.2 formulation in (i) and (ii), with slight adaptation. (i)

X and Y are the daughters of Z iff Z={X,Y}

(ii)

X contains Y iff: a.

Y is X’s daughter or

b.

Y is the daughter of an element Z that X contains.

7 The exposition here, including (8) is based on Plesniak 2022: 2.2.2.

122 

 Hajime Hoji

expressed by the output of application of Merge.8 Suppose, for example, that two items, V (verb) and NP (noun phrase), are merged. The result of the Merge is {V,NP}; V and NP are the two members of the set {V,NP}. If we want to conceive a result of the application of Merge in terms of a tree representation, we should therefore understand that a tree representation of {V,NP} can be thought of as represented like a mobile in 3D space (which I will call a “3D representation” for short), as illustrated in (9) by different images of the same set, {V,NP}, represented in this 3D sense.9 (9)

NP

V

NP

V

V

NP

V

NP

V

NP

V

NP

(8) should therefore be understood as a particular “2D image” of a 3D representation of the set {D, {C,{A,B}}}, in the sense that the linear order among D, C, A, and B in (8) is due to arbitrary choices among different 2D images.10 A sequence of sounds, such as a sentence, on the other hand, comes with a linear order specified among the audible elements within the sentence. Consider the Japanese sentence in (10), for example. (10) Mary-ga susi-o tabe-ta. Mary-NOM sushi-ACC eat-past ‘Mary ate sushi’ If a particular order of morphemes (e.g., tabe and ta) were reversed (showing up as “ta-tabe”, that would no longer express the concept corresponding to ‘ate’. Likewise, different orders of phonemes among /t/, /a/, /b/, and /e/, for tabe (‘eat’), would result in something that would no longer express the concept of ‘eat’. The I-language of a speaker of Japanese must include (morpho-phonemic) knowledge of that sort, and it must also include the knowledge that (10) is a specific instance 8 One may have a linearization algorithm (either learned and language specific or innate and universal) that operates based on the structures built by Merge in a consistent way. That would mean that Merge does come to define a(n at least partial) linear order, but this is not the core property of Merge itself. 9 This metaphor is made possible by the fact that our sets are binary; if we had a set {a, b, c}, for example, this physical metaphor would not be possible. 10 The choices will not be arbitrary once we adopt I-language-particular hypotheses about “linearization” (between “head” and “its complement”, for example).

The Key Tenets of Language Faculty Science 

 123

of (11), using NP to stand for the kind of phrase that can typically be followed by “case markers/particles” such as ‘-ga’, and ‘-o’. (11)

NP1-ga NP2-o V

Such knowledge must be a basis for native speakers’ intuition that an expression corresponding to something like ‘yesterday’ cannot occur between Mary and -ga or between susi and -o in (10) while it can in any of the places marked by “__” in (12). (12)

__ Mary-ga __ susi-o __ tabeta.

Now, schematically speaking, in reference to (11), and ignoring -ga and -o, the schematic phonetic sequence of NP1, NP2, and V by itself “does not come” with information about how NP1, NP2, and V are “structurally organized”, such as whether (11) corresponds to (13a) or (13b), for example. (13)

a. NP1 NP2 b.

NP1

NP2

V V

Under the conception of Merge adopted here, however, (11) cannot correspond to the structural representation in (13b) because Merge takes two items at a time and forms one (and hence, it cannot take three items at a time to form one, as in (13b)). According to (what was) an influential view (in generative grammar) in the early 1980s, however, the structural representation of (11) was said to be as in (13b), and thus “arguments” had to be made against this view in support of (13a).11 As noted above, what is intended by (13a), as output of application of Merge, can be viewed as a 3D representation, which we might try to approximate as in (14), for example, stressing that the linear order among NP1, V and NP2 is not considered fixed by the 3D representation itself.

11 Chapter 1 of this volume provides a critical review of the arguments presented in Hoji 1985, a representative work of this type (that is, arguing for strictly binary branching structure), from the perspectives of LFS.

124 

 Hajime Hoji

(14) NP1 V

NP2

What we are concerned with is the c-command relation(s) between the elements in such structures. C-command, however, is an abstract relation, not directly observable. What we can directly observe is linear precedence relation between the pronounced elements corresponding to, say, NP1 and NP2. However, a representation like (14), where NP1 c-commands NP2, does not itself encode the precedence relation, as noted above. Studying the properties of the language faculty by the basic scientific method thus requires, at the most fundamental level, hypotheses about how the surface and observable sequence of sounds are related to the abstract representations created by the recursive application of Merge.

3 Relating Sounds and Meaning As noted, given that the language faculty relates linguistic sounds and meaning (see (15)), hypotheses and predictions in LFS must be about what underlies this ability of ours. (15)

signs

/ sou

nds Language Faculty

meanin g Figure 3: The language faculty relating linguistic sounds/signs and meaning.

Now that we have the theoretical concept of c-command, we want to formulate hypotheses about the language faculty in terms of it; if we obtain and replicate experimental results as predicted under such hypotheses, this will serve as c-command detection. That is the basic logic of our approach.

The Key Tenets of Language Faculty Science 

 125

The c-command relation is a relation between two “elements”. For the purpose of c-command detection, we should therefore be dealing with linguistic intuitions pertaining to two such elements, at least at the most fundamental level.12 This leads us to seek to identify meaning relation (MR) pertaining to X and Y (MR(X, Y)) in a sentence such that the MR(X, Y) can arise only if the requisite c-command condition is satisfied, as long as unwanted noise has been eliminated.13 The availability of such MR(X, Y) should be indicative of the c-command relation holding between X and Y,14 as indicated in (16).15 (16) a phonetic sequence containing X and Y

c-command? Yes No

MR(X, Y) possible

MR(X, Y) not possible Figure 4: The availability of MR(X, Y) should be indicative of the c-command relation.16

Under this view, we see that the type of meaning that we are concerned with in LFS is not the meaning of the entire sentence in question but a part of its meaning that has to do with X and Y, specifically that which is crucially based on c-command, at least at the most fundamental level.17

12 As research proceeds, we may deal with linguistic intuitions that may not be directly related to c-command, but such intuitions should still relate (indirectly) to c-command (or at least Merge). 13 We are focusing on cases where the sequence of sounds in question is a sentence. 14 Strictly speaking, the c-command relation is not between two linguistic expressions X and Y in a phonetic sequence, but it is between two abstract objects in an abstract representation corresponding to X and Y, as will be discussed shortly. 15 The meaning relation pertaining to ‘every engineer’ and ‘his’ in ‘Every engineer praised his robot’, which is referred to as BVA(every engineer, his) in Hoji 2015 and earlier works and also in other chapters of this volume, is an instance of MR(X, Y). Specific instances of MR(X, Y) and related issues will be discussed more in depth in other chapters of this volume; see Chapter 5 in particular. 16 The c-command relation is asymmetrical, as will be addressed immediately below in relation to Figure 5. 17 As remarked in footnote 12, we expect success of research focusing on c-command to help us explore other aspects of the language faculty and related issues beyond “c-command” per se.

126 

 Hajime Hoji

When we view a tree structure like (13a), repeated here, we “see” that NP1 precedes NP2. (13) a. NP1 NP2

V

Given that a tree structure like (13a) is meant to be a 2D image of a “3D representation”, however, we cannot simply read such a precedence relation off the tree representation. It may be useful, therefore, to clearly distinguish elements in the phonetic sequence and those in a 3D representation like (14). Elements in a phonetic sequence can be directly observed, but we cannot directly observe elements in an abstract representation, such as a 3D representation like (14). We can only “observe” the latter by means of hypotheses and experiments. To make the distinction, I shall employ the notation LF(X) to refer to an abstract object in a 3D representation corresponding to a linguistic expression X in a phonetic sequence. Our use of “LF” follows its historical use in generative research.18 The clarification given above and the notation of LF(X), “c-commands?” in (16) can be given a more specific content, as in (17).19

18 Chomsky 1976, May 1977, and Chomsky 1981, 1995, for example. 19 Recall that “MR(X, Y)” is a meaning relation holding between two linguistic expressions X and Y. As pointed out to me by Daniel Plesniak (p.c., February, 2022), one may understand that we are having X (and Y) represent not only a linguistic object in a PS, but also “arguments” of MR, hence presumably something that expresses a meaning. LFS research as presented here, however, focuses on the meaning relation between X and Y without (directly) addressing the “meanings” of X and Y. As discussed, the most basic property of the language faculty that can be made testable has to do with the c-command relation, whose effects we seek to detect based on the (un)availability of MR(X, Y). Our focus on the meaning relation between X and Y, as a tool for exploring the language faculty, is thus a consequence of LFS seeking to accumulate knowledge about the language faculty by the basic scientific method. It is stated earlier (below (16)) that “the type of meaning that we are concerned with in LFS is not the meaning of the entire sentence in question but a part of its meaning that has to do with X and Y, specifically that which is crucially based on c-command, at least at the most fundamental level.” We can state likewise that the type of meaning that we are concerned with in LFS is not the meaning of specific linguistic expressions, at least at the most fundamental level.

The Key Tenets of Language Faculty Science 

 127

(17) a phonetic sequence (PS) containing X and Y

Does LF(X) c-command LF(Y) in LF(PS)?

Language Faculty

Yes No

MR(X, Y) possible

MR(X, Y) not possible Figure 5: The language faculty relating a sentence containing linguistic expressions X and Y to a meaning relation holding between X and Y (MR(X, Y)), based on whether LF(X) c-commands LF(Y).

Our general concern with the ability to relate linguistic sounds and meaning (see (15)) has now been articulated as concern with the ability to relate a phonetic sequence (PS) containing linguistic expressions X and Y to a 3D representation corresponding to it (i.e., LF(PS)). Ideally, whether LF(X) does or does not c-command LF(Y) will determine the possibility or the impossibility of MR(X, Y) for the PS in question. This articulation has been prompted by the adoption of the basic scientific method. We recognize that we can identify facts in LFS only by deducing definite predictions from hypotheses and obtaining and replicating experimental results precisely in line with such predictions. Based on this recognition, we seek to detect c-command effects in any I-language and replicate successful c-command detection in other I-languages. For two “elements”, x and y, in a 3D representation generated by application of Merge, x either c-commands y or it does not.20 As long as y is contained in what is merged with x, as in (18), x c-commands y.21

20 The use of small letters, x and y, is to indicate that we are considering “objects” in the LF representation of a sentence. This distinguishes them from capital letters, which, as stated, represent linguistic expressions in a PS. It is not intended that x have a systematic relationship with X such that x=LF(X); x is simply a variable for LF objects, just as X is a variable for PS objects. 21 See note 6.

128 

 Hajime Hoji

(18) X

X y y

If x is not merged with something that contains y, as in (19) or (20), on the other hand, x does not c-command y. (19) y

y

X

y

X

y

X

X

(20)

y

x

x

y

In (19), four images of the same 3D/LF representation are given, and in (20), two images of the same 3D/LF representation are given. There are thus two types of LF representations with regard to two “objects” in the LF representation; one type where x c-commands y, as in (18), and the other type where it does not, as in (19) and (20). Making reference to this, we can classify any sentence that contains X and Y into two types, [-cc]S and [+cc]S, as in (21). (21) a. b.

S: a schematized PS that cannot correspond to LF(PS) in which LF(X) c-commands LF(Y) [-cc]

S: a schematized PS that can correspond to LF(PS) in which LF(X) c-commands LF(Y) In order to relate a PS containing X and Y to LF(PS), and specify whether LF(X) does or does not c-command LF(Y), we must have hypotheses about correspondences between a PS (e.g., a sentence) and its LF representation. [+cc]

The Key Tenets of Language Faculty Science 

 129

As noted, whether LF(X) c-commands LF(Y) is a function of where LF(X) and LF(Y) appear in the LF representation. At a minimum, therefore, our hypotheses must specify how a given PS is mapped to LF(PS), specifically to what parts in the structure X and Y are mapped, so that we can address whether LF(X) c-commands LF(Y). This makes it necessary for us to analyze a PS as consisting of “chunks” of things, not just a sequence of sounds.22 Among the numerous types of sentence patterns and possibilities regarding where in the sentence X and Y may appear, we can consider the so-called transitive-sentence pattern, as in (11), repeated here. (11) NP1-ga NP2-o V Suppose that we have a hypothesis that (11) must correspond to the 3D representation in (14), where LF(NP1) asymmetrically c-commands LF(NP2).23, 24 If X is NP1 and Y is (or is contained in) NP2, as indicated in (22a, b), LF(X) c-commands LF(Y). (22) a. [NP1X]-ga [NP2 Y]-o V b. [NP1X]-ga [NP2 Y-no N]-o V On the other hand, if X is NP2 and Y is (or is contained in) NP1, as indicated in (23a, b), LF(X) does not c-command LF(Y). (23) a. b.

[NP1Y]-ga [NP2 X]-o V [NP1Y-no N]-ga [NP2 X]-o V

The presence of a “case marker” such as -ga in (23) is assumed not to prevent LF(NP1) from c-commanding LF(NP2).25 In term of our [+cc]S and [-cc]S notations, the sentence patterns (which we may refer to as “schemata”) in (24) are instances of [+cc]S in whose corresponding LF rep-

22 We leave aside, at least for now, how the speaker analyzes, i.e., parses, a PS as a sequence of such “chunks and what underlies the speaker’s ability to do so, hoping that we might be able to address the relevant issues in a testable manner once we have established an effective means to detect c-command effects. 23 Since the c-command relation as defined here is asymmetrical, ‘asymmetrically’ here is redundant; its mention here is for an emphasis. 24 Recall that as 3D (i.e., LF) representations, (14) and (13a) are identical. 25 The assumption is compatible with the view that morphological realizations of case markers, for example, are “visible” in the PS but not at LF.

130 

 Hajime Hoji

resentations of which LF(X) can, and in this case in fact does, c-command LF(Y), and those in (25) are instances of [-cc]S, in the corresponding LF representations of which LF(X) does not c-command LF(Y).26 (24) a. X-ga Y-o V b. X-ga Y-no N-o V (25) a. Y-ga X-o V b. Y-no N-ga X-o V 3D (i.e., LF) representations corresponding to (24a, b) and (25a, b) are provided in (26)–(29).27 (26)

(For (24a))

Y-o (27)

X-ga

X-ga

X-ga

Y-o

V

V

V

Y-o

(For (24b))

X-ga

X-ga

X-ga V

Y-no

N-o

V

V

Y-no

N-o

Y-no

N-o

26 (24) and (25) are the same as (22) and (23), respectively, except that “NP1” and “NP2” are removed in the former for simpler presentation. 27 “Case markers” such as -ga, -o, and -no, in (26)–(29) are only for convenience; see note 25 and the text remark thereabout.

The Key Tenets of Language Faculty Science 

(28)

(For (25a))

Y-ga

Y-ga

Y-ga

X-o (29)

 131

X-o

V

V

V

X-o

(For (25b))

Y-no

Y-no

N-ga

X-o

V

N-ga

Y-no

V

X-o

V

N-ga

X-o

4 Hypotheses and Judgments We have so far introduced concepts/notations in (30). (30) a. b. c. d. e. f. g.

PS: phonetic sequence X and Y: linguistic expressions in PS MR(X, Y): meaning relation pertaining to X and Y28 LF(PS): a 3D (i.e., LF) representation corresponding to PS LF(X): what corresponds to X in LF(PS) [-cc] S: a schematized PS that cannot correspond to LF(PS) in which LF(X) c-commands LF(Y)29 [+cc] S: a schematized PS that can correspond to LF(PS) in which LF(X) c-commands LF(Y)

28 See note 15. 29 The inclusion of the modal (can) in (30g) is due to the possibility that a single PS can correspond to more than one LF (and a single LF can correspond to more than one PS, for that matter), as extensively discussed in Ueyama 1998, in relation to Japanese.

132 

 Hajime Hoji

Let us review the discussion so far, in reference to (17). The adoption of the language faculty as our object of inquiry and the pursuit of the basic scientific method have led to the postulation of Merge and c-command. In light of this reasoning, we take c-command detection, more precisely put, establishing/ developing an effective means to detect c-command effects, as our most basic research goal in LFS. We seek to detect c-command effects (initially) by considering the relation between a PS containing X and Y and the possibility and the impossibility of MR(X, Y) in the PS via the hypothesized presence and the absence of the c-command relation between LF(X) and LF(Y) in LF(PS). We must therefore have hypotheses that determine how a PS containing X and Y is related to, i.e., is represented as, LF(PS), where LF(X) either c-commands or does not c-command LF(Y). We must also consider hypotheses that determine how the presence or the absence of the c-command relation in question is related to the possibility or impossibility of MR(X, Y). In relation to the Figure 5 in (17), let us use “MR(S, X, Y):J” to indicate an individual speaker’s Judgment on the availability of MR(X, Y) in a particular PS (such as a sentence).30 J has the value yes or no.31 (31) hence covers both (32a) and (32b). (31) MR(S, X, Y):J (32) a. MR(S, X, Y):yes b. MR(S, X, Y):no Let us use “S” in (31) and (32) to refer either to a particular phonetic sequence (PS) or a schematized representation of it (=a schema), like a “sentence pattern” as in (24) and (25), for example; used as a schema, it covers both instances of [-cc]S and instances of [+cc]S. The notations introduced above allow us to express any of (33).32

30 Notations such as (31) and (32) are adapted (with modification) from Plesniak 2022: 2.2.2. 31 The categorical nature of J (i.e., the speaker judgment on the availability of MR(X, Y) in a PS) is in line with our pursuit of the basic scientific method, as we have discussed earlier, and it will play an important role when we later address how we check our correlational (i.e., conditional) prediction and its contrapositive. 32 (33d) is included in (33) to exhaust the possibilities, but it is not discussed here because this case is not nearly as significant as the other three cases. It will be included when an experimental result is analyzed, as in Chapters 6 and 7 of this volume.

The Key Tenets of Language Faculty Science 

(33) a. b. c. d.

 133

MR([-cc]S, X, Y):no MR([+cc]S, X, Y):yes MR([-cc]S, X, Y):yes MR([+cc]S, X, Y):no

(33a) corresponds to “MR(X, Y) not possible” in (17), indicating an individual speaker’s judgment that MR(X, Y) is not possible in a sentence that instantiates the [-cc]S. (33b) corresponds to “MR(X, Y) possible” in (17), indicating an individual speaker’s judgment that MR(X, Y) is possible in a sentence that instantiates the [+cc] S. Under the hypothesis that MR(X, Y) requires that LF(X) c-command LF(Y), (33a) is a predicted judgment and (33b) is not unexpected.33 Under that hypothesis, (33c) is predicted not to arise. According to the preceding discussion, if we obtain both (33a) and (33b) from the same speaker at roughly the same time, this counts as c-command detection. As briefly discussed above, to determine whether a particular PS is [-cc]S or [+cc] S, i.e., whether LF(X) can c-command LF(Y) in LF(PS), we need hypotheses connecting PS and LF. Likewise, we need hypotheses connecting LF and MR. This is indicated in (34). (34) PS (phonetic sequence) containing X and Y

Hypotheses connecting LF and PS

LF(PS) Language Faculty Hypotheses connecting LF and MR If LF(X) c-commands LF(Y) If LF(X) does not c-command LF(Y)

MR(X, Y) not possible

MR(X, Y) possible

MR(X, Y)

Figure 6: Need for hypotheses connecting PS and LF and ones connecting MR and LF.

Since PS is I-language-particular, what connects PS and LF is (ultimately) an I-language-particular hypothesis. Likewise, hypotheses connecting MR(X, Y), with specific choices of X and Y, and LF are also I-language-particular because X and Y in MR(X, Y) are linguistic expressions in the PS (necessarily, of a particular I-language). Recall our basic view of the language faculty, as schematically indicated

33 (33b) is not predicted because, according to the hypothesis in question, LF(X) c-commanding LF(Y) is a necessary condition for MR(X, Y) but not a sufficient condition.

134 

 Hajime Hoji

in (2). Given this view, the speaker judgment as in (33a) (corresponding to “MR(X, Y) not possible” in (17)) or as in (33b) (corresponding to “MR(X, Y) possible” in (17))  – must be based in part on the initial state of the language faculty (i.e., universal properties), in addition to properties due to exposure to a linguistic environment (i.e., I-language-particular properties). When a language faculty scientist (=LFStist) tries to study contributions of the initial state of the language faculty in (2) by looking into his/her own I-language(s), s/he necessarily deals with properties specific to his/her own I-language(s). It may be useful (for a proper understanding of LFS) to imagine a countless number of instances of (2), reflecting different linguistic exposures different individuals have had, yet with the inner circle being invariant in all those instances. As noted, the inner circle in (2) is the LFStist’s primary concern. The methodology for LFS being articulated here and to be illustrated in subsequent chapters is thus designed to enable us to accumulate knowledge about the inner circle in (2), despite vast differences among individual linguistic experiences. This leads us to recognize that we must have CS(=Computational System)-internal, universal, hypotheses, in addition to hypotheses connecting LF and PS and those connecting LF and MR, as indicated in (35). (35)

PS (phonetic sequence) containing X and Y

Hypotheses that are CSinternal and universal

Hypotheses connecting LF and PS

LF(PS) Language Faculty Hypotheses connecting LF and MR

MR(X, Y) possible or not possible

MR(X, Y)

Figure 7: Three types of hypotheses.

As noted, the addition of “hypotheses that are CS-internal and universal” in (35) is crucial because our primary concern is with contributions of the initial state of the language faculty in (2), hence the universal aspects of the language faculty. The addition of these CS-internal (universal) hypotheses in (35) reminds us that what is most directly tested, when we pursue replication across (substantially) different I-languages, is the validity of CS-internal, universal, hypotheses, rather than those connecting PS and LF and those connecting LF and MR, with specific choices of X and Y, because the latter two types of hypotheses are I-language-particular.

The Key Tenets of Language Faculty Science 

 135

As noted, our most basic research goal in LFS is to establish/develop an effective means to detect c-command effects. For this purpose, we consider the CS-internal, universal, hypotheses in (36). (36)

FR(x, y) is possible only if x c-commands y.

FR(x, y) is a formal relation based on an LF representation,34 where, as previously discussed, x and y are two abstract “objects” in an LF representation. The deduction of predictions regarding speaker intuitions about the (im) possibility of MR(X, Y) thus requires, minimally, three types of hypotheses, as indicated in (37).35 (37)

Three types of hypotheses necessary for prediction-deduction in LFS: a. Hypotheses connecting PS (containing X and Y) and LF(PS): they have the consequence of specifying whether or not LF(X) c-commands LF(Y) in LF(PS) b. Hypotheses about what formal relation (FR) must be “established” based on the LF representation in order for MR(X, Y) to be possible c. CS-internal and universal hypotheses about FR: what conditions must be satisfied for an FR to be established

As noted, one type of PS is [+cc]S (for which there can be LF(PS) where LF(X) c-commands LF(Y)) and the other type is [-cc]S (for which there cannot be LF(PS) where LF(X) c-commands LF(Y)). While whether a given PS is [+cc]S or [-cc]S is determined by hypotheses connecting PS and LF, whether we predict (33a) is determined by hypotheses connecting LF and MR, which refer to the presence or the absence of FR(LF(X), LF(Y)), such as (38). (33) a. b.

MR([-cc]S, X, Y):no MR([+cc]S, X, Y):yes

34 FR abbreviates Formal Relation. 35 The formulation of (37b) is based the erroneous assumption that every instance of MR(X, Y) is based on FR. The recognition that such is not the case, i.e., there is non-FR-based-MR(X, Y) as well as FR-based-MR(X, Y), would lead us to change “MR(X, Y)” in (37b) to “FR-based-MR(X, Y)”, leading us to a seemingly tautological statement in (i), as opposed to (ii). (i)

FR-based-MR(X, Y) is possible only if there if FR(LF(X), LF(Y)).

(ii)

MR(X, Y) is possible only if there if FR(LF(X), LF(Y)).

How we can pursue rigorous testability under (i) will be addressed when we turn to the correlational methodology to be introduced shortly.

136 

 Hajime Hoji

(38) MR(X, Y) is possible only if there is FR(LF(X), LF(Y)). Notice that there may be conditions on FR(x, y) other than x c-commanding y, and in that case, what has to be checked is no longer just whether LF(X) c-commands LF(Y) but whether LF(PS) satisfies all the conditions on FR(x, y). Such a possibility has indeed been addressed in some depth in the case of the FR(x, y) hypothesized to underlie at least one type of MR(X, Y), mentioned in note 15, in works such as Hoji 1995, 1997a, 1997b, and 1998, mostly in relation to Japanese. Figure 7 in (35) is thus modified as in Figure 8 in (39).36 (39)

Hypotheses connecting LF and PS: specifying whether LF(X) ccommands LF(Y) in LF(PS)

PS (phonetic sequence) containing X and Y

Hypothesis that is CS-internal and universal: FR(x, y) only if: (i) x c-commands y (ii) ...

LF(PS) Language Faculty

If LF(PS) does not satisfy one of the conditions on FR(x, y)

Hypothesis connecting LF(PS) and FR-MR: FR-MR(X, Y) is possible only if FR(LF(X), LF(Y)) If LF(PS) satisfies all the conditions on FR(x, y)

FR-MR(X, Y) FR-MR(X, Y) not possible

FR-MR(X, Y) possible

Figure 8: Focusing on the possibility of FR-based-MR(X, Y).

Note further that we could maintain that the absence of FR(LF(X), LF(Y)) necessarily results in (33a) only if there are no sources of MR(X, Y) other than FR(LF(X), LF(Y)). Our basic view of the language faculty, as schematized in (2), however, suggests that we cannot rule out the possibility that there may be sources of MR(X, Y) other than FR(LF(X), LF(Y)). Such sources are in fact extensively discussed in Ueyama 1998: Appendix D.2, and also addressed in Chapters 2, 5, and 6 of this volume, and Plesniak 2022: 2.8. We are thus led to conclude that the MR(X, Y) whose possibility or impossibility we have discussed so far is in fact specifically

36 It is important to clearly distinguish the theoretical status of FR(x y) and the non-theoretical status of MR(X, Y). FR(x, y) is a purely theoretical concept hypothesized to hold between abstract objects x and y only when x c-commands y. MR(X, Y), on the other hand, is a meaning relation holding between two linguistic expressions X and Y. The properties of FR(x, y) is part of the object of inquiry in LFS, but those of MR(X, Y) is not. We consider properties of MR(X, Y) in hopes that we might obtain insight into properties of FR(x, y), as will be addressed further in the rest of this chapter and subsequent chapters.

The Key Tenets of Language Faculty Science 

 137

FR-based-MR(X, Y), rather than just “any type” of MR(X, Y). We can have intuitions about whether MR(X, Y) is possible or not in a given sentence, e.g., whether ‘Every engineer praised his robot’ allows the interpretation that each of the engineers in question praised his own robot. We cannot, however, have similar types of intuitions about whether the MR in question is FR-based or not. That can be determined only by our hypotheses about how our linguistic judgments come about and by experiments that test the validity of those hypotheses. It seems useful at this point to consider the model of judgment making, as schematized in (40), adapted from Ueyama 2010.37 (40)

A Hypothesis about what goes on in the mind of a speaker when s/he judges a sentence with regard to the availability of MR(X, Y) (adapted from Ueyama 2010) I Is this sentence acceptable under MR(X, Y)? Is MR(X, Y) possible in this sentence? Mental Lexicon

Does LF(PS) satisfy the condition(s) for FR(x, y)?

Check A set of items taken from the Mental Lexicon

MR(X, Y) Sentence (PS)

N

CS PF

LF

Judgment!

Compare

Figure 9: The Ueyama model of judgment making.

37 When the researcher checks his/her own intuitions about the possibility of MR(X, Y) in “self-experiment”, the researcher “constructs” sentences that he/she finds acceptable, and the only issue to consider is whether the MR(X, Y) in question is possible. In a “self-experiment”, therefore, we can focus on the possibility of the MR(X, Y). When a non-researcher’s intuitions are checked, on the other hand, it is not immediately obvious how we can differentiate “MR(X, Y) is not possible although the sentence is acceptable without the MR in question” from “The sentence is unacceptable and hence MR(X, Y) is not possible”. It is for this reason that the Figure in (40) gives two ways of asking the questions, i.e., “Is this sentence acceptable under MR(X, Y)?” and “Is MR(X, Y) possible in this sentence?” This has already been addressed to a degree in Chapter 3 of this volume. Subsequent chapters will continue to address it further. Chapter 5 of this volume discusses “self-experiment” where the researcher checks the availability of MR(X, Y). Chapter 6 of this volume discusses a demonstration in Japanese, where the non-researcher participants are asked about the acceptability of the sentence in question with MR(X, Y). Chapter 7 of this volume discusses a demonstration in English in which the participants are asked separately about the availability of MR(X, Y) and about the acceptability of the sentence in question.

138 

 Hajime Hoji

Embedded in (40) is Chomsky’s (1993) model of the Computational System (CS), as schematized in (41), which serves as a central conceptual basis for LFS, along with (40) and its elaborations to be given below. (41) A set of items taken from the Mental Lexicon

Computational System (CS)

LF (the mental representation serving as a basis of meaning)

PF (the mental representation serving as a basis of sounds/signs) Figure 10: Chomsky’s (1993) Model of the Computational System.38

According to (41), the CS takes as its input a set of items from the Mental Lexicon and generates two output representations; one is called an LF representation (serving as a basis of meaning, as discussed above) and the other is called a PF representation (serving as a basis of a phonetic sequence (PS)). The generation of the LF representation is achieved by recursive applications of Merge. A pair of PF and LF in (41) corresponds to the PS-LF pairing in our preceding exposition, which in turn corresponds to sound-meaning pairing in the Figure in (15). (41) is intended to state that the CS, given a particular input, does or does not generate a particular LF, and this categorical nature of (41) suits our pursuit of LFS by the basic scientific method.39 38 Choices among specific implementations of the leading idea behind (41), as they have been suggested and pursued in works subsequent to Chomsky 1993, are not consequential to the present discussion. According to the conception of the CS as discussed above, the iterative application of Merge yields an LF representation that is to be “externalized” to a PF/PS and to serve as the basis of meaning, and in that sense, it differs from (41), where LF refers specifically to the representation that serves only as the basis of meaning. In the spirit of LFS pursued here, specific proposals will have to be evaluated by the basic scientific method, and among the crucial questions to be raised is whether they make definite and testable predictions. 39 We do not address exactly how LF(PS) is constructed based on what set of lexical items taken from the Mental Lexicon or exactly how LF(SP) is “externalized” as the corresponding PF in (41). At the moment, we are only concerned with hypotheses that lead to rigorously testable predictions. The relevant operations most likely make reference to the “head-complement” relation, the “subcategorizational” properties of each verb, and the “linearization” algorithm based on something like Fukui and Takano’s (1998) “DeMerge”. Such issues, suppressed here, are pursued in works by Ueyama, including Ueyama 2015. It is hoped that successful establishment of an

The Key Tenets of Language Faculty Science 

 139

(40) is a general hypothesis about what goes on in the mind of a speaker when he/she judges a PS containing two linguistic expressions X and Y with regard to the possibility of MR(X, Y), and it is understood to be a special and simplified model of how a speaker comprehends a sentence. A fuller version of that process is schematized in (42). (42)

(perceived) phonetic strings frequent patterns Numeration Extractor

Lexicon

(formal) features numeration Computational System

PF

Phonology

(generated) phonetic strings

LF Information Extractor SR Information Database

Concepts Working Space

Inference rules

input/output process system

influence

reference (dynamic) database

Figure 11: The Ueyama 2015 model of sentence comprehension.40

(42) is a specific elaboration of our general conception of the language faculty serving the purpose of placing (40) in a broader context, although it does not specify “where” discourse-related information is managed or represented. Such considerations seem crucial when we turn to MR(S, X, Y):yes judgments that is based on something other effective means to detect c-command effects in line with the basic scientific method will aid such research endeavors. 40 This is based on a draft version of Ueyama 2015. Ueyama 2015 presents a simplified version of (42) and its “sentence production” counterpart.

140 

 Hajime Hoji

than FR(LF(X), LF(Y)). Instances of MR(S, X, Y) of this type will be referred to here as “non-FR-(based-)MR(S, X, Y), as opposed to FR-(based-)MR(S, X, Y). The most basic aspect of what is intended by (40) in the context of (42) is as follows; when asked about the availability of MR(X, Y) in a sentence containing X and Y, the cognitive agent, whether it is the researcher him/herself (checking his/her own judgments) or someone else who the researcher is asking to judge the sentence, comes up with a set of items from his/her Mental Lexicon and the CS takes the set as its input and generates two output representations, LF and PF representations. Focusing on LF, a semantic representation (SR) obtains based on the LF representation via further mapping operations,41 and the availability of MR(X, Y) in the sentence is based on the SR, influenced by the “world knowledge” stored in the “Information database”, for example. An individual speaker’s judgment (yes or no) on (the availability of) MR(S, X, Y) can thus be affected not only by whether LF(X) c-commands LF(Y), but also by (potentially numerous) other factors, including how discourse-related information is managed or represented. In light of this, we modify (40), as in (43). (43)

A Hypothesis about what goes on in the mind of the participant when they judge a sentence (adapted from Ueyama 2010) II Is this sentence acceptable under MR(X, Y)? Is MR(X, Y) possible in this sentence?

Mental Lexicon

Do the SR and related cognitive representations, combined, satisfy the condition(s) for non-FRbased MR(X, Y)?

A set of items taken from the Mental Lexicon

MR(X, Y) Sentence (PS)

N

CS

LF

SR

PF

Judgment! PS

Compare

Figure 12: Model of judgment making, focusing on non-FR-based-MR(X, Y).

41 Ueyama 2015 is an attempt to spell out operations that map LF to SR. See the first

paragraph of note 39.

The Key Tenets of Language Faculty Science 

 141

(40), given earlier, is a model of judgment making, focusing on FR-based-MR(X, Y). As such, it is “applicable” only if there are no sources for MR(X, Y) other than FR or if the effects of such possible sources have been properly controlled for. (43), on the other hand, is the model of judgment making, focusing on possible sources of MR(X, Y) other than FR(x, y). As such, it should be understood that (40) and (43) combined constitute a more complete model of judgment making. The two Figures in (40) and (43) are thus intended to be seen “on top of each other”, so to speak, with (43) on top of (40). According to our hypothesis about what goes on in the mind of a speaker when s/he judges the availability of MR(X, Y) for a given PS, what s/he checks includes, in principle, whether the condition(s) for FR(LF(X), LF(Y)) is/are satisfied in LF(PS) and whether non-FR-sources are available for the MR(X, Y) in question. (40) concerns the former and (43) concerns the latter. The speaker’s MR(S, X, Y):J can, in principle, arise based on what goes on in (40) and/or what goes on in (43). Because we are concerned with c-command detection in LFS, for the reasons discussed in Section 1, we care most about what goes on in (40). We can, however, only check what goes on in (40) through the cloudy world of (43), so to speak. In other words, we can obtain c-command detection (by the basic scientific method) only if we can manage to reduce the effects of non-FR-based MR(X, Y) to zero, in some way. The main tenets of LFS include the thesis that we can indeed do so and the remainder of this chapter is concerned with conceptual articulation about how we can do that. As indicated above, whether LF(PS) satisfies the requirement on FR(x, y) that LF(X) c-command LF(Y), i.e., whether the PS in question is [+cc]S or [-cc]S, is a function of our PS-LF correspondence hypotheses, such as the one discussed in relation to (14) above. Similar remarks apply to whether LF(PS) satisfies other conditions on FR(x, y) insofar as those conditions are stated in terms of properties of the CS, hence as universal properties of the language faculty. For a given PS containing X and Y, it is therefore possible to determine, by inspecting the PS, whether it is possible for LF(PS) to satisfy the condition(s) on FR(x, y); this is what (40) “addresses”. Whether a PS containing X and Y allows MR(X, Y), on the other hand, involves a great deal more, as it involves not only (40) but also (43). It is in fact not possible to formally determine whether a given PS allows MR(X, Y) for an individual just by inspection of the PS itself precisely because there may be  possible sources of MR(X, Y) other than FR(x, y), as discussed in Chapter 2 of this volume, and Chapter 5: Section 8 of this volume; see also Plesniak 2022: 2.8. To illustrate, MR(S, X, Y):no arises if (i) LF(PS) does not satisfy one (or more) of the condition(s) for FR(x, y) and (ii) no non-FR-sources for the MR(X, Y) in question are available in the SR based on LF(PS). MR(S, X, Y):yes, on the other hand, arises if either (i) LF(PS) satisfies all the condition(s) for FR(x, y) or (ii) a non-FRsource is available for the MR(X, Y) in question. Recall that “S”, under one of its

142 

 Hajime Hoji

two uses, represents a schematized PS, covering both [-cc]S and [+cc]S.42 We have noted that [+cc]S may not necessarily “license” FR-based-MR(X, Y) because there could be conditions on a particular FR(x, y) other than the c-command condition in (36), repeated here. (36) FR(x, y) is possible only if x c-commands y. Let us use “FR-✶S” to refer to an S that does not satisfy at least one of the conditions for FR(x, y).43 Likewise, we can use “FR-okS” to refer to an S that satisfies all the conditions for FR(x, y), including the c-command condition in (36).44 By using “FR-✶S” and “FR-okS” in place of S in MR(S, X, Y), we can discuss how J=yes or no obtains for a given MR(S, X, Y):J more effectively than before. Consider the four possibilities in (44). (44) a. b. c. d.

MR(FR-✶S, X, Y):no MR(FR-okS, X, Y):yes MR(FR-✶S, X, Y):yes MR(FR-okS, X, Y):no

If it has been ensured that there are no sources of MR(X, Y) available other than FR, we predict (44a), and our prediction would be disconfirmed by (44c).45 If there are effects of sources of non-FR-based-MR(X, Y), however, that can result

42 Under the other use, it represents an actual sentence instantiating such a schematized PS. 43 In order for FR-MR(S, X, Y) to arise, the S must be [+cc]S, hence [-cc]S is necessarily a “FR-✶S”. 44 “✶S” and “okS” here are notational variants of what Hoji 2015: Chapter 2 called “✶Schema” and “okSchema”. 45 According to the “fundamental schematic asymmetry” between “✶Schema-based prediction” and “okSchema-based prediction” in the terms of Hoji 2015, only (44a), but not (44b), would be a disconfirmable prediction, hence (44c), but not (44d), would constitute disconfirmation of our prediction.

The Key Tenets of Language Faculty Science 

 143

in (44c). MR(S, X, Y):yes can therefore arise under any of the conditions specified in (45a, b, c).46, 47 (45)

MR(S, X, Y):yes can arise under any of the conditions in: a. LF(PS) satisfies all the condition(s) for the relevant FR(LF(X), LF(Y)) and no non-FR-sources are available for the MR(X, Y). b. LF(PS) does not satisfy one (or more) of the condition(s) for the relevant FR(x, y) but a non-FR-source is available for the MR(X, Y). c.

LF(PS) satisfies all the condition(s) for the relevant FR(LF(X), LF(Y)) and a non-FR-source is available for the MR(X, Y).

(45a) and (45c) can result in (44b), and (45b) in (44c). MR(S, X, Y):no, on the other hand, can arise only if LF(PS) does not satisfy one (or more) of the condition(s) for FR(LF(X), LF(Y)) and no non-FR-sources are available for the MR(X, Y), resulting in (44a), leaving aside the cases where FR(LF(X), LF(Y)) was in principle available to the individual making the judgment but they failed to apply it for some reason.48 The chart in (46) summarizes this state of affairs.49 (46) Predicted judgments on MR(S, X, Y) Effects of sources for MR(X, Y) other than FR(LF(X), LF(Y)) absent present FR-✶S no (44a) yes (44c) FR-okS yes (44b) yes (44b)

46 Here and elsewhere, issues are suppressed regarding cases of J=yes (and for that matter, cases of J=no as well) arising due to errors, inattentiveness, etc. The recognition of such issues leads us to the conception of J of “MR(S, X, Y):J” such that it is not just a result of a single act of judgment making, but based on multiple acts of judgment making, perhaps even involving different “subtypes” of [-cc]S or [+cc]S in question, and, in the case of a non-researcher judging sentences, also based on results of “preliminary” experiments in which we test the participant’s attentiveness, for example, as will be addressed in Chapters 6 and 8 of this volume. 47 For some instances of “MR(S, X, Y):yes”, it is not straightforward to determine which of (45a, b, c) is its source, so to speak, and the relevant determination, when possible, constitutes substantive progress in our understanding of the properties of the language faculty, as will be discussed in Chapter 5. 48 When we turn to a given experiment, we try to identify such cases of (44d) by having multiple instances of a given schema judged, ensuring that MR(FR-okS, X, Y) is not being blocked by some “incidental factor”, but that it consistently fails to obtain for that individual at that time with that particular S; see note 46. If it does not consistently fail to obtain, that would force us to reexamine our hypotheses. 49 What is meant by “yes” in (46) is that the J can be yes; see note 45.

144 

 Hajime Hoji

We see that we can consider yes/no as revealing about c-command effects only if it is ensured that effects of the sources of non-FR-based-MR(X, Y) are absent. If MR(X, Y) is available when at least one of the conditions on FR(LF(X), LF(Y)) is not satisfied, as in (44c), it may be called Quirky-MR, as in in Ueyama 1998 and Chapter 2 of this volume.50 The correlational methodology to be discussed in the next section is an attempt to control for effects that would lead to non-FR-basedMR(X, Y) so that we could obtain (44a), as predicted.

5 The Correlational Methodology To proceed further with our discussion about c-command detection, I shall now make the simplification that FR-✶S and FR-okS are equivalent to [-cc]S and [+cc]S, respectively, i.e., I will proceed with the assumption that the only condition on FR(x, y) is the c-command condition as in (36).51 Under this simplification, c-command detection based on MR(X, Y) obtains only if we get both (47a) and (47b). (47) C-command detection based on MR(X, Y) obtains only if we get both (a) and (b). a. MR([-cc]S, X, Y):no b. MR([+cc]S, X, Y):yes In light of the fact that the condition(s) on FR(x, y)52 are stated as necessary condition(s), and not as sufficient and necessary condition(s), the prediction about J=yes in MR([+cc]S, X, Y):J has a radically different status from the prediction about J=no in MR([-cc]S, X, Y):J. The former is, in principle, not a disconfirmable prediction because the necessary condition(s) for FR(x, y) to be satisfied do not

50 We are using “MR” as a general term covering specific MRs discussed in Ueyama 1998 and Chapter 2 of this volume. Although discussion of “Quirky-MR” in Ueyama 1998: Appendix D.2 and Chapter 2 of this volume is (mostly) limited to FR-✶S, as in (44c), sources for non-FR-basedMR(X, Y) can also be a basis for J=yes in FR-okS, as in (44b). It is in fact indicated in (46) that the yes judgment can be based on FR or something other than FR, or even both. Further discussion requires considerations of the sort addressed in Chapter 5 of this volume. 51 If we did not make this simplification, our discussion about what counts as c-command detection would be significantly more complicated. With discussion of the relevant complication, however, comes additional empirical coverage and theoretical understanding, as well as an enhancement of our experimental tool, which will be discussed in Chapter 5 of this volume. 52 With the simplification just made, we are focusing just on the c-command condition (x c-commanding y).

The Key Tenets of Language Faculty Science 

 145

guarantee J=yes; see Hoji 2015: 2.4 and Chapter 3 of this volume for relevant discussion. Its confirmation, however, is significant for c-command detection with MR(X, Y) because c-command detection based on MR(S, X, Y):J obtains only if we get both (47a) and (47b), as discussed in some depth in Hoji 2015: 2.4.53 In short, (47a) obtaining, by itself, does not constitute c-command detection with MR(X, Y), and it is necessary for both (47a) and (47b) to obtain at the same time in order to constitute c-command detection with MR(X, Y). The relevant factors that affect the J (yes or no by an individual) in (48) include, among others, the choices of MR, S, X, and Y, as indicated in (49).5455 (48) MR(S, X, Y):J (49) a. MR: what MR is considered54 b. S: whether it is [-cc]S or [+cc]S55

53 It is MR([+cc]S, X, Y):yes that gives MR([-cc]S, X, Y):no (full) “significance” in terms of demonstrating that it arises from the absence of the requisite c-command relation rather than from some factors unrelated to c-command (such as the complexity or the unnaturalness of the sentence in question). 54 This includes how we understand (i.e., “characterize” or “define”) the MR(X, Y) in question, which affects how we convey it to participants in an experiment, as will be addressed in Chapter 6–8 of this volume. As noted, we consider the kind of meaning relation (MR) such that its availability may lead to c-command detection, and which MR we consider leads to general constraints as to what can be X and what can be Y of that particular MR(X, Y). For any type of MR we might consider, it gives constraints as to what can possibly be X and what can possibly be Y. It should be stressed that MR is an observational, not a theoretical, concept; we use whatever MR, with whatever characterization in hopes that the use of such an MR, with its characterization, leads to c-command detection in terms of yes and no for the J of MR(S, X, Y):J. If we consider MR(X, Y) such that the number of entities expressed by Y “is dependent” upon the number of entities expressed by X, for example, both X and Y must express some quantity in the relevant sense. Suppose furthermore that included in the characterization of the MR(X, Y) is that “the number of Y” for “each of X” is two or more. ‘Three girls’ and ‘two boys’, for example, can serve as X and/or Y in that case, but ‘that boy’ or ‘John’ cannot. That is to say, regarding a sentence containing ‘three girls’ and ‘two boys’, we can ask whether MR(three girls, two boy) is possible, with the MR understood in the above-mentioned sense. (MR(three girls, two boys) just addressed is referred to as “DR(three girls, two boys)” in Hoji 2003, Hayashishita 2004, among other places; see Chapter 5: Section 2 for more details. We cannot, however ask whether MR(three girls, that boy) is or is not possible, with the MR understood in the same way. 55 Here, we are suppressing other aspects of S that might be relevant, for the purposes of a straightforward discussion.

146 

 Hajime Hoji

c. X: what is used as X56 d. Y: what is used as Y If the choice between [-cc]S and [+cc]S (more accurately, between FR-✶S and FR-okS) were the only factor affecting the J of MR(S, X, Y):J, we would necessarily obtain MR([-cc]S, X, Y):no and could, in principle, obtain MR([+cc]S, X, Y):yes. Such, however, is not the case, as noted above in relation to the possibility of nonFR-based-MR(X, Y). For example, what is used as X and what is used as Y can induce non-FR-based-MR(X, Y), i.e., certain choices of X and Y can lead to MR(X, Y) without LF(X) c-commanding LF(Y), but what choices do so vary among speakers and even within a speaker at different times.567 Nevertheless, the correlational methodology to be presented makes it possible to obtain and replicate c-command detection, despite this seemingly unsurmountable challenge. In order for MR(X, Y) to serve as a basis for c-command detection, or more precisely, in order for the judgments as indicated in (47) to count as c-command detection, it is necessary to identify a choice of X and a choice of Y that do not induce non-FR-based-MR(X, Y) for a given speaker at a given time, and the relevant identification must be independent of the MR(X, Y) in question. A reasonable way to identify such choices would rely on two other types of MR(X, Y) such that one type helps us identify a choice of X that does not induce non-FR-based-MR(X, Y) and the other type a choice of Y that does not induce non-FR-based-MR(X, Y) for a given speaker at a given time. To differentiate the general concept of MR

56 The possibility of BVA(X, Y), as introduced in footnote 15, for example, can be affected by what is used as X and what is used as Y in (i) (among other factors), with (ii-a) and (ii-b) being specific instantiations of (i). (i)

Y’s N V X

(ii)

a. b.

His robot praised every engineer. His robot praised at least one engineer.

Variable effects of choices of X and Y, and other related issues will be discussed in some depth in Chapter 5 of this volume, with respect to a single native speaker of Japanese (myself), and in Chapters 6 and 7 of this volume, with respect to a number of native speakers of Japanese (Chapter 6) and a number of native speakers of English (Chapter 7). 57 There are other factors other than the choices of X and Y that affect the J of MR(S, X, Y):J. One notable factor is how the S containing the X and Y in question is embedded in a larger syntactic or discoursal context, as discussed in Ueyama 1998: Appendix D, and in Chapter 5 of this volume; see also Plesniak 2022: 2.2.2. How a combination of X and Y, not the choice of X and that of Y in isolation, might affect the J of MR(S, X, Y) is another factor, not addressed in this chapter but will be discussed in Chapter 5 of this volume.

The Key Tenets of Language Faculty Science 

 147

from its specific instances, I will now refer to specific instances of MR as MR1, MR2, MR3, etc. while using “MR” itself to refer to the general concept of MR. We can now address three types of MR(X, Y) as in (50) and speaker judgments on each of (50) as in (51). (50) a. MR1(X, Y) b. MR2(X, beta) c. MR3(alpha, Y) (51)

a. MR1(S, X, Y):J b. MR2(S, X, beta):J c. MR3(S, alpha, Y):J

“beta” is meant to be an instance of Y and “alpha” an instance of X, with the understanding that what can serve as X of MR1 perhaps may not be able to serve as X of MR3 (hence we call the latter alpha), and what can serve as Y of MR1 perhaps may not be able to serve as Y of MR2 (hence we call the latter beta). On the other hand, MR2 is chosen so that it does share X with MR1, and likewise MR3 is chosen so that it does share Y with MR1. We seek to attain c-command detection with MR1 by making reference to MR2 and MR3 such that MR2 shares the same X as MR1(X, Y), but not the same Y, as indicated in (50b) and (51b), and MR3 shares the same Y as MR1(X, Y), but not the same X, as indicated in (50c) and (51c).58

58 The specific instances of MR1, MR2, and MR3 that we deal with in subsequent chapters of this volume are “BVA(X, Y)”, “DR(X, Y)”, and “Coref(X, Y)”, respectively. BVA(X, Y) is mentioned briefly in notes 15 and 56; the J in BVA([-cc]S, every engineer, his):J is about the possibility of the interpretation for (i), as given in (ii). (i)

His robot praised every boy.

(ii)

Each of the boys in question was praised by his own robot.

DR(X, Y) is briefly mentioned in footnote 54. The J in DR([-cc]S, three girls, two boys):J is about the possibility of the interpretation for (ii), as given in (iv). (iii)

Two boys praised three girls.

(iv)

Each of the three girls in question was praised by a different set of two boys.

The meaning relation of Coref(X, Y) is that of “coreference”. The J in Coref([-cc]S, John, his):J is about the possibility of the interpretation for ‘His robot praised John’, where ‘John’ and ‘his’ are meant to refer to the same individual. These three MRs will be addressed extensively in Chapters 3, 5–7 of this volume.

148 

 Hajime Hoji

We attain c-command detection with MR1(X, Y) based on the identification of X that leads to a c-command pattern with MR2(X, beta) and the identification of Y that leads to a c-command pattern with MR3(alpha, Y) for a given speaker at a given time, under the assumptions that (i) a given choice of X has the same effects on the J of MR2(S, X, beta):J as it does on the J of MR1(S, X, Y) (with the S being identical in terms of whether it is [-cc]S or [+cc]S), (ii) a given choice of Y has the same effects on the J of MR3(S, X, beta):J as it does on the J of MR1(S, X, Y) (with the S being identical in terms of whether it is [-cc]S or [+cc]S), (iii) the choice of alpha does not affect the J of MR2(S, X, beta), and (iv) the choice of beta does not affect the J of MR3(S, alpha, Y), as the Figure in (52) is intended to indicate. (52)

The same effects of X: The choice of X has the same effects on the J of MR2(S, X, beta):J and MR1(S, X, Y):J.

MR2(S, X, beta):J MR3(S, alpha, Y):J MR1(S, X, Y):J The same effects of Y: The choice of Y has the same effects on the J of MR3(S, alpha, Y):J and MR1(S, X, Y):J. Figure 13: Correlations among MR2(S, X, beta):J, MR3(S, alpha, Y):J, and MR1(S, X, Y):J, under the assumption that there are no effects of a particular choice of alpha/beta.

Seeking the existence of MR such that MR([-cc]S, X, Y):no and MR([+cc]S, X, Y):yes obtain as a consistent pattern is a conceptual consequence of having the language faculty as our object of inquiry and adopting the basic scientific method as our method of inquiry. If it were not possible to identify such MR(S, X, Y) (for a given speaker at a given time), LFS by the basic scientific method, focusing on c-command detection, would not be possible. In order for such MR(S, X, Y) to constitute c-command detection in the spirit of the basic scientific method, it must be accompanied by rigorous testability, i.e., it must be based on a disconfirmable prediction. The claim about the existence of MR such that MR([-cc]S, X, Y):no and MR([+cc]S, X, Y):yes obtain cannot be disconfirmed empirically. Disconfirmable predictions can be made only about MR([-cc]S, X, Y):no provided that the choice of X and the choice of Y do not induce non-FR-based-MR(X, Y). The identification of such choices of X and Y, independently of the MR(X, Y) in question is thus crucial. This then leads us to seek c-command detection with MR1(X, Y) based on the identification of such a choice of X (determined using MR2) and that of such a choice of Y (determined using MR3). The key hypothesis is that the choice of X has the same effects on the J of MR2(S, X, beta):J and the J of MR1(S, X, Y) and that the

The Key Tenets of Language Faculty Science 

 149

choice of Y has the same effects on the J of MR3(S, alpha, Y):J and the J of MR1(S, X, Y). It is this hypothesis that makes it possible to predict and obtain c-command detection with MR1(S, X, Y) in a rigorously testable manner. Adopting the correlational methodology for c-command detection is thus a consequence of having the language faculty as our object of inquiry and adopting the basic scientific method as our method of inquiry. In line with (49), the factors that affect the J of each of (51a, b, c) are: (i) MR1, MR2, and MR3, (ii) whether the S is [-cc]S or [+cc]S, (iii) what is used as X, (iv) what is used as Y, (v) what is used as beta, and (vi) what is used as alpha. In line with the preceding discussion, along with the simplifications we have made, the J in MR2(S, X, beta):J depends in part upon the specific choices of X and beta; likewise, the J in MR3(S, alpha, Y):J depends in part upon the specific choices of alpha and Y. For the purpose of the present illustration of the correlational methodology, however, I make the further simplification that specific choices of beta do not affect the J of MR2(S, X, beta):J and specific choices of alpha do not affect the J of MR3(S, alpha, Y):J. Under these simplifications, if (53) obtains for a given speaker at a given time, that indicates that the specific choice of X does not give rise to non-FR effects for MR2(S, X, beta):J for that speaker at that time.59 (53)

MR2([-cc]S, X, beta):no

Likewise, if (54) obtains for a given speaker at a given time, that indicates that the specific choice of Y does not give rise to non-FR effects on the J of MR3(S, alpha, Y):J for that speaker at that time. (54) MR3([-cc]S, alpha, Y):no So far, we discussed each of (49).60 (49) a. b. c. d.

MR: what MR is considered60 S: whether it is [-cc]S or [+cc]S X: what is used as X Y: what is used as Y

59 Chapter 5 and 6 of this volume addresses consequences of eliminating these simplifications, providing further illustration of the correlational methodology proposed here. 60 See note 54.

150 

 Hajime Hoji

There are other factors affecting the J of MR(S, X, Y):J, especially, but not limited to, those inducing non-FR-based-MR(S, X, Y):yes, but they are suppressed in this chapter.61 To make this introduction to the essentials of the correlational methodology manageable, we are making the simplification that the only factors that affect the J of MR(S, X, Y):J are whether the S is [-cc]S or [+cc]S, the choice of the X and that of Y.62 Under this simplification, for a given individual at a given time, the J of MR(S, X, Y):J is the function of the choice between [-cc]S and [+cc]S, the choice of X, and that of Y.63 All these factors (and simplifications) considered, we arrive at a testable correlational prediction, which can be summarized as in (55), which states that if the MR2 test determines that X can participate in FR-MR but cannot induce non-FR MR and the MR3 test determines that Y can participate in FR-MR but cannot induce non-FR MR, then we predict that MR1(S, X, Y) can only be available when FR(LF(X), LF(Y)) is possible, i.e., for S such that LF(X) c-commands LF(Y) in LF(S). (55)

Testable correlational/conditional prediction: MR2([+cc]S, X, beta):yes ∧ MR2([-cc]S, X, beta):no ∧ MR3([+cc]S, alpha, Y):yes ∧ MR3([-cc]S, alpha, Y):no → MR1([-cc]S, X, Y):no

According to (55), a c-command pattern with MR2(X, beta), i.e., a combination of MR2([+cc]S, X, beta):yes and MR2([-cc]S, X, beta):no, and one with MR3(alpha, Y), i.e., a combination of MR3([+cc]S, alpha, Y):yes and MR2([-cc]S, alpha, Y):no, entail MR1([-cc]S, X, Y):no. However, MR1([-cc]S, X, Y):no (the consequent in (55)) does not by itself count as c-command detection, because the J=no can be due to factors other than c-command. To obtain full correlational c-command detection, we also need the MR1([+cc]S, X, Y):yes, for the reason noted at the beginning of this section. We thus modify (55) as in (56), adding MR1([+cc]S, X, Y):yes to the antecedent. (56) Testable correlational/conditional prediction about c-command-detection: MR2([+cc]S, X, beta):yes ∧ MR2([-cc]S, X, beta):no ∧ MR3([+cc]S, alpha, Y):yes ∧ MR3([-cc]S, alpha, Y):no ∧ MR1([+cc]S, X, Y):yes → MR1([-cc]S, X, Y):no

Not only is (56) disconfirmable, and hence rigorously testable, but its confirmation would the desired c-command detection with MR1(X, Y), with MR1([-cc]S, X, Y):no 61 They will be addressed in Chapter 5 of this volume. 62 Under this simplification, the choice between ‘every engineer’ and ‘at least one engineer’ for X is considered to be relevant for variable effects on the J of MR(S, X, Y):J, but not the choice between ‘engineer’ and ‘father’ as they appear in ‘every engineer’ and ‘every father’. The choice of the latter kind and other factors will be addressed in some depth in Chapters 5 and 6 of this volume. 63 See note 57.

The Key Tenets of Language Faculty Science 

 151

in the consequent of the conditional now being accompanied by the newly added MR1([+cc]S, X, Y):yes in the antecedent clause. We will be testing the prediction in (56) in Chapters 5–7 of this volume. We can test it by ensuring all the conditions in the antecedent clause of the conditional are met, that is, by using the X and the Y that have been identified, by the MR2 test and the MR3 test, as not inducing non-FR-based-MR2 or non-FR-based-MR3. We then check to see if, as predicted, the J on MR1([-cc]S, X, Y):J is necessarily no. In those chapters, we will also be checking what happens if MR1([-cc]S, X, Y):yes obtains, which we predict will happen only when the conditions in the antecedent clause are not met. That is, MR1([-cc]S, X, Y):yes is possible only as an instance of non-FRbased-MR1, hence only in cases when X or Y (or both) induce non-FR-based-MR(X, Y); this, in turn, is only predicted to occur when non-FR-based-MR2(X, beta) or non-FRMR3(alpha, Y) or both are possible. We test this secondary prediction in Chapters 5–7 by checking the contrapositive of (56). In anticipation of discussion about the contrapositive of (56), let us restate (56) as in (57).64 (57) Testable correlational/conditional prediction about c-command-detection: Provided that MR2([+cc]S, X, beta):yes ∧ MR3([+cc]S, alpha, Y):yes ∧ MR1([+cc]S, X, Y):yes; MR2([-cc]S, X, beta):no ∧ MR3([-cc]S, alpha, Y):no → MR1([-cc]S, X, Y):no

The form of the conditional in (57) (P → Q) is equivalent to its contrapositive (not Q → not P); hence (58) is equivalent to (59a), which is, by De Morgan’s Law, equivalent to (59b), which in turn is equivalent to (59c). (58) MR2([-cc]S, X, beta):no ∧ MR3([-cc]S, alpha, Y):no → MR1([-cc]S, X, Y):no (59) a.

not MR1([-cc]S, X, Y):no → not [MR2([-cc]S, X, beta):no ∧ MR3([-cc]S, alpha, Y):no]

b. not MR1([-cc]S, X, Y):no → not MR2([-cc]S, X, beta):no ∨ not MR3([-cc]S, alpha, Y):no

c.

MR1([-cc]S, X, Y):yes → MR2([-cc]S, X, beta):yes ∨ MR3([-cc]S, alpha, Y):yes

64 In case one is confused about slight changes in the way I have worded the implication, note that (i) is equivalent to (ii) as well as to (iii), for example. (i) P∧Q→R (ii) Provided P, Q → R (iii) P ∧ Q → P ∧ Q ∧ R

152 

 Hajime Hoji

(57) is thus equivalent to (60).65 (60) The contrapositive of (57):65 Provided that MR2([+cc]S, X, beta):yes ∧ MR3([+cc]S, alpha, Y):yes ∧ MR1([+cc] S, X, Y):yes; MR1([-cc]S, X, Y):yes → MR2([-cc]S, X, beta):yes ∨ MR3([-cc]S, alpha, Y):yes

While (57) is a testable correlational/conditional prediction about c-command detection, (60) is a testable correlational/conditional prediction about the presence of the sources behind non-FR1-MR1, non-FR2-MR2, and non-FR3-MR3. Chapter 5 of this volume discusses how (57) and (60) have both been confirmed in my self-experiment, with every combination of choices of X and Y of MR(X, Y) considered there, with three type of MRs (see footnotes 15, 54, and 59), and with various forms of S’s. Chapter 6 of this volume discusses a “large-scale” non-self-experiment containing a small subset of the choices of X and Y, and S’s used in my self-experiment, and reports that the correlational/conditional prediction about c-command detection in (57) and its contrapositive in (60) have both been confirmed, replicating the result of my self-experiment. Chapter 7 of this volume discusses how further replication has been attainted with experiments involving (hundreds of) native speakers of English.

6 Summary and Conclusion 6.1 Summary The aim of this chapter is to articulate how we can accumulate knowledge about contributions of the initial state of the language faculty to our linguistic intuitions by the basic scientific method. The relevant features of the basic scientific method are that we deduce definite and testable predictions from hypotheses, obtain and replicate definite experimental results in line with such predictions, which is done while stating our hypotheses in terms of a minimal number of theoretical concepts. The conceptual articulation in question has thus taken the form of providing an answer to each of (3a–c), repeated here.

65 The significance of checking the contrapositive of our correlational prediction was brought to my attention by Daniel Plesniak (p.c., December 2019).

The Key Tenets of Language Faculty Science 

 153

(3) a.

How we can deduce definite predictions about an individual’s linguistic intuitions regarding the relation between linguistic sounds and meaning b. How we can obtain experimental results pertaining to an individual precisely in line with such definite predictions c. How we can replicate the experimental results alluded to in (3b)

In relation to (3a), the deduction of definite predictions about an individual’s linguistic intuitions about the relation between linguistic sounds and meaning crucially relies on what theoretical concepts we can make use of in formulating our hypotheses. For this, we have turned to Chomsky’s proposal that the language faculty contains a computational system (CS), the sole structure-building operation of which is Merge, which forms a set based on two items. This is taken along with Chomsky’s claim that the abstract representations created by recursive application of Merge serve as a basis for our linguistic intuitions about those aspects of meaning that are directly based on CS operations. We have then adopted the abstract structural relation “c-command”, defined based on the recursive application of Merge, as given in (7), repeated here. (Section 1).66 (7) x c-commands y if and only if x is merged with something that contains y66 For a given pair of objects, x and y, in the abstract representation created by recursive application of Merge, either x c-commands y or it does not, and we want to capitalize on this categorical distinction in our efforts to accumulate knowledge about the contributions of the initial state of the language faculty to our linguistic intuitions, by the basic scientific method. In order to achieve what is alluded to in (3a), we need the presence or the absence of a c-command relation to be related to certain aspects of our linguistic intuitions. In relation to the concept of the I-language (a steady state of the language faculty), as schematized in (2), the primary concern of language faculty science (LFS) is the aforementioned contributions of the initial state of the language faculty to our linguistic intuitions. In order to find out about them, we must plow our way through the “cloud” of various effects due to exposure to a specific “language”. In other words, in order to find out about contributions of the initial state of the language faculty (universal properties of the language faculty) to our linguistic intuitions, we must consider properties due to exposure to a particular linguistic environment (I-language-specific properties). More specifically, hypothesizing how what is observ-

66 See note 6.

154 

 Hajime Hoji

able is related to what is not observable is a key part of our answer to (3a). For a conceptual articulation of this, we differentiated two distinct types of concepts; one for what is observable and the other for what is abstract and not directly observable. The former includes: two linguistic expressions X and Y in a sentence S and certain meaning relations pertaining to X and Y (MR(X, Y)); the latter includes: LF(X), LF(Y), and LF(S), the abstract objects in the abstract representation corresponding to X, Y, and S, respectively. We also introduced hypotheses such as one stating that LF(S), where the S in Japanese is of the pattern in (11), is represented as in (14), with regard to the c-command relation between LF(NP1) and LF(NP2). (Section 2) (11) is repeated here. (11) NP1-ga NP2-o V In addition to concepts for observable linguistic expressions and those for abstract objects in the abstract representation and hypotheses that relate them, we introduced CS-internal hypotheses regarding a certain type of abstract relation (which we called an FR (formal relation)) pertaining to two abstract objects, x and y. A hypothesis such as (36), repeated here, is indeed about contributions of the initial state of the language faculty in (2), and, hence, it is the most important type of hypothesis in LFS because it plays the most critical role in our attempt to establish a reliable means to detect c-command effects by the basic scientific method. (36) FR(x, y) is possible only if x c-commands y. Ueyama’s (2010) model of judgment making and Chomsky’s (1995) model of the CS were introduced as conceptual bases for our discussion, aimed at c-command detection, as in Hoji 2015. (Section 4) It is the recognition of the existence of non-FR-based-MR that has led us to adopt the correlational methodology, as in (57), repeated here. (Section 5) (57) Testable correlational/conditional prediction about c-command-detection: Provided that MR2([+cc]S, X, beta):yes ∧ MR3([+cc]S, alpha, Y):yes ∧ MR1([+cc]S, X, Y):yes; MR2([-cc]S, X, beta):no ∧ MR3([-cc]S, alpha, Y):no → MR1([-cc]S, X, Y):no

It is the testable correlational/conditional prediction about c-command-detection in (57) that makes it possible for us to plow through the “cloud” of I-languageparticular properties (see (2)) to accumulate knowledge about the contributions to our linguistic intuitions of the initial state of the language faculty, shared by all members of the human species, by the basic scientific method.

The Key Tenets of Language Faculty Science 

 155

As noted, we can obtain c-command detection by the basic scientific method only if we can manage to reduce to zero the effects of sources of non-FR-basedMR(X, Y), in some way. Sources of non-FR-based-MR(X, Y), by hypothesis, are part of what is outside the inner “circle” in (2). They most likely stem from information stored in the “Lexicon” and “Information database” in (42) and how our mind organizes our thoughts beyond what obtains based on a “single sentence” (e.g., how “discourse-related information” is “managed” or “represented”). One can try to state under what conditions non-FR-based-MR(X, Y) can arise without FR(LF(X), LF(Y)), but those conditions will invariably be internal to the individual in question and their applicability affected by the cognitive state of the individual at the given time in question, and so we will need to check things about that individual in order to find out about the conditions; see Chapter 5: Section 8 of this volume. That the conditions in question are internal to the individual in question and are subject to the individual’s linguistic experience leads us to realize that replication in LFS cannot be sought simply in terms of replication of J=yes or no on MR(S, X, Y):J in isolation. That is how we are led to seek to attain and replicate c-command detection by pursuing the correlational methodology as discussed above. C-command detection with MR1(X, Y) based on the X that has been identified as leading to a c-command pattern with MR2(X, beta) and the Y that has been identified as leading to a c-command pattern with MR3(alpha, Y) was crucially based on the assumptions that (i) a given choice of X has the same effects on the J of MR2(S, X, beta):J and the J of MR1(S, X, Y) (with the S being identical in terms of whether it is [-cc]S or [+cc]S), (ii) a given choice of Y has the same effects on the J of MR3(S, X, beta):J and the J of MR1(S, X, Y) (with the S being identical in terms of whether it is [-cc]S or [+cc]S), (iii) the choice of alpha does not affect the J of MR2(S, X, beta), and (iv) the choice of beta does not affect the J of MR3(S, alpha, Y), as discussed in the preceding section; see (52). The assumptions in question will be supported insofar as we obtain experimental results in line with the correlational/conditional prediction and its contrapositive. Conversely, if they lead to incorrect predictions, that will lead us to modify or abandon the assumptions, hopefully by being able to identify which of the assumptions is responsible for the failure of the prediction, as discussed in Chapter 3 of this volume.67, 68 67 The same can be said about another assumption not mentioned, having to do with whether we consider slightly different subsets of [-cc]S (or [+cc]S) to be identical. The relevant issues are suppressed in this chapter but will be addressed in Chapter 5 in reference to a specific I-language (namely my own). 68 Relevant examination will likely lead us to consider the effects of a given choice in (i)–(iv) (in the preceding paragraph) in two different respects; one with regard to the possibility of FR-

156 

 Hajime Hoji

The preceding illustration of the basic tenets of LFS has stayed almost strictly at an abstract level, so that the applicability of the proposed correlational methodology to the study of the language faculty by the basic scientific method can be seen as general as possible. A more concrete illustration and extensive discussion of (3) will be provided in subsequent chapters, based on actual experiments: (3b) will be illustrated in Chapter 5, and (3c) will be illustrated more fully in Chapter 6 (with respect to Japanese) and in Chapter 7 (with respect to English). How one can implement or execute the proposed correlational methodology will be addressed in those chapters. Before concluding this chapter, I would like to briefly address another important aspect of the proposed methodology, having to do with replication.69 Consider again the correlational/conditional prediction in (57), repeated here. (57)

Testable correlational/conditional prediction about c-command-detection: Provided that MR2([+cc]S, X, beta):yes ∧ MR3([+cc]S, alpha, Y):yes ∧ MR1([+cc]S, X, Y):yes; MR2([-cc]S, X, beta):no ∧ MR3([-cc]S, alpha, Y):no → MR1([-cc]S, X, Y):no

The testing of the prediction requires that we first obtain (61), (62), and (63a), and then check whether (63b) also obtains. (61) a. MR2([+cc]S, X, beta):yes b. MR2([-cc]S, X, beta):no (62) a. MR3([+cc]S, alpha, Y):yes b. MR3([-cc]S, alpha, Y):no (63) a. MR1([+cc]S, X, Y):yes b. MR1([-cc]S, X, Y):no One of the first tasks for the LFStist is therefore to check what MR1(X, Y) might result in his/her own judgments as indicated in (63), at least with certain choices of X and Y.70 This task requires the postulation of hypotheses about PSs (e.g.,

based-MR and that of non-FR-based-MR (for each of MR1, MR2, and MR3). If there are, at least, two distinct types of sources of non-FR-based-MR(X, Y), as will be discussed in Chapter 5 of this volume, the examination will have to be even more fine-grained. 69 This will be elaborated on in Chapters 5–7 of this volume. 70 This applies also to (61) and (62).

The Key Tenets of Language Faculty Science 

 157

sentences) containing X and Y, regarding what PS is [+cc]S and what PS is [-cc]S.71 What MRs and what PS’s one might consider, along with what PS-LF correspondence hypotheses one might pursue, depends largely on one’s past research experience.72 For example, the availability of at least some instances of MR(X, Y) in at least some cases seems to be “licensed” by X preceding Y.73 To make sure that MR([+cc]S, X, Y):yes is not due to X preceding Y and MR([-cc]S, X, Y):no is not due to X not preceding Y (i.e., Y preceding X), we can focus strictly on a PS where Y precedes X, at least as our crucial PS’s in relation to our pursuit of c-command detection by the basic scientific method. It is during the process of trying to obtain judgments as indicated in (63) that the researcher typically observes his/her own judgments fluctuate based on choices of X, Y, specific subtypes of the S, and other factors to be addressed in Chapter 5 of this volume. The goal behind (57) is obtaining c-command detection by the basic scientific method. The basic scientific method, however, also allows us to deepen our understanding of factors that contribute to the availability of non-FR-based-MR(S, X, Y), by the use of (57) and its contrapositive (especially the latter). This can be achieved only through the checking of effects of a combination of various factors, which include not only c-command-based, hence CS-related, factors but also non-c-command-based, hence non-CS-related, factors. Such checking can be conducted most effectively by the researcher on him/herself, given his/her familiarity with various issues involved. (57) is to check if we obtain c-command detection, based on the predicted correlation of the absence of the non-FR effects. Its contrapositive in (60), repeated here, is to check the predicted correlations of the presence of the non-FR effects. (60) The contrapositive of (57): Provided that MR2([+cc]S, X, beta):yes ∧ MR3([+cc]S, alpha, Y):yes ∧ MR1([+cc]S, X, Y):yes; MR1([-cc]S, X, Y):yes → MR2([-cc]S, X, beta):yes ∨ MR3([-cc]S, alpha, Y):yes As such, investigation based on (60) is in fact focusing on non-FR factors.

71 Recall that [-cc]S is an S for which it is not possible for LF(X) to c-command LF(Y) and [+cc]S is an S for which it is possible for LF(X) to c-command LF(Y) in LF(S), according to our hypotheses relating the PS to an abstract LF representation. 72 One way for a “novice” researcher to proceed is to first try to replicate other researchers’ reported judgments by using MRs, choices of X and Y, and PSs in his/her own I-language, by using “similar” choices to what those researchers have considered. Whether or not that is actually possible or how difficult that might turn out to be depends in part on, hence is a good indication of, how clearly the other researchers have reported the design of their (self-) experiments. 73 See Ueyama 1998: Chapter 5 and references therein.

158 

 Hajime Hoji

Even for the researcher him/herself, obtaining judgments as indicated in (63) in his/her own self-experiment can be a challenge, and replicating them with other speakers of the same linguistic community is even more of a challenge. (63)

a. MR([-cc]S, X, Y):no b. MR([+cc]S, X, Y):yes

As noted earlier, being concerned with contributions of the initial state of the language faculty to our linguistic intuitions (universal properties of the language faculty), we seek to obtain c-command detection in a given I-language and to replicate successful experimental results obtained in this way in any other I-languages we choose to examine. The “ultimate object” of replication in LFS is thus the validity of universal hypotheses, hence, that of the hypotheses about FR(x, y).74 Once the LFStist has attained c-command detection by the correlational method outlined above (in a reproducible manner) in his/ her own self-experiment and then in non-self-experiment with speakers of the same linguistic community as his/her own, s/he seeks to replicate it with speakers of other “languages”. It is possible that some people get clear judgments in line with (63) in isolation (i.e., without making reference to two other types of MR’s (MR2 and MR3 in the preceding discussion) in their own self-experiment, making them feel not particularly compelled to adopt the correlational methodology indicated in (57). Once they try to attain replication beyond their own I-language, however, they will realize that there is a great deal of judgmental variation and fluctuation, as reported and discussed in some depth in Chapters 5–7 of this volume, and hence that there is a need to adopt a correlational methodology.

6.2 Conclusion An individual speaker’s judgment on MR(S, X, Y) depends not only on whether S is [-cc] S or [+cc]S but also on the choices of X and Y, among many other factors. Unless we have an effective means to control various such factors, we cannot expect to obtain definite and categorical judgments in line with (63). We can understand that the model of judgment making in (40) focuses on c-command-based judg74 Note that the I-language-particular hypotheses must also be valid in order for our predictions to be supported experimentally because our predictions are deduced from a combination of I-language-particular hypotheses and universal hypotheses.

The Key Tenets of Language Faculty Science 

 159

ments, with the important, and as it turns out, incorrect, assumption that it is consistently possible to obtain c-command detection with just one MR(X, Y), by choosing the “right” pair of X and Y (and other relevant factors). The results of “multiple-non-researcher-informant experiments” reported in Hoji 2015 in fact came quite close to obtaining experimental results in support of the categorical prediction in line with (63), with the level of experimental success not attained previously, as far as I know.75 A (perhaps even “the”) serious shortcoming of Hoji 2015, however, is its failure to focus on self-experiment. With the options of X and Y for MR(X, Y) in the “multiple-non-researcher-informant experiments” in Hoji 2015, I do not obtain judgments in line with (63a), i.e., my own judgments would result in disconfirmation of the definite prediction indicated in (63a). With different choices of X and Y, ones not used in the “multiple-non-researcher-informant experiments” reported in Hoji 2015, it is possible for me to obtain judgments in line with (63), at a given time, but the same judgments may not obtain at different times, even with the same choices of X and Y, and my judgments can be affected by other factors, some of which are to be addressed in some depth in Chapter 5 of this volume. Attributing judgmental fluctuation within a single speaker, such as myself, to distinct types of contributions to our judgment, we pursue the correlational methodology in LFS, as summarized in (57) and its contrapositive in (60).76 The correlational methodology in question is crucially based on two types of contributions to our judgment, one having to do with FR(x, y), based most crucially on x c-commanding y, and the other having to do with sources for non-FR-based-MR, belonging to the part of the language faculty outside the CS. Of course, the hypothesized distinction between the two types of factors must be justified by empirical/experimental observations by the basic scientific method. Experimental successes achieved thus far, to be discussed in Chapters 5–7 of this volume, do seem to support this distinction, although it may ultimately not be maintained in its current form in the future as our understanding deepens and broadens.

75 See Chapter 1 of this volume for related discussion. 76 Judgmental variations and fluctuations are often addressed in relation to different groups of speakers or (simply) among different speakers, but, in light of its internalist thesis  – stemming from the fact that the language faculty is internal to an individual – LFS also addresses judgmental variations and fluctuations within the same speaker, at different times. The internalist and individualist perspective in LFS in fact suggests that study of judgmental variations and fluctuations within the same speaker has epistemological primacy over study of judgmental variations and fluctuations among different (groups of) speakers. Discussion in Chapters 5 and 6 of this volume provides an illustration of this point.

160 

 Hajime Hoji

The correlational methodology, which we have been led to adopt as a consequence of pursuing the basic scientific method, allows us to obtain c-command detection, without necessarily reducing to zero the effects of sources of a particular non-FR-based-MR, in self-experiment as well as in non-self-experiment. That the conditions on non-FR-based-MR are internal to the individual in question leads to the view that the most effective way to learn about the sources of non-FRbased-MR(X, Y) is through self-experiment. When an LFStist conducts self-experiment on him/herself, s/he can “measure” his/her intuitions by checking various types of S and by manipulating various factors that might affect his/her judgment in ways that are likely not possible for someone who is not as familiar with the various issues under consideration. My self-experiment in fact has led to the realization that there are, at least, two distinct types of sources of non-FR-basedMR(X, Y), making it possible to make a more fine-grained prediction than in (57), as will be discussed in Chapter 5 of this volume. This is a significant result for the following reason: the initial function of the correlational methodology, propelled by the basic scientific method, was to obtain c-command detection by controlling for effects of sources of non-FR-based-MR(X, Y), not to understand their exact nature. The fact that we seem to be able to attain an understanding about sources of non-FR-based-MR(X, Y) to the extent that we can make definite correlational predictions about the availability of a particular type of non-FR-based-MR(X, Y), however, indicates that the correlational methodology has a wider application and usefulness than its initial conception. That is in fact a good illustration of the ways in which our research, the initial concern of which was strictly with c-command detection and properties of the CS, can move (much) beyond its initial domain of inquiry.

References Chomsky, Noam. 1976. Conditions on rules of grammar. Linguistic Analysis 2. 303–351. Chomsky, Noam. 1981. Lectures on government and binding: The Pisa lectures. Dordrecht: Foris Publications. Chomsky, Noam. 1986. Knowledge of language: Its nature, origin, and use. Westport, CT: Praeger. Chomsky, Noam. 1993. A minimalist program for linguistic theory. In Kenneth Hale and Samuel Jay Keyser (eds.), The view from building 20: Essays in linguistics in honor of Sylvain Bromberger, 1–52. Cambridge, MA: MIT Press. Chomsky, Noam. 1995. The minimalist program. Cambridge, MA: MIT Press. Chomsky, Noam. 2017. The Galilean challenge. Inference: International review of science 3(1). https://inference-review.com/article/the-galilean-challenge (accessed 14 February 2022) Feynman, Richard. 1994 [1965]. The character of physical law. New York: The Modern Library.

The Key Tenets of Language Faculty Science 

 161

Fukui, Naoki and Yuji Takano. 1998. Symmetry in syntax: Merge and demerge. Journal of East Asian Linguistics 7. 27–86. https://doi.org/10.1023/A:1008240710949 (accessed 10 March 2022) Hayashishita, J.-R. 2004. Syntactic and non-syntactic scope. Los Angeles, CA: University of Southern California dissertation. Haider, Hubert. 2018. On minimalist theorizing and scientific ideology in grammar theory. Unpublished ms., University of Salzburg. https://www.researchgate.net/ publication/327601284_On_Minimalist_theorizing_and_scientific_ideology_in_ grammar_theory (accessed 10 March 2022) Hoji, Hajime. 1985. Logical form constraints and configurational structures in Japanese. Seattle, WA: University of Washington dissertation. Hoji, Hajime. 1995. Demonstrative binding and principle B. In Jill N. Beckman (ed.), NELS 25, 255–271. Amherst, MA: University of Massachusetts, Amherst, GLSA Publications. Hoji, Hajime. 1997a. Sloppy identity and principle B. In Hans Bennis, Pierre Pica and Johan Rooryck (eds.), Atomism and binding, 205–235. Dordrecht: Foris Publications. Hoji, Hajime. 1997b. Sloppy identity and formal dependency. In Brian Agbayani and Sze-Wing Tang (eds.), Proceedings of the 15th West Coast Conference on Formal Linguistics, 209–223. Stanford, CA: CSLI Publications. Hoji, Hajime. 1998. Formal dependency, organization of grammar, and Japanese demonstratives. Japanese/Korean Linguistics 7. 649–677. Hoji, Hajime. 2003. Falsifiability and repeatability in generative grammar: A case study of anaphora and scope dependency in Japanese. Lingua 113. 377–446. Hoji, Hajime. 2015. Language faculty science. Cambridge: Cambridge University Press. May, Robert. 1977. The grammar of quantification. Cambridge, MA: MIT dissertation. Newmeyer, J. Fredrick. 2008. A review of linguistic minimalism: Origins, concepts, methods, and aims. by Cedric Boeckx. Language 84. 387–395. Plesniak, Daniel. 2022. Towards a correlational law of language: Three factors constraining judgment variation. Los Angeles, CA: University of Southern California dissertation. Ueyama, Ayumi. 1998. Two types of dependency. Los Angeles, CA: University of Southern California dissertation. Ueyama, Ayumi. 2010. Model of judgment making and hypotheses in generative grammar. In Shoichi Iwasaki, Hajime Hoji, Patricia Clancy and Sung-Ock Sohn (eds.), Japanese/Korean Linguistics 17, 27–47. Stanford, CA: CSLI Publications. Ueyama, Ayumi. 2015. Tōgo imiron [Syntactic semantics]. Nagoya: Nagoya University Publishing.

Hajime Hoji

Detection of C-command Effects 1 Introduction 1.1 The Correlational Prediction in Language Faculty Science: A General Formulation Language faculty science (LFS) seeks to accumulate knowledge about the language faculty by deducing definite/categorical predictions and by obtaining and replicating experimental results in line with such predictions. Since the language faculty is internal to an individual, our predictions are about an individual. Since the language faculty underlies our ability to relate linguistic sounds/signs (henceforth, simply sounds) to meaning, our predictions are about the relation between sounds and meaning, at least at the most fundamental level. We thus deal with an individual speaker’s judgments about the relation between sounds and meaning; we deduce definite predictions about them, and try to obtain and replicate experimental results in line with such predictions. As discussed in Chapter 4, our attempts to obtain empirical support for our basic hypotheses about the language faculty have led us to seek to obtain c-command detection by checking the correlational/conditional prediction of the form in (1).1 (1) Correlational/conditional prediction about c-command-detection (a general formulation): Provided that MR2([+cc]S, X, beta):yes ∧ MR3([+cc]S, alpha, Y): yes ∧ MR1([+cc]S, X, Y): yes; MR2([-cc]S, X, beta): no ∧ MR3([-cc]S, alpha, Y): no → MR1([-cc]S, X, Y):no

The relevant definition and notations are given below, taken from Chapter 4 of this volume. (2) x c-commands y if and only if x is merged with something that contains y2

1 See Chapter 4 of this volume: Section 5. 2 As in Chapter 4 of this volume, I adopt (i) and (ii) for the definition of contain, based on Plesniak’s (2022: 2.2.2) formulation, with slight adaptation (i)

X and Y are the daughters of Z iff Z={X,Y}

(ii) X contains Y iff: a. Y is X’s daughter or b. Y is the daughter of an element Z that X contains.. https://doi.org/10.1515/9783110724790-005

164 

 Hajime Hoji

(3) a. b. c. d. e. f.

PS: phonetic sequence X and Y: linguistic expressions in PS MR(X, Y): meaning relation pertaining to X and Y LF(PS): a 3D (i.e., LF) representation corresponding to PS LF(X): what corresponds X in LF(PS) [-cc] S: a schematized PS that cannot correspond to LF(PS) in which LF(X) c-commands LF(Y),3 g. [+cc]S: a schematized PS that can correspond to LF(PS) in which LF(X) c-commands LF(Y)4

(4) a. J in “MR(S, X, Y):J”: an individual speaker’s judgment (yes or no) on the availability of MR(X, Y) in a particular PS5,6 b. yes in “MR(S, X, Y):yes”: an individual speaker’s judgment that the MR (X, Y) is available in a particular PS c. no in “MR(S, X, Y):no”: an individual speaker’s judgment that the MR (X, Y) is not available in a particular PS We check an individual speaker’s judgments (J=yes or no) at a given time on the availability of a specific type of MR, let us call it MR1(X, Y), in a PS, to see if such judgments are in line with the definite prediction in (1). To do so, judgments

3 S in [-cc]S and [+cc]S can be understood as signifying “schema”, as noted in Chapter 4. 4 The inclusion of the modal (can) in (3g) is due to the possibility that a single PS can correspond to more than one LF and a single LF can correspond to more than one PS, as extensively discussed in Ueyama 1998, in relation to Japanese, as noted in Chapter 4. 5 We use “S” in (4) to refer either to a particular phonetic sequence (PS) or a schematized representation of it (=schema), like a “sentence pattern”, as in (16) below, for example). 6 Chapter 4 of this volume: note 46 points to the need of conceiving of J of “MR(S, X, Y):J” such that it is not just a result of a single act of judgment making (“performed” as in the models of judgment making discussed in Chapter 4), but based on multiple acts of such judgment making, possibly involving different “subtypes” of the S in question and more. Under that conception, we may introduce a somewhat different notation, such as “MR(S∧, X∧, Y∧):J∧” where we clearly distinguish a Schema (S∧) from its specific instance (S), the “general form” (X∧) distinguishing it from its specific instance (X), and the “general form” (Y∧) from its specific instance (Y), and J∧ (Yes or No) is based on a number of instances of J (yes or no). The relevant distinction will in fact prove to be useful and possibly crucial when we consider how the language faculty scientist (LFStist) “arrives at” his/her judgments and how s/he evaluates reported judgments by an experiment-participant on specific sentences for the purpose of obtaining reliable “measurements” of the participant’s judgments, hence for the purpose of reliable replication. In light of the general level of granularity of the presentation in this volume, I will refrain from introducing such new notations. While I will stick to the notations in (4) in the ensuing discussion, what is expressed in terms of (4) is often intended as expressing what would be expressed in terms of such new notations.

Detection of C-command Effects 

 165

on the availability of MR1(X, Y) must be checked in relation to judgments on the availability of two other types of MR, MR2(X, beta) and MR3(alpha, Y). The prediction in (1) would be disconfirmed if we obtained MR1([-cc]S, X, Y):yes with the X such that we obtained the judgments in (5) and the Y such that we obtained the judgments in (6) from a given speaker at a given time. (5) a. MR2([+cc]S, X, beta):yes b. MR2([-cc]S, X, beta):no (6) a. MR3([+cc]S, alpha, Y):yes b. MR3([-cc]S, alpha, Y):no If we additionally get “MR1([+cc]S, X, Y):yes” (as stated in the last conjunct of the “Provided that . . .” clause in (1)), this means that we have obtained c-command detection with MR1(X, Y) from that speaker at that time. This method of c-command detection ultimately derives from the combination of the hypotheses in (7) and (8), and from the fact that the only difference between the two schemata in (9) is that LF(X) c-commands LF(Y) in [+cc]S in (9a) but not in [-cc]S in (9b); see (3f) and (3g). (7) FR(x, y) is possible only if x c-commands y. (8) MR(X, Y) can arise based on FR(LF(X), LF(Y)) or something essentially nonstructural. (9) a. MR([+cc]S, X, Y):yes b. MR([-cc]S, X, Y):no The discussion in Chapter 4 about what hypotheses lead to (1) and how it would be tested was mostly conceptual and abstract, as we stayed away from addressing a specific I-language in most of the chapter so as to be able to emphasize the general applicability of the proposed correlational methodology. In this chapter, I will provide a concrete illustration of the proposed correlational methodology for LFS, dealing with my own I-language as a native speaker of Japanese. Reference to a specific I-language (my own I-languages), allows me to discuss details necessary to consider when trying to find out about the language faculty by the basic scientific method. Recall that the ultimate concern of LFS is properties of the initial state of the language faculty, i.e., universal properties of the language faculty, a basic part of the human nature. As noted in Chapter 4, however, a great deal of “work” needs to be done about a specific I-language because we can only get to the world of the initial state through the “cloudy” world of a steady state, so

166 

 Hajime Hoji

to speak; see specifically the discussion in Chapter 4: Sections 4 with regard to the “two models” of judgment making.7 Addressing properties of a specific I-language (e.g., my own I-language(s), as in this chapter) in fact leads to a new understanding about different types of sources for MR(X, Y), in particular the ones that are essentially non-structural and thus distinct from FR (see (7) and (8)), making it possible to provide a not only more concrete but also more satisfactory illustration of what is intended by both (1) and its contrapositive, as will be discussed below.

1.2 Specific MRs (Meaning Relations) and Specific I-language-particular Hypotheses To illustrate what is meant by (1) and related concepts, I now turn to specific MRs. As noted in Chapter 4, what specific MRs one chooses in (1) depends largely on one’s past research experience, which may well reflect the “collective research experience” of the field. I choose to “work with” BVA(X, Y), as I have done since the early 1980s. BVA(X, Y) is illustrated by the English sentences in (10), with X=‘every engineer’ and Y=‘his’. (10)

BVA(every engineer, his): a. Every engineer praised his robot. b. His robot praised every engineer. c. His robot, every engineer praised.

In this chapter, we will be dealing with my I-language as a native speaker of Japanese and the English sentences in (10) are employed just to illustrate what is meant by BVA(X, Y).8 What is intended by BVA(X, Y) is that the “value” of a singular-denoting expression X “depends upon” that of a non-singulardenoting expression Y.9 Take (10b), for instance; what is intended by BVA

7 One is an abstract model concerned with properties of the initial state, more specifically, properties of the CS (the Computational System of the language faculty) and the other addresses factors outside the CS. 8 It is therefore not of relevance whether BVA(every engineer, his) obtains in (10b) and (10c) for an individual native speaker of English. (10a) is even less relevant because X precedes Y in that sentence can introduce a confound, as will be addressed. As such, (10a) is included here as a “reference sentence”, so to speak. 9 By this “definition” of BVA, in order for ‘every engineer’ to serve as X of BVA(X, Y), it must correspond to more than one engineer. The meaning relation between ‘every engineer’ and ‘his’ thus does not count as an instance of BVA(every engineer, his) if there is only one engineer.

Detection of C-command Effects 

 167

(every engineer, his) in (10b) is an interpretation like (11a), for example, often expressed as in (11b). (11) a. Assuming that there are three engineers, A, B, and C. A’s robot praised A, B’s robot praised B, and C’s robot praised C. b. for every x such that x is an engineer, x’s robot praised x With various options for X and Y, along with actual sentences instantiating various [+cc]S and [-cc]S in Japanese, I have checked over the years (since the early 1980s) whether I obtain judgments as indicated in (12).10 (12) a. BVA([+cc]S, X, Y):yes b. BVA([-cc]S, X, Y):no If judgments in line with (12) clearly obtain, that points to effects of c-command because the yes judgment in (12a) and the no judgment in (12b) would be in line with consequences of the PS-LF correspondence hypotheses we adopt, as given in (13), here focusing on a sentence that contains a ga-marked NP (=NP2) and an o-marked NP (NP1).11 (13)

PS-LF correspondence hypotheses: a. NP2-ga NP1-o V (roughly, Subject Object Verb) in Japanese must correspond to a “3D” (i.e., LF) representation where LF(NP2) asymmetrically c-commands LF(NP1), as in (14). b. NP1-o NP2-ga V (roughly, Object Subject Verb) in Japanese can correspond to a “3D” (i.e., LF) representation where LF(NP2) asymmetrically c-commands LF(NP1), as in (14).

(14) a. NP2 NP1

V

10 In this volume, Chapter 1: Section 2. 11 What “case markers” (such as -ga, -o, -ni) are used for NPs with a particular (type of) verb seems determined in part by “lexical specification” of the (type of) verb in question. We will return to this later when we address two other types of “case-marking” patterns in Japanese.

168 

 Hajime Hoji

b. NP2

NP2

NP2

NP1

V

V

NP1

V

NP1

In (14), NP1 is merged with the V, resulting in {NP1, V}, and {NP1, V} in turn is merged with NP2, resulting in {NP2, {NP1, V}}. When the verb has more than one “argument”, the one that is merged with the V is what receives the V’s “internal theta role” (which we can informally understand as its “object theta role”).12 In case the V has two “arguments”, we can say that the NP that is merged with {NP1, V} receives its “subject theta role”. (13) thus states (15a) and (15b), in effect. (15) a. The LF representation as in (14) can be “externalized” as “NP2-ga NP1-o V” (in the Subject-Object-Verb order) or as “NP1-o NP2-ga V” (in the Object-Subject-Verb order). b. While “NP2-ga NP1-o V” must correspond to an LF representation as in (14), “NP1-o NP2-ga V” (of the Object-Subject-Verb order) can correspond to it.13

12 If, for example, the V is like hihansuru ‘to criticize’ (given here as its “dictionary” form), its “internal theta role” (or the “object theta role”) “assigned” to the NP merged with the V (NP1 in (14)) is the “semantic role of a criticizee” and the “subject theta role” “assigned” to the NP merged with {NP1, V} (NP2 in (14)) is the “semantic role of a criticizer”. 13 As indicated in the Figure in (i), “NP1-o NP2-ga V” (of the Object-Subject-Verb order) can also correspond to another LF where LF(Object) c-commands LF(Subject), which is the “Deep OS type” in the terms of Ueyama 1998. (i) 

LF

PS NP2-ga NP1-o V (e.g., John-ga susi-o tabeta. ‘John ate sushi.’) NP2

NP2

NP2 V

NP1

V

NP1

V

NP1

LF(NP2) c-commands LF(NP1)

NP1-o NP2-ga V (e.g., Susi-o John-ga tabeta. ‘(Lit.) Sushi, John ate.’)

Another LF where LF(NP1) c-commands LF(NP2)

Detection of C-command Effects 

 169

Note that the numbers (1 and 2) on the NPs indicate the “order” in which the two NPs are merged, not the surface order of these two NPs.14 According to (13), (16a) can, but (16b) cannot, correspond to an LF representation in which LF(X) c-commands LF(Y), where we understand that “[ . . . Y . . . ]” in (16) is a phrase containing Y. (16)

a. b.

S: [ . . . Y . . . ]-o X-ga V(erb) S: [ . . . Y . . . ]-ga X-o V(erb)

[+cc] [-cc]

(16a) is an instance of [+cc]S because of (13b); it can correspond to an LF representation where Y is part of what is merged with V, i.e., part of the set {[ . . . Y . . . ], V}, and X is further merged with that set, resulting in the set {X, {[ . . . Y . . . ], V}}, at least roughly speaking.15 (16b), on the other hand, is an instance of [-cc]S because of (13a); it must correspond to an LF representation that results from X and V first merging to form {X, V}, which is then merged with a phrase that contains Y (i.e., [ . . . Y . . . ]), forming {[ . . . Y . . . ], {X, V}}. Restricting our discussion to cases where each of LF(X) and LF(Y) is, or is contained in, NP1 or NP2 in the LF representation in (14), recall that LF(X) c-commands LF(Y) iff LF(X) is merged with something that contains LF(Y). Given that (13), (17a) and (17b) can correspond to {X, {[ . . . Y . . . ], V}} and {X, {Y, V}}, respectively, these are all cases where LF(X) is merged with something that contains LF(Y), making these PSs instances of [+cc]S. (17) Instances of [+cc]S a. [ . . . Y . . . ]-o X-ga V b. Y-o X-ga V

When we consider the “Deep OS type”, the simplified presentation above that “NP1” is necessarily merged with the V will have to be modified, along with what hypotheses lead to the correspondences in the Figure in (i). The non-self-experiment, to be discussed in Chapter 6 of this volume, includes sentences of the “Deep OS type”. Reference to the “Deep OS type” is necessary for our intended point about the FR-MR (“FR-MR” abbreviates “FR-based-MR”) ([+cc]S, X, Y):yes being due to LF(X) c-commanding LF(Y) rather than some theta-role-based considerations. See Plesniak’s Chapter 7 of this volume for related discussion. 14 Strictly speaking, what is merged are not NPs, but what corresponds to each of the NPs (LF(NP1) and LF(NP2)) in LF(PS). 15 The assumption that “. . . Y . . .” in a PS in (16) necessarily corresponds to LF(. . . Y . . .) is a simplification necessary for our current discussion. It is hoped that successively more reliable means for c-command detection established in the future will help us clarify the nature of this assumption.

170 

 Hajime Hoji

On the other hand, according to (13), there is no LF(PS) where LF(X) c-commands LF(Y) corresponding to (18a) or (18b), because LF(X) is merged with just V, making (18) instances of [-cc]S.16 (18)

Instances of [-cc]S a. [ . . . Y . . . ]-ga X-o V b. Y-ga X-o V

16 While (17) exhaust the “major” instances of [+cc]S for a sentence whose major constituents are NP-ga and NP-o, (18) and (i) do not exhaust the “major” instances of [-cc]S. (i) Instances of [-cc]S a. [ . . . Y . . . ]-ga [ . . . X . . . ]-o V b. Y-ga [ . . . X . . . ]-o V c. [ . . . Y . . . ]-o [ . . . X . . . ]-ga V d. Y-o [ . . . X . . . ]-ga V (i-a) and (i-b) are instances of [-cc]S because LF(X) is not merged with something that contains LF(Y) in their LF representations. The PSs in (i-c, d) can correspond to an LF representation where LF([. . . X . . . ]) c-commands LF(Y), according to (13b), but LF(X) fails to c-command LF(Y); hence (i-c, d) are instances of [-cc]S. (17), (18) and (i), combined, exhaust the “major” instances of [+cc] S and [-cc]S for a sentence whose major constituents are NP-ga and NP-o, and they give us an idea about how to check the “major” instances of [+cc]S and [-cc]S for sentences of a different type. It is important to bear in mind that one of the minimal requirements for pursuing the basic scientific method is that we check things as exhaustively as possible so as to be able to check if there is any case where the predicted entailments fail to obtain. Since we are considering an infinite array of possibilities (due to the digital infinity of the language faculty), we can check things exhaustively only by creating an exhaustive list of what is being checked. In the case at hand, we have an initial exhaustive list of schemata consisting just of [+cc]S and [-cc]S, and each of these has an exhaustive list of its “major” subtypes, as just discussed. Each “major” subtype can have its own subtypes. “[. . . Y . . . ]” in (17a), for example, can be of the form “Y-no N” (Y’s N), where Y is “connected to” the head N by the “genitive case marker” -no, but it can also be of the form where Y is embedded in a clause (such as a relative clause) modifying the head N. When we consider various possibilities for X and Y, in addition to the possibilities of different schemata, what needs to be checked can increase (almost) exponentially. A “hypothesized” classification of types of X and Y is thus an integral part of (experimental) research in LFS. In my self-experiment, I check my judgments exhaustively in terms of (17), (18) and (i), and try to check my judgments as exhaustively as possible with regard to types of X and Y, although space limitations allow me to provide here only a very small portion of the results of my self-experiments. In all the PSs in (17), (18), and (i), Y precedes X. As addressed briefly in Chapter 4: 6.1, the availability of at least some instances of MR(X, Y) in at least some cases seems to be “licensed” by X preceding Y. In my self-experiments discussed in this chapter, I avoid PSs where X precedes Y to exclude the possibility of the precedence relation (X preceding Y) affecting the availability of the MR in question. Once an effective and reliable means for c-command detection has been established, we can investigate the possible effects of precedence. If we do that, we will be considering the variant of each PS in (17), (18), and (i) where X and Y are switched.

Detection of C-command Effects 

 171

Given in (19) are some of the different types of X for BVA(X, Y).17 What is provided below is by no means an exhaustive list. (19)

a. b. c. d.

QE-no N18 conjoined NP dono N-mo ‘Lit. which N-also’ ‘every N’ NP-igai ‘ones other than NP’

Given in (20) are among the different types of Y for BVA(X, Y). (20) a. b. c. d. e.

soko ‘that place, it’ soitu ‘that guy’ sono N ‘that N’ kare ‘he, him’ kare ka kanozyo ‘he or she, him or her’

Returning to the choices of X, various quantity expressions (“QE”) that can be used in (19a) include those listed in (21). (21)

Various instances of QE in (19a): a. subete every b. kanarino kazu considerable number c. numeral + cl(assifier) e.g., 3-nin 3-cl

There are “variations” of each of (21). For example, hotondo ‘almost’ can be added to (21a), sukunakutomo ‘at least’ and izyoo ‘more than’ can be added to (21c), as in ‘almost every N’ and ‘at least three or more Ns’ in English. “Cl” in (21c) can be replaced by paasento ‘percent’. Kanarino ‘considerable’ in (21b) can be replaced by a “regular” noun-modifying clause, as corresponding to ‘that John expected/ 17 Hayashishita and Ueyama 2012 provides a comprehensive review of the types of expressions that can potentially serve as X of BVA(X, Y) in Japanese, along with references to past research on the topic. 18 QE: quantify expression. This follows the terminology used in Hayashishita and Ueyama 2012. See below. The particle -no is the so-called “genitive marker,” typically needed between a non-clausal modifier/complement and a head noun.

172 

 Hajime Hoji

announced/etc.’ in ‘(the/a) number that John expected/announced/etc.” in English. Furthermore, (19a) is one of the three possible patterns given in (22). (22) a. (=(19a), with the addition of CM (=case marker or a postposition)) QE-no N-CM b. N-CM QE c. N QE CM Let us consider (16) again. (16) a. b.

S: [ . . . Y . . . ]-o X-ga V(erb) S: [ . . . Y . . . ]-ga X-o V(erb)

[+cc] [-cc]

(16a) is the [+cc]S in (17a) and (16b) is the [-cc]S in (18a). By choosing a pair of X and Y among the various choices for X and Y, introduced above, and by choosing a particular verb and particular lexical items for the “. . .” parts in (16), we can “construct” numerous sentences corresponding to [+cc]S and [-cc]S, as in (23) and (24), for example. (23) a. BVA([+cc]S, subete-no kaisya, soko): soko-no robotto-o subete-no kaisya-ga hihansita it-gen robot-acc every-gen company-nom criticized ‘its robot, every company criticized’ b. BVA([-cc]S, subete-no kaisya, soko): soko-no robotto-ga subete-no kaisya-o hihansita it-gen robot-nom every-gen company-acc criticized ‘its robot criticized every company’ (24) a. BVA([+cc]S, sukunakutomo 55%-izyoo-no gisi, sono otoko): sono otoko-no robotto-o sukunakutomo 55%-izyoo-no that man-gen robot-acc at:least 55%-more:than-gen gisi-ga hihansita engineer-nom criticized ‘that man’s robot, at least 55% or more engineers each criticized’ b. BVA([-cc]S, sukunakutomo 55%-izyoo-no gisi, sono otoko): sono otoko-no robotto-ga sukunakutomo 55%-izyoo-no that man-gen robot-nom at:least 55%-more:than-gen gisi-o hihansita engineer-acc criticized ‘that man’s robot criticized each of at least 55% or more engineers’

Detection of C-command Effects 

 173

If I obtain judgments in line with (12), repeated below, with the [+cc]S and the [-cc] S therein being as given in (16), also repeated here, that will point to c-command effects. (12) a. BVA([+cc]S, X, Y):yes b. BVA([-cc]S, X, Y):no (16) a. b.

S: [ . . . Y . . . ]-o X-ga V(erb) S: [ . . . Y . . . ]-ga X-o V(erb)

[+cc] [-cc]

Whether I obtain judgments in line with (12), however, depends upon factors such as (i) the choice of X, (ii) the choice of Y, (iii) the choice of the N of (19a), if (19a) is chosen for X, and (iv) the choice of the N of (20c), if (20c) is chosen for Y, among other factors. Furthermore, my judgments are not always stable; they change over time, as will be discussed in Sections 2–4 below.19 That is not surprising, given that BVA(X, Y) can arise not only based on FR(LF(X), LF(Y)) but also due to something essentially non-structural, which is a specific instance of (8), repeated here. (8)

MR(X, Y) can arise based on FR(LF(X), LF(Y)) or something essentially nonstructural.

Let us refer to any source of MR(X, Y) that is essentially non-structural as a non-formal source (NFS), and restate the BVA version of (8) as (25). (25) BVA(X, Y) can arise based on FR(LF(X), LF(Y)) or an NFS.20

19 As discussed in Chapter 1 of this volume, Hoji 1985, for example, was an attempt to obtain judgments (from myself and my colleagues) in line with (12), with various choices of X and Y (and various choices of [+cc]S and [-cc]S). The attempt was not particularly successful, which is not surprising in light of both the judgmental fluctuation we know to be possible in an individual speaker and judgmental variation among speakers. An attempt is made in Hoji 2015 to obtain judgments in line with (12) from non-researcher participants in a large-scale “experiment” in a reproducible manner. Although the “multiple-non-researcher-informant experiments” reported in Hoji 2015 came quite close to obtaining results in support of the categorical prediction in line with (12), Hoji 2015, however, suffered from a serious shortcoming by not focusing on self-experiments, as pointed out in Chapter 4: 6.2. 20 As will be addressed briefly in Section 8, there seem to be at least two types of NFS, and the availability of one is crucially based on the choice of X and how X is construed and that of the other is crucially based on the choice of Y and how Y is construed.

174 

 Hajime Hoji

While the choices of X and Y are not the only factors that affect the possibility of NFS-based-BVA(X, Y),21 they are the most easily “controllable” factors, in addition to having most clearly discernable effects; see below. For this reason, our initial illustration of the correlational methodology will focus on the choices of X and Y. Given that we do not have a good/full understanding of how speaker judgments are affected by NFS factors, it is highly unlikely that we can reduce their effects to zero when considering the availability of BVA(X, Y). Our correlational methodology, which we pursue in the form of (1), is meant to allow us to focus on effects of FR by controlling for effects of NFS. It is hoped that initial success of the correlational methodology, pursued in the form of (1), will make it possible for us to not only explore further properties of the FR in question (through investigation of FR-MR’s) but also investigate various NFS factors. Research in that direction so far seems quite promising, as will be discussed briefly in later sections including Section 8. In summary, judgments in line with (12) obtain only with a few of the combinations of X and Y considered, and even when they do, the judgments do not always remain the same, providing an example of the instability of speaker judgments. In the case at hand, because the instability of judgments is my own, it cannot be attributed to the inattentiveness of the speaker or the speaker failing to understand what is intended by BVA(X, Y) (since I am neither inattentive nor failing to understand my own intentions). I now turn to instances of MR2(X, beta) and MR3(alpha, Y) to illustrate the correlational/conditional prediction in the form of (1). Just as the choice of BVA depends largely upon one’s research experience in the past, so do the choices of MR2 and MR3.22 I choose DR(X, beta) for MR2, and Coref(alpha, Y) for MR3, which are exemplified by English sentences in (26) and (27), respectively. These choices largely reflect what MRs have been discussed in the field of generative linguistics when structural conditions on MRs are addressed.23 21 “NFS-based-BVA” will often be abbreviated as “NFS-BVA”, as in the case of similar abbreviations such as “FD-BVA”, “FR-MR”, etc. 22 Daniel Plesniak (p.c., February 2022) suggested to me that successful choices of MR2 and MR3 might depend largely on the choice of MR1. The suggestion is reasonable in light of the roles played by MR1, MR2 and MR3 in the correlational prediction in (1) and its specific instance to be given in (29). While the specific choices of three types of MRs discussed here might appear to be due to “historical accident”, the logic of the correlational prediction points to the non-accidental nature of their choices, as suggested. 23 Judgmental variation, instability, etc. on the availability of such MRs seem to have led researchers to abandon their attempt to seek to obtain definite and categorical speaker judgments on the availability of the relevant MRs, within a speaker or across speakers, which seems to have

Detection of C-command Effects 

(26)

DR(every engineer, three robots): a. Every engineer praised three robots. b. Three robots praised every engineer. c. Three robots, every engineer praised.

(27)

Coref(that engineer, his): a. That engineer praised his robot. b. His robot praised that engineer. c. His robot, that engineer praised.

 175

Take (26b), for instance; what is intended by DR(every engineer, three robots) in (26b) is an interpretation as in (28). (28)

Each of the engineers in question was praised by a distinct set of three robots.

What is intended by Coref(alpha, Y), on the other hand, is that alpha and Y are understood to be referring to the same individual/object (e.g., ‘his’ is understood as ‘that engineer’s’).24

1.3 A Specific Instance of the Correlational Prediction With BVA, DR, and Coref, as introduced above, serving as MR1, MR2, and MR3, respectively, we now have (29), as a specific instance of (1).

prompted the emergence of the currently-prevailing research concerned with tendencies of judgments among speakers, as discussed in some depth in Plesniak 2021. 24 It is perhaps useful to point out that MRs, such as BVA, DR, and Coref, are not part of the object of inquiry, but they are observational tools for finding out about properties of the language faculty by the basic scientific method, more specifically for the purpose of c-command detection. BVA, for example, is not the same as what has been discussed as “quantificational binding”, “pronominal binding”, etc. For example, the meaning relation between someone and his does not count as BVA(someone, his) because someone is singular-denoting; likewise, the meaning relation between everyone and their does not count as BVA(everyone, their) unless their is understood to be singular-denoting; see Plesniak 2022: Chapter 4 for discussion of the singular-denoting use of their. Likewise, DR is not the same as “scope dependency” as discussed in the field. The “scope interaction” between someone and everyone, where the former is said to take scope over the latter, does not count as DR(someone, everyone) because we cannot have an interpretation like: for each of those corresponding to someone, there is a distinct set of “everyone”.

176  (29)

 Hajime Hoji

Correlational/conditional prediction about c-command-detection with BVA(X, Y): Provided that DR([+cc]S, X, beta): yes ∧ Coref([+cc]S, alpha, Y): yes ∧ BVA([+cc]S, X, Y): yes; DR([-cc]S, X, beta):no ∧ Coref([-cc]S, alpha, Y): no → BVA([-cc]S, X, Y):no

To clarify the meaning of (29), let us consider the significance of ①, ② and ③ (= ④ and ⑤) in (30). (30)

Provided that

DR(

S, X, beta):yes + Coref(

S, alpha, Y):yes

+

BVA(

S, X, Y):yes

DR(

S, X, beta):no

S, alpha, Y):no



BVA(

S, X, Y):no

Identification of X that gives rise to DR(X, beta) but does not give rise to NFS-DR(X, beta), i.e., X that gives rise to DR(X, beta) only based on FR.

+ Coref(

Identification of Y that gives rise to Coref(alpha, Y) but does not give rise to NFS-Coref(alpha, Y), i.e., Y that gives rise to Coref(alpha, Y) only based on FR.

;

When accompanied by and , it is c-command detection with BVA(X, Y), with the X and Y identified in and . In isolation, identification of a pair of X and Y that gives rise to BVA(X, Y), but does not give rise to NFS-BVA(X, Y), i.e., a pair of X and Y that gives rise to BVA(X, Y) only based on FR.

Figure 1: Illustrating the correlational/conditional prediction about c-command detection with BVA(X, Y) in (29).

Recall that judgments addressed in (30) are judgments by a given speaker (myself in my self-experiment) at a given time. ① constitutes the identification of an X that does not give rise to NFS-DR(X, beta) for a given speaker at a given time because we have the hypothesis that DR(X, beta) can arise based on FR or NFS, as in (31b). (31)

a.

(=(25)) BVA(X, Y) can arise based on FR(LF(X), LF(Y)) or NFS. b. DR(X, beta) can arise based on FR(LF(X), LF(beta)) or NFS. c. Coref(alpha, Y) can arise based on FR(LF(alpha), LF(Y)) or NFS.

DR([+cc]S, X, beta):yes, along with DR([-cc]S, X, beta):no in ①, indicates that DR([-cc] S, X, beta):no in ① is due to the absence the c-command relation required for

Detection of C-command Effects 

 177

FR-DR(X, beta). Likewise, Coref([+cc]S, alpha, Y):yes, along with Coref([-cc]S, alpha, Y):no in ②, indicates that Coref([-cc]S, alpha, Y):no in ② is due to the absence the c-command relation required for FR-DR(X, beta).25 ① and ②, combined, are predicted to entail ④, and this predicted entailment is the disconfirmable correlational/conditional prediction in (30). Assumed here are (32a) and (32b).26 (32)

a.

X gives rise to NFS effects with BVA(X, Y) only if it gives rise to NFS effects with DR(X, beta).

b. Y gives rise to NFS effects with BVA(X, Y) only if it gives rise to NFS effects with Coref(alpha, Y). ① and ②, combined, however, are not predicted to entail ⑤ because the c-command condition on FR(x, y) is only a necessary condition, not a sufficient condition (see (7)); hence ① and ②, combined, are not predicted to entail ③, either. Let us turn to ③, which, in isolation, would also constitute the identification of a pair of X and Y that gives rise to BVA(X, Y) but does not give rise to NFSBVA(X, Y). This is similar to how ① and ② constitute the identification of X that gives rise to DR(X, beta) but does not give rise to NFS-DR(X, beta) and the identification of Y that gives rise to Coref(alpha, Y) but does not give rise to NFS-Coref(alpha, Y), respectively. ③, however, is different from each of ① and ②, because ③ involves both X and Y. Each of ① and ② provides us with a means to identify, for a given speaker at a given time, whether a given choice of X does or does not give rise to NFS-DR(X, beta) and whether a choice of Y does or does not give rise to NFS-Coref(alpha, Y). When taken separately, they do not provide us with a means

25 Assumed at this point is (i), as discussed in Chapter 4: Section 5. (i) The choice of beta and that of alpha do not lead to NFS effects with DR(X, beta) and Coref(alpha, Y). It is also assumed here that, if there are no NFS effects with DR or Coref in [-cc]S, then there are no NFS effects with DR or Coref in the “corresponding” [+cc]S and that other possible factors are controlled for; see Plesniak 2022: 2.7, 2.8, and 3.2 for more details. 26 (32) is a simplification for the initial illustration of the correlational methodology. Section 5, where the absence of other types of entailment is discussed, will address what is simplified in relation to my self-experiment. The design of the “large-scale” non-self-experiment discussed in Chapter 6 of this volume is based on (32) and its result is precisely as predicted, indicating that effects of simplification such as this can be effectively controlled by our correlational methodology. An advancement of our understanding of different types of NFS effects might make it possible for us to maintain stronger versions of (32), with “iff” replacing “only if”.

178 

 Hajime Hoji

to identify (again for a given speaker at a given time) what combination of X and Y gives rise to NFS-BVA(X, Y). In order for the combination of ① and ② to entail ④, it must be assumed that NFS effects with BVA(X, Y) cannot be given rise to by an effect stemming from the particular combination of X and Y, when X and Y alone cannot give rise to NFS effects with DR(X, beta) or Coref(alpha, Y), respectively.27 Let us review this in somewhat more concrete terms. According to (13), (33a) can, but (33b) cannot, correspond to an LF representation in which LF(X) c-commands LF(Y). (33)

a. [NP1 Y]-o [NP2 X]-ga V (an instance of NP1-o NP2-ga V) b. [NP2 Y]-ga [NP1 X]-o V (an instance of NP2-ga NP1-o V)

(33a) is thus an instance of [+cc]S and (33b) an instance of [-cc]S. Judgments in line with (34) (i.e., an instance of ①), therefore, point to c-command effects, as discussed, identifying a choice of X that does not lead to NFS-based-DR(X, beta) for a given speaker at a given time. (34) a. DR((33a), X, beta):yes b. DR((33b), X, beta):no Similar remarks apply to the pattern of judgments in line with (35) (i.e., an instance of ②), which points to c-command effects, identifying a choice of Y that does not lead to NFS-based Coref(alpha, Y) for a given speaker at a given time. (16)

a. b.

(35)

a. Coref((16a), alpha, Y):yes b. Coref((16b), alpha, Y):no

S: [ . . . Y . . . ]-o X-ga V(erb) S: [ . . . Y . . . ]-ga X-o V(erb)

[+cc] [-cc]

It is the identification of such X and Y that leads to the predicted entailment in (30), i.e., the entailment of ④ from a combination of ① and ②. ④ in turn constitutes c-command detection, when accompanied by ⑤. In summary, obtaining each of ①, ②, and ③ points to c-command effects, in the form of confirmation of an existential prediction, so to speak, that (i) there exists a choice of X such that ① obtains, (ii) that there exists a choice of Y such 27 This is a simplification, closely related to (32), to make the present discussion manageable. Remarks similar to what is noted in footnote 25 apply here. I will return to this briefly in Section 8.

Detection of C-command Effects 

 179

that ② obtains, and (iii) that there exist choices of X and Y such that ③ obtains, for a given speaker at a given time. Since such predictions are “existential predictions”, they cannot be disconfirmed, in isolation. What leads to a disconfirmable prediction is considerations of correlations of judgment as indicated in (30).

1.4 Outline of the Rest of the Chapter In this chapter, I will first illustrate how I obtain results in my self-experiments in line with the correlational/conditional prediction in (30). (Sections 2–4) More than twenty options for X of BVA(X, Y) and DR(X, beta) and more than ten options for Y for BVA(X, Y) and Coref(alpha, Y) are considered. For each one of the more than 20x10=200 combinations of X and Y for BVA(X, Y), I have obtained judgments in line with (30) in a number of schemata. The discussion below only covers a small, but representative, cases of what I have checked in my self-experiments. The self-experiment has two general stages.28 In the first stage, I identify Xs of DR(X, beta) that lead to ① (Section 2) and Ys of Coref(alpha, Y) that lead to ② (Section 3). The purpose of this is to identify which choices of X do not lead to NFS-DR(X, beta) and which choice of Y do not lead to NFS-Coref(alpha, Y) for myself at a given time. In the second stage, I test the prediction in (30) that I necessarily obtain ④ (=BVA([-cc]S, X, Y):no) with the X and Y identified as such during the first stage (Section 4). If I were instead to obtain BVA([-cc]S, X, Y):yes, with the X and Y in question, that would disconfirm the definite correlational prediction in (30). As will be discussed below, the choices of X and Y that lead to c-command detection are not always the same, even for me at different points in time, but once such X and Y are identified for me at a given point in time, based on DR(X, beta) and Coref(alpha, Y), they also lead to c-command detection with BVA(X, Y). This then counts as c-command detection not only in the existential sense, but in a rigorously testable manner, which is enabled by the correlational methodology. With the remarkable effectiveness of the correlational methodology, one might wonder whether the correlational/conditional prediction needs to be as specific/ complex as what is presented in (29) and (30). In Section 5, I will address other types of “correlational/conditional” predictions, simpler than that in (29) and (30)

28 Not to be confused with different stages of my “mental states” affecting my judgments, as will be discussed in some depth in subsequent sections.

180 

 Hajime Hoji

and show that the definite/categorical prediction in such cases is disconfirmed in my self-experiment, by the demonstration of the existence of the absence of the predicted entailment.29 In that section, we will begin to address part of what is simplified in the preceding sections. In Section 6, I will illustrate how the contrapositive of (29)/(30) is checked in my self-experiment.30 The empirical coverage will be expanded in Section 7, in relation to different sentence patterns and additional structural requirement (in addition to the c-command requirement) on some instances of FR(x, y). Section 8 provides further discussion to give the reader a glimpse at the further issues that have been, and are currently being, considered and investigated. It should be noted that the main point of Sections 7 and 8 is to provide a small demonstration of the fact that behind the simplified presentation of the proposed correlational methodology for LFS in Sections 1–6, there are a great deal of empirical as well as theoretical/conceptual considerations not (fully) covered in this chapter. Section 9 concludes the chapter.

2 DR(X, beta): Identifying Effective Choices of X for C-command Detection Among the more than twenty choices of X and close to ten choices of Y for BVA (X, Y), DR(X, beta) and Coref(alpha, Y) that I have checked in my recent selfexperiments, the following illustration uses only three choices for X as given in (36), and three choices for Y to be given later in (42). (36)

Choices for X for BVA(X, Y) and DR(X, beta): a. subete-no N ‘every N’ b. sukunakutomo numeral-cl-no N ‘at least # Ns’ c. NP-igai ‘ones other than NP’

While the choice of X of DR(X, beta) has significant effects on the availability of the DR(X, beta), the choice of beta of DR(X, beta) does not. In the ensuing discussion, #-cl-no N ‘# N’s’ will be used as beta in DR(X, beta). The schemata considered here are the [+cc]S in (37a) and the [-cc]S in (37b), and I consider my judgments (J) in (38a) and (38b). 29 The result of a “large-scale” non-self-experiment, discussed in Chapter 6 of this volume, replicates the result of my self-experiment. 30 The contrapositive of (29)/(30) will also be addressed in Chapter 6 of this volume, in relation to a “large-scale” non-self-experiment.

 181

Detection of C-command Effects 

S: beta-o X-ga V S: beta-ga X-o V

(37) a. b.

[+cc]

(38) a. b.

DR([+cc]S, X, beta):J DR([-cc]S, X, beta):J

[-cc]

My judgments in (38a) and (38b), with the three choices of X in (36), are summarized in (39). (39) My judgment (J) on the availability of DR(X, beta) on the [+cc]S in (37a) and the [-cc]S in (37b):31 X beta

(36a)

(36b)

(36c)

subete-no N

sukunakutomo #-cl-no

NP-igai ‘ones

‘every N’

N ‘at least # Ns’

other than NP’

S

#-cl-no N

[+cc]

[-cc]

S

[+cc]

S

[-cc]

S

[+cc]

S

[-cc]

yes

yes

yes

no/yes

yes

no

S

I consistently get J=yes in [+cc]S, regardless of the choice of X, usually without any effort. My judgments on [-cc]S, on the other hand, are much less clear or consistent. I get J=no in [-cc]S with X=(36c), even with a great deal of effort. My J in [-cc]S with (36b) is not always the same; I sometimes find it difficult to accept the DR, but, with some effort, I can accept it, which I indicate by the “no/yes” under [-cc]S under (36b) in the chart in (39).32, 33 Given in (40) are some example sentences instantiating [+cc]S.

31 I focus here on sentences with the “o-marked” object, but the same patterns of judgments obtain with the sentences with the “ni-marked object”, as will be discussed in Section 7. This also applies to sentences with Coref and BVA, to be addressed below. 32 The “shift” of my judgments on DR with (36b) is not nearly as systematic as the shifts of my judgments on Coref([-cc]S, alpha, Y) and BVA([-cc]S, X, Y), which are affected by a particular choice of Y, as will be discussed below. 33 The possibility of J=yes in DR([-cc]S, X, beta):J seems to be affected by whether and how easily the sentence can be understood as corresponding to a “categorical-thought” such that the X of the intended DR(X, beta) is understood as the “Subject” of the sentence about which the rest of the sentence expresses a property, in the sense of “Subject” of the “Subject-Predicate” structure in Kuroda’s (1972, 1992) discussion of “categorical judgment”, not in the sense of the “subject” of “subject-object-verb”. I will return to this in Section 8.

182 

 Hajime Hoji

(40) a. DR([+cc]S, subete-no gisi, 3-tu-no robotto): 3-tu-no robotto-o subete-no gisi-ga hihansita three-cl-gen robot-acc all-gen engineer-nom criticized ‘three robots, every engineer criticized’ b.

DR([+cc]S, sukunakutomo 5-nin-no gisi, 3-tu-no robotto) 3-tu-no robotto-o sukunakutomo 5-nin-no three-cl-gen robot-acc at least 5-cl-gen gisi-ga hihansita engineer-nom criticized ‘three robots, at least 5 engineers criticized’

c.

DR([+cc]S, ano gisi igai, 3-tu-no robotto): 3-tu-no robotto-o ano gisi-igai-ga hihansita three-cl-gen robot-acc that engineer-other:than-nom criticized ‘three robots, ones other than that engineer criticized’ (Intended as: (while we may not even know who “falls under” “ones other than that engineer”) each such individual is such that s/he criticized a distinct set of three robots)

Some example sentences instantiating [-cc]S are given in (41). (41)

a.

DR([-cc]S, subete-no gisi, 3-tu-no robotto): 3-tu-no robotto-ga subete-no gisi-o three-cl-gen robot-nom all-gen engineer-acc ‘three robots criticized every engineer’

hihansita criticized

b. DR([-cc]S, sukunakutomo 5-nin-no gisi, 3-tu-no robotto) 3-tu-no robotto-ga sukunakutomo 5-nin-no three-cl-gen robot-nom at least 5-cl-gen gisi-o hihansita engineer-acc criticized ‘three robots criticized at least 5 engineers’ c.

DR([-cc]S, ano gisi igai, 3-tu-no robotto): 3-tu-no robotto-ga ano gisi-igai-o hihansita three-cl-gen robot-nom that engineer-other:than-acc criticized ‘three robots criticized ones other than that engineer’ (Intended as: (while we may not even know who “falls under” “ones other than that engineer”) each such individual is such that a distinct set of three robots criticized him/her)

Detection of C-command Effects 

 183

Among the choices of X for DR(X, beta), NP-igai ‘ones other than NP’ seems to be the best choice of X for me for obtaining ① (with DR(X, beta)) in (30).34 (In the ensuing discussion, when a number inside a circle, such as ① and ②, is mentioned, it is in reference to (30), and I may not always repeat “in (30)”.)

3 Coref(alpha, Y): Identifying Effective Choices of Y for C-command Detection In this initial illustration of my self-experiment, we focus on the three choices for Y for Coref(alpha, Y) and BVA(X, Y), as given in (42). (42)

Choices for Y for BVA(X, Y) and Coref(alpha, Y): a. soko ‘it’ b. soitu ‘that guy’ c. sono otoko ‘that man’

As in the case with the effects of the choice of beta in DR(X, beta), I assume that the choice of alpha of Coref(alpha, Y) does not affect the judgments on the availability of the Coref,(alpha, Y).35 The schemata considered here are the [+cc]S in (43a) and the [-cc]S in (43b), and I check my judgments (J) in (44a) and (44b). (43)

a. b.

S: [ . . . Y . . . ]-o alpha-ga V(erb) S: [ . . . Y . . . ]-ga alpha-o V(erb)

[+cc] [-cc]

(44) a. Coref([+cc]S, alpha, Y):J b. Coref([-cc]S, alpha, Y):J

34 Of the three choices for X being considered here, the “worst” choice for the purpose of c-command detection is (36a). The difference between (36a) and (36c) seems to stem from how easily they can be understood as corresponding to a “Subject” in the sense addressed in footnote 32, which in turn seems to be related to their (in)ability to be used to refer to a specific group of entities, as discussed extensively in the works by Ueyama and Hayashishita. 35 Strictly speaking, this is contrary to fact. For some speakers, including myself, Coref(John, soitu) is more difficult to obtain than Coref(ano gisi, soitu), for example, but it is possible to “overcome” the difficulty. The same goes with BVA(John-igai, Y) and BVA(ano gisi-igai, Y), with Y=soitu/sono otoko. I adopt this assumption here to focus on the issues most pertinent to our current concern, as is the case with the assumption made about the absence of the effects of the choice of beta in DR(X, beta).

184 

 Hajime Hoji

With ano N ‘that N’ as alpha, my judgments on the availability of Coref(alpha, Y) are as summarized below.36 When I have not been judging relevant sentences for a while, my judgments on (43a) and (43b) are as indicated in (45). I refer to this “stage” as “Stage 1”. (45)

My judgments on Coref(alpha, Y) at Stage 1 on the [+cc]S in (43a) and the [-cc] S in (43b):      

alpha

ano N

Y

‘that N’

(42a) (soko)

[+cc] S yes

[-cc] S yes

(42b) (soitu)

yes

no

(42c) (sono otoko)

yes

no

At this stage, I obtain ② (with Coref(alpha, Y)) with Y=soitu and sono otoko, but I do not obtain it with soko. When I have been checking my judgments fairly extensively for a while, I start to obtain judgments as indicated in (46). My judgments at this stage, which we call Stage 2, are as indicated in (46). (46) My judgments on Coref(alpha, Y) at Stage 2 on the [-cc] S in (43b):        alpha Y

S in (43a) and the

[+cc]

ano N ‘that N’ [+cc]

(42a) (soko)

S yes

[-cc] S yes

(42b) (soitu)

yes

yes

(42c) (sono otoko)

yes

no

36 The shifts of judgments as reported here are for my own judgments. It is not claimed that everyone goes through the same shifts of judgments. It is expected, rather, that each speaker undergoes different shifts of judgments, and perhaps (slightly) different shifts of judgments even within a single-speaker. It is possible for a given speaker that a certain choice of Y never induces NFS effects at a given time, but at other times, it does, and it can later “revert” to the stage where that is not possible. The “general direction” of the shifts may not even be the same for a single speaker. The claim made under the correlational methodology is that, despite a great deal of (seemingly intractable) judgmental variations and fluctuation, the definite correlational prediction in (29)/(30) never gets disconfirmed (as long as experiments are conducted in a sufficiently rigorous manner).

Detection of C-command Effects 

 185

At this stage, I obtain ② in (30) only with Y=sono otoko, failing to do so with Y=soko or soitu. Finally, once I have spent even more time checking my judgments, I start getting judgments as indicated in (47). (47) My judgments on Coref(alpha, Y) at Stage 3 on the [-cc] S in (43b):       alpha Y

S in (43a) and the

[+cc]

ano N ‘that N’

(42a) (soko)

[+cc] S yes

[-cc] S yes

(42b) (soitu)

yes

yes

(42c) (sono otoko)

yes

yes

At this stage, I no longer obtain ② in (30), with any of the three Ys given in (42). Given in (48) are some example sentences of [+cc]S. (48) a. Coref([+cc]S, ano kaisya, soko): soko-no robotto-o ano kaisya-ga hihansita it-gen robot-acc that company-nom criticized ‘its robot, that company criticized’ b. Coref([+cc]S, ano gisi, soitu): soitu-no robotto-o ano gisi-ga hihansita that:guy-gen robot-acc that company-nom criticized ‘that guy’s robot, that engineer criticized’ c. Coref([+cc]S, ano gisi, sono otoko): sono otoko-no robotto-o ano gisi-ga hihansita that man-gen robot-acc that company-nom criticized ‘that man’s robot, that engineer criticized’ Some example sentences of [-cc]S are given in (49). (49) a.

Coref([+cc]S, ano kaisya, soko): soko-no robotto-ga ano kaisya-o it-gen robot-nom that company-acc ‘its robot criticized that company’

hihansita criticized

186 

 Hajime Hoji

b. Coref([+cc]S, ano gisi, soitu): soitu-no robotto-ga ano gisi-o that:guy-gen robot-nom that company-acc ‘that guy’s robot criticized that engineer’ c.

hihansita criticized

Coref([+cc]S, ano gisi, sono otoko): sono otoko-no robotto-ga ano gisi-o hihansita that man-gen robot-nom that company-acc criticized ‘that man’s robot criticized that engineer’

Most of the time, for me, sono otoko is the best choice among the different options of Y for Coref(alpha, Y) for obtaining ② in (30).

4 BVA(X, Y): Testable C-command Detection 4.1 Obtaining C-command Detection Consider again (29). (29)

Correlational/conditional prediction about c-command-detection with BVA(X, Y): Provided that DR([+cc]S, X, beta):yes ∧ Coref([+cc]S, alpha, Y):yes ∧ BVA([+cc]S, X, Y):yes; DR([-cc]S, X, beta):no ∧ Coref([-cc]S, alpha, Y):no → BVA([-cc]S, X, Y):no

What is crucially assumed in relation to (29) is (32), repeated here. (32)

a.

X gives rise to NFS effects with BVA(X, Y) only if it gives rise to NFS effects with DR(X, beta).

b. Y gives rise to NFS effects with BVA(X, Y) only if it gives rise to NFS effects with Coref(alpha, Y). From (32), we can see that the use of X that does not give rise to NFS-DR(X, beta) and that of Y that does not give rise to NFS-Coref(alpha, Y) necessarily results in the absence of NFS-BVA(X, Y). That is the prediction in (29)/(30). As discussed above, what choice of X leads to ① and what choice of Y lead to ② can vary for a single speaker, including myself.37 There do not seem to be

37 As does what combination of X and Y leads to ③.

Detection of C-command Effects 

 187

particularly discernable stages as to when (36b) stops being an effective choice of X for obtaining ① for me, especially in relation to DR([-cc]S, X, beta). Which choice(s) of Y are effective for obtaining ② (with Coref(alpha, Y)), on the other hand, seems to be affected by how extensively I have been judging the relevant types of sentences. The correlational prediction about the J of BVA(S, X, Y):J in (29)/(30) is thus tested by making use of whichever X and Y are such that, at the given time, they lead to ① and ②. In relation to ①, we considered the choices of X as in (36) and the [+cc]S and [-cc] S in (37), and my judgments are summarized in (39). (36)

Choices for X for BVA(X, Y) and DR(X, beta): a. subete-no N ‘every N’ b. sukunakutomo #-cl-no N ‘at least # Ns’ c. NP-igai ‘ones other than NP’

(37)

a. b.

S: beta-o X-ga V S: beta-ga X-o V

[+cc] [-cc]

(39) My judgment (J) on the availability of DR(X, beta) on the [+cc]S in (37a) and the [-cc]S in (37b): X beta

(36a)

(36b)

(36c)

subete-no N

sukunakutomo

NP-igai ‘ones

‘every N’

#-cl-no N ‘at

other than NP’

least # Ns’ S yes [+cc]

#-cl-no N

S yes

[-cc]

S yes

[+cc]

S no/yes

[-cc]

S yes [+cc]

S

[-cc]

no

For ②, we considered the choices of Y as in (42) and the [+cc]S and [-cc]S in (43), and my judgments at the three different stages discussed above are summarized in (45)–(47), repeated here. (42) Choices for Y for BVA(X, Y) and Coref(alpha, Y): a. soko ‘it’ b. soitu ‘that guy’ c. sono otoko ‘that man’ (43)

a. b.

S: [ . . . Y . . . ]-o alpha/X-ga V(erb) S: [ . . . Y . . . ]-ga alpha/X-o V(erb)

[+cc] [-cc]

188 

 Hajime Hoji

(45) My judgments on Coref(alpha, Y) at Stage 1 on the [-cc] S in (43b):      alpha Y

S in (43a) and the

[+cc]

ano N ‘that N’ [+cc]

[-cc]

(42a) (soko)

S yes

S yes

(42b) (soitu)

yes

no

(42c) (sono otoko)

yes

no

(46) My judgments on Coref(alpha, Y) at Stage 2 on the [-cc] S in (43b):       alpha Y

S in (43a) and the

[+cc]

ano N ‘that N’ [+cc]

[-cc]

(42a) (soko)

S yes

S yes

(42b) (soitu)

yes

yes

(42c) (sono otoko)

yes

no

(47) My judgments on Coref(alpha, Y) at Stage 3 on the [-cc] S in (43b):     alpha Y

S in (43a) and the

[+cc]

ano N ‘that N’ [+cc]

[-cc]

(42a) (soko)

S yes

S yes

(42b) (soitu)

yes

yes

(42c) (sono otoko)

yes

yes

4.2 Predictions At Stage 1, disconfirmable predictions about the J of BVA(X, Y) are made only with X=soitu or sono otoko, but not with X=soko, because I obtain ② with Coref(alpha, soitu) and Coref(alpha, sono otoko), but not with Coref(alpha, soko) at this stage, as indicated in (45). At Stage 2, we make disconfirmable predictions only with X=sono otoko, because I obtain ② with Y=sono otoko, but not with Y=soko or soitu, as indicated in (46). At Stage 3, we do not make a correlational predic-

Detection of C-command Effects 

 189

tion about c-command detection because I do not obtain ②, at least not with the limited choices of Y given in (42).38 The following predictions are thus made about my judgments on the J in BVA([-cc]S, X, Y):J at each of these Stages, depending upon the choices of Y. We will assume I am using NP-igai ‘ones other than NP’ as X of BVA(X, Y), since it gives the clearest instance of ① for me. (50) Predictions about my J in BVA([-cc]S, NP-igai, Y):J at each of the three Stages”:39        Stages

Stage 1

Stage 2

Stage 3

Y (42a) (soko)

yes

yes

yes

(42b) (soitu)

no

yes

yes

(42c) (sono otoko)

no

no

yes

The disconfirmable predictions at Stages 1 and 2 are indicated by the “no” in (50). If I obtained BVA([-cc]S, NP-igai, soitu/sono otoko):yes at Stage 1 that would constitute disconfirmation of the prediction. Likewise, if I obtained BVA([-cc]S, NP-igai, sono otoko):yes at Stage 2, that would constitute disconfirmation of the prediction. My J in BVA([-cc]S, NP-igai, Y):J at each of the three Stages were in line with the predictions in (50). Hence the predictions are not disconfirmed, which is to say that they have survived attempts at disconfirmation. Furthermore, I obtained

38 One might wonder if there is a choice of Y that leads to ③ (with BVA(X, Y)) at Stage 3 for me. With the so-called overt pronoun kare ‘he’ as Y, I in fact obtain ③ with BVA(NP-igai, kare) at Stage 3. (See Hoji et al. 1999 and reference cited, including Hoji 1991, for relevant discussion.) In order for that to count as a case of c-command detection, it is necessary that I obtain ② with Coref(alpha, kare) (along with ① (with DR(NP-igai, beta). The fact that kare is inherently “referential” at least as the “unmarked option” (for most speakers) (being “D-indexed”, in the terms of Ueyama 1998), however, makes it necessary to force kare to be “non-referential”. The use of aru gisi ‘a certain engineer’, as alpha of Coref(alpha, Y), as in Coref(aru gisi, kare) ‘Coref(a certain engineer, he)’, for example, can serve that purpose (for me) if we add something like “but I don’t know who”, making both aru gisi ‘that engineer’ and kare non-referential. Although further discussion and illustration is not possible here, the relevant considerations about BVA/Coref(X/ alpha, kare) provide further support for validity of the correlational/conditional prediction in (29), not only with regard to the schemata ([-cc]S and [+cc]S) considered so far, but also with regard to additional schemata such as those to be discussed in Sections 7 and 8. 39 For reasons addressed in relation to ⑤ in (30), presenting yes in (50) as a prediction is not accurate and the presentation in (50) is “simplified” for the purpose of exposition.

190 

 Hajime Hoji

BVA([+cc]S, NP-igai, X):yes, with each of the choices in (42) at each of the three Stages in question, as summarized in (51).40 (51) My judgments (J) in BVA(S, NP-igai, Y):J at each of the three Stages:      Stages

Stage 1

Stage 2

Stage 3

Y S yes

[+cc]

(42a) (soko)

S yes

[-cc]

S yes

[+cc]

S yes

[-cc]

S yes [+cc]

S yes [-cc]

(42b) (soitu)

yes

no

yes

yes

yes

yes

(42c) (sono otoko)

yes

no

yes

no

yes

yes

BVA(NP-igai, soitu/sono otoko) at Stage 1 and BVA(NP-igai, sono otoko) at Stage 2 thus lead to c-command detection, for me. In summary, although my judgments on BVA(S, X, Y) are affected by the choices of X and Y, and they do not remain the same even with the same choices of X and Y, definite predictions can still be made about my judgments in the correlational way as indicated in (29), repeated here. (29)

Correlational/conditional prediction about c-command-detection with BVA(X, Y): Provided that DR([+cc]S, X, beta):yes ∧ Coref([+cc]S, alpha, Y):yes ∧ BVA([+cc]S, X, Y):yes; DR([-cc]S, X, beta):no ∧ Coref([-cc]S, alpha, Y):no → BVA([-cc]S, X, Y):no

Given in (52) are some example sentences of [+cc]S with BVA(ano NP-igai, Y). (52)

a.

BVA([+cc]S, ano kaisya-igai, soko): soko-no robotto-o ano kaisya-igai-ga hihansita it-gen robot-acc that company-other:than-nom criticized ‘its robot, ones other than that company criticized’

b. BVA([+cc]S, ano gisi-igai, soitu): soitu-no robotto-o ano gisi-igai-ga hihansita that:guy-gen robot-acc that engineer-other:than-nom criticized ‘that guy’s robot, ones other than that engineer criticized’

40 Prior to the self-experiments that this chapter is based on, I did not always get J=yes in BVA([+cc]S, X, sono otoko), as will be discussed in Section 8.

Detection of C-command Effects 

c.

 191

BVA([+cc]S, ano gisi-igai, sono otoko): sono otoko-no robotto-o ano gisi-igai-ga that man-gen robot-acc that engineer-other:than-nom hihansita criticized ‘that man’s robot, ones other than that engineer criticized’

Some example sentences of [-cc]S are given in (53). (53)

a.

BVA([-cc]S, ano kaisya-igai, soko): soko-no robotto-ga ano kaisya-igai-o it-gen robot-nom that company-other:than-acc ‘its robot criticized ones other than that company’

hihansita criticized

b. BVA([-cc]S, ano gisi-igai, soitu): soitu-no robotto-ga ano gisi-igai-o hihansita that:guy-gen robot-nom that engineer-other:than-acc criticized ‘that guy’s robot criticized ones other than that engineer’ c.

BVA([-cc]S, ano gisi-igai, sono otoko): sono otoko-no robotto-ga ano gisi-igai-o that man-gen robot-nom that engineer-other:than-acc hihansita criticized ‘that man’s robot criticized ones other than that engineer’

5 The Correlational Prediction and the Absence of Other Types of Entailment Although the preceding discussion focuses on the experimental results supporting (i) the effectiveness of the correlational methodology in LFS, which is a universal conceptual/methodological proposal and (ii) the validity of the universal hypotheses and the I-language-particular hypotheses that lead to the correlational predictions, I would like to make the following point: The correlational methodology and the identification of effective choices of X and Y for c-command detection by that method in fact helps us accumulate knowledge about a particular I-language through the checking of additional PS-LF correspondence hypoth-

192 

 Hajime Hoji

eses (and, in addition, through the checking of the validity of more basic hypotheses that lead us to such hypotheses).41 As summarized in (54), I obtain DR([-cc]S, ano gisi-igai, 3-tu-no robotto):no, along with DR([+cc]S, ano gisi-igai, 3-tu-no robotto):yes.42 (54)

(Cf. (39).) My judgment (J) on DR([+cc]S, X, 3-tu-no robotto):J and DR([-cc]S, X, 3-tu-no robotto):J     X beta

subete-no gisi ‘every ano gisi-igai ‘ones other engineer’ S yes [+cc]

3-tu-no robotto

than that engineer’ S yes [-cc]

S yes [+cc]

S

[-cc]

no

‘three robots’

Because LF(ano gisi-igai) does not c-command LF(3-tu-no robotto) in [-cc]S, FRDR(ano gisi-igai, 3-tu-no robotto) is predicted to be impossible in [-cc]S. DR([-cc]S, ano gisi-igai, 3-tu-no robotto):no indicates that, not only FR-DR(ano gisi-igai, 3-tu-no robotto) but also NFS-DR(ano gisi-igai, 3-tu-no robotto) is not possible for me. Assuming that the choice of beta in DR(X, beta) does not affect the possibility of the DR, we are led to conclude that ano gisi-igai as X will not induce NFS effects with DR([-cc]S, X, beta), for me. According to (32a), repeated below, this indicates that using ano gisi-igai as X will not induce NFS effects with BVA([-cc]S, ano gisiigai, Y), for me.43 (32)

a.

X gives rise to NFS effects with BVA(X, Y) only if it gives rise to NFS effects with DR(X, beta).

41 Such investigation would be essentially “syntax research” that pursues rigorous testability (rather than compatibility); see Chapter 1 of this volume for discussion about compatibility-seeking vs. testability-seeking research. 42 Subsequent to the preparation of the submitted draft of the chapter, I came to obtain DR([-cc]S, ano gisi-igai, 3-tu-no robotto):yes. An updated presentation of results of my self-experiments should thus include what choice of beta leads to DR([-cc]S, ano gisi-igai, beta):no at this stage. The judgments reported in this chapter were all at a time when I got DR([-cc]S, ano gisi-igai, 3-tuno robotto):no, corresponding to a slightly different I-language than the one I have when I get DR([-cc]S, ano gisi-igai, 3-tu-no robotto):yes. What is most crucial is that predicted correlations hold with slightly different I-languages. Hoji in preparation contains relevant discussion and provides further articulation and illustration of the correlational methodology. 43 What is crucial here is that the [-cc]S in DR([-cc]S, X, beta), i.e., “beta-ga X-o V” and the one in (55b) are considered “identical” for the purpose at hand.

Detection of C-command Effects 

 193

Without the DR([+cc]S, ano gisi-igai, beta):yes judgment, the J=no in DR([-cc]S, ano gisi-igai, beta):no could be due to factors other than c-command. As summarized in (54), however, my DR([-cc]S, ano gisi-igai, 3-tu-no robotto):no is accompanied by DR([+cc]S, ano gisi-igai, 3-tu-no robotto):yes. This leads us to conclude that the former is indeed due to the absence of c-command, and the combination of DR ([+cc]S, ano gisi-igai, 3-tu-no robotto):yes and DR([-cc]S, ano gisi-igai, 3-tu-no robotto):no is thus taken as a sign of c-command effects.44 Likewise, as summarized in (56), I obtain Coref([-cc]S, ano gisi, sono otoko):no at Stage 2, along with Coref([+cc]S, ano gisi, sono otoko):yes. This leads us to conclude that sono otoko does not induce NFS effects with Coref(alpha, sono otoko) for me at Stage 2 in the [-cc]S in (55b). (55)

(Cf. (43).) For BVA(X, Y) and Coref(alpha, Y): a. [+cc]S: [Y-no N]-o alpha/X-ga V(erb) b. [-cc]S: [Y-no N]-ga alpha/X-o V(erb)

(56)

(Cf. (46).) Stage 2: My judgments (J) on Coref([+cc]S, ano gisi, Y) and Coref([-cc]S, ano gisi, Y):J      alpha Y

ano gisi ‘that engineer’ [+cc]

[-cc]

soitu ‘that guy’

S yes

S yes

sono otoko ‘that man’

yes

no

Being accompanied by my Coref([+cc]S, ano gisi, sono otoko):yes judgment, my Coref([-cc]S, ano gisi, sono otoko):no is due to the absence of c-command.45 The combination of Coref([+cc]S, ano gisi, sono otoko):yes and Coref([-cc]S, ano gisi, sono otoko):no is thus taken as a sign of c-command effects. According to (32b), repeated here, this predicts that sono otoko does not induce NFS effects with BVA(X, sono otoko) in the same [-cc]S for me at this stage. (32) b. Y gives rise to NFS effects with BVA(X, Y) only if it gives rise to NFS effects with Coref(alpha, Y).

44 See the first “conjunct” of the “Provided that . . .” clause in (29). 45 See the second “conjunct” of the “Provided that . . .” clause in (29).

194 

 Hajime Hoji

With ano gisi-igai ‘ones other than that engineer’ not inducing NFS effects with BVA(ano gisi-igai, Y) and with sono otoko not inducing NFS effects with BVA(X, sono otoko), we are led to conclude that BVA([-cc]S, ano gisi igai, sono otoko) cannot arise due to NFS. Since FR-based BVA(X, Y) is not possible in [-cc]S, we now have a prediction that BVA(ano gisi igai, sono otoko) is not possible in [-cc]S for me at this stage.46 If I indeed get BVA([-cc]S, ano gisi-igai, sono otoko):no, as predicted, along with BVA([+cc]S, ano gisi-igai, sono otoko):yes, that will result in c-command detection with BVA(ano gisi-igai, sono otoko). My judgments at Stage 2, summarized in (57), indeed confirm the BVA([-cc]S, ano gisi-igai, sono otoko):no prediction. (57)

(Cf. (56).) Stage 2: My judgments (J) on BVA([+cc]S, ano gisi-igai, Y):J and BVA([-cc]S, ano gisi-igai, Y):J       X Y

ano gisi igai ‘ones other than that engineer’ [+cc]

S yes

[-cc]

soitu ‘that guy’

S yes

sono otoko ‘that man’

yes

no

Furthermore, with the BVA([+cc]S, ano gisi-igai, sono otoko):yes judgment, I obtain c-command detection with BVA(ano gisi-igai, sono otoko) at Stage 2. As noted, the use of soitu ‘that guy’ as Y can also result in c-command detection for me at Stage 1 though not at Stage 2. In the immediately preceding illustration, we have considered only two choices of X and two choices of Y for BVA(X, Y), DR(X, beta) and Coref(alpha, Y). As suggested in (19) and (20) and by the remarks thereabout, there are many more choices of X and Y that we can consider and that I have in fact considered. Whether we obtain (testable) c-command detection with BVA(X, Y) depends not only upon the choices of X and Y but also upon the speaker in question and the particular point in time for the speaker judging the relevant sentences. By using the correlational methodology, as summarized in (29), we can nonetheless deduce definite predictions, and obtain and replicate experimental results precisely in line with the predictions, despite the great deal of “uncertainty” just noted.

46 NFS effects with BVA(X, Y) can also arise, at least for some people, including myself, when the head N of X and that of Y are “identical”. Since NP-igai, unlike NP-igai-no N ‘N other than NP’, does not have a head N, this type of NFS effects do not arise with BVA(NP-igai, Y).

Detection of C-command Effects 

 195

As a further illustration of the entailment in (29), I would now like to consider why other types of entailments, such as (58a) and (58b), do not hold.47 (58) Conceivable entailments that do not hold: a. Provided that DR([+cc]S, X, beta):yes; DR([-cc]S, X, beta):no → BVA([-cc]S, X, Y):no b. Provided that Coref([+cc]S, alpha, Y):yes; Coref([-cc]S, alpha, Y):no → BVA([-cc]S, X, Y):no The absence of the entailment in (58a) and (58b) is due to (59a) and (59b), respectively. (59) a. A particular choice of X not inducing NFS effects with DR(X, beta) does not guarantee the absence of NFS-BVA(X, Y) because a certain choice of Y might induce NFS effects with BVA(X, Y). b. A particular choice of Y not inducing NFS effects with Coref(alpha, Y) does not guarantee the absence of NFS-BVA(X, Y) because a certain choice of X might induce NFS effects with BVA(X, Y). I always obtain DR([-cc]S, ano gisi-igai, 3-tu-no robotto):no (along with DR([+cc]S, ano gisi-igai, 3-tu-no robotto):yes. This, however, does not guarantee that I always obtain BVA([-cc]S, ano gisi-igai, Y):no, regardless of the choice of Y. Even with the “best choice” of X for c-command detection with BVA(X, Y) in my I-language, i.e., ano gisi-igai ‘ones other than that engineer’, choosing an NFS-inducing Y can result in NFS-BVA. I in fact obtain BVA([-cc]S, ano gisi-igai, soitu):yes at Stage 2. As to (59b), recall that I that get Coref([-cc]S, ano gisi, sono otoko):no (along with Coref([+cc]S, ano gisi, sono otoko):yes) at Stage 2. This, however, does not guarantee that I always obtain BVA([-cc]S, X, sono otoko):no, regardless of the

47 If we wanted to make (58) look more like (29), we could add “and BVA([+cc]S, X, Y):yes” in the “Provided that . . .” clause. The entailments in (58a) and (58b) (which do not hold) would then become (i) and (ii), respectively, which would be the (wrong) prediction about c-command detection with BVA(X, Y). (i) If we obtain ① (with DR(X, beta)), we obtain ③ (with BVA(X, Y)). (ii) If we obtain ② (with Coref(alpha, Y)), we obtain ③ (with BVA(X, Y)). Similar remarks apply to (60) below.

196 

 Hajime Hoji

choice of X. With X=subete-no otoko ‘every man’, for example, I obtain BVA([-cc]S, subete-no otoko, sono otoko):yes.48 Similar considerations point to the absence of entailments in (60). (60) We do not have entailments such as: a. Provided that BVA([+cc]S, X, Y):yes; BVA([-cc]S, X, Y):no → DR([-cc]S, X, beta):no b. Provided that BVA([+cc]S, X, Y):yes; BVA([-cc]S, X, Y):no → Coref([-cc]S, alpha, Y):no For example, at Stage 2, I obtain judgments as indicated in (61). (61)

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

BVA([+cc]S, subete-no gisi, sono otoko):yes BVA([-cc]S, subete-no gisi, sono otoko):no DR([+cc]S, subete-no gisi, 3-tu-no robotto):yes DR([-cc]S, subete-no gisi, 3-tu-no robotto):yes Coref([+cc]S, ano otoko, sono otoko):yes Coref([-cc]S, ano otoko, sono otoko):yes

(61a) and (61b) satisfy the “Provided that . . .” clause and the antecedent in (60a). (61d), however, indicates that the consequent in (60a) does not necessarily hold, because subete-no gisi ‘every engineer’ (and presumably, subete-no N, in general) induces NFS effects with DR, but does not seem to necessarily induce NFS effects with BVA. Turning to the entailment in (60b), its failure to hold is indicated by (61a), (61b), and (61f), combined. We see in (61f) that the use of sono otoko ‘that man’, by itself, does not necessarily exclude NFS effects; see Section 8 for further discussion, where I address different types of NFS.

48 As noted before, what choices of X and Y lead to (testable) c-command detection varies among speakers and also among different points in time. Likewise, what choices of X and Y would effectively illustrate the absence of the entailments in (59) vary among speakers and also among different points in time.

Detection of C-command Effects 

 197

6 Contrapositive Consider again (29). (29)

Correlational/conditional prediction about c-command-detection with BVA(X, Y): Provided that DR([+cc]S, X, beta):yes ∧ Coref([+cc]S, alpha, Y):yes ∧ BVA([+cc]S, X, Y):yes; DR([-cc]S, X, beta):no ∧ Coref([-cc]S, alpha, Y):no → BVA([-cc]S, X, Y):no

If the antecedent clause of the conditional in (29) is not satisfied, for example if I obtain DR([-cc]S, X, beta):yes or Coref([-cc]S, alpha, Y):yes, BVA([-cc]S, X, Y):no (the consequent of (29)) is not entailed, even if the “Provided . . .” clause is satisfied. What judgment I get on BVA([-cc]S, X, Y):J in that case is, therefore, of no significance with regard to the evaluation of our prediction. In my self-experiment, many choices of X and Y do not result in ① (with DR(X, beta)) or ② (with Coref(alpha, Y)) and in fact only a few choices do.49 That means that we perhaps have limited combinations of X and Y that allow us to actually test our predictions in the form of (29)/(30) even if we consider many more choices for X and Y. If we happen to choose a “wrong” pair of X and Y, we may in fact end up being unable to test our predictions in the form of (29). From the perspective of seeking testability, this may be a source of concern, as some people has pointed out over the years. The contrapositive of (29), as in (62), however, provides us with a means to test our predictions in a different way.50 (62)

The contrapositive of (29): Provided that DR([+cc]S, X, beta):yes ∧ Coref([+cc]S, alpha, Y):yes ∧ BVA([+cc]S, X, Y):yes; BVA([-cc]S, X, Y):yes → DR([-cc]S, X, beta):yes ∨ Coref([-cc]S, alpha, Y):yes

Unlike the testing of (29), which is based on DR([-cc]S, X, beta):no and Coref([-cc]S, alpha, Y):no, the testing of (62) is based on BVA([-cc]S, X, Y):yes, which, according to the preceding discussion, must be based on NFS. The NFS effects with BVA(X, Y), 49 Based on my past experiences of conducting non-self-experiments, this also seems to be the case with other speakers as well. Chapter 6 of this volume contains relevant discussion. 50 I would like to thank Daniel Plesniak for the extensive discussion about the relevant issues at least since December of 2019, which has eventually led me to adopt the correlational/conditional prediction of the form in (29), which makes it straightforward to address its contrapositive. See Chapter 4 of this volume: Section 5.

198 

 Hajime Hoji

as indicated in the antecedent clause in (62), can be due to the choice of X or that of Y or both.51 As a result, it predicts that X must induce NFS effects (with DR(X, beta)), Y must induce NFS effects (with Coref(alpha, Y)), or both. The prediction in (62) is that BVA([-cc]S, X, Y):yes is necessarily accompanied by DR([-cc]S, X, beta):yes, Coref([-cc]S, alpha, Y):yes, or both. Recall that MR([-cc]S, X, Y):yes is due to NFS effects. The contrapositive in (62) thus provides us with a means to study what gives rise to NFS effects, in addition to giving us plenty of opportunities to directly test (29) in its contrapositive form in (62). The prediction in the form of (62) has been supported in every single relevant case (of BVA([-cc]S, X, Y):yes) in my self-experiment.52 Checking predictions in the form of (62) has in fact resulted in a significant advancement of our understanding about what contributes to the nature of NFS effects and what gives rise to them, as will be discussed briefly in Section 8.53

7 Expanding the Empirical Coverage 7.1 Different “Case Marking” Patterns As remarked in note 11, what “case markers” (such as -ga, -o, -ni) are used for NPs with a particular type of verb is determined in part by “lexical specification” of the type of verb in question. Hihansuru ‘to criticize (given here as its “dictionary form”), for example, is the type of verb that “marks” its direct object NP with -o. This is the type of verb that we have been focusing on in the preceding discussion. The initial empirical discussion in this chapter has been limited to sentences of the Subject-Object-Verb and Object-Subject-Verb forms in Japanese, with this type of verb. See (52) and (53) for some examples. I now expand the empirical coverage by considering sentences with a transitive verb that “marks” its object NP with -ni (instead of -o). Renrakusuru ‘to contact’ (also given here as its “dictionary” form), is the type of verb that “marks” its object

51 It can also be due to a combination of X and Y, as discussed in the preceding section. To simplify the discussion, however, let us assume that that NFS effects due to a combination of X and Y are excluded, by making sure that the “N-head identity” requirement is not satisfied. 52 We will see in Chapter 6 of this volume that the same holds in a large-scale non-self-experiment. 53 This includes the recognition that there are (at least) two types of NFS, each having distinct sets of conditions, leading to a new understanding, empirically as well as theoretically, as will be addressed briefly in Section 8.

Detection of C-command Effects 

 199

NP with -ni, as illustrated in (63) and (64), which are the renrakusuru versions of (52) and (53), with the same choices of X and Y. (63) a. BVA([+cc]S, ano kaisya-igai, soko): soko-no robotto-ni ano kaisya-igai-ga renrakusita it-gen robot-dat that company-other:than-nom contacted ‘its robot, ones other than that company contacted’ b. BVA([+cc]S, ano gisi-igai, soitu): soitu-no robotto-ni ano gisi-igai-ga renrakusita that:guy-gen robot-dat that engineer-other:than-nom contacted ‘that guy’s robot, ones other than that engineer contacted’ c.

(64) a.

BVA([+cc]S, ano gisi-igai, sono otoko): sono otoko-no robotto-ni ano gisi-igai-ga that man-gen robot-dat that engineer-other:than-nom renrakusita contacted ‘that man’s robot, ones other than that engineer contacted’ BVA([-cc]S, ano kaisya-igai, soko): soko-no robotto-ga ano kaisya-igai-ni renrakusita it-gen robot-nom that company-other:than-dat contacted ‘its robot contacted ones other than that company’

b. BVA([-cc]S, ano gisi-igai, soitu): soitu-no robotto-ga ano gisi-igai-ni that:guy-gen robot-nom that engineer-other:than-dat renrakusita contacted ‘that guy’s robot contacted ones other than that engineer’ c.

BVA([-cc]S, ano gisi-igai, sono otoko): sono otoko-no robotto-ga ano gisi-igai-ni that man-gen robot-nom that engineer-other:than-dat renrakusita contacted ‘that man’s robot contacted ones other than that engineer’

My judgments on the availability of the BVA in these sentences are exactly the same as in the case of (52) and (53). They “shift” in the same way as my judgments “shift” in the case of (52) and (53). In other words, the correlational/conditional

200 

 Hajime Hoji

prediction in (29) and its contrapositive in (62) are both supported experimentally in my self-experiment, not only with regard to (37) and (43), repeated here, but also with regard to (65) and (66), with various combinations of X and Y. (37)

a. b.

[+cc]

(43)

a. b.

[+cc]

(65)

a. b.

[+cc]

(66) a. b.

[+cc]

S: beta-o X-ga V S: beta-ga X-o V

[-cc]

S: [ . . . Y . . . ]-o alpha/X-ga V(erb) S: [ . . . Y . . . ]-ga alpha/X-o V(erb)

[-cc]

S: beta-ni X-ga V S: beta-ga X-ni V

[-cc]

S: [ . . . Y . . . ]-ni alpha/X-ga V(erb) S: [ . . . Y . . . ]-ga alpha/X-ni V(erb)

[-cc]

This means that (13) or its restatement in (15), along with (14), should be generalized to cover both (65) and (66), in addition to (37) and (43). I repeat (13), (14), and (15) for convenience. (13)

PS-LF correspondence hypotheses: a. NP2-ga NP1-o V in Japanese must correspond to a 3D (i.e., LF) representation where LF(NP2) asymmetrically c-commands LF(NP1), as in (14). b. NP1-o NP2-ga V in Japanese can correspond to a 3D (i.e., LF) representation where LF(NP2) asymmetrically c-commands LF(NP1), as in (14).

(14)

a. NP2 NP1

V

b. NP2

NP2

NP2

NP1

V

V

NP1

V

NP1

Detection of C-command Effects 

(15)

 201

a. The 3D (i.e., LF) representation as in (14) can be “externalized” as “NP2-ga NP1-o V” (in the Subject-Object-Verb order) or as “NP1-o NP2-ga V” (in the Object-Subject-Verb order). b. While “NP2-ga NP1-o V” (of the Subject-Object-Verb order) must correspond to the 3D (i.e., LF) representation as in (14), “NP1-o NP2-ga V” (of the Object-Subject-Verb order) can correspond to it.54

We can, for example, restate (13) and (15) as in (67) and (68), respectively, for this purpose. (67) PS-LF correspondence hypotheses, where cm= “case marker” -o or -ni: a. NP2-ga NP1-cm V in Japanese must correspond to an LF representation where LF(NP2) asymmetrically c-commands LF(NP1), as in (14). b. NP1-cm NP2-ga V in Japanese can correspond to an LF representation where LF(NP2) asymmetrically c-commands LF(NP1), as in (14). (68) where cm= “case marker” -o or -ni: a. The LF representation in (14) can be “externalized” as “NP2-ga NP1-cm V” (in the Subject-Object-Verb order) or as “NP1-cm NP2-ga V” (in the Object-Subject-Verb order). b. While “NP2-ga NP1-cm V” (of the Subject-Object-Verb order) must correspond to the LF representation in (14), “NP1-cm NP2-ga V” (of the Object-Subject-Verb order) can correspond to it. We can also articulate this as in (69). (69) Considering the 3D (i.e., LF) representation below, where LF(NP2) c-commands LF(NP1):

NP2

NP2

NP2

NP1 54 See note 13.

V

V

NP1

V

NP1

202 

 Hajime Hoji

a.  It can be “externalized” in the linear order of “NP2 NP1 V” or “NP1 NP2 V”.55 b. With the “o-marking transitive verb”, NP1 and NP2 appear with -o and -ga, respectively. c.  With the “ni-marking transitive verb”, NP1 and NP2 appear with -ga and -ni, respectively.56 In addition to sentences of the regular “o-marking transitive verbs” and “nimarking transitive verbs”, there are sentences that have been said to have the so-called “ergative” case-marking pattern, such as (70), as discussed in Kuroda 1978 ; see Mukai’s chapter 3 of this volume, Section 6.4 for more details.57 (70)

a.

John-ni ano robotto-ga mieta John-dat that robot-nom was:visible ‘John saw that robot’ ‘that robot was visible to John’

b. John-ni okane-ga John-dat money-nom ‘John needs money’

iru. need

Just as with the “transitive” verb that marks its object with -ni (see (63) and (64)), two surface orders are possible between NP-ni and NP-ga with the ergative verb; in addition to the NP-ni NP-ga order in (70), the NP-ga NP-ni order is also possible. In other words, (71a) and (71b) are both possible for either a “transitive ni-marking” verb or the ergative verb. (71)

a. NP-ga NP-ni V b. NP-ni NP-ga V

In the case of the “transitive ni-marking” verb, we have said that it is NP1 in the sense of (69) that is marked with -ni. Based on the results of my self-experiment, 55 The “NP2 NP1 V” order is assumed to be the result of DeMerge of Fukui and Takano (1998) (along with the “head-initial” property of I-languages of JP), and the “NP1 NP2 V” order is assumed to be due, in addition to DeMerge, to a PF operation as in Ueyama’s analysis of the Surface OS type. 56 We are restricting our discussion to cases where NP1 and NP2 are both “phonetically realized”. 57 See Kuroda 1992 [1978]: p. 238, note 3 for concern expressed by Kuroda for the use of “ergative” for this type of verb. I am using “ergative” here with a similar concern, adopting it as a descriptive term without accepting any particular analysis of it as parallel to other so-called “ergatives” in other languages.

Detection of C-command Effects 

 203

however, it seems that in the case of the “ergative verb”, it is in fact NP2 in the sense of (69) that is marked with -ni.58 We thus modify (69) by adding (72) “as (69d)”. (72)  (To be added to (69), as (69d)) With the “ergative verb”, NP1 and NP2 appear with -ga and -ni, respectively.59 With the modification just introduced, we make the standard correlational/ conditional prediction in the form of (29) and its contrapositive in (62), with the [+cc] S and the [-cc]S for ergative verbs given in (73) and (74), as opposed to those for transitive -ni-marking verbs, as given in (75) and (76). (29) Correlational/conditional prediction about c-command-detection with BVA(X, Y): Provided that DR([+cc]S, X, beta):yes ∧ Coref([+cc]S, alpha, Y):yes ∧ BVA([+cc]S, X, Y):yes; DR([-cc]S, X, beta):no ∧ Coref([-cc]S, alpha, Y):no → BVA([-cc]S, X, Y):no (62) The contrapositive of (29): Provided that DR([+cc]S, X, beta):yes ∧ Coref([+cc]S, alpha, Y):yes ∧ BVA([+cc]S, X, Y):yes; BVA([-cc]S, X, Y):yes → DR([-cc]S, X, beta):yes ∨ Coref([-cc]S, alpha, Y):yes (73) With an ergative verb a. [+cc]S: beta-ga X-ni V b. [-cc]S: beta-ni X-ga V

(74) With an ergative verb a. [+cc]S: [ . . . Y . . . ]-ga alpha/X-ni V b. [-cc]S: [ . . . Y . . . ]-ni alpha/X-ga V 58 This result is consistent with the thesis put forth in the past literature (Kuroda 1978, 1986) that “NPni NP-ga V” is the “base order” and “NP-ga NP-ni V” is the “derived order” of sentences with ergative verbs. This seems to apply also to the “existential sentence pattern” of “NP2-ni NP1-ga aru” (there is NP1 in NP2), as also discussed in Kuroda 1978. The results of my self-experiment in support of (72) can thus be understood as a compelling demonstration, by the basic scientific method, of the validity of the thesis put forth in the past literature, assuming we take the “base order” to be the one wherein linear precedence relationships match c-command relationships (at least between the two NPs (and what they correspond to in LF representations)). As such, we expect it to be replicated in non-self-experiments (in the form of a “demonstration” of the sort to be discussed in Chapter 6 of this volume). 59 See note 55.

204 

 Hajime Hoji

(75) With a “transitive ni-marking” verb (cf. (65)) a. [+cc]S: beta-ni X-ga V b. [-cc]S: beta-ga X-ni V (76) With a “transitive ni-marking” verb (cf. (66)) a. [+cc]S: [ . . . Y . . . ]-ni alpha/X-ga V b. [-cc]S: [ . . . Y . . . ]-ga alpha/X-ni V Notice that the [+cc]S and the [-cc]S in (75) are reversed in (73), and those in (76) are reversed in (74). I obtain judgments exactly as predicted, by (73), (74), (75), and (76). For a brief illustration, consider (77). (77)

a. NP2-ga NP1-o hituyoo to site iru ‘NP2 needs NP1’ b. NP2-ni NP1-ga hituyoo da ‘NP2 needs NP1’ ‘NP1 is necessary for NP2’

As indicated, the two sentences in (77) express very similar, in fact almost identical, meanings, at least with regard to who is needing what. Once we start considering the availability of MRs, such as BVA, DR, and Coref, however, very different (in fact completely opposite) patterns of judgments emerge, precisely as predicted. Let us consider BVA(ano gisi-igai, sono otoko) ‘BVA(ones other than that engineer, that man)’, DR(ano gisi-igai, 3-tu-no robotto) ‘DR(ones other than that engineer, three robots)’, and Coref(ano gisi, sono otoko) ‘Coref(that engineer, that man)’, where the best choices of X and Y, among the choices considered above at Stage 2, for me are used for c-command detection; see (36) and (42). With X=ano gisi-igai ‘ones other than that engineer’ and Y=sono otoko ‘that man’, I obtain ①, ②, and ③, crucially with the [+cc]S and the [-cc]S in (78) (for DR) and (79) (for Coref and BVA), with a “transitive o-marking” verb, such as hituyoo to site iru ‘needs’ in (77a). (78)

With a “transitive o-marking” verb a. [+cc]S: beta-o X-ga V b. [-cc]S: beta-ga X-o V

(79)

With a “transitive o-marking” verb a. [+cc]S: [ . . . Y . . . ]-o alpha/X-ga V b. [-cc]S: [ . . . Y . . . ]-ga alpha/X-o V

Likewise, with these same choices of X and Y, I also obtain ①, ②, and ③ with an ergative” verb, such as hituyoo da ‘needs’ in (77b), but this time, crucially with the

Detection of C-command Effects 

 205

S and the [-cc]S in (73) (for DR) and (74) (for Coref and BVA), repeated here, not with the [+cc]S and the [-cc]S in (78) and (79).

[+cc]

(73)

With an ergative verb a. [+cc]S: beta-ga X-ni V b. [-cc]S: beta-ni X-ga V

(74)

With an ergative verb a. [+cc]S: [ . . . Y . . . ]-ga alpha/X-ni V b. [-cc]S: [ . . . Y . . . ]-ni alpha/X-ga V

As noted before, I also obtain ①, ②, and ③ with the [+cc]S and the [-cc]S in (75) (for DR) and (76) (for Coref and BVA), repeated here, with a transitive ni-marking verb, such as renrakusuru ‘contact’; see (63) and (64). (75)

With a “transitive ni-marking” verb a. [+cc]S: beta-ni X-ga V b. [-cc]S: beta-ga X-ni V

(76)

With a “transitive ni-marking” verb a. [+cc]S: [ . . . Y . . . ]-ni alpha/X-ga V b. [-cc]S: [ . . . Y . . . ]-ga alpha/X-ni V

With many other combinations of X and Y, at various stages, I do not obtain ③ (with BVA(X, Y)). Suppose that the failure to obtain ③ is due to BVA([-cc]S, X, Y):yes. In such cases, the contrapositive in (62) is always supported by results of my self-experiment. That is, with the particular choices of X and Y that lead to BVA([-cc]S, X, Y):yes for me at a given time, I always obtain DR([-cc]S, X, beta) or Coref([-cc]S, alpha, Y) or both at that time. With certain ergative verbs other than the ones discussed thus far, there are complications, albeit ones our methodology allows us to handle. Consider (80b) (in contrast to (80a)), for example. (80) a.

NP2-ga NP1-o mita. ‘NP2 saw NP1’ ‘NP 2 looked at NP1’

206 

 Hajime Hoji

b.  NP2-ni NP1-ga mieta.60 ‘NP2 saw NP1’ ‘NP1 was visible to NP2’ I do obtain ①, ② and ③ with mieta, as in the ergative verb case discussed above, with the [+cc]S and the [-cc]S as in (81) and in (82).61 (81)

(Cf. (73).) With “can see” mieta a. [+cc]S: beta-ga X-ni mieta b. [-cc]S: beta-ni X-ga mieta

(82)

(Cf. (74).) With “can see” mieta a. [+cc]S: [ . . . Y . . . ]-ga alpha/X-ni mieta b. [-cc]S: [ . . . Y . . . ]-ni alpha/X-ga mieta

However, one complication with mieta is that it allows the form in (83a), where we have an “extra” NP-ni, in addition to NP2-ni and NP1-ga in (80b). (83)

a.

NP2-ni NP1-ga NP-ni mieta. ‘NP2 saw NP1 as NP’ ‘NP1 appeared to NP2 to be like NP’ b. (Cf. (a).) NP1-ga NP-ni mieta ‘NP1 appeared (to someone) like NP”

If NP2 in (83a) is missing phonetically, we have the form in (83b), which looks identical to the “NP1-ga NP2-ni mieta” (i.e., the OSV) version of (80b). We can ask whether the sentence with the mieta with an extra NP-ni, as in (83b) “behaves like” the SOV or OSV. Depending upon the answer, we have either (84) and (85), on the one hand, or (86) and (87), on the other hand. To distinguish this “type” of mieta from the type in (81) and (82), I call it “appear like” mieta in (84)–(87).

60 The relevant case marking properties of an ergative” verb are perhaps related to the “potential” and “spontaneous” (zihatu) forms of the verb in question, but we cannot address the relevant issues further here. Some, if not many, of the case-marking alternation issues mentioned above and the case-marking mechanisms needed are discussed in some depth in past literature. It is hoped that research focusing on c-command detection by the basic scientific method, as is being pursued here, will shed new light on the relevant issues (among many others). 61 I have added ‘can see” before mieta in (81) and (82) in anticipation of another “type” of mieta to be discussed immediately.

Detection of C-command Effects 

 207

(84) With “appear like” mieta, possibility 1 a. [+cc]S: beta-ga X-ni mieta b. [-cc]S: beta-ni X-ga mieta (85)

With “appear like” mieta, possibility 1 a. [+cc]S: [ . . . Y . . . ]-ga alpha/X-ni mieta b. [-cc]S: [ . . . Y . . . ]-ni alpha/X-ga mieta

(86) With “appear like” mieta, possibility 2 a. [+cc]S: beta-ni X-ga mieta b. [-cc]S: beta-ga X-ni mieta (87)

With “appear like” mieta, possibility 2 a. [+cc]S: [ . . . Y . . . ]-ni alpha/X-ga mieta b. [-cc]S: [ . . . Y . . . ]-ga alpha/X-ni mieta

As discussed above, the two possibilities (i.e., (84)+(85) and (86)+(87)) make radically different predictions. The checking of my judgments on the relevant sentences with BVA, DR, and Coref, with various options for X and Y, clearly indicate that we should adopt (86)+(87). Under (86)+(87), I obtain ①, ②, and ③, hence c-command detection with BVA(X, Y), with the best choices of X and Y (for c-command detection); likewise, with various other combinations of X and Y, I also obtain results that are in line with the correlational/conditional prediction in (29)/(30), and hence with its contrapositive in (62).62 As discussed, definite/categorical experimental results come in the form of correlation of judgments. When we focus on judgments on a particular MR (such as BVA, DR, and Coref, for example) and with particular choices of X and Y, my judgments are not always stable, but the “murky” judgments also correlate across different MRs, across types of verbs, etc., and how they become clear also correlate across different MRs, across types of verbs, etc. We are here focusing on my self-experiments, dealing with particular I-language(s) of mine. The fact that I obtain judgments precisely in line with the correlational/conditional prediction and its contrapositive in all these different constructions and with different combinations of X and Y points to the general effectiveness of the correlational approach pursued and its usefulness in accumulating our knowl62 Although it is not entirely clear how we can “derive” this, we must assume that the extra “NP(-ni)” (not “NP1-ga”) in (83) is merged with the Verb, probably not because it is the “internal argument” of the verb but perhaps it is part of the predicate itself (which can somehow be separated from the verb at PF).

208 

 Hajime Hoji

edge about the language faculty, not only about its initial state but also about its steady states.63

7.2 The [-loc] Requirement on FD(x, y) The hypothesis in (7), repeated here, is the only hypothesis about FRs that we have considered.64 (7) FR(x, y) is possible only if x c-commands y. As discussed in Chapter 4, we can understand that our research is an attempt to accumulate knowledge about the initial state of the language faculty by studying FR(x, y) through our investigation of FR-based-MR(X, Y) in a particular I-language, and this endeavor requires that we control for NFS effects. The adoption of the correlational methodology serves to achieve this control of NFS effects, as discussed above. When we addressed FR-based-BVA, FR-based-DR, and FR-based-Coref in the preceding discussion, we did not consider the possibility of different types of FR’s. As it turns out, however, there do seem to be different types of FR’s. Restricting our discussion just to BVA, DR, and Coref, there seems to be one particular FR(x, y) that underlies FR-based BVA(X, Y) and Coref(alpha, Y) (but not DR(X, beta). This particular FR(x, y) seems to have an additional condition on top of x c-commanding y, i.e., the condition that x and y cannot be “co-arguments” of the same predicate/verb. For example, x and y cannot be the subject and the object (respectively) of the same verb. The FR that underlies FR-based DR does not have this property. Let us refer to FR(x, y) that can underlie BVA(X, Y) and Coref(alpha, Y) as FD(x, y) and the FR(x, y) that can underlie DR(X, beta) as DD(x, y).65 To reiterate, the structural conditions on FD are as given in (88). 63 I have replicated my judgments in several non-self-experiments on my researcher-colleagues regarding a “transitive ni-marking” verb vs. an ergative verb, although not regarding the case of complication just discussed. 64 The thesis that the correlational/conditional prediction about c-command detection and its contrapositive are an effective means to accumulate knowledge about the language faculty by the basic scientific method can be considered as a conceptual universal hypothesis. What the LFStist tries to replicate in non-self-experiment dealing with speakers of a language (substantially) different from his/her own is the validity of this conceptual universal hypotheses, along with hypotheses about FRs such as (7). 65 “FD” is an abbreviation of “Formal Dependency” and the term has been used since the mid 1990s; it was used in Hoji 1995, for example. “DD” is an abbreviation of “Distributive Dependency”, modeled after FD but modified to reflect its relationship to DR.

Detection of C-command Effects 

 209

(88) FD(x, y) is possible only if: (i) x c-commands y (ii) x and y are not co-arguments Under (88), (89b) is a schema where FD-BVA66/Coref(X/alpha, Y) is not possible, as compared to (89a), where it is, precisely because X and Y are not co-arguments in (89a) but are co-arguments in (89b). What is indicated by “[+loc(al)]” is that LF(X/ alpha) and LF(Y) are in a “local relation” as they are “co-arguments” of the V. (89)  a. [+cc, -loc]S: [NP1 . . . Y . . .] [NP2 X/alpha] V67 b. [+cc, +loc]S: [NP1 Y ] [NP2 X/alpha] V We thus obtain a “schematic contrast” not only between (90a) and (90b), as before, where one schema satisfies the c-command condition on FD and the other does not, but also between (89a) and (89b), where both schemata satisfy the c-command condition on FD but one does not satisfy the “anti-locality” condition on FD(x, y) in (88-ii).68

66 “FD-BVA” abbreviates “FD-based BVA”. 67 A subtype of (89a) is (i), where NP1 consists of Y + -no (the “genitive marker”) + the head Noun. (i)

S: [NP1 Y-no N] [NP2 X/alpha] V

[+cc, -loc]

Because the choice between (89a) and (i) does not result in any significant difference in my judgments, I generally consider the simpler version of the two in my self-experiments, i.e., (i). Likewise, I generally consider the “Y-no N” subtype of (90a) and (90b), as I have in fact done in the preceding discussion. 68 An actual sentence instantiating (90a) and one instantiating (90b) are provided in (i) and (ii), respectively. (i)

(A sentence instantiating (90a)) soitu-no robotto-o subte-no gisi-ga suisensita that:guy-gen robot-acc every-gen engineer-nom recommended ‘that guy’s robot, every engineer recommended’ (ii) (A sentence instantiating (90b)) soitu-o subte-no gisi-ga suisensita that:guy-acc every-gen engineer-nom recommended ‘that guy, every engineer recommended’ See Hoji 2016: Sections 2 and 3 for some general remarks about related issues, as addressed in the papers collected in Hoji 2013. See also Chapter 1 of this volume: Sections 2 and 3. Hoji in preparation will illustrate how a breakthrough may be possible in the relevant area of investigation by the correlational methodology proposed here.

210  (90) a. b.

 Hajime Hoji

S: [NP1 . . . Y . . . ] [NP2 X/alpha] V S: [NP2 . . . Y . . . ] [NP1 X/alpha] V

[+cc, -loc] [-cc, -loc]

Recall our earlier discussion, when we had not yet sub-divided FR into FD and DD; at that point, we had only one structural condition on FR(x, y), namely that must x c-command y. If MR([-cc]S, X, Y):yes obtained, we regarded it as being an instance of NFS-MR(X, Y), rather than as an instance of FR-MR(X, Y). With regard to BVA(X, Y) and Coref(alpha, Y), we considered BVA([-cc]S, X, Y):yes and Coref([-cc]S, alpha, Y):yes instances of NFS-BVA and NFS-Coref, respectively. In the same vein, now that we have introduced the anti-locality condition on FD, we can consider BVA([+cc, +loc]S, X, Y):yes and Coref([+cc, +loc]S, alpha, Y):yes instances of NFS-BVA and NFS-Coref, respectively. As discussed above, the correlational/conditional prediction about c-command-detection with BVA(X, Y), involving two other MRs (DR(X, beta) and Coref(alpha, Y)), is meant to control for NFS effects by identifying a choice of X and a choice of Y that do not seem to induce NFS effects with DR or with Coref. We used the X and Y thus identified when considering BVA(X, Y). With (88b), and with the notation of [±loc], we can now consider the correlational/conditional prediction in (91), where we try to identify the choice of Y that does not seem to give rise to NFS effects with Coref and use a Y thus identified, along with an X that has been determined with the DR test as not inducing NFS effects with DR, to test the correlational/conditional prediction. (91) Correlational/conditional prediction with BVA(X, Y), with reference to [±loc]: Provided that DR([+cc]S, X, beta):yes ∧ Coref([+cc, -loc]S, alpha, Y):yes ∧ BVA([+cc, -loc]S, X, Y):yes; DR([-cc]S, X, beta):no ∧ Coref([+cc, +loc]S, alpha, Y):no → BVA([+cc, +loc]S, X, Y):no

(91) is not, at least in a direct way, a prediction about c-command detection with BVA(X, Y), because BVA([+cc, +loc]S, X, Y):no cannot be due to the absence of c-command because it is about [+cc, +loc]S, where X does c-command Y. This contrasts with BVA([-cc]S, X, Y):no, where we have [-cc]S. (91) thus contrasts with (92), itself an “updated” version of (29), which is a prediction about c-command detection; the consequent in (92) has “[-cc]S” while that in (91) has “[+cc, +loc]S”.69

69 The J=no in BVA([-cc]S, X, Y):no in (92) must be, in part, due to the absence of c-command while the J=no in BVA([+cc, +loc]S, X, Y):no in (91) is not due to the absence of c-command.

Detection of C-command Effects 

 211

(92) Correlational/conditional prediction about c-command-detection with BVA (X, Y), with reference to [-loc] just in the “Provided . . .” clause: Provided that DR([+cc]S, X, beta):yes ∧ Coref([+cc, -loc]S, alpha, Y):yes ∧ BVA([+cc, -loc]S, X, Y):yes; DR([-cc]S, X, beta):no ∧ Coref([-cc]S, alpha, Y):no → BVA([-cc]S, X, Y):no The theoretical consequences about FD-BVA and FD-Coref, as given in (93) and (94), nonetheless, allow us to look into various issues in ways that have not been possible.70 (93) a. FD-BVA([+cc, -loc]S, X, Y):yes b. FD-BVA([+cc, +loc]S, X, Y):no (94) a. FD-Coref([+cc, -loc]S, X, Y):yes b. FD-Coref([+cc, +loc]S, X, Y):no The correlational prediction in (91) is indeed supported in my self-experiment, with regard to sentences with a “transitive o-marking” verb, ones with a “transitive ni-marking” verb, and ones with an ergative verb, providing yet further experimental evidence for the usefulness of the correlational methodology being pursued here although I cannot provide the relevant illustration due to space limitations.71

8 Further Issues There are a number of (important) issues that are not addressed in this chapter due to space limitations. In this section, I would like to make brief remarks on a few of them just to give the reader a glimpse at some of the further issues that have been, and are currently being, considered and investigated.72 As to what can induce NFS effects, we have so far mostly addressed choices of X and choices of Y. There are, however, other factors, including (i) whether the head N of X and that of Y are “identical” and (ii) whether the sentence in question is construed 70 For discussion about how the [-loc] requirement on FD(x, y) is related to the anti-locality condition on so-called Principle B of the Binding Theory (Chomsky 1981 and many works before and after), see Hoji 1995, 1997a, 1997b, and 1998. 71 The [-loc] requirement on FD(x, y) will be further discussed in the following section, in relation to two types of NFS’s. 72 A fuller discussion of the points below will be presented in a separate work, such as Hoji in preparation.

212 

 Hajime Hoji

as expressing a “categorical” or a “non-categorical” thought. Both of these seem relevant only to one of the two types of NFS; see note 52. To facilitate the ensuing exposition, let us refer to the type of NFS whose effects get affected by (i) and (ii) as “NFS1” and the other type whose effects do not get affected by (i) or (ii) as “NFS2”. Although the discussion thus far in this chapter simply states that the choice of X and that of Y both induce NFS effects, there seems to be an important difference between NFS1 and NFS2; the choice of Y seems to have a much stronger effect on NFS2 than X does on NFS1, to the point that it seems like X needs some sort of “support” from the sentence/context to achieve NFS1 whereas there is no such requirement for NFS2 and Y.73 Observationally (i.e., mostly based on my self-experiments), NFS1-BVA/Coref and NFS2-BVA/Coref are both possible in [+cc, +loc]S, as in (95a), (96a), and (97a). (95)

With a “transitive o-marking” verb a. [+cc, +loc]S: Y-o X/alpha-ga V b. [+cc, -loc]S: [Y-no N]-o X/alpha-ga V

(96)

With a “transitive ni-marking” verb a. [+cc, +loc]S: Y-ni X/alpha-ga V b. [+cc, -loc]S: [Y-no N]-ni X/alpha-ga V

(97)

With an ergative verb a. [+cc, +loc]S: Y-ga X/alpha-ni V b. [+cc, -loc]S: [Y-no N]-ga X/alpha-ni V

For example, recall that at Stage 2, as discussed in Sections 2 and 3, soitu can induce NFS2 effects for me with BVA/Coref(X/alpha, Y).74 It can be Y of (a) and (b) in (95)–(97). Sono otoko ‘that man’ at that stage, however, cannot induce NFS2 effects with BVA/Coref(X/alpha, Y); hence it cannot be Y of (a) of (95)–(97) because the [-loc] requirement on FD(x, y) is not satisfied, but it can be Y of (b)

73 The crucial difference between NFS1 and NFS2 is therefore that the former is based on the choice of X of MR(X, Y) and the latter on that of Y of MR(X, Y). NFS1’s reference to (i) and (ii), addressed above, and its crucial reference to X of MR(X, Y) will be understood to be a consequence of what underlies NFS1, and the difference between NFS1 and NFS2 turns out to play an important role in accumulating knowledge about the language faculty by the basic scientific method, as will be discussed in Hoji in preparation. 74 That is indicated by the fact that, at Stage 2, I get BVA([-cc, -loc]S, X, soitu):yes, with X being NP-igai ‘ones other than NP’, which, for me, cannot induce NFS1, as indicated by the J =no in DR([-cc]S, NP-igai, beta):no.

Detection of C-command Effects 

 213

in (95)–(97), where the [-loc] requirement on FD(x, y) is satisfied along with the c-command condition. The NFS2-BVA/Coref in question, however, becomes impossible in a [-cc, +loc]S, as in (98)–(100). (98)

(99)

With a “transitive o-marking” verb [-cc, +loc] S: Y-ga X/alpha-o V With a “transitive ni-marking” verb S: Y-ga X/alpha-ni V

[-cc, +loc]

(100) With an ergative verb [-cc, +loc] S: Y-ni X/alpha-ga V NFS1-BVA/Coref, which has the “N head identity” requirement, is possible in (98)–(100) as well as in other S, including [+cc, +loc]S, as in (a) of (95)–(97), and [-cc, -loc] S. NFS1-BVA/Coref, however, becomes impossible in a “non-categorical context” (see below), although the possibility of NFS2-BVA/Coref is not affected by being placed in such a “non-categorical context”.75 At Stage 3, unlike Stages 1 and 2, sono otoko can induce NFS2 effects for me. The pattern of my judgments with sono otoko as Y of BVA/Coref(X/alpha, Y) at this stage is exactly the same as the pattern of my judgment with soitu as Y of BVA/Coref(X/alpha, Y) at Stage 2. In addition to choices of X and those of Y, the possibility of DR([-cc]S, X, beta):yes seems to be affected by whether, and how easily, the sentence can be understood as corresponding to a “categorical-thought”, as pointed out in note 32. A “categorical-thought” is such that the X of the intended DR(X, beta) is understood as the “Subject” of the sentence about which the rest of the sentence expresses a property.76 Depending on the speaker, there are ways in which we can make it more likely or less likely (or even simply impossible) for a given sentence to be understood as corresponding to a “categorical-thought”. For example, for 75 There is no space for illustrating this based on concrete materials or for discussing possible accounts of the pattern of judgments in question. I only mention here that typical instances of NFS1-BVA/Coref, which are possible even in (98)–(100) but become impossible in a “noncategorical context”, are of the form BVA(. . . N, . . . N) and Coref(. . . N, . . .N), where the two head Ns are identical. 76 Relevant issues have been extensively discussed in works by Ueyama (1998) and Hayashishita (2004, 2013). Reference to “categorical” and “thetic” judgments (not to be confused with judgments such as J=yes and J=no in the preceding discussion) in relation to what is here called NFS-MR is first made in Ueyama 1998, drawing from Kuroda’s (1972, 1992) work on such judgments and their relation to sentence forms.

214 

 Hajime Hoji

many speakers, including myself, embedding the sentence in question in a To yuu koto-wa, ___ koto-ni naru ne ‘In that case, we can conclude ___’ context seems to serve the former purpose, and putting the sentence in the A! ___! ‘Look! ___!’ context seems to serve the latter purpose. Let us refer to the former context a “categorical context” and the latter as a “non-categorical context”. The use of these two contexts (or similar contexts) allows us to conduct more fine-grained experiments and can lead us to a better understanding of the nature of our judgments. For example, regarding DR([-cc]S, sukunakutomo 5-nin-no gisi, beta):J (see (41b), I get J=yes quite easily in a categorical context, but impossible in a non-categorical context.77 While the correlational methodology in LFS, as summarized in (29), (91), and (92), allows us to control for NFS effects for MR, the more we understand about the nature of NFS’s, the better position we will be in to accumulate knowledge about the language faculty by the basic scientific method. I should note that speakers’ judgments seem to differ as to how clearly the particular contexts mentioned above serve as a categorical context or as a non-categorical context. This makes it quite important for our further research that we have an effective and reliable means to “create” (or at least determine what is) a non-categorical context for a given speaker more effectively and reliably than simply using the particular non-categorical context mentioned above. To that end, I would like to make brief remarks here about how the reference to my own dialect of Japanese (the Kochi dialect) offers a useful means to “create” a “non-categorical context”. The Kochi dialect has two distinct verbal ending forms, ‘V-yuu’ (“progressive”) and ‘V-tyuu’ (“resultative”), corresponding to the V-te-iru form, which is ambiguous between “progressive” and “resultative” in Standard Japanese.78 It seems that the ‘V-yuu’ form provides me with a very effective means to create a non-categorical context, more effectively and reliably than the non-categorical context mentioned above. In that context, I clearly get DR([-cc]S, X, beta):no, even with X=subete-no N ‘every N’, which is normally the most NFS-inducing choice of X for me; see (39). With V-yuu’ (as opposed to ‘V-tyuu’), I also detect clear cases of non-FD-BVA/Coref becoming impossible in a non-categorical context even when my judgments become blurry with a non-

77 As pointed out in Hayashishita (2004, 2013) in relation to DR(X, alpha) and in Ueyama 1998 in relation to BVA(X, Y), the absence of effects of the choice of X on the availability of DR(X, alpha) and BVA(X, Y) is a hallmark of FR-DR/BVA. The fact that the J=yes in DR([+cc]S, X, beta):J (e.g., DR([-cc]S, sukunakutomo 5-nin-no gisi, beta):J) remains readily available in the “non-categorical context”, regardless of the choice of X, indicates that DR([+cc]S, X, beta) in a “non-categorical context” can be FR-DR (more precisely, DD-DR). 78 Many Western dialects in Japan have a similar distinction.

Detection of C-command Effects 

 215

categorical context such as the one mentioned above. This is another instance where we see a clear benefit for the LFSist in looking at his/her own I-language(s). Space limitations, however, do not allow me to discuss further how we can use these and other means to enhance the effectiveness of our experiment. The preceding discussion has focused on how the choice of Y influences NFS effects with BVA/Coref(X/alpha, Y); see Sections 3 and 4. The choice of Y, however, also affects the J of FD-BVA(X, Y):J, as discussed extensively in Ueyama 1998 and Hoji et al. 1999. As pointed out in Ueyama 1998: 6.1.4.2, this seems related to how FD(x, y) is interpreted; FD(x, y) is interpreted in such a way that y is to be interpreted as being “identical to” x, and the interpretation of FD(x, y) thus results in the complete elimination of the “semantic content” of y. In order for a linguistic expression to be Y of FD-BVA(X, Y), its semantic content must, therefore, be such that the speaker can “tolerate” its complete elimination due to how FD(LF(X), LF(Y)) is interpreted, as also pointed out in Ueyama 1998. Among the clearly singular-denoting overt nominal expressions in Japanese, the semantic content of soko ‘it’ seems the “smallest” and that of soitu ‘that guy” the next smallest. Ueyama 1998: 3.1 and Hoji et al. 1999 observe that speakers generally find it easier to use soko than soitu as Y of BVA(X, Y), at least so long as their judgments are such that BVA([+cc, -loc]S, X, Y):yes and BVA([-cc, -loc]S, X, Y):no, i.e., so long as they are considering FD-BVA(X, Y).79, 80 I was not able to use sono otoko as Y of FD-BVA(X, Y) until relatively recently, but I am now able to do so. The three stages discussed in Sections 2 and 3 do not include the “shifts” of my judgments in this regard, because by the time I started recording my judgments systematically, I had already come to accept sono otoko as Y of FD-BVA(X, Y) relatively easily. For most speakers, however, I suspect that, if they engaged themselves in a series of intensive acts of judgment making, there would be different stages as to when they could accept a given item as Y of FD-BVA(X, Y).81 If these individuals only participate in a non-intensive (non-self-)experiment, and judge the relevant 79 As in all other crucial schemata discussed here, we have Y preceding X in [+cc, -loc]S and [-cc, -loc]S, to avoid possible effects of X preceding Y on the availability of MR(X, Y). 80 FD(x, y) having two conditions (the c-command condition and the anti-locality condition) as in (88) leads us to predict a new, this time BVA-internal, correlation (with a certain assumption that I cannot get into here) that the Y that leads to a combination of BVA([+cc, -loc]S, X, Y):yes and BVA([-cc, -loc]S, X, Y):no also leads to a combination of BVA([+cc, -loc]S, X, Y):yes and BVA([+cc, +loc]S, X, Y):no. BVA-internal correlations are also discussed in Hoji 2003: Section 3. 81 Shifts from one stage to another in this respect interact, in a very interesting way, with the shifts with regard to what can serve as Y of NFS2-BVA/Coref(X/alpha, Y). In a nutshell, it seems that the latter shifts lag behind the former shifts. This is fortunate, as, if they shifted simultaneously, we would not be able to isolate FD-BVA(X, Y), because it could not be differentiated from

216 

 Hajime Hoji

sentences only one or two times, I expect the following, based on my years of experience in judging the relevant sentences myself and having other speakers judge them: As Y of FD-BVA(X, Y), most of them will accept soko, only some will accept soitu, and most will reject sono otoko. The chart in (101) indicates the relative ease with which a given Y can be Y of FD(LF(X), LF(Y)).82, 83 (101)

As just discussed, the choice of the Y in (101) can affect the availability of FD-based BVA/Coref, i.e., the J in (102), provided that no NFS effects intervene. A typical instance of [+cc, -loc]S is given in (103).84 (102) BVA/Coref([+cc, -loc]S, X/alpha, Y):J (103)

S: [Y-no N]-o X/alpha-ga V(erb)

[+cc, -loc]

As discussed earlier, the choice of the Y in (101) can also affect the J in (104) where the BVA/Coref must be due to NFS effects.

NFS2-BVA/Coref(X/alpha, Y). More extensive discussion will be provided in a separate work, e.g., Hoji in preparation. 82 While the “positioning” of soko, soitu, and sono otoko in the chart seems to be shared by other speakers, the relative positioning between sono otoko and kare may turn out to be more “idiosyncratic” among speakers. 83 As hinted at in note 79, for a more complete description of correlations of judgments regarding DR, Coref and BVA, especially between Coref and BVA, and also within BVA with regard to different verb types, different schema (sub-)types, and various other options, we must consider how different stages such as those discussed in Sections 2–4 “interact” with different stages with regard to (101), for a given speaker; this is yet another challenge and source of excitement in LFS. 84 Although the preceding discussion has tended to focus on cases where the BVA/Coref available in (103) is FD-BVA/Coref(X/alpha, Y), we cannot rule out the possibility that BVA/Coref(X/ alpha, Y) in (103) (and other subtypes of [+cc, -loc]S) can also be due to NFS if we have choices of X and Y that can induce NFS effects. In other words, the choice of a schema alone (such as (103)) does not guarantee that the BVA/Coref available there is FD-based, giving us another instance of uncertainty. As in the cases discussed above, this type of uncertainty does not prevent us from making definite correlational/conditional predictions and obtaining and replicating experimental results precisely in line with the predictions. In fact, uncertainty such as this allows us to further test the validity of our general correlational methodology and to learn both about the nature of the FR’s in question, as well as that of NFS’s in question.

Detection of C-command Effects 

 217

(104) a. BVA/Coref([-cc]S, X/alpha, Y):J b. BVA/Coref([+cc, +loc]S, X/alpha, Y):J Typical instances of the respectively.85 (105) a. b.

S and the

[-cc]

[+cc, +loc]

S are given in (105a) and (105b),

S: [Y-no N]-ga alpha/X-o V(erb) S: Y-o X/alpha-ga V(erb)

[-cc]

[+cc, +loc]

One might wonder if such judgment shifts within a speaker, as well as judgmental variations among speakers, would make it impossible for us to accumulate knowledge about the language faculty by the basic scientific method, which demands that we deduce definite predictions from our hypotheses and obtain and replicate experimental results accordingly. As discussed in the preceding pages, it is the correlational methodology, as summarized in our correlational/conditional predictions in (29), (91), and (92), and their contrapositives, that makes it possible for us to pursue the basic scientific method in the face of such judgment shifts and variations. Such considerations as above are meant to illustrate what challenges we face in LFS research and how we deal with them by our correlational methodology. We try to control for the NFS factors by the correlational methodology. To the extent that the testable correlational prediction never gets disconfirmed, and we additionally obtain c-command detection, we are able to provide support for both the universal and the I-language-particular hypotheses that lead to the prediction. Such results also provide support for the correlational methodology itself. This is how we can accumulate knowledge about the language faculty without understanding the full nature of any given NFS. After all, NFS effects are not part of our object of inquiry; we are concerned with the CS, which we observe through FR-based MRs. Nonetheless, understanding the nature of NFS is important in LFS. Which factors are considered part of the FRs and which are considered part of the NFS(s) is determined only by hypotheses and experiments. The same can be said of the question of what belongs to the CS and what belongs outside the CS, as schematized (106). 85 Space limitations prevent me from discussing further details about what happens to my I-language when I “move” from Stage 1 to Stage 2 to Stage 3, as discussed in Sections 3 and 4, which have to do with NFS2 effects, or when my judgments shift in relation to (101), which has to do with FD-BVA/Coref. The relevant issues are discussed in some depth in Ueyama 1998: Appendix D and Chapter 2 of this volume, in relation to the availability of BVA(X, Y) in (105a) and in Ueyama 1998: 3.4 and Hoji et al. 1999, in relation to the availability of BVA(X, Y) in (103).

218 

 Hajime Hoji

(106)

The “faculties” of the mind, or whatever our mind makes use of when we put the language faculty (more specifically, steady states of the language faculty) to use, plus the language faculty proper, which includes the CS the language faculty CS

Figure 2: The Computational System (=CS), the language faculty and what is “outside” the language faculty.86

What we are currently controlling for by our correlational methodology and calling “NFS factors” may well turn out to include FR factors, once we have attained a better understanding about how our linguistic judgments come about. We therefore want to expand the coverage of our research program to address what we currently understand as NFS factors. We can make hypotheses about NFS factors, make testable correlational predictions based on those hypotheses in combination with other hypotheses (including the universal hypotheses and the I-language-particular hypotheses whose validity has already been experimentally supported) even if the formulation of such hypotheses requires the use of concepts that are not (currently) part of our “theoretical concepts”.87 How such a research endeavor has already yielded new insight has been alluded to in various parts of the preceding discussion, albeit only schematically and briefly. A fuller discussion of the relevant

86 The two outer “circles” indicate that the borders indicated by them are “fluid”; see Figure 2 in Chapter 4 of this volume and the discussion thereabout. The use of the broken lines for the outer-most “circle” indicate that the border indicated are even more “fluid” than the border of the language faculty. 87 Recall that our “theoretical concepts” are for addressing properties of the CS or those directly related to the CS.

Detection of C-command Effects 

 219

investigation of intricate interaction of hypotheses about FR factors and ones about NFS1 factors and NFS2 factors and its success will illustrate the usefulness of the basic scientific method much beyond what has been addressed in this chapter, not only for our better understanding about properties of the CS proper but also for a better understanding about properties outside the CS.88

9 Summary and Conclusion Central to the preceding discussion is (29), repeated here. (29)

Correlational/conditional prediction about c-command-detection with BVA(X, Y): Provided that DR([+cc]S, X, beta):yes ∧ Coref([+cc]S, alpha, Y):yes ∧ BVA([+cc]S, X, Y):yes; DR([-cc]S, X, beta):no ∧ Coref([-cc]S, alpha, Y):no → BVA([-cc]S, X, Y):no

This chapter has addressed the design of my self-experiment. With every combination of X and Y, and in a number of constructions, I have obtained judgments in line with (29), and its slightly altered (91) and (92), and their contrapositives although the illustration provided above covers only a small portion of what has been considered. More specifically, in this chapter, I have illustrated how I have obtained judgments precisely in line with the correlational/conditional prediction in (29), repeated above; see also (30). The self-experiment, as described above, has two general stages. In the first stage, I identified choices of X that lead to ① and choices of Y that do ② for myself at a given time.89 In the second stage, I tested the prediction in (30) that I necessarily obtain BVA([-cc]S, X, Y):no) with the X and Y identified in the first stage. With the X and Y in question, I indeed obtained ③, hence c-command detection with BVA(X, Y).90 As discussed, what choices of 88 I plan to provide a fuller illustration and related discussion in Hoji in preparation. I will briefly address the “usefulness” of making reference to NFS1 and NFS2 when I turn to the demonstration of the absence of “other types of entailment” in Chapter 6 of this volume, based on a large-scale non-self-experiment. 89 This is the stage where I try to confirm the existential predictions that there exist at least one combination of MR, X, Y and schemata that gives clear signs of c-command effects. 90 In the preceding discussion, (i-a) and (i-b) are considered as “the same type” of [-cc]S and (ii-a) and (ii-b) as “the same type” of [-cc]S, for DR and BVA/Coref, here focusing on cases where the V is the “regular o-marking” transitive verb.

220 

 Hajime Hoji

X and Y lead to c-command detection is not always the same, even for the same speaker, such as myself, but once such X and Y are identified at a given point in time, based on DR(X, beta) and Coref(alpha, Y), they lead to c-command detection with BVA(X, Y). If I were instead to obtain BVA([-cc]S, X, Y):yes, with the X and Y in question, that would disconfirm the definite correlational prediction in (30). As stated, the self-experiment is the most basic experiment in language faculty science, and this chapter has provided a brief illustration of how I have obtained c-command detection in my own self-experiment, utilizing the correlational methodology. Objections may be raised, however, against the use of the researcher’s own intuitions as data for or against the hypotheses that the researcher himself entertains, as have in fact been raised by a number of people in the past, not only by those critical of Chomsky’s general conception of the language faculty (e.g., Gibson and Fedorenko (2013)) but also by those sympathetic to Chomsky’s generative enterprise (e.g., Schütze (1996)). What might be the most effective response to such an objection is a demonstration that the successful c-command detection in the self-experiment can be reliably and testably replicated with other speakers, and even in other “languages”. In Chapter 6, I will discuss how the correlational, but definite, experimental results obtained in my self-experiment have been replicated in a large-scale non-self-experiment.

(i)

(ii)

S

[-cc]

a. b.

For DR(X, beta): beta-ga X-o V For BVA/Coref(X/alpha, Y): beta-no N-ga X-o V

S For DR(X, beta): beta-o X-ga V For BVA/Coref(X/alpha, Y): beta-no N-o X-ga V

[+cc]

a. b.

An important issue remains as to what counts as “the same type” of schemata. The issue needs to be addressed both from the empirical perspective (e.g., based on what answers to the question lead to the most reliable and effective means for c-command detection) and from the conceptual perspective (e.g., what conceptual basis there can be for the answers that we will have arrived at). Such line of research will be guided by the basic scientific method, and more in particular, by the correlational methodology in LFS, as outlined above.

Detection of C-command Effects 

 221

References Chomsky, Noam. 1981. Lectures on government and binding: The Pisa lectures. Dordrecht: Foris Publications. Fukui, Naoki and Yuji Takano. 1998. Symmetry in syntax: Merge and demerge. Journal of East Asian Linguistics 7. 27–86. https://doi.org/10.1023/A:1008240710949 (accessed 10 March 2022) Gibson, Edward and Evelina Fedorenko. 2013. The need for quantitative methods in syntax and semantics research. Language and Cognitive Processes 28(1)/(2). 88–124. Hayashishita, J.-R. 2004. Syntactic and non-syntactic scope. Los Angeles, CA: University of Southern California dissertation. Hayashishita, J.-R. 2013. On the nature of inverse scope readings. Gengo Kenkyu 143. 29–68. Hayashishita, J.-R. and Ayumi Ueyama. 2012. Quantity expressions in Japanese. In Edward Keenan and Denis Paperno (eds.), The handbook of quantification in natural language, 535–612. New York: Springer. Hoji, Hajime. 1985. Logical form constraints and configurational structures in Japanese. Seattle, WA: University of Washington dissertation. Hoji, Hajime. 1991. KARE. In Carol Georgopoulos and Roberta Ishihara (eds.), Interdisciplinary approaches to language: Essays in honor of S.-Y. Kuroda, 287–304. Dordrecht: Kluwer Academic Publishers. Hoji, Hajime. 1995. Demonstrative binding and principle B. In Jill N. Beckman (ed.), NELS 25, 255–271. Amherst, MA: University of Massachusetts, Amherst, GLSA Publications. Hoji, Hajime. 1997a. Sloppy identity and principle B. In Hans Bennis, Pierre Pica and Johan Rooryck (eds.), Atomism and binding, 205–235. Dordrecht: Foris Publications. Hoji, Hajime. 1997b. Sloppy identity and formal dependency. In Brian Agbayani and Sze-Wing Tang (eds.), Proceedings of the 15th West Coast Conference on Formal Linguistics, 209–223. Stanford, CA: CSLI Publications. Hoji, Hajime. 1998. Formal dependency, organization of grammar, and Japanese demonstratives. Japanese/Korean Linguistics 7. 649–677. Hoji, Hajime. 2003. Falsifiability and repeatability in generative grammar: A case study of anaphora and scope dependency in Japanese. Lingua 113. 377–446. Reprinted in: Hoji 2013. Hoji, Hajime. 2013. Gengo kagaku-o mezashite [Towards linguistic science]: Issues on anaphora in Japanese. (Ayumi Ueyama and Yukinori Takubo (eds.).) Shiga: Ohsumi Publisher. Hoji, Hajime. 2015. Language faculty science. Cambridge: Cambridge University Press. Hoji, Hajime. 2016. Towards language faculty science: Remarks on the papers collected in Hoji 2013. Preface to the e-edition of Hoji 2013. Hoji, Hajime. In preparation. Experiment in language faculty science. Hoji, Hajime, Satoshi Kinsui, Yukinori Takubo and Ayumi Ueyama. 1999. Demonstratives, bound variables, and reconstruction effects. Proceedings of the Nanzan GLOW: The Second GLOW Meeting in Asia, September 19–22. 141–158. Kuroda, S.-Y. 1972. The categorical and the thetic judgments. Foundations of Language 9. 153–185. Kuroda, S.-Y. 1978. Case-marking, canonical sentence patterns and counter equi in Japanese (A preliminary survey). In John Hinds and Irwin Howard (eds.), Problems in Japanese syntax

222 

 Hajime Hoji

and semantics, 30–51. Tokyo: Kaitakusha. Reprinted in S.-Y. Kuroda 1992, Japanese syntax and semantics: Collected papers, 222–239. Dordrecht: Kluwer. Kuroda, S.-Y. 1986. Movement of noun phrases in Japanese. In Takashi Imai and Mamoru Saito (eds.), Issues in Japanese linguistics, 229–272. Dordrecht: Foris Publications. Reprinted in S.-Y. Kuroda 1992, Japanese syntax and semantics: Collected papers, 253–292. Dordrecht: Kluwer. Kuroda, S.-Y. 1992. Judgment forms and sentence forms. In S.-Y. Kuroda, Japanese syntax and semantics: Collected papers, 13–77. Dordrecht: Kluwer. Plesniak, Daniel. 2021. Experiments beyond statistics: How generative linguistics can beat the replication crisis. Ms, available at: https://www.researchgate.net/ publication/357063705_Experiments_Beyond_Statistics_How_Generative_Linguistics_ Can_Beat_the_Replication_Crisis (accessed 21 March 2022) Plesniak, Daniel. 2022. Towards a correlational law of language: Three factors constraining judgment variation. Los Angeles, CA: University of Southern California dissertation. Schütze, Carson. 1996. The empirical base of linguistics: Grammaticality judgments and linguistic methodology. Chicago: University of Chicago Press. Reprinted in 2016. Berlin: Language Science Press. Ueyama, Ayumi. 1998. Two types of dependency. Los Angeles, CA: University of Southern California dissertation.

Hajime Hoji

Replication: Predicted Correlations of Judgments in Japanese 1 Introduction In my self-experiments, I obtain experimental results precisely as predicted by the correlational/conditional prediction in (1) and its contrapositive in (2), with each of the many options for X and Y I consider, and with a number of subtypes of [+cc]S and [-cc]S, as discussed in Chapter 5.1 (1)

Correlational/conditional prediction about c-command-detection with BVA(X, Y): Provided that DR([+cc]S, X, beta):yes ∧ Coref([+cc]S, alpha, Y):yes ∧ BVA([+cc] S, X, Y):yes; DR([-cc]S, X, beta):no ∧ Coref([-cc]S, alpha, Y):no → BVA([-cc]S, X, Y):no

(2) The contrapositive of (1): Provided that DR([+cc]S, X, beta):yes ∧ Coref([+cc]S, alpha, Y):yes ∧ BVA([+cc] S, X, Y):yes; BVA([-cc]S, X, Y):yes → DR([-cc]S, X, beta):yes ∨ Coref([-cc]S, alpha, Y):yes

In this chapter, I will illustrate how some of the results of my self-experiments have been replicated in non-self-experiments. An experiment in LFS is always about an individual; this is obvious in a self-experiment, but it is also the case in non-self-experiments as well; the latter should be conceived of as an experiment conducted on someone other than the researcher him/herself in an attempt to replicate the result of his/her self-experiment. A given non-self-experiment may be conducted on multiple different people, giving the impression of multiple participants in the same experiment, but this situation is more properly understood as multiple non-self-experiments occurring in parallel, each on a particular individual. For reasons discussed in Chapters 4 and 5, self-experiments are the most important form of experiment in LFS for deepening our understanding of the lan-

1 See Chapter 4 for the initial conceptual motivation for (1) and Chapter 5 for how predictions of the form (1) have received experimental support in my self-experiment. Some notations crucially used in this chapter are “explained” in Appendix II. https://doi.org/10.1515/9783110724790-006

224 

 Hajime Hoji

guage faculty by the basic scientific method. Nevertheless, as in any scientific research program, replication of experimental results is required in order for the results to count as being reliable (enough). Given our correlational approach, replication in LFS is not concerned with the replication of specific judgments, i.e., judgments on the availability of a specific instance of MR(X, Y) with specific choices of X and Y in a specific sentence in a specific language. Rather, replication in LFS concerns correlations of judgments in the form of (1) and (2). The focus on correlations of judgments in the form of (1) and (2) is what makes it possible for us to pursue replication of results of a self-experiment (with a particular I-language) with non-self-experiment (with any other I-language). As briefly addressed in Chapter 5, within his/her own self-experiments, a “LFStist” tries to replicate the results of an initial self-experiment by extending the initial empirical coverage, e.g., by considering additional sentence patterns, additional choices of X and Y, additional MRs, etc. This is done so as to obtain more confidence in the validity of the hypotheses that have led to apparent success in the initial self-experiment. Along with replication via self-experiments, a LFStist also ought to obtain replication in non-self-experiments with speakers in the same linguistic community (i.e., speakers of the same language) to obtain further confidence in, and demonstrate to others about, the validity of our hypotheses about properties of the Computational System (CS) of the language faculty (which are hypothesized to be universal) and the correlational methodology. Finally, the LFStist wants to obtain replication in non-self-experiments with speakers of other languages to obtain yet further confidence in, and demonstrate to others about, the validity of our universal hypotheses and the correlational methodology that are behind the correlational/conditional predictions. In addition, non-self-experiments can be designed to investigate I-language-specific properties of speakers other than the self (i.e., the researcher his/herself).2 Replication of results of the self-experiment in non-self-experiments is thus part of the essential activities in LFS. To be convincing to others, replication may have to take the form of a number of non-self-experiments, preferably with non-researcher-participants (i.e., “naïve” participants), to preempt possible objections to the use of introspective judgments by researchers, let alone the researcher him/herself, who has designed the experiments based on the hypotheses that s/he is putting forth and/or adopting. This chapter addresses replication in non-self-experiments with speakers of the same linguistic community as the LFStist (myself), i.e., a speaker of Japanese. These 2 Such investigation, based on non-self-experiments or self-experiments, can be carried out effectively only if we have attained self-experimental results in support of our universal hypotheses and the correlational methodology. How this can be pursued involves issues that go beyond the scope of this chapter, and will not be addressed here; see, however remarks in Section 4.

Replication: Predicted Correlations of Judgments in Japanese 

 225

experiments took the form of what is sometimes called a “large-scale experiment”, by which I mean that the “experiment” (really multiple experiments as discussed above) involves the checking, in various ways, of linguistic judgments by a “large” group of participants. I will refer to such a set of non-self-experiments as “demonstration” because it is intended to demonstrate the validity and viability of the relevant hypotheses and methodology by replication of results of my self-experiment. It is important to stress, yet again, that we still focus on an individual, more precisely, on definite correlational/conditional predictions about each individual’s judgments, rather than the average responses from the group of participants.3 The “large-scare experiment” in question, consisting of close to 200 non-selfexperiments (see below), was conducted on-line in a large General-Education course at Kyushu University during the spring semester of 2017.4 The participants are all undergraduate students with little to no prior background in linguistics. In addition to illustrating how the result of my self-experiment is replicated in these non-self-experiments, in relation to (1) and (2), this chapter addresses replication of the demonstrated absence of “other types of entailment” discussed in Chapter 5: Section 5. Non-self-experiments of this type differ from self-experiments in large part due to the different degrees of “resourcefulness” that we can assume of a researcher and of an (average) non-researcher. We might consider two distinct types of “resourcefulness”, having to do with (i) the unbounded willingness and patience to judge sentences and (ii) the familiarity and awareness of what is being checked, both of which we can in principle assume of a researcher him/herself but not necessarily of a non-researcher. Given that the researcher has (i) but a non-researcher presumably does not, it follows that a self-experiment, in principle, has no constraint as to how many sentences are checked, how many times the same or very similar sentences are checked, how unnatural or complicated sentences can be, etc. Non-self-experiments in this kind of demonstration, on the other hand, have severe constraints in all these regards. What we can expect to check in such non-self-experiments must therefore be a small subset of what we can expect to check in selfexperiments.

3 See Plesniak 2022: 3.1.2 for important discussion about possible, or even likely, confusion due to the use of the term “experiments” here, because what is meant by the term in the field of linguistics at large is “typicality-seeking”, which is qualitatively different from what is meant by the term in LFS, which is essentially “possibility-seeking”, in the terms of Plesniak’s discussion. This is also addressed briefly in Plesniak’s Chapter 8 of this volume. 4 I am deeply indebted to Ayumi Ueyama for having her students participate in these experiments as part of their course work.

226 

 Hajime Hoji

Given that the researcher has (ii) but a non-researcher, in principle, does not, it is necessary for such non-self-experiment to include “components” specific to non-self-experiment in order to “compensate” for the absence of (ii) on the part of a non-researcher, to ensure as much reliability as possible for results of non-selfexperiment. Specific discussion about this will be provided in Section 2.4.

2 The Design of Non-self-experiment 2.1 General Features of the Non-self-experiment In my self-experiment, I have checked the correlational/conditional predictions of the form in (1) and its contrapositive in (2), using the schemata for three distinct types of verbs, as given in (3)–(8)5 for BVA(X, Y), DR(X, beta) and Coref(alpha, Y), and with a number of choices of X and Y. (3) For DR(X, beta), with a “transitive o-marking” verb: a. [+cc]S: beta-o X-ga V b. [-cc]S: beta-ga X-o V (4) For DR(X, beta), with a “transitive ni-marking” verb: c. [+cc]S: beta-ni X-ga V d. [-cc]S: beta-ga X-ni V (5) For DR(X, beta), with an “ergative” verb: e. [+cc]S: beta-ga X-ni V f. [-cc]S: beta-ni X-ga V (6) For BVA(X, Y) and Coref(alpha, Y), with a “transitive o-marking” verb a. [+cc]S: [ . . . Y . . . ]-o alpha/X-ga V b. [-cc]S: [ . . . Y . . . ]-ga alpha/X-o V (7) For BVA(X, Y) and Coref(alpha, Y), with a “transitive ni-marking” verb a. [+cc]S: [ . . . Y . . . ]-ni alpha/X-ga V b. [-cc]S: [ . . . Y . . . ]-ga alpha/X-ni V

5 For the hypotheses that lead to the relevant predictions, see Chapter 5 of this volume: Section 1.

Replication: Predicted Correlations of Judgments in Japanese 

(8)

 227

For BVA(X, Y) and Coref(alpha, Y), with an “ergative” verb a. [+cc]S: [ . . . Y . . . ]-ga alpha/X-ni V b. [-cc]S: [ . . . Y . . . ]-ni alpha/X-ga V

I have also checked the correlational/conditional prediction in (9) and its contrapositive in (10), with the additional schemata for BVA and Coref, as given in (11).6 (9) Testable correlational/conditional prediction with BVA(X, Y), based on DR(X, beta) and Coref(alpha, Y): Provided that DR([+cc]S, X, beta):yes ∧ Coref([+cc, -loc]S, alpha, Y):yes ∧ BVA([+cc, -loc]S, X, Y):yes; DR([-cc]S, X, beta):no ∧ Coref([+cc, +loc]S, alpha, Y):no → BVA([+cc, +loc]S, X, Y):no (10) Contrapositive of (9): Provided that DR([+cc]S, X, beta):yes ∧ Coref([+cc, -loc]S, alpha, Y):yes ∧ BVA([+cc, -loc]S, X, Y):yes; BVA([+cc, +loc]S, X, Y):yes → DR([-cc]S, X, beta):yes ∨ Coref([+cc, +loc]S, alpha, Y):yes S: [NP1 . . . Y . . .] [NP2 X/alpha] V6 S: [NP1 Y ] [NP2 X/alpha] V

(11) a. b.

[+cc, -loc]

[+cc, +loc]

As reported in Chapter 5, in my self-experiment, the testable/disconfirmable predictions never get disconfirmed, i.e., there has not been any instance of the predicted entailments not holding, with every combination of X and Y considered, and with every one of the three types of verb. Furthermore, my judgments make the third conjunct of the “Provided that .  .  .” clause true, depending upon the choice of Y and depending upon the time of judgment-making, as discussed in Chapter 5: Sections 2–4, resulting in c-command detection in my self-experiment.

6 A subtype of (11a) is (i), where NP1 consists of Y-no (the so-called genitive marker) + the head Noun. (i)

S: [NP1 Y-no N] [NP2 X/alpha] V

[+cc, -loc]

The choice between (11a) and the structure in which Y is embedded in a (relative) clause modifying the head N, for example, does not result in any significant difference in my self-experiment or in earlier non-self-experiments, such as those discussed in Hoji 2015. In the subsequent discussion, I will therefore consider the simpler version of the two, i.e., (i).

228 

 Hajime Hoji

As noted, the researcher, at least in theory, has the unbounded willingness and patience to judge sentences, but a non-researcher does not. This necessarily makes the “empirical coverage” significantly smaller in a non-self-experiment than in a self-experiment. The non-self-experiments conducted at Kyushu University in the spring semester of 2017 only use the [-cc]S and [+cc]S in (3) and (4) and specific instances of (6), and (7), as given in (12) and (13), not considering (5), (8), (11b), or the many other schemata considered in my self-experiment. (12)

For BVA(X, Y) and Coref(alpha, Y), with a “transitive o-marking” verb a. [+cc]S: [Y-no N]-o alpha/X-ga V b. [-cc]S: [Y-no N]-ga alpha/X-o V

(13) For BVA(X, Y) and Coref(alpha, Y), with a “transitive ni-marking” verb a. [+cc]S: [Y-no N]-ni alpha/X-ga V b. [-cc]S: [Y-no N]-ga alpha/X-ni V From this point, I will refer to the totality of the non-self-experiments conducted at Kyushu University in the spring semester of 2017 as “K17s”. K17s also only includes the choices of X and Y in (14) and (15), thus excluding the many other choices of X and Y considered in my self-experiment.7 (14) Choices of X for BVA/DR(X, Y/beta) in K17s: a. conjoined NPs b. subete-no N ‘every N’ c. #-cl-no N ‘# Ns’ d. sukunakutomo #-cl-izyoo-no N ‘at least # or more N’ e. the FQ version of (d) = N-cm sukunakutomo #-cl-izyoo7 (15)

Choices of Y for BVA(X/alpha, Y) in K17s: a. soko b. soitu

As discussed in Chapter 5, the best pair of X and Y for me for c-command detection, is most of the time a pair of X=NP-igai and Y=sono otoko. These choices of X and Y are not, however, included in (14) and (15). Given the choices of X and Y included in K17s, if I had been a participant, my judgments would not have led to

7 See Hayashishita and Ueyama 2012: Section 10.2.1 for difference between the two different forms as (14d) and (14e).

Replication: Predicted Correlations of Judgments in Japanese 

 229

c-command detection with BVA(X, Y) (though they would not have disconfirmed the key predictions either, because the antecedent clause in (1) is not satisfied for me with any pair of X and Y included in K17s). For purposes of illustration, let us refer to Figure 1, copied from Chapter 5 of this volume.

Provided that

DR(

S, X, beta):yes + Coref(

S, alpha, Y):yes

+

BVA(

S, X, Y):yes

DR(

S, X, beta):no

S, alpha, Y):no



BVA(

S, X, Y):no

Identification of X that gives rise to DR(X, beta) but does not give rise to NFS-DR(X, beta), i.e., X that gives rise to DR(X, beta) only based on FR.

+ Coref(

Identification of Y that gives rise to Coref(alpha, Y) but does not give rise to NFS-Coref(alpha, Y), i.e., Y that gives rise to Coref(alpha, Y) only based on FR.

;

When accompanied by and , it is c-command detection with BVA(X, Y), with the X and Y identified in and . In isolation, identification of a pair of X and Y that gives rise to BVA(X, Y), but does not give rise to NFS-BVA(X, Y), i.e., a pair of X and Y that gives rise to BVA(X, Y) only based on FR.

Figure 1: Illustrating the correlational/conditional prediction about c-command detection with BVA(X, Y) in (1).

As pointed out in Sections 2–4 of Chapter 5 of this volume, I obtain  only with certain choices of X and Y, and which choices of X and Y lead to such judgments can vary depending upon the given time of judgment making. I expect this to hold of other speakers as well, as I have in fact observed over the years. With the limited choices of X and Y in K17s, as in (14) and (15), we may therefore not obtain  (with DR(X, beta)) or  (with Coref(alpha, Y)) for some (or even many) speakers in K17s, with the particular choices of X and Y in (14) and (15).8 If that turns out to be the case, (1) does not make any prediction about those speakers’ judgments on BVA([-cc]S, X, Y), i.e., there is no entailment that we can check for those speakers in line with (1).9 This, of course, does not disconfirm our definite correlational prediction; we simply cannot check the entailment because the antecedent clause of

8 As in Chapter 5, whenever a number in a circle, such as ,  and , is used, reference is made to Figure 1. 9 This, however, does not necessarily mean that those individuals whose judgments do not constitute  or  have no role to play in checking our predictions because they can potentially be

230 

 Hajime Hoji

(1) is not satisfied. Only if the antecedent clause of (1) is satisfied, i.e., only if we obtain  and  for a given speaker, do we make a definite correlational prediction about that speaker’s judgment on BVA([-cc]S, X, Y); more precisely, we predict that that speaker will give the BVA([-cc]S, X, Y):no judgment, with the X and Y that have led to  and . As will be illustrated below, in every case in K17s where we obtain  and  for a given speaker, the definite correlational prediction about that speaker’s judgment about BVA([-cc]S, X, Y):no is not disconfirmed. In other words, when the entailment is made for a given speaker that the J of that speaker in BVA([-cc]S, X, Y):J is no, the J is indeed no for every such speaker, i.e., the prediction has not been disconfirmed. Furthermore, many of those instances of BVA([-cc]S, X, Y):no are accompanied by judgments instantiating the third conjunct of the “Provided that . . .” clause, i.e., “BVA([-cc]S, X, Y):yes”, which thus gives us c-command detection with BVA(X, Y). C-command detection with BVA(X, Y) in my self-experiment is thus replicated in K17s, though not with the same choices of X and Y. We have also discussed the contrapositive of (1), repeated here, along with (1). (1)

Correlational/conditional prediction about c-command-detection with BVA(X, Y): Provided that DR([+cc]S, X, beta):yes ∧ Coref([+cc]S, alpha, Y):yes ∧ BVA ([+cc]S, X, Y):yes; DR([-cc]S, X, beta):no ∧ Coref([-cc]S, alpha, Y):no → BVA([-cc]S, X, Y):no

(2) The contrapositive of (1): Provided that DR([+cc]S, X, beta):yes ∧ Coref([+cc]S, alpha, Y):yes ∧ BVA([+cc] S, X, Y):yes; BVA([-cc]S, X, Y):yes → DR([-cc]S, X, beta):yes ∨ Coref([-cc]S, alpha, Y):yes

Since (2) is the contrapositive of (1), (2) should be supported experimentally as long as (1) is. In my self-experiment, I have confirmed that, as noted in Chapter 5. It is only with very limited pairs of X and Y that satisfy the antecedent clause in (1) that I can check the predicted entailment in (1), i.e., whether I obtain BVA([-cc] S, X, Y):no. In the case of (2), the situation is reversed; with a number of pairs of X and Y, I obtain BVA([-cc]S, X, Y):yes, i.e., BVA(X, Y) without LF(X) c-commanding LF(Y). That satisfies the antecedent clause in (2), making it possible to check the predicted entailment in (2), i.e., whether I necessarily get DR([-cc]S, X, beta):yes

used to support contrapositive predictions in the form of (2), as discussed in Chapter 5 of this volume: Section 6.

Replication: Predicted Correlations of Judgments in Japanese 

 231

or Coref([-cc]S, alpha, Y):yes or both, with the choices of X and Y that lead to BVA ([-cc]S, X, Y):yes. Since most pairs of X and Y lead to BVA([-cc]S, X, Y):yes for me, I have plenty of opportunities to test the validity of (2), contrasting quite sharply with the situation regarding (1) (where I can check the entailment only with a very limited number of pairs of X and Y). Both (1) and its contrapositive in (2) are checked in K17s, and whenever the antecedent clause is satisfied, the predicted entailment is confirmed, without an exception. This is significant in light of the fact that speaker judgments can be affected by various factors and whose instability and variation seem to have led many to concede to the thesis that we cannot deal with speaker judgments by the basic scientific method.10 Judgmental instability is discussed with regard to my self-experiments in Chapter 5, and judgmental variability will be discussed with regard to K-17s in the following sections. Despite such judgmental instability and variability, we obtain categorical experimental results both in my self-experiments and K17s, in support of (1) and its contrapositive in (2), which seems to point to the basic validity of the correlational methodology being pursued here.

2.2 Design of the Non-self-experiment As noted, K17s uses the [-cc]S and the [+cc]S in in (3), (4), (12) and (13), and the choices of X and Y in (14) and (15). Hoji 2015: Chapter 7 provides a detailed illustration of the actual implementation of non-self-experiment, on the basis of the [+cc] S and the [-cc]S in (6) and (7) (which subsume the schemata in (12) and (13)). The general design of each component of a non-self-experiment discussed in Hoji 2015 is duplicated in the design of K17s.11 I will therefore not provide too many details about the exact implementation of K17s here, except for the most crucial aspects of the non-self-experiments in question and the features newly added in K17s.12

10 See Plesniak’s Chapter 7 of this volume and Plesniak 2021 for much relevant discussion. 11 Hoji 2015 does not pursue the correlational methodology discussed here, dealing only with BVA(X, Y) (not with DR(X, beta) or Coref(alpha, Y)). Chapter 1 of this volume contains a brief history of research from Hoji 1985 to Hoji 2015 to the correlational methodology as it is currently pursued. 12 The reader is thus referred to Hoji 2015: 6.3 and 7.3 for actual implementation of specific non-self-experiments discussed in Hoji 2015. Mukai’s Chapter 3 of this volume reviews what Hoji 2015 pursues and provides a concrete illustration of how the research can be conducted in the spirit of Hoji 2015.

232 

 Hajime Hoji

The main features of the actual non-self-experiment both in Hoji 2015 and K17s include the following: First, each experiment consists of a/multiple set(s) of three sentence types as indicated in (16), with MR being BVA, DR, or Coref.1314 (16) Three sentence types in K17s (and in Hoji 2015):14 a. a sentence instantiating [+cc]S, with an intended MR(X, Y) b. a sentence instantiating [-cc]S, with the intended MR(X, Y) c. the same sentence as (b) but without the MR(X, Y) For (16a) and (16b), intended to check c-command effects on the MR(X, Y) in question, the participant is instructed to report whether s/he finds the sentence acceptable at all with the specified MR(X, Y).15 (16c) is included in non-self-experiment, to make sure that a participant’s J=no in MR([-cc]S, X, Y):J is not merely due to the participant’s not being able to accept the sentence in general, regardless of MR(X, Y).16 Two distinct types of verb are used for each set of sentences, one being a regular o-marking transitive verb and the other being the regular ni-marking transitive verb. With the three sentence types in (16) and these two types of verbs, we have 3×2=6 sentences as a minimal set for each MR(X, Y). Using the same two verbs across all items dealing with BVA(X, Y), DR(X, beta) and Coref(alpha, Y), six sentences are presented to the participant, in a random order.17 K17s includes experiments with DR(X, beta) and Coref(alpha, Y), in addition to BVA(X, Y). While what is intended by Coref(alpha, Y) (that alpha and Y are used to refer to the same individual/object) is relatively straightforward for participants to understand, what is intended by BVA(X, Y) and DR(X, beta) does not seem to be, and that makes it even more important with BVA(X, Y) and DR(X, beta) than with Coref(alpha, Y) that the intended interpretations are properly

13 Recall that Hoji 2015 does not discuss DR or Coref. 14 Mukai’s Chapter 3 of this volume refers to this as “the minimum paradigm requirement”. 15 See Hoji 2015: 7.2 for exactly how the instructions are given. See also Plesniak’s Chapter 8 of this volume for related discussion. 16 Checking each participant’s judgments on (16a) and (16b), for each of the three MRs in question, is a required part of self or non-self-experiment, to check the validity of (1). 17 In Hoji 2015, the participants were able to choose this design or an alternative design where one sentence is shown at a time, in a random order. They were allowed to participate in a given experiment under either “design” or both. The choice between these two “designs” did not yield a significant difference at all, and in K17s, we only had one of the two “designs”. See Hoji 2015: 5.3 and also “The one-sentence-at-a-time test type vs. the three-sentences-at-a-time test type” under “EPSA [31]-1 (=[31]-8)” at http://www.gges.xyz/hojiCUP/English/31-1.shtml, for example. What is given there is in reference to a particular English experiment but the same applies to other experiments in English and also to experiments in Japanese discussed in Hoji 2015.

Replication: Predicted Correlations of Judgments in Japanese 

 233

understood by the participants. The way in which intended DR(X, beta) is provided in K17s is exemplified in (17), which corresponds to the case of DR(every local government, three politicians) for Japanese sentences such as one corresponding to ‘Every local government criticized three politicians’.18 (17) Under the interpretation “Having criticized three politicians seems to hold of every local government, and which three politicians differs depending upon each local government”. The number of the criticized politicians is three times as many as the number of the local governments. The way in which intended BVA(X, Y) is provided in K17s is exemplified in (18), which corresponds to the case of BVA(every local government, it), again for Japanese sentences such as one corresponding to ‘Every local government criticized its employees’. (18) Under the interpretation “Having criticized their own employees holds true of every local government.” There are at least as many criticized employees as the number of the local governments.’ In K17s, unlike in Hoji 2015, each sentence has rasii (“it seems” or “I/we hear”) added to the end of it. This is meant to facilitate the acceptability of the sentence in question, by making a “non-specific interpretation”, which is required for the BVA or the DR interpretation in question, easier to obtain.

2.3 What is Considered “Yes” and What is Considered “No” in Non-self-experiment To evaluate whether the result of a non-self-experiment in K17s is in line with the definite correlational/conditional prediction in (1), it is necessary to specify exactly what reported judgment(s) by a participant on a particular sentence, e.g., those in (16a) and (16b), would count as yes or no. (1) and (16) are repeated here.

18 Appendix I provides some actual sentences included in K17s, as presented to the participants. These include those with BVA(X, Y) and DR(X, beta) and their intended interpretations, along with sentences with Coref(alpha, Y).

234  (1)

 Hajime Hoji

Correlational/conditional prediction about c-command-detection with BVA(X, Y): Provided that DR([+cc]S, X, beta):yes ∧ Coref([+cc]S, alpha, Y):yes ∧ BVA([+cc]S, X, Y):yes; DR([-cc]S, X, beta):no ∧ Coref([-cc]S, alpha, Y):no → BVA([-cc]S, X, Y):no

(16) Three sentence types in K17s (and in Hoji 2015): a. a sentence instantiating [+cc]S, with an intended MR(X, Y) b. a sentence instantiating [-cc]S, with the intended MR(X, Y) c. the same sentence as (b) but without the MR(X, Y)

The claim is that if X gives rise to DR(X, beta) but does not give rise to NFS-DR(X, beta) and if Y gives rise to Coref(alpha, Y) but does not give rise to NFS-Coref(alpha, Y), there will be no NFS effects with BVA(X, Y), which means BVA([-cc]S, X, Y):no, with the X and Y in question.19, 20 The particular [+cc]S and [-cc]S used are as in (12) for BVA/Coref(X/alpha, Y) and as in (3) for DR(X, beta), repeated below, focusing here on sentences with a “transitive o-marking” verb.21 (12) For BVA(X, Y) and Coref(alpha, Y), with a “transitive o-marking” verb a. [+cc]S: [Y-no N]-o alpha/X-ga V b. [-cc]S: [Y-no N]-ga alpha/X-o V (3) For DR(X, beta), with a “transitive o-marking” verb: a. [+cc]S: beta-o X-ga V b. [-cc]S: beta-ga X-o V The consequent of the conditional in (1), applicable only if said conditions are met, states, that BVA(X, Y) is not available in sentences of the form in (12b) hence such sentences are unacceptable with BVA(X, Y). 19 ‘NFS’ abbreviates “non-formal source”, as opposed to a “formal source”, of MR(X, Y), such as BVA(X, Y), DR(X, beta), and Coref(alpha, Y). The “formal source” is called “FR” (formal relation), and “FD” and “DD”, discussed in Chapter 5 of this volume: Section 7.2, are instances of FR. 20 This is under the simplification made in Chapter 5 of this volume: Sections 1–4, as stated in (i) and further addressed in Section 5 thereof. (i)

(=Chapter 5: (33)) a. b.

X gives rise to NFS effects with BVA(X, Y) only if it gives rise to NFS effects with DR(X, beta). Y gives rise to NFS effects with BVA(X, Y) only if it gives rise to NFS effects with Coref(alpha, Y).

21 See (13) and (4) for the schemata with a “transitive ni-marking” verb.

Replication: Predicted Correlations of Judgments in Japanese 

 235

K17s was conducted online and the participants were allowed to return to the experiment website and report their judgments in the same experiment more than once. As noted, for each of (16a) and (16b), the participant is instructed to report whether s/he finds it acceptable at all with the specified MR(X, Y). If the reported judgments by a given participant on every sentence instantiating (12b) is “not acceptable at all”, that participant’s overall judgment counts as the J=no in MR([-cc]S, X, Y):J, in line with the prediction. If any of the reported judgments on any sentence instantiating (12b) is “acceptable at least to some extent”, on the other hand, that participant’s overall judgment counts as the J=yes in MR([-cc] S, X, Y):J. Such judgments have the potential to disconfirm our main prediction; such disconfirmation would happen specifically provided that their judgments are such that, by the tests we have established, X does not lead to NFS effects with DR(X, beta) and Y does not lead to NFS effects with Coref(alpha, Y). The judgments that are considered as indicating that X does not lead to NFS effects with DR(X, beta) are those that constitute  in Figure 1, with the [+cc]S and [-cc] S as in (3), again focusing on sentences with a “transitive o-marking” verb. The judgments that are considered as indicating that Y does not lead to NFS effects with Coref(alpha, Y) are those that constitute  in Figure 1, with the [+cc]S and [-cc] S as in (12). What counts as DR([-cc]S, X, beta):no is straightforward. If a participant’s reported judgment on every instance instantiating (3b) is “Not acceptable at all”, that counts as DR([-cc]S, X, beta):no for that participant; otherwise (i.e., if the participant answers “Acceptable at least to some extent” on any sentence instantiating (3b), even once), that participant’s overall judgment does not count as DR([-cc]S, X, beta):no. What counts as DR([+cc]S, X, beta):yes, however, is not so straightforward because the difference between (3a) and (3b) is that, according to our hypotheses, the former can, but the latter cannot, correspond to an LF representation in which LF(X) c-commands LF(Y), reflecting the general difference between [-cc]S and [+cc]S, as given in (19). (19)

(=Chapter 5: (3f, g)) a. [-cc]S: a schematized PS that cannot correspond to LF(PS) in which LF(X) c-commands LF(Y) b. [+cc]S: a schematized PS that can correspond to LF(PS) in which LF(X) c-commands LF(Y)

(19a) guarantees DR([-cc]S, X, beta):no unless the X induces NFS effects. (19b), on the other hand, does not guarantee DR([+cc]S, X, beta):yes because LF(X) c-commanding LF(beta) is a necessary but not a sufficient condition for FR-based

236 

 Hajime Hoji

DR(X, beta).22 What counts as DR([+cc]S, X, beta):yes, therefore, cannot be a given participant’s judgments that sentences instantiating (3a) are always “Acceptable at least to some extent”. As noted, K17s was conducted online and the participants were allowed to return to the “experiment site” and report their judgments in the same experiment more than once. (Not many participants reported their judgments more than once but some participants did.) While for MR([-cc]S, X, Y):J, the J=no is predicted regardless of how many times the participant reports his/her judgments provided that there are no NFS effects, the J=yes for MR([+cc]S, X, Y):J is not predicted in the same way because LF(X) c-commanding LF(Y) is a necessary but not a sufficient condition for FR-based MR(X, Y). MR([-cc]S, X, Y):no is significant for the purpose of c-command detection only if it is accompanied by MR([+cc]S, X, Y):yes at least sometimes. In K17s, we consider that a participant’s giving “yes” at least 50% of the time counts as J=yes for this purpose.23 In other words, only if a participant’s reported judgment on (a)  sentence(s) instantiating (3a) is “Acceptable at least to some extent” at least 50% of the time, do the participant’s judgments count as DR([+cc]S, X, beta):yes for that participant. The same applies to what counts as Coref([+cc]S, alpha, Y):yes and what counts as BVA([+cc]S, X, Y):yes.

22 Furthermore, the [+cc]S in question does not necessarily correspond to an LF representation in which LF(X) c-commands LF(beta), as pointed out in Chapter 5: note 13. 23 From a theoretical/conceptual perspective, this decision is arbitrary, contrary to what we take as J=no in MR([-cc]S, X, Y):J. From a practical standpoint, however, it is not arbitrary, in the sense that the decision about what we take as J=yes in MR([+cc]S, X, Y):J is based on considerations about maintaining a “good balance” between the degree of reliability of the experimental results and the degree of “effectiveness” in the demonstration. If we adopted a higher figure, e.g., 100%, the combination of MR([-cc]S, X, Y):no and MR([+cc]S, X, Y):yes would constitute the clearest pattern of judgments indicating c-command effects, i.e., MR never being possible without c-command and always being possible with c-command. Considering the possibility that actual sentences instantiating [+cc]S can be quite involved, MR([+cc]S, X, Y):yes under the “100%” criterion arise only if the participant could overcome the possible difficulty stemming from various complications in the sentences in question, and that makes the participant’s MR([-cc]S, X, Y):no all the more significant because the participant, despite being able to overcome the difficulty in question, never accepts sentences instantiating [-cc]S, thereby making the experimental results most reliable. The “100%” criterion, however, we will likely make the number of the “qualified” participants significantly smaller than some other criterion, like “10%”, thereby reducing the degree of “effectiveness” in the demonstration.

Replication: Predicted Correlations of Judgments in Japanese 

 237

2.4 Sub-preliminary Experiments As noted in Section 1, we can assume the researcher, at least “in theory”, to have (i)  unlimited willingness and patience to judge sentences and (ii) familiarity and awareness of what is being checked. We cannot, however, assume the same of a non-researcher. As discussed in Section 2.1, the absence of (i) on the part of non-researchers makes the empirical coverage (significantly) smaller in non-self-experiments than in self-experiments.24 Furthermore, to “compensate” for the absence of (ii) on the part of a non-researcher participant and ensure as much reliability as possible for results of a non-self-experiment, it is necessary to include “components” specific to non-self-experiment. Such “components” often include certain forms of “sub-preliminary-experiments” which test things such as (20) and (21) for each participant.25 (20) whether the participant is paying close attention to the instructions (21) whether the participant understands the intended MR(X, Y) as specified As noted, there is (usually) no need to check (20) and (21) for the researcherparticipant in a self-experiment. After all, s/he is the one who has designed the experiment and ought to know exactly what is meant by MR(X, Y) in question. We also assume that the researcher-participant in a self-experiment knows when s/he is not attentive for whatever reason and only judges the relevant sentences when s/he is attentive (enough). For non-researcher participants in non-self-experiment, on the other hand, we cannot necessarily assume (20) and (21). We can try to assess (20) by conducting “preliminary experiments” (largely) independent of our main concern. In the online non-self-experiments dealing with English, where we tested (1) with native speakers of English, we included a “preliminary experiment” that asked the participants whether the sentence in (22) is “acceptable” under the interpretations in (23a) and (23b).26 24 See Plesniak’s Chapters 7 and 8 of this volume and Plesniak 2022 for other ways to conduct sub-preliminary experiments and general methodological discussion about non-self-experiments. 25 In this chapter, I am leaving aside non-self-experiments on other researchers who have as much of (i) and (ii) as the researcher. 26 The “preliminary experiment” was designed by Daniel Plesniak and has been conducted since the spring of 2019, mostly in a large General Education course at USC, but also in other venues. The on-line non-self-experiments in question were designed following the design of K17s. How K17s was designed, however, is, obviously, not the only way to conduct a demonstration to test the correlational/conditional prediction in (1). Plesniak’s Chapter 8 of this volume illustrates other ways of conducting non-self-experiment and concludes that the particular design chosen for K17s is not necessary for obtaining the predicted experimental results.

238 

 Hajime Hoji

(22) Two teachers each praised a different set of three students. (23) Under the interpretation that: a. the total number of students in question is three b. the total number of students in question is six In the experiments on DR, the intended interpretation for DR(X, beta) (for English sentences instantiating the [-cc]S and [+cc]S analogous to those in (3)) was conveyed to the participants by using a version of (22). If a participant reported that (22) was acceptable under the interpretation in (23a), we took that as indicating that the participant was not attentive or that something like (22) is not a good way to convey the intended interpretation of DR(X, beta) for this participant, or potentially both.27 In K17s there are BVA-related “sub-preliminary experiments” and DR-related ones. Let us first consider BVA-related “sub-preliminary experiments”. As noted, the intended BVA(X, Y) is provided in K17s as in (18), which corresponds to the case of BVA(every local government, it), for Japanese sentences such as one corresponding to ‘Every local government criticized its employees’. (18) Under the interpretation “Having criticized their own employees holds true of every local government.” There are at least as many criticized employees as the number of the local governments.’ In the actual Japanese “instructions” given to the participants, zibun(-no tokoro) is used as corresponding to Y of BVA(X, Y), as in (24).28 (24) A sentence instantiating [+cc]S with BVA(X, Y) corresponding to BVA(every local government, it): (“Zibun-no tokoro-no syokuin-o hihansita toyuu koto-ga own-gen place-gen employee-acc criticized that fact-nom subete-no tihoozititai-ni atehamaru rasii” toyuu all-gen local:government-dat be:true it:seem that imi-de.

27 More than a quarter of the participants (just focusing on the native speakers of English) in fact have reported that (22) is acceptable at least to some extent under the interpretation in (23a). (61 out of 283, in one of the instances of this “preliminary experiment”.) For related, and further, discussion, see Chapter 8 of this volume and Plesniak 2022. 28 See Appendix I below and Hoji 2015: 226–229 for more details.

Replication: Predicted Correlations of Judgments in Japanese 

 239

meanig-with Hihansareta syokuin-wa sukunakutomo tihoozititai-no was:criticized employee-top at:least local:government-gen kazu dake iru.) number as:many:as exist ‘Under the interpretation that having criticized their own employees holds true of every local government’. There are at least as many criticized employees as the number of the local governments.’ So-ko-no syokuin-o subete-no tihoozititai-ga that-place-gen employee-acc all-gen local:government-nom hihansita rasii. criticized it:seems ‘It seems that every local government criticized its employees.’ To check if this way of conveying the intended BVA(X, Y) is effective for a given participant, one type of “sub-preliminary experiment” checks whether a participant can use a-NP (such as asoko ‘that place’) as Y of BVA(X, Y). For most native speakers of Japanese, an a-NP cannot be Y of BVA(X, Y), unlike a so-NP (e.g., soko ‘that place’ or soitu ‘that guy’), as discussed extensively in Ueyama 1998, Hoji 2003, Hoji et al. 2003, Hoji 2015, among other places. We thus expect most speakers to answer “no” to the question of whether a Japanese sentence corresponding to (25a) is acceptable at all under the interpretation in (26).29 (25)

a. Every local government criticized asoko’s employees. b. Every local government criticized soko’s employees.

(26) (Cf. (18).) Under the interpretation “Having criticized their own (=zibun-no tokoro-no) employees holds true of every local government.” There are at least as many criticized employees as the number of the local governments.’ This type of “sub-preliminary experiment” thus checks the effectiveness of this way of conveying the BVA(X, Y) by checking whether a participant reports BVA([+cc]S, X, a-NP):no while reporting BVA([+cc]S, X, so-NP):yes. If a participant gives BVA([+cc]S, X, asoko):yes, that means that the participant is not attentive or this is not an effective way to convey the intended interpretation for BVA(X, Y) for this

29 See Hoji 2015: 7.2 for actual sentences used in Hoji 2015 for this type of “sub-preliminary experiment” and relevant discussion.

240 

 Hajime Hoji

participant, or both, provided that the MR(X, Y) in question indeed qualifies as an instance of BVA(X, Y). To address the question of whether the MR(X, Y) in question indeed qualifies as an instance of BVA(X, Y), we must recall from Chapter 5: Section 1.2 that what is intended by BVA(X, Y) is that the “value” of a singular-denoting expression Y “depends upon” that of a non-singular-denoting expression X. To determine whether something like (26) (see also (24)) is an effective way to convey the intended BVA(X, Y) for the participant in question, it is, therefore, necessary to make sure that the a-NP in question is necessarily singular-denoting.30 If a-NP can be plural-denoting, an apparent BVA([+cc]S, X, a-NP) might arise because it could allow a kind of “collective possession” reading, where the group of local governments in question were criticized by the employees of said local governments, which is compatible with a state of affairs of the particular employees criticizing any particular local government that they work for.31 As in Hoji 2015 and earlier works, where this was checked for so-NPs such as soko ‘it’ or soitu ‘that guy’, K17s tests this by checking whether a participant can accept sentences corresponding to (27), where asoko ‘it’ or aitu ‘that guy’ is used in “__”, taking A and B as its “split antecedents”.32 30 What is used in K17s as Y of BVA(X, Y) are soko ‘it’ and soitu ‘that guy’, and their a-NP counterparts are asoko ‘it’ and aitu ‘that guy’, respectively; see Hoji 2015: 7.2.1 and the references given there, especially those in note 2. K17s does not include sono otoko ‘that man’, as Y of BVA/ Coref(X/alpha, Y). As noted, the use of sono otoko is necessary for me to obtain c-command detection with BVA(X, Y) in line with (1), most of the time. If we include sono otoko in a non-self-experiment, we also need to include its a-NP counterpart ano otoko ‘that man’ for the reasons just addressed. 31 As noted in Chapter 5, I get clear BVA([-cc]S, ano gisi-igai, sono otoko):no, along with clear BVA([+cc]S, ano gisi-igai, sono otoko):yes at Stage 2. My judgments, however, become unclear if I use sono otoko-tati ‘that man and others (or those men)’ (which can have “split antecedence” in contexts such as (27)) in place of sono otoko ‘that man’. (This applies both in the [-cc]S case and in the [+cc]S case; the J of the intended BVA([-cc]S, ano gisi-igai, sono otoko-tati):J is not a clear no and the J of the intended BVA([+cc]S, ano gisi-igai, sono otoko-tati):J is not a clear yes. The judgments are blurry. It seems that what is intended by BVA(ano gisi-igai, sono otoko) can be inferred when sono otoko is replaced by sono otoko-tati in the sense that what is intended by BVA(ano gisi-igai, sono otoko) can be inferred as being compatible with the situation expressed by the “collective” interpretation of BVA(ano gisi-igai, sono otoko-tati), where what is expressed by ano gsi-igai ‘ones other than that engineer’ and what is expressed by sono otoko-tati ‘that man and others (or those men)’ are understood to refer to the same group of individuals. A similar “inferred” interpretation is possible for me even with Y=ano gisi-tati ‘that gisi and others”, in sharp contrast to Y=ano gisi ‘that engineer’. 32 The “split antecedence test” in question, which I will sometimes refer to below as “singular-denoting test”, is analogous to checking whether a speaker clearly rejects “split antecedence” in (i-a), unlike (i-b).

Replication: Predicted Correlations of Judgments in Japanese 

 241

(27) a A told B about __’s C. b. A told B that . . . __ . . . If a participant allows split antecedence in sentences analogous to (27), with asoko ‘it’ in “__”, for example, asoko can be plural-denoting for that participant, and his/her J=yes on BVA([+cc]S, X, asoko) does not necessarily mean that the participant is inattentive or that the specification of the intended BVA(X, Y) is not effective for that participant.33 K17s also has a split antecedence sub-preliminary experiment with sentences analogous to (27), with soko ‘it’ appearing in the “__” position. If a participant allows split antecedence with soko ‘it’, soko can be plural-denoting for that participant. If a participant never allows split antecedence with asoko, we predict that that participant will never report BVA(S, X, asoko):yes.34 That is exactly what we find in K17s. Among the 190 participants who reported judgments on BVA([+cc]S, subete-no N, asoko), 31 participants reported BVA([+cc]S, subete-no N, asoko):yes. There were 14 participants who never allowed split antecedence for either asoko or soko. None of the 14 participants who never allowed “split antecedence” for asoko or soko reported BVA([+cc]S, subete-no  N, asoko):yes. When we check the availability of BVA(X, soko) for the purpose of c-command detection, we specifically check the judgments given by those 14 participants, because BVA(X, soko) for those participants cannot be due to the “collective possession” reading. These participants also seemed attentive (based on the a-NP vs. so-NP test, along with the split antecedence test), so that is not a concern for them. When we check (1), with X=subete-no N ‘every N’ and Y=soko, as in (28), we therefore focus on the reported judgments by those 14 participants in the relevant experiments with DR, Coref, and BVA.

(i)

a. b.

A1 asked B2 about his1+2/her1+2/its1+2 (joint) project. A1 asked B2 about their1+2 (joint) project.

33 The split antecedence sub-preliminary experiments with asoko and aitu are not included in Hoji 2015. I am indebted to Haley Wei (p.c., Spring, 2016) for making me realize that those experiments are needed for the a-NP vs. so-NP sub-preliminary experiment to serve its intended purpose. 34 Regardless of whether the S is [-cc]S or [+cc]S because asoko, unlike soko, can never be Y of FDbased BVA(X, Y).

242  (28)

 Hajime Hoji

Testable correlational/conditional prediction about c-command-detection with BVA(subete-no N, soko), based on DR(subete-no N, beta) and Coref (alpha, soko): Provided that DR([+cc]S, subete-no N, beta):yes ∧ Coref([+cc]S, alpha, soko):yes ∧ BVA([+cc]S, subete-no N, soko):yes; DR([-cc]S, subete-no N, beta):no ∧ Coref([-cc]S, alpha, soko):no → BVA([-cc]S, subete-no N, soko):no

These 14 participants were deemed “qualified” by the sub-preliminary experiments serving as the split antecedence tests for asoko and soko and one serving as the a-NP vs. so-NP test, as just discussed. If we relied on the sub-preliminary experiments serving as the split antecedence tests for aitu and soitu, instead of asoko and soko, when checking (28), the results will be less reliable because whether soko can be used as Y of BVA(X, Y) for a given participant is not checked by those particular sub-preliminary experiments (because they check soitu instead). K17s has another type of sub-preliminary experiment dealing with DR. This experiment checks whether a participant can use ano #-cl-no N ‘those # N’s’ as beta of DR(X, beta). This is effectively checking whether a participant answers yes to the question of whether (the Japanese version of) (29a) can be acceptable at all under the interpretation in (30). (29) a. Every local government criticized those three politicians. b. Every local government criticized three politicians. (30) (Cf. (17).) “Having criticized three politicians holds of every local government, and which three politicians differs depending upon each local government”. The number of the criticized politicians is three times as many as the number of the local governments. We intend DR(X, beta) to have an interpretation like (30) (DR(every local government, three politicians)). For most speakers, ano #-cl-no N ‘those # N’s’ cannot be beta of DR(X, beta), just as (29a) is not acceptable under the interpretation in (30) (in contrast to (29b)) for most speakers of English. This type of sub-preliminary experiment checks the participant’s attentiveness to the (general) instructions or his/her understanding of what is expressed by a statement like (30). Earlier in this subsection, I noted that sub-preliminary-experiments are needed for non-self-experiment to check the attentiveness of a participant and the compatibility of the particular experimental design with that participant. We

Replication: Predicted Correlations of Judgments in Japanese 

 243

addressed the compatibility issue with respect to how the intended MR(X, Y) is conveyed to the participant and whether a given Y of MR(X, Y) can indeed serve as Y of the intended MR(X, Y) for the participant. For the researcher-participant (in a self-experiment), there are, at least in theory, no issues related to attentiveness and how the intended MR(X, Y) is conveyed to the participant, as noted. Whether a given Y of MR(X, Y) can indeed serve as Y of the intended MR(X, Y), for example, whether a given Y for BVA(X, Y) is necessarily singular-denoting, is something that needs to be checked for the researcher-participant (in self-experiment) as well. It is in fact not impossible for me to have split antecedence with soko ‘it’, though I cannot have it with soitu ‘that guy’ or sono otoko ‘that man’. I can, for example, accept a sentence analogous to (31), with the understanding that soko stands for an imaginary company that comprises Toyota and Nissan. (31) Toyota told Nissan that the FBI was investigating soko. Such split antecedence is not possible with soitu ‘that guy’ or sono otoko ‘that man’, for me. This might well be due to the difficulty of imaging a person “comprising of two people”. Be that as it may, this indicates that the split antecedence test is, in principle, needed even for the researcher-participant (in self-experiment). One may wonder whether the possibility of soko ‘it’ being “plural-denoting” is the source of my BVA([-cc]S, X, soko):yes, i.e., whether this is the same as my J=yes being due to NFS2 in the terms of Chapter 5: Section 8. There is reason to believe that it is not (or at least not the only such source). First, when I get BVA([-cc]S, X, soitu) at Stage 2 (see Chapter 5: Sections 3 and 4), I still cannot take soitu as “plural-denoting” based on the split antecedence test. The same applies when I get BVA([-cc]S, X, sono otoko) at Stage 3, while not being able to take sono otoko as plural-denoting, based on the split antecedence test. Second, and more importantly, my BVA([-cc]S, X, soko):yes at Stages 1, 2, and 3, my BVA([-cc]S, X, soitu):yes at Stages 1 and 2, and my BVA([-cc]S, X, sono otoko):yes at Stage 3 are quite clear intuitions, similar to equally clear intuitions of my BVA([cc]S, X, soitu):no and BVA([+cc]S, X, soitu):yes at Stage 1 and my BVA([-cc]S, X, sono otoko):no and BVA([+cc]S, X, sono otoko):yes at Stage 2. Such clear intuitions of mine are in sharp contrast to my very blurry intuitions both on the intended BVA([-cc]S, X, sono otoko-tati) and BVA ([+cc]S, X, sono otoko-tati), even with the best choice of X for me for the purpose of c-command detection, i.e., NP-igai ‘ones other than NP’.35 These observations

35 As noted in note 31, sono otoko-tati ‘that man and others (or those men)’ can have split antecedence in contexts such as (27), in contrast to sono otoko ‘that man’.

244 

 Hajime Hoji

would be rather unexpected if my BVA([-cc]S, X, soko):yes at Stages 1, 2, and 3 is due to soko being “plural-denoting”.36 In summary, “sub-preliminary experiments” are, generally speaking, meant to serve two distinct purposes, namely to assess (i) the attentiveness of a participant and (ii) the compatibility between the participant and the design of the experiment. While some of them serve just one purpose – for example, the type of sub-preliminary experiment mentioned in relation to (29) and (30) only tests (i) – in most cases, (i) and (ii) are tested by a combination of these two types of “sub-preliminary experiments in K17s. A sub-preliminary experiment that checks whether a given Y for BVA(X, Y) is necessarily singular-denoting for a given participant (the split antecedence test discussed in relation to (31)) is for checking (ii), but another sub-preliminary experiment that checks whether the participant allow so-NP, but not a-NP, as Y of BVA(X, Y) (the a-NP vs. so-NP test) can serve to check (i) only if the participant in question has “passed” the split antecedence test, disallowing a given choice of Y from being taken as plural-denoting. When testing (1) with Y=soko ‘it’, the relevant sub-preliminary experiments thus include the particular split antecedence tests and the a-NP vs. so-NP test that deal with soko and asoko. When testing (1) with Y=soitu, they include the split antecedence tests and the a-NP vs. so-NP test that deal with soitu and aitu. Such careful checking is part of how we try to ensure the reliability of the result of non-self-experiments as much as possible.37

36 To draw a firmer conclusion about this, we must conduct systematic investigation of different types of NFS (effects), building on the result of our investigation centering around (1), in which NFS effects are something that is to be controlled for and are not the object of our investigation. It is hoped that as the LFS research program advances, NFS (effects) will be subject to investigation by the basic scientific method, building on the success of our research focusing on (1). 37 The types of sub-preliminary experiment just mentioned make reference to choices of X as well as Y, so that we can check the experimental results in a “fine-tuned” way, paying attention to issues of compatibility between a given participant and the design of the experiment, often in relation to particular choices of X and Y, thereby enhancing the reliability of the result of non-self-experiment. The ensuing discussion, however, only addresses the choices of Y because we seem to be able to detect their effects more clearly than the effects of the choices of X included in K17s, as discussed in Chapter 5. Whether or not we use the same Y in the sub-preliminary experiments as the Y in BVA(X, Y) in (1) can result in a difference, as will be discussed in Section 3.

Replication: Predicted Correlations of Judgments in Japanese 

 245

3 Results of K17s 3.1 Checking the Correlational/Conditional Prediction 3.1.1 How We can Check the Result of Non-self-experiments In K17s, we check the prediction in (1) for each participant, based on the schemata, repeated here. (1)

Testable correlational/conditional prediction about c-command-detection with BVA(X, Y), based on DR(X, beta) and Coref(alpha, Y): Provided that DR([+cc]S, X, beta):yes ∧ Coref([+cc]S, alpha, Y):yes ∧ BVA([+cc]S, X, Y):yes; DR([-cc]S, X, beta):no ∧ Coref([-cc]S, alpha, Y):no → BVA([-cc]S, X, Y):no

(3)

For DR(X, beta), with a “transitive o-marking” verb: a. [+cc]S: beta-o X-ga V b. [-cc]S: beta-ga X-o V

(4)

For DR(X, beta), with a “transitive ni-marking” verb: c. [+cc]S: beta-ni X-ga V d. [-cc]S: beta-ga X-ni V

(12) For BVA(X, Y) and Coref(alpha, Y), with a “transitive o-marking” verb a. [+cc]S: [Y-no N]-o alpha/X-ga V b. [-cc]S: [Y-no N]-ga alpha/X-o V (13) For BVA(X, Y) and Coref(alpha, Y), with a “transitive ni-marking” verb a. [+cc]S: [Y-no N]-ni alpha/X-ga V b. [-cc]S: [Y-no N]-ga alpha/X-ni V The actual testing/checking in K17s involves specifying various choices for X and Y in (14) and (15), choices for beta, alpha and their “internal constitution”, as well as choices for V, all of which are given/repeated below.38

38 Although there is a great deal of details to be addressed in relation to many of these choices, the relevant discussion would take us too afield; see Hayashishita and Ueyama 2012, Hoji et al. 2003, Mukai 2012, and the references therein. It is hoped that various issues such as those can

246 

 Hajime Hoji

(14)

Choices of X for BVA/DR(X, Y/beta) in K17s: a. conjoined NPs b. subete-no N ‘every N’ c. #-cl-no N ‘# Ns’ d. sukunakutomo #-cl-izyoo-no N ‘at least # or more N’ e. the FQ version of (d) = N-cm sukunakutomo #-cl-izyoo

(15)

Choices of Y for BVA(X/alpha, Y) in K17s: a. soko b. soitu

Given in (32) are the actual instances of conjoined NP’s used in K17s, and the actual N’s in (14b-e) used in K17s are given in (33). (32)

The conjoined NPs used as X for BVA/DR(X, Y/beta) in K17s: a. asoko to koko ‘that place and this place b. aitu to koitu ‘that guy and this guy’

(33)

The N’s used in K17s for (14b-e): a. tihoo zititai ‘local government’ b. kaisya ‘company’ c. zyooin giin ‘senator’ d. kokkai giin “parliament member’ e. kyuudan ‘ballclub’

The choices for other “variables” used in K17s are given in (34)–(37). (34) N of [Y-no N] in (12) and (13): a. syokuin ‘employee’ b. bengosi ‘attorney’ c. daiiti hisyo ‘chief secretary’ d. kantoku ‘manager (as of a ballclub, for example)’

be looked into by applying the basic scientific method, on the basis of the effective and reliable means for c-command detection.

Replication: Predicted Correlations of Judgments in Japanese 

 247

(35) V in (3), (4), (12), and (13): a. “transitive no-marking” verb hihansita ‘criticized’ b. “transitive ni-marking” verb rihuzin-na yookyuu-o sita ‘made an unreasonable demand upon’ (36) beta in (3) and (4): a. dareka ‘someone’ b. 3-nin-no seizika ‘three policians’ (37) alpha in (12) and (13): a. koko ‘this place/institution/etc.’ b. asoko ‘that place/institution/etc.’ c. koitu ‘this guy’ d. aitu ‘that guy’ As discussed, the claim made in (1) is as follows: If we obtain at a given time from a given speaker DR([+cc]S, X, beta):yes and DR([-cc]S, X, beta):no with a specific choice of X (which is  in Figure 1 given in Section 2.1), and Coref([+cc]S, alpha, Y):yes and Coref([-cc]S, alpha, Y):no with a specific choice of Y (which is  in Figure 1), then, with those choices of X and Y, BVA([-cc]S, X, Y):no obtains for that speaker at that time. If we obtain BVA([+cc]S, X, Y):yes, in addition, that constitutes c-command detection with BVA(X, Y) with the choices of X and Y, for that speaker at that time.39 This applies to any choices of X and Y. Let us examine a specific instance of (1) in order to illustrate how K17s is designed and how we check the correlational/conditional prediction of the form in (1). Consider the case of DR(subete-no N, 3-nin-no seizika) ‘DR(every N, three politicians)’, Coref(aitu, soitu) ‘Coref(that guy (with the a-demonstrative), that guy (with the so-demonstrative))’, and BVA(subete-no N, soitu) ‘BVA(every N, that guy)’. With these choices of X and Y, and of alpha and beta, we are then considering the correlational/conditional prediction in (38), as a specific instance of (1).

39 Because our focus is on obtaining c-command detection, we have added the third conjunct in the “Provided that . . .” clause in (1) so that what is entailed in (1) is indeed c-command detection.

248 

 Hajime Hoji

(38) Correlational/conditional prediction about c-command-detection with BVA(subete-no N, soitu), based on DR(subete-no N, 3-nin-no seizika) and Coref(aitu, soitu): Provided that DR([+cc]S, subete-no N, 3-nin-no seizika):yes ∧ Coref([+cc]S, aitu, soitu):yes ∧ BVA([+cc]S, subete-no N, soitu):yes; DR([-cc]S, subete-no N, 3-nin-no seizika):no ∧ Coref([-cc]S, aitu, soitu):no → BVA([-cc]S, subete-no N, soitu):no For a given speaker, we can thus check his/her J of each of (39) and (40), to see if they are in line with the correlational/conditional prediction in (38). (39) a. DR([+cc]S, subete-no N, 3-nin-no seizika):J b. Coref([+cc]S, aitu, soitu):J c. BVA([+cc]S, subete-no N, soitu):J (40) a. DR([-cc]S, subete-no N, 3-nin-no seizika):J b. Coref([-cc]S, aitu, soitu):J c. BVA([-cc]S, subete-no N, soitu):J According to (38), if the speaker’s J on any of (39) is no, we do not make the correlational/conditional prediction in (38). Similarly, if his/her J on (40a) or (40b) is yes, we do not make the prediction either. Only if his/her J on all of (39) is yes, and if J=no on both (40a) and (40b), we make the prediction that his/her J on (40c) is no. We thus first check, for each speaker, (i) whether the J on all of (39) is yes, (ii) whether the J on (40a) is no, and (iii) whether the J on (40b) is no, and if the participant “passes” all these “tests”, we can check his/her J on (40c), to see if it is in line with the correlational/conditional prediction in (38). We can visualize this as follows: For a given speaker,40 we check whether s/he is a member of the set indicated by the circle in (41).41 I belong to this set myself.

40 “Speaker”, “individual”, and “participant” are used interchangeably here. 41 I will refer to the ellipse shape as a “circle” to simplify the exposition.

Replication: Predicted Correlations of Judgments in Japanese 

 249

(41)

Those speakers whose J=yes in all of: DR([+cc] S, subete-no N, 3-nin-no seizika):J Coref([+cc]S, aitu, soitu):J BVA([+cc] S, subete-no N, soitu):J

We then check whether the speaker is a member of the set indicated by the circle in (42) and whether s/he is a member of the set indicated by the circle in (43). (42)

Those speakers whose J=no in: DR([-cc]S, subete-no N, 3-nin-no seizika):J

(43)

Those speakers whose J=no in: Coref([-cc]S, aitu, soitu):J

250 

 Hajime Hoji

I am not a member of either of the sets in (42) and (43). We make the testable (i.e., disconfirmable) prediction in (38) for a given speaker only if the speaker belongs to the middle intersection in (44). Since I am not one of those speakers, with my membership in these sets being as indicated by the position of the smile mark in (44), the correlational/conditional prediction in (38) is not applicable to me; it is only applicable to those in the intersection of all three of these “circles”. (44) Those speakers whose J=yes in all of: DR([+cc]S, subete-no N, 3-nin-no seizika):J Coref([+cc]S, aitu, soitu):J BVA([+cc]S, subete-no N, soitu):J

Those speakers whose J=no in: DR([-cc]S, subete-no N, 3-nin-no seizika):J

Those speakers whose J=no in: Coref([-cc]S, aitu, soitu):J

To summarize, we check each participant’s reported judgments and determine whether s/he belongs to the center intersection in (44). If s/he does, we can test the correlational/conditional prediction in (38) for that participant by checking his/her J on (40c). We do this classification and checking with every combination of X and Y and for every participant. The correlational/conditional prediction in (1) is that we will not find a single case where a speaker in the center circle in (45) has BVA([-cc] S, X, Y):yes, provided that the speaker is attentive, understands what is meant by the instructions intended to convey DR(X, alpha) and BVA(X, Y), and there are no “compatibility issues” between the speaker and aspects of the design of the experiment, based on results of the sub-preliminary-experiments (see Section 2.4).

Replication: Predicted Correlations of Judgments in Japanese 

 251

(45) Those speakers whose J=yes in all of: DR([+cc]S, X, beta):J Coref([+cc]S, alpha, Y):J BVA([+cc]S, X, Y):J

Those speakers whose J=no in: DR([-cc]S, X, beta):J

Those speakers whose J=no in: Coref([-cc]S, alpha, Y):J

For ease of exposition, if a speaker is determined by the results of sub-preliminary-experiments to be attentive, to understand the instructions, and to not have any compatibility issues with aspects of the design of the experiment, let us refer to such an individual as a “qualified speaker”. For each combination of X and Y used in K17s, we first check where in the graph in (46) each participant belongs. (46) Those speakers who are qualified and whose J=yes in all of: DR([+cc]S, X, beta):J Coref([+cc]S, alpha, Y):J BVA([+cc]S, X, Y):J

Those speakers whose J=no in: DR([-cc]S, X, beta):J

Those speakers whose J=no in: Coref([-cc]S, alpha, Y):J

252 

 Hajime Hoji

If a participant belongs in the center intersection, we check this participant’ J on BVA([-cc]S, X, Y):J to see if the prediction J=no is borne out. With every single combination of X and Y used in K17s, the prediction is indeed borne out for all participants who belong in the center intersection. Recall that (1) does not make a  direct prediction for a participant who does not belong in the center intersection.42 Suppose that there are 180 speakers who have reported judgments on the sentences relevant to (46), which involve questions on DR(X, beta), Coref(alpha, Y), and BVA(X, Y), with particular pairs of X and Y, as well as sub-preliminary experiments that are intended to check (i) the attentiveness of the participant and (ii) the “compatibility” between the participant and the design of the experiment. We then have 180 graphs like (46), one graph for each participant, indicating where in the graph the participant belongs, for that particular pair of X and Y, based on the specific “threshold” chosen for determination of J=yes in the top circle in (46).43

42 With many of the combinations of X and Y used in K17s, it is not even necessary to add the “specification” of the “qualified (speaker)” in the top circle in order to find reliable c-command detection with BVA(X, Y); that is, in those cases, looking at the individuals who are in the intersection of the two bottom circles, regardless of whether or not they are in the top one, results in finding no individuals who report BVA([-cc]S, X, Y):yes. It is, however, necessary to do so when we consider the top circle with at least some of the combinations of X and Y; otherwise, individuals who appear to disconfirm the predictions are found. One may take the fact that reference to the top circle is not always necessary as suggesting that the reference to the correlation in (1) alone is sufficient to ensure the reliability of a participant (with respect to the experiment in question), at least for the purpose of c-command detection. The more sub-preliminary experiments we make use of for the determining whether a given participant is “qualified”, the smaller the number of the participants in the top circle is likely to become, thus shrinking the number in the center intersection, which reduces the number of participants we make predictions about. It is necessary to consider what might be the “optimal use” of sub-preliminary experiments (including which (types of the) of sub-preliminary experiments are effective for ensuring the participant “quality”); see Chapter 8 of this volume for much related discussion. 43 The threshold that we used in analyzing the Kyudai17s result was 50, as discussed at the end of Section 2.3. The number of possible choices regarding the relevant “threshold” to be chosen is, in principle, infinite. If we consider that, the possible number of the graphs in question is, therefore, infinite even with a particular set of choices of X/alpha and Y/beta. How we can arrive at the optimal threshold for the determination of J=yes in question is among what needs to be addressed when we try to articulate the methodology for demonstration of the type that is being considered here. See Chapter 8 of this volume and Plesniak 2022 for related discussion.

Replication: Predicted Correlations of Judgments in Japanese 

 253

As noted, we make the correlational/conditional prediction in (1) with regard to participants who belong in the center intersection. For a concrete illustration, let us go back to the choices of X and Y that lead the correlational/conditional prediction in (38), repeated here. (38) Correlational/conditional prediction about c-command-detection with BVA(subete-no N, soitu), based on DR(subete-no N, 3-nin-no seizika) and Coref(aitu, soitu): Provided that DR([+cc]S, subete-no N, 3-nin-no seizika):yes ∧ Coref([+cc] S, aitu, soitu):yes ∧ BVA([+cc]S, subete-no N, soitu):yes; DR([-cc]S, subete-no N, 3-nin-no seizika):no ∧ Coref([-cc]S, aitu, soitu):no → BVA([-cc]S, subete-no N, soitu):no A graph like (47) allows us to indicate where in the graph each participant belongs, as well as whether his/her J is yes or no in BVA([-cc]S, subete-no N, soitu):J and in BVA([+cc]S, subete-no N, soitu):J, such as the ones in (48), for example, which are instances where (47) has been specified with participant-specific information about where they stand with relation to the circles (though not yet what their judgments on the crucial BVA sentences are, which is why they are given as just “J”). In (48)–(50), I revert to the use of X, Y, alpha, and beta in place of their specific choices in (38), to make the graphs manageable spacewise. (47) Those speakers who are qualified and whose J=yes in all of: DR([+cc]S, X, beta):J Coref([+cc]S, alpha, Y):J BVA([+cc]S, X, Y):J

Those speakers whose J=no in: DR([-cc]S, X, beta):J

Those speakers whose J=no in: Coref([-cc]S, alpha, Y):J

254 

 Hajime Hoji

(48) a. Those speakers who are qualified and whose J=yes in all of: DR([+cc]S, X, beta):J Coref([+cc]S, alpha, Y):J BVA([+cc]S, X, Y):J BVA([+cc]S, X, Y):J BVA([-cc]S, X, Y):J

Those speakers whose J=no in: DR([-cc]S, X, beta):J

Those speakers whose J=no in: Coref([-cc]S, alpha, Y):J

b. Those speakers who are qualified and whose J=yes in all of: DR([+cc]S, X, beta):J Coref([+cc]S, alpha, Y):J BVA([+cc]S, X, Y):J

Those speakers whose J=no in: DR([-cc]S, X, beta):J

BVA( BVA(

S, X, Y):J S, X, Y):J

Those speakers whose J=no in: Coref([-cc]S, alpha, Y):J

Replication: Predicted Correlations of Judgments in Japanese 

 255

c. Those speakers who are qualified and whose J=yes in all of: DR([+cc]S, X, beta):J Coref([+cc]S, alpha, Y):J BVA([+cc]S, X, Y):J

Those speakers whose J=no in: Coref([-cc]S, alpha, Y):J

Those speakers whose J=no in: DR([-cc]S, X, beta):J BVA( BVA(

S, X, Y):J S, X, Y):J

d. Those speakers who are qualified and whose J=yes in all of: DR( S, X, beta):J Coref( S, alpha, Y):J BVA( S, X, Y):J

BVA( BVA(

Those speakers whose J=no in: DR( S, X, beta):J

S, X, Y):J S, X, Y):J

Those speakers whose J=no in: Coref( S, alpha, Y):J

The prediction in (1) is as indicated in (49); namely, if a participant is as in (48d), the particular J of their BVA judgments ought to be as indicated in (49).

256 

 Hajime Hoji

(49) Those speakers who are qualified and whose J=yes in all of: DR([+cc]S, X, beta):J Coref([+cc]S, alpha, Y):J BVA([+cc]S, X, Y):J

BVA( BVA(

S, X, Y):yes S, X, Y):no

Those speakers whose J=no in: Coref([-cc]S, alpha, Y):J

Those speakers whose J=no in: DR([-cc]S, X, beta):J

If we were to obtain (50), that would disconfirm the specific correlational/conditional prediction in (38), and hence the general correlational/conditional prediction in (1). (50) Those speakers who are qualified and whose J=yes in all of: DR([+cc]S, X, beta):J Coref([+cc]S, alpha, Y):J BVA([+cc]S, X, Y):J

BVA( BVA(

Those speakers whose J=no in: DR([-cc]S, X, beta):J

S, X, Y):yes S, X, Y):yes

Those speakers whose J=no in: Coref([-cc]S, alpha, Y):J

Replication: Predicted Correlations of Judgments in Japanese 

 257

(38) does not make any direct predictions about the judgments by the participants  who do not belong in the center intersection, as in the case of (48a, b, c), for example.44 To summarize, the claim made by the prediction in (1) is that if given choices of X and Y “place” a given speaker in the center intersection in (48d), we necessarily obtain BVA([-cc]S, X, Y):no from this speaker.45 In other words, if the combination of judgments as indicated in (51) were to obtain with particular choices of X and Y, with any speaker who is deemed “qualified”, that would show the failure of the predicted entailment to hold, hence constituting disconfirmation of the prediction made in (1). (51) A combination of judgments that would disconfirm the prediction in the form of (1): a. DR([+cc]S, X, beta):yes b. DR([-cc]S, X, beta):no c. Coref([+cc]S, alpha, Y):yes d. Coref([-cc]S, alpha, Y):no e. BVA([+cc]S, X, Y):yes f. BVA([-cc]S, X, Y):yes This possibility of the experimental disconfirmation of our prediction is what makes the set of hypotheses chosen rigorously testable.

44 That, however, does not mean that the judgments of those speakers do not have any role to play in LFS. In fact, the checking of the contrapositive of (38) addresses the judgment of a speaker who belongs to a “section” of the top circle, other than the middle intersection, such as the speaker in (48a). Furthermore, it is the judgments of those speakers just alluded to that will play the most important role when we start investigating why they get MR([-cc]S, X, Y):yes. Further, it should be recalled that, while reference is being made to “speaker”, as has been pointed out, we are addressing a speaker at a given time; the same speaker might belong to different “sections” of the graph, depending upon what “Stages” that person is at (in the sense of Chapter 5: Sections 3 and 4). 45 If we do not, something must be wrong with some of the hypotheses that have led to the prediction, some aspect of the design of the experiment, or both.

258 

 Hajime Hoji

3.1.2 Results Let us consider the result of checking the specific correlational/conditional prediction in (38), repeated here. (38) Correlational/conditional prediction about c-command-detection with BVA(subete-no N, soko), based on DR(subete-no N, 3-nin-no seizika) and Coref(aitu, soko): Provided that DR([+cc]S, subete-no N, 3-nin-no seizika):yes ∧ Coref([+cc] S, aitu, soitu):yes ∧ BVA([+cc]S, subete-no N, soitu):yes; DR([-cc]S, subete-no N, 3-nin-no seizika):no ∧ Coref([-cc]S, aitu, soitu):no → BVA([-cc]S, subete-no N, soitu):no As discussed at the end of the preceding subsection, if we obtained from a “qualified” speaker the set of judgments indicated in (52), that would demonstrate the failure of the predicted entailment in (38) to hold. (52) A combination of judgments that would disconfirm the prediction in (38): a. DR([+cc]S, subete-no N, 3-nin-no seizika):yes b. DR([-cc]S, subete-no N, 3-nin-no seizika):no c. Coref([+cc]S, aitu, soko):yes d. Coref([-cc]S, aitu, soko):no e. BVA([+cc]S, subete-no N, soko):yes f. BVA([-cc]S, subete-no N, soko):yes (52a-e) are the content of the “Provided that . . .” clause and the conditional in (38). What is checked here is then whether or not (52f) obtains for a “qualified” speaker who has reported (52a-e). If it does, that will disconfirm the prediction in (38); if it does not, that is, if BVA([-cc]S, subete-no N, soko):no obtains, that will constitute c-command detection with BVA(subete-no N, soko). We consider every speaker who participated in all the relevant non-selfexperiments as addressed in (52), where sentences instantiating the [+cc]S and the [-cc] S are included for each of DR, Coref, and BVA. We then check, and indicate for each speaker, where in the graph in (53) the speaker belongs to.

Replication: Predicted Correlations of Judgments in Japanese 

 259

(53) A: Those speakers whose J=yes in all of: DR([+cc]S, subete-no N, 3-nin-no seizika):J Coref([+cc]S, aitu, soitu):J BVA([+cc]S, subete-no N, soitu):J

A

AC

AB

B: Those speakers whose J=no in: DR([-cc]S, subete-no N, 3-nin-no seizika):J

ABC

B

BC

C

C: Those speakers whose J=no in: Coref([-cc]S, aitu, soitu):J

If the speaker belongs to ABC, i.e., if the speaker’s reported judgments are as indicated in (53) (=(52a-e)), we make the prediction in (54) about that speaker as in (54), which (52f) would disconfirm. (54) Prediction about a speaker in ABC in (53): BVA([-cc]S, subete-no N, soitu):no Given that our concern is with c-command-detection, we also want to indicate whether a speaker’s reported judgments on the experiment would count as (i) potential c-command detection, (ii) potential disconfirmation of the prediction, or (iii) neither, regardless of which “circle” a speaker is in. In line with the discussion in Section 2.3, we adopt (55) for the determination of what counts as MR([-cc]S, X, Y):no and what counts as MR([+cc]S, X, Y):yes in relation to (52), (53) and (54) in K17s. (55)

How we determine what counts as MR([-cc]S, X, Y):no and what counts as MR([+cc]S, X, Y):yes in K17s:46 a. Only if a participant’s reported judgment on every token of sentence instantiating MR([-cc]S, X, Y) is “Not acceptable at all”, does that participant’s overall judgment count as “MR([-cc]S, X, Y):no”; otherwise, the overall judgment counts as “MR([-cc]S, X, Y):yes” for that participant.

46 How yes and no are distinguished in (55a) is for a principled reason, but how yes and no are distinguished in terms of “50%” in (55b) is somewhat arbitrary; see note 23. I am stating (55b) as given here, to make the ensuing exposition about results of experiments concrete.

260 

 Hajime Hoji

b. Only if a participant’s reported judgment on token(s) of sentence instantiating MR([+cc]S, X, Y) is “Acceptable at least to some extent” at least 50% of the time, does that participant’s overall judgment count as “MR([+cc]S, X, Y):yes” for that participant; otherwise, the overall judgments counts as “MR([+cc]S, X, Y):no” for that participant. Based on (55), we can determine (56) for every speaker who has participated in all the relevant experiments “mentioned” in (38). (56)

a. where s/he belongs in the graph in (53) b. what judgments s/he has given on BVA([+cc]S, subete-no N, soko) and BVA([-cc]S, subete-no N, soko)

If a speaker belongs to ABC (i.e., the middle intersection) and reported judgments as in (57), this counts as c-command detection. If a speaker reported judgments as in (58) and belongs to ABC, this counts as disconfirmation of the prediction in (54). (57) a. BVA([+cc]S, subete-no N, soko):yes b. BVA([-cc]S, subete-no N, soko):no (58) BVA([-cc]S, subete-no N, soko):yes In the case at hand, i.e., in the part of K17s that tests (38), no speaker inside of ABC has the pattern of judgments in (58), and indeed, they all have the pattern in (57), demonstrating c-command detection across all such individuals. For our visual presentation of experimental results in line with the preceding discussion, we represent each speaker whose reported judgments are as in (57) with a green circle, each speaker whose reported judgments are as in (58) with a red star (“✶”), and any other speakers by a blue “x” (see below).47 This is summarized in (59) for testing (38) in particular, and its generalized version is given in (60).48

47 We give both the color (green, red, and blue) and shape (circle, ✶, and x) because the differences in color may not always be available. 48 It is important to remember, for a proper understanding of the result graphs to be given below, that each speaker will be represented in one of these three ways, regardless of which circle they might belong in; the classification of speakers into the three types as in (59) is independent of the classification of speakers in terms of which circle they belong in.

Replication: Predicted Correlations of Judgments in Japanese 

(59)

 261

How we visually present experimental results for testing (38): BVA([+cc]S, subete-no N, soitu):yes BVA([-cc]S, subete-no N, soitu):no

*

BVA([-cc]S, subete-no N, soitu):yes

x

BVA([+cc]S, subete-no N, soitu):no BVA([-cc]S, subete-no N, soitu):no

(60) How we visually present experimental results for testing (1). BVA([+cc]S, X, Y):yes BVA([-cc]S, X, Y):no

*

BVA([-cc]S, X, Y):yes

[+cc] x BVA([-cc] S, X, Y):no

BVA(

S, X, Y):no

As indicated in (60), a speaker is represented with a blue “x” if his/her reported judgments are BVA([-cc]S, X, Y):no, hence not disconfirming our prediction, combined with BVA([+cc]S, X, Y):no, hence not counting as c-command detection. As indicated in (61), there is no red star in ABC, hence our disconfirmable prediction in (38) (see also (54)) is not disconfirmed, and furthermore, there are only green circles in ABC, hence we obtain c-command detection from every relevant speaker.49

49 As noted earlier, I would belong outside B at any of Stages 1, 2, or 3, and outside C at Stages 2 and 3; as such, I belong outside ABC at any of the three Stages discussed in Chapter 5. (38), therefore, does not make any prediction about my judgments about BVA([-cc]S, subete-no N, soitu) at any of those Stages, though as might be expected, I would likely show up as a red star.

262  (61)

 Hajime Hoji

a.

b. Results

Green

Blue

Red

Total

ABC

4

0

0

4

AB

3

0

0

3

AC

1

0

0

1

BC

7

82

11

100

A

0

0

12

12

B

5

21

7

33

C

5

17

5

27

None

1

2

3

6

Total

26

122

38

186

The chart in (61b) indicates, under the “Total” column, how many speakers “show up” in each of the sections in the graph (see (53) for the partitioning), and how many of those speakers in each section are Green (circles), Blue (xs) and Red (stars). It is important to recall that the graph in (61a) is a collection of 186 graphs

 263

Replication: Predicted Correlations of Judgments in Japanese 

(placed on top of each other), such as those given in (62), where each graph represents which section a given speaker belongs in and what judgments the speaker reports (in terms of our criteria for J=yes and J=no given in (55)). Recall too that what is predicted by (38) is the absence of the Red star in the middle intersection in any of these graphs. (62)

The graph in (61a) is a composite of 186 graphs like the following. *

x

*

*

x

x

* x

*

...

For the checking of the prediction in (38), the crucial graphs are then those in which the dot appears in the middle “ABC” intersection and hence the figures in the row ABC in (61b). In the above “analysis” of the experimental results, we paid no attention to judgments reported on sentences of the (16c) type, i.e., sentences identical to the (16b) type ([-cc]S) but without involving the MR in question. We nonetheless obtained c-command detection with BVA(subete-no N, soitu), based on DR(subete-no N, 3-nin-no seizika) and Coref(aitu, soitu). If we included judgments on sentences of the (16c) type, e.g., focusing on participants who reported “yes” on sentences of the (16c) type 50% of the time or even 100% of the time, we will have exactly the same individuals inside the ABC intersection. That is what happens generally; reference to the judgments on sentences of the (16c) type does not seem to affect the crucial part of the result, except that it only very occasionally reduces the number of speakers in the ABC intersection, which is as expected.

264 

 Hajime Hoji

This contrasts sharply with the sub-preliminary experiments and the effects they can have, as will be discussed immediately below.50

3.1.3 Some Combinations of X and Y Require Reference to “Sub-preliminary” Experiments It is remarked in note 42 that “[w]ith many of the combinations of X and Y used in K17s, it is not even necessary to add the “specification” of the “qualified (speaker)” in the top circle in order to find reliable c-command detection with BVA(X, Y); that is, in those cases, looking at the individuals who are in the intersection of the two bottom circles, regardless of whether or not they are in the top one, results in finding no individuals who report BVA([-cc]S, X, Y):yes. It is, however, necessary to do so when we consider the top circle with at least some of the combinations of X and Y; otherwise, individuals who appear to disconfirm the predictions are found.”. As just illustrated above, the combination of subete-no N and soitu is one of the combinations that do not require the use of sub-preliminary experiments, such as a split antecedence test (see Section 2.4) for determining whether a given participant is “qualified (speaker)” (belonging in Circle A) in order to find participants in the center circle and obtain c-command detection. The combination of subete-no N and soko does, however. With the combination of subete-no N and soko, we are considering the prediction in (63). What is different in (63) from (38) is indicated in bold in (63). (63) Correlational/conditional prediction about c-command-detection with BVA(subete-no N, soko), based on DR(subete-no N, 3-nin-no seizika) and Coref(asoko, soko): Provided that DR([+cc]S, subete-no N, 3-nin-no seizika):yes ∧ Coref([+cc]S, asoko, soko):yes ∧ BVA([+cc]S, subete-no N, soko):yes; DR([-cc]S, subete-no N, 3-nin-no seizika):no ∧ Coref([-cc]S, asoko, soko):no → BVA([-cc]S, subete-no N, soko):no If we do not consider the “qualified speaker” criterion, then what is in the A, B, and C circles in the graph representing the experimental results related to (63) is as indicated in (64). What is represented by different colors/shapes in the graph is indicated in (65), but this is independent of considerations of “qualification”.

50 In the ensuing discussion, we will continue to not refer to judgments on sentences of the (16c) type.

Replication: Predicted Correlations of Judgments in Japanese 

 265

(64) A circle: DR([+cc]S, subete-no N, 3-nin-no seizika):yes Coref([+cc]S, asoko, soko):yes BVA([+cc]S, subete-no N, soko):yes B circle: DR([-cc]S, subete-no N, 3-nin-no seizika):no C circle: Coref([-cc]S, asoko, soko):no (65) Red star: BVA([-cc]S, subete-no N, soko):yes Green circle: BVA([+cc]S, subete-no N, soko):yes and BVA([-cc]S, subete-no N, soko):no Blue x: Neither Red nor Green (66) a.

266 

 Hajime Hoji

b.51 Results

Green

Blue

Red

Total

ABC

3

0

2

5

AB

4

0

2

6

AC

2

0

0

2

BC

13

70

13

96

A

0

0

9

9

B

5

16

7

28

C

7

10

7

24

None

2

4

3

9

Total

36

100

43

179

Two out of the five speakers in ABC gave BVA([-cc]S, subete-no N, soko):yes. While that is contrary to the prediction in (63), our prediction is about “qualified speakers” not about every speaker. If we refer to results of three sub-preliminary experiments (the asoko vs. soko test, and two split antecedence tests, one with asoko and one with soko), and include “passing” these tests as part of the criteria for A, as indicated in (67), we obtain the result indicated in (68).51 (67) A circle: DR([+cc]S, subete-no N, 3-nin-no seizika):yes Coref([+cc]S, asoko, soko):yes BVA([+cc]S, subete-no N, soko):yes

51 The fact that the total number is 186 in (61b) but it is 179 in (66b) perhaps calls for an explanation. As implied by the preceding discussion, the non-self-experiment conducted on each of around 180 students consisted of a number of “component experiments”, such as one testing the possibility of DR(subete-no N, 3-nin-no seizika), one testing the possibility of Coref(aitu, soitu), and one testing the possibility of BVA(subete-no N, soitu), in the case of the testing of the correlational prediction in (38). See Appendix II for actual sentences for the three “component experiments” in question. The number of the component experiments in the non-self-experiment in Kyudai17s was more than 100, including many “sub-preliminary experiments”; the students were asked to participate in them in the course of the semester, and some students did not participate in some of those “component experiments”. The total number of 186 in (61b) means that 186 students participated in all the three “component experiments” (for the testing of the correlational prediction in (38). When we changed soitu ‘that guy’ in BVA(subete-no N, soitu) to soko ‘it’, and consider BVA(subete-no N, soko), not only do we change the “component experiment” on BVA(X, Y), but also the one on Coref(alpha, Y) because soko ‘it/the place’ requires a different choice of alpha than soitu ‘that guy’ (the latter is for a human while the former is not). When we check the correlational prediction in (63), we use the same “component experiment” on DR, but different “component experiments” on BVA and Coref. The total number of 179 means the number of students who participated in those three “component experiments”.

Replication: Predicted Correlations of Judgments in Japanese 

 267

Passing the asoko vs. soko test Passing the split antecedence test with soko Passing the split antecedence test with asoko B circle: DR([-cc]S, subete-no N, 3-nin-no seizika):no C circle: Coref([-cc]S, asoko, soko):no (68)

The result of checking (63), with three “sub-preliminary experiments” (the asoko vs. soko test, and the split antecedence test with asoko and the one with soko): a.

268 

 Hajime Hoji

b. Results

Green

Blue

Red

Total

ABC

1

0

0

1

AB

0

0

1

1

AC

1

0

0

1

BC

15

70

15

100

A

0

0

0

0

B

9

16

8

33

C

8

10

7

25

None

2

4

12

18

Total

36

100

43

179

There is now only one speaker in ABC, and that speaker’s reported judgments do not disconfirm the prediction in (63); in fact, they constitute c-command detection with BVA(subete-no N, soko). There being only one speaker in ABC is of no concern. The judgment(s) of this speaker being in line the predicted entailment is what is significant. If this speaker had given BVA([-cc]S, subete-no N, soko):yes, that would be a source of serious concern because that would be an “existential” demonstration of the failure of the predicted entailment to hold, provided that the speaker is indeed a “qualified speaker”. Consider again the result in (61). As noted, this result is not based on reference to any sub-preliminary experiments. One may thus wonder if the four participants in the center circles in (61) are indeed all qualified speakers. If we refer to three of the sub-preliminary experiments in K17s (the aitu vs. soitu test, and two split antecedence tests, one with aitu and one with soitu), we obtain the results indicated in (69).52

52 These sub-preliminary experiments deal with aitu and soitu (rather than asoko and soko), because (38) deals with experiments where Y of BVA/Coref(X/alpha, Y) is soitu (rather than soko). Likewise, the aitu vs. soitu test uses subete-no N as X of BVA(X, Y) because that is the X used in the BVA experiment in (38). When “wrong choices” are made in this regard when choosing which sub-preliminary experiments to consider when checking a particular prediction with particular choices of X and Y, it can result in the reduction of the effectiveness of sub-preliminary experiments, resulting in apparent disconfirmation. I cannot illustrate the point here due to space limitations.

Replication: Predicted Correlations of Judgments in Japanese 

 269

(69) a.

b.

Results

Green

Blue

Red

Total

ABC

3

0

0

3

AB

1

0

0

1

AC

1

0

0

1

BC

8

82

11

101

A

0

0

0

0

B

7

21

7

35

C

5

17

5

27

None

1

2

15

18

Total

26

122

38

186

Sub-preliminary experiments are intended to allow us to focus on qualified speakers, and because of the use of the sub-preliminary experiments, the result in (69) is more reliable than that in (61), which we obtained without using any of the sub-preliminary experiments. As indicated, the reference to the sub-preliminary

270 

 Hajime Hoji

experiments has resulted in a smaller number of speakers in the middle intersection; four has become three. What is noteworthy is that while the reference to the “sub-preliminary experiments” made the number of speakers in ABC smaller, none of the remaining speakers are reds. With the considerations given in Section 2.4, we consider the result indicated in (69) more reliable than the result indicated in (61). The prediction of the form in (1) is supported in K17s, with every combination of X and Y in K17s. (1)

Testable correlational/conditional prediction about c-command-detection with BVA(X, Y), based on DR(X, beta) and Coref(alpha, Y): Provided that DR([+cc]S, X, beta):yes ∧ Coref([+cc]S, alpha, Y):yes ∧ BVA([+cc]S, X, Y):yes; DR([-cc]S, X, beta):no ∧ Coref([-cc]S, alpha, Y):no → BVA([-cc]S, X, Y):no

As noted, in many cases we seem to obtain c-command detection with BVA(X, Y), based on DR(X, beta) and Coref(alpha, Y), without reference to sub-preliminary experiments. However, even when we obtain a result that is in support of (1) (i.e., no red ✶ and some green circle(s) in ABC) without reference to “sub-preliminary experiments”, we likely want to refer to some sub-preliminary experiments to see if we continue to have some green circle(s) in ABC. If reference to sub-preliminary experiments results in ABC having no speakers, that does not disconfirm our prediction, but it does not provide c-command detection with BVA(X, Y), either; it suggests that the green circles that were in ABC before such reference was made were actually individuals who were not fully “qualified” for their judgments to serve as testing grounds for our predictions, at least by the “qualification” criteria we set. There is, however, a question of how we should determine these criteria. The more sub-preliminary experiments we rely on when analyzing results of non-self-experiments, the more reliable their results can become, at least in principle. But it can also result in a smaller number of speakers in ABC, increasing the chance of “eliminating” evidence of genuine c-command effects.53 Given the main purpose of non-self-experiments like those in K17s is “demonstration”, not for scientific discovery per se, we perhaps want to seek a good balance between the degree of reliability and the “effectiveness” of demonstration.54 If seeking an 53 Sometimes, people fail sub-preliminary experiments because of some limitation of the sub-preliminary experiment (or problem with the reporting, for example, by pressing the wrong radio button by mistake), and that can result in the elimination of possible evidence for genuine c-command effects. 54 What is meant by the “effectiveness” of the demonstration is the degree to which the result of the demonstration might be convincing to others as replication of the result of my self-experiment.

Replication: Predicted Correlations of Judgments in Japanese 

 271

extremely high degree of reliability of the experiment results in no speakers in ABC, for example, that goes against the purpose of demonstration.

3.1.4 More Results There are five choices of X and two choices of Y (soko ‘it’ and soitu ‘that guy’) included in K17s, as indicated in (70) and (70). (70)

a. b. c. d. e.

(71) a. b. c. d. e.

BVA(asoko to koko, soko) ‘BVA(that place and this place, it)’ BVA(subete-no N, soko) ‘BVA(every N, it)’ BVA(#-cl-no N, soko) ‘BVA(# N’s, it)’ BVA(sukunakutomo #-cl-no N, soko) ‘BVA(at least # or more N, it)’ BVA(N-cm sukunakutomo #-cl, soko) (the same as (d) except that the X is in the so-called “floating numeral” form) BVA(aitu to koitu, soitu) ‘BVA(that guy and this guy, that guy)’ BVA(subete-no N, soitu) ‘BVA(every N, that guy)’ BVA(#-cl-no N, soitu) ‘BVA(# N’s, that guy)’ BVA(sukunakutomo #-cl-no N, soitu) ‘BVA(at least # or more N, that guy)’ BVA(N-cm sukunakutomo #-cl, soitu) (the same as (d) except that the X is in the so-called “floating numeral” form)

Of these ten combinations, five combinations seem to give us c-command detection with BVA(X, Y), based on DR(X, beta) and Coref(alpha, Y), without reference to sub-preliminary experiments; they are all the options in (71) except for (71a), (i.e., (71b, c, d, e)), and (70a). The remaining five combinations (i.e., (70b, c, d, e) and (71a)) result in one or two speakers whose reported judgments seemingly constitute disconfirmation of the prediction in (1). If we consider results of the relevant sub-preliminary experiments, however, the speakers whose reported judgments seemed to disconfirm the prediction in (1) are no longer in the ABC intersection, while there is still at least one speaker in the ABC intersection whose reported judgments give us c-command detection with BVA(X, Y). Reference to the sub-preliminary experiments also reduces the number of the speakers in ABC in most of (71b, c, d, e)), and (70a) as well. Different sets of sub-preliminary experiments are used for each of the ten combinations of X and Y of BVA(X, Y), using the same choice of X and the same choice of Y (and the a-NP counterpart of the Y), for the reasons discussed in Section 2.4. If we used a “wrong” set of “sub-preliminary experiments”, e.g., if we

272 

 Hajime Hoji

used sub-preliminary experiments with soitu and aitu, when checking the result of an experiment with BVA(X, soko), the efficacy of these tests is reduced. Recall that five of the ten combinations of X and Y of BVA(X, Y) had at least one red in ABC. Specifically, there were two for (70b), two for (70c), one for (70d), two for (70e) and one for (71a). As noted, reference to the “right” sub-preliminary experiments reduced these numbers to zero. Reference to the “wrong” sub-preliminary experiments does reduce the numbers to zero in four of the five combinations, but for (70e), one of the two reds remains in ABC. It is significant that reference to the “wrong” sub-preliminary experiments has resulted in one red remaining in ABC in (70e), and hence would have demonstrated the failure of the entailment in (1). This in turn points to the importance of the use of the “right” sub-preliminary experiments to identify qualified speakers. The results of the experiments mentioned above are provided in Appendix III.

3.1.5 Checking the Contrapositive Consider (61) again, which is the result of the experiments testing (38) (without reference to sub-preliminary experiments). (38) Correlational/conditional prediction about c-command-detection with BVA (subete-no N, soko), based on DR(subete-no N, 3-nin-no seizika) and Coref (aitu, soko): Provided that DR([+cc]S, subete-no N, 3-nin-no seizika):yes ∧ Coref([+cc]S, aitu, soko):yes ∧ BVA([+cc]S, subete-no N, soko):yes; DR([-cc]S, subete-no N, 3-nin-no seizika):no ∧ Coref([-cc]S, aitu, soko):no → BVA([-cc]S, subete-no N, soko):no The contrapositive of (38) is (72). (72) The contrapositive of (38): Provided that DR([+cc]S, subete-no N, 3-nin-no seizika):yes ∧ Coref([+cc]S, aitu, soko):yes ∧ BVA([+cc]S, subete-no N, soko):yes; BVA([-cc]S, subete-no N, soko):yes → DR([-cc]S, subete-no N, 3-nin-no seizika):yes ∨ Coref([-cc]S, aitu, soko):yes

(72) states that if a speaker belongs to A, and if the speaker is a red ✶, that speaker should be outside B and/or outside C, i.e., outside the ABC intersection. Since (72) is the contrapositive of (38), as long as (38) holds of every relevant speaker, (72) should also hold of every relevant speaker. As such, (72) holding of every rel-

Replication: Predicted Correlations of Judgments in Japanese 

 273

evant speaker in (61a) is not surprising, given that we have already seen that (38) does. The significance of considering the contrapositive is that it (can) allow us to check our prediction beyond those speakers in the ABC intersection, more precisely, those speakers who belong in A, AB or AC in the terms of the partitioning in (53), repeated here as (73).55 (73)

A

AC

AB ABC B

BC

C

The position of each red star (i.e., the speaker represented with one) would tell us the source of the speaker’s BVA([+cc]S, X, Y):yes – whether it is due to NFS effects on DR, Coref or both. If we focus on the speakers who belong in the top circle, a red star in AC would indicate that the BVA([+cc]S, X, Y):yes is due to NFS effects on DR, one in AB would indicate that it is due to NFS effects on Coref, and one in A both.

55 Speakers who are represented by a blue “x” are those who reported BVA([+cc]S, subete-no N, soitu):no and BVA([-cc]S, subete-no N, soitu):no (neither an instance of potential c-command detection nor potential disconfirmation of the prediction in (38)); see (59) and (60). Since the speakers in the top circle have necessarily reported BVA([+cc]S, subete-no N, soitu):yes (see the last conjunct of the “Provided that . . .” clause in (38)), all the blue x’s are necessarily outside the top circle, as in (61a). This is what makes the checking of the contrapositive of (38) straightforward. Readers may want to note that the inclusion of the reference to MR([+cc]S, X, Y):yes, i.e., DR([+cc]S, X, beta):yes, Coref([+cc]S, alpha, Y):yes, and BVA([+cc]S, X, Y):yes in A is a relatively new innovation, and the graphs presented in Chapter 7 follow a slightly older convention.

274 

 Hajime Hoji

3.2 Checking the Absence of Other Types of Entailment 3.2.1 Introduction Consider again the correlational/conditional prediction in (1). (1) Correlational/conditional prediction about c-command-detection with BVA(X, Y): Provided that DR([+cc]S, X, beta):yes ∧ Coref([+cc]S, alpha, Y):yes ∧ BVA([+cc] S, X, Y):yes; DR([-cc]S, X, beta):no ∧ Coref([-cc]S, alpha, Y):no → BVA([-cc]S, X, Y):no

As discussed in Chapter 5, there has not been a single case so far in my selfexperiment, based on various combinations of X, Y, [-cc]S, [+cc]S, etc., that would constitute empirical disconfirmation of this prediction, i.e., a demonstration of the failure of the antecedent to entail the consequent. Of course, it can always be the case that the prediction is false but we have not yet encountered a piece of data that would contradict it; disconfirmation of universal predictions is always existential, meaning that we never can show that no such a counter example exists. If we want to refute a given prediction, we therefore need to search until we find an existential demonstration of the failure of the relevant entailment to hold (for the sake of brevity, we may refer to this as demonstration of the “failure” of a given entailment). We may thus say that in my self-experiment, I have failed to find any data which demonstrates the failure of the entailment in (1). The same also holds true in non-self-experiments like K17s, which involve a greater number of speakers but significantly fewer choices of options than in my self-experiment. As just noted, there is no guarantee that the failure of the entailment will never be existentially demonstrated in the future. It is, however, noteworthy that we have not found a single combination of various options in my self-experiment that would constitute a demonstration of the failure of the predicted entailment in (1) to hold or a single “qualified” speaker in K17s whose reported judgments would existentially demonstrate the failure of the predicted entailment in (1). One might wonder if this success is simply due to a tight correlation between DR and BVA and/or one between Coref and BVA. This might suggest that the correlations indicated in (74) and (75) might hold independently, rendering (1) merely a consequence of those correlations.56

56 Without “BVA([+cc]S, X, Y):yes” in the “Provided that . . .” clause, (74) would only state that  (with DR(X, beta)) entails BVA([-cc]S, X, Y):no, not c-command detection with BVA(X, Y). Likewise,

Replication: Predicted Correlations of Judgments in Japanese 

 275

(74) Provided that DR([+cc]S, X, beta):yes ∧ BVA([+cc]S, X, Y):yes; DR([-cc]S, X, beta):no → BVA([-cc]S, X, Y):no (75) Provided that Coref([+cc]S, alpha, Y):yes ∧ BVA([+cc]S, X, Y):yes; Coref([-cc]S, alpha, Y):no → BVA([-cc]S, X, Y):no In Chapter 5: Section 5, I noted that my own judgments constituted a demonstration of the failure of the predicted entailment to hold not only in (74) and (75) but also in (76) and (77). (76)

Provided that BVA([+cc]S, X, Y):yes ∧ DR([+cc]S, X, beta):yes; BVA([-cc]S, X, Y):no → DR([-cc]S, X, beta):no

(77)

Provided that BVA([+cc]S, X, Y):yes ∧ Coref([+cc]S, alpha, Y):yes; BVA([-cc]S, X, Y):no → Coref([-cc]S, alpha, Y):no

Roughly speaking, (1) claims that a combination of c-command pattern with DR (X, beta) and c-command pattern with Coref(alpha, Y) entails c-command detection with BVA(X, Y). As noted, we have not found a single case that would demonstrate the failure of the entailment. On the other hand, the entailments indicated in (74)–(77) are, again roughly speaking, (74), that c-command pattern with DR(X, beta) entails c-command detection with BVA(X, Y), (75), that c-command pattern with Coref(alpha, Y) entails c-command detection with BVA(X, Y), (76), that c-command pattern with BVA(X, Y) entails c-command detection with DR (X, beta), and (77), that c-command pattern with BVA(X, Y) entails c-command detection with Coref(alpha, Y). As noted, the failure of the entailments indicated in (74)–(77) has been existentially demonstrated in my self-experiment, in sharp contrast to the failure of the entailment in (1). In this subsection, I will discuss the absence of the entailments in (74)–(77) in relation of K17s, as further illustration of the significance of the correlational/conditional prediction in (1).

without “BVA([+cc]S, X, Y):yes” in the “Provided that . . .” clause, (75) would only state that  (with Coref(alpha, Y)) entails BVA([-cc]S, X, Y):no, not c-command detection with BVA(X, Y). Similar remarks apply to (76) and (77) below.

276 

 Hajime Hoji

3.2.2 The C-command Detection with DR(X, alpha) does not Entail BVA([-cc]S, X, Y):no The claim in (1) is that a combination of (78) and (78) entails (78), making reference to a portion of Figure 1 given in Section 2.1. (78)

1 Provided that

2

DR(

S, X, beta):yes ˄ Coref(

S, alpha, Y):yes ˄

DR(

S, X, beta):no

S, alpha, Y):no

˄ Coref(



3 5

BVA(

S, X, Y):yes

4

BVA(

S, X, Y):no

;

What is crucially assumed is that if a particular choice of X does not give rise to NFS effects with DR(X, beta), as indicated by (78), it also does not give rise to NFS effects with BVA(X, Y), and if a particular choice of Y does not give rise to NFS effects with Coref(alpha, Y), as indicated by (78), it also does not give rise to NFS effects with BVA(X, Y). This assumption alone accounts for the absence of the entailment in (74), as demonstrated by my judgments (see Chapter 5: Section 5), as indicated in (79). (79)

My judgments (at all three Stages) that demonstrated the failure of the entailment in (74): a. DR([+cc]S, ano kaisya-igai, 3-tu-no robotto):yes b. DR([-cc]S, ano kaisya-igai, 3-tu-no robotto):no c. BVA([+cc]S, ano kaisya-igai, soko):yes d. BVA([-cc]S, ano kaisya-igai, soko):yes

The claim made by (74) is that the entailment indicated therein holds with every choice of X, Y, and beta, with every speaker, and at any point in time. That is to say, the failure of the entailment in (74) would be existentially demonstrated by a set of judgments as in (80) by any speaker at any point in time, with any choice of X, Y, beta. (80) a. b. c. d.

DR([+cc]S, X, beta):yes; DR([-cc]S, X, beta):no BVA([-cc]S, X, Y):yes BVA([-cc]S, X, Y):yes

Replication: Predicted Correlations of Judgments in Japanese 

 277

(79) is an instance of (80); it thereby constitutes an existential demonstration of the failure of the entailment in (74) to hold, with myself being the speaker in question. I also obtain a similar demonstration with soitu ‘that guy’ in place of soko ‘it’ in (79), with an appropriate change of kaisya ‘company’ to gisi ‘engineer’, for example. While K17s includes soko ‘it’ and soitu ‘that guy’ for Y, it does not include NP-igai ‘ones other than NP’ as a choice of X. When checking the result of K17s, we therefore need to turn to what may be the best choice of X for obtaining DR([-cc] S, X, beta):no, along with DR([+cc]S, X, beta):yes (i.e.,  in Figure 1) for some participants in K17s so as to be able to check if there is a speaker in K17s whose judgments existentially demonstrate the failure of the entailment in (74). Let us use subete-no N ‘every N’ as X because it is used as X of BVA/DR(X, Y/beta) in the preceding illustration. We can thus check if any speaker in K17s reported judgments as in (81) or (82). (81) a. b. c. d.

DR([+cc]S, subete-no N, #-cl-no N):yes DR([-cc]S, subete-no N, #-cl-no N):no BVA([+cc]S, subete-no N, soko):yes BVA([-cc]S, subete-no N, soko):yes

(82) a. b. c. d.

DR([+cc]S, subete-no N, #-cl-no N):yes DR([-cc]S, subete-no N, #-cl-no N):no BVA([+cc]S, subete-no N, soitu):yes BVA([-cc]S, subete-no N, soitu):yes

If we find such a speaker, we have a demonstration of the failure of the entailment in (74) to hold, based on someone other than myself. We first check whether there are speakers in K17s whose reported judgments are as in (81a, b, c), we then check and see if any of those speakers reported BVA([-cc]S, subete-no N, soitu):yes. In the terms of the preceding discussion, we first check if there is any speaker in the middle intersection in (83).

278 

 Hajime Hoji

(83) BVA([+cc]S, subete-no N, soitu):yes

DR([+cc]S, subete-no N, #-cl-no N):yes

DR([-cc]S, subete-no N, #-cl-no N):no

We then check if there is any speaker in the middle intersection who reported BVA([-cc]S, subete-no N, soitu):yes. As noted, the existence of such a speaker would demonstrate the failure of the entailment in (74) to hold. As indicated in (84), there are indeed five such speakers. (84) a.

Replication: Predicted Correlations of Judgments in Japanese 

b. Results

Green

Blue

Red

 279

Total

ABC

11

0

5

16

AB

5

0

19

24

AC

9

0

8

17

BC

0

33

2

35

A

2

0

3

5

B

0

13

0

13

C

0

73

3

76

None

0

6

1

7

Total

27

125

41

193

(85) provides the information about what is included in A, B, and C, and reviews what is meant by “red ✶”, “green circle”, and “blue x”.57 (85) A: BVA([+cc]S, subete-no N, soko):yes B: DR([+cc]S, subete-no N, 3-nin-no seizika):yes C: DR([-cc]S, subete-no N, 3-nin-no seizika):no Red “✶”: BVA([-cc]S, subete-no N, soko):yes Green circle: BVA([+cc]S, subete-no N, soko):yes and BVA([-cc]S, subete-no N, soko):no Blue “x”: neither Red “✶” nor Green circle. If we add (86) to A in (85), however, we obtain the result in (87). (86) a. b. c.

Passing the aitu vs. soitu test Passing the singular-denoting test with aitu Passing the singular-denoting test with soitu

57 By the nature of the prediction being considered, A, B, and C in (85) are distinct from A, B, and C in the partitioning in (73). A in (85), for example, includes ABC, AB, AC, and A in the terms of the partitioning in (73). The table in (87b) is in line with the portioning in (73). Similar remarks apply to graphs such as (90) and (93).

280  (87)

 Hajime Hoji

a.

b. Results

Green

Blue

Red

Total

ABC

4

0

0

4

AB

2

0

2

4

AC

1

0

2

3

BC

7

33

7

47

A

0

0

0

0

B

3

13

17

33

C

7

71

9

87

None

2

6

4

12

Total

26

123

41

190

The K17s result with BVA(subete-no N, soitu) is thus compatible with the entailment in (74). As noted, the results of my own self-experiments, discussed in Chapter 5 (cf. Stages 1 and 2), indicate that NFS2 effects with BVA(X, Y) can arise more easily/readily with Y=soko than with Y=soitu. We have in fact known, based

Replication: Predicted Correlations of Judgments in Japanese 

 281

on our less systematic studies over the past 25 years, that soko induces NFS effects more easily/readily than soitu.58 We thus expect that we might find a case that demonstrates the failure of this type of entailment in K17s if we turn to BVA(X, soko). That is exactly what we find. Focusing on BVA(X, soko), we now check (88), instead of its “soitu version” in (89), both of which are instances of (74). (88) Provided that DR([+cc]S, subete-no N, beta):yes ∧ BVA([+cc]S, subete-no N, soko):yes; DR([-cc]S, subete-no N, beta):no → BVA([-cc]S, subete-no N, soko):no (89) Provided that DR([+cc]S, subete-no N, beta):yes ∧ BVA([+cc]S, subete-no N, soitu):yes; DR([-cc]S, subete-no N, beta):no → BVA([-cc]S, subete-no N, soitu):no Without reference to “sub-preliminary experiments”, we obtain the result given in (90).

58 This is in line with the “shift” of my judgments from Stage 1 (where soko induce NFS2 effects but soitu does not) to Stage 2 (where both soko and soitu induce NFS2 effects).

282 

 Hajime Hoji

(90) a.

b.

Results

Green

Blue

Red

Total

ABC

13

0

8

21

AB

9

0

18

27

AC

13

0

14

27

BC

0

27

3

30

A

2

0

1

3

B

0

7

3

10

C

0

62

3

65

None

0

7

2

9

Total

37

103

52

192

As in (85), (91) provides the relevant information in relation to (90). (91) A: BVA([+cc]S, subete-no N, soko):yes B: DR([+cc]S, subete-no N, 3-nin-no seizika):yes

Replication: Predicted Correlations of Judgments in Japanese 

 283

C: DR([-cc]S, subete-no N, 3-nin-no seizika):no Red “✶”: BVA([-cc]S, subete-no N, soko):yes Green circle: BVA([+cc]S, subete-no N, soko):yes and BVA([-cc]S, subete-no N, soko):no Blue “x”: neither Red “✶” nor green circle. As indicated in (90), 8 out of 21 in ABC are Reds. Even when we add (92) to A in (91), we still have Reds in ABC, as indicated in (93). (92) a. Passing the asoko vs. soko test b. Passing the singular-denoting test with asoko c. Passing the singular-denoting test with soko (93)

a.

284 

 Hajime Hoji

b.

Results

Green

Blue

Red

Total

ABC

1

0

2

3

AB

1

0

1

2

AC

0

0

0

0

BC

12

27

9

48

A

0

0

0

0

B

8

7

20

35

C

13

61

15

89

None

2

7

3

12

Total

37

102

50

189

Two out of the three speakers in ABC are Red ✶s, which demonstrates the failure of the entailment in (88), and hence of the general entailment (74). Crucially, this demonstration is based on judgments by two speakers in K17s, i.e., speakers other than myself. Let us summarize the preceding discussion and its significance, making reference to my self-experiments and K17s. I am a “qualified speaker”, by definition, as to attentiveness, and sono otoko ‘that man”, the choice of Y that gives me c-command detection with BVA(X, Y), is singular-denoting for me based on the split antecedence test. In principle, my own judgments alone should suffice as the demonstration of the failure of the entailment in question to hold. For various reasons, however, there is general skepticism about the use of the researcher’s own introspective judgments as evidence for or against hypotheses in studies related to language.59 Non-self-experiments such as K17s are conducted mainly for the purpose of convincing others about entailment in (1)  and the absence of entailment as (74) and other such cases, not for the purpose of scientific discovery. What is observed above about the demonstration of the failure of entailment in (74) to hold contrasts qualitatively with the failure to obtain an existential demonstration of the failure of the entailment in (1). The disconfirmation of the prediction in (1) would take the form of finding just a single case that demonstrates the failure of the predicted entailment to hold, just as the absence of the entailment in (74) has been demonstrated exis-

59 While the skepticism seems partially justified by the nature of the study, when the object of inquiry is internal to an individual, as in LFS, the skepticism seems to be also due to how introspective judgments (the researcher’s own or other speakers’) have generally been treated in relevant research. The correlational methodology proposed in this volume is meant to provide a way to dispel, or at least counter to, the skepticism of the latter type. See Chapter 9 of this volume for further discussion.

Replication: Predicted Correlations of Judgments in Japanese 

 285

tentially. As discussed, there has been no instance found thus far that would demonstrate the absence of the entailment in (1), based on a number of combinations of X, Y, [-cc]S, [+cc]S, etc., in my self-experiment; likewise, there has not been a single instance in K17s that would demonstrate the failure of the entailment in (1), in sharp contrast to the case of (74). For further illustration, let us compare the two “predictions” in (94) (an instance of (74)) and (63) (an instance of (1)). (94) Entailment that does not hold (an instance of (88), which in turn is an instance of (74)): Provided that DR([+cc]S, subete-no N, 3-nin-no seizika):yes ∧ BVA([+cc]S, subete-no N, soko):yes; DR([-cc]S, subete-no N, 3-nin-no seizika):no → BVA([-cc]S, subete-no N, soko):no (63) Correlational/conditional prediction about c-command-detection with BVA(subete-no N, soko), based on DR(subete-no N, 3-nin-no seizika) and Coref(asoko, soko): Provided that DR([+cc]S, subete-no N, 3-nin-no seizika):yes ∧ Coref([+cc]S, asoko, soko):yes ∧ BVA([+cc]S, subete-no N, soko):yes; DR([-cc]S, subete-no N, 3-nin-no seizika):no ∧ Coref([-cc]S, asoko, soko):no → BVA([-cc]S, subete-no N, soko):no (94) and (63) are both about BVA(subete-no N, soko) and both refer to the same sub-preliminary experiments as indicated in (92), repeated here. (92) a. Passing the asoko vs. soko test b. Passing the singular-denoting test with asoko c. Passing the singular-denoting test with soko When checking (94), we “used” (95), and when checking (63), we “used” (96). (95) “Information” for (94): A circle: a. BVA([+cc]S, subete-no N, soko):yes b. Passing the asoko vs. soko test c. Passing the singular-denoting test with asoko d. Passing the singular-denoting test with soko

286 

 Hajime Hoji

B circle: DR([+cc]S, subete-no N, 3-nin-no seizika):yes C circle: DR([-cc]S, subete-no N, 3-nin-no seizika):no Red star: BVA([-cc]S, subete-no N, soko):yes Green circle: BVA([+cc]S, subete-no N, soko):yes and BVA([-cc]S, subete-no N, soko):no Blue “x”: Neither Red nor Green. (96) Information for (63): A circle: a. DR([+cc]S, subete-no N, 3-nin-no seizika):yes b. Coref([+cc]S, asoko, soko):yes c. BVA([+cc]S, subete-no N, soko):yes d. Passing the asoko vs. soko test e. Passing the singular-denoting test with asoko f. Passing the singular-denoting test with soko B circle: DR([-cc]S, subete-no N, 3-nin-no seizika):no C circle: Coref([-cc]S, asoko, soko):no Red star: BVA([-cc]S, subete-no N, soko):yes Green circle: BVA([+cc]S, subete-no N, soko):yes and BVA([-cc]S, subete-no N, soko):no Blue x: Neither Red nor Green The result of checking (63) is as indicated in (68), with three “sub-preliminary experiments” (the asoko vs. soko test, and the split antecedence test with asoko and the one with soko).60 We see that there are no Reds and one Green in the ABC intersection in (68). The attempt in K17s to demonstrate the failure of the predicted entailment to hold has just failed, thereby providing experimental support for the prediction in (63). Since (63) is an instance of (1), this provides further experimental support for (1). Compare (68) with (93), the result of testing (94), with the “sub-preliminary experiments” in (92). In (93), two out of the three speakers in ABC are Reds, constituting a demonstration of the failure of the predicted entailment in (94) to hold. Since (94) is an instance of (74), the result in (93) demonstrates the failure of the predicted entailment in (74). When we check predictions of the forms in (1) and (74), the failure of the predicted entailment to hold can be demonstrated existentially, by a single (reliable) instance of the failure of the entailment. That is to say, the disconfirmation

60 These “sub-preliminary experiments” are what is listed in (92).

Replication: Predicted Correlations of Judgments in Japanese 

 287

of the prediction can be existentially demonstrated. The validity of the prediction cannot be demonstrated existentially; in fact, its validity cannot be demonstrated because they make a universal claim that whenever the antecedent of the conditional is met, the consequent is satisfied, but we cannot check every possible case where the antecedent of the conditional is met. This in fact is the content of the “fundamental predicted schematic asymmetry” in Hoji 2015, which plays a central role in our conceptual articulation of the nature of rigorous testability in LFS. The failure of the entailment in (74) to hold can be explained by the fact that a certain choice of Y alone can give rise to NFS effects (NFS2 effects in the terms of Chapter 5: Section 8) with BVA(X, Y). (74)

Provided that DR([+cc]S, X, beta):yes ∧ BVA([+cc]S, X, Y):yes; DR([-cc]S, X, beta):no → BVA([-cc]S, X, Y):no

For example, even with the choice of X that seems to most strongly resist NFS effects with DR(X, beta), such as NP-igai ‘ones other than NP’, for me, the use of soko as Y gives rise to BVA([-cc]S, NP-igai, soko):yes, which we attribute to NFS2 effects with BVA(X, Y) due to Y=soko. What choice of Y (easily) gives rise to NFS2 effects with BVA(X, Y) for me depends upon different Stages, as discussed in Chapter 5, and this seems to hold of other speakers as well. Based on my own experiences of obtaining my own judgments (see Chapter 5: Sections 3 and 4) and obtaining judgments from others, on various relevant sentences, I expect that, between the two choices for Y of BVA/ Coref(X/alpha, Y) in K17s (soko and soitu), soko can give rise to NFS2 effects with BVA more easily/readily than soitu. If soko indeed can give rise to NFS2 effects with BVA for a given speaker and if one of the choices of X for DR(X, beta), among those given in (14), never gives rise to NFS effects with DR(X, beta) for the same speaker, we should thus be able to demonstrate the failure of the entailment in (74) based on judgments by that speaker. It in fact turns out, as indicated in (93), that there are two speakers in K17s for whom subete-no N never gives rise to NFS effects of DR(X, beta) but for whom soko can give rise to NFS2 effects with BVA. Their reported judgments on BVA(subete-no N, soko) and DR(subete-no N, beta), as indicated by the two red stars in the center intersection (the intersection of A, B and C circles in (95)) in (93), constitute demonstrations of the failure of the entailment in (74) to hold, like my judgments on BVA(NP-igai, soko) and DR(NP-igai, beta).61

61 Recall that the use of BVA(subete-no N, soko) and DR(subete-no N, beta) does not result in a demonstration of the failure of the entailment in (74) in my self-experiment because the antecedent of (74) does not hold for me.

288 

 Hajime Hoji

3.2.3 Other Cases of the Absence of Entailment In Chapter 5: Section 5, I noted that my own judgments demonstrated the failure of entailments in cases other than the one in (74), including those in (75), (76) and (77), repeated here. (75)

Provided that Coref([+cc]S, alpha, Y):yes ∧ BVA([+cc]S, X, Y):yes; Coref([-cc]S, alpha, Y):no → BVA([-cc]S, X, Y):no

(76)

Provided that BVA([+cc]S, X, Y):yes ∧ DR([+cc]S, X, beta):yes; BVA([-cc]S, X, Y):no → DR([-cc]S, X, beta):no

(77)

Provided that BVA([+cc]S, X, Y):yes ∧ Coref([+cc]S, alpha, Y):yes; BVA([-cc]S, X, Y):no → Coref([-cc]S, alpha, Y):no

Instances of my judgments that constituted demonstrations of the failure of the entailments indicated in (75), (76) and (77), are given in (97), (98), and (99), respectively.62 (97)

My judgments at Stage 2 that demonstrated the failure of the entailment in (75) to hold: a. Coref([+cc]S, ano gisi, sono otoko):yes b. Coref([-cc]S, ano gisi, sono otoko):no c. BVA([+cc]S, subete-no otoko, sono otoko):yes d. BVA([-cc]S, subete-no otoko, sono otoko):yes

(98) My judgments at Stage 2 that demonstrated the failure of the entailment in (76) to hold: a. BVA([+cc]S, subete-no gisi, sono otoko):yes b. BVA([-cc]S, subete-no gisi, sono otoko):no c. DR([+cc]S, subete-no gisi, 3-tu-no robotto):yes d. DR([-cc]S, subete-no gisi, 3-tu-no robotto):yes (99) My judgments at Stage 2 that demonstrated the failure of the entailment in (77) to hold: a. BVA([+cc]S, subete-no gisi, sono otoko):yes

62 What is meant by [+cc]S here is [+cc, -loc]S, but the presentation is simplified.

Replication: Predicted Correlations of Judgments in Japanese 

 289

b. BVA([-cc]S, subete-no gisi, sono otoko):no c. Coref([+cc]S, ano otoko, sono otoko):yes d. Coref([-cc]S, ano otoko, sono otoko):yes The account of the judgments that constituted a demonstration of the failure of the entailment in (74), my own or others’, can be straightforwardly presented without needing to refer to any particularly “new” concepts, as discussed in the preceding sub-section. On the other hand, an account of the judgments that constituted a demonstration of the failure of the entailment in (75), (76) and (77), as given in (97), (98), and (99), respectively, makes it necessary to address certain issues that have not been addressed. Although we do not have enough space to address these issues in (full) detail here, I will discuss what seem to me to be the most relevant issues, which I hope will serve as a basis for further investigation into the relevant issues. Let us first consider (75). What would demonstrate the failure of the entailment in (75) is a set of judgments by a (reliable) speaker as indicated in (100), which my (97) above is an instance of. (100)

What would constitute demonstrate the failure of the entailment in (75): a. Coref([+cc]S, alpha, Y):yes b. Coref([-cc]S, alpha, Y):no c. BVA([+cc]S, X, Y):yes d. BVA([-cc]S, X, Y):yes

It is stated in the earlier discussion of (1) that, if a given choice of Y does not give rise to NFS effects with Coref(alpha, Y), the same choice of Y does not give rise to NFS effects with BVA(X, Y), either. A combination of (100b) and (100d), which a combination of (97b) and (97d) is an instance of, however, indicates that something additional has to be said. Recall the judgments in (101), which I obtain at Stage 2, as discussed in Chapter 5: Section 4. (101) a. BVA([+cc]S, ano gisi-igai, sono otoko):yes b. BVA([-cc]S, ano gisi-igai, sono otoko):no (101b) indicates that sono otoko cannot give rise to NFS2 effects with BVA for me at that stage, which is the most typical “stage” for me. Given this, my (97d) BVA([-cc]S, subete-no otoko, sono otoko):yes (at that stage) must be due to something other than NFS2 effects. As it turns out, BVA(X, Y) and Coref(alpha, Y) can arise not only based on NFS2 effects, which seem to be solely due to the choice of Y, but also

290 

 Hajime Hoji

based on another type of NFS effects that seem to be facilitated by the N head of X/alpha and the N head of Y being “identical”.63 Although we only have a limited understanding of this type of NFS, in fact NFS in general, the effects of this type of NFS seem to become unavailable when the sentence in question is embedded in a “non-categorical context” as in the case of NFS1 effects with DR; see Chapter 5: Section 9. For the ease of exposition, let us refer to this type of NFS effects with BVA/Coref as “NFS1 effects with BVA/Coref” because of their sensitivity to the “categorical vs. non-categorical context” distinction, as in the case of NFS1 effects with DR.64 Under the reasonable assumption that NP-igai does not have a head N (as it in fact does not overtly), the “N-head identity” “requirement”65 is not satisfied in (97b) while it can be in (97d) BVA([-cc]S, subete-no otoko, sono otoko):yes, accounting for the different judgments of mine in (97b) and (97d) (at Stage 2). My judgments in (98) and (99) can be accounted for in a similar way. In the case of (98), (98b) BVA([-cc]S, subete-no gisi, sono otoko):no obtains, in contrast to (97d) BVA([-cc]S, subete-no otoko, sono otoko):yes, because the N-head identity “requirement” cannot be satisfied in the former but it can in the latter, giving us the c-command pattern with BVA(subete-no gisi, sono otoko). When I get those judgments, I still get (98d) DR([-cc]S, subete-no gisi, 3-tu-no robotto):yes (along with (98c) DR([+cc]S, subete-no gisi, 3-tu-no robotto):yes. In the case of (99), we have the same “situation” with BVA(subete-no gisi, sono otoko), giving us a c-command pattern, but I get (99d) Coref([-cc]S, ano otoko, sono otoko):yes, which can satisfy the N-head identity requirement for the NFS in question with Coref(alpha, Y).66 Turning to the non-self-experiments under discussion, K17s does not include NP-igai ‘ones other than NP’, which is the best choice of X for c-command detection for me, and all the choices of X in K17s have an N head, which could (potentially) lead to NFS1 effects due to “N-head identity”. K17s also does not have sono otoko

63 The reason why “identical” is in quotation marks is because what is required does not seem to be the strict morphological identity. For example, BVA(subete-no hihonzin gisi, sono gisi) ‘BVA(every Japanese engineer, that engineer)’ can arise in [-cc]S, despite the non-identity between the N head of X, i.e., the compound noun, nohonzin gisi ‘Japanese engineer’, and the N head Y, gisi ‘engineer’. What the exact nature of this “requirement” might be and how it might even be related to “Condition D’” (Condition Dprime ) in Hoji et al. 2003 are not well understood at this point. 64 The fact that NFS1 requires the “categorical context” suggests that subete-no N does not by itself give rise to NFS1 effects. 65 In the sense addressed above. 66 The “N-head identity” considerations above suggest that, just as the choice of X affects the availability of NFS1-BVA(X, sono otoko), the choice of alpha also affects the availability of NFS1-Coref(alpha, sono otoko). The possibility of NFS1-BVA/Coref has been suppressed in the earlier discussion in Chapter 4 and much of Chapter 5.

Replication: Predicted Correlations of Judgments in Japanese 

 291

‘that man’ as Y of BVA/Coref(X/alpha, Y), having just two choices of Y for BVA/ Coref(X/alpha, Y), soko ‘it’ and soitu ‘that guy’, both of which lead to NFS2 effects (relatively) easily (for me). Given the choices of X and Y in Kyudai17s, it is therefore not possible to directly replicate in K17s my judgments that demonstrated the absence of the entailment in (75), for example. We can, however, check if the pattern of my judgments that demonstrated the failure of the entailment in (75) gets replicated in K17s, with different choices of X and Y. (75)

Provided that Coref([+cc]S, alpha, Y):yes ∧ BVA([+cc]S, X, Y):yes; Coref([-cc]S, alpha, Y):no → BVA([-cc]S, X, Y):no

What would demonstrate the failure of the entailment in (75) to hold is as given in (102), which (97), repeated below, is an instance of. (102) What would demonstrate the failure of the entailment in (75): a. Coref([+cc]S, alpha, Y):yes b. Coref([-cc]S, alpha, Y):no c. BVA([+cc]S, X, Y):yes d. BVA([-cc]S, X, Y):yes (97) My judgments at Stage 2 that demonstrated the failure of the entailment in (75): a. Coref([+cc]S, ano gisi, sono otoko):yes b. Coref([-cc]S, ano gisi, sono otoko):no c. BVA([+cc]S, subete-no otoko, sono otoko):yes d. BVA([-cc]S, subete-no otoko, sono otoko):yes By using subete-no N ‘every N’ as X of BVA(X, Y), as in the preceding discussion of K17s, and asoko ‘it’ and aitu ‘that guy’ as alpha of Coref(alpha, Y), we can check to see if any of the participants in K17s who have passed the sub-preliminary experiments mentioned in (92) or (86), reported judgments as in (103) or (104), respectively. (92) and (86) are repeated below. (103) A specific instance of (102): a. Coref([+cc]S, asoko, soko):yes b. Coref([-cc]S, asoko, soko):no c. BVA([+cc]S, subete-no tihoozititai, soko):yes ‘BVA(every local government, it)’ d. BVA([-cc]S, subete-no tihoozititai, soko):yes

292  (104)

(92)

 Hajime Hoji

Another specific instance of (102): a. Coref([+cc]S, aitu, soitu):yes b. Coref([-cc]S, aitu, soitu):no c. BVA([+cc]S, subete-no zyooin giin, soitu):yes ‘BVA(every senator, that guy)’ d. BVA([-cc]S, subete-no zyooin giin, soitu):yes a. Passing the asoko vs. soko test b. Passing the singular-denoting test with asoko c. Passing the singular-denoting test with soko

(86) a. Passing the aitu vs. soitu test b. Passing the singular-denoting test with aitu c. Passing the singular-denoting test with soitu If such a pattern of judgments obtains from any of the relevant participants, that demonstrates the failure of the entailment in (75) to hold, based on someone other than myself. In order for the judgments in (104b) to arise, soitu must not give rise to NFS2 effects with Coref(alpha, soitu). (104d) would therefore have to be due to NFS1 effects. It is, however, highly unlikely that the N head of subete-no zyooin giin ‘every senator’, i.e., zyooin giin ‘senator’, and the bound morpheme -itu in soitu can satisfy the N-head identity requirement for NFS1 for BVA. Similar considerations apply to how the judgments in (103) could possibly arise. Our expectation is therefore that it is highly unlikely that we find a demonstration in K17s of the failure of the entailment in (75) to hold, based on (103) or (104).67 Before turning to the relevant results of K17s, let us consider (77), repeated here. (77)

Provided that BVA([+cc]S, X, Y):yes ∧ Coref([+cc]S, alpha, Y):yes; BVA([-cc]S, X, Y):no → Coref([-cc]S, alpha, Y):no

What would demonstrate the failure of the entailment in (77) to hold is a set of judgments by a (reliable) speaker as indicated in (105), which (99), repeated below, is an instance of.

67 Since we are dealing with NFS effects, about which we only have a rather limited understanding, this is just an expectation not a full prediction.

Replication: Predicted Correlations of Judgments in Japanese 

 293

(105) What would demonstrate the failure of the entailment in (77): a. BVA([+cc]S, X, Y):yes b. BVA([-cc]S, X, Y):no c. Coref([+cc]S, alpha, Y):yes d. Coref([-cc]S, alpha, Y):yes (99) My judgments at Stage 2 that demonstrated the failure of the entailment in (77): a. BVA([+cc]S, subete-no gisi, sono otoko):yes ‘BVA(every engineer, that guy)’ b. BVA([-cc]S, subete-no gisi, sono otoko):no c. Coref([+cc]S, ano otoko, sono otoko):yes d. Coref([-cc]S, ano otoko, sono otoko):yes Applying basically the same considerations as above, (99) must be due to: (i) sono otoko not giving rise to NFS2 effects of BVA, (ii) BVA(subete-no gisi, sono otoko) not giving rise to NFS1 effects with BVA, and (iii) Coref(ano otoko, sono otoko) satisfying the N-head identity requirement, giving rise to NFS1 effects with Coref. Now turning to K17s, consider two specific instances of (105), as given in (106) and (107). (106) A specific instance of (105): a. BVA([+cc]S, subete-no tihoozititai, soko):yes ‘BVA(every local government, it)’ b. BVA([-cc]S, subete-no tihoozititai, soko):no c. Coref([+cc]S, asoko, soko):yes d. Coref([-cc]S, asoko, soko):yes (107) Another specific instance of (105): a. BVA([+cc]S, subete-no zyooin giin, soitu):yes ‘BVA(every senator, that guy)’ b. BVA([-cc]S, subete-no zyooin giin, soitu):no c. Coref([+cc]S, aitu, soitu):yes d. Coref([-cc]S, aitu, soitu):yes As noted, in order for (106b)/(107b) (along with (106a)/(107a)) to arise, soko/soitu should not give rise to NFS2 effects with BVA(X, soko/soitu), hence also not giving rise to them with Coref(alpha, soko/soitu). (106d)/(107d) therefore would have to be due to NFS1 effects with Coref. In order for that to happen, the use of the identi-

294 

 Hajime Hoji

cal bound morpheme -ko in a-soko68 and so-ko in the case of (106) and -itu in a-itu and so-itu in the case of (107) must satisfy the N-head identity requirement. Since our understanding of NFS1 effects with BVA/Coref is rather limited, we do not know if such a situation obtains consistently for some speakers.69 If we compare the (remote) possibility of NFS1 effects arising in (107d) (with Coref(aitu, soitu) ‘Coref(that guy, that guy))’, for example, and (106d) (with Coref(asoko, soko) ‘Coref(that place, it)’), however, the chances might be better with the former than with the latter, given that -itu seems to have more “semantic content” than -ko, including its clearly derogatory connotation, which -ko does not have. Let us now turn to (76). (76) Provided that BVA([+cc]S, X, Y):yes ∧ DR([+cc]S, X, beta):yes; BVA([-cc]S, X, Y):no → DR([-cc]S, X, beta):no A set of judgments by a (reliable) speaker that would demonstrate the failure of the entailment in (77) to hold is as indicated in (108), which (98) is an instance of. (108) What would demonstrate the failure of the entailment in (76): a. BVA([+cc]S, X, Y):yes b. BVA([-cc]S, X, Y):no c. DR([+cc]S, X, beta):yes d. DR([-cc]S, X, beta):yes (98)

My judgments at Stage 2 that demonstrated the failure of the entailment in (76): a. BVA([+cc]S, subete-no gisi, sono otoko):yes ‘BVA(every engineer, that may)’ b. BVA([-cc]S, subete-no gisi, sono otoko):no c. DR([+cc]S, subete-no gisi, 3-tu-no robotto):yes ‘DR(every engineer, three robots)’ d. DR([-cc]S, subete-no gisi, 3-tu-no robotto):yes

68 As noted in Hoji 2003: note 10, so in a-soko comes from si in a-siko that appeared in the Tyuuko period (A.D. 794–1192), and is unrelated to the demonstrative prefix so-, according to Satoshi Kinsui (p.c., August, 1997). This seems to be shared knowledge among those who work on the history of the Japanese language in light of entries for the relevant items in a dictionary such as日本国語大辞典 (Shogakukan Unabridged Dictionary of the Japanese Language) (Satoshi Kinsui, p.c., March 2022). 69 I cannot use my own judgments in the relevant deliberation except for Stage 1, which generally lasts only for a rather short period of time after I have not judged relevant sentences for some time. At Stages 2 and 3, soitu can give rise to NFS2 effects with BVA/Coref.

Replication: Predicted Correlations of Judgments in Japanese 

 295

The preceding considerations suggest that (98b) is due to (i) sono otoko not giving rise to NFS2 effects with BVA and (ii) subete-no gisi ‘every engineer’ and sono otoko ‘that man’ not satisfying the N-head identity requirement for NFS1 effects with BVA, and that (98d) is due to subete-no gisi ‘every engineer’ giving rise to NFS1 effects with DR.70 Turning to K17s, to see the possible demonstration of the failure of the entailment in (76), we can, for example, check to see if there is any speaker who reported judgments as in (109). (109) A specific instance of (76): a. BVA([+cc]S, subete-no zyooin giin, soitu):yes ‘BVA(every senator, that guy)’ b. BVA([-cc]S, subete-no zyooin giin, soitu):no c. DR([+cc]S, subete-no tihoozititai, 3-nin-no seizika):yes ‘DR(every local government, three politicians)’ d. DR([-cc]S, subete-no tihoozititai, 3-nin-no seizika):yes In line with the suggested account of (98b) above, we can entertain the possibility that (109b) could be due to (i) soitu being disallowed as Y of NFS2-BVA(X, Y), as in the case of my Stage 1, and unlike my Stages 2 and 3,71 and (ii) BVA(subete-no zyooin giin, soitu) not satisfying the N-head identity requirement for NFS1 effects with BVA. In light of the considerations given above, it is not difficult to imagine that there is a speaker for whom such holds. (109d) must be due to subete-no tihoozititai ‘every local government’, and more generally subete-no N ‘every N’, giving rise to NFS1 effects with DR, as in the case of myself and many other speakers in fact.72 The considerations given above thus suggest that, among (75), (76), and (77), a demonstration of the failure of entailment is more likely to be “found” for (76) than for (75) and (77), and it is least likely to be “found” for (75), in K17s. When we check K17s, that is exactly what we find. In the case of (76), with BVA(subete-no N, soitu) ‘BVA(every N, that guy)’, there are six participants in the middle ABC intersection,

70 (98c) does not seem to require anything “special”, as explicitly pointed out in Hayashishita 2004, 2013, among other places. For BVA([+cc]S, X, sono otoko):yes to arise, together with BVA([-cc] S, X, sono otoko):no, sono otoko must satisfy the (“smallness”) requirement for Y of FD-BVA(X, Y), as discussed in Chapter 5: Section 8. That issue has been suppressed in the preceding discussion. 71 At Stage 2, I obtain the pattern of judgments as in (109a, b), with sono otoko, but not with soitu. 72 Remarks similar to what is noted in footnote 70 apply here. (109c) does not seem to require anything “special”, and for BVA([+cc]S, X, soitu):yes to arise, together with BVA([-cc]S, X, soitu):no, soitu must satisfy the (“smallness”) requirement for Y of FD-BVA(X, Y).

296 

 Hajime Hoji

with the sub-preliminary experiments in (86), and two of them are red ✶s, demonstrating the failure of entailment in (76).73 In the case of (75), with Y=soitu (see (104)), there are four speakers in the ABC intersection, and all of them are green circles; with Y=soko (see (103)), we have two speakers in the ABC intersection, and both of them are green circles. In the case of (77), with Y=soko (see (106)), we have two speakers in the ABC intersection, and both of them are green circles, but with Y=soitu (see (107), six speakers are in the ABC intersection and two of them are red ✶. As noted, my own judgments demonstrated the absence of the entailments in (74), (75), (76) and (77). For (76) and (77), we did find speakers in K17s whose judgments existentially demonstrated the failure of the entailments therein, as in the case of (74) discussed earlier.74 In the case of (75), however, we did not find any speaker whose reported judgments demonstrate the failure of the entailment therein; that was not unexpected given the particular design of K17s, as discussed above.75

3.2.4 Summary The discussion in this subsection has suggested that, contrary to what was assumed as a simplification in Chapter 4 (and elsewhere), the choice of alpha 73 Reference to another sub-preliminary experiment (in addition to the sub-preliminary experiments in (92)) that checks whether a speaker clearly rejects the use of the Japanese analogue of ‘those three politicians” as beta of DR(X, beta) (see (29) and (30) and discussion thereabout) does not change the result at all, suggesting that the relevant judgments are already relatively robust. 74 As noted earlier in relation to (74), the most crucial in the demonstration of the failure of the entailment in (74) is that Y of BVA(X, Y) is one that can give rise to NFS2 effects. We expected there to be speakers in K17s for whom soko can give rise to NFS2 effects, and this expectation was met when we checked the result of K17s. 75 The demonstration of the failure of the entailment in (74) to hold is due to the combination of the impossibility of NFS1-DR(X, beta) (leading to (80b)) and the possibility of NFS2-BVA(X, Y) (leading to (80d)). The possibility of NFS2-BVA/Coref(X/alpha, Y) is sensitive only to the choice of Y) while that of NFS1-DR(X, beta) is sensitive only to the choice of X. This is in line with the discussion in Chapters 4 and 5, where the correlational prediction in (1) is presented under the assumptions that the possibility of NFS2-Coref(alpha, Y) is solely due to the choice of Y and that of NFS1-DR(X, beta) is solely due to the choice of X. The relevant judgments of mine are clear, and so are the relevant judgments by others including those in K17s, it seems. The demonstration of the failure of the entailment in (75), (76), and (77) to hold, on the other hand, requires reference to the new concept of “N-head identity”, not mentioned in (1); the demonstration in question is crucially based on the possibility and the impossibility of the N-head-identity-sensitive BVA/ Coref(X/alpha, Y). Just as the nature of “N-head identity” in question is not entirely clear, so the relevant judgments on its effects are not as clear as those on NFS2 effects. The fact that the K17s results are consistent with the account suggested here of the failure of the entailment in (74), (75), (76), and (77) to hold should be understood accordingly.

Replication: Predicted Correlations of Judgments in Japanese 

 297

of Coref(alpha, Y) is not really irrelevant for the availability of Coref(alpha, Y). It seems that one type of NFS-Coref can arise when the head N of alpha and the head N of Y are “identical”. The head N of X and the head N of Y being “identical” also seems to give rise to NFS effects with BVA(X, Y). A more accurate formulation of (1) should then be something like (110), where “and the head N of alpha, that of X, and that of Y are not “identical” is added to the “Provided that . . .” clause.76 (110) Correlational/conditional prediction about c-command-detection with BVA(X, Y), based on DR(X, beta) and Coref(alpha, Y): Provided that DR([+cc]S, X, beta):yes ∧ Coref([+cc]S, alpha, Y):yes ∧ BVA([+cc]S, X, Y):yes, and the head N of alpha, that of X, and that of Y are not “identical”; DR([-cc]S, X, beta):no ∧ Coref([-cc]S, alpha, Y):no → BVA([-cc]S, X, Y):no Reference to different types of NFS effects allowed us to consider what seems to have given rise to the judgments of mine that demonstrated the failure of the entailments in (74), (75), (76) and (77) to hold. The considerations in fact led to different expectations as to how “easily” we might find speakers whose judgments (would) demonstrate the absence of entailment in (74), (75), (76) and (77). Because of our limited understanding of different NFS effects, we only had expectations, not predictions. I find it rather remarkable, however, that the expectations are indeed “borne out”. Of course, we must attain a better understanding about (different) NFS effects, both empirically and theoretically, but what we have seen in this subsection gives us the sense that the usefulness of the basic scientific method is not limited to investigation of core properties of the Computational System (CS) of the language faculty, but rather, it is applicable even when we turn our focus on properties outside the CS.

4 Concluding Remarks An experiment in LFS is about an individual, whether it is a self-experiment or a non-self-experiment. This chapter has illustrated how the results of my self-experiments in line with (1), as discussed in Chapter 5, have been replicated in a particular set of non-self-experiments, collectively referred to here as “K17s”. Section 76 On the other hand, see discussion of pairs of X and Y in Plesniak 2022, like ‘more than one professor’ and ‘that professor’, which did not seem to show sensitivity to N-head identity in relation to their Coref-DR-BVA correlations.

298 

 Hajime Hoji

2 addressed issues specific to non-self-experiments, including the ways in which they are limited when compared to self-experiments. The features of K17s were then introduced, building on the “multiple-non-researcher experiments”, as discussed in Hoji 2015, which includes “sub-preliminary experiments”. To make the subsequent discussion concrete, we addressed how we “measure” an individual participant’s reported judgments so that we can use the “measurement” in evaluating the validity of our correlational/conditional prediction in (1). (1)

Correlational/conditional prediction about c-command-detection with BVA(X, Y): Provided that DR([+cc]S, X, beta):yes ∧ Coref([+cc]S, alpha, Y):yes ∧ BVA([+cc] S, X, Y):yes; DR([-cc]S, X, beta):no ∧ Coref([-cc]S, alpha, Y):no → BVA([-cc]S, X, Y):no

Section 3 discussed the general proposal in Section 2 in concrete terms, illustrating how the result of K17s replicate the result of my self-experiment, without a single exception. The absence of other types of predictions, “variations” of (1), are then considered, as further illustration of the promise and the effectiveness of the proposed correlational methodology in LFS. It is particularly noteworthy that the result of K17s replicated the demonstration in my self-experiment of the failure of the entailment in those “variations” of (1), despite severe limitations in (the design of) K17s. The logic behind the absence of the entailment in question involves reference to something we have only limited understanding of, due to its “non-formal” nature. The replication in question, though limited for reason, has thus points to the applicability of the basic scientific method beyond the CS proper. While, as illustrated in Section 3.2, the absence of entailment can be demonstrated “existentially”, i.e., by providing a single (reliable) case where the antecedent is true and the consequent is false, the presence of entailment cannot be demonstrated in the same way. In fact, such a demonstration, i.e., the confirmation of the validity of a statement such as (1), is not possible, as pointed out in Chapters 4 and 5. Empirical/experimental support for such a prediction comes from the repeated failure to find a single case that would demonstrate that the predicted entailment does not hold. A single (reliable) case where the antecedent is true and the consequent is false shows the failure of the predicted entailment/prediction, as my judgments in fact did with regard to (74), (75), (76) and (77). I find it quite remarkable that I have not found a single case in my selfexperiment that would show that the predicted entailment/prediction in the form of (1) fails to hold, despite numerous such attempts I have made, with a number of choices of X and Y and a number of sub-types of [-cc]S, as long as I focus on the sentence patterns discussed in Chapter 5 and the hypotheses about them also

Replication: Predicted Correlations of Judgments in Japanese 

 299

discussed there. The more (and also, more rigorous) attempts we make to find a case showing the failure of the predicted entailment/prediction to hold and fail at our attempt, the more empirical/experimental support we consider ourselves to have obtained for the hypotheses that lead to (1) and for the general correlational methodology underlying (1). This applies within self-experiments, where the researcher can consider more and more choices for X and Y, alpha and beta, and other lexical and structural (and even discoursal) options available in the sentences in question (including different sub-types of [-cc]S (along with the corresponding [+cc]S), even additional hypotheses that lead to new correlational/conditional predictions of the general form as (1); see Chapter 5: Section 7 for related discussion). As we expand our empirical coverage, not just in terms of different lexical items for X, Y, etc., but in the sense just noted (for example, with various new structural hypotheses about various new sentence patterns), we expect to find a case that would demonstrate the failure of the entailment in question to hold, and that will provide us something new.77

77 As Poincaré’s (1952: Ch. 9: 150–151) remarks suggest, we should in that case “rejoice, for [we] found an unexpected opportunity of discovery. [Our] hypothesis . . . had not been lightly adopted. It took into account all the known factors which seem capable of intervention in the phenomenon. If it is not verified, it is because there is something unexpected and extraordinary about it, because we are on the point of finding something unknown and new. Has the hypothesis thus rejected been sterile? Far from it. It may be even said that it has rendered more service than a true hypothesis. Not only has it been the occasion of a decisive experiment, but if this experiment had been made by chance, without the hypothesis, no conclusion could have been drawn; nothing extraordinary would have been seen; and only one fact the more would have been catalogued, without deducing from it the remotest consequence.” The disconfirmation of our prediction could provide us with an opportunity of discovery only if it is explicitly articulated what set of hypotheses lead to the prediction and how each of the hypotheses, sometimes as a combination thereof, has been tested in prior experiments. What has been presented in Chapters 4 and 5 of this volume and this chapter is an illustration of how. When this happens in self-experiment, the LFStist can immediately check if the demonstration of the failure of the entailment in question replicates with different choices of one of the “variables” – lexical variables, structural variables, variables in terms of CS-related hypotheses, variables in terms of hypotheses/assumptions about NFS, etc. With the unlimited willingness and patience that are assumed of the researcher, the LFStist can indeed make this a good opportunity for discovery. Situations are different in the case of non-self-experiment. If a given non-researcher experimental participant, deemed reliable based on sub-preliminary experiments, reports judgments that (seemingly) constitute a demonstration of the failure of the entailment in question, we may not be in a position to ask the speaker a host of “follow-up questions” (if the speaker is no longer available for non-selfexperiment, we may not even be able to check whether the speaker would give the same set of judgments again). We should also be concerned with the possibility that the reported judgments in question might be affected by some aspect of the design of the non-self-experiment. It thus seems that the difference between self-experiment and non-self-experiment manifests itself in a

300 

 Hajime Hoji

If I had such a case in my self-experiment, that would suffice as a demonstration of the failure of the entailment in question,78 even if no such case has been observed with other speakers. If we consider my self-experiment and K17s for simple illustration, a possible demonstration of the failure of entailment in my self-experiment could not possibly be observed in K17s, in the same form, for the reason that the entailment is made in my self-experiment only with choices of X and Y that are not included in K17s.79 Note that the failure of the entailment in question (such as in (74), (75), (76) and (77)) involves the “demonstration” of J=yes, more specifically MR([-cc]S, X, Y):yes; hence, according to our hypotheses, such judgment must be based on NFS effects. The “detection” of NFS effects seems to generally correlate with the “resourcefulness” of a speaker, including the ability to “manipulate” various essentially non-structural factors. It is thus reasonable to assume that the more “resourceful” the speaker is, the easier it is to find a case that would existentially demonstrate the absence of the entailment in question. In Chapters 4 and 5, it was noted that a speaker’s judgment can change at different times, and that is why reference was made to “at a given point in time” when we addressed speaker judgments. In the researcher’s self-experiment, the researcher can try to make different points in time effectively the same as the “same point in time” by judging the relevant sentences almost simultaneously (judging one sentence right after another) and/or judging a set of relevant sentences (which can become a sizable set) back and forth, by making systematic changes in them, to see how the predicted correlations would hold up. If a given participant judges the sentences included in non-self-experiments over a span of a few to several months, as in the case of K17s, there is no guarantee that his/her judgments are based on the same “cognitive state” and that might well result in distinct “stages” analogous to my different stages discussed in Chapter 5 in relation to shifts of my judgments. The successful replication of the result of my self-experiment in K17s is, therefore, all the more remarkable for this reason. It is, in fact, remarkable that the examination of the result of K17s has not produced a single case that would show the failure of the predicted entailment/prediction in the form of (1) while it has produced cases that demonstrated

significantly more acute manner in the event of the “failure” of our prediction than in the event of its “success”. 78 And I would proceed as indicated in note 76. 79 Recall that, no combination of choices of X and Y in K17s makes the antecedent clause of (1) true for me, and that means that, given the design of K17s, there is no entailment for me to check. I would never be in the ABC intersection in K17s, with any of the ten combinations of X and Y for BVA(X, Y).

Replication: Predicted Correlations of Judgments in Japanese 

 301

the failure of the entailment in (74), (76) and (77).80 Again, this seems to indicate how effective the correlational methodology may be. The discussion of the correlational methodology in this volume, as it pertains to (1), has focused on how it effectively controls for variable effects due to the choice of X and that of Y, and enables c-command detection with BVA(X, Y), based on the DR test and the Coref test. The result of K17s seems to indicate that the correlational methodology in fact controls for variable effects due to factors other than the choices of X and Y.81 Given that the most crucial experiments in LFS are self-experiments, the most crucial replication in LFS is replication of results of self-experiments. Given the difference between what we can expect of a researcher and what we can expect of a non-researcher, in terms of their willingness and patience, however, replication of result of self-experiment in non-self-experiment on non-researchers faces serious challenges.82 K17s, in fact, replicates the result of my self-experiment in a very specific way, strictly focusing on the correlational prediction, not dealing with the lexical items crucially used in my self-experiments at all. Given the general skepticism against the use of the researcher’s own introspective judgments in language-related fields, replication in non-self-experiments is perhaps a necessary part of demonstration in LFS, and for this reason, it is perhaps necessary for us to further articulate methodology for non-self-experiment, as addressed in Chapter 8 of this volume and Plesniak 2022.83

80 As discussed in Section 3.2.3, the fact that we did not find a single speaker in K17s whose reported judgments would have demonstrated the absence of the entailment in (75) was not particularly surprising. 81 Plesniak 2022: 2.7 and 3.2 discusses how the effectiveness of the correlational methodology in question is derived. 82 Similar issues can arise with non-self-experiments on researchers unless they are just like the self in all the relevant respects; in this sense, it may be more useful to focus on the dichotomy between self-experiments and non-self-experiments, regardless of whether the latter is on a researcher or a non-researcher. 83 Based on the further articulation, we should be able to prepare a manual for non-self-experiments, as we should be able to prepare a manual for self-experiments, based on further articulation of the methodology for self-experiment, in the spirit of Chapter 3 of this volume.

302 

 Hajime Hoji

Appendix I: Samples of Actual Sentences in K17s (111)

DR(subete-no tihoozititai, 3-nin-no seizika) ‘DR(every local government, three politicians)’ a. Sentence instantiating [+cc]S: (“Dono 3-nin-no seizika-ka-wa zititai-goto-ni   which 3-cl-gen politician-Q-top local:government-each-for kotonaru-ga, differ-but 3-nin-no seizika-o hihansita toyuu koto-ga subete-no 3-cl-gen politician-acc criticized that fact-nom every-gen tihoozititai-ni atehamaru rassi toyuu imi-de. local:government-dat applies it:seems that meaning-with hihansareta seizika-no kazu-wa tihoozititai-no was:criticized politician-gen number-top local:government-gen kazu-no 3-bai.) number-gen 3 times ‘Under the interpretation that it seems that having criticized three politicians holds of every local government, and which three politicians differs depending upon each local government. The number of the criticized politicians is three times as many as the number of the local governments’ 3-nin-no seizika-o subete-no tihoozititai-ga 3-cl-gen politician-acc every-gen local:government-nom hihansita rasii. criticized it:seems ‘It seems that three politicians, every local government criticized.’ b. Sentence instantiating [-cc]S: (“Dono 3-nin-no seizika-ka-wa zititai-goto-ni Which 3-cl-gen politician-Q-top local:government-each-for kotonaru-ga, differ-but 3-nin-no seizika-ni hihansareta toyuu koto-ga 3-cl-gen politician-dat was:criticized that fact-nom subete-no tihoozititai-ni atehamaru rassi toyuu every-gen local:government-dat applies it:seems that

Replication: Predicted Correlations of Judgments in Japanese 

 303

imi-de. hihansita seizika-no kazu-wa meaning-with criticized politician-gen number-top tihoozititai-no kazu-no 3 bai.) local:government-gen number-gen 3 times ‘Under the interpretation that it seems that having been criticized by three politicians holds of every local government, and which three politicians differs depending upon each local government. The number of the politicians who did the criticizing is three times as many as the number of the local governments’ 3-nin-no seizika-ga subete-no tihoozititai-o 3-cl-gen politician-nom every-gen local:government-acc hihansita rasii. criticized it:seems ‘It seems that three politicians criticized every local government.’ c. The same form as (b), but without DR(X, beta): (Tokuteino 3-nin-no seizika-ni hihansareta toyuu specific 3-cl-gen politician-dat was:criticized that koto-ga subete-no tihoozititai-ni atehamaru rassi fact-nom every-gen local:government-dat applies it:seems toyuu imi-de. hihansita seizika-no kazu-wa 3-nin.) that meaning-with criticized politician-gen number-top 3–cl ‘Under the interpretation that it seems that having been criticized by a particular group of three politicians seems to hold true of every local government. The number of the politicians is three.’ 3-nin-no seizika-ga subete-no tihoozititai-o 3-cl-gen politician-nom every-gen local:government-acc hihansita rasii. criticized it:seems ‘It seems that three politicians criticized every local government.’ (112)

Coref(aitu, soitu) ‘Coref(that guy, this guy)’: a. Sentence instantiating [+cc]S: (‘Aitu’ to ‘soitu’-ga onazi hito-o sasu that guy and that guy-nom same person-acc refer:to kaisyaku (‘aitu = ‘soitu’   de)) interpretation that:guy that:guy with

304 

 Hajime Hoji

‘Under the interpretation that ‘aitu’ and ‘soitu’ refer to the same person (‘aitu’ = ‘soitu’).’ Soitu-no bengosi-o aitu-ga hihansita rasii. that:guy-gen attorney-acc that:guy-nom criticized it:seems ‘It seems that guy criticized that guy’s attorney.’ b. Sentence instantiating [-cc]S: (‘Aitu’ to ‘soitu’-ga onazi hito-o sasu that guy and that guy-nom same person-acc refer:to kaisyaku (‘aitu = ‘soitu’   de)) interpretation that:guy that:guy with ‘Under the interpretation that ‘aitu’ and ‘soitu’ refer to the same person (‘aitu’ = ‘soitu’).’ Soitu-no bengosi-ga aitu-o hihansita rasii. that:guy-gen attorney-nom that:guy-acc criticized it:seems ‘It seems that that guy’s attorney criticized that guy.’ c.

(113)

The same form as (b), but without Coref(alpha, Y): (‘Aitu’ to ‘soitu’-ga betubetu-no hito-o sasu that guy and that guy-nom different-gen person-acc refer:to kaisyaku (‘aitu ≠ ‘soitu’   de)) interpretation that:guy that:guy with ‘Under the interpretation that ‘aitu’ and ‘soitu’ refer to different people (‘aitu’ ≠ ‘soitu’).’ Soitu-no bengosi-ga aitu-o hihansita rasii. that:guy-gen attorney-nom that:guy-acc criticized it:seems ‘It seems that that guy’s attorney criticized that guy.’

BVA(subete-no zyooin giin, soitu) ‘BVA(every senator, that guy)’ a. Sentence instantiating [+cc]S: (“Zibun-no daiiti hisyo-o hihansita toyuu koto-ga    own-gen first secretary-acc criticized that fact-nom

Replication: Predicted Correlations of Judgments in Japanese 

 305

subete-no zyooin giin-ni atehamaru rasii” toyuu imi-de. all-gen Senator-dat be:true it:seem that meanig-with Hihansareta daiiti hisyo-wa sukunakutomo was:criticized first secretary-top at:least zyooin giin-no kazu dake iru rasii.) Senator-gen number as:many:as exist it:seems ‘Under the interpretation that having criticized their own first secretary holds true of every senator. There are at least as many criticized first secretaries as there are senators’. So-itu-no daiiti hisyo-o subete-no zyooin giin-ga that-guy-gen first secretary-acc all-gen Senator-nom hihansita rasii.84 criticized it:seems ‘It seems that every senator criticized his or her first secretary’. b. Sentence instantiating [-cc]S: (“Zibun-no daiiti hisyo-ni hihansareta toyuu koto-ga  own-gen first secretary-dat was:criticized that fact-nom subete-no zyooin giin-ni atehamaru rasii” toyuu imi-de. every-gen Senator-dat be:true it:seems that meanng-with Hihansita daiiti hisyo-wa sukunakutomo zyooin giin-no criticized first secretary-top at:least Senator-gen kazu dake iru.) number as:many:as exist ‘Under the interpretation that having been criticized by their own first secretaries holds true of every senator’’. There are at least as many first secretaries who did criticizing as there are senators.’ So-itu-no daiiti hisyo-ga subete-no zyooin giin-o That-guy-gen first secretary-nom all-gen Senator-acc hihansita rasii. criticized it:seems ‘It seems that his or her first secretaries criticized every senator’.

84 In K17s, the intended anaphoric relation is indicated by means of underlining instead of italics.

306 

 Hajime Hoji

c.

The same form as (b), but without BVA(X, Y): (“So-itu”-ga senkoobunmyaku-de kisyutu-no  that-guy-nom previous:context-in already:mentioned-gen aru zinbutu-o sasu kaisyaku-de) certain person-acc refer:to interpretation-with ‘Under the interpretation where “so-itu” refers to someone who has already been mentioned in the prior discourse’ So-itu-no daiiti hisyo-ga subete-no zyooin giin-o That-place-gen first secretary-nom all-gen Senator-acc hihansita rasii. criticized it:seems ‘It seems that his/her first secretary criticized every senator’.

Appendix II: Some Notations The relevant notations introduced in Chapters 4 and 5, and crucially used in this chapter, are given in (114) and (115). (114) (=Chapter 5: (3)) a. PS: phonetic sequence b. X and Y: linguistic expressions in PS c. MR(X, Y): meaning relation pertaining to X and Y d. LF(PS): a 3D (i.e., LF) representation corresponding to PS e. LF(X): what corresponds X in LF(PS) f. [-cc]S: a schematized PS that cannot correspond to LF(PS) in which LF(X) c-commands LF(Y) g. [+cc]S: a schematized PS that can correspond to LF(PS) in which LF(X) c-commands LF(Y) (115) (=Chapter 5: (4)) a J in “MR(S, X, Y):J”: an individual speaker’s judgment (yes or no) on the availability of MR(X, Y) in a particular PS b. yes in “MR(S, X, Y):yes”: an individual speaker’s judgment that the MR(X, Y) is available in a particular PS c. no in “MR(S, X, Y):no”: an individual speaker’s judgment that the MR(X, Y) is not available in a particular PS

Replication: Predicted Correlations of Judgments in Japanese 

 307

Appendix III More Results BVA(asoko to koko, soko) ‘BVA(that place and this place, it)’ (116) For (70a) (BVA(asoko to koko, soko) ‘BVA(that place and this place, it)’) a. Without “sub-preliminary experiments”

308 

 Hajime Hoji

Results

Green

Blue

Red

Total

ABC

6

0

0

6

AB

2

0

10

12

AC

3

0

0

3

BC

14

63

19

96

A

1

0

6

7

B

0

10

18

28

C

3

13

6

22

None

3

2

0

5

Total

32

88

59

179

b. With “sub-preliminary experiments” using the same choices of X and Y of BVA(X, Y) and the a-NP counterpart of the Y.

Replication: Predicted Correlations of Judgments in Japanese 

Results

Green

Blue

Red

Total

ABC

2

0

0

2

AB

1

0

2

3

AC

0

0

0

0

BC

18

63

19

100

A

0

0

0

0

B

1

10

26

37

C

6

13

6

25

None

4

2

6

12

Total

32

88

59

179

BVA(subete-no N, soko) ‘BVA(every N, it)’ (117) For (70b) (BVA(subete-no N, soko) ‘BVA(every N, it)’) a. Without “sub-preliminary experiments”

 309

310 

 Hajime Hoji

Results

Green

Blue

Red

Total

ABC

3

0

2

5

AB

4

0

2

6

AC

2

0

0

2

BC

13

70

13

96

A

0

0

9

9

B

5

16

7

28

C

7

10

7

24

None

2

4

3

9

Total

36

100

43

179

b. With “sub-preliminary experiments” using the same choices of X and Y of BVA(X, Y) and the a-NP counterpart of the Y.

Replication: Predicted Correlations of Judgments in Japanese 

Results

Green

Blue

Red

Total

ABC

1

0

0

1

AB

0

0

1

1

AC

1

0

0

1

BC

15

70

15

100

A

0

0

0

0

B

9

16

8

33

C

8

10

7

25

None

2

4

12

18

Total

36

100

43

179

BVA(#-cl-no N, soko) ‘BVA(# N’s, it)’ (118) For (70c) (BVA(#-cl-no N, soko) ‘BVA(# N’s, it)’) a. Without “sub-preliminary experiments”

 311

312 

 Hajime Hoji

Results

Green

Blue

Red

Total

ABC

7

0

2

9

AB

5

0

2

7

AC

0

0

2

2

BC

4

70

12

86

A

2

0

10

12

B

4

14

4

22

C

7

11

10

28

None

2

4

5

11

Total

31

99

47

177

b. With “sub-preliminary experiments” using the same choices of X and Y of BVA(X, Y) and the a-NP counterpart of the Y.

Replication: Predicted Correlations of Judgments in Japanese 

Results

Green

Blue

Red

 313

Total

ABC

2

0

0

2

AB

0

0

0

0

AC

0

0

0

0

BC

9

70

14

93

A

1

0

1

2

B

9

14

6

29

C

7

11

12

30

None

3

4

14

21

Total

31

99

47

177

BVA(sukunakutomo #-cl-no N, soko) ‘BVA(at least # or more N, it) (119) For (70d) (BVA(sukunakutomo #-cl-no N, soko) ‘BVA(at least # or more N, it)’) a. Without “sub-preliminary experiments”

314 

 Hajime Hoji

Results

Green

Blue

Red

Total

ABC

6

0

1

7

AB

4

0

6

10

AC

2

0

3

5

BC

14

49

13

76

A

2

0

14

16

B

4

4

6

14

C

8

21

10

39

None

1

7

4

12

Total

41

81

57

179

b. With “sub-preliminary experiments” using the same choices of X and Y of BVA(X, Y) and the a-NP counterpart of the Y.

Replication: Predicted Correlations of Judgments in Japanese 

Results

Green

Blue

Red

Total

ABC

2

0

0

2

AB

1

0

2

3

AC

0

0

0

0

BC

18

49

14

81

A

0

0

0

0

B

7

4

10

21

C

10

21

13

44

None

3

7

18

28

Total

41

81

57

179

 315

BVA(N-cm sukunakutomo #-cl, soko) (120) For (70e) (BVA(N-cm sukunakutomo #-cl, soko) (the same as (70d) except that the X is in the so-called “floating numeral” form)) a. Without “sub-preliminary experiments”

316 

 Hajime Hoji

Results

Green

Blue

Red

Total

ABC

8

0

2

10

AB

5

0

3

8

AC

1

0

0

1

BC

16

59

18

93

A

1

0

13

14

B

3

6

10

19

C

2

8

11

21

None

1

2

8

11

Total

37

75

65

177

b. With “sub-preliminary experiments” using the same choices of X and Y of BVA(X, Y) and the a-NP counterpart of the Y.

Replication: Predicted Correlations of Judgments in Japanese 

Results

(121)

Green

Blue

Red

Total

ABC

2

0

0

2

AB

1

0

1

2

AC

0

0

0

0

BC

22

59

20

101

A

0

0

0

0

B

7

6

12

25

C

3

8

11

22

None

2

2

21

25

Total

37

75

65

177

 317

With “wrong” “sub-preliminary experiments” (using ones with soitu/aitu (and subete-no N as X of BVA(X, soitu/aitu)), instead of using ones with soko/asoko (and N-cm sukunakutomo #-cl as X of BVA(X, soko/asoko)):

318 

 Hajime Hoji

Results

Green

Blue

Red

Total

ABC

3

0

1

4

AB

2

0

1

3

AC

0

0

0

0

BC

21

59

19

99

A

0

0

1

1

B

6

6

12

24

C

3

8

11

22

None

2

2

20

24

Total

37

75

65

177

BVA(aitu to koitu, soitu) ‘BVA(that guy and this guy, that guy)’ (122) For (71a) (BVA(aitu to koitu, soitu) ‘BVA(that guy and this guy, that guy)’) a. Without “sub-preliminary experiments”

Replication: Predicted Correlations of Judgments in Japanese 

Results

Green

Blue

Red

ABC

4

0

1

5

AB

1

0

2

3

AC

0

0

2

2

BC

6

77

18

101

A

2

0

13

15

B

2

11

10

23

C

4

12

4

20

None

1

3

9

13

Total

20

103

59

182

 319

Total

b. With “sub-preliminary experiments” using the same choices of X and Y of BVA(X, Y) and the a-NP counterpart of the Y.

320 

 Hajime Hoji

Results

Green

Blue

Red

Total

ABC

3

0

0

3

AB

0

0

0

0

AC

0

0

0

0

BC

7

77

19

103

A

1

0

1

2

B

3

11

11

25

C

4

12

6

22

None

2

3

21

26

Total

20

103

58

181

BVA(subete-no N, soitu) ‘BVA(every N, that guy)’ (123) For (71b) (BVA(subete-no N, soitu) ‘BVA(every N, that guy)’) a. Without “sub-preliminary experiments”

Replication: Predicted Correlations of Judgments in Japanese 

Results

Green

Blue

Red

Total

ABC

4

0

0

4

AB

3

0

0

3

AC

1

0

0

1

BC

7

82

11

100

A

0

0

12

12

B

5

21

7

33

C

5

17

5

27

None

1

2

3

6

Total

26

122

38

186

 321

b. With “sub-preliminary experiments” using the same choices of X and Y of BVA(X, Y) and the a-NP counterpart of the Y.

322 

 Hajime Hoji

Results

Green

Blue

Red

Total

ABC

3

0

0

3

AB

1

0

0

1

AC

1

0

0

1

BC

8

82

11

101

A

0

0

0

0

B

7

21

7

35

C

5

17

5

27

None

1

2

15

18

Total

26

122

38

186

BVA(#-cl-no N, soitu) ‘BVA(# N’s, that guy)’ (124) For (71c) (BVA(#-cl-no N, soitu) ‘BVA(# N’s, that guy)’) a. Without “sub-preliminary experiments”

Replication: Predicted Correlations of Judgments in Japanese 

Results

Green

Blue

Red

ABC

5

0

0

5

AB

1

0

3

4

 323

Total

AC

1

0

2

3

BC

10

76

6

92

A

3

0

8

11

B

2

18

6

26

C

5

16

9

30

None

2

5

6

13

Total

29

115

40

184

b. With “sub-preliminary experiments” using the same choices of X and Y of BVA(X, Y) and the a-NP counterpart of the Y.

324 

 Hajime Hoji

Results

Green

Blue

Red

Total

ABC

3

0

0

3

AB

1

0

0

1

AC

0

0

0

0

BC

12

76

6

94

A

0

0

1

1

B

2

18

9

29

C

6

16

11

33

None

5

4

13

22

Total

29

114

40

183

(BVA(sukunakutomo #-cl-no N, soitu) ‘BVA(at least # or more N, that guy)’ (125) For (71d) (BVA(sukunakutomo #-cl-no N, soitu) ‘BVA(at least # or more N, that guy)’) a. Without “sub-preliminary experiments”

Replication: Predicted Correlations of Judgments in Japanese 

Results

Green

Blue

Red

 325

Total

ABC

3

0

0

3

AB

3

0

6

9

AC

2

0

2

4

BC

10

62

9

81

A

1

0

14

15

B

3

10

3

16

C

5

28

11

44

None

2

7

4

13

Total

29

107

49

185

b. With “sub-preliminary experiments” using the same choices of X and Y of BVA(X, Y) and the a-NP counterpart of the Y.

326 

 Hajime Hoji

Results

Green

Blue

Red

Total

ABC

3

0

0

3

AB

0

0

1

1

AC

0

0

0

0

BC

10

62

9

81

A

0

0

0

0

B

6

10

8

24

C

7

27

13

47

None

3

7

18

28

Total

29

106

49

184

BVA(N-cm sukunakutomo #-cl, soitu) (126) For (71e) (BVA(N-cm sukunakutomo #-cl, soitu) (the same as (71d) except that the X is in the so-called “floating numeral” form) a. Without “sub-preliminary experiments”

Replication: Predicted Correlations of Judgments in Japanese 

Results

Green

Blue

Red

Total

ABC

6

0

0

6

AB

5

0

4

9

AC

0

0

1

1

BC

26

53

20

99

A

3

0

13

16

B

8

3

8

19

C

5

5

13

23

None

2

0

7

9

Total

55

61

66

182

 327

b. With “sub-preliminary experiments” using the same choices of X and Y of BVA(X, Y) and the a-NP counterpart of the Y.

328 

 Hajime Hoji

Results

Green

Blue

Red

Total

ABC

4

0

0

4

AB

1

0

1

2

AC

0

0

0

0

BC

28

53

20

101

A

0

0

0

0

B

12

3

11

26

C

5

5

14

24

None

5

0

20

25

Total

55

61

66

182

References Hayashishita, J.-R. 2004. Syntactic and non-syntactic scope. Los Angeles, CA: University of Southern California dissertation. Hayashishita, J.-R. 2013. On the nature of inverse scope readings. Gengo Kenkyu 143. 29–68. Hayashishita, J.-R. and Ayumi Ueyama. 2012. Quantity expressions in Japanese. In Edward Keenan and Denis Paperno (eds.), The handbook of quantification in natural language, 535–612. New York: Springer. Hoji, Hajime. 1985. Logical form constraints and configurational structures in Japanese. Seattle, WA: University of Washington dissertation. Hoji, Hajime. 2003. Falsifiability and repeatability in generative grammar: A case study of anaphora and scope dependency in Japanese. Lingua 113. 377–446. Hoji, Hajime. 2015. Language faculty science. Cambridge: Cambridge University Press. Hoji, Hajime, Satoshi Kinsui, Yukinori Takubo and Ayumi Ueyama. 2003. The demonstratives in modern Japanese. In Yen-Hui Audrey Li and Andrew Simpson (eds.), Functional structure(s), form and interpretation, 97–128. New York: Routledge. Mukai, Emi. 2012. Binding and scope dependencies with ‘floating quantifiers’ in Japanese. Los Angeles, CA: University of Southern California dissertation. Plesniak, Daniel. 2021. Experiments beyond statistics: How generative linguistics can beat the replication crisis. An early draft available at: https://www.researchgate.net/ publication/357063705_Experiments_Beyond_Statistics_How_Generative_Linguistics_ Can_Beat_the_Replication_Crisis (accessed 21 March 2022) Plesniak, Daniel. 2022. Towards a correlational law of language: Three factors constraining judgment variation. Los Angeles, CA: University of Southern California dissertation. Poincaré, Henri. 1952. Science and hypothesis. New York: Dover Publications. (The English translation of La science et l’hypothèse (1902).) Ueyama, Ayumi. 1998. Two types of dependency. Los Angeles, CA: University of Southern California dissertation.

Daniel Plesniak

Predicted Correlations of Judgments in English 1 Introduction In this chapter, I discuss experiments performed in English according to the correlational methodology discussed by Hoji in Chapters 4–6 of this volume. While the experiments conducted are implemented somewhat differently than Hoji’s and are conducted in English rather than Japanese, the basic result is the same: our categorical predictions about the (un)acceptability of ✶-schemata receive definite experimental support once we take into account correlations of judgments. This chapter discusses the results of three experiments. The first one, discussed in the section “Replicating Correlations”, centers around the same sort of “weak crossover” construction that Hoji used to establish the correlational methodology in Chapters 5 and 6. As noted, it turns out that, despite differences in implementation and language, the results obtained in English essentially replicate what was found by Hoji in Japanese. This success raises one potential concern about the correlational methodology, however, namely that it might be giving us the “good-looking” results for the wrong reasons. That is, Hoji claims that the correlations come out as predicted only as a result of the experimenter having correctly identified the relevant c-command relations between elements in the schemata in question, but how do we know this is true? What if, for example, the correlations obtain regardless of whether we have correctly guessed the correct c-command relations in the schemata under investigation? If that were the case, then the methodology would not be useful for detecting the presence of absence of c-command relations at all; the same result would obtain whether or not we had come up with the correct structural hypotheses for the sentence-types involved, allowing us to get a “positive result” even if we had made a wildly inaccurate guess. The second experiment discussed in this chapter, in the section “Absence of Correlations with Bad Hypotheses” examines such a case, namely what would happen if we erroneously hypothesized that the object (O) in an OSV (topicalized) construction is never c-commanded by the subject (S). This incorrect assumption would lead us to believe that [Y Noun, X Verb] should not be compatible with a c-command-based BVA(X, Y) interpretation, and thus BVA(X, Y) should not be possible after correlations are taken into account. As it turns out, experiments conducted in keeping with the correlational methodhttps://doi.org/10.1515/9783110724790-007

330 

 Daniel Plesniak

ology do not simply support all hypotheses no matter how wrong; in this case, they fail to provide support for our erroneous hypothesis. The results of the OSV experiment look categorically different from the results of the “weak crossover” experiment, correctly failing to support an incorrect hypothesis. This lends credence to the notion that the correlational methodology yields supportive results only when we have correctly identified the c-command relations in the schemata involved. The third experiment discussed in this chapter, in the section “Checking New Constructions”, addresses yet another concern: does the correlational methodology work for constructions other than weak crossover and reconstruction, using which it was initially developed? If the answer were “no”, then it would not be as generally useful as hoped, but thankfully, the answer is “yes”: I show that so-called “spec-binding” constructions, wherein the binder is a possessor inside of the subject, and thus is hypothesized to not c-command the object, show the exact same categorical pattern of results as the “weak crossover” constructions. In short, once correlations of judgments are taken into account, all cases of apparent binding from possessors can be accounted for in a principled manner such that they are as predicted by c-command-based hypotheses. The correlational methodology thus can give categorical results regarding constructions it was not initially developed to test, suggesting its long-term viability as a probe into the language faculty.1

2 Preliminary Remarks Before moving on to the discussion of the experiments, I provide a brief summary and discussion of the relevant theoretical and methodological matters addressed in pervious chapters. The intent is to review and make clear how these matters pertain to the experiments at hand, and thus to give readers a clear sense as to how these experiments bear on the concepts that have been discussed. Following discussion in Chapters 4–6, the most essential points to keep in mind are: (I) the object of inquiry is the Computational System (CS, Chomsky 1995), specifically as it pertains to the link between the external form of a sentence/utterance

1 In particular, this demonstrates the usefulness of the correlational methodology when we try to determine correct structural analyses of various constructions, as briefly addressed in the “Expanding the empirical coverage” section of Hoji’s Chapter 5 of this volume. That section made use of Hoji’s self-experiment to make the intended point, whereas this chapter shows that similar results can be obtained via non-self-experiments.

Predicted Correlations of Judgments in English  

 331

and its “internal” meaning/interpretation. By finding what form-meaning pairs are possible, we see what the generative procedure can produce; likewise, by finding what form-meaning pairs are not possible, we find its limitations. This last consideration regarding impossibility and limitations informs (II): following Hoji (2015 and Chapter 4 of this volume2), I assume that only negative predictions about such pairs are rigorously testable; if we predict a certain form-meaning pair to be generable, there are a myriad of reasons why an individual might not successfully generate such a form-meaning pair at a given time. Some of these reasons may have to do with purely CS-internal issues, such as the individual failing to notice a possible structural “parse” for a given string, or they may arise from more complex issues involving other factors beyond the CS. On the other hand, if a form-meaning pair is predicted to be un-generatable, then no individual should be able to generate it, no matter what other factors intervene. In other words, while possible things may end up not happening, impossible things, by definition, never happen. Finally, (III): such predictions must be checked at the level of the individual. As Hoji (2017) notes, in the spirit of Chomsky’s (1986a) conception of the “I-language”, each individual has their own internal grammar. As such, terms like “English” are really labels of convenience for groups of similar I-languages. The question of whether a given form-meaning pair is acceptable in “English” thus cannot necessarily be answered with a direct yes or no. Instead, we expect there to be a great deal of variation from speaker to speaker, especially regarding aspects of the language faculty that are not themselves part of the CS itself; indeed, as noted in the preceding chapters, even within a single speaker, we expect judgments can, and likely will, vary across time. Our task then is to articulate theories that can survive speaker-to-speaker (or time-to-time) variation, yet still maintain rigorous testability. The key hypothesis underlying these experiments, identical to those in Chapters 4–6, is that the CS constrains the form-meaning mappings through c-command relations. We thus need to address two questions: (A) for a given utterance or utterance type, what syntactic structure(s) do speakers of the relevant language assign to it, and thus what c-command relations obtain between its elements, and (B) in what way do these c-command relations constrain the utterance’s meaning? To do this, using a term from Hoji’s Chapters 4–6 in this volume,

2 The discussion in these works draws on the Ueyama (2010) model of judgment-making; the relevant points underlying this focus on negative predictions perhaps ultimately trace back to Hoji 2003.

332 

 Daniel Plesniak

we need to find some meaning relation between elements X and Y (MR(X, Y)) such that MR(X, Y) is reliant on X c-commanding Y.3 One such MR is provided by BVA (“Bound Variable Anaphora”). Following Ueyama 1998 (p. 2), it is important to keep in mind that, as an MR, BVA is itself a purely observational concept, rather than a theoretical one. While it is of longterm importance to formulate a coherent and useful definition of what sorts of interpretations “count” as BVA, for now, we can simply get a general sense, as the cases to be considered are quite “canonical” instances of BVA, and thus there is no reason for concern as to their proper classification.4 Consider, for example: (1) Every boy loves his mother. In (1), there can be an interpretation wherein each boy loves his own mother, e.g., boy A loves boy A’s mother, boy B loves boy B’s mother, etc. We will refer to such a reading as BVA(every boy, his); it is also called the “quantificationally bound” reading, where ‘every boy’ “binds” ‘his’. Reinhart (1983) proposed that such a reading is permissible only under certain structural conditions, namely that the binder (e.g., ‘every boy’) must c-command the bindee (e.g., ‘his’). Using various tests (Lasnik 1990 provides a compact summary, for example), it has long been argued that subjects c-command objects, but not vice versa. If Reinhart’s (1983) hypothesis for BVA is correct, then BVA(every boy, his) should be possible in sentences like (1),5 where the binder is the subject, ‘every boy’, and the bindee, ‘his’, is within the object, recalling that if the subject c-commands the object, it definitionally c-commands everything contained within the object. Supporting this prediction, (1) is indeed widely judged acceptable with the BVA(every boy, his) reading. As noted, however, our focus cannot be solely on sentences where an MR is acceptable; what we need to predict is unacceptability. To do so, Hoji uses cases

3 To be more precise, the elements corresponding to X and Y in an individual’s structural representation of the sentences in question should have this property. I will use “X” and “Y” to refer to both the external form and these internal, structural representations interchangeably in this chapter. Of course, they are not the same, but it should generally be clear which one is being referred to in a given context. 4 If one wants a “rough” definition for BVA, we can provisionally accept Hoji 2015’s for now: BVA(α, β) is when “the reference invoked by a singular-denoting expression β co-varies with what is invoked by [a] non-singular denoting expression α”. 5 At least if nothing else blocks it. Reinhart’s theory holds that c-command is necessary, not sufficient. See further discussion in Hoji’s Chapters 4–6 of this volume.

Predicted Correlations of Judgments in English  

 333

of “weak crossover,” (Postal 1971, Wasow 1972). For our purposes, we can instantiate the relevant pattern simply by reversing subject and object6 from (1): (2) His mother loves every boy. Given that the object is hypothesized to not c-command the subject, the prediction of Reinhart 1983’s account is that BVA(every boy, his) should be impossible in this sentence. However, it has been known since the identification of such constructions that this is not always the case (hence the use of “weak” in “weak crossover”, signifying weak effects). As such, the original pure c-command approach cannot be maintained without some sort of modification. To achieve this, we will follow Hoji’s discussion in Chapters 4–6 of this volume in assuming that there are multiple ways to achieve BVA(X, Y), with X c-commanding Y being just one of them. One way, which I will return to in discussing Experiment 3 is X preceding Y; following the hypotheses of Ueyama 1998, such a configuration can permit BVA(X, Y) even when X does not c-command Y. This, however, cannot be the source here, as ‘every boy’ does not precede ‘his. The more salient way or ways to achieve BVA in (2) are via certain effects obtaining on either X or Y, which cause X and/or Y to become effectively structurally insensitive. Hoji’s Chapter 5 of this volume develops a more detailed account of such effects, under the label of the effects “non-formal sources” (NFS). However, what underlies such structural insensitivity, while of great importance generally, is not of key import for our purpose here (see Plesniak 2022 for discussion of various proposals, including those laid out in Ueyama (1998 and Chapter 2 of this volume), and Hoji (Chapter 5 of this volume)). What is essential is how such insensitivity can be detected. Following the general program laid out in Hoji 2017 (which has been expanded on in Hoji’s Chapter 4 of this volume), in order to make definite predictions about the acceptability of given MR(X, Y), e.g. BVA(X, Y), it is possible to use other MR’s to independently assess the structural (in)sensitivity of X and Y. Specifically, to detect possible interference in MR1(X, Y) arising from structural insensitivity of X or Y, we can check an individual’s judgments on some MR2(X, Y’) and MR3(X’, Y), where X’ and Y’ are elements different than X or Y. If the individual’s judgments on MR2 and MR3 conform to c-command-based predictions, and we believe the factors

6 There are some who do not consider the term “crossover” applicable except in cases where there is overt displacement of a wh-element, as was the case in its original formulation. However, this broader usage has already entered into scholarly parlance (see for example Barker 2012), so I will continue to use it here for lack of a better term.

334 

 Daniel Plesniak

that enable MR1 are a subset of those that enable MR2 and MR3, then MR1(X, Y) should also conform to c-command-based predictions. For MR1=BVA(X, Y), Hoji (Chapters 4–6 this volume) uses MR2=DR(X, Y’) and MR3=Coref(X’, Y). This creates a paradigm like the one below: (3) BVA ok Every teacher spoke to his student. With BVA(every teacher, his) (4) BVA ✶ His student spoke to every teacher With BVA(every teacher, his) (5) DR ok Every teacher spoke to three students. With DR(every teacher, three students) (6) DR ✶ Three students spoke to every teacher With DR(every teacher, three students) (7) Coref ok John spoke to his student. With Coref(John, his) (8) Coref ✶ His student spoke to John. Coref(John, his) In (5) and (6), DR(every teacher, three students) refers to the MR where the ‘three students’ in question vary in interpretation based on the teacher, e.g., Teacher A spoke to students 1–3, Teacher B to students 4–6, etc.; there will be three times as many students spoken to as there are teachers who spoke. In (7) and (8), Coref(John, his) refers to the interpretation where ‘his’ can be understood as meaning ‘John’s’. If an individual accepts the relevant DR and Coref ok items at least sometimes and rejects the relevant DR and Coref ✶ items always, then that individual has demonstrated that X and Y (in this case ‘every teacher’ and ‘his’) are not structurally insensitive; X seems to need to c-command Y’ for DR(X, Y’) to be possible, and Y seems to need to be c-commanded by X’ for Coref(X’, Y) to be possible, suggesting that both X and Y are sensitive to structure in terms of

Predicted Correlations of Judgments in English  

 335

their ability to participate in MR interpretations. As such, when we turn to BVA, we predict that BVA(X, Y) must be impossible in cases like (2), where X does not c-command Y (or precede it). As such, the individual in question must either consistently reject the ✶ schema, or, if that does not happen, we must conclude that our hypotheses were wrong. Hoji (Chapters 5 and 6 of this volume) finds that, as far as weak crossover in Japanese is concerned, individuals using “structurally sensitive” X and Y universally do indeed reject the relevant ✶-schema instantiations (weak crossover constructions) as predicted. One significant but nevertheless reasonable caveat is that, as discussed in Hoji 2015, achieving these results relies on various “subexperiments” to ensure that participants understand and are attentive to the questions being asked; those who show signs of inattentiveness/miscomprehension are largely excluded from consideration, as their reported judgments may not, in fact, accurately correspond to what their intuitions about the availability of BVA(X, Y) would be if they were attentive and understood the questions being asked. As noted, Hoji’s experiment, while it covers a great deal of ground, especially in terms of its use of many different choices of X, is confined to speakers of Japanese. Given, however, that most of the hypotheses Hoji relies on are claimed to be universal, e.g., that connection between BVA and c-command and the asymmetric c-command of objects by subjects, we expect that such results ought to obtain in other languages too. As will be shown by Experiment 1, this is indeed the case, at least for English.

3 Experiment 1: Replicating Correlations To reiterate, for our purposes, “weak crossover” constructions will refer to pairings of sentences where Y is within the subject of the sentence and X is the object of the sentence with interpretations of the type MR(X, Y). Assuming the hypothesis that objects do not c-command subjects, we expect BVA(X, Y) in such sentences to be rejected if the availability of BVA for an individual requires X to c-command Y. Hoji (chapter 6 of this volume) gives the results of experiments in Japanese showing that, assuming correlations with DR and Coref are considered alongside filtering by attentiveness/comprehension-checking sub-experiments, then c-command-based predictions about BVA in weak crossover constructions survive rigorous testing across all individuals considered. That is, if a Japanese sentence had the form Y-no Noun-ga X-o/ni Verb (English equivalent: Y’s Noun Verb X), then all individuals: (I) rejected BVA(X, Y), or (II) had judgments on DR/Coref that suggested that non-structurally sensitive X/Y’s may be disrupting c-command-based

336 

 Daniel Plesniak

patterns in BVA judgments, or (III) had judgments on sub-experiments that suggested the possibility of attentiveness/comprehension issues. That is, weak crossover constructions do indeed disallow BVA, just so long as no sources of non-structural noise intervene (be those sources of “grammatical” noise like the insensitivity of X/Y or of “incidental” noise like a person not paying attention). The experiment performed for this section was conducted to check whether analogous results to Hoji’s might be found in English. Its scale is somewhat smaller than Hoji’s; two choices of X were used, ‘every teacher’ and ‘more than one teacher’, with approximately 100 individuals judging each choice (200 total). There was only one choice of Y, ‘his’. The crucial sentences were thus of the form Y’s Noun Verb X, as in: (9) His student spoke to every teacher with BVA(X, Y) A similar experiment to this was performed earlier in Plesniak (2021),7 but notably, it featured far fewer experimental items (essentially only one instantia-

7 I put below the results as given in Plesniak (2021), which are presented in a similar (but slightly different) format than the ones to be given in the rest of chapter. All Participants

Supporting Neutral Violating

BVA Inst-Sub

Non-Exceptional Y

Non-Exceptional X

The three circles, clockwise from the top, represent: (i) whether the participants gave the intended answers on the attentiveness/comprehension checking question that tested whether they understood the phrases being used to convey a BVA reading as intended (the “BVA Inst-Sub”), (ii) whether they got “c-command detection” with Coref (see Chapter 6), indicating “non-exceptional” Y (that is Y that is not structurally insensitive/subject to NFS effects) and (iii) whether they got c-command detection with DR (again see Chapter 6), indicating “non-exceptional” X (indicating the same as with Y). Color indicated what the respondent’s pattern of judgments showed: (green) accepting the instance of the BVA ok-schema and rejecting the instance of the BVA ✶-schema,

Predicted Correlations of Judgments in English  

 337

tion per schema, for a total of ten questions, three of which were not used in the analysis), as well as having only one choice of X, ‘every teacher’. Plesniak (2021) did indeed find that Hoji’s Japanese results replicated in English, but because that experiment was much more limited than Hoji’s, it is reasonable to ask whether its results were simply “lucky”. The results found here, however, suggest that this was not the case. Despite using three times as many BVA items (doubling the number of instantiations of ✶- and ok-schemata and adding a second ok schemata, though the latter items did not end up factoring into the results8), six times as many DR and Coref items (quadrupling the number of ✶- and ok-schema instantiations, and again, adding the second ok schema as with BVA), and also including far more attentiveness/comprehension-checking sub-experiments (ten vs. one/ four previously), qualitative results are essentially identical in this experiment, both to Plesniak (2021) and to Hoji (Chapter 6 of this volume). Space in this chapter does not permit discussion of all the items individually, but readers may consult Chapter 8 of this volume for detailed discussion of these items, as well as and the general considerations behind their design and presentation. One point is worth mentioning about the ok-schemata used here, which are somewhat different from the earlier Hoji/Plesniak ok-schemata. Specifically, they do not control for precedence, i.e., manipulate the ok-schemata such that the order of X and Y was consistent with the order of X and Y in the ✶-schemata. This can be seen in the ok-schema in the previous section where the order of X and Y in the ok-schemata was reversed relative to the ✶-schemata (‘every boy loves his mother’ vs. ‘his mother loves every boy’); this contrasts with both Hoji’s experiments and Plesniak’s (2021) where the objects of the ok-schemata were scrambled/topicalized in order to keep the X-Y order consistent between ok- and ✶-schemata. In this experiment, in contrast to the earlier Hoji/Plesniak experiments, what is controlled for is thematic role, rather than precedence. Both this experiment and its predecessors use the ‘weak crossover’ schemata, (10), as their ✶-schema. The previous experiments created their ok-schema by “transforming” the sentence type in (10) to the sentence type in (11), which preserves the order of X and Y but changes both the structure so that X c-commands Y and which theta-roles X and Y receive. This experiment, on the other hand, uses passive sentences, as

(yellow) rejecting both, skipping the question, indicating they thought one of the sentences malformed, etc., and (red) accepting the ✶-schema instantiation. We can see that, though there are red dots in all other locations, in the center, when all three criteria are met, there are seven green dots, three yellow dots, and no red dots, precisely the sort of result Hoji obtains in Chapter 6. 8 These were, following Hoji, versions of the ✶-schema sentences that were paired with non-MR interpretations, e.g., ‘his student spoke to every teacher’ with ‘his’ referring to a specific individual. See Chapter 8 for further discussion.

338 

 Daniel Plesniak

in (12), keeping the theta roles constant, but changing both the c-commands relation between X and Y and the linear order of X and Y. See the examples below. (10)

schema Y’s Noun Verb X e.g., “his student praised every teacher” ✶

(11) ok-schema (precedence-controlled) Y’s Noun, X Verb e.g., “his student, every teacher praised” (12) ok-schema (theta-role controlled) X was Verb by Y’s noun e.g., “every teacher was praised by his student” The switch from ok-schema sentences in the form of (11) to those in the style of (12) means that this experiment does not precisely replicate the ok-schematic aspects of Plesniak (2021) or Hoji (this volume), but makes a new contribution in terms of demonstrating that BVA(X, Y) rejection in sentences like (10) cannot simply be reduced to X not receiving an agentive theta-role.9 Sentences of the type in (11) are in fact dealt with in Experiment 2 in the next section. However, as ok-schema predictions are not inherently testable, and essentially only serve to increase the significance of the ✶-schema results,10 this change does not fundamentally affect the question of whether Hoji’s (Chapter 6) qualitative results will replicate in English. To be clear, this is not to say that a lack of control of precedence is not a limitation of the experiment; it is indeed one, though by the same token, a lack of control of theta-roles can be said to be a shortcoming of previous experiments. Both have various potential downsides. Both types of experiments taken together, however, gives us replication of a basic prediction across two different domains; whether one controls for either precedence OR theta-roles, the same qualitative sort of result obtains after all analysis is done. Namely, when all relevant non-structural factors are controlled for, there are no individuals who accept BVA(X, Y) in sentences where X does not c-command Y and there are indi9 Hoji does include the so called “Deep OS” schema in the Japanese experiments, wherein the object is hypothesized to be base-generated in a high position, such that it c-commands the subject, rather than merely being displaced to the front of the sentence. This does allow for investigation of theta-role-related issues, though in a somewhat different way than what is presented here. 10 See discussion in Mukai’s Chapter 3 of this volume, Section 2.3 (which is, as it notes in its initial footnote, is based on Hoji 2015, Chapter 2, Section 2.4).

Predicted Correlations of Judgments in English  

 339

viduals who reject BVA(X, Y) in sentences where X does not c-command Y, but do accept BVA(X, Y) in sentences where X does c-command Y, and this obtains regardless of whether we choose to match the latter sentences with the former in terms of precedence or theta-roles.11 The relevant BVA-DR-Coref paradigm for ok- and ✶-schemata for this “thetarole-controlled” version are as exemplified below: (13) BVA ok Every teacher was spoken to by his student. With BVA(every teacher, his) (14) BVA ✶ His student spoke to every teacher With BVA(every teacher, his) (15) DR ok Every teacher was spoken to by three students. With DR(every teacher, three students) (16) DR ✶ Three students spoke to every teacher With DR(every teacher, three students) (17) Coref ok John was spoken to by his student. With Coref(John, his) (18) Coref ✶ His student spoke to John. Coref(John, his) To understand the experimental results regarding (13)–(18) to be shown below, consider three groups of individuals. The first are those who accept BVA(X, Y) in weak crossover constructions (X not c-commanding Y), such as in (14). Given that the c-command (and precedence) relationships required to facilitate such a reading

11 See also Plesniak 2022, where both theta-role and precedence are controlled for, both separately and together. That is, not only are sentences like (10)–(12) considered, but also sentences like “by his student, every teacher was praised”. The result is again qualitatively the same.

340 

 Daniel Plesniak

are not present in weak crossover constructions, our hypotheses force us to assume that such an individual must have made use of some structural insensitivity of X or Y (or both) in order to effect it. Another type of individual is one who never accepts BVA(X, Y) in weak crossover constructions. Such an individual, it would seem, is not making use of any sort of structural insensitivity, and thus cannot accept the reading, given the lack of c-command or linear precedence. However, as Hoji notes, we must be a little cautious in classifying such an individual, as it is possible that a given participant’s rejection of BVA in weak crossover is not due to the syntactic structure of the sentences at all, but rather, some conceptual difficulties with BVA readings in general, or at least the way they are specified in the instructions the participant received. The rejection of BVA by a person with such confounding issue is not particularly revealing about whether or not they have a “weak crossover effect”, given that such individuals will always reject BVA(X, Y) regardless of the sentence it is matched with. To check whether this is indeed the case, we can examine the individual’s judgment of the BVA ok-schema sentences, such as (13). If informants consistently reject these sentences too, then it would seem they are simply rejecting BVA(X, Y) out of hand, regardless of the structure of the sentence in question; such individuals will constitute the second of the three groups of individuals we will be considering. If, however, an individual accepts (at least some of) these sentences, it shows that that person does allow BVA(X, Y) in other contexts. That person’s rejection of BVA in weak crossover is therefore something about the specific pairing between weak crossover constructions and BVA, not just BVA itself. These are the individuals who constitute the third group of individuals we will be consider, which are those who truly behave consistently with a Reinhartian theory of BVA, i.e., BVA(X, Y) only when X c-commands Y. If we focus exclusively on those participants who show no signs of structural insensitivity (i.e., those whose DR(X, Y’) and Coref(X’, Y) judgments on analogous sentences follow strictly c-command-based patterns), then there ought to be no individuals belonging to the first group described above. That is, no individual who does not show signs of structural insensitivity in either X or Y should accept BVA(X, Y) in weak crossover constructions; rather, any such person should reject BVA in weak crossover constructions, and ideally such an individual will accept BVA in sentences (like (13)) where X c-commands Y, though this second part is not strictly predicted. Such a result is indeed what we find. Consider three further (overlapping) categories of participants, categorized independently of their BVA judgments. Category 1 is attentive/comprehending participants, which we can identify via subexperiments. (The relevant sub-experiments are reviewed in Chapter 8). Those who did not “fail” any of these sub-experiments i.e., those whose answers are consistent with attentiveness and understanding of the instructions, are deemed

Predicted Correlations of Judgments in English  

 341

“attentive” here. Category 2 is those who showed no signs of structural insensitivity on X, as diagnosed by DR(X, Y’). That is, they consistently rejected DR(X, Y’) in weak crossover-style sentences, such as (16), but accepted it at least sometimes in the ok-schema sentences, such as (15). The importance of rejecting DR(X, Y’) in weak crossover sentences should be clear, as accepting such an interpretation would signify that, for the individual in question, X can participate in MR(X, Y) when it does not c-command Y. The reason for needing to accept the ok-schema is a bit subtler. As with BVA, we have to be a bit careful here when measuring “rejection” in the ✶-schema cases; an individual might simply reject DR-type interpretations in all sentences, regardless of the sentence it is paired with. As noted above, such an individual’s rejections do not really tell us much, as they may signify that the individual has difficulty either with conceptualizing the intended reading or perhaps simply with understanding the directions. As such, we cannot really tell from their judgments whether they have structurally (in)sensitive X’s in weak crossover constructions or not. We focus, therefore, on participants who accept DR readings a majority of the time such readings are presented along with sentences where X c-commands Y’ (the ok-schema instantiations, which in this case are passives, e.g., ‘every teacher was spoken to by three students’ in (15)), but consistently reject such readings when they appear alongside a weak crossover construction, as in (16). Such individuals seem to be clearly indicating that they do accept DR(X, Y’), but only when X c-commands Y’; the rejections of the ✶-schema must be specifically because of some aspect of the ✶-schema, and not a general rejection of DR. A similar procedure is followed with coreference, giving us Category 3, which consists of those who clearly accept Coref(X’, Y), but do not accept it in weak crossover constructions, meaning that Y is also structurally sensitive for them (e.g., accepting (17) but not (18)). Our crucial prediction is that anyone who belongs to all three of these categories ought to be someone who rejects BVA(X, Y) in weak crossover. This is because such a person is attentive, does not have structurally insensitive X’s or Y’s, and the weak crossover constructions are such that X does not precede or c-command Y in them, so our hypotheses lead us to conclude that there is nothing that can enable a BVA(X, Y) reading for such an individual on such a sentence (and thus they must reject it). The experimental results of this classification are shown in (19), bearing out this prediction. The three circles represent each of the three diagnosed properties just discussed: attentiveness, DR determining that X is not structurally insensitive, and Coref determining that Y is not structurally insensitive. If an individual has all three such properties, it indicates that individual’s judgments on BVA(X, Y) have bearing on our c-command based predictions. Each shaped dot represents an individual; if that individual is inside a circle, they have the given property, e.g., being attentive. A dot’s colors (and form) correspond to

342 

 Daniel Plesniak

the relevant individual’s judgments regarding BVA(X, Y). Red squares represent those who accept BVA in weak crossover constructions (at least sometimes), green circles are those who always reject BVA specifically in weak crossover constructions, and yellow triangles are those who always reject BVA(X, Y) regardless of whether the corresponding sentence is in a weak crossover configuration: (19)

a. Venn Diagram

Weak Crossover

Rejects BVA (WC Specifically) Rejects BVA (Always)

Attentiveness

Accepts BVA In WC

Coref

DR

b.

Venn Diagram information

Weak Crossover

Green

Yellow

Red

All Three

8

1

0

9

Attentive, No Insensitive X

2

1

1

4

Attentive, No Insensitive Y

17

0

5

22

1

0

1

2 54

No Insensitive X, No Insensitive Y Attentive Only

Total

31

3

20

No Insensitive X Only

0

0

1

1

No Insensitive Y Only

7

2

11

20

None of the Three

29

6

54

89

Total

95

13

93

201

Predicted Correlations of Judgments in English  

 343

Crucially, in the central intersection, there are no red squares, while there are several (eight) green circles. That means that all individuals who were identified as attentive and for whom neither X nor Y were structurally insensitive in weak crossover constructions, based on DR and Coref tests, reject BVA(X, Y) in weak crossover constructions, exactly as predicted. Such a result suggests that those who do accept BVA(X, Y) in weak crossover constructions only do so by making use of structural insensitivity of X or Y. Note that, while we are indeed focusing on those in the central intersection, we are not discarding anyone’s judgments. We can make various analytically meaningful statements about those not in the center; for example, for anyone represented by a red dot, i.e., one who accepts weak crossover BVA, we can now give a both plausible and non-circular account of why they accept it. Consider, for example, the red square in the intersection of “Attentiveness” and “DR”; because they are represented by a red square, we know that that individual accepted weak crossover BVA at least once. Because they are inside the “Attentiveness” circle, we also know that that person was attentive, so the judgments he or she indicated most likely do represent his or her judgments on DR/Coref/BVA accurately. Finally, because they are in the “DR circle, we also know that this individual showed no signs of a structurally insensitive X, so that cannot be the source of BVA(X, Y) acceptance in weak crossover. However, this individual is outside of the “Coref” circle, meaning that he or she did accept coreference in a weak crossover configuration, allowing for the possibility of a structurally insensitive Y (‘his’). We can thus conclude that this structurally insensitive ‘his’ facilitated the acceptance of BVA in weak crossover constructions for this individual. Such results are exactly in keeping with the results found by Hoji for Japanese. The Japanese results thus do indeed replicate in a reliable and straightforward manner in English, lending support to Hoji’s main hypotheses about structure and BVA, as well as to the viability of the correlational methodology for probing structure in languages other than Japanese.

4 Experiment 2: Absence of Correlations with Bad Hypotheses Though experiments like the previous one are argued to provide evidence for the c-command-based hypotheses that they purport to test, it is not obvious this is true. For example, an important aspect of the argumentation in both the preceding section, as well as in Hoji (Chapter 6) is that if, in the central intersection, there are some “green circles” (individuals for whom BVA(X, Y) demonstrably

344 

 Daniel Plesniak

requires X to c-command Y), and no “red squares” (individuals for whom BVA(X, Y) does not require X to c-command Y), then this is evidence that, once noise control is applied, all individuals conform to c-command-based predictions. However, any scientific test is only useful as such if it distinguishes correct hypotheses from incorrect ones; for the correlational methodology in particular, the “some greens, no reds in the center” result ought to generally not occur if a prediction based on incorrect hypotheses is tested. If this is not the case, and the “some greens, no reds in the center” result obtains regardless of whether correct or incorrect hypotheses are tested, then there is essentially no significance to that result. To address this issue directly, an experiment was conducted that was similar to Experiment 1 in essentially every way, except with a crucial difference: instead of the ✶-schema being “weak crossover”, it was one involving the Object-SubjectVerb (OSV) word order, i.e.: (20) Y’s Noun, X Verb E.g., “his student, every teacher praised” As discussed above and in Hoji (Chapters 5 and 6), there is reasonably strong evidence that such a sentence can correspond to a “reconstructed” structure where the object, Y’s Noun is structurally represented as if it were a “regular” object, i.e., c-commanded by the subject, X. If so, then X does c-command Y, so BVA(X, Y) should be in principle acceptable, even for individuals using X and Y that are structurally sensitive for them. Let us suppose though that we did not know this, and erroneously believe that in such sentences, X never c-commands Y. Then we expect the sentences to behave like weak-crossover sentences; once proper noise control is applied, no one should accept BVA(X, Y) when paired with such a sentence. We can then “contrast” this with the canonical Subject-Verb-Object (SVO) version of such sentences as instantiations of an ok-schema where we do hypothesize X to c-command Y: (21) X Verb Y’s Noun E.g. “Every teacher praised his student” The question is then whether there are individuals who accept sentences like (21) but consistently reject their “reconstructed” counterparts like (20), and whether the same sort of noise control can pick out such individuals in the same way as was done in Experiment 1. If so, then we have successfully “tricked” the methodology into giving us a false positive; it diagnoses that X does not c-command Y in (20), when in fact it does (or at least can).

Predicted Correlations of Judgments in English  

 345

The full “paradigm”, exemplified across all three MR’s, is as below: (22)

BVA ok Every teacher spoke to his student. With BVA(every teacher, his)

(23) BVA ✶ To his student, every teacher spoke. With BVA(every teacher, his) (24) DR ok Every teacher spoke to three students. With DR(every teacher, three students) (25) DR ✶ To three students, every teacher spoke. With DR(every teacher, three students) (26) Coref ok John spoke to his student. With Coref(John, his) (27) Coref ✶ To his student, John spoke. Coref(John, his) As with the previous experiment, we divide the participants according to their judgments on BVA, this time with regards to whether it is: (a) specifically unavailable under reconstruction, as in sentences corresponding to (20), but available in sentences corresponding to (21), (b) always unavailable regardless of sentence type, or (c) available under reconstruction in at least some corresponding to (20). We also classify participants as to whether they are: (i) attentive, (ii) show no signs of having a structurally insensitive X, this time diagnosed by their rejecting reconstruction-based DR (while accepting its SVO counterpart), and likewise (iii) show no signs of having a structurally insensitive Y, diagnosed by their rejection of reconstruction-based Coref (and their acceptance of its SVO counterpart). In a counterfactual world where our ✶-schema hypothesis was correct, we would expect the results for this experiment to look qualitatively like the previous ones. Namely, if an individual possesses the three key properties (attentiveness, structurally sensitive X, and structurally sensitive Y), they should reject recon-

346 

 Daniel Plesniak

struction BVA. Further, at least some subset of these individuals should accept BVA in the ok-schema, i.e., SVO, where X is S and Y is in O (as in (21)). Since we in fact believe that the hypothesis is wrong, and OSV is not in fact a ✶-schema, we hope that such a result does not obtain. As it turns out, it indeed does not: (28)

a. Venn diagram

Reconstruction

Rejects BVA (RC Specifically) Rejects BVA (Always)

Attentivenes s

Accepts BVA In RC

Coref

DR

z

b. Venn diagram information Reconstruction All Three

Green

Yellow

Red

Total

0

0

0

0

Attentive, No Insensitive X

6

1

4

11

Attentive, No Insensitive Y

0

0

0

0

2

0

0

2

15

5

69

89

No Insensitive X, No Insensitive Y Attentive Only No Insensitive X Only

3

0

3

6

No Insensitive Y Only

0

0

1

1

None of the Three

11

1

79

91

Total

37

7

156

200

Predicted Correlations of Judgments in English  

 347

In fact, there is absolutely no one in the central intersection whatsoever. This is because no one in the entire set of 200 participants was such that they were attentive and consistently rejected reconstruction coreference, suggesting that essentially no individuals have Y’s that consistently resist coreference in OSV but not in SVO (with X being S and Y being in O in both cases). A few did consistently reject reconstruction DR, suggesting some individuals did have X’s which permitted DR(X, Y’) when X was S of SVO but not when X was S of OSV. However, some of those individuals did allow BVA(X, Y) under reconstruction, so there was no qualitative correlation between DR and BVA alone. Given the differences between this diagram and the previous one, we can see that the overall pattern of judgments contrasts both qualitatively and quantitatively with what was found for weak crossover constructions in the previous experiment,12 as is consistent with hypothesis that OSV reconstruction sentences are not, in fact, instantiations of a ✶-schema. In short, Hoji’s correlational methodology fails to give us a “supporting” result when our hypotheses are bad, suggesting that when it does give us such a result, this is not merely a fluke of the methodology itself, but reflects positively on the correctness of the hypotheses we are testing with it.13

5 Experiment 3: Checking New Constructions So far, the results of the weak crossover and reconstruction experiments have concurred with what has been commonly found in the theoretical and experimental literature, though with an added level of detail in terms of participant classification, which allows us to account for all apparent “exceptions” to the predicted judgment as per the Hoji’s correlational methodology. However, a new 12 One limitation here is that we cannot compare these individual’s responses on this experiment to their behavior on an experiment where OSV served, correctly, as an ok-schema. As can be seen from Hoji’s (this volume) data (see next footnote) as well as the data from Plesniak 2022, when such a comparison is possible, things come out “as expected”, i.e., the experiment(s) based on incorrect hypotheses does not give support to those hypotheses, while the experiment(s) based on correct hypotheses does support them. 13 Space here does not permit a full discussion, but this experiment can effectively be repeated with Hoji’s experimental data as well. If we do so, what obtains is essentially the same. For most choices of X and Y, there is no individual in the center if the OSV sentences are treated as instantiations of a ✶-schema. In a very small number of cases, one individual does make it into the center, but those individuals are in fact “red dots”, meaning that they accept OSV BVA. Thus, it seems true that the correlational methodology consistently fails to support the incorrect hypothesis that OSV cannot correspond to a structure where S c-commands O, and sometimes seems to overtly falsify it.

348 

 Daniel Plesniak

contribution is made in this section with regard to so-called “spec-binding” sentences, specifically spec-binding in OSV “reconstruction” constructions as discussed in previous sections (I will refer to this simply as “spec-binding” for the sake of brevity). Noted as early as Higginbotham 1980, so-called “spec-binding” constructions allow BVA(X, Y) where there is no obvious c-command, as in (29): (29) Every author’s mother loves his books. Speakers often accept sentences like (29) with BVA(every author, his), even though ‘every author’ is contained with the subject ‘every author’s mother’, and as such, definitionally cannot c-command anything outside of the subject. BVA(every author, his) is thus established without ‘every author’ (apparently) c-commanding ‘his’, but as seen previously, this should not surprise us too much. Historically, Reinhart (1983) was aware of such cases and revised her generalizations to allow for possessors to bind out of subjects. Building on such notions, work like that of Kayne 1994 has tried to redefine the structural status of “specifiers” (for our purposes possessors) and the nature of structural relations such that possessors do actually c-command out of the phrases they are inside of, giving us the notion of “spec(ifier) binding”.14 However, as works such as Barker 2012 catalogue,

14 In particular, Kayne defines c-command not in terms of sisterhood, but in terms of “exclusion”, following May 1985 and Chomsky 1986b. Further, domination is given an added condition that makes reference to “segments”, essentially two nodes sharing the same label, as is said to happen in an adjunction structure, where the result of merging XP with YP, while structurally the set {XP, YP}, is interpreted as if it were simply XP (or YP). In such a case, both XP and {XP, YP} are “segments” of XP. Kayne 1994’s and revised definitions flowing from these conceptions are as follows: (i)

(a)

X dominates Y iff every segment of X dominates Y.

(b)

Exclusion: X excludes Y iff no segment of X dominates Y.

(c)

C-command: X c-commands Y iff X and Y are categories and X excludes Y and every category that dominates X dominates Y.

Kayne further argues that specifiers are structurally adjuncts, and thus that possessive constructions and subjects ought to be analyzed as having the rough representation in (ii) (I am ignoring the details having to do with which functional projections there are or how they are labeled): (ii)

[VP [NP1 [NP2 Every boy’s] [NP1 mother]] [VP loves [NP3 him]]]

No segment of NP2 dominates NP3, under either the traditional definition of domination or the revised one, thus NP2 excludes NP3. Under the revised definition of domination (but not the traditional one), neither VP nor NP1 dominate NP2, as there are segments of both VP and NP1 that do not dominate NP2, even though there are other segments of VP and NP1 that do. As such, NP2 is not dominated by anything within VP. If we embed VP within a larger constituent, then that constituent will dominate NP2, but it will also dominate NP3. As such, any category that dominates NP2 also dominates

Predicted Correlations of Judgments in English  

 349

such attempts have repeatedly failed to capture the range of possible exceptions. For example, Kayne’s approach does not explain why “inverse linking” constructions (May 1985) also allow BVA(X, Y) without X c-commanding Y, as in (10): (30) Someone from every city hates it. ‘Every city’ in (30) is not a possessor but is in fact inside of a prepositional phrase inside of the subject, and yet (30) is uncontroversially accepted with BVA(every city, it). Barker 2012 reviews further attempts to accommodate these facts, such as Hornstein’s (1995) “almost c-command” and shows that these too fail to account for the general pattern that BVA(X, Y) when X is inside of a subject (or indeed elsewhere) and Y is (within) a following element is often possible, regardless of c-command. Before we descend down this rabbit hole of continuing exceptions, we should pause to consider the status of the initial problem, i.e., sentences like (29). In the case of spec-binding, such constructions typically have a feature that weak crossover constructions lack, namely, word-order-based linear precedence between binder and bindee. As mentioned briefly in the introduction, precedence-based theories of BVA have a long history. For example, Chomsky (1976) initially proposed what has been called the “leftness condition” (Higginbotham 1980), namely: (31) A variable cannot be an antecedent of a pronoun on its left. Essentially, this prohibits BVA(X, Y) if X does not precede Y. Such a condition is not correct as written, however, as shown by examples like (25): (32) A: Which of his parents do you think every boy loves the most? B: Probably his mother. (32)’s A is frequently accepted allowing a reading where ‘his’ is bound by ‘every boy’, despite ‘his’ preceding ‘every boy’. This makes sense under a purely c-command-based approach, as the underlying structure is such that ‘which of his parents’ is an object in the embedded clause ‘every boy loves which of his parents the most’ and is thus c-commanded by ‘every boy’ under reconstruction. This is precisely the motivation for reconstruction effects; ‘his’ in (32) behaves “as if” ‘his mother’ were the object of ‘loves’, despite not occupying the canonical object position in the sentence-string. Object-hood, at least for the purposes of BVA, is there-

NP3, and thus both conditions for c-command are met, and NP2 c-commands NP3. This allows Kayne to claim that possessors in subjects of a given clause actually do c-command objects in that clause.

350 

 Daniel Plesniak

fore better understood as a matter of structure, not just word order. Displacement of this sort, as we have seen, is not limited to wh-items. For example: (33) Every teacher praised his student. (34) His student, every teacher praised. As results from Experiment 2 show, BVA(every teacher, his) is possible for many individuals in both sentences like (33) and sentences like (34). As a negative constraint then, the leftness condition is false, but this does not necessarily mean that linear precedence has no effect on BVA. Bruening’s (2014) recent work, for example, on “precede and command”, reviving ideas from Langacker (1969), Jackendoff (1972), and Lasnik (1976), tries to reduce c-command and precedence effects (at least in the domain of coreference) to a similar source, involving phases and left-to-right processing. Ueyama 1998, on the other hand, considers c-command and precedence to be separate factors that can both independently facilitate BVA. Consider, for example, the pair of “donkey anaphora” (Geach 1962) sentences in (35) and (36): (35) Every farmer who owns a donkey beats it. (36) Every farmer who owns it beats a donkey. While (35) seems to clearly permit ‘it’ to vary across donkeys, such a reading is far more difficult, if not impossible, to achieve in (36). In neither case is there clear c-command of either binder or bindee. As such, given that precedence does appear to be able to enable BVA independently of c-command, it is crucial to assess exceptions to c-command-based predictions only when such a confound is absent. If spec-binding really is due to possessors being able to c-command out of subjects, then we expect that topicalizing the object should not block BVA, just as it does not block BVA in simple OSV topicalization (the reconstruction cases in the previous section). That is, as with “regular” OS sentences like (37), sentences like (38) (the topicalized counterpart of (29)), should be acceptable with BVA(every author, his), even after controls for structurally insensitive X and Y are applied: (37) His books, every author loves. (38) His books, every author’s mother loves. Put another way, if possessors really did c-command “out of” their containing subjects, there should be no qualitative contrast in patterning in terms of acceptability

Predicted Correlations of Judgments in English  

 351

between the “regular” OS type and the “spec-binding” OS type, as exemplified by (37) and (38). As it turns out, however, once the usual factors are indeed controlled for, spec-binding can indeed be shown to be a ✶-schema, unlike the “regular OS” case. That is, when the X of BVA(X, Y) is a not subject, but rather merely a possessor in the subject, BVA(X, Y) can only be based on non-structural factors and/or linear precedence. “Reconstruction effects” like those observed for simple OSV do not occur, suggesting that possessors do not in fact have the same c-command domain as the phrases they are embedded in. Reinhart, Kayne, and others were therefore overly quick to accept these apparent counterexamples and revise their hypotheses around them; the potential confound of word order effects via linear-precedence was overlooked, as were the potential idiosyncratic effects of X and Y. Such a result represents a victory for the orthodox view of c-command, corrects the longstanding misconception of spec-binding in the theoretical literature, and once again shows that multiple factors play a role in facilitating BVA readings. It further emphasizes that we can successfully control for such factors via the correlational methodology, even in constructions other than the ones with which it was initially developed. Experiment 3 thus repeats the same sort of experiment as the previous ones, but uses sentences like the ones below (where the ok-examples are SVO sentences with X as S and Y in O, similar to ones we have seen before, but designed to semantically “match” the spec-binding sentences): (39) BVA ok Every teacher has a colleague who spoke to his student. With BVA(every teacher, his) (40) BVA ✶ To his student, every teacher’s colleague spoke. With BVA(every teacher, his) (41) DR ok Every teacher has a colleague who spoke to three students. With DR(every teacher, three students) (42) DR ✶ To three students, every teacher’s colleague spoke. With DR(every teacher, three students) (43) Coref ok John has a colleague who spoke to his student. With Coref(John, his)

352 

 Daniel Plesniak

(44) Coref ✶ To his student, John’s colleague spoke. Coref(John, his) With such sentences, we find the same sort of distribution as predicted given the hypothesis that the ✶-schema sentences can only be accepted due to structural insensitivities of X or Y. As is shown below, the results of this spec-binding experiment pattern with those of weak crossover experiment (Experiment 1), rather than those of Experiment 2. This result supports the hypothesis that possessors do not in fact c-command out of the phrases that contain them: (45) a. Venn diagram

Spec-Binding

Rejects BVA (SB Specifically) Rejects BVA (Always)

Attentiveness

Accepts BVA In SB

Coref

DR

b. Venn diagram info Spec-Binding All Three

Green

Yellow

Red

Total

4

0

0

4

Attentive, No Insensitive X

2

0

6

8

Attentive, No Insensitive Y

0

0

1

1

4

0

0

4

17

5

45

67

No Insensitive X, No Insensitive Y Attentive Only

Predicted Correlations of Judgments in English  

Spec-Binding

Green

Yellow

Red

Total 10

No Insensitive X Only

3

1

6

No Insensitive Y Only

2

0

2

4

None of the Three

23

2

77

102

Total

55

8

137

200

 353

In the central intersection, there are four green circles and no red squares, suggesting that all cases of individuals accepting spec-binding are due to either inattentiveness/confusion or structurally insensitive X/Y. Stated another way, there is no attentive participant who accepts BVA(X, Y) in spec-binding constructions and who does not display evidence that they either: (i) have a structurally insensitive X, as diagnosed by DR, and/or (ii) have a structurally insensitive Y, as diagnosed by Coref. Unlike with reconstruction in Experiment 2, there are individuals who reject both DR and Coref in spec-binding constructions, and all individuals who do so also reject spec-binding BVA. As such, the evidence points to spec-binding BVA having a similar status to weak crossover BVA: acceptable to certain individuals under certain, clearly diagnosable conditions, namely the individual having structurally insensitive X and/or Y, but otherwise, spec-binding BVA is not acceptable. As such, there is no need to revise the definition of c-command or the relationship between c-command and BVA to account for such cases. Spec-binding thus presents a case where non-c-command-based noise was historically mistaken for the effects of c-command, leading researchers to pursue various accounts of such “false effects of c-command”, as briefly reviewed above; this emphasizes the importance of using tools like the correlational methodology to diagnose what the source of a given BVA judgment is. We thus see that this methodology not only can, but indeed should, be used on constructions other than the few on which it was initially developed, if we want to be able to use the (un)acceptability of BVA readings as a reliable indicator of c-command relations.

6 Conclusion The experiments reported here demonstrate several key aspects of both the CS and the correlational methodology employed to study it. First and most basically, we have seen in Experiment 1 that the Japanese results discussed in Chapter 6 replicate in a near-identical manner in English. That such replication occurs lends credence to the notion of cross-linguistically universal syntactic properties, such as subjects asymmetrically c-commanding objects in the types of simple SOV/SVO sentences examined. While speculation and argumentation to this effect is not

354 

 Daniel Plesniak

new, this, along with Plesniak (2021), is one of the first times universality of this type on the level of the individual has been shown experimentally. That this is possible provides a far more direct argument for a universal and cross-linguistically constant CS that is instantiated literally in each individual, as maintained by the basic hypotheses of generative grammar. Moreover, these experiments have shown that the basic correlational methodology used to make this demonstration is highly sensitive to the hypotheses that underlie a given experiment; as demonstrated by Experiment 2 in particular, incorrect hypotheses about the underlying c-command relations fail to be supported by experimental results derived via this methodology. It is thus fair to say that “successful” results as discussed, including Hoji’s (in Chapter 6), those in Plesniak (2021), and Experiment 1 are not simply due to some inherent “trick” of the design of the experiments but do in fact reflect the presence or absence of the hypothesized c-command relations. (See also discussions in Chapters 5 and 6 of this volume of failures of non-predicted entailments to hold.) While some may have considered the possibility of such a “trick” far-fetched, it is nevertheless useful to have provided evidence that such concerns can be laid to rest, so that further research can build on our initial results in ways that (continue to) ensure rigorous testability. Finally, as shown in Experiment 3, we can expand this methodology beyond the types of sentences on the basis of which it was first developed. The correlational methodology is thus not merely some sort of “weak crossover” detector; all indications at this point are that it is widely applicable to a variety of constructions where c-command based hypotheses can be formulated. By doing this, we can put to the test various nuanced hypotheses not only as to the CS representation of the structures of various constructions, but also thereby the fundamental properties of CS relations like c-command. An example of this was seen in Experiment 3, the results of which are consistent with the hypothesis that possessor does not c-command outside of subjects, which helps to resolve a longstanding debate as to the nature of c-command. Continued application of this form of inquiry to other such longstanding debates in the syntactic literature promises to usher in a new paradigm for structural analysis, wherein variation in judgments is addressed by reference to judgment correlations.

References Barker, Chris. 2012. Quantificational binding does not require c-command. Linguistic Inquiry 43(4). 614–633. Bruening, Benjamin. 2014. Precede-and-command revisited. Language 90. 342–388.

Predicted Correlations of Judgments in English  

 355

Chomsky, Noam. 1976. Conditions on rules of grammar. Linguistic Analysis 2. 303–351. Chomsky, Noam. 1986a. Knowledge of language: Its nature, origin and use. Westport, CT: Praeger. Chomsky, Noam. 1986b. Barriers. Cambridge, MA: MIT Press. Chomsky, Noam. 1995. The minimalist program. Cambridge, MA: MIT Press. Geach, Peter. 1962. Reference and generality: An examination of some medieval and modern theories. Ithaca, NY: Cornell University Press. Higginbotham, James. 1980. Anaphora and GB: Some preliminary remarks. In J. Jensen (ed.), Proceedings of NELS 10, 223–236. Ottawa: Cahiers Linguistiques d’Ottawa, University of Ottawa. Hoji, Hajime. 2003. Falsifiability and repeatability in generative grammar: A case study of anaphora and scope dependency in Japanese. Lingua 113. 377–446. Hoji, Hajime. 2015. Language faculty science. Cambridge: Cambridge University Press. Hoji. Hajime. 2017. Galileo’s other challenge. Inference Review 3(2). https://inference-review. com/letter/galileos-other-challenge (accessed 21 February 2022) Hoji, Hajime. This volume: Chapter 4. The key tenets of language faculty science. In Hajime Hoji, Yukinori Takubo and Daniel Plesniak (eds.), The theory and practice of language faculty science. Berlin and Boston: De Gruyter Mouton. Hoji, Hajime. This volume: Chapter 5. Detection of c-command effects. In Hajime Hoji, Yukinori Takubo and Daniel Plesniak (eds.), The theory and practice of language faculty science. Berlin and Boston: De Gruyter Mouton. Hoji, Hajime. This volume: Chapter 6. Replication: Predicted correlations of judgments in Japanese. In Hajime Hoji, Yukinori Takubo and Daniel Plesniak (eds.), The theory and practice of language faculty science. Berlin and Boston: De Gruyter Mouton. Hornstein, Norbert. 1995. Logical form: From GB to minimalism. Oxford: Blackwell. Jackendoff, Ray. 1972. Semantic interpretation in generative structure. Cambridge, MA: MIT Press. Kayne, Richard. 1994. The antisymmetry of syntax. Cambridge, MA: MIT Press. Langacker, Ronald. 1969. On pronominalization and the chain of command. In David Reibel and Sanford Schane (eds.), Modern studies in English, 160–186. Englewood Cliffs, NJ: Prentice-Hall. Lasnik, Howard. 1976. Remarks on coreference. Linguistic Analysis 2. 1–22. Lasnik, Howard. 1990. Syntax. In Daniel N. Osherson and Howard Lasnik (eds.), Language (An invitation to cognitive science, Volume 1), 5–21. Cambridge, MA: A Bradford Book/MIT Press. May, Robert. 1985. Logical form: Its structure and derivation. Cambridge, MA: MIT Press. Plesniak, Daniel. 2022. Towards a correlational law of language: Three factors constraining judgment variation. Los Angeles, CA: University of Southern California dissertation. Plesniak, Daniel. 2021. Experiments beyond statistics: How generative linguistics can beat the replication crisis. Ms, available at: https://www.researchgate.net/ publication/357063705_Experiments_Beyond_Statistics_How_Generative_Linguistics_ Can_Beat_the_Replication_Crisis (accessed 21 March 2022) Plesniak, Daniel. This volume: Chapter 8. Implementing experiments on the language faculty. In Hajime Hoji, Yukinori Takubo and Daniel Plesniak (eds.), The theory and practice of language faculty science. Berlin and Boston: De Gruyter Mouton. Postal, Paul. 1971. Cross-over phenomena. New York: Holt, Rinehart and Winston.

356 

 Daniel Plesniak

Reinhart, Tanya. 1983. Anaphora and semantic interpretation. Chicago: University of Chicago Press. Ueyama, Ayumi. 1998. Two types of dependency. Los Angeles, CA: University of Southern California dissertation. Ueyama, Ayumi. 2010. Model of judgment making and hypotheses in generative grammar. In Shoichi Iwasaki, Hajime Hoji, Patricia Clancy and Sung-Ock Sohn (eds.), Japanese/Korean Linguistics 17, 27–47. Stanford, CA: CSLI Publications. Ueyama, Ayumi. This volume: Chapter 2. On non-individual denoting so-words. In Hajime Hoji, Yukinori Takubo and Daniel Plesniak (eds.), The theory and practice of language faculty science. Berlin and Boston: De Gruyter Mouton. Wasow, Thomas. 1972. Anaphoric relations in English. Cambridge, MA: MIT dissertation.

Daniel Plesniak

Implementing Experiments on the Language Faculty 1 Introduction This chapter serves to illustrate the kinds of decisions one faces when trying to implement the type of correlational experiment described by Hoji and Plesniak in Chapters (4–7) in this volume and gives practical advice as to how to address them. In order to give readers a tangible example of how to do so, I walk through the precise implementation of the English-based “weak crossover experiment” from Chapter 7 and compare and contrast the implementational decisions made there with those made in other similar experiments. These examples are coupled with general design considerations, which should be applicable to many different styles of implementation for experiments of this sort. Some concerns addressed are, just to name a few examples, how to elicit judgments, how similar or different the sentences used in the experiments should be from one another, and how to decide how many instantiations of each schema to use. Special attention is paid to what Hoji (2015 and also Chapter 6 in this volume) terms “sub-experiments” and the various ways one might choose to implement them. Additionally, I discuss the ways in which the interpretation of the results of a given experiment depends partially on the implementational choices made. This in turn leads us to consider how to deal with the consequences of certain implementational choices, which I again address using the specific example of the weak crossover experiment. By clearly elucidating the challenges one faces, as well as giving practical examples demonstrating at least some ways of overcoming them, this chapter serves as a “how to” manual for experimenters desiring to utilize the correlational methodology articulated by Hoji in Chapters 4–6 of this volume. In addressing these issues, we will come to see several of the different strategies for implementing this methodology that have been utilized across different experiments. As such, this chapter can be of use both to those wanting to implement an analogous experiment and to those wishing to understand the particular ways in which the experiments discussed in this volume (and other related experiments) have been performed. In Section 2, I will first focus primarily on how the items to be judged are to be presented to participants. Given that experiments of this sort involve asking participants whether a given sentence can have a certain interpretation, specifyhttps://doi.org/10.1515/9783110724790-008

358 

 Daniel Plesniak

ing the relevant interpretation(s) in an effective manner is crucial. In Section 3, I explore how the participant’s response can be elicited, both in terms of the literal form responses are to take (e.g., “yes”-“no” options), as well as how elicitation of one judgment is juxtaposed with elicitations of others. In Section 4, I go beyond these more narrow concerns to examine broader issues of experimental design, involving what sorts of items should be included and how they should be structured relative to one other. In Section 5, I conclude by discussing high-level decisions regarding certain tradeoffs experiment designers have to make and offer advice to aid in navigating that choice. Before continuing, I will briefly make a point of comparison and contrast. A key notion to keep in mind in this chapter is that the design of any experiment should reflect what that experiment is designed to test, which in turn reflects what the experimenter is trying to discover and/or prove something about. Many experiments in linguistics dealing with acceptability judgments are concerned with finding what the judgment of a typical or average speaker is. For example, take Schütze 2016’s advice that “subjects must be sufficient in number in order for the assumptions of the required statistical tests to be met and to avoid distorting the results with atypical speakers.” (emphasis mine, Schütze 2016: 184). Indeed, though there is not space to review them one-byone, throughout Schütze 2016, many suggestions are made to the effect that various aspects of the experiment (e.g., the lexical items used) should be both kept average or typical and be quite numerous and varied, in order that a general picture might emerge via statistical analysis; such analysis will effectively average over the responses of many individuals to the various choices to create said general picture. As emphasized by the quote, this procedure allows one to get a measure of what a typical speaker is likely to think about a typical instance of a given sentence/item type. While such information is valuable for a wide variety of purposes, acquiring knowledge of what is typical is not the goal of the program laid out in this particular volume. Rather, our interest lies in checking definite predictions against the judgments of individuals; as such, we want to test whether our predictions are correct for each individual, and indeed, whether they are true for each judgment rendered by each individual.1 As such, if an individual is atypical and gives 1 See discussion in both Chapters 4 and 7; this is a consequence of our desire to study the universal Computational System (CS) as instantiated in each individual. The natural question is of course whether this might be done via consideration of averages that emerge from the data, rather than specific (correlations of) judgments. To give a brief answer: the primary requirements for any program that seeks to rigorously study the CS is that it make definite predictions deduced from hypotheses about the CS that can be subjected to empirical tests. This volume as a whole

Implementing Experiments on the Language Faculty  

 359

a judgment inconsistent with what the majority of other people would give, or even inconsistent with most other judgments given by that same individual, such an occurrence is potentially highly significant for us; for this particular sort of experiment, we in fact do not want to design things so that such data can be “averaged out”. Rather, we intend our predictions to be sophisticated enough that they survive rigorous testing via all (relevant) judgments of all individuals who participate in our experiments. Thus, our concern is with the range of patterns of judgment manifested in the data we gather, rather than what pattern of judgments is statistically average or modal in that data. In this sense, we can say that our research is “possibility-seeking”, focused on the range of possible patterns of judgments a given individual might have, as opposed to “typicality-seeking” research, which investigates which specific judgments an individual is most likely to have. This is not to say that the two methodologies should produce wildly differing results; we would in fact expect that what judgments are typical relates to what patterns of judgments are possible. Somewhere “down the road”, the careful application of both systems of inquiry may yield convergent results in many cases. However, they will contrast in what their concerns are, and thus

discusses one way to achieve this, but this discussion does not preclude other ways, including ways that might be based on assessments of what is statistically typical. What would be necessary is (a) for the predicted statistical tendencies to be rigorously deduced from hypotheses about the CS, the language faculty, and/or other relevant factors, and (b) for the predicted statistical tendencies to be definite enough so that we can clearly assess whether (or at least to what degree) observed data matches up to these predictions. A discussion of the degree to which this is doable, or indeed, is being done in other experimental work, is beyond the scope of this chapter; it seems likely to me, however, that successfully producing such predictions would require far more sophisticated hypotheses than we have adopted in Chapters 4–7, so as to enable the deductive “leap” from the universal properties of the CS to some sort of behavioral tendencies. Given that the results in Chapters 5–7 suggest that our non-tendency-based predictions are “doing well enough”, I have not seen the need to pursue such an option, but as Language Faculty Science is a research program in its infancy, it seems premature (at least to me) to rule out such an approach. Indeed, if we should ever need to account for something like “rate of human error”, it seems clear that statistical analysis of some sort would have to play a role in doing so. We do not seem to have reached the point where such considerations are crucial yet, however, and there may be ways of circumventing them, at least for the time being.

360 

 Daniel Plesniak

what sorts of implementational choices they make. Many of the pieces of advice for a Schütze 2016-style experiment, designed to aid in averaging through differences, will actually result in dumping a variety of different sources of noise on individual judgments, rendering the correlational style of investigation (as pursued in this volume) unusable. Conversely, many of the design decisions described below may be rightly accused to biasing the average result of a given judgment one way or another, which would render them unusable for many purposes in a “typicality-seeking” experiment; this particular sort of bias is often not an issue for us, as the average judgment on an item never directly factors into our analysis to begin with. Thus, while inspiration can and should be drawn between different modes of inquiry, it should be clarified that this chapter is not intended to serve as a manual for any other type of experiment.

2 Conveying Interpretations In this section, I describe strategies for conveying interpretations, and further, how the effectiveness of the chosen method of conveyance can be checked. Given that the experiments conducted using the correlational methodology have (thus far) primarily made use of correlations of DR, Coref, and BVA (see previous chapters), I will focus on these interpretations, though the general considerations should be relevant for any MR’s used. In the implementation of the English “weak crossover experiment” discussed in Chapter 7, BVA(every teacher, his) was conveyed as below: (1) “His student spoke to every teacher.” ◯ In this sentence, “his student” can refer to each teacher’s own student. ◯ In this sentence, “his student” cannot refer to each teacher’s own student. In essence, the phrase “each teacher’s own” is doing the job of conveying to the participant that ‘his’ is to be interpreted not as referring to a specific individual but as changing in reference relative to each teacher in question. This approximates a BVA reading; in an ideal world, we would be able to teach every participant exactly what we meant by BVA, and then ask them simply whether BVA(every teacher, his) is possible. For reasons of both time and difficulty, this has not been seriously attempted yet in an experiment in this style. Plesniak 2022 did include a “tutorial” section where participants were given various examples of sentences that could or could not have BVA readings, as well as DR and Coref, and

Implementing Experiments on the Language Faculty  

 361

were provided feedback on how accurately they assessed these tutorial items. As Plesniak 2022 notes, this and other design factors seemed to significantly improve attentiveness and comprehension, so it is perhaps a promising venue for future endeavors. As will be a common theme, however, such intensive instruction comes at a cost, namely that it is more work for the experimenter(s); Plesniak 2022 in particular required the experimenter to be in communication with each participant as the experiment was performed, severely increasing the amount of time taken to collect data, as only one participant could take the experiment at once. In contrast, the experiments discussed in Chapter 7 were completed by all participants essentially simultaneously, with the whole process taking a manner of hours; Plesniak 2022’s experiments took weeks to complete, despite having roughly 1/20 of the number of participants. The use of ‘own’ and/or ‘each’ to help convey the BVA interpretation is common to multiple experiments, including Plesniak (2021) and Hoji 2015 (and further unpublished English experiments by Hoji). A contrasting approach, however, can be seen in Plesniak 2022, where the intended interpretations were conveyed pictorially: (2)

His student was spoken to by every professor.

(II)

(I) A

Spoke to

A

B C D E F G

The same student

Spoke to

A’s Student

B

B’s Student

C

C’s Student

D

D’s Student

E

E’s Student

F

F’s Student

G

G’s Student

In (2), (II) represents the BVA reading, the differing lines indicating that A-F (each a different professor) each spoke to a different student, which is identified as A-F’s own respective student. While this graphical presentation may be easier for some participants to understand than an abstract verbal presentation, it is worth noting that there is nevertheless a potential downside: while the use of ‘each teacher’s own student’ is essentially a paraphrase of what a BVA reading is, the arrow drawings only give a picture of a specific situation compatible with a BVA reading. That is, if a participant accurately understands the intended meaning of the instructions, and that participant says that, in the sentence in (1), the phrase in question (‘his student’) “cannot” have the specified interpretation (‘each teacher’s own student’), we can have confidence that they did not accept

362 

 Daniel Plesniak

BVA readings of that sentence. On the other hand, if a participant says that the sentence in (2) is incompatible with the situation depicted in (II) there, we do not know for sure that it is not compatible with some other BVA-consistent situation (e.g., one in which some teachers share students, so that it is still true that each teacher spoke to “his own student”, but sometimes the “his own student” for different teachers overlaps). Such issues become heightened in situations as described in Plesniak 2022, where in Mandarin Chinese experiments, quantified expressions meaning things like ‘many students’ caused some difficulties due to participants needing to be instructed as to whether the depicted number of students in certain pictures could be considered ‘many’ or not; this was not directly related to what Plesniak 2022 intended to test, yet it may nevertheless have influenced the results if participants answered inaccurately due to confusion about what could be counted as ‘many’ in the particular scenario depicted. Such an issue would not arise in the “paraphrase” style of presenting interpretations, as there is no particular situation depicted under that approach. Paraphrases also offer advantages when dealing with certain situations that are not easily pictured, e.g., how to express the notion of ‘even’ in “even John praised his student.” Graphically depicting “even-ness” in this sense is not an easy problem, whereas if we are paraphrasing, we can simply just use ‘even’ in the paraphrase (e.g., “even John is such that he spoke to his own student.”), circumventing the issue. However, paraphrases come with their own limitations, namely being generally much easier for an individual to misunderstand. In order to check that a given paraphrase is accurately understood by an individual, following Hoji 2015, a “BVA-Inst(ruction)-Sub(experiment)” can be employed, as was done for the experiments in Chapter 7 as below: (3) “John praised each teacher’s own student.” ◯ This sentence can mean that John praised only one person. ◯ This sentence cannot mean that John praised only one person.

Here, the concern addressed is that participants might not understand ‘each teacher’s own student’ to mean something like a BVA reading, but might understand it referentially, perhaps as something like “the one student of all the teachers”. If an individual could understand it in such a way, then presumably that individual would choose the “can” option in (3). If so, then we could not consider any of that individual’s judgments on BVA questions as significant; such an individual may be understanding “each teacher’s own student” to somehow signal a referential reading, rather than the BVA reading intended. If this is the case, then the wording of the questions is not conveying the correct interpretations to

Implementing Experiments on the Language Faculty  

 363

this individual, and as such, the judgments provided are not (necessarily) BVA judgments, meaning they have no direct bearing on our predictions. On the other hand, if the individual chooses the “cannot” option, then we have evidence that they do not (mis)understand the phrase in this undesirable way. Of course, we cannot be sure; in the worst-case scenario, a participant could simply be choosing randomly, for example. However, this risk will always be present to a greater or lesser degree; though we become more confident in the reliability of a participant’s judgments the more such tests they pass, there is always a chance of passing by random luck or other such factor. On the other hand, though, if a participant “fails” such a test, then we are fairly sure that something has gone wrong, one way or another. A similar BVA-Inst-Sub can be seen in Hoji 2015, where the exact way in which the BVA instructions are tested is different, as shown in this excerpt: (4) (Under the interpretation where his refers to Mike) Every boy criticized his own father. Here, rather than the phrase ‘each teacher’s own’, it is only ‘his own’ that is being used in the directions, and rather than asking if it can refer to a single individual, what is specifically being asked is whether it can refer to the individual Mike. The crucial similarity though is that if an individual happens to accept such a sentence-interpretation pairing, then that individual is not understanding the instructions as conveying a BVA reading; in this case, it would mean that they can accept “Every boy criticized his own father” such that “his own father” is referential, rather than it forcing a BVA reading of ‘his’. In common to both BVA-Inst-Subs displayed above is that the “disqualifying” answer is of the “can accept” variety; this is crucial, as there are various reasons why an individual might not accept something (see discussion of ✶- vs. ok-schemata in previous chapters), some of which may be completely tangential to what is being investigated. For example, if an individual finds it weird, say, to have a boy named “Mike”, then saying “cannot accept” in an analogous sub-experiment to the above would not be a reliable indicator of intended comprehension or lack thereof of the instructions. Acceptance, on the other hand, does not have the same confounding issue; there are many aspects of a sentence(-interpretation pair) that might cause rejection, but if one accepts a sentence, presumably one finds all aspects (more or less) acceptable. That is, any single rejection-causing issue will (definitionally) cause rejection, so if there is not rejection, we can safely assume that there are no rejection-causing issues. As such, if a “cannot accept” version of the sub-experiment is to be used for some reason, it should generally

364 

 Daniel Plesniak

be done in conjunction with a “can accept” counterpart, as is briefly discussed in Section 4. BVA-Inst-Subs are perhaps the most important of the Hoji 2015-style subexperiments; Plesniak (2021) in fact uses only a BVA Inst-sub to check for comprehension, and that works sufficiently well in that case, demonstrating just how powerful that sub-experiment is. Crucially, it helps to ensure that BVA judgments reported actually reflect a participant’s BVA judgments; if this is not the case, then testing any hypothesis using BVA is essentially meaningless, at best an exercise in seeing how somewhat confused people will answer the questions, which certainly does not have direct bearing on diagnosing the presence/absence of conditions like c-command. Inst-subs are also possible for MR’s that are being used to diagnose the properties of X and Y of the main MR(X, Y), e.g., DR and Coref, as was the case of the weak crossover experiment in Chapter 7. DR in that the experiments for that chapter was conveyed as below: (5) “Two students praised every teacher.” ◯ This sentence can mean that a different set of two students praised each teacher. ◯ This sentence cannot mean that a different set of two students praised each teacher. Here, the phrase “a different set of two students verbed each teacher” conveys, or at least is intended to convey, the sense of “distribution” of one quantified expression over another. For some reason, it seems that this way of expressing DR is somewhat easier to understand than the previously discussed way of expressing BVA (as in (1)); Plesniak (2021) notes that other sub-experiments besides BVA-Inst were included in the experiment reported there but were not used in the analysis because they did not qualitatively affect the outcome. In fact, the DR-Inst-Sub was failed by exactly 0 out of the approximately 100 participants in that experiment, indicating that the vast majority (perhaps even all of them) understood the DR instructions as intended. For Chapter 7’s experiments, a few did fail the DR-InstSub, but this was relatively rare. Given the way in which DR was expressed in (5), the DR-Inst-Sub took the following form: (6)

“Two teachers each spoke to a different set of three students.” ◯ In this sentence, the total number of students can be three. ◯ In this sentence, the total number of students cannot be three.

Implementing Experiments on the Language Faculty  

 365

While this may seem like a simple math problem, and perhaps for many it is, (which is why so many can “pass” it), we can note two possible issues that might lead to a participant not passing it. One, the individual might simply not be paying close attention, and/or two, the individual might understand ‘a different set of three students’ to mean something like “a different set of three students than the original set of three students we were previously talking about”. This would yield a reading where there are only three students total involved, which would not be the DR reading. Of course, because there is no “original set of three students” in the first place, this is unlikely, but an individual may nevertheless imagine it; especially when we are soliciting judgments from hundreds of individuals, “unlikely” ways of interpreting sentences are bound to come up. Individuals not paying attention or having atypical understandings of the instructions are not guaranteed to be judging the acceptability of a DR interpretation when they encounter items like (5), and thus their reported judgments on such items cannot be considered reliable indicators of whether they find the intended DR interpretation(s) possible for the sentence(s) in question; the intent of the DR-Inst-Sub is to aid in the detection of these types of individuals. Coref is perhaps the easiest of the three interpretations to convey. In the Chapter 7 experiments, this was done as below: (7) “His student spoke to John.” ◯ In this sentence, “his” can refer to John. ◯ In this sentence, “his” cannot refer to John.

Given this ease of understanding, no Inst-sub was included for Coref. The overall results of the experiments still came out as predicted, meaning that we have no reason to believe that poor understanding of the Coref instructions leads to any problems; as such, it seems that “‘his’ can refer to John” is indeed clear enough that either (a) everyone understands it correctly, or (b) those who do not understand it also do not understand other things, causing them to fail different sub-experiments. As such, we can see that Inst-subs are not necessarily always required, and their use is motivated by practical, rather than theoretical, concerns. Plesniak 2022 conveyed DR and Coref interpretations pictorially, in analogous ways to the pictorial BVA interpretations discussed above, with the (II) option corresponding to DR and Coref respectively:

366 

 Daniel Plesniak

(8)

Two students were spoken to by every professor.

(I) Prof. #1

Student #1

Prof. #2 Prof. #3

Student #2

Prof #4.

(9)

(II)

Spoke to

Prof. #1

Spoke to

Prof. #2 Prof. #3 Prof #4.

Student #1 Student #2 Student #3 Student #4 Student #5 Student #6 Student #7 Student #8

His student was spoken to by that new professor.

(I)

(II) Spoke to

That new professor

Spoke to

Someone else’s student That new professor

That new professor’s student

Because it is far harder for someone to misunderstand these diagrams than it is to misunderstand the verbal paraphrases, and because participants were given tutorials ensuring that they understood how to read the diagrams as well, no Instsubs were used in Plesniak 2022. Nevertheless, the overall results of the experiments still matched predictions perfectly, suggesting that participants were indeed understanding the interpretations as intended. As such, it is fair to say that the way in which one chooses to convey the MR interpretations has a direct impact on the structure of the experiment, potentially requiring or benefiting from the inclusion of Inst-subs, depending on the clarity and level of training that can be provided to participants. As noted above though, clarity can come at the expense of reliability, potentially shifting focus from judgments on general interpretations to judgments on compatibility with specific situations, and as will be explored in Section 5, there are tradeoffs associated with greater levels of training that might be undesirable, depending on the experimenter’s purpose.

Implementing Experiments on the Language Faculty  

 367

3 Obtaining Judgments Having established various ways that interpretations can be presented to participants, I now turn to the question of how to elicit their judgments on such items. The most basic aspect of this process is what options the participants will have for their responses. For example, the experiments in Chapter 7 had options as shown below: (10) “John was spoken to by his student” ◯ In this sentence, “his” can refer to John. ◯ In this sentence, “his” cannot refer to John. ◻ Check this if you feel the sentence cannot be judged because it is not English. As noted in the previous section, there were two main options, whether the interpretation “can” or “cannot” be as stated. As can be seen in (10), there was a secondary question asked which allowed participants to indicate that the sentence was not a proper English sentence. Similar versions of this option were included in Plesniak (2021) and Plesniak (2022), the logic being to distinguish between sentences that are simply rule-violating strings of words for the participant from those which are possible sentences but which cannot have the interpretation in question; for the type of analysis to be performed, we will be primarily concerned with issues of string-MR pairings, not simple string unacceptability. While for a sentence like “John was spoken to by his student”, this question is rather trivial, for others, especially those involving things like topicalization, e.g., “by his student, every professor was spoken to” (a sentence in Plesniak 2022), the possibility of individuals feeling the sentence is unusual enough to be non-English arises, and indeed, some participants (albeit a fairly small minority of them) do indicate that is their judgment. Participants were also able to skip the question entirely if they were unsure of their answer. While perhaps not strictly required, such an option is generally desirable because of the sensitive nature of possibility-seeking analysis; if, for example, a participant is simply having a hard time judging a BVA ✶-schema instantiation and decides to just “pick an answer” and chooses the “can accept” option, then that individual might potentially look as if they have disconfirmed the prediction at hand, when in reality, that participant was unable to even come to a clear judgment. A skip option helps to mitigate this possibility. Hoji (for experiments referenced in Hoji 2015 and Chapter 6 of this volume) builds in such an option explicitly:

368 

 Daniel Plesniak

(11)

This design is different from the one presented in (10) in several other respects. First, rather than giving options of the form “this sentence can X” or “this sentence cannot X”, it simply presents the sentence along with “under the interpretation that X”. Then, for all questions, there are three options: “unacceptable at all”, “(more or less) acceptable”, and “skip”. There is also explicit instruction at the top, (which was repeated for each such set of items) explaining again that the distinction should be between “totally unacceptable” and “acceptable at least to some extent”. Though the success of experiments like those in Chapter 7 suggests such reminders and language are not strictly needed in all cases, they may well help reduce the noise resulting from confusion of participants as to what sort of standard they should set for acceptance/rejection of sentence-interpretation pairings. (Indeed, for some experimental designs, this may be crucial). The Chapter 7 experiments in particular handled this via directions at the beginning of the experiment as well as a brief tutorial, but whether this is as effective a method as Hoji’s is unclear at this point, even though it apparently was sufficient. Another key difference visible in (11) is that, unlike in the Chapter 7 experiments, where each item was presented on its own, three related items are presented together in (11).2 There is something of a spectrum possible in terms of how much is presented together; in Plesniak (2021) for example, there was only one page, which contained all the experimental items to be judged (though there were very few such items). On the one hand, the more items are presented together, the clearer it will be for participants the relevant factors for making their judgments, e.g., if they see ‘John spoke to his student’ and ‘his student spoke to John’ together, this will make clearer that it is the pairing of the sentence structure with the interpretation that they should be judging than if these items appeared in isolation. In the isolated case, it is possible that participants might be more inclined to judge based on some other criterion, such as naturalness or plausibil-

2 In the experiments for Hoji 2015, participants were given an option between this presentation style and a “one-sentence-at-a-time” style; in 5.4.6 of that work, Hoji reports the two options did not seem to result in any significant difference in responses.

Implementing Experiments on the Language Faculty  

 369

ity; assuming that the participants have no prior linguistic training, judgments based on these sorts of criteria may well seem much more “normal” to them. In this sense, presenting items together can help participants understand the intended task. On the other hand, however, presenting items together runs the risk of making participants think that they must be contrasting the items against one another, creating the incorrect impression of a “forced choice” task. This may potentially cause them to “reject” items they would otherwise have accepted, because they are less easy to accept than their co-presented counterparts. Hoji’s wording in (11), of course, essentially tells them not to do this, but confused or careless participants may fail to read carefully, understand what is meant, or keep the instructions in mind. A related issue whether there should be randomization between the different items in the experiment in terms of order, i.e., to what degree should thematically related items be presented together vs. being “scattered” throughout the experiment at irregular intervals. One reason sometimes given for such randomization is that it makes it harder for participants to guess what is being tested; indeed, the same logic is given to argue against presenting like items together at once. However, given that what is being tested in this type of experiment is a complex correlation across many different item types, rather than judgments on a specific item(s)/item type(s), it is not clear how a participant could come to “guess” that. Anecdotally speaking, participants generally seem to think that our primary interest is knowing their judgments on the different sentence-interpretation pairs individually, rather than any connection across such pairs. Further, their reported judgments frequently vary wildly across what are for us closely analogous sentences, implying that they are not changing their judgments based on having figured out the “intended” pattern. Further, because the experiments are possibility-seeking in nature, note that it only takes one participant violating said correlations to yield a result that goes against the predictions. As such, even if a particular participant were to guess the correlations in question, and then that participant went on to give answers that they felt the experimenter “wanted to hear”, rather than accurately reporting their judgments, this would not ensure overall results that support the predictions in question; other participants, who did not make such a guess, could still easily provide a disconfirming judgment. As such, unless one is willing to argue that every single participant (or at least a very large subset) is guaranteed both to figure out the intended correlation and strictly match their judgments to it, such concerns will not directly impact the results in a qualitative sense. Nevertheless, similar to the issue of co-presentation, we may nevertheless be concerned that certain orderings or groupings of presentations may bias towards certain patterns of responses. However, given the (admittedly limited) data available, this does not seem to be the case. The items in the Chapter 7

370 

 Daniel Plesniak

experiments were largely3 randomized, and yet the results essentially converge with those of Plesniak (2021), which presented all items together, Plesniak 2022, which presented items separately but in a fixed repeating pattern, and Hoji’s Chapter 6 experiment, which also had randomization of items within certain set groups. There is qualitative convergence across these different methods, suggesting that, for the correlational methodology, the various forms of obfuscation (or the lack thereof) do not have a particularly strong effect, a stark contrast with their reputation in the sort of “typicality-seeking” research discussed in Section 1. In a similar vein, numeric ranking scales are also popular in many other forms of experimental collection of judgments, but they have thus far played a minimal role in the development of the correlational methodology (though see Hoji 2010 for an attempt to make use of them in an individual-focused experiment, albeit one without the correlational methodology). This is not to say such rating scales are unusable, merely that their use is no simple matter. This is true even in the “typicality-seeking” research domain; see discussion and references in Dillon and Wagers 2021, where it is pointed out that we have no guarantees that many common assumptions hold when using such scales, e.g., that participants treat the distance between “2” and “3” as the same as the distance between “3” and “4”, which will frustrate overly simplistic forms of statistical analysis. This problem is directly echoed in research focusing on the judgments of individuals as well: what definite judgment sensation does “this sentence interpretation pair is a 2 out of 5” correspond to? We could perhaps map a scale to “unacceptable” and “acceptable” through some procedure, e.g. 1“unacceptable”, everything else“acceptable”, but in that case, why not simply ask about “unacceptable” and “acceptable”?4 If we have a set of hypotheses that predicts certain gradient judgments, such that we can say things like “such and such a sentence-interpretation pair will be rated no higher than a 2 if. . .” and “at least some instantiations of this schema should be rated at least a 4 in order to. . .”, then such a scale will become quite useful. 3 In earlier iterations of the experiments, participants sometimes complained about having to “judge the same item multiple times”, which was concerning given that no item appeared more than once. However, some items were quite similar to one another, and it seemed that participants were assuming that such items were just the same. This raised the possibility that they might not thoroughly read the second such item, assuming it to be a copy, and just repeat the same answer as before, which could cause their answers to be essentially invalid. To eliminate this possibility, similar sentences were moved closer to one another, though in such a way that order between them remained randomized, so no pattern could be reasonably detected. 4 Earlier versions of the Chapter 7 experiment attempted to take just this sort of numerical approach, but it seemed that there was no simple way of achieving this that would yield anything like the qualitative results found by the “binary” method.

Implementing Experiments on the Language Faculty  

 371

As it stands though, our model of judgment-making is relatively limited, allowing only for sentence-interpretation pairs to be generated or not generated (see Hoji’s Chapter 4 in this volume), and until a way of extending this model to incorporate such numerical “levels” of acceptability is fruitfully articulated and demonstrated, the binary choice between “acceptable” and “unacceptable” (potentially with a few auxiliary options, e.g., “this is not an English sentence at all.”), seems to be doing well enough that it has yielded multiple predictions that receive definite and replicable support from experimental data.5 This situation exemplifies the ways in which measurements must be constrained by the hypotheses under consideration, and how a change in hypotheses may allow for a change in measurement type.

4 Design Across Items Having established options for both conveying interpretations as well as eliciting participants’ judgments on whether those interpretations are or are not possible for given sentences, in this section, I discuss how those sentences should be designed with respect to one another. Following the basic tenets laid out in Hoji 2015 and reviewed in earlier chapters of this volume, at a minimum it is necessary to have sentences that instantiate both ok-schemata and ✶-schemata. In the case of the weak crossover experiment, these were, for the BVA case, ‘X was Verb by Y’s Noun’ and ‘Y’s Noun Verb X’ respectively, both with BVA(X, Y). There were supporting DR and Coref schemata of the same sort, respectively ‘X was Verb by # Noun’ (# indicating a number) and ‘# Noun Verb X’, both with DR(X, # Noun), and ‘Noun/NounPhrase was Verb By Y’s Noun’ and ‘Y’s Noun Verb Noun/NounPhrase’, both with Coref(Noun/NounPhrase, Y). An immediate question arises as to why these sentences are so similar. What we should remember is that DR and Coref are playing the role of performing the tests on X and Y that will tell us whether X and Y are structurally sensitive6 or not for the individual; the results of these tests will inform how we interpret the 5 A perhaps simpler and more likely to succeed extension would be to incorporate notions of “more/less acceptable than”; I, for example, have not had intuitions like “this is an acceptability level of 2”, but I have had intuitions of some items being “not totally unacceptable, but much harder to accept than this other one”. It is possible some hypotheses might fruitfully generate such predictions, but as with the numerical scales, it remains to be demonstrated that such hypotheses would lead to replicable results. 6 See discussion of this term in Chapter 7, which itself refers back to discussion of “NFS” in Hoji’s Chapters 4 and 5 of this volume and “quirky” in Ueyama’s Chapter 2 of this volume. What is meant is essentially whether X and Y participate in MR(X,Y) when X does not c-command Y.

372 

 Daniel Plesniak

(un)availability of BVA(X, Y) in the relevant ✶-schema instantiations. As such, one might think it would be ideal for the DR and Coref schemata to be rather different than the BVA ones, so as to avoid any sense of circularity. More clearly articulated, the concern is something along the lines of the following: because the DR and Coref ✶- and ok-schemata are so similar in form to the BVA ✶- and ok-schemata, then perhaps using judgments on the DR and Coref schemata to “predict” something about judgments on the BVA schemata is tautological, inferring “A” from “A”. While this is a reasonable concern, the claim that what is being inferred is tautological is inaccurate; as implemented in experiments like those of Chapter 7, the DR and Coref tests are not circular in their predictions; though the sentence structure is similar, the DR and Coref items are not the same as the BVA items, nor is there an obvious a priori requirement that judgments on Coref and/or DR items ought to be the same as the judgments on their BVA analogues. Consider the following sentences from Chapter 7’s “weak crossover” experiments, which are the BVA/DR/Coref “equivalents” of one another: (12) With a BVA(every teacher, his) interpretation: a. Every teacher praised his student. b. His student praised every teacher. (13) With a DR(every teacher, three students) interpretation: a. Every teacher praised three students. b. Three students praised every teacher. (14) With A Coref(that teacher, his) interpretation: a. That teacher praised his student. b. His student praised that teacher. These sentences are all clearly distinct and have different interpretations; there is no a priori obvious reason that an individual who accepts (13a) and rejects (13b) with the specified interpretation will have the same pattern of judgments on (12a) and (12b), and further, we have identified hundreds of individuals who in fact do not have the same patterns of judgments on analogous items across different MR’s. There is thus no “circularity” induced by using an individual’s judgments on the sentences in (13) and (14) to predict that individual’s judgments on the sentences in (12).7 Rather, the similarity between sentences is a consequence of our

7 Note further that what was done in the experiments presented in Chapter 7 was more complicated than this simple comparison, involving using judgments on multiple DR/Coref sentences,

Implementing Experiments on the Language Faculty  

 373

attempt to minimize potential sources of noise that might creep in when aspects of the sentences change; as such, unless we feel very confident we understand these sources of noise well enough and are confident that we have a robust way of identifying them across different kinds of constructions, the most straightforward solution is to keep all sentences as similar as possible, allowing for only minimal difference between them when necessary. In some sense, this is simply an extension of the common usage of “minimal pairs” as a basis for data in the various sub-fields of generative linguistics; we are not necessarily expecting these minimally-differing pairs of sentences to contrast with one another, but the basic reasoning is the same, namely minimizing the chance for some sort of “outside factor” to confound judgments by keeping the sentences under consideration as similar as possible to one another. While perhaps it would be indeed ideal, or at least impressive in a certain sense, to use more different-looking schemata for DR and Coref as compared to BVA, implementing such an idea faces immediate difficulties; as emerges in the analysis of Plesniak 2022’s data, as well as several experiments that went unpublished precisely because of this failure, it is not trivial to use a non-minimally different schema to predict how X or Y will behave in the BVA schemata. Trying to use the pattern of acceptability of DR in passives to predict the pattern of acceptability of BVA active voice sentences (all for a given individual with a specific choice of X) seems unreliable, and has failed to replicate previously established results thus far, albeit it has not been tried very extensively.8 This difficulty extends not only to the structural aspects of the sentences, but also to the words within them. While there seems to be greater leeway here (Hoji’s experiment in Chapter 6 does not always perfectly match all lexical items

all corresponding to given schemata, to predict judgments on BVA sentences corresponding to the analogous schemata. 8 That is to say, while an individual’s ability to accept DR(every teacher, two students) in (i) has been reliably useful in predicting that individual’s ability to accept BVA(every teacher, his) in (ii), that individual’s ability to accept DR(every teacher, two students) in (iii) has not been reliably useful in the same way: (i)

Two students spoke to every teacher.

(ii)

His student spoke to every teacher.

(iii)

Two students were spoken to by every teacher.

By “reliably useful”, I am considering their use as diagnosers of the possibility of non-structurally-sensitive BVA after all the other steps required for the correlational methodology have been taken (consulting with Coref sentences as well, for example). It may well be that sentences like (iii) can be used in this way once the additional steps are taken, but whatever those are, they seem to be somewhat more intensive than what is required to use (i) in this way.

374 

 Daniel Plesniak

between conditions, and this seems to not have had much of an adverse effect on the results, though it is hard to know for sure), experiments like those in Chapter 7 have erred on the side of caution, changing words between conditions only when absolutely necessary. On the one hand, this restricts our results to only a very narrow range of sentences; all sentences in the Chapter 7 experiments involved the verbs ‘spoke’ or ‘praised’, for example, so we may well wonder if the same results will obtain if we use different verbs. However, within that small set, the reliability of our results is greatly enhanced, because if an individual’s judgment consistently changes between two otherwise identical items, then we can feel confident that this change is due to the minimal difference between those items; if we not only changed the crucial aspect of interest between sentences, but also gave each sentence a different verb, if judgments change when we move from one sentence to another, we would be unsure which difference was responsible: the crucial aspect of interest or the different choice of verb? That is not to say that such experiments should never be done, but rather, that figuring out what “incidental” changes are able to take place without disrupting the overall correlations is a currently a frontier in the development of this methodology, and until more is known, a conservative approach is likely prudent unless one wishes to address that question in particular. Keeping with the theme of using minimally differing sentences, some but not all experiments of this type make use of a second set of ok-schemata, which are matched to the ✶-schemata in their sentences, but which feature non-MR interpretations, as shown in the below examples: (15) “His student spoke to every teacher.” ◯ In this sentence, “his student” can refer to a specific teacher’s student, such as Mike’s student. ◯ In this sentence, “his student” cannot refer to a specific teacher’s student, such as Mike’s student. (16) “Two students praised every teacher.” ◯ This sentence can mean that the same two students praised each teacher. ◯ This sentence cannot mean that the same two students praised each teacher. (17) “His student spoke to John.” ◯ In this sentence, “his” can refer to someone other than John, such as Bill. ◯ In this sentence, “his” cannot refer to someone other than John, such as Bill.

Implementing Experiments on the Language Faculty  

 375

(15)–(17) have the same form as the ✶-schema examples in (1), (5), and (7) respectively, but unlike in those cases, the interpretation asked about is not a BVA/DR/ Coref interpretation. Hoji 2015 originally used these to ensure that a given sentence was acceptable without the MR interpretation, which makes rejection of the sentence with the MR interpretation more significant, given that we know that this is not merely a rejection of the string of words that form the sentence itself (see Mukai’s discussion in Chapter 3: Section 3 of this volume). However, given that the Chapter 7 experiments had an explicit option that participants could choose if a sentence seemed non-English to them, the inclusion of such questions was essentially redundant, and they did not figure into the analysis. Indeed, such items may run afoul of similar concerns to those voiced in Section 2 about graphical depictions of situations rather than paraphrases of MR’s, namely that it may be the case that a participant does in fact accept the sentence under some non-MR interpretation, but not the particular non-MR interpretation displayed. As such, just generally ensuring that a participant agrees that the sentences that participant is judging is indeed a sentence of their language is perhaps a more widely applicable way of warding off the concerns of “string unacceptability”, though as always, there may be experiment-specific reasons to prefer other methods. Yet another approach to presenting related sentence-interpretation pairs can be seen in the examples from Plesniak 2022 (see (18) below, which is (2) repeated), where both a BVA and non-BVA (or whichever MR) picture are presented side by side, with a participant being asked to say whether option I, option II, both, or neither is possible; if a participant answered “neither”, a follow-up question was asked as to whether the sentence was simply a “bad sentence”, or whether there was some third interpretation the participant thought was possible: (18)

His student was spoken to by every professor.

(II)

(I) A

Spoke to

A

B C D E F G

The same student

Spoke to

A’s Student

B

B’s Student

C

C’s Student

D

D’s Student

E

E’s Student

F

F’s Student

G

G’s Student

In some sense, this is a “best of both worlds” approach, as it alleviates the concerns mentioned above about rejection of a specific interpretation not generally implying that all possible interpretations are rejected. It preserves, however, one aspect of

376 

 Daniel Plesniak

the original Hoji 2015 formulation which is easy to overlook, namely that providing a non-MR interpretation of ✶-schemata may help participants better evaluate the MR interpretation. Anecdotally, sometimes a participant does in fact find BVA(X, Y) impossible for a given sentence, yet because BVA(X, Y) seems so outlandish paired with that sentence, the participant assumes we must have been asking them about some other potential reading. This leads to the participant essentially “making up” a plausible (possibly non-BVA) reading and reporting their judgments on the availability of that reading instead, yielding an inaccurate reporting of their “BVA” judgment. Essentially, the participant assumes we must be asking them about a “sensible” reading, and they find BVA totally nonsensical (in the sense of it being a totally impossible reading for the sentence), so they change the reading to a more easily acceptable one without realizing it, likely causing them to report acceptance when in fact they strongly reject BVA. In this “side-by-side” way, such an outcome is inhibited, as the interpretation provided in (I) is usually uncontroversially acceptable, meaning that participants are able to understand what the MR reading would contrast with, and thus better identify if they themselves can accept the MR reading; that is, by seeing at least one “sensible” reading, they are more easily able to understand that the other reading might indeed seem nonsensical and/or impossible and correctly reject it. Further, this side-by-side presentation is done for both ok- and ✶ -schema items, reducing the asymmetry between the two inherent in the original Hoji 2015 design, where it was (mostly) only the ✶-schemata that had non-MR counterparts. Nevertheless, this way of presenting things is not without drawbacks; see discussions in Sections 1 and 5. These different schemata being established, the natural question arises as to how many instantiations of each schema is necessary. Successful experiments have been conducted where the answer is as low as one (e.g., Plesniak 2021), but this is generally inadvisable, if for no other reason that there is a chance for human error in relaying the judgment (e.g., pushing the wrong radio button without noticing), and with only one instantiation, it becomes impossible to distinguish such situations from accurately reported judgments. However, there is no one-size-fits-all answer; after various trial and error attempts, four instantiations of each Coref and DR schema and two of each BVA schema were used for the Chapter 7 experiments. This seemed to be the minimum necessary to ensure a “clean” result, and roughly matches with Hoji’s (Chapter 6) methodology. One may even go one step further, and have the same items be judged twice (or more), as has been done in certain unpublished experiments (p.c. Hajime Hoji, May 2021). This would further eliminate the possibility of an erroneous button press throwing off the experiment and allow us to examine an individual’s consistency on the same item. This is naturally desirable in principle, though as a downside, it doubles experiment length and may be confusing to participants, and as such it has a number of associated costs.

Implementing Experiments on the Language Faculty  

 377

As to why the DR, Coref, and BVA schemata were each instantiated by 4 sentences, 4 sentences, and 2 sentences respectively, this derives from the earlier discussed need to keep lexical items in sentences minimally different from one another, so as to not introduce new noise between conditions. As mentioned, there were two “predicates” used, ‘praised’ and ‘spoke to’. For the DR(X, Y’) cases, the choices of X was the same as in BVA(X, Y) (either ‘every teacher’ or ‘more than one teacher’), but there were two options given for Y’(=# Noun): ‘two students’ and ‘three students’; the use of these two options helped increase confidence that a rejection of DR(X, Y’) was not due merely to the choice of Y’. Similarly, for Coref(X’, Y), where Y was the same as BVA(X, Y), there were two choices of X’, ‘John’, and ‘that teacher’, also serving to increase confidence that the choice of X’ was not driving rejections of Coref(X’, Y) for a given individual. As such, there were four instantiations of DR and Coref (2 verbs multiplied by 2 choices for Y’/X’), but only two of BVA, as there were no alternate choices for X and Y (for a given participant). These numbers were not theoretically derived but were merely the numbers that seemed minimally necessary to ensure that a given MR had been sufficiently probed for a given participant to ensure that a consistent rejection of that MR in a ✶-schema was not due to some idiosyncratic fluke of the particular item (or the participant answering randomly/without sufficient caution or consideration). Another such non-theoretical number is how many instantiations of an ok-schema a participant must accept in order for us to count them as generally accepting the ok-schema. If one is very confident in the reliability of one’s participants (as was the case in Plesniak 2022), then perhaps only one such acceptance is necessary, but in an online setting like the ones the Chapter 7 experiments were deployed in, there is a high chance of participants answering carelessly or becoming confused. As such, it seemed that in order to be sure of consistency, it was necessary for participants to accept at least a majority of the instances of a given ok-schema before it could be concluded that they did indeed accept that schema (2 or 3 items, depending on whether it was a BVA or DR/Coref schema). Once again, this aspect of the design/analysis is not derived from the basic hypotheses being tested, but must be established for a given style of experiment by trial and error; replication of previously established results has a role to play here, as these can serve as a kind of “control” by which to set what thresholds are necessary to replicate the results.9

9 At this point, perhaps the only result well established enough for such a purpose is that of weak-crossover (see Hoji’s Chapter 6, Plesniak’s Chapter 7, Plesniak 2021, Plesniak 2022), though even that has only be “well established” in a very limited domain.

378 

 Daniel Plesniak

One further area which, although certainly connected to hypotheses, has aspects that need to be empirically worked out on an experiment-by-experiment basis is the question of which sub-experiments to include and how many instantiations of each to include. While Hoji 2015 gives a theoretical motivation for each of the sub-experiments used, it is nevertheless true that it is helpful to have “catch trials” or “control questions”, which have a correct, or at least a “standard”, answer and can be used to determine which participants are not giving reliable judgments. This is true regardless of whether these questions are theoretically motivated by the experiment at hand or not. When participants are in settings where many are likely to be incautious and/or confused, such considerations become even more important; Plesniak 2022’s in-person, one-on-one experiment makes little to no use of them, for example, whereas the Chapter 7 experiments, based on mass surveys of anonymous online test-takers, use ten sub-experiments each, five basic types represented by two instances each. Two basic types have been discussed before, namely the BVA and DR Inst-subs. Borrowing from Hoji 2015, three others were used, primarily to ensure participant attentiveness and comprehension: (19)

BVA Lex Sub “Every teacher spoke to John’s student.” ◯ In this sentence, “John’s student” can refer to each teacher’s own student. ◯ In this sentence, “John’s student” cannot refer to each teacher’s own student.

(20) DR Lex Sub “Every teacher praised those two students.” ◯ This sentence can mean that each teacher praised a different set of two students. ◯ This sentence cannot mean that each teacher praised a different set of two students. (21)

Top Sub “That teacher, Bill spoke to Mary.” ◯ This sentence can mean that Bill spoke to that teacher. ◯ This sentence cannot mean that Bill spoke to that teacher.

Readers can refer to Hoji 2015: 4.4 for specific explanations of the original rationales for such questions, but the basic notion is that none of these sentences should be acceptable with the interpretations in question. (21) is particularly egregious, and it is hard to see how anyone could find the pairing acceptable,

Implementing Experiments on the Language Faculty  

 379

unless such a person was stretching the notion of “mean” immensely, in which case, that person may not be not answering the questions as intended to begin with, so it is desirable to not assign much significance to their judgments. Some of these items are a bit less clear cut, however; while most individuals do indeed reject the DR(every teacher, those two students) reading asked about in (20), I myself can imagine contexts where it is possible (namely if the students are understood as representing a “kind” of student). Plesniak 2022 addresses this by adding ‘over there’ after ‘those two students’, which makes that already unlikely reading much harder to get. Likewise, with (19), I have had participants ask me whether it might be the case that every teacher is named “John”, and thus who “John’s student” is varies with the teacher in question.10 Rather than an example of poor attentiveness or comprehension, this is a case where participants are being extremely observant and considering very atypical yet possible situations, which is what we would generally like them to do given our focus on what is or is not possible at all over what is typical. As such, these sub-experiments are somewhat over-eliminative; it is generally the case that even a well-designed sub-experiment may risk “filtering out” some individuals who are very attentive and comprehending, but just very creative or insightful.11 This is a disadvantage, to be sure, and it suggests we should use such sub-experiments as minimally as possible; nevertheless, sub-experiments have proven generally necessary when dealing with non-expert participants, especially in non-face-to-face situations, so we may sometimes need to be over-eliminative, as this only reduces the strength of our results, whereas being under-eliminative can lead to results that suggest a given judgment is reliable for our purposes when it is not, a far worse result. 10 Another, perhaps even more extreme, anecdote is that in Plesniak 2022’s experiment, there was a series of tutorial sentences of the form: (i) (ii) (iii)

John likes his roommate, and Bill does too. John likes Bill’s roommate, and Bill does too. John likes pizza, and Bill does too.

It is intended that participants see that (i) can have both “strict” and “sloppy” interpretations of ‘does too’, namely (in this case) ‘Bill likes John’s roommate’ and ‘Bill likes Bill’s roommate’. (ii) is supposed to only be able to have the latter interpretation (though a persistent minority insists the former is possible, and I admit that I can see what they mean, to a degree), whereas (iii) is not supposed to be acceptable with either interpretation. However, at least one participant insisted (iii) could mean that Bill likes either Bill or John’s roommate, on the condition that the roommate is named “Pizza”. This demonstrates how hard it is, indeed perhaps how futile it is, to try to come up with a sentence-interpretation pair that can “never be accepted”, (at least so long as the sentence string itself is not totally malformed). 11 See the “resourcefulness” entry in Hoji 2015: Glossary and “informant resourcefulness” in the Index, as well as Ch. 4: note 20.

380 

 Daniel Plesniak

There are of course permutations of this basic pattern; Plesniak 2022 uses modified versions of such sub-experiments primarily as tutorials, for example, giving participants feedback on their responses to help them better understand the main questions they will be asked in the experiment. On the other hand, Hoji 2015 and Chapter 6 of this volume include the equivalent of ok-schemata in the sub-experiments, as in the first sentence below, which a participant must accept up to a certain threshold to be considered to “pass” the sub-experiment. For example, the topic sub, where individuals had to accept the first choice and reject the second to pass, thus incorporating both an ok-schema and a ✶-schema into the sub-experiment. (22) (Under the interpretation “Mary praised Bill”) Bill, Mary praised. (Under the interpretation “Mary praised Bill”) Bill, Mary praised John. (Under the interpretation “Mary praised Bill”) Bill, Mary praised him. As in many areas, there is a tradeoff here between careful control on the one hand and over-elimination of subjects on the other (as well as a cost in time spent on sub-experiments). These tradeoffs are the subject of the Section 5. Before moving on to that, let me sum up this section by noting that the primary consideration for determining how the various items should relate to one another is noise control. This manifests both in controlling noise in ways we can derive based on our hypotheses, e.g., minimizing the differences between ✶-schemata so as to ensure that no new “non-structural effects” are introduced, and in controlling noise in ways that must be addressed on a “trial and error” basis, e.g., how many instantiations of a given sub-experiment are necessary to ensure that an individual has “genuinely” passed it. By examining the sources of noise inherent to a given task, experimenters can refine what types of noise control are needed for the specific experiment being conducted; generally one wants to use the minimum amount necessary, but in many cases, that minimum amount is still quite considerable.

5 Conclusion: Design Constraints I close with a brief discussion of the relationship between the overall intent of a given experiment and the design decisions made by the experimenter. In the previous sections, I discussed how choices can be made with regard to item pres-

Implementing Experiments on the Language Faculty  

 381

entation, data elicitation, and inter-item experimental design, with a focus on the pros and cons of the various options employed in previous experiments. A running theme that I hope has emerged is that the design of a particular experiment should be dictated by the overall intent of the experimenter. As I mentioned in Section 1, correlational experiments in the style laid out by Hoji in previous chapters differ in their object of inquiry from many other linguistic experiments, and thus are subject to different design considerations. Going further, however, even within those experiments that share the general goal of testing hypotheses and their predictions in a “possibility-seeking” universalist and individualist manner, there are many possible sub-goals for a given experiment. A clear example of one such potential “split” can be seen between (i) experiments that prioritize robust noise control and (ii) those that prioritize maximizing the number of potential replications of a given prediction. (i) tends to sacrifice volume of “crucial” data gathered for reliability of said data, whereas (ii) does the reverse. The experiments in Chapter 7 are of the former type; out of 200 individuals, only a handful (about 10 in the case of the weak crossover experiment) “survive” the various types of noise control; that is, the vast majority of participants are shown to have some factor which may cause their judgment on the key BVA items to be unreliable for our purposes, either because of an attentiveness/comprehension issue diagnosed by Hoji 2015-style sub-experiments, or because of I-language specific noise in the form of structural insensitivity of X and/or Y of BVA(X, Y) (as diagnosed by their DR and Coref judgments respectively). That is not to say that the judgments of such individuals are “thrown out”; that data in fact features in the ultimate analysis, especially those individuals who do accept the ✶-schema items (see discussion in Chapter 7). However, even if we do include such individuals in the analysis in some way, it is clear that there are many individuals whose judgments did not directly (positively) replicate the predictions but were merely consistent with them, in the sense of providing neutral results which do not strongly support the predictions, even though they do not contradict them. Plesniak 2022 takes precisely the opposite approach, eschewing most forms of noise control (few if any sub-experiments in the style of Hoji 2015, and very few instantiations of each schema) in favor of providing far more (and more varied) BVA items for participants to judge, all in an effort to replicate the crucial “schematic asymmetries” across as great a proportion of individuals as possible; in many cases, this replication is achieved across the majority of participants in Plesniak 2022’s experiments. On some level, this result is more pleasing; evidence for a universal claim about individuals is strengthened when we can show that the predicted pattern does indeed obtain across most, if not all, individuals when the correct conditions are met. In some sense, however, something has been sacrificed, namely reliability. In the theme of “learning from mistakes” (see Mukai’s

382 

 Daniel Plesniak

Chapter 3 in this volume), we must wonder what would happen if things did not work out as planned for Plesniak 2022, i.e., if the predictions being tested there seemed to get disconfirmed by certain data. The relative lack of noise control, though it enabled greater replication, would also confound attempts to discover whether or not the apparent disconfirmation of predictions was really due to the predictions being wrong, or just due to Plesniak 2022’s experiments not having strict enough noise control. Thus, while experiments in the style of Plesniak 2022 are perhaps more convincing when they show the intended result, experiments in the style of Chapter 7 are more likely to be reliable when things do not come out as planned. Neither style is inherently better in all ways, and the choice must be made based on the confidence the experimenter has that the results will come out in accordance with predictions. Indeed, as Plesniak 2022 was, at least in part, replicating the results from the experiments in Chapter 7, such confidence was perhaps warranted; these results were already well-established (especially so in the case of weak crossover), so the risk of disconfirmation due to basic hypotheses being fundamentally wrong was low. Plesniak 2022 was also confident that disconfirmation due to poor attentiveness/comprehension would be alleviated by a much more participant-friendly approach, whereby the experimenter both gave tutorials directly to the participants and was available to answer questions they posed during the experiment. Naturally, the rate of error due to confusion was lower in such an experiment than in the ones described in Chapter 7, where participants were simply given a form with questions over the internet, with no chance to ask questions or receive feedback (see Section 2). As noted in Section 2, however, the drawback is that Plesniak 2022’s experiments had an order of magnitude fewer participants than did those of Chapter 7. One could of course try to combine these styles, with a wide array of questions, an abundance of checking and re-checking, many participants, and heavy experimenter instruction/availability for each of them, but the costs in resources, in terms of money, time, and human energy expended, grow quickly. Sometimes, it is possible to meet some of these expanded demands due to special circumstances, as was true in the case of the experiment conducted for Chapter 6 (see discussion therein12), but even that did not meet all of these desiderata; excepting cases where one is blessed with an exceptionally large budget, some degree of compromise will need to be made. Plesniak 2022’s concluding chapter offers some thoughts as to how best to budget for each of these different aspects but generally concurs with the thesis put forth here: depending on what one is trying to do, there

12 Crucially the fact that they were taken by hundreds of undergraduate students working on them over the course of a semester.

Implementing Experiments on the Language Faculty  

 383

is not a set way of making these decisions. Experiments seeking to discover “new laws” of the language faculty will necessarily differ from those seeking to provide strong demonstrations of the accuracy of existing predictions. Ultimately, number of questions, repetitions, or types of control are less important than the rationale used for deciding what is included; careful trial and error can give us a good sense of how many instances of a given thing is needed, but general design principles must flow from basic considerations of what the object of inquiry is, and what the specific goal of the experiment is to be. This is consistent with the theoretical considerations discussed in preceding chapters: in both the theoretical and practical domains, careful deduction and patient trial and error are the key to increasing our understanding of the language faculty via experimentation.

References Dillon, Brian and Matthew Wagers. 2021. Approaching gradience in acceptability with the tools of signal detection. In Grant Goodall (ed.), The Cambridge handbook of experimental syntax. Cambridge: Cambridge University Press. Hoji, Hajime. 2010. Evaluating the lexical hypothesis about Otagai. Linguistic Research 4. Hoji, Hajime. 2015. Language faculty science. Cambridge: Cambridge University Press. Hoji, Hajime. This volume: Chapter 4. The key tenets of language faculty science. In Hajime Hoji, Yukinori Takubo and Daniel Plesniak (eds.), The theory and practice of language faculty science. Berlin and Boston: De Gruyter Mouton. Hoji, Hajime. This volume: Chapter 5. Detection of c-command effects. In Hajime Hoji, Yukinori Takubo and Daniel Plesniak (eds.), The theory and practice of language faculty science. Berlin and Boston: De Gruyter Mouton. Hoji, Hajime. This volume: Chapter 6. Replication: Predicted correlations of judgments in Japanese. In Hajime Hoji, Yukinori Takubo and Daniel Plesniak (eds.), The theory and practice of language faculty science. Berlin and Boston: De Gruyter Mouton. Mukai, Emi. This volume: Chapter 3. Research heuristics in language faculty science. In Hajime Hoji, Yukinori Takubo and Daniel Plesniak (eds.), The theory and practice of language faculty science. Berlin and Boston: De Gruyter Mouton. Plesniak, Daniel. 2021. Experiments beyond statistics: How generative linguistics can beat the replication crisis. Ms, available at: https://www.researchgate.net/ publication/357063705_Experiments_Beyond_Statistics_How_Generative_Linguistics_ Can_Beat_the_Replication_Crisis (accessed 21 March 2022) Plesniak, Daniel. 2022. Towards a correlational law of language: Three factors constraining judgment variation. Los Angeles, CA: University of Southern California dissertation. Plesniak, Daniel. This volume: Chapter 7. Predicted correlations of judgment in English. In Hajime Hoji, Yukinori Takubo and Daniel Plesniak (eds.), The theory and practice of language faculty science. Berlin and Boston: De Gruyter Mouton. Schütze, Carson. 2016. The empirical base of linguistics: Grammaticality judgments and linguistic methodology. Berlin: Language Science Press.

Part 3: LFS as an Exact Science

Hajime Hoji and Daniel Plesniak

Language Faculty Science and Physics 1 Introduction 1.1 Chapter Objectives This chapter explores the consequences of two basic notions.1 The first such notion is that the “scientific method” can be applied anywhere reflexes of a formal system can be measured, regardless of whether or not those reflexes are physical or mental. In other words, so long as the object of inquiry has definite properties, and those properties give rise to reliably measurable phenomena, it can be studied via the process of deducing definite predictions from basic hypotheses and comparing those predictions to measured results. In this sense, Language Faculty Science (LFS), where the object of inquiry is a specific aspect of the mind, can be pursued using the same basic approach as is used in physics to study the properties of nature. The second, derivative, notion is that, despite superficial differences between such domains, the challenges faced in a primarily “mental” field like LFS and a primarily “physical” field like physics are instantiations of the same fundamental challenges. Indeed, not only are these challenges the same, but so are the basic methods of overcoming them. As will be explored, these challenges and their solutions have their basis in issues surrounding the comparison of definite, deduced predictions against carefully measured data; what differentiates fields is the nature of those predictions and the data to be measured. This chapter will illustrate these two notions with regard to the case studies of LFS and physics, drawing on examples from both domains. LFS is by no means the only relevant “mental” field. For example, the formal study of vision also deals with a primarily mental phenomenon, relating light hitting the retina to the mental images we “see”. The process by which a particular pattern of light gets mapped to a mental image involves mental computation, and thus requires some sort of a formal underlying system (see, for example, Marr 1982). Our purpose, however, is not to compare “mental” fields with one another, but rather to demon-

1 Much of what will be presented in this chapter is grounded in what has been addressed in preceding chapters. As such, it is based on various conceptual and methodological proposals and experimental illustrations of their viability provided there; for that reason, we will provide pointers in footnotes as to where the reader can turn to for relevant discussion in preceding chapters of topics we do not have space to fully address here. https://doi.org/10.1515/9783110724790-009

388 

 Hajime Hoji and Daniel Plesniak

strate that direct comparison is possible between “mental” and “physical” fields. As such, while fruitful comparisons between language and vision (and indeed between vision and physics) can certainly be made, we will focus on the comparison between language and physics. Similarly, we will not make much comparison between LFS and other types of “language”-related research; having articulated here the ways in which LFS represents a formal scientific endeavor, we leave it to future works to compare and contrast LFS and other approaches to languagerelated research along this dimension.2

1.2 Conceptual Review of LFS Because this chapter involves comparison across different disciplines, and is considerably less “technical” than other chapters in this volume, we expect there may be some readers who wish to read this chapter in particular, without having read the rest of this volume. To accommodate such readers, in this section, we provide a brief overview of LFS as described in this volume, which assumes no background reading or knowledge of linguistics. Those who have read the previous chapters or otherwise know the relevant background should feel free to skip to the following section (though some may find it a helpful “summary” of key points from the previous chapters). Language Faculty Science (LFS) is a research program that investigates the human ability to map between external linguistic symbols to our own internal representations of meaning. Such linguistic symbols can come in various forms, be it the spoken word, the signs of sign language, or even written text. While we tend to think of these forms as bearing meaning, they themselves are simply waves of sound or light from a physical perspective; it is our minds which assign particular meanings to those waves. We term the component(s) of the mind involved in this process the language faculty; our research question, in the broadest terms, is how the language faculty works. Some readers unfamiliar with language-related research may (quite reasonably) find it hard to believe that there is anything systematic about this process; after all, phrases like “that’s a big dog!” might mean very different things to different individuals: someone who has grown up around German Shepherds probably has a very different idea of what a “big dog” is when compared with someone 2 Though see the brief comments at the beginning of Chapter 8 regarding differences between the program we lay out in this volume and other common programs for “experimental syntax”; consideration of these differences and their consequences would likely be a good starting point for such an endeavor.

Language Faculty Science and Physics 

 389

who raises Beagles. Similarly, depending on the context, the identity of whatever is being referred to as “that” will certainly be different, and indeed, even within a specific context, it is potentially unclear which dog the speaker is referring to. We cannot, therefore, reasonably maintain that the mind has an algorithm for translating a sentence deterministically to a single particular “total” meaning. There are, however, certain constraints imposed by the sentence on the possible interpretations the sentence might correspond to. Regardless of the nuances of our interpretation, we nevertheless understand that “that’s a big dog!” states that whatever is referred to as “that” is being described as a “big dog”, which we in turn understand to be a “dog” that is in some way “big”, whatever that might mean to us. It would thus be very confusing, to the point of being a lie, for someone to tell us “that’s a big dog!” while pointing at a parakeet. Given this, we can see that the mapping between sentences and their meanings, while not totally deterministic, is not totally random either. There must therefore be some rules that govern this process. What kind of rules might these be? Some, surely, are simply memorized equations between words and concepts or categories, e.g., knowing that “dog” refers to a member of the species Canis familiaris. There must be more than simply these equations, however, as we do not only speak in one-word phrases, but also in much larger phrases and sentences. Forming such phrases and sentences requires combining words together. As a result, even if we have memorized the meanings of “that”, “is”, “a”, “big”, and “dog”, there must also be something that lets us know how to combine them to produce the meaning of “that is a big dog”; note that this cannot simply be “guesswork” based on the contents of the words, as we could also use the same words to create the sentence “a dog is that big” (which might be said while pointing at a roughly dog-size object). The meanings of these two sentences are different, even though they are “made of” the same words. Whatever rules we have thus must take into account the particular order the words come in, not only their meanings. A first guess might simply be that we just also memorize what each sentence means, but it is easy to see that such an approach is not feasible. As pointed out decades ago by Noam Chomsky, there is no finite set of sentences; even if we restrict ourselves just to English, for any given sentence, we could, for example, simply append another sentence to it with “and”, and do this over and over again, forming longer and longer sentences for as long as we like. This non-finiteness is not restricted to simple sentence conjunction; for example, we can iterate modifying expressions; for example, there is a well-known (among certain English speakers, at least) children’s song about a woman who swallowed a series of various animals in order to catch one another. We might describe her as, “the lady who swallowed a cow to catch the goat, which she swallowed to catch the dog,

390 

 Hajime Hoji and Daniel Plesniak

which she swallowed to catch the cat, which etc.”. We can go on this way for as long as we like; the lady can swallow a whole taxonomy book’s worth of animals, and we can keep adding on a modifying expression for each one. This repeatability property poses problems if we believe that rules for interpreting sentences are simply memorized sentence by sentence. First of all, if we assume that those rules are stored in the brain, then the brain would require infinite storage space, which, being finite in size, it does not have. Even if we could overcome this objection somehow, a more basic issue is that language is creative; we encounter new sentences all the time. Most people reading this chapter, for example, have presumably never heard or seen this sentence before, and yet you are (we hope) able to understand it. Even if someone has encountered this particular sentence before, both common sense and basic logic tell us that there must be some sentences that are new to them. The problem of non-finiteness recurs again here; if there are infinitely many understandable sentences, no one could possibly have had time to encounter all sentences before and memorize their meanings. There must, therefore, be some way to know what a sentence means besides having encountered it and memorized its meaning in the past. We could suppose that people are simply born knowing what all sentences mean, but this does not seem to be true; unless they have studied it as foreign language, children who grow up in exclusively English-speaking communities do not know what sentences of Japanese mean, and vice versa. Further, when children are born, it does not seem that they know what any sentences mean; certainly, we see them “learning” their language as they grow up. This fact draws our attention to yet another question: how do they learn their language in the first place, especially given that they cannot simply be just memorizing everything? We are left in a bit of a bind: children cannot possibly acquire all of their language-interpreting abilities simply by memorizing various things during their lifetimes, nor does it seem they are born with the ability to interpret strings of language fully formed. A compromise solution seems to be the only possible answer: children are born knowing certain things, but not others. While they are going to memorize things like particular word meanings, other pieces of knowledge need to be inborn. In particular, children must be born with certain tools that are going to enable them to learn their language, including tools that are going to let them understand the way in which words are grouped into phrases and sentences in a predictable way (rather than just memorizing each sentence individually). As Chomsky points out, this sort of issue was discussed as early as Classical Greece by the philosopher Plato. In the first paragraph of the Preface in Chomsky 1986, “explaining how we can know so much given that we have such limited evidence” is referred to as “Plato’s problem”. As Plato recognized, it cannot just be that we have extrapolated much from little data, because that simply leads to

Language Faculty Science and Physics 

 391

the question of how we know how to extrapolate. If we consider carefully, it is clear that one cannot “learn how to learn”. This is true simply because of what “learning” is; if a person could learn to learn, then that person would evidently already know how to learn, leading to a paradox. The ability to learn must therefore be pre-existent, and language learning should be no exception. Indeed, there must be quite a few things infants and children unconsciously know to acquire their language(s), but for our purposes, the most important such thing will be whatever knowledge they are born with that allows them to get around needing to memorize each sentence. Considering the examples we have just seen, the basic answer seems to be that there is something about the order of the words in a sentence that can be used as a key to how that sentence “maps” onto its meaning. ‘Iguanas eat leaves’ is a statement about the diet of iguanas, whereas ‘leaves eat iguanas’ is a (somewhat more terrifying) statement about the diet of leaves. The intuition here is that the word before ‘eat’ expresses what does the eating, and the word after ‘eats’ expresses what is eaten. This detail of information, however, cannot be what children are born knowing: different languages use different word orders to convey the same information (Japanese, for example, usually puts the verb last, giving us Iguwana-wa happa-o taberu ‘iguanas leaves eat’). Even in English, which word comes first is not always a reliable clue. Consider, for example, the sentence ‘leaves, iguanas eat, but fish, they don’t’: here, ‘leaves’ actually comes before both ‘iguanas” and ‘eat’, but ‘leaves’ is still what is eaten. What seems to be important for the interpretation of sentences is not so much the linear order of words, but the basic “structure” to which those words correspond. That is, regardless of whether ‘leaves’ appears before or after ‘eat’, the important part is that it occupies what is traditionally called the “object” position. Further, as English speakers, we have learned various basic templates, such as ‘X verb Y’ ( e.g., ‘iguanas eat leaves’) and ‘Y, X verb’ (e.g., ‘leaves, iguanas eat’), and that either one can be understood as a structure where Y (=‘leaves’) is the object of the sentence and X(=‘iguanas’) is the subject. The Japanese speakers have presumably learned similar things, e.g., ‘X Y verb’ being able to correspond to X being the subject and Y being the object, just as English speakers have learned this about ‘X verb Y’. Of course, the subject-verb-object structure is just one among many structures that our languages make use of; our primary interest is thus not limited to categories like subject and object. Rather, what seems to be most striking is the fact that there is an abstract structural representation that can contain/represent such categories at all. The basic intuition of the account just given is that there is some intermediate level of representation between sentence as pronounced and the sentence as interpreted, where categories like “subject” and “object” are assessed according

392 

 Hajime Hoji and Daniel Plesniak

to learned rules. The existence of such an intermediate level solves the learning problem we have been discussing: the existence of the intermediate level and the way in which sentence structures can be represented there are pre-known. Other things, like the meanings of individual words or the rules for identifying which structure corresponding to which pattern of words, are learned based on experience. This allows for different individuals to learn different languages; indeed, under this conception, different languages are simply different learned systems for mapping between sentences and meanings via this shared inborn intermediate level. As such, learning is possible, but it starts from a core base of knowledge that allows us to overcome the difficulties summarized in Plato’s problem. If we want to investigate what underlies the human ability to map between sentences and meanings, we therefore ought to be studying the properties of this intermediate level. Whatever these properties are, they will have to allow for the infinity of possible sentences addressed above. Let us return to that issue in a somewhat more technical sense. We may say that language exhibits what is termed “digital infinity”, meaning that language makes use of various discrete (“digital”) elements (words, syllables, letters, etc.), which are finite in number, but can create an infinite variety of different sentences from these basic elements. Accounting for the digital infinity of language is what Chomsky terms “the Galilean challenge”, inspired by Galileo’s musings on the question of how the few letters of the alphabet could be variously combined to represent all the thoughts of all human beings. Considering our intermediate structural level, the Galilean challenge implies we cannot simply be born with all possible structures pre-formed; there must be some way of building them “on the fly”, to account for new sentences. Chomsky’s solution is to hypothesize that, somewhere within the language faculty (which, as noted, is just the name for the assemblage of mental modules that allows for the translation between sentences and meaning) there exists a structure-building module. Chomsky calls this structure-building module the “Computational System” (CS), as it takes pre-existing elements and/or pieces of structure and combines them to “compute” larger structures. These structures are formed using an operation that Chomsky calls “Merge”, which is simply a set-building operation that takes two objects and combines them into a set. These elements can be “atomic”, like words, or they can be elements previously built by Merge, giving the CS recursive capabilities; that is, the CS can apply computations to the outputs of its previous computations, meaning that an unbounded number of structures can be built. If we take the example of our ‘iguanas eat leaves’, we can imagine the derivation proceeding as follows: first, we combine ‘eat’ and ‘leaves’ together to create what is traditionally called the ‘predicate’ of the sentence; in set notation, this is simply {eat, leaves}. Now that we have the ‘eat leaves’ set, we can combine it

Language Faculty Science and Physics 

 393

with the subject to derive the full sentence, {iguanas, {eat, leaves}}. We can keep going if we want: suppose we want to say, ‘iguanas eat leaves, but skinks eat insects’. We can follow the same procedure as before, building first {eat, insects}, and then {skinks, {eat, insects}}. We then combine this second sentence (technically a “clause”) with ‘but’, {but, {skinks, {eat, insects}}}, and then combine that with our {iguanas, {eat, leaves}}, yielding {{iguanas, {eat, leaves}}, {but, {skinks, {eat, insects}}}}. Long series of nested sets like this can be hard to read, so they are often represented by tree diagrams as below: {{iguanas, {eat, leaves}},{ but, {skinks, {eat, insects}}}} {iguanas, {eat, leaves}} iguanas

{eat, leaves} eat

leaves

{but, {skinks, {eat, insects}}} but

{skinks, {eat, insects}} skinks

{eat, insects} eat

insects

We can keep merging new structures together endlessly, creating a new structure every time, solving the Galilean challenge of digital infinity. The structures built are then be handed off both to modules that translate them into a pronounceable form (using rules acquired by experience, but also possibly inborn principles) and to modules that interpret them (this is presumably largely based on inborn principles, but it also takes into account things learned by experience, such as the meanings of specific words). Chomsky’s theory of the CS thus provides an account of the human ability for language that is compatible with the various considerations discussed above. Many questions, however, arise based on this theory of the language faculty. Most basically, is it even correct, and if so, how could we demonstrate that? If it is correct, then we will also want to know more about the CS, the structures it builds, and the rules behind how they are both pronounced and interpreted. As noted, the structures the CS builds are hypothesized to serve as the intermediaries between external “form” and internal “meaning”; as such evidence for the CS and its properties would presumably be found in the way(s) in which humans map between form and meaning. More specifically, the CS ought to manifest itself in terms of rules and/or constraints on how sentences can be interpreted. For example, perhaps certain sentences might only be interpretable in a certain way that would be unexpected if Chomsky’s theory was not correct.

394 

 Hajime Hoji and Daniel Plesniak

Given that the CS is argued to be an inborn component of the human language faculty, these “CS-constraints” on the form-meaning mapping ought to be universal to all humans, regardless of language. There has been much speculation about “linguistic universals” for some time, but as mentioned above, we are not interested in just any universals, only those that would be unexpected if there was not a Merge-based CS. As noted, we want to find constraints on form-meaning mappings that are predicted only if we assume that sentence structures are built in the way exemplified above, so what might such predictions be? First, it is crucial to note that, because of the order in which various elements are merged, not all words have co-equal relations to one another in a given structure. For example, in ‘iguanas eat leaves’, ‘eat’ and ‘leaves’ form a set of their own; we term such a grouping of words a “constituent”. ‘iguanas’ and ‘eat’ do not form a constituent, as there is no set that includes them to the exclusion of ‘leaves’, so we might expect that ‘eat’ and ‘leaves’ have some sort of close relationship that ‘eat’ and ‘iguanas’ do not. Further, because ‘iguanas’ is combined with the {eat, leaves} set, this means that it is combined with a constituent that contains ‘leaves’, so, as the tree diagram suggests, we can think of it as being “higher” or “less embedded” than ‘leaves’. This kind of structural relationship is called “c(onstituent)-command”. The notion that ‘iguanas’ c-commands ‘leaves’ is indeed unique to Chomsky’s theory (or at least to theories quite like it); there is no way to define c-command without assuming structure like the kind built by Merge. One way to verify the existence (and probe the properties) of the CS is thus to find effects on sentence interpretation that are somehow based on c-command; if there was no CS and thus no c-command, then such effects would certainly not be expected to exist! One early case of claimed “c-command effects” was observed by Tanya Reinhart (who originated the term “c-command”), who investigated the availability of various types of sentence interpretation. To take just one particular such interpretation (which has been of crucial use in this volume) consider the sentence ‘every student praised his teacher’. (Let us assume we are in an allboys school, so that the use of ‘his’ is appropriate.) There are two different ways we might interpret ‘his teacher’ here. First, ‘his teacher’ might be a specific person’s teacher; maybe we have been talking about John, and we are noting John’s teacher was exceptionally nice to everyone, such that all the students praised John’s teacher. In such a context, one might well say ‘every student praised his (=John’s) teacher’. This usage, which we can call the “referential” interpretation of ‘his teacher’, however, is probably not the one most readers first thought of when reading the sentence. The more obvious reading is that each student praised that student’s own teacher, John praised John’s teacher, Bill praised Bill’s teacher, etc. Essentially, in this usage, the interpretation of ‘his teacher’ varies across each member of the group denoted by ‘every student’. In that sense, ‘his’ serves as

Language Faculty Science and Physics 

 395

a “bound variable”, an entity getting its interpretation from a quantificational expression like ‘every student’. In ‘every student praised his teacher’, ‘every student’ is the subject and ‘his teacher’ is the object. Now consider flipping this relationship: ‘his teacher praised every student’. For many, perhaps most, individuals, the bound variable reading is no longer possible; that is, the sentence cannot mean that John’s teacher praised John, Bill’s teacher praised Bill, etc.; instead, it must have the referential reading, namely that someone specific’s teacher praised each of the students. This is a somewhat unexpected contrast: why should being a subject vs. being an object affect what ‘his teacher’ can mean? Adding to the mystery, subsequent work has shown that this distinction seems to exist in most, likely all, languages (albeit with certain caveats to be discussed below). If so, then we may have a linguistic universal on our hands; the question then is whether it is uniquely predicted by hypotheses involving c-command. As Reinhart took pains to show, this is indeed the case. What we can note from our iguana example is that subjects c-command objects but not vice versa (at least if we assume that the relevant structures are always roughly the same as the one given there, even in other languages). Reinhart thus proposed that, in the CS-structure corresponding to the sentence in question, bound variables must be c-commanded by the quantificational expressions that bind them. As such, bound variable objects with quantificational subjects are possible, but the reverse is not true; since the object does not c-command the subject, a quantificational object cannot bind the subject. This account makes many further predictions beyond just one sentence or sentence type; any time the quantified expression does not c-command the intended bound variable, the relevant bound variable interpretation should be impossible. Reinhart demonstrated this across a variety of sentence types, and also demonstrated that other accounts, such as those based on word order alone, do not make the correct predictions. To the extent that we accept her evidence and argumentation, the distribution of bound variables does indeed seem to be explainable only by considerations of c-command, thus providing evidence for the existence of Chomsky’s Merge-based CS. Argumentation in favor of the CS has often proceeded along this model, as judgments about the acceptability of sentences with various interpretations frequently seem to depend on considerations of constituency and/or c-command. Further, if we are confident that the availability of interpretative phenomena like bound variables coincides with the presence of certain structural configurations, we can then use their (un)availability as a diagnostic tool, a sort of “probe” for the structures of new sentence types. Additionally, if a type of interpretation suddenly becomes available or unavailable in surprising places, this can point to new prop-

396 

 Hajime Hoji and Daniel Plesniak

erties of the CS/language faculty. For example, most people do not accept bound variable readings of sentences like ‘every student praised him’ (which would yield an interpretation like “every student praised himself”), even though ‘him’ is the object and ‘every student’ is the subject; we can thus conclude that there must be some additional constraints on bound variables besides just c-command. However, there is a problem lurking just below the surface here. Namely, the patterns of judgment claimed to demonstrate to c-command-based constraints often do not come out as “cleanly” as predictions would suggest. That is, people often have judgments that do not appear, at least superficially, to coincide with c-command relations in the way expected if a Reinhart-style approach were correct. As noted above, sometimes, when these exceptions are limited and systematic, they can lead to fruitful refinements of the original theory. Unfortunately, as has increasingly become clear over the years, many of these exceptions are not particularly limited, and even the most basic of datapoints can be called into question. For example, not everyone even agrees that phrases like ‘his teacher’ in ‘his teacher praised every student’ cannot have a bound variable interpretation; for these individuals, the sentence can indeed mean something like ‘every student was praised by his own teacher’, exactly what Reinhart’s account claims should be impossible. That individuals might vary in their judgments, even if they speak the same language, is not unexpected. As discussed earlier in this section, our language faculties contain a mixture of inborn traits and traits acquired by experience. When children grow up in wildly different language communities (e.g., monolingual English vs. monolingual Japanese), the differences in the traits acquired due to experience are obvious; they speak in ways that are not mutually intelligible. Even within a given language community however, each individual is exposed to slightly different linguistic stimuli, which we would expect might result in the development of slightly different “languages” in their minds (what Chomsky calls each person’s “I-language”).3 A child may pronounce certain words differently than their parents, for example, or use phrases in idiosyncratic ways their siblings do not. Even the same person at a given time might have a different I-language; we can learn new words, for example, and other, subtler properties can change or shift too. There is thus no reason to expect all individuals to have the same judgments on all sentences, and indeed, some variation is reasonably expected. Such considerations, however, do not allow us to simply dismiss the issues at hand. To serve as evidence for the existence of a universal CS, the purported c-command effects must themselves be universal. They are not directly comparable to say, the different ways people might pronounce “tomato”; such pronun-

3 Chomsky intends “I” of “I-language” to express “internal”, “individual”, and “intensional”.

Language Faculty Science and Physics 

 397

ciation is acquired based on experience, whereas sensitivity to c-command is claimed to be inborn. If it were learned, that would defeat the entire purpose of postulating an inborn CS in the first place. So, what is a CS-believer to do? The intuition behind the LFS program as presented in this volume is that, while c-command-based constraints on the form-meaning mapping are indeed universal, their influence on a given individual’s judgments is sometimes masked by I-language-specific factors (which are affected by experience and thus manifest in a non-universal manner). As stated above, exceptions to c-command-based theories can be accommodated if they are either limited or systematic; while variation on sentences like ‘his student praised every teacher’ does not appear to be particularly limited, what we have shown in this volume is that such variation is, in fact, systematic. That is, the I-language-specific factors that interfere with the observation of c-command effects are not random; they are, in fact, highly predictable, assuming one has enough data about the individual in question. Precisely how these interferences can be predicted is a complicated and technical issue, and has occupied the bulk of this volume, so we will not repeat them here. To give a brief intuition, however, such interferences are caused by certain I-language-specific properties, which can be diagnosed via certain tests. These tests involve having the individual make judgments about other sentences besides the main sentence in question; depending on their judgments on these other sentences, it is determined whether or not the relevant interfering factors are affecting their judgments on the sentence(s) in question. If such interfering factors are not affecting their judgments, then an individual’s judgments on the main sentence in question can be predicted with 100% accuracy, a claim we have supported with extensive data (see Chapters 5–7, for example). As such, while we cannot consistently predict how an individual will judge a single sentence in isolation, we can consistently predict the possible patterns of judgments that individual may have on certain sets of sentences; that is, the way in which judgments are implicationally correlated with one another, with certain sentences serving as “predictors” for other sentences. These correlations are themselves uniquely reliant on c-command for their explanation, thus giving us both an effective demonstration of the existence of the universal CS as well as a powerful probe into the properties of the language faculty.

1.3 Chapter Overview Having summarized the basic idea of LFS as presented in this volume, we now return to the core goal of this chapter, the comparison of LFS and physics. To establish this comparison, we precede in the upcoming sections as follows:

398 

 Hajime Hoji and Daniel Plesniak

First, we describe the basic “scientific” features of LFS (Section 2), which stem from the application of the scientific method to a mental organ (or a set of mental organs), the language faculty. In this basic accounting, LFS looks “on paper” much like a “hard” science like physics (albeit in a much less “developed” state than modern physics is), because of its adherence to the strict hypothesis-prediction-experimentation procedure, as well as related notions like “replication”, “testability”, and “noise control”. However, there are several areas of key interest where these notions, so familiar from fields like physics, may seem less than obvious when considering LFS. The first such area (Section 3) is the relationship between the hypotheses and predictions of LFS and the results of its experiments. While LFS and physics are similar in that facts are discovered in a process that is dependent on hypotheses and predictions, LFS differs from physics, in that in physics, much is “observable4” without precise theories; what is observable in this way is much more limited when dealing with the language faculty (Section 3.1). This is partially because LFS focuses primarily on the Computational System (CS) of the mind, rather than the entirety of the mind itself, in contrast to physics, which takes as its domain the entirety of the physical universe. This lack of “observables” may give rise to the notion that certain hypotheses about the CS are unfalsifiable/untestable. There is a grain of truth to this, but only to the extent that other “hard core” hypotheses are untestable, or at least not usually tested, in their respective research programs. As we point out, this is fundamentally no different than physics (Section 3.2). It is the function of any scientific endeavor to generate auxiliary hypotheses that bridge from the hard-core hypotheses to testable predictions. As such, the fundamental relationship between hypotheses, predictions, and experimental results is the same in physics and LFS, but LFS differs from physics both due to its immaturity and to the mind-internal nature of its object of inquiry, which adds a barrier of observability that was not present in the infancy of physics. A second area of interest is the difference between measurements in the two fields (Section 4). Physics, at its core, deals with various quantities, while LFS is fundamentally categorical, dealing with (correlations of) qualitative judgments.

4 Note that we are implicitly holding here that a “fact” and an “observation” are two distinct concepts. One may of course say that “we observed X” is factual in the sense that X was really observed, but until we have done the appropriate tests, we do not know if that fact is merely “incidental”, an artifact of having done our observations at a certain time or place, or whether it represents something more than that. Specifically, the “something more” which constitutes a fact for us must have a degree of reliability and repeatability; that is, not just that X was observed, but that if one follows all the necessary steps, X will be observed. We are thus focusing on this type of repeatable and reliable fact, rather than a “fact” in the sense of an observation being factual.

Language Faculty Science and Physics 

 399

As we make clear, the latter can be predicted in as definite a manner as the former (Section 4.1). As noted above, however, LFS deals only with a specific sub-part of the mind, so whether a given measured value has relevance for a given prediction is somewhat more complicated in LFS than in physics (Section 4.2). This concern is in fact the basis of this entire volume, and is an issue with which LFS is still very much grappling. As will be noted, however, such issues are not entirely foreign to physics; nevertheless, they are of far more central concern at this early point in the development of LFS than they are in modern physics, highlighting the former’s status as a still maturing field. Another area of interest is what counts as “demonstrating” a result as well as what counts as “replicating” a result (Section 5). In terms of the former, LFS and physics are similar, but because physicists are not (generally) interested in performing experiments on themselves, they do not have the same sort of self vs. non-self distinction in experiments as is found in LFS (Section 5.1). Indeed, as addressed in Chapter 4, the object of inquiry in LFS is internal to each human (and common among all members of the human species), and LFS requires the use of “correlations”; these have some resemblance to the laws of physics, but also appear to differ from physical laws in certain ways. Specifically, these correlations are designed to render the results of experiments conducted within different individuals meaningfully comparable with one another, which is not obviously a concern in physics. As such, replication in LFS is fundamentally the replication of correlations. However, we show through various analogies that this is not at all alien to what goes on in physics (Section 5.2). Finally, noise control is an area where similarities between LFS and physics are clear, but where the two also have key differences arising from the same distinction already discussed multiple times, namely that the CS is only a part of the mind while there is no (known) equivalent to the CS in the physical universe (that is, a part of the physical universe which works according to a specific formal system in contrast to other parts that do not). This affects not only what forms of noise control are pursued, but also how basic research rigor and prediction-deduction are pursued as well (Section 6). The ultimate consequence is that LFS requires far more “replication” than is necessary in physics, in the form of replication across different (types of) I-languages. In particular, LFS has to contend with the presence of I-language-specific non-CS noise,5 that is, the effects from systems, either within or otherwise interacting with the language faculty, which are not formal in the same way that the CS is. The type of noise disallows hypotheses in LFS from

5 See Chapters 4 and 5; “non-CS” may be understood as “non-formal”, as in Chapter 5’s Non-Formal Source (NFS).

400 

 Hajime Hoji and Daniel Plesniak

being as exclusively “formal” as those in physics, even though the object of inquiry is a formal system in both. The underlying existence of a formal system enables the application of the scientific method in both LFS and physics, though the formal system about which LFS seeks to learn is shrouded in external noise in ways that are not traditionally found in physics. While immature in many ways, in this respect, LFS is working at the forefront of scientific exploration along a particular frontier, namely the domain of the “mental world”. As it tackles problems in this domain, facts that have significance far beyond LFS, perhaps even providing significance for physics, maybe discovered (Section 7). The better we can understand the nature of the mental world, the better we can understand the formal properties undergirding all scientific endeavors, suggesting the possibility that the work of LFS may one day play a central role in some future scientific revolution.

2 The Scientific Basis of Language Faculty Science Language Faculty Science is a deliberately and self-consciously scientific endeavor. We may take its basic features, as articulated in the previous chapters of this volume, to roughly correspond to the following list of notions (as applied to the language faculty): (1) a. b. c. d. e. f.

hypotheses predictions and prediction-deduction experiments/observation/measurement replication testability (as opposed to compatibility) noise control

(1a) expresses LFS’s commitment to the evaluation of (definite) first principles. As set forth in the introduction, the scientific method requires the presence of a formal system; in order to be licensed to use such a method, one must first postulate such an underlying formal system exists, necessitating hypotheses that attempt to characterize the basic properties of that system. (1b) is a necessary follow-up to (1a); once hypotheses are established, any would-be investigator must determine what the observable consequences of those basic hypotheses are. Given that the hypotheses are definite in nature, the predictions must be definite as well. Note that definite need not be taken to mean deterministic as to the outcome of a single

Language Faculty Science and Physics 

 401

measurement; the approach laid out in previous chapters is in fact correlational but nevertheless definite. More specifically, for a given sentence-interpretation pair,6 there is no single predicted judgment; in fact, it is anticipated that there will be variation between individuals, and even within an individual, as to whether the pairing is accepted. Nevertheless, the predictions of LFS are definite, in that specific correlations between judgments are deduced, such that space of possibilities for overall patterns of judgments that an individual might have is restricted from the space of all conceivable such patterns. That is, there are certain patterns of judgments that might be logically possible but which we predict never actually occur, due to the particular properties of the CS and/or language faculty. Such correlational predictions lie at the heart of the equations of fields like physics as well: Newton’s Second Law, f=ma, for example, is a definite correlation said to hold between force, mass, and acceleration. Specifically, it is a correlation because it does not predict in isolation the force, mass, or acceleration of a given body, but nevertheless limits the space of possible combinations of force, mass, and acceleration a given body can have. Likewise, the correlational methodology, as discussed in preceding chapters, does not make an exact prediction about a given individual’s judgment on any particular type of sentence in isolation. Rather, when we consider whether a given sentence can be interpreted such that it (in part) expresses a particular “meaning relation” (MR), this issue is ultimately only settleable by examining the interpretative possibility of other sentences. In essence, we do not predict the (im)possibility of a given MR-sentence pair, but rather, we predict an individual’s pattern of judgments on a given type of MR-sentence pair based on that same individual’s judgments on other types of MR-sentence pairs. This correlation is not an equation per se as there are no quantities to equate, but this distinction is a reflection of the difference between the objects of inquiry, the categorical CS of the language faculty and the quantitative physical universe. (1c) expresses the experimental nature of LFS, as well as the distance between what is measured, namely acceptability judgments, and what is hypothesized, namely various mind-internal structures and operations. By (1b), hypotheses and experimental results are linked via definite predictions, allowing experimental results to be meaningfully categorized as to whether they support or go against the hypotheses of (1a). Crucially, however, there is no point at which one can 6 For the purposes of this volume, such a pairing is between (i) a given string, say “every man sharpened his pencil” and (ii) an MR (“meaning relation”) like the bound variable anaphora interpretation of ‘his’ in that sentence, which we term BVA(S, every man, his) in previous chapters. (ii) represents a conceivable way in which (part of) the sentence could be interpreted, leaving aside whether such an interpretation is actually possible for the sentence. Having a speaker “judge” whether it is actually impossible or not impossible forms the core content of experiments in LFS.

402 

 Hajime Hoji and Daniel Plesniak

declare with certainty that a given hypothesis has been definitively proven true. (1d) expresses the need to repeatedly validate experimental results; as a given prediction survives repeated attempts at disconfirmation, confidence in the correctness of its basic hypotheses grows, though never to certainty. Indeed, because of the imperfect nature of experiments, it is not even possible to be certain that a given set of basic hypotheses is false based on the results of an experiment; there is always the chance that the experiment was set up incorrectly or suffered some internal error, resulting in erroneous results. Here again, replication allows greater certainty; if an observed result which seemed to contradict a given prediction cannot be reproduced, one may be licensed to believe it was the result of experimental error rather than a genuine disconfirmation of the hypotheses that generated the prediction in question. To be clear, such limitations is true of any experimental program of discovery and thus is not unique to LFS. Despite such limitations, it is a crucial property of LFS that its predictions (and thus its hypotheses) are testable, as expressed by (1e). That is, while we cannot exclude possibilities like there being some sort of “error”, our analysis can tell us, in a definite manner, whether the observed results are consistent with the basic hypotheses or are in contradiction with them. Such a formulation still leaves room for doubting the test in question. It is possible that the test was not rigorous enough, or the experimenter “got lucky”, in which case, we may not assign a “supporting” result much credence. Equally, perhaps the test was poorly designed or suffered some sort of failure, in which case, a “contradicting” result is not reliable. Repeatability of course helps to alleviate these concerns, but as noted, it can never fully overcome them. However, testability nevertheless differs from mere compatibility,7 in the sense that the results of the experiment are definite, no matter what we think of the likelihood that these results correspond to “reality.” If we only care about achieving compatibility, no definite test need be made; the task is to explain how obtained results can be understood to be consistent with hypotheses, not to determine whether the obtained results support or contradict them. While certainly there may be times when such approaches are necessary and useful, they do not, and should not, form the heart of a scientific endeavor. Finally, (1f) expresses the challenge with which LFS must contend, namely the various factors that can obscure the output of the formal system in question. The correlational methodology detailed in Chapters 4 and 5, and the rigorous use of attentiveness/comprehension-checking sub-experiments addressed in Chapters 6 and 8, (as well as in Hoji 2015), are both ways in which LFS attempts to overcome such noise.

7 See discussion in Chapter 1.

Language Faculty Science and Physics 

 403

At this point in time, knowledge of the CS is quite limited, the formal recognition of its existence only having begun a few decades prior. This contrasts greatly to laws of the physical universe, which have been studied extensively by the scientific method for centuries. Through this long process, a great many “facts” about the physical world have been accumulated (in the sense of physical predictions of a definite nature, well-substantiated by repeated experimentation, see footnote 3). In LFS, by sharp contrast, we are still at the stage of identifying basic “facts”, while simultaneously trying to establish an effective means for such “fact identification”. As such, it is not appropriate to compare the “results” of the two fields; rather, this chapter notes some of the relevant established “facts” of physics and aspects of the process by which they were established, and compares them with current attempts in LFS to establish facts by analogous means. In this sense, it is crucial to understand that one of the central goals of this volume has been to articulate how c-command detection can be used as a basis for “fact identification”. As such, if we wish to support hypotheses about the CS and the structures it builds, we must focus on the type of “meaning” that reflects c-command and whose (un)availability can be understood in a definite sense, i.e., clearly possible or clearly impossible. These types of meaning include the MRs of previous chapters. As discussed in those chapters, the focus on c-command directly leads to the need to a focus on Schemata (that is, sentence patterns), rather than individual sentences, and that the focus on definiteness leads particularly to a focus on ✶Schema-based predictions (predictions that certain MRs will be impossible with certain schemata). This is coupled with the understanding that there are factors other than c-command that can give rise to similar, perhaps identical, MR readings of a given sentence, requiring careful noise control to tease them apart, thus necessitating the correlational methodology. LFS, as described in this volume, is thus an attempt to use (a) hypotheses based on c-command, (b) predictions based on the unavailability of certain types of “meaning”, and (c) an experimental/analytical methodology to enable the predictable detection of c-command-based patterns; (c) in particular must be such that it can be used even if there is non-c-command-based noise that might obscure such patterns. Used together, (a)–(c) make it possible to test hypotheses about the CS of the language faculty, specifically by establishing “c-command-based” facts of the type discussed above (or showing that a given purported “fact” is contradicted by data). While this approach is highly specific to the language faculty, it flows naturally from the general aspects of science mentioned in (1), which, as noted throughout this section, might also be taken as basic core aspects of a field like physics. As will be shown, despite dealing with very different types of hypotheses, as well as being in different stages of maturation as fields, these unifying factors create parallels in the way in which research is carried out in both LFS and physics.

404 

 Hajime Hoji and Daniel Plesniak

3 Hypotheses, Predictions, and their Relationship to Data 3.1 Hypotheses and Observations In a certain sense, the notion from the previous section that LFS seeks to find MRs that can be used to make ✶Schema-based predictions, and in turn to establish facts, may seem backwards. In the typical conception of science, the process of hypothesization begins with observation of some phenomenon, inviting an explanation thereof. Of course, LFS has the observations like that of digital infinity and the mapping between forms and meanings, but these are of a very basic nature. Certainly, as the well-repeated anecdote of the “Newton and the apple” demonstrates, there were and are specific, observable phenomena in the physical universe about which one can form hypotheses (like objects falling towards the ground); in LFS on the other hand, it is generally true that “meaning-related”, and certainly “structural”, phenomena of language are generally not identified until after the relevant basic hypotheses have been proposed. Unless one is looking for such things, it is quite unlikely that one will notice that we make use of MR interpretations; further, because humans do not naturally produce unacceptable sentence-MR pairings, one would certainly not stumble upon the existence of said unacceptable pairings, much less see that an abstract structural relation like c-command seemed to be playing a crucial role in determining such unacceptability. Such investigation relies on the consideration of things that are not naturally occurring, and thus which cannot (or at least are very unlikely) be observed “by happenstance”. If one is looking to find out about the rules governing certain types of interpretation, on the other hand, then there is a reasonable chance of finding out things like the above, but to be engaged in such a search in the first place, one has to have already hypothesized that such rules exist (or at least might exist). As such, at a minimum, one will have had to first hypothesize that there must be some basic systematic rules governing how elements in a sentence are interpreted before one can begin to consider “non-naturally occurring” sentence-MR pairs, and thus make the relevant kind(s) of observations. Nevertheless, it is not the case that the hypotheses that predict a given phenomenon always follow the observation of that phenomenon in physics. Indeed, if that were the case in a very literal sense, physical theory would be essentially useless; we might as well just write down a list of all observations, and this would tell us as much as any theory. Laws, such as the aforementioned f=ma, may be based on observations; in the case of f=ma in particular, these might be observations such as that a “heavy” object requires more force to bring up to a certain

Language Faculty Science and Physics 

 405

speed than a “lighter” object, but once the law is formulated, it predicts the force required to accelerate any object of any mass. One may naturally quibble though that this is not a new “phenomenon”, but rather, new instances of the same phenomenon, namely the interaction of weight/mass and force. However, especially in modern physics, there are even more dramatic reversals of the observationhypothesis order. An often cited case is the discovery of the positron: the Dirac equation was initially hypothesized in order to derive the behavior of electrons, which are negatively charged, yet it was noted that it admitted a solution involving positively charged particles as well. Initial reactions were varied: some dismissed this as a fault of the equation that had to be ruled out by some other factor or a correction of the equation, while others attempted to identify the positively-charged “anti-electron” with some already known particle, such as the likewise positively-charged proton. Another response was to take the equation’s predictions literally and began to search for a new particle. This final camp eventually won out, discovering the positron, the positively charged anti-particle counterpart of the electron, and in the process inventing the notion of “anti-matter” that lies at the heart of much of modern physics. In this case, hypothesis and prediction, namely the Dirac equation and the positron solution to it, preceded the observation of anti-matter. In general, when the object of inquiry is a formal system, it is possible that two (related) types of things may occur: (2) a. Theory leads to a discovery of a new “phenomenon”. b. Establishing/building a detector for observing something not directly observable advances our understanding of the subject matter. The discovery of the positron certainly was an example of (2a), though it was less so an example of (2b); that is, though theory preceded observation, that observation was accomplished largely through already established means of detection. As such, rather than comparison with the discovery of the positron, a perhaps more analogous pair of LFS and physics examples might be as displayed in (3): (3) a. LFS: Merge and c-command detection b. Physics: General relativity and the detection of gravitational waves As noted in chapters throughout this volume, (3a) is certainly an instance of both (2a) and (2b). (2a), in the sense that the various correlations between ✶- and ok-schema judgments in various MR’s were only observed after the concept of c-command was formulated; indeed, there would essentially have been no salient reason to check judgments regarding any MR in any schema unless there were a

406 

 Hajime Hoji and Daniel Plesniak

theory suggesting that the availability of those MRs in those schemata were significant. (2b) in the sense that, in attempting to achieve such detection, it has been necessary to contend with various issues related to noise control, which has led to the development of the correlational methodology described in previous chapters, a tool for detecting syntactic structure. Turning to (3b), Einstein’s theory of General Relativity, like Chomsky’s postulation of the Merge-based CS, provides a framework which (at least in conjunction with other hypotheses) leads to the deduction of numerous new predictions. In particular, various new phenomena, such as black holes, were predicted, conforming to (2a), yet black holes can be “seen” by telescopes. (More specifically, they can be inferred by their property of blocking other bodies from being seen.) While the type of telescope required to do so is certainly a sophisticated one, such telescopes are not fundamentally different from other telescopes used in modern astronomy in that they rely on detection of electromagnetic waves, such as light or radio-waves, emanating from distant astronomical bodies. On the other hand, Einstein’s theory also predicted the existence of gravitational waves, subtle8 ripples in spacetime caused by motion of massive objects relative to one another. Such waves were first detected about a century after first being predicted by Einstein, via a detection method specifically designed to measure distortions in spacetime on a scale small enough so that gravitational waves could be measured. The first such observation was made by the Laser Interferometer Gravitational-Wave Observatory (LIGO), which in effect used the slight phase-shift in light caused by a sub-atomic-sized change in spacetime due to a gravitational wave passing through the Earth (see Abbott et al 2016). What is noteworthy is that this LIGO-initiated observational device for detection of gravitational waves has opened up a new way to observe what was not observable before; massive objects like black holes, though they do not emit light, do emit gravitational waves under the correct conditions, and so LIGO allows them to be observed directly, rather than the indirect methods required by light-based telescopes. It is worth noting, however, that General Relativity by itself did not make obvious either the existence of black holes or the methods by which gravitational waves could be detected. To date, two separate Nobel prizes in physics, both 2017 (Rainer Weiss, Kip Thorne, and Barry Barish, for the development of the theory behind LIGO as well as its implementation), and in 2020 (co-awarded to Roger Penrose, for mathematical work deriving black holes from General Relativity) have paid tribute to what is but a fraction of the large amount of research that has

8 At least by the time they reach us on Earth.

Language Faculty Science and Physics 

 407

been involved with this decades-long theoretical and experimental enterprise. We can (optimistically) analogize this to the situation in generative linguistics, wherein Chomsky’s initial formulation of Merge (which allows us to derive c-command) has been and continues to be used by other theoreticians and experimentalist to try to establish what its empirical consequences might be and how to detect them, as the LFS program endeavors to do. The analogy is not perfect; linguists had glimpses of c-command effects in “early days” of generative grammar; this was largely because intuitions about things like syntactic constituents occurred to linguists before the articulation of Chomsky’s basic idea that Merge creates purely hierarchical mental representations, which derives why these constituents exist. Among the various notions expressing different structural relations was a version of “c-command” that had essentially the same formal definition as the “c-command” that is based on Merge. Reinhart (see especially Reinhart 1976 and Reinhart 1983) developed the concept in part due to Chomsky’s notions of abstract syntactic structure, but this earlier notion of “c-command” was crucially tied neither to purely hierarchical mental representations nor to binary branching; it thus differed somewhat from the version presented earlier in this chapter, which is formulated under Chomsky’s Merge-based conception. Further, some MRs were indeed discussed in relation to, i.e., as a reflection of, some structural relations – often in the context of choosing which of the possible structural relations were to be adopted. This was especially true of both scope-related phenomena, of which the MR DR is a sub-type (DR being a crucial ingredient in the c-command detections discussed in previous chapters), and pronominalization, itself an instance of the MR Coref, (which is also a crucial ingredient in current LFS c-command detections as discussed in previous chapters). See for example Langacker 1969, which formulates one of the earliest versions of what would later come to be known as “c-command”. Given the above considerations, it seems that physics is in fact “ahead” of LFS in terms of (2a–b); rather than observation preceding theory in physics but not LFS, we have seen that physics contains clear historical cases where theory preceded observation (e.g., positrons and gravitational waves). On the other hand, during its relatively brief history, theories in LFS have often followed observations made “pre-theoretically” (e.g., Reinhart’s formulation of c-command, which was not originally derived from first principles like Merge). While physics has had instances of observations that were completely novel and only encountered due to their existence first having been deduced from hypotheses, this has never happened in a “pure” way in LFS. This perhaps reflects physics’ maturity as a field; Einstein had a rich base of established facts on which to base his hypotheses, and as such, the sophisticated theory that emerged had a rich body of entail-

408 

 Hajime Hoji and Daniel Plesniak

ments. LFS does not have such body of facts, and as such, it has not yet been able to articulate a theory with anywhere near the predictive power of Einstein’s.9 9 To elaborate a bit further: as the field of physics began, observations were made in increasingly accurate ways, but in many cases without what we would normally call ‘theory’ in modern terms; nevertheless, these observations were considered facts. Indeed, these initial observations have proved quite accurate; while we have “complicated” the picture somewhat since the initial days of physics, the world as described by Galileo’s contemporaries and their various observations is still essentially compatible with even the most advanced of modern physical theories. The latter are based on far more accurate and precise measurements than were Galileo’s, and thus certain things that would have seemed obvious “facts” centuries ago have to be amended, but nothing gets really “turned upside down”, so to speak. This is not to say there were not various errors made, but these were generally corrected simply with more observation, e.g., the first telescopes were rudimentary by today’s standards, and the data acquired using them was not always reliable, but this was rectified simply by building better telescopes. Thus, for physics, pre-theoretical observations were essentially sufficient for fact-establishment; it is only once we reach the rather intricate level of physics found in say, the 20th century, that theories start to become far more obviously necessary for fact-establishment. As such, the fact that people like Einstein were able to create theories that predicted never-before-seen kinds of phenomena is a reflection of the centuries of fact-establishment via systematic observation that provided the “empirical basis” for such theories. In LFS, on the other hand, there has been no analogous way to establish facts, and the field does not have the centuries of history that physics does, so there are essentially no “facts” to base such a predictive theory off of (or at least very few of them). To clarify, the basic scientific method described in (1) forces us to deal with definite predictions and definite experimental results and to replicate definite experimental results as predicted. Facts in LFS are, for this reason, definite and replicated experimental results predicted by our hypotheses about the language faculty, including specifically our hypotheses about universal properties of the language faculty. It is in this sense that we hold that there are, as of yet, few “facts” established in LFS. To establish facts in the above sense, results in LFS must be definite (for us, in the sense of “X is definitely unacceptable”), deducible (from hypotheses about the language faculty), and replicable (ultimately, universally across all (relevant) I-languages). Previous research has sometimes met one or two of these criteria, but rarely all three: many patterns of judgment that have been deduced and replicated are somewhat vague, some definite and replicable patterns have an unclear relationship to our basic hypotheses, and still other definite and deduced patterns fail to replicate in a consistent manner. In this volume, we have primarily been concerned with this last case, and our main claim is that such failure to replicate comes from inadequate consideration of the sources of a given judgment. This leads us to try to “home in” on a particular source, which necessitates the correlational methodology (which allows us to diagnose which potential sources are operative for a given individual at a given time considering a given sentence). We expect, however, that advancements in the area of replication will lead to (and indeed necessitate) advances in making vaguer predictions definite and in determining how a given prediction can be rigorously deduced. When all this is done, then many of the “pre-fact” observations made over the years can be upgraded to full facts in LFS (which in fact we have been trying to do to an extent in this volume, especially in Chapters 5–7). Despite these current efforts, we must still concede (and indeed want to emphasize) that:

Language Faculty Science and Physics 

 409

Before closing this section, it is worth further commenting on the distinction between observations and facts, which the above discussion has touched on to a degree. Facts are established as the result of hypotheses being supported when predictions survive rigorous experimental testing. Facts are thus part of what must be accounted for, explained, etc., by any new theory that is proposed. If a new theory does not account for, or at least manage to be compatible with, previously established facts, then it is generally not worth adopting, even if it might explain certain observations. For LFS under the correlational methodology, facts take the form of predicted correlations of schematic asymmetries based on the hypotheses we have; the addition of new hypotheses or modification of the existing hypotheses should therefore be evaluated in terms of whether its empirical consequences are better than before. Historically, scope-related observations (of which observations about MRs like DR are a subtype) have (generally) been about what scope relations are possible or preferred, but not about what scope relations are impossible, in certain “configurations”. There were thus numerous observations, but few, if any, “facts”, at least as that term is to be understood in the context of LFS. Indeed, we may speculate that the long tradition of scope-related discussion seems to have been guided (at least) in part not so much by considerations of the CS but by considerations of logic;10 this can be seen within a long tradition of philosophy of quantifiers, stretching back to at least Aristotle. This distinction highlights another difference between LFS and physics, which we have already briefly discussed, namely that (at least modern) physics assumes that physical law underlies, at some level, all observable physical phenomena, whereas LFS deliberately assumes that the CS is only a portion of the mind, and thus underlies only some observable mental phenomena. As such, while both physics and LFS utilize data in a way which is crucially driven by theory, LFS requires theory to discriminate between those phenomena which are a reflex of a property of the CS and those which are not, whereas physics needs no such heuristics. Of course, for all practical purposes, a phys-

(i) LFS is far behind physics in terms of fact-identification activities/results. (ii) LFS is far behind physics in terms of prediction-deduction activities/results. and fact-identification and prediction-deduction are actually (almost) inseparable in LFS, but they were separable in physics unless/until we consider “modern-day” physics. 10 Or perhaps we may say “the faculty of logic”; even if practitioners have not recognized it as such, certainly the field of logic relies primarily on our intuitions of what is or is not a valid deduction, which seem to be universal across the human species; there must therefore be some inborn system that allows for such intuitions to arise. With this understanding, some of what has been termed “non-formal source” in Chapter 5 may in fact be rather a different formal source, namely whatever formal system is the “CS-equivalent” in the faculty of logic.

410 

 Hajime Hoji and Daniel Plesniak

icist who is not pursuing a “theory of everything” must discriminate between phenomena/effects relevant to whatever the current topic of investigation is and those that are not relevant (and must control for the influence of the latter if it causes noise), so there is a degree of overlap between the two fields in this sense. However, a physicist is free to observe some novel natural phenomenon, declare, “I want to find the physical explanation for this”, and to feel confident that such an explanation does indeed exist; those studying the CS cannot be so certain that every novel language phenomenon has a CS-based explanation. As such, physics can make much more liberal use of mere observations, rather than facts, but LFS cannot so easily do so; Newton can be confident his apple falling reflects a physical law, whereas an LFS practitioner cannot easily be sure whether given quantifier scope intuitions reflect CS properties or not.

3.2 Hard-core and Auxiliary Hypotheses How can the practitioners of LFS be sure, though, that there really is a Mergebased CS? There may be some sense in which it seems that, unlike what is done in physics, practitioners of LFS are merely “assuming” the validity of the most central tenet of the field, rather than testing it experimentally. However, both LFS and physics share a distinction between, as Lakatos (1978) puts it, “hardcore” and “auxiliary hypotheses”. That is, both involve certain “research programs”, that have at their heart certain hypotheses, the “hard core” ones, which are not discarded no matter what experimental results obtain. Rather, these hypotheses are used in conjunction with “auxiliary hypotheses” to make testable predictions, and it is these auxiliary hypotheses that are revised and/ or discarded when experimental results do not come out as predicted. Of course, at times, even the “hard core” hypotheses are dropped, but this does not come about due to the results of any particular experiment, but rather an accumulated weight of data and insight presenting a convincing enough case that a fundamentally different research program is necessary. Such “revolutionary” events are rare, at least in a stable field; an example would be the change from “Newtonian” to “modern” physics, which occurred only centuries after the former was established. To exemplify the distinction between hard-core and auxiliary hypotheses, let us consider further Newtonian physics. In the Newtonian research program, the hard-core hypotheses include Newton’s three laws of motion, as well as the universal law of gravitation. Auxiliary hypotheses might include, for example, hypotheses about the existence and “positions” of certain planets, along with their specific properties, such as their masses. Note that the former group of

Language Faculty Science and Physics 

 411

hypotheses is insufficient to make any definite predictions; only when specific values are supplied by the latter group are such predictions made. If actual observations, say about an orbit of some planet, for example, were not in line with all these hypotheses, the Newtonian physicists did not generally try to modify the hard-core hypotheses, but they did try to modify auxiliary hypotheses. Considering the orbit of Uranus example, deviations from certain expected values led researchers not to throw out Newton’s law of gravitation but rather to hypothesize that the solar system included more planets than were initially observed, which were exerting gravitational influences according to (rather than in contradiction to) Newton’s laws. While doing so may sound “circular”, it in fact led to the discovery of Neptune. In this sense, maintaining the hard-core hypotheses can predict that previously unaccounted-for factors must exist, which guides the search and discovery process. In comparison, in LFS, hypotheses about the role of c-command in establishing the kinds of “formal objects” (in particular, those which can lead to certain MR’s, e.g., “FD”, “DD”, etc.11) form the hard-core hypotheses of the program, or at least come quite close to it; we might instead take the idea of Merge as the hard-core and c-command-based objects as slightly auxiliary, but such gradations of “hard-core-ness” are not problematic. Other hypotheses, such as those relating directly observable sentence strings to corresponding syntactic structures, as well as those spelling out which MR’s are enabled by which factors, form the body of auxiliary hypotheses. This latter group of hypotheses is not limited to merely when c-command-based formal objects can lead to MR’s, but can also include cases when MR’s can arise due to factors not based on c-command or the CS at all.12 Using these hypotheses, (correlational) predictions can be made, and when/if experimental results turn out contrary to these definite correlational predictions, it is the set of the auxiliary hypotheses, rather than the hard-core ones, that is modified by revising, adding, or discarding hypotheses. For example, if correlations do not come out as predicted, the hypothesized structure of a given sentence type might be revised, or the conditions under which a given MR may be attributed to a non-c-command-MR reading might be updated, so that new correlations are predicted, continuing the process of “learning from errors” (see Chapter 3), just as the Newtonian physicists did. Returning to the discussion near the end of the previous sub-section, recall that a fact in LFS is something that is predicted by hypotheses and obtained and

11 These being c-command-based sources for meaning relations (MRs); see Chapter 5. 12 See again discussion of “NFS” in Chapter 5 or of “quirky binding” in Chapter 2.

412 

 Hajime Hoji and Daniel Plesniak

replicated experimentally. Beyond simply replicating exactly what has been demonstrated before, replication in LFS can be pursued by keeping intact most of the hypotheses that have given initial success and modifying/altering just one of the hypotheses (or at least a minimal number of them) and trying to obtain experimental success with the modification/alteration. Success in such a case will constitute replication of the viability of the hypotheses that are kept intact, and failure suggests that the content of the modification/alteration is what is most likely responsible for the failure. This is one of the ways to accumulate our knowledge about the subject matter in LFS, i.e., if a prediction does not get supported by experimental results, it constitutes accumulation of knowledge as long as the structure of our prediction deduction has been made completely explicit and as long as it has been made clear which of the hypotheses is being checked, with prior experiments having provided support for the validity of the other hypotheses in question (see the flowchart in Chapter 3). This method is also central to physics: Poincaré (1952) states: The physicist who has just given up one of his hypotheses should, on the contrary, rejoice, for he found an unexpected opportunity of discovery. His hypothesis, I imagine, had not been lightly adopted. It took into account all the known factors which seem capable of intervention in the phenomenon. If it is not verified, it is because there is something unexpected and extraordinary about it, because we are on the point of finding something unknown and new. Has the hypothesis thus rejected been sterile? Far from it. It may be even said that it has rendered more service than a true hypothesis. Not only has it been the occasion of a decisive experiment, but if this experiment had been made by chance, without the hypothesis, no conclusion could have been drawn; nothing extraordinary would have been seen; and only one fact the more would have been catalogued, without deducing from it the remotest consequence. (Poincaré 1952: 150–151)

We can thus see the centrality of hypotheses, including incorrect hypotheses, in the accumulation of facts in both physics and language faculty science. It is possible, of course, that notions like Merge, or at least c-command, may one day need to be refined in LFS, just as Poincaré himself was part of the movement which abandoned classical/Newtonian physical hard-core hypotheses and moved to modern ones. However, until the point comes at which it is clear that such a change must be made, the regular process of accumulation of facts by the formulation, testing, and refinement of auxiliary processes in LFS continues, allowing for the regular process of science to go on, just as is now the case in physics. The central hypotheses of physics, however, have far more empirical support for them at this time, while, as noted, LFS is still in its infancy.

Language Faculty Science and Physics 

 413

4 Prediction and Measurement 4.1 Prediction In this section, the two linked notions of prediction and measurement are considered. Crucially, in both LFS and physics, predictions are definite, and this constrains the ways in which measurement is done. Such constraints, in turn, allow for both testability and replicability. However, LFS faces certain challenges that are less prominent in physics, primarily relating to the source of a measured value, which can be difficult to determine when dealing with mind-internal phenomena. As noted in the introduction, both physics and LFS deduce and test definite predictions, though these predictions are often definite not in the sense of predicting an absolute value, but rather, definite correlations between values. Physics may in this sense be said to encompass a wider variety of types of definiteness; as it also has absolute predictions, e.g., the charge of all electrons claimed to be the same, and so when we measure the charge of an electron, we know exactly what it is predicted to be. Another different type of definiteness in physics involves definite predictions as to probability, for example as represented by the probability density functions of quantum mechanics. Despite being probabilistic, such predictions are nevertheless definite; a certain observed distribution can be quantitively measured as to how much it is said to differ from a predicted distribution, just as an observed value may be measured as to how much it differs from the predicted value of an absolute or correlational prediction. In all cases, the crucial factor is that there is a specific prediction, and the degree to which the results of an experiment differ from that prediction is unambiguously quantifiable; if the difference falls within acceptable limits, whatever those may be,13 the result is said to be consistent with the prediction, and if it does not, then the result is said to contradict the prediction. For LFS, at least as described in this volume, the “acceptable limits” of deviation from prediction are essentially infinitesimal; results should match predictions exactly, in that if it is predicted that no individuals with properties A, B, or 13 In the case of fields like modern particle physics, the gold standard is the “5 sigma” standard, meaning roughly that if one considers a distribution of results that would be expected if all that was being observed was random fluctuations, the results actually found would be five standard deviations away from the mean (about 1 in 3.5 million), making it highly implausible that the observed results are indeed random fluctuations; depending on the analysis, some of the results of LFS may rise to this standard, but a consistent evaluation has not yet been developed to assess this.

414 

 Hajime Hoji and Daniel Plesniak

C should accept a given sentence with a given interpretation, then if there is an individual with those properties who does accept such a pairing, the experimental result contradicts the prediction. This lack of tolerance for deviation is consistent with LFS’s reliance on categorical data. While physics primarily makes use of numerical data, in that properties, such mass, charge, velocity, etc., are associated with some numerical value, LFS makes use of qualitative judgment data, “acceptable” or “unacceptable”. This is not to suggest that LFS is non-mathematical and physics is mathematical. Rather, this split reflects the different forms of mathematics that are hypothesized to underlie each: physical laws are hypothesized to make reference to values that can enter into various combinations with one another, e.g. addition, multiplication, etc., depending on the property in question, whereas LFS makes no such assumption; two “acceptable” judgments do not combine to make an “unacceptable” judgment or a “doubly acceptable” judgment. Rather, LFS assumes that the CS, via Merge, builds a hierarchical structure of sets, and that the relations between elements in those sets determine what sort of CS-based dependencies can be established. The establishment of those dependencies is an all-or-nothing proposition; the structural configuration either has the requisite properties or it does not, namely either X c-commands Y or it does not. As such, predictions based on c-command must be, at their heart, binary; in effect, the only meaningful outcomes of a judgment on whether an MR is compatible with a sentence string are “possible” or “impossible”.14 If c-command were replaced with some relation that encoded some aspect of “degree”, e.g., X “c-commands” Y to degree 5 vs. to degree 4, then this would no longer be the case. In general, whatever form predictions take is dependent on the hypotheses that generate them. Because LFS’s basic predictions are of the form “possible” or “impossible”, there is little room for error; the value 1.05 might be considered “close enough” to the value 1 for the sake of a certain physics experiment, but “impossible” and “possible” are qualities that are both opposite and absolute, so there is no equivalent way to derive a situation wherein “impossible” might be considered “close enough” 14 That we should focus on “possible” vs. “impossible” is addressed in more detail in Hoji 2015: Chapter 2, where the concept of “fundamental schematic asymmetry” is proposed. (Research that focuses on “possible” vs. “impossible” is referred to in Plesniak’s Chapter 7 of this volume and Plesniak 2022: Chapter 3 as “possibility-seeking research”, as opposed to “typicality-seeking research.) The correlational methodology proposed and illustrated in this volume makes use of that concept, with the crucial addition of the concept of correlation. The correlational/conditional prediction about c-command detection discussed in this volume is a predicted correlation of schematic asymmetries, and having been confirmed (at least to an extent), it is now a confirmed predicted correlation of schematic asymmetries, if we use the concepts in Hoji 2015, along with the concept of correlation.

Language Faculty Science and Physics 

 415

to “possible”. There is of course some room for numbers when establishing “how much” replication/evidence is enough to make the decision as to whether an individual finds something to be possible (see discussion in Chapter 8), but this is a problem of trying to make an interpretation, not of prediction itself.

4.2 Measurement, Testability, and Replication As with predictions, measurements in both LFS and physics are definite in nature. In each case, a measurement yields a single, specific value: a number for physics, a judgment of “acceptable” or “unacceptable” for LFS. However, we must make the caveat that, while in physics, there is never concern that the measurement taken might reflect a “non-physical value” (except for a failure of equipment), this concern is very much at the heart of LFS. That is because MR’s can arise due to factors outside of the CS. As a result, an individual’s finding a given sentence acceptable with a given MR does not necessarily mean that this acceptability reflects CS properties. This is further complicated by the recognition that I-languages of speakers in the “same linguistic community” are most likely not identical; as noted earlier in this chapter, “faculties of the mind” outside the CS itself frequently affect linguistic behaviors of a given individual at a given time (such as how they judge the possibility of a meaning-sound pairing) in ways that may not always be the same for different individuals or even for the same individual at different times (see Chapter 5 for further discussion). Tools like correlations of judgments are thus necessary to determine the source of a given judgment because they allow the experimenter to check what aspects of the mind an individual’s judgments seem to be reflecting, whether that is the CS or something else. This understanding that the nature of a linguistic judgment may be based, at least partially, on non-CS sources separates the approach laid out in this volume from that of Hoji 2015. The latter is concerned primarily with an absolute ✶Schema-based prediction getting disconfirmed or surviving a rigorous attempt at disconfirmation. Hoji 2015 is in this sense much closer to the physics-style interpretation of results, where it is assumed they must correspond to the value assigned by the object of inquiry (at least if there are no problems with the experimental device, which Hoji 2015 attempts to deal with via sub-experiments, as do some of the chapters in this volume). Under the correlational approach, however, we are concerned with a ✶Schema-based correlational prediction; given that our correlations are written as implicational statements, they hold only if the antecedent of the implication holds; the purpose of that antecedent is precisely to limit us to cases where experimental diagnostics suggest that linguistic judgments reflect CS properties. Correlations are thus a

416 

 Hajime Hoji and Daniel Plesniak

tool for mitigating the “multifaceted” nature of acceptability judgments; only under the proper conditions can we act “like physicists” and take measurements at face value as reflections of the object of inquiry, and correlations tell us precisely when said conditions have been met. There remains, however, an “issue of interpretation” in LFS as to what to “count” as an individual’s judgment on a given schema; that is, how many judgments must be checked before determining that someone finds a given MR impossible/possible with a given sentence type. Currently, this determination is simply a matter of the experimenter’s own “judgment call” as to what seems to us to be “good/clear enough” indication about a given individual’s judgments. In a given multiple-participant experiment, this sort of judgment call is made in a consistent and definite manner, but what standards one picks to do so are still a matter of what seems best to the experimenter. There is also a split between judgments of “acceptable” and “unacceptable”. Generally, in a self-experiment, when the experimenter finds a given instance of something being judged “acceptable” when it was predicted to be judged “unacceptable”, this is typically taken as an indication of the disconfirmation of the prediction in question, assuming that judgment can be replicated (a notion to be clarified in the following section). In a non-self experiments/demonstrations, such replication has not always been possible due to the limited number of questions that can be posed to participants, so one acceptance of such an acceptance has been considered sufficient grounds for the prediction in question to be considered disconfirmed. Analogously, when a self-experimenter finds one instance of something being judged acceptable when it is indeed predicted to be acceptable, this has typically been considered as reasonably strong evidence that it is acceptable (though replication across different instantiations, or even the same instantiation at different times, is still important for the purposes of evaluating the claim), but in non-self experiments, because the participants cannot be considered as reliable as the experimenter him/herself, multiple instances of acceptance have often been required to determine that the thing in question is indeed acceptable to that speaker. As we can see from the above (and as discussed in previous chapters), there is an asymmetry between cases where the predicted judgment is “unacceptable” and cases where the predicted judgment is “acceptable” (or more technically, “possibly acceptable”; see Chapter 4). In the former case, a single “acceptable” judgment has been taken as sufficiently contradicting the prediction in question, whereas in the latter, a single “acceptable” judgment may not be sufficient for us to consider the prediction well-supported; this differs from physics to the extent that, in physics, there are not particular predicted numerical values that are treated asymmetrically with respect to one another. However, if we understand this asymmetry

Language Faculty Science and Physics 

 417

as a manifestation of the desire to try as hard as possible to disconfirm our own predictions, and support them only by failing to successfully disconfirm them, then LFS and physics are parallel in that sense. That is, it is not that the predictions are treated asymmetrically in LFS, but that supporting vs. contradicting data are treated asymmetrically, with just one (reliable) instance of the latter being enough to disconfirm predictions, but only a large amount of the former being considered “significant” support. Even with these guiding principles, the decision of what counts as either (i) a sufficiently reliable/replicable contradiction of predictions or (ii) a sufficiently numerous amount of support, remains a fundamentally arbitrary one at this point; it is not derived from any particular theory but is made to the practical needs of the moment, as addressed in Chapter 8 of this volume. These are problems that have been dealt with for centuries in physics, and have led to the development of various standards and techniques for quantifying results. As LFS matures, such techniques will surely be relevant, and may in turn give numerical values greater prominence in LFS, not as direct measurements of judgments, but as determinants of how many judgment patterns of a given type (supporting or contrasting with predicted judgment patterns) may be taken to constitute sufficient evidence of acceptability or unacceptability of a given schema (at least in the context of non-self experiments).

5 Experiment, Demonstration, and Replication 5.1 Experiment and Demonstration We now further examine the relationship between experimental results and the hypotheses/predictions that they bear on. Both physics and LFS maintain a division between experiments on the one hand and demonstrations on the other. The key difference is that experiments allow for truth seeking, in the form of truly testing predictions against observed values, while demonstrations are attempts to provide evidence to corroborate such predictions. It is trivial to see that this is true in physics; physics demonstrations are viewable at most science museums, e.g., pendulum clocks, Van de Graff generators, etc. In LFS, this distinction is also present, if we assume that self-experiments, or extremely detailed experiments with non-self participants, constitute true experiments, and what are called non-self-experiments in previous chapters are actually multi-person demonstrations intended to replicate (part of) what has been found in the experiments. As noted in Chapter 8, to make such demonstrations, sufficient trial and error are

418 

 Hajime Hoji and Daniel Plesniak

necessary to correctly calibrate what sorts of sub-experiments to use, how many instantiations of each schema should be judged, etc. Because of the “error” aspect of trial and error, it is quite plausible that some of these attempts will fail, and yet, this does not cause trepidation to the experimenter, as the experimenter “knows” that the intended result is basically correct, based on the experiments they have conducted, and if the demonstration fails, it is a failure of setup, not substance. Of course, if the same failure repeats itself in a reliable manner, the experimenter may begin to re-examine what was assumed to be “known”, but one seeming disconfirmation in isolation in a demonstration-style non-self-experiments will not cause great concern, whereas a disconfirmation in a self-experiment, or very detailed non-self-experiment, will cause much greater concern. Physics does not have self- vs. non-self-experiments in the same way, as its experiments are not on individual humans. In this sense, there are challenges in LFS posed by the fact that the language faculty is internal to an individual, and we must deal with different I-languages to find out about universal properties of the language faculty. There is some analogy with physics, in that the same results ought to replicate no matter where the experiment is conducted assuming proper noise control is done, an experiment on Earth should yield the same result as one on Mars, and if the result only obtains on Earth, then it would appear that it was some Earth-specific property causing it. However, that Earth-specific property, by assumption, must also be a physical property, whereas if a result obtains in one I-language and not another, this may not be a reflection of a CS property at all. Attempts to engage in LFS before the innovation of the correlational methodology pursued here, such as Hoji 2015, struggled because of this issue. Such attempts assumed that, as long as an individual meets certain basic criteria (tested for via sub-experiments), that individual’s judgments on the availability of MR(X,  Y) (e.g., BVA(X, Y)) must correspond to the possibility or impossibility of the CS generating a structure in which the required structural relation for the MR(X, Y) in question obtains. While, in a sense, this assumption remains valid even when working with correlations, the “certain basic criteria” are much expanded to ensure a lack of interference from non-CS-based effects; this expansion allows us to rigorously account for individuals whose judgments on a given MR do not derive exclusively from the CS. This development is based on the recognition that, while the CS itself may be universal, speakers do not make equal use of the CS when reporting their judgments on the availability of a certain MR(X, Y) in a given sentence at a given time. The correlational methodology is thus an attempt to rectify this issue, allowing for control of non-CS factors across I-languages; in a certain sense this makes LFS more like physics, as now experiments run on different “planets”(=I-languages) are indeed expected to give the same results in the correlational sense.

Language Faculty Science and Physics 

 419

5.2 Replication and Discoveries We can recall that a fundamental tenet of the scientific method is that nothing is ever proven to be true; a given hypothesis, when evaluated in conjunction with other hypotheses, is supported when it survives rigorous attempts to disconfirm it; the more rigorous, varied, and numerous these attempts are, the stronger the support, but there is no threshold at which it becomes “certain”. Even if we are sure that there has not been a failure of experimental setup, deduction, etc., and that results really are in accordance with what is predicted by the hypotheses in question, there is always a chance that there is a competing set of hypotheses that will explain the data equally well and possibly outperform the current hypotheses in future experiments. On the other hand, if we assume no failure of the experimental setup, deduction, and other such things, then we can definitely prove a hypothesis incorrect by showing that observed results do not match predictions; of course, we cannot be sure that no such failure has occurred, so we cannot really prove things definitely false either, but the conceptual distinction between “proving true” and “proving false” is nevertheless important (see the discussion near the end of Section 4, for example). Namely, the former is not even possible in theory via the scientific method, while the latter is possible in theory, albeit not in practice. Indeed, with the various caveats above, a theory being supported by experimental results teaches us very little; it only increases our confidence in what is already believed. A failure of predictions, on the other hand, is a discovery, namely that something about the hypotheses is wrong or incomplete. This distinction is true not only in novel experiments but also in experiments intended to replicate results of previously conducted experiments, again emphasizing the centrall importance of ✶Schema-based predictions (correlational or not), as these provide the locus of such testability.15 Returning to the example of LIGO, consider the case of gravitational wave detection resulting from the merger of two massive objects (which involves quite a lot of movement relative to one another, thus producing the best known source of detectable gravitational waves). We may ask: how can the detection of gravitational waves from a particular event of two particular massive objects merging at a particular place/time be reproduced? Naturally, what is reproduced is not the 15 As discussed in some depth in Chapters 5 and 6 of this volume, a ✶Schema-based prediction (correlational or not) can be disconfirmed, at least in principle, by an existential demonstration of the failure of the predicted entailment in question to hold, i.e., a single instance of judgment by a reliable speaker that a single sentence instantiating the ✶Schema in question (with the specified MR(X, Y)) is acceptable, contrary to the prediction.

420 

 Hajime Hoji and Daniel Plesniak

same measurement but the theoretically predicted correlation between (i) measurements and (ii) an event of two massive objects merging at a different place/ time. Thinking this way, the detection of gravitational waves and the c-command detection are quite close to one another. Fundamentally, both involve a predicted link between what actually occurs and what is measured, as is true of any rigorous prediction-testing experiment. Regardless of the subject matter though, what is reproduced when we “reproduce” this link is not the original measurement itself, but the correlational pattern that the original measurement instantiated; that is, we cannot observe the exact same thing twice, but we can observe the same type of link between a type of situation and a type of outcome (such as one which corresponds to a particular measurement), which is itself a correlation. As we have mentioned, though, it is, however, a somewhat ironic fact about science in general that we can only “learn” through failing to replicate predictions. That is, when predictions do replicate, they provide support for a given theory; this may make us more likely to believe the theory, but unless there is an alternative theory that made a different prediction (in which case, the results would have been a replication failure of this other theory), we have not truly “learned” anything definitive, only become more confident in what we already knew. LIGO, for example, increased confidence in General Relativity by detecting gravitational waves, but had it not done so, General Relativity would not have been abandoned. Indeed, during LIGO’s initial eight-year run, it found no evidence of gravitational waves, but this merely prompted researchers to investigate further. Their ultimate detection served as an excellent demonstration of Einstein’s theory, but precisely because it was a demonstration, it provided support to an existing idea, rather than leading to a new discovery (at least in terms of fundamental physical theory). We may contrast LIGO’s results with those of another famous interferometry experiment, the Michelson-Morley experiment, performed over a century earlier. This experiment sought to utilize the movement of the Earth relative to the hypothesized “luminiferous aether”, which was theorized to be the stationary medium through which light waves travelled. As understood through modern physics, no such aether is necessary; light, in the form of photons, is an excitation of the electromagnetic field, which itself permeates all of space, so no carrier medium is necessary. At the time however, because this “aether” was assumed to be stationary, the movement of the Earth through it would cause distortive effects on light moving parallel to the Earth’s motion, not unlike the distortions that were eventually leveraged in the LIGO experiments. Michelson-Morley and similar experiments, however, failed: the anticipated interference effect was not detected. The fact that such predictions failed to be experimentally supported cast grave doubt on the popular theories of its time, and opened the door for new

Language Faculty Science and Physics 

 421

hypotheses to explain this departure from predictions, leading to the proposal of length contraction by Lorentz, which itself was ultimately utilized in Einstein’s Special Relativity. Likewise, the most dramatic new discoveries of LFS will come when predicted correlations break down, rather than when they are demonstrated. A central purpose of LFS experiments is thus to search for reliable ways of discrediting various auxiliary hypotheses, identifying holes in our knowledge which can be used as the basis for new theories.16 This reminds us of Feynman’s (1994 [1965]: 152) “[W]e are trying to prove ourselves wrong as quickly as possible, because only in that way can we find progress”.17 To the extent that no such counterexamples are found, we can consider current hypotheses to be sufficient, and indeed, we would like our hypotheses to be as sufficient as possible, but it is not our role to defend such hypotheses from disconfirmation, only to ensure as reliable a test as possible. Noise control is thus used to ensure reliability of any result, supporting or disconfirming, so that we can be increasingly confident that the conclusion the result seems to point to is indeed an accurate one.

6 Noise Control As mentioned, both physics and LFS suffer from the same problem of noise; measurements are only meaningful if we have confidence that the measured value of a property reflects the “true” value of that property. With physics, this usually comes in the form of calibrating measuring devices and assuring that there is not some sort of mechanical impediment to them. There is an oft-repeated anecdote that Wilson and Penzias, the astronomers who first detected the background “noise” of the big bang, did not believe their results until they had thoroughly cleaned pigeon droppings off their instruments; the “noise” persisted despite all attempts to remove sources of error from the machine, ultimately convincing them of the reality of what they were detecting. In LFS this sort of “noise” takes the form of individuals misunderstanding (or not attending to) the task at hand, which is addressed by Hoji 2015-style sub-experiments. However, as previously 16 We also conduct in LFS experiments to see if new auxiliary hypotheses lead to new predictions that can be supported experimentally; if they do, that yields another form of progress. 17 This is part of the following: “One of the ways of stopping science would be only to do experiments in the regions where you know the law. But experimenters search most diligently, and with the greatest efforts, in exactly those places where it seems most likely that we can prove our theories wrong. In other words we are trying to prove ourselves wrong as quickly as possible, because only in that way can we find progress.”

422 

 Hajime Hoji and Daniel Plesniak

mentioned, LFS contains noise of another kind, interference from mental systems outside the CS, which must be addressed via tools like the correlational methodology. Once again, this difference flows naturally from the differing subject-matters of the two fields; physics studies “all” the universe, whereas LFS studies but a relatively small “corner” of the human mind. A further such difference, again flowing from the nature of the object of inquiry, is that noise in physics is (usually) measured in terms of numbers and its control is based on theoretical understanding of the sources of noise. Galileo, contrary to the popular story of dropping objects from a high tower, instead examined the properties of objects rolling down an inclined plain, which introduced various sources of noise such as the friction of the plane, which had to be accounted for mathematically. In keeping with the non-numerical nature of the data in LFS, noise in LFS is not numerical. It thus cannot be “factored out” through mathematics. Instead, as described, correlations are used to determine when such noise is or is not present; when it is not present, we can get judgment results that reliably show properties of the CS, and when it is present, we can study the noise itself. However, crucially, to factor out the noise, it is not a priori necessary to understand it, merely to be able to detect it in a reliable manner; understanding it helps to detect it of course, but initially at least, our understanding of the noise has been quite limited compared to our understanding of the CS. In the future, however, there are prospects of the correlational methodology and the reliable c-command detector helping us obtain a better understanding of the nature and the source of the “noise” that interferes with the “signal” of c-command. That is, if we can observe systematic deviations from c-command patterns by using our system for studying c-command, then we can study those systematic deviations as well. Thus, what was once “noise” may become core subject matter; this is only possible, however, once basic patterns of c-command can be detected in a reliable manner, allowing us to then detect systematic deviations away from those patterns. This “non-numericity” has further consequences than that, however. Consider the case of “research rigor.” In physics, we can think of such rigor coming in two types, of the mathematical sort, as well as basic scientific integrity and honesty. In LFS, there is rigorous deduction, and this deductive process takes on an analogous role to that of the mathematical rigor achieved by the use of equations in physics. Naturally, scientific integrity and honesty can be pursued in LFS as well. However, there are unique challenges to pursuing such integrity and honesty without (quantity-based) mathematics. Returning again to Poincaré 1952: Now, under what conditions is the use of hypothesis without danger? The proposal to submit all to experiment is not sufficient. Some hypotheses are dangerous, – first and foremost

Language Faculty Science and Physics 

 423

those which are tacit and unconscious. And since we make them without knowing them, we cannot get rid of them. Here again, there is a service that mathematical physics may render us. By the precision which is its characteristic, we are compelled to formulate all the hypotheses that we would unhesitatingly make without its aid. Let us also notice that it is important not to multiply hypotheses indefinitely. If we construct a theory based upon multiple hypotheses, and if experiment condemns it, which of the premises must be changed? It is impossible to tell. Conversely, if the experiment succeeds, must we suppose that it has verified all these hypotheses at once? Can several unknowns be determined from a single equation? (Poincaré 1952: 151–152)

The uncompromising requirement to arrive at something like an expected number/ set of numbers/distribution forces physicists to examine their own tacit assumptions in order to formulate them in a way legible to mathematical physics. In principle, because LFS lacks the same sort of mathematical system, the practitioner of LFS risks having unconscious assumptions “creep in”, thus allowing the situation Poincaré warns against, wherein it is unclear whether the success or failure of an experiment is due to the hypotheses the experimenter thinks it is due to, or whether it might be the hidden assumptions which have caused it to succeed or fail. However, as has been mentioned previously, LFS does have as its core a basic formal notion, namely hierarchically nested sets, in particular those created by Merge. Requiring as much as possible to be formulated in terms of Merge-derived/ Merge-dependent notions to some extent alleviates this challenge for LFS, making its deductive process more similar to that of mathematical physics. However, as just discussed, it seems true at this point that neither Merge nor even the CS can be understood to be solely responsible for the outcome of linguistic judgments; there is “noise” that fundamentally interrupts the basic “CS patterns”. As has been additionally noted, this provides challenges as to determining what even counts as a “fact” for LFS, and if one does not have a clear sense of what counts as a fact, it is not clear how one can try to pursue scientific integrity and honesty. As a result, LFS pursues a type of research rigor that is less central to physics, namely the need for replication (and thus the opportunity for failure to replicate) of the basic results in different settings, which for LFS are different I-languages. This includes not only different I-languages of speakers of the same linguistic community, but also across different types of I-languages, such as “Japanese”, “English”, etc. This allows for an increased rigor of testing, hopefully allowing researchers to discover anything that has been “concealed” by the informal deduction surrounding I-language-specific “noise.” While of course, physicists do frequently attempt to show that a given law or principle applies across all relevant environments, cross-I-language replication is a somewhat different process, more analogous to, say, repeating all the classic

424 

 Hajime Hoji and Daniel Plesniak

experiments done on Earth on Mars, for example (as discussed in Section 5.1). Of course, doing so would be a strong demonstration of the universality of the result in question, and experiments are indeed sometimes repeated in, say, outer space, that are done on Earth, but the fact remains that most physicists do not feel they need to see such a thing to feel “convinced” that the result in question is valid. On the other hand, an LFS practitioner may reasonably be concerned that results of a given experiment may not replicate in a (significantly or minorly) different I-language than the ones in which they were first devised, precisely because it is not clear which factors that make up a given “language” derive from the Mergebased CS and which do not. Ultimately, this difference in the way in which research rigor is pursued derives from the same basic split between LFS and physics: though the two are unified in pursuing the scientific method, they apply the scientific method to very different subject matters. Crucially, physicists do not look into their own minds, or the minds of other human beings, to find out about laws of universe, while “LFStists” do look into their own mind and the minds of others to find out about what laws govern the language faculty. This not only transposes the challenges physicists face to a new domain, but also introduces new challenges with which physicists do not typically need to grapple.

7 A Science of the Mental This chapter has discussed the ways in which LFS resembles and departs from physics. Both take as their object of inquiry a formal system, and thus their inquiry consequently follows the scientific method. The two differ, however, not only in terms of their development, with physics being far more mature than LFS, but also in terms of the nature of their subject matter. Physics is the study of the entirety of the physical universe, while LFS studies (primarily) the CS of the language faculty, a subset of the “mental universe”. Of course, a given physicist does not study all of the universe at once, so for a given inquiry, one might understand that, just like in LFS, a physicist is studying a specific part of the physical universe. The key difference, though, is that, by assumption, all the physical universe is undergirded by the same formal system, whereas this seems not to be true in the mental domain; other components of the mind seem to act in ways quite different from the CS, obscuring its formal properties with noise of a “nonformal” nature. Even though this added difficulty makes pursuing LFS challenging, such an endeavor is worthwhile (at least in our opinion) because LFS represents some-

Language Faculty Science and Physics 

 425

thing not found in modern physics, namely a definite science of (part of) the mind. One might argue, though, that the mind ought to be reducible to the physical, and as such, physics is already a science of the mind. However, there is yet to be formulated any physical theory that derives the predictions that LFS has thus far successfully borne out; perhaps that will change in the future, but the relationship between the mental and the physical is not currently well enough understood that one can simply derive the latter from the former (at least not in ways that can be supported by the scientific method as discussed). One useful way of conceptualizing the situation comes from Penrose’s (2003) suggestion to divide reality into three “worlds”, the physical, the mental, and the mathematical/Platonic; that is, there are at least three related but conceptually different “domains” that we seem to have experience with: the external world, our internal, experiential world, and the world of abstractions. A sphere in the physical world, our mental representation of the sphere, and the mathematical equation describing that sphere can be understood as different things; indeed they can conflict, as our perception of the sphere might not match its physical properties, and it may turn out that the (physical) sphere itself is not perfectly spherical (in the mathematical sense). The fact that we can talk about mismatches between physical and mental and mathematical representations of a sphere (and many other things) suggests that these are not merely different ways of viewing the same thing, but rather, distinct domains, at least as far as we experience them as human beings. These three worlds, Penrose goes on to say, give rise to three mysteries, namely: how the physical world generates the mental, how the mental world (mind) can represent/understand the mathematical/Platonic world, and how the mathematical/Platonic world can underlie the physical world via physical law (as represented in Figure 1 below). Penrose concedes that perhaps each of these worlds might only “partially” contain/generate the other (as represented in Figure 2 below); perhaps there are aspects of the mind not derived from the physical world, Platonic truths outside the grasp of our minds, and/or aspects of the physical world not governed by systematic abstractions like mathematics. (Penrose notes he prefers the simpler picture wherein all three are in some sense “contained” in one another). See the figures below (which correspond to Penrose’s Figures 1.3 and 1.4 from his pages 18 and 20 respectively). Regardless of how each world is contained (partially or completely), it is only a subset of the preceding world; the aspects of the physical world that generate the mind are only a subset of all the physical world, and so forth. As such, only a portion of the mental world “contributes” to the mathematical/Platonic world. We may understand that the CS is part of the mental world that contributes to the mathematical/Platonic world; investigation of the properties of the CS has

426 

 Hajime Hoji and Daniel Plesniak

Figure 1: The Three Worlds Underlying One Another.

Figure 2: The Three Worlds Only Partially Underlying One Another.

led to the postulation of Merge, which, being a formal set-building operation, is part of the mathematical/Platonic world. This conception, in fact, underscores why LFS does not take as its object of inquiry the entire “mental world.” LFS focuses

Language Faculty Science and Physics 

 427

on the CS, precisely because the CS has a direct connection to the world of formal mathematics, and thus the scientific method may fruitfully be applied to it in order to gain knowledge about the language faculty. LFS correspondingly does not focus on other aspects of the mind, or even the language faculty beyond the CS, where such a connection either does not exist or has not yet been made clear (although it deals with them in the context of noise control). In particular, areas of the language faculty wherein it is not possible to deduce and successfully support via experiment definite and testable predictions regarding how given sentences will be judged are not amenable to the investigative methodologies employed by LFS; if nothing like a (correlational) schematic asymmetry can be deduced and/or replicated about a given aspect of the language faculty, then LFS can bear only indirectly on its study. By analogy, those dealing with mathematics who want to learn something about physics “through” (or in relation to) mathematics will have to “set aside” many aspects of mathematics which simply do not happen to correspond to anything physical; the very belief that there are basic laws of nature implies that only very particular types of mathematical objects occur in them; the cardinality of the set of conceivable mathematical objects is infinite, and thus no finite set of finite laws could employ even a significant fraction of them. Focusing on a narrow subset of the total whole is thus an essential aspect of relating the mind, mathematics, and the physical world, deriving from the interrelated nature of the three worlds that Penrose points out. We do not know, however, what aspects of the mind may indeed be reflexes of formal systems, and which may, therefore, be within the purview of LFS. As noted above, the determining factor will ultimately be whether facts, in the sense of experimentally supported predicted (correlations of) definite judgment patterns, can be established, specifically in the form of schematic asymmetries. Given the discussion in previous sections, such facts can only be so established if the predictions flow from definite hypotheses (and receive experimental support); if such definite hypotheses are possible for a given domain, it is perhaps necessarily true that the domain is, or at least contains, some reflex of a formal system, as definite hypotheses are, in essence, hypotheses stated in formal terms. The degree to which such definite hypotheses can be applied across different areas of the mind remains to be seen, but we may hold our hope that there may be many aspects of the mind which have the relevant properties. In conclusion, consider that advances in physics not only give insight into the physical world, but also help us understand the physical basis of the mental world as well. Likewise, advances in the understanding of mental world, in fields like LFS, not only give insight into the human mind, but may also provide insight into how that mind can grasp, generate, or otherwise interact with, the abstract Platonic/mathematical world as well. The better we understand the formal prop-

428 

 Hajime Hoji and Daniel Plesniak

erties of the human mind, the better we can understand how we as humans can grasp and apply concepts like “formal properties” at all. This in turn may lead us to understand aspects of the mathematical world in ways we have not imagined before, which may even open up new avenues of analysis in areas like physics. Just as Penrose’s three worlds each contain (all or part of) one another, work on any aspect of the mysteries surrounding their connections has the potential to shine light on all three worlds.

References Abbott, B.P., et al. 2016. Observations of gravitational waves from a binary black hole merger. Physical Review Letters 116. https://physics.aps.org/featured-article-pdf/10.1103/ PhysRevLett.116.061102 (Accessed 22 February 2022) Chomsky, Noam. 1986. Knowledge of Language: Its nature, origin, and use. Westport, CT: Praeger. Chomsky, Noam. 2017. The Galilean challenge. Inference: International review of science 3(1). https://inference-review.com/article/the-galilean-challenge (accessed 14 February 2022) Feynman, Richard. 1994 [1965]. The character of physical law. New York: The Modern Library. Hoji, Hajime. 2015. Language faculty science. Cambridge: Cambridge University Press. Lakatos, Imre. 1978. The methodology of scientific research programmes (Philosophical papers Volume 1). (John Worrall and Gregory Currie (eds.).) Cambridge: Cambridge University Press. Langacker, Ronald. 1969. On pronominalization and the chain of command. In David Reibel and Sanford Schane (eds.), Modern studies in English, 160–186. Englewood Cliffs, NJ: Prentice-Hall. Marr, David. 1982. Vision: A computational investigation into the human representation and processing of visual information. Cambridge, MA: MIT Press. Penrose, Roger. 2003. The road to reality: A complete guide to the laws of the universe. New York: Random House. Poincaré, Henri. 1952. Science and hypothesis. New York: Dover Publications. (The English translation of La science et l’hypothèse (1902).) Reinhart, Tanya. 1976. The syntactic domain of anaphora. Cambridge, MA: MIT dissertation. Reinhart, Tanya. 1983. Anaphora and semantic interpretation. Chicago: University of Chicago Press.

Hajime Hoji and Daniel Plesniak

Conclusion: Why Language Faculty Science? 1 Introduction Language Faculty Science (LFS) is the scientific study of what underlies the human ability to relate linguistic signs/sounds/symbols to meaning, focusing on definite predictions and definite experimental results about an individual. In this volume, we have presented and motivated the basic tenets of our current approach to LFS and both explained and demonstrated how they can be executed and implemented. At this point, we would like to close this volume by highlighting the basic conceptual challenges and concerns that serve as motivation to this approach and thus the reasons why we choose (and recommend) to pursue LFS in this manner. In Chapter 1, a key distinction was made between “compatibility-seeking” research, as exemplified by Hoji 1985, and “testability-seeking” research, as exemplified by this volume. The key difference is that testability-seeking research aims to maximize the chances of reliably disconfirming incorrect hypotheses, whereas compatibility-seeking research aims to provide plausible explanations that reconcile a given set of data with a given set of hypotheses. While this contrast may make it seem that the former approach focuses on disconfirming incorrect hypotheses and the latter focuses on supporting correct hypotheses, Chapter 1 argues that testability-seeking in fact provides opportunities for both. Maximizing the chances of reliably disconfirming incorrect hypotheses is not itself the fundamental goal of testability-seeking, but rather, it is part of the attempt to accumulate knowledge about our subject matter; if the attempted disconfirmation fails, the hypotheses being tested receive experimental support and are taken provisionally to be correct. If, on the other hand, the prediction made under the hypotheses is contradicted by experimental results, this can inform us as to what went wrong. Testability-seeking thus provides us with a method of discriminating between supported and unsupported hypotheses, and in the case of the latter, a method for improving the set of hypotheses until we find some that will be supported. We have emphasized the value of the testability-seeking approach throughout this volume, yet it must also have become clear to readers that achieving experimental “success” when following this approach is no simple matter. Chapter 1, for example, mentioned the attempt to pursue testability-seeking in Hoji 2015, which, while promising, failed to produce any predictions that survived rigorous testing. The issues behind this failure were explored in greater depth in Chapter 2’s discussion of “quirky” phenomena, the upshot of which is that there is genuine variation in judgments; specifically, sentences which are crucially predicted to be unaccepthttps://doi.org/10.1515/9783110724790-010

430 

 Hajime Hoji and Daniel Plesniak

able with a given MR1 may become acceptable to certain individuals under certain circumstances, frustrating any straightforward prediction of universal unacceptability. In some sense then, the failure of Hoji 2015’s predictions provided an opportunity to “learn from mistakes”, to use Chapter 3’s terminology. Further to this end, Chapter 3 makes several other recommendations that are adopted in the following chapters, including two crucial ones that we will briefly review here. The first is that, because our hypotheses are about the language faculty, which relates externalized sentences to meanings, our predictions should ultimately be about the (un)acceptability of the pairing of a sentence template on the one hand and a potential interpretation of the sentences instantiating that template on the other; for our purposes, such interpretations are specifically ones that interpretatively “link” two elements in said sentences. Calling the two related elements X and Y, we have frequently expressed the intended interpretation as MR(X, Y), e.g., in Coref(John, his) in ‘John praised his student’, sometimes adding “S” to refer to the particular sentence (type) in question, as in Coref(S, John, his). The second crucial recommendation made in Chapter 3 is that we should focus on two basic types of predictions about the un(availability) of schema-MR pairings. These are as repeated below: (1)

(Chapter 3: (6)) A ✶Schema-based prediction:    Speakers judge any ✶Example conforming to a ✶Schema to be completely unacceptable with meaning relation MR(X, Y).

(2) (Chapter 3: (8)) An okSchema-based prediction:     Speakers judge some okExamples conforming to an okSchema to be acceptable (to varying degrees) with meaning relation MR(X, Y). The ✶Schema is the locus of testability for us, as it constitutes a definite prediction that can be disconfirmed with a single counterexample. The okSchema, on the other hand, cannot be directly disconfirmed, but we can succeed or fail in supporting it; the purpose of having okSchemata is to provide minimal contrasts with a given ✶ Schema, thus demonstrating the effects of the presence and absence of the formal factors conditioning MR availability. The formal factor that we have been most concerned with in this volume is the presence/absence of the relevant c-command rela-

1 “MR” standing for meaning relation.

Conclusion: Why Language Faculty Science? 

 431

tion. The reasons for our focus on this relation are overviewed in Chapter 4, namely that it is a direct reflection of the way in which the hypothesized Computational System of the language faculty builds structural representations. As such a ✶Schema attains its status, most crucially, due to the absence of the relevant c-command relation, and likewise the okSchema attains its status again most crucially, due to the (possible) presence of the relevant c-command relation. If the okSchema-based prediction is confirmed and the ✶Schema-based prediction does get disconfirmed, then we have found strong evidence not only for the specific c-command-based hypotheses that lead to the ✶Schema and okSchema predictions being tested, but also for the basic hypothesis of the structure-building Computational System as well. (1) and (2) above are themselves based on Hoji 2015: Section 2, and as such, rely on Hoji 2015’s non-variation-based conception of judgments, which does not permit them to be relativized to a given speaker at a given time with specific choices of, among other things, X and Y of MR(X, Y), all of which we now know to be relevant to the possibility of MR(X, Y) for a given sentence. As such, (1) and (2) need to be revised to handle the variability/quirkiness problem. Chapter 3 does in fact provide the first inkling as to the solution pursued in this volume to this problem. Chapter 4 specifically calls this the problem of dealing with “non-formal sources” (NFS), which provides us with a useful term for talking about said problem’s solution. In particular, the key insight of the solution first touched on in Chapter 3 is that, if we want to ensure that MR acceptability judgments reflect c-command relations and not NFS, we must first diagnose certain properties on the level of the individual. Specifically, because any given MR is such that it relates two elements, X and Y, we must, at a minimum, establish which choices of X and Y consistently permit MR(X, Y) only when X c-commands Y, for that individual (at that time). Only if we have ensured this, are we licensed to use that particular MR(X, Y) to test the relevant c-command-based hypotheses. Recognizing the need to “relativize” (1) and (2) in the sense just noted, Chapter 4 adopts the following outlook: a given MR can come about through multiple sources, including formal relations (FR), which are contingent on some definite configuration of X and Y holding, such as X c-commanding Y,2 and NFS, which do not have definite configuration-based requirements. Because of the nature of FR, assuming we have hypotheses as to what structure a given schema corresponds to, we can deduce whether FR-based MR is possible or not for a given schema, but the same is not true of NFS-based MR. Thus, ✶Schema and okSchema are predictive only 2 The c-command relation is technically between LF(X) and LF(Y), as first discussed in some depth in Chapter 4: Section 3 and addressed in subsequent Chapters. This important point is simplified here for the ease of exposition. The same simplification will be made in the ensuing discussion including remarks after (3), in relation to DR(X, beta), Coref(alpha, Y) and BVA(X, Y), to be mentioned there.

432 

 Hajime Hoji and Daniel Plesniak

if we know that a given MR interpretation can arise solely based on FR (for the individual in question, with the particular choices of X and Y, etc.). Because we definitionally cannot know whether or not NFS-based MR is possible solely by looking at the schema itself, we also cannot predict whether a given MR is or is not possible for a given schemata in general. We thus cannot use terms like ✶Schema and okSchema unless we have more information (specified to the individual in question). We can, however, label schemata according to their predicted formal properties, which we can assess without requiring such “further information”. Chapter 4 thus labels schemata as [+cc]S or [-cc]S, indicating schemata where X either does ([+cc]) or does not ([-cc]) c-command Y. [+cc]S and [-cc]S become okSchemata and ✶Schemata (respectively), only if we are able to determine that NFS-based MR is not possible.3 The question, of course, is how to ensure that MR arises solely based on FR. From Chapter 4 onwards, this volume focused on one such way, namely the use of three MRs, with the first two serving to detect whether X and Y were fit to use as “probes” for c-command detection with the third. Chapter 4 discusses this as a “correlational methodology”; its most general form is as given in (3), which is applicable to any three MR’s with the relevant properties. In Chapter 5, this formula is “filled in” with the three specific MRs (BVA, DR, and Coref) on which the majority of this volume focuses, as displayed in (4). (3) (Chapter 4: (57)) Testable correlational/conditional prediction about c-command-detection:   Provided that MR2([+cc]S, X, beta4):yes ˄ MR3([+cc]S, alpha, Y):yes ˄ MR1([+cc] S, X, Y):yes;   MR2([-cc]S, X, beta):no ˄ MR3([-cc]S, alpha, Y):no → MR1([-cc]S, X, Y):no (4) (Chapter 5: (29)) Correlational/conditional prediction about c-command-detection with BVA   (X, Y):   Provided that DR([+cc]S, X, beta):yes ˄ Coref([+cc]S, alpha, Y):yes ˄ BVA([+cc] S, X, Y):yes;   DR([-cc]S, X, beta):no ˄ Coref([-cc]S, alpha, Y):no → BVA([-cc]S, X, Y):no

3 And a given [+cc]S may still not become an okSchemata if the FR in question has other requirements, such as the FR “FD”, said to underlie c-command-based BVA and Coref, which, as discussed in Chapter 5, Section 7.2, seems also to have a “anti-locality” requirement as well. 4 As noted in Chapter 4: Section 5, ““beta” is meant to be an instance of Y and “alpha” an instance of X, with the understanding that what can serve as X of MR1 perhaps may not be able to serve as X of MR3 (hence we call the latter alpha), and what can serve as Y of MR1 perhaps may not be able to serve as Y of MR2 (hence we call the latter beta).”

Conclusion: Why Language Faculty Science? 

 433

(5) illustrates the basic logic behind our pursuit of c-command detection by the correlational methodology, as outlined in this volume, in reference to (4). (5) (Chapter 5: (30))

1 Provided that

3

2

DR(

S, X, beta):yes + Coref(

S, alpha, Y):yes

+

5

BVA(

S, X, Y):yes

DR(

S, X, beta):no

S, alpha, Y):no



4

BVA(

S, X, Y):no

1 Identification of X that gives rise to DR(X, beta) but does not give rise to NFS-DR(X, beta), i.e., X that gives rise to DR(X, beta) only based on FR.

+ Coref(

2 Identification of Y that gives rise to Coref(alpha, Y) but does not give rise to NFS-Coref(alpha, Y), i.e., Y that gives rise to Coref(alpha, Y) only based on FR.

;

3 When accompanied by 1 and 2, it is c-command detection with BVA(X, Y), with the X and Y identified in 1 and 2. In isolation, identification of a pair of X and Y that gives rise to BVA(X, Y), but does not give rise to NFS-BVA(X, Y), i.e., a pair of X and Y that gives rise to BVA(X, Y) only based on FR.

Figure 1: Illustrating the correlational/conditional prediction about c-command detection with BVA(X, Y) in (4).

By obtaining judgments in , i.e., accepting DR(X, beta) only in a sentence/ schema where X c-commands beta, from a given speaker at a given time, we identify X that is a good probe (i.e., free of NFS effects) for c-command detection for that speaker at that point in time. Likewise, obtaining judgments in , i.e., accepting Coref(alpha, Y) only in a sentence/schema where alpha c-commands Y, from a given speaker at a given time, we identify Y that is a good probe for c-command detection for that speaker at that point in time. Once we have identified such X and Y, we make a definite and categorical prediction that BVA(X, Y) is possible only in a sentence/schema where X c-commands Y, from this speaker at that time.5 This “correlational methodology” was explained abstractly in Chapter 4, and then its effectiveness was demonstrated experimentally in Japanese in both a detailed self-experiment (Chapter 5) and a replicational “demonstration” of some of the results of that self-experiment with “naïve” participants (Chapter 6). Chapter 7

5 See Plesniak 2022: Chapters 2 and 3 for an articulation of what assumptions are needed for the deduction of (4).

434 

 Hajime Hoji and Daniel Plesniak

then demonstrated that analogous results can be found in English6 (with quite minimal modification), as well as providing other evidence that this methodology is both robust and widely applicable through other types of follow-up experiments. We therefore hope at this point it is clear to readers that this methodology does indeed “work”, and thus that it is indeed possible to both pursue rigorous testability-seeking in research on language and make predictions that survive such testing (overcoming the flaws in Hoji 1985 and Hoji 2015 respectively). Nevertheless, even if readers believe that such a program is possible, they may still worry that it would be quite complicated and difficult to pursue, and one would be well-justified in wondering whether, in the end, it is worth all the trouble. Chapter 8 provided some practical advice on how to implement (non-self-)experiments of this sort, and Chapter 9 discussed the ways in which this program shares many basic aspects with the sort of research done in physics. Hopefully, the former has made the prospect of doing such research seem a little less challenging, and the latter has made the motivation behind this program clearer. We cannot deny, however, that pursuing Language Faculty Science as described in this volume is not easy. Throughout this volume, we have justified the importance of doing so in terms of the “basic scientific method”, the main features of which we understand are as stated in (6)–(8), repeated here from Chapter 4: (4)–(6), respectively. (6) Guess-Compute-Compare (based on Feynman 1994 [1965]: 150): a. Guess-Compute: Deduce definite and testable predictions from hypotheses. b. Compare: Obtain definite experimental results and compare them with the definite and testable predictions. (7) Methodological Minimalism:7       Hypotheses must be stated in terms of the minimal number of theoretical concepts necessary. (8) Replication: Experimental results must be replicable. Because the language faculty is internal to an individual and is understood to underlie our ability to relate sounds and meaning, the relevant investigation 6 As we have noted at times, readers are referred to Plesniak 2022, which demonstrates analogous results in Korean and Mandarin Chinese as well. 7 As noted in Chapter 4: note 3, what is stated in (7) is the aspect of methodological minimalism that is most relevant to our present discussion.

Conclusion: Why Language Faculty Science? 

 435

must be concerned, at least at the most fundamental level, with an individual’s linguistic intuitions about the relation between sounds and meaning, as noted in Chapter 4. In order to accumulate knowledge about the language faculty by the basic scientific method, we must therefore articulate each of (9a–c). (9) (Chapter 4: (3)) a. How we can deduce definite predictions about an individual’s linguistic intuitions regarding the relation between linguistic sounds and meaning b. How we can obtain experimental results pertaining to an individual precisely in line with such definite predictions c. How we can replicate the experimental results alluded to in (9b) As we have argued throughout this volume, adherence to this basic scientific method forces us to study the language faculty in the manner described above (or a close variation thereof). Compared to a more compatibility-seeking approach, this imposes a number of restrictions on what topics we can easily address and how we can go about addressing them. In its current formulation, it limits us to topics that can be addressed by obtaining reliable acceptable vs. unacceptable contrasts regarding a given MR in different related schemata. If one desires to test hypotheses that do not lead to predictions about definite unacceptability of an MR with a given schema (under the right conditions), then it likely cannot be studied in this way; certain extensions may be possible, of course, but it seems probable that most language-related topics cannot be investigated in this way. In our view, this is an unavoidable consequence of wanting to study a specific thing (the language faculty) in a specific way (via the basic scientific method outlined above). There is a certain reciprocity here: the types of hypotheses we want to evaluate constrain the methods we use to evaluate them, and the methods we use to evaluate hypotheses constrain the types of hypotheses we can evaluate. This is true under any scientific research program, but (rigorous) testability-seeking when pursued in a particular domain of inquiry imposes fairly stringent restrictions on how such “evaluation” may take place. We understand that there can be good-faith disagreements about how science should be pursued, and that, as such, not everyone will agree that correlational methodologies like the one explored in this volume are required, or even ideal, for studying the language faculty at this point in time. It is not our desire to use the term “science” as a club with which to beat all our skeptics as “unscientific”. Rather, in the remainder of this conclusion, we would like to argue that adherence to this vision of language faculty research is not merely a constraint that we have imposed on ourselves for the sake of “scientific purity”, but it actually empowers us to study the language faculty in ways that have never been done before. While it is true that, compared to the wide-ranging and rather straightforward

436 

 Hajime Hoji and Daniel Plesniak

argumentation provided in works like Hoji 1985, the correlational methodology narrows our focus and makes us do quite a bit of work. However, to make another physics analogy, much the same thing can be said of a lens-powered microscope or telescope;8 they concentrate our focus entirely on one specific area and require quite a lot of work to build. The payoff, however, is that the level of detail we are able to observe in that focused area is enormously increased compared to our “naked eye” observations. In this sense, the adoption of the correlational methodology can be regarded as a way of enhancing our observational power in a given domain, rather than simply a limitation and/or burden.

2 Some Benefits of Adopting the Correlational Methodology 2.1 Introduction: Improvements over Hoji 1985 (and Other Similar Works) For the sake of demonstrating the ways in which the program put forth in this volume acts as a lens and brings tangible improvement over previous ways of studying the language faculty, let us compare once again with Hoji 1985.9 Besides simply producing more robust predictions, this volume tangibly improves over Hoji 1985 in a number of areas, including (I) the discovery of the existence and properties of NFS, which Hoji 1985 essentially missed entirely, (II) the ability of the correlational methodology to use the availability of MR(X, Y) as a probe for qualitatively more effective and reliable investigation of various structures/phe-

8 For an extended analogy between the correlational methodology and the telescope, readers are referred to Plesniak’s “Building the Linguistic Telescope” presentation at https://sites.google.com/usc.edu/danielplesniak/papers-and-presentations. 9 Hoji 1985, which is discussed in some depth in Chapter 1, is a work with a compatibility-seeking research orientation that deals with many of the same topics we have dealt with in this volume, which (in addition to the identity of its author) is why we use it here as a point of comparison. We do not intend to claim that Hoji 1985 is an exemplary, or even a representative, instance of compatibility-seeking research in this area. The ways in which subsequent (or even prior) such research has or has not brought improvements in the areas to be discussed, we leave for consideration in future work. The body of literature on these topics is both vast and heterogenous, so to some extent, this must be assessed on a case-by-case basis. See, for example, Hoji 2006, which presents a comparison along these lines of two works dealing with related topics in Japanese syntax.

Conclusion: Why Language Faculty Science? 

 437

nomena, old and new, which Hoji 1985’s approach does not allow for, and (III) the (re)evaluation of purported cross-linguistic contrasts between, e.g., those between Japanese and English that were suggested (but not given any principled explanation) in Hoji 1985.

2.2 The Discovery of NFS Taking each of these three related improvements in turn, we start first with (I), the discovery of the existence and properties of NFS. Following earlier works, Hoji 1985 held that what we term BVA(X, Y) could only arise if X c-commanded Y in the structural representation of the relevant sentence. Coupled with the hypothesis that subjects c-command objects, this led to the prediction that sentences of the form “Y-no N-ga X-o V” (Y’s Noun Verb X, e.g., Soko-no sitauke-ga subete-no kaisya-o uttaeta ‘Its subsidiaries sued every company’) should be unacceptable with a BVA(X, Y) (e.g., with BVA(subete-no kaisya, soko) ‘BVA(every company, it)’ reading), whereas sentences of the form X-ga Y-no N-o V (X Verb Y’s Noun, e.g., Subete-no kaisya-ga soko-no sitauke-o uttaeta ‘Every company sued its subsidiaries’) should be acceptable with the same BVA(X, Y) reading.10 A contrast between the two sentence types is thus predicted. Consistent with the compatibility-seeking mode of inquiry, Hoji 1985 sought to identify cases where this prediction was borne out, and for cases where the prediction was not borne out (as clearly as predicted), proceeded to consider a wider range of empirical materials (until finding some that would seem to confirm the predictions), allowing the basic hypotheses to be defended. There were various instances of this, often involving specific choices of X or Y of BVA(X, Y). For example, the use of kare ‘he’ (the so-called overt pronoun in Japanese) for Y seemed to result in BVA(X, Y) becoming unacceptable regardless of c-command, whereas the use of zibun (roughly ‘oneself’) as Y or subete-no N ‘every N’ as X seemed to result in BVA(X, Y) becoming acceptable regardless of c-command. In each of these cases, Hoji 1985 (and/or subsequent works like Hoji 2003 and Hoji 2015) gave word/ phrase-specific explanations of why the use of such words/phrases did not “count” as contradicting the basic predictions, suggesting that those predictions could only be reliably checked with other choices of X and Y.11 10 To be historically accurate, the so-called empty pronoun was used in place of soko ‘it’ as Y of BVA(X, Y) in Hoji 1985, and the structure of the relevant sentence was more involved as a result, but the intended point remains; see Chapter 1. 11 Hoji 1985: 2.2, for example, maintained that kare was not suitable for use as Y of BVA(X, Y) because “overt pronominals such as kare ‘he’ can never become semantic variables [i.e., Y of BVA(X, Y) in the terms of the current discussion]”, citing Nakai 1976 and Nakayama 1982. (Nakai,

438 

 Hajime Hoji and Daniel Plesniak

This strategy relies on the assumption that we can obtain from speakers consistent and stable judgments on the availability of MR(X, Y) without reference to correlations of judgments if we work with the right choices of X and Y. As we have argued extensively in this volume, this is not possible, and correlations are indeed needed. Once we recognize this, we are led to expect that, we might in fact be able to use those “excluded” lexical items for c-command detection. That is precisely what has happened, as illustrated in the case of kare by Hoji’s judgments at “Stage 3”, as discussed in Chapter 5; see also Chapter 1: note 21, which includes remarks on zibun. That subete-no N ‘every N’ can be X of BVA/DR(X, Y/beta) for purposes of c-command detection is illustrated in Chapter 6 by the fact that c-command detection with BVA(subete-no N, Y) obtained with some speakers. As such, while Hoji 1985 and like works were able to provide plausible accounts for specific exceptions to c-command-based patterns of MR, they were missing something quite fundamental, namely the existence of general alternative sources to MR like BVA besides c-command. That is, it is not simply the case that there are a handful of isolated exceptions to, say, the requirement that X c-command Y in order for BVA(X, Y) to be acceptable, but rather, there are multiple systematic exceptions, which we can in fact reliably diagnose if we look for them. Several of these sources have been described in Chapter 5 of this volume, including what are called there NFS1 and NFS2 (as well as a source having to do with linear precedence). Importantly, it is not simply that Hoji 1985 was wrong about which sources could lead to BVA; rather, by essentially dismissing any non-c-command-based BVA (via posthoc explanations), Hoji 1985 in fact failed to realize that things like NFS1 and NFS2 existed at all. By “explaining away” or “brushing off” what was in fact evidence of a new discovery, an opportunity to gain a much more comprehensive picture of the potential sources for BVA was lost (which prevents us from using BVA as an effective

S. 1976. “A study of anaphoric relations in Japanese,” ms., University of Massachusetts, Amherst, and Nakayama, S. 1982. On English and Japanese pronouns, MA thesis, University of Tokyo, Tokyo, as given in Hoji 1985; these are not included in the References because we have not been able to locate these unpublished manuscripts.) Hoji 1985: 2.2 also remarked, in part based on considerations of its “emphatic use” and “discourse topic” use, prompted by discussion in Kuno 1985, that, “[i]n regard to its bound variable interpretation, zibun seems to behave less “systematically” than the empty pronominal, . . . [and] [f]or this reason, it might be more fruitful at this point to examine the cases of empty pronominals as bound variables to see more clearly in what configurations the intended bound variable interpretation is allowed in Japanese”. (As discussed in Chapter 1 of this volume, as well as in Hoji 2003: 2.2.2.1, the “empty pronoun” turns out to be not a reliable Y of BVA(X, Y) for the purpose of c-command detection.) In other works, e.g., in Hoji 2003, 2015, it is suggested that we should avoid using subete-no N ‘every N’ as X of DR(X, beta) because it allows a reading of a specific group of individuals, making it possible for many speakers to obtain DR based on what we now understand is the source called NFS1 in Chapter 5.

Conclusion: Why Language Faculty Science? 

 439

tool for probing the nature of the language faculty). On the other hand, as we have seen in Chapter 5, investigations via the correlational methodology have already proved useful in revealing the properties like NFS1 and NFS2. The focus on data that seem to contradict c-command-based predictions resulting from the pursuit of testability-seeking has thus enabled us to make discoveries about the sources of a given MR in ways that Hoji 1985’s style of inquiry forced its author to overlook.

2.3 Reliable Expansion of Our Empirical Coverage Let us now turn to (II), the ability of the correlational methodology to use the availability of MR(X, Y) as a probe for qualitatively more effective and reliable investigation of various structures/phenomena, old and new. Chapters 1, 5, and 6 illustrated this in relation to the simple Subject Object Verb sentence pattern and its corresponding Object Subject Verb pattern.12 The key hypotheses about those patterns we entertained were (I) that in the Subject Object Verb pattern, the object never c-commands the subject (and thus does not c-command anything contained in the subject), (II) that in the Object Subject Verb pattern, it is possible for the subject to c-command the object, (and thus to c-command what is contained in the object). By controlling for the NSF factors that lead to murky judgments, by using the correlational methodology, we have in fact been able to attain completely categorical patterns of judgments, in relation to the availability of BVA(X, Y), that support these basic hypotheses, as discussed in Chapters 4 and 5.13 Chapter 5: Section 7 discussed how we can expand the empirical domain to different sentence patterns (Section 7.1) and a related phenomenon, often addressed in the field as “local disjointness effects” in relation to the “Principle B of the Binding Theory” (Section 7.2). While the sentence pattern and the phenomenon in question have been studied in the past literature, hence an “old” structure/phenomenon (see Chapter 5: Section 7, as well as Chapters 1 and 3: Section 6.4), it is the use of the correlational methodology that has made it possible to deduce definite correlational predictions and attain definite experimental results in line with the predictions, significantly improving our understanding about issues that have been addressed in the past literature. We will focus on the first of these examples, briefly reviewing a subset of the discussion in Chapter 5: Section 7.1. As discussed in Chapter 5, in Hajime Hoji’s 12 The illustration of the correlational methodology in Chapter 6 was strictly based on the simple Subject Object Verb and its corresponding Object Subject Verb patterns in Japanese, and that in Chapter 5 was largely (though not entirely) based on those patterns. 13 And we saw in Chapter 7 that something quite analogous is the case in English as well.

440 

 Hajime Hoji and Daniel Plesniak

self-experiments (when his judgments are in the state he calls “Stage 2” there), the pattern of availability of DR(X, beta) and Coref(alpha, Y) suggest that X=NPigai and Y=sono otoko resist NFS effects, and as such, BVA(NP-igai, sono otoko) should strictly require NP-igai to c-command sono otoko14 (at least so long as Hoji remains at that “stage” and assuming that there are no other confounding factors present). As expected, if we use those choices of X and Y in paradigm below, where V is a “transitive o-marking” verb (which marks the subject with -ga and the object with -o), such as hituyoo to site iru ‘needs’, BVA(X, Y) is available only with [+cc]S in (10a) and never with [-cc]S in (10b). (10)

(Modified from Chapter 5: (79)) With a “transitive o-marking” verb a. [+cc]S: [ . . . Y . . . ]-o X-ga V b. [-cc]S: [ . . . Y . . . ]-ga X-o V

With a “transitive ni-marking” verb, such as renrakusuru ‘to contact’, we obtain the relevant [+cc]S and [-cc]S by simply substituting -ni for -o in (10), and Hoji’s judgments remain the same. Judgments on transitive -o-marking and -ni-marking verbs thus correspond to our basic hypotheses we have held about subjects and objects throughout. The different sentence patterns considered in Chapter 5: Section 7.1 involve “ergative” verbs, such as hituyoo da ‘needs’. As indicated below, the two sentences in (11) express very similar, in fact almost identical, meanings, at least with regard to who is needing what. (11)  (Based on Chapter 5: (77))15 a. (With a “transitive ni-marking verb) NP2-ga NP1-o hituyoo to site iru ‘NP2 needs NP1’

14 Technically, as mentioned in footnote 2, if we follow the convention of Chapter 4, we should say LF(NP-igai) c-commanding LF(sono-otoko), but we are suppressing this technical detail for ease of exposition. 15 (i-a) and (i-b) below correspond to (11a) and (11b), respectively. (i) a. John-ga kuruma-o hituyoo to siteiru. ‘John needs a car.’ b. John-ni kuruma-ga hituyoo da. ‘‘John needs a car.’

Conclusion: Why Language Faculty Science? 

 441

b. (With an “ergative” verb) NP2-ni NP1-ga hituyoo da ‘NP2 needs NP1’ ‘NP1 is necessary for NP2’ It is claimed in works such as Kuroda 1978, 1986 that, unlike with the transitive ni-marking verbs, “NP-ni NP-ga V” is the “base order” and “NP-ga NP-ni V” is the “derived order” of sentences with an ergative verb; see Chapter 5: note 62. Assuming that base order reflects c-command structure in that NP-ni c-commands NP-ga in ergatives (the opposite of what obtains in the otherwise schematically identical transitive-ni-marking sentences, where “NP-ga NP-ni V” is the base order), we obtain the following predicted distinctions: (12) (Based on Chapter 5: (74)) With an ergative verb a. [+cc]S: [ . . . Y . . . ]-ga X-ni V b. [-cc]S: [ . . . Y . . . ]-ni X-ga V (13)

(Based on Chapter 5: (76)) With a “transitive ni-marking” verb a. [+cc]S: [ . . . Y . . . ]-ni X-ga V b. [-cc]S: [ . . . Y . . . ]-ga X-ni V

Strong evidence for such a claim, however, has proved difficult to provide in the past. Attempts to use the acceptability of MR for such a purpose have been confounded due to the notoriously slippery and variable nature of the relevant judgments. Such issues, we now understand, result from failure to control for non-FRsources of MR. As noted above, if we control such sources for “Stage 2 Hoji”, and use the resulting choices of X=NP-igai and Y=sono otoko, we can put the hypothesized [±cc]S status of the relevant schemata to the test. As noted above, the “transitive ni-marking” cases come out as predicted; we then expect the pattern ought to “flip” in the ergative case. As shown in Chapter 5, this is precisely what happens; with a “transitive ni-marking” (e.g., renrakusuru ‘to contact’; see Chapter 5: (63) and (64) for actual examples), the pattern of judgments are “reversed” with an ergative verb, so long as the “c-command-sensitive” X and Y identified by the correlational methodology are used. As a result, we are able to provide direct and reliable support in favor of the longstanding hypotheses about such ergative verbs. We can thus see an example of the way in which the correlational methodology provides us with a decisive experimental means to test and potentially support the longstanding intuitions of native speaker theoreticians. In particular, it allows us to identify the sorts of

442 

 Hajime Hoji and Daniel Plesniak

choices (such a particular X and Y of MR(X, Y)) that will ensure that a given MR can be used as a reliable probe for the presence of a given c-command relation in a sentence. It goes without saying that Hoji 1985, where such identification was at best post-hoc and stipulation-based, could neither produce predictions of this kind nor test them; when judgments are murky, investigations of structural hypotheses are bound to be inconclusive, but if we can find ways to control for the factors that cause said “murkiness” (such as the correlational methodology presented in this volume), then clear progress on such issues is possible.16

16 Chapter 5: Section 7.1 provided further illustration of the effectiveness of the correlational methodology by addressing the structural ambiguity of (i), which is an instance of a sentence with an ergative verb, due to the possibility of (ii). (i)

(ii)

John-ni Bill-ga mieta. John-dat Bill-nom saw/looked ‘Bill was visible to John. NP-ni Bill-ga John-ni -dat Bill-nom John-dat ‘To NP, Bill looked like John’

mieta. saw/looked

If NP-ni in (ii) is phonetically unrealized (not pronounced), and is understood as expressing “to me”, for example, Bill-ga John-ni mieta in (ii) would mean “To me, Bill looked like John”, but that would be the same phonetic sequence as the “inverted word order” of (i). That in turn raises questions about the correlations of judgments on the availability of BVA(X, Y), DR(X, beta) and Coref(alpha, Y), when X/alpha and Y/beta appear as an NP in place of Bill or John or inside such an NP. As reported in Chapter 5: Section 7.1, the predicted correlations of judgments do obtain based on Hoji’s judgments, pointing at the correct structural analysis of (ii) in relation to Bill-ga and John-ni. Recall that our empirical investigation started out with what seem to be among the most basic sentence patterns (i.e., SOV and OSV with a regular transitive verb), and reliable expansion of empirical coverage was made possible by the identification of effective tools for obtaining definite experimental results based on our investigation of those basic patterns by the correlational methodology. Although we have not been able to discuss them in this volume due to space limitations, the use of the correlational methodology also makes it possible to attain a qualitatively better understanding than before about many other “phenomena”, many of which involve “displacement” of some sort, such as the “multiple-object” construction, “long-distance scrambling”, and “multiple scrambling”, as well as other related phenomena such as the sloppy identity reading, the cleft construction, the relative clause construction, island constraints just to mention a few, all by focusing on the availability of MR(X, Y). With the ever-expanding inventory of experimental tools that we will have accumulated through such investigation, not only about properties of specific FRs but also about properties of NFSs (and even other factors) that contribute to how our linguistic judgments come about, we will be in a good position to seek a proper analysis of any new “construction” or “phenomenon”. Under a research orientation, such as the one pursued in Hoji 1985, we could not expect anything even close to this, because such research proceeds based on empirical bases that are not rigorously established, and, as indicated in preceding chapters, the inability to establish a rigorous empirical basis stems from not pursuing the correlational methodology.

Conclusion: Why Language Faculty Science? 

 443

2.4 (Re)evaluation of Purported Cross-linguistic Contrasts Finally, we turn to (III), the (re)evaluation of purported cross-linguistic contrasts. We can take the example of the claim that the scope order “rigidly” reflects the surface word order in Japanese but not in English; that is, that the availability MR(X, Y) like BVA(X, Y) and DR(X, Y) are restricted in Japanese based on what positions X and Y occupy in the sentence, as claimed in Hoji 1985, and that this is not the case in English. As we can see from Chapters 5–7 of this volume, not only is this simply not the case, but in fact, it is missing a very basic fact about “rigidity”. In particular, if we compare results of the various non-self-experiments reported in Chapters 6 and 7, which address Japanese and English respectively, we see that, for a given choice of X and Y, there are some individuals from each language who seem to behave in a “scope rigid” manner, i.e., to allow MR(X, Y) only when X is the subject and Y is (in) the object, and there are some from each language who do not behave in this way, allowing MR(X, Y) more permissively. Zooming in on Chapter 6’s Japanese results in particular, we can see that which individuals have “scope-rigidity” changes based on the choices of X and Y we consider; further, if we look at the results of the self-experiment reported in Chapter 5, we see that which X and Y behave “rigidly” actually changes over time for the same individual. There seem simply to be no grounds on which to describe “English” or “Japanese” as being or not being scope-rigid. Indeed, we feel at this point it is fair to suppose that much of what has been traditionally attributed to cross-linguistic variation is in fact inter-and intra-speaker variation. This is not to say that there is no such thing as a cross-linguistic difference: there may well be factors about Japanese that encourage scope rigidity or of English that discourage it, and some of these factors may well turn out to be very significant. However, unless one can control for effects at the level of the individual, it is hard to see how we can reliably evaluate hypotheses about such factors. Indeed, if, as Hoji 1985 did, we persist in taking the judgment of one (or a few) particular speaker(s) at one particular time on one particular sentence as reflecting “the judgment in Japanese” on the availability of an MR in a given schema, then we are doomed to find many spurious “cross-linguistic differences”, which will be mixed indistinguishably with the reflexes of any such difference that does exist. Research working on the level of the individual allows us to separate out the former from the latter, vastly improving our ability to evaluate claims of cross-linguistic differences.

3 A Dream Takes Flight From just the small sampling of examples given above, it is clear that, far from being a cumbersome limitation, adherence to the principles underlying Lan-

444 

 Hajime Hoji and Daniel Plesniak

guage Faculty Science as put forward in this volume have expanded our investigative power considerably. The results detailed in this volume represent only a few years of initial investigations using the correlational methodology, and yet we have already uncovered a wealth of knowledge about the language faculty; we cannot but anticipate that future research will yield even greater such dividends. To answer the question posed by the title of this conclusion, “Why Language Faculty Science?”, we need not invoke “philosophy of science” issues at all; there are clear benefits to pursuing such a program, both in terms of its unprecedented ability to produce reliably reproducible definite results and in its “magnification power” to uncover a myriad of never-before-seen discoveries. Of course, even if we had yet to achieve such results and were still, like Hoji 2015, pursuing testability-seeking without yet having found the proper means to do so, we would still maintain that there is great value in doing so. In science, it is better to have an honest failure than a misleading success, as pointed out, in effect, in Poincaré 1952: 150–15117 and Feynman 1994 [1965]: 152,18 for example. Though we think it fair to say that Hoji 1985 had a number of useful insights, it portrayed itself as an empirical investigation when it was more akin to data-sensitive speculation. As discussed in Chapter 1, while it did (as we now understand it) get a number of things right, it also got other things disastrously wrong, and the mistaken and ill-grounded generalizations about Japanese it claims to have supported have been misleading language researchers for decades now. That is the consequence of claiming that hypotheses have been empirically supported without having made sufficient efforts to attain testability. It is not our claim that all compatibility-seeking research is deceptive or not useful. Such an argument would immediately be undermined by our repeated citing of various ideas of Chomsky’s, including Merge, which forms the backbone of essentially all of our hypotheses. As we have described, Merge was Chomsky’s solution to a conceptual problem, that of “digital infinity” (see Chapter 4), which also had to be compatible with observations about sentence structure and c-command (many of which also stemmed from Chomsky). The initial arguments for Merge were thus entirely compatibility-seeking in nature. There is no reason that compatibility-seeking approaches cannot co-exist with testability-seeking ones, and indeed, as we are demonstrating, the latter can build off the former, just as we have endeavored in this volume to show how

17 See Chapter 9: Section 3. 18 See Chapter 9: Section 5.2; see also Feynman 1985: 340–343, mentioned briefly in Chapter 1: Section 2.

Conclusion: Why Language Faculty Science? 

 445

Merge-based predictions can be made testable and supported by definite experimental results. In physics, mathematical and highly theoretical speculation often flies far ahead of current experimental capability, and in turn, new experimental observations sometimes lead mathematical and highly theoretical speculation in new directions. The scientific goal of ultimately arriving at, as Einstein 1982 [1936]: 294 puts it, “a system of the greatest conceivable unity and of the greatest poverty of concepts of the logical foundations, which is still compatible with the observations made of our senses”, clearly points to the need to pursue both compatibility-seeking and testability-seeking in the terms of our discussion. There is no reason this should not be just as true in language-based research as it is in physics, at least in principle. Nevertheless, it is our belief that pursuit of compatibility without concern for testability has a long-term deleterious effect on scientific fields. Evidence of this in our own field is already apparent: we have at times heard from colleagues and fellow researchers that they doubt that the theories they work on are “real” (which we take to mean being concerned that the theories would not survive rigorous empirical testing) or that they would like to focus more on testability but fear that it would be damaging to their career (because they would not be able to successfully obtain results that support any meaningful set of hypotheses). We hope that the already substantial successes of the program laid out in this volume offer such individuals a path forward.19 It is possible to pursue experiments rigorous and definite enough that they leave little room for one to doubt the reality of their outcomes while at the same time obtaining results that ought to do credit to anyone’s career.20 As we write this,

19 Further materials, in the form of manuals regarding how to conduct both self- and non-self-experiments, are currently under preparation and should hopefully be available in the near future after this book is published. We expect the guidance they provide will be of additional use for at least some of those wishing to pursue this “path forward”. 20 Both a challenge and an opportunity come from the fact that LFS is not yet a mature science; a challenge, because it means there is not yet a time-tested way of addressing many basic problems, but an opportunity because it gives practitioners a chance of making major contributions to the overcoming of said problems. We expect it may take quite a long time for LFS to become a “mature science”. After all, it took thousands of years for physics to develop into what it is today. Remarks like the following may be suggestive in this regard, and also with regard to the critical importance of testability: The history of the thing, briefly, is this. The ancients first observed the way the planets seemed to move in the sky and concluded that they all, along with the earth, went around the sun. This discovery was later made independently by Copernicus, after people had forgotten that it had already been made. Now the next question that came up for study was: exactly how do they go around the sun, that is, with exactly what kind

446 

 Hajime Hoji and Daniel Plesniak

an altered version of John Lennon’s song “Imagine” rings in our minds, “You may say we are dreamers but we are not the only ones. We hope someday you will join us.” Hoji 2015 attempted to show how we can deduce hard predictions and how we can identify hard facts in language faculty science in the sense of those concepts used in Feynman’s (1999: 198–199) “In the strong nuclear interaction, we have this theory of colored quarks and gluons, very precise and completely stated, but with very few hard predictions. It’s technically very difficult to get a sharp test of the theory, and that’s a challenge. I feel passionately that that’s a loose thread; while there’s no evidence in conflict with the theory, we’re not likely to make much progress until we can check hard predictions with hard numbers”. Hoji 2015’s

of motion? Do they go with the sun as the centre of a circle, or do they go in some other kind of curve? How fast do they move? And so on. This discovery took longer to make. The times after Copernicus were times in which there were great debates about whether the planets in fact went around the sun along with the earth, or whether the earth was at the centre of the universe and so on. Then a man named Tycho Brahe evolved a way of answering the question. He thought that it might perhaps be a good idea to look very very carefully and to record exactly where the planets appear in the sky, and then the alternative theories might be distinguished from one another. This is the key of modern science and it was the beginning of the true understanding of Nature―this idea to look at the thing, to record the details, and to hope that in the information thus obtained might lie a clue to one or another theoretical interpretation. So Tycho, a rich man who owned an island near Copenhagen, outfitted his island with great brass circles and special observing positions, and recorded night after night the position of the planets. It is only through such hard work that we can find out anything. When all these data were collected they came into the hands of Kepler, who then tried to analyse what kind motion the planets made around the sun. And he did this by a method of trial and error. At one stage he thought he had it; he figured out that they went around the sun in circles with the sun off centre. Then Kepler noticed that one planet, I think it was Mars, was eight minutes of arc off, and he decided this was too big for Tycho Brahe to have made an error, and that this was not the right answer. So because of the precision of the experiments he was able to proceed to another trial and ultimately found out three things [i.e., Kepler’s three laws of planetary motion]. (Feynman 1994 [1965]: 5–6), content in brackets added by us. We may also recall, as discussed in Chapter 9, that it took nearly 100 years from Einstein’s postulation of the key hypotheses behind the deduction of the existence of gravitational waves to the first empirical demonstration of their existence. Indeed, achieving said demonstration took the concerted efforts by a group of some 950 scientists at universities in 16 countries for at least 20 years. This just shows “how hard it is to get to really know something, how careful you have to be about checking the experiments, how easy it is to make mistakes and fool yourself” as Feynman (1999: 22–23) puts it. When our subject matter is part of the human mind, we should not expect our challenges to be any less.

Conclusion: Why Language Faculty Science? 

 447

attempt, however, fell short of that goal, because it inadequately understood the degree to which judgments vary between individuals (and even within an individual at different times). To remedy this failing, this volume has addressed how we can deduce definite predictions about the judgments of an individual speaker on the basis of universal and language-particular hypotheses and obtain and replicate experimental results precisely in accordance with such predictions, based on the correlational methodology. In this conclusory chapter, we have highlighted just a few of the benefits that come from working with this methodology, but these represent merely the initial results of this promising new program for language faculty research. Hoji (2015: 318) “envisage[s] [. . .] a time when we will be able to deduce definite and categorical predictions (predicted schematic asymmetries) in various languages, evaluate by experiments the validity of our universal and language-particular hypotheses, and formulate hypotheses of a successively more general nature, without losing rigorous testability. When something like that has become the norm of the research program, an experiment dealing with one language can be understood clearly in terms of the universal hypotheses (along with language-particular hypotheses) so that the implications of the result of an experiment dealing with a particular language can be transparent with respect to other languages. Researchers “working with” different languages will at that point share (many of) the same puzzles and issues pertaining to universal properties of the language faculty. They will know precisely what necessary care and checks they need to do in order to design effective experiments for testing the validity of the same universal hypotheses and how they should interpret the [e]xperimental results in accordance with the way the predictions have been deduced by hypotheses. That will enable us to proceed in a way much more robust than what has been presented in the preceding chapters, but still on the basis of confirmed predicted schematic asymmetries.” The correlational methodology indeed has yielded much more robust experimental results. Hoji (2015: 318) continues, “The field will at that point be widely regarded as an exact science, and everyone will take that for granted. And I also suspect that, at that point, other fields of research that deal with the brain and the mind pay close attention to the research results and methodology in language faculty science because they find it useful to try to learn from how categorical experimental results obtain in language faculty science and how its methodology has guided its research efforts.” This hope for the future is what we mean by referring to ourselves as “dreamers” above. With enough people joining our efforts, we can make this vision a reality. Indeed, the research results reported in this volume make us hopeful that the dream may already be coming true.

448 

 Hajime Hoji and Daniel Plesniak

References Einstein, Albert. 1936. Physics and reality. The Journal of the Franklin Institute 221(3). 349–382. Reprinted in 1955, Ideas and opinions. New York: Crown Publishers. (The page references are to the 1982 edition. New York: Crown Trade Paperbacks.) Feynman, Richard. 1985. Surely you’re joking, Mr. Feynman!: Adventures of a curious character. New York: W.W. Norton & Company. Feynman, Richard. 1994 [1965]. The character of physical law. New York: The Modern Library. Feynman, Richard. 1999. The pleasure of finding things out. New York: Basic Books. Hoji, Hajime. 1985. Logical form constraints and configurational structures in Japanese. Seattle, WA: University of Washington dissertation. Hoji, Hajime. 2003. Falsifiability and repeatability in generative grammar: A case study of anaphora and scope dependency in Japanese. Lingua 113. 377–446. Reprinted in Hajime Hoji 2013, Gengo kagaku-o mezashite [Towards linguistic science]: Issues on anaphora in Japanese (Ayumi Ueyama and Yukinori Takubo (eds.)), 185–260. Shiga: Ohsumi Publisher. Hoji, Hajime. 2006. Assessing competing analyses: Two hypotheses about “scrambling” in Japanese. In Ayumi Ueyama (ed.), Theoretical and empirical studies of reference and anaphora: Toward the establishment of generative grammar as an empirical science (A report of the Japan Society for the Promotion of Science Grant-in-Aid for Scientific Research (B), Project No. 15320052), 139–185. Fukuoka: Kyushu University. Hoji, Hajime. 2015. Language faculty science. Cambridge: Cambridge University Press. Kuno, Susumu. 1985. Anaphora in Japanese. Paper presented at the Japanese Syntax Workshop, University of California San Diego, March. (Published as: Anaphora in Japanese. In S.-Y. Kuroda (ed.), Working Papers From the First SDF Workshop in Japanese Syntax, 11–70. La Jolla, CA: University of California, San Diego.) Kuroda, S.-Y. 1978. Case-marking, canonical sentence patterns and counter equi in Japanese (A preliminary survey). In John Hinds and Irwin Howard (eds.), Problems in Japanese syntax and semantics, 30–51. Tokyo: Kaitakusha. Reprinted in S.-Y. Kuroda 1992, Japanese syntax and semantics: Collected papers, 222–239. Dordrecht: Kluwer. Kuroda, S.-Y. 1986. Movement of noun phrases in Japanese. In Takashi Imai and Mamoru Saito (eds.), Issues in Japanese linguistics, 229–272. Dordrecht: Foris Publications. Reprinted in S.-Y. Kuroda 1992, Japanese syntax and semantics: Collected papers, 253–292. Dordrecht: Kluwer. Plesniak, Daniel. 2022. Towards a correlational law of language: Three factors constraining judgment variation. Los Angeles, CA: University of Southern California dissertation. Poincaré, Henri. 1952. Science and hypothesis. New York: Dover Publications. (The English translation of La science et l’hypothèse (1902).)

Index Example 62–64, 66–72, 430 Schema 62–64, 66–74, 82, 85, 87, 92–94, 99, 100, 102, 106, 108, 110, 111, 142, 335, 337–338, 419, 430–432 ✶ Schema-based prediction 18, 62–64, 66–69, 71, 72, 74, 76–80, 87, 89, 92, 99, 109, 142, 403, 404, 415, 419, 430, 431 #-cl-no 95, 100, 101, 104, 180, 228, 242, 246, 271, 277, 311, 313, 322, 324 % izyoo-no 24, 95, 99, 100, 104 2-wari izyoo-no 25 3D representation 86, 94, 122, 123, 126–131, 164, 167, 200, 201, 306 3-nin-no seizika 247, 248, 253, 258, 263–267, 272, 279, 282, 283, 285, 286, 295, 302, 303 3-tu no 24–26 3-tu-no robotto 182, 192, 193, 195, 196, 204, 276, 288, 290, 294 ✶ ✶

Abbott, B.P. 406 Absolute ✶Schema-based prediction 415 Absolute predictions 413 Abstract structural representation 58–59, 391 Acceptability judgment 358, 401, 416, 431 Accumulation of facts 412 Across-I-language replication 423 Aitu 20, 25, 240–242, 244, 246–248, 253, 258, 263, 266, 268, 271, 272, 279, 291–294, 303, 304, 317, 318 All-or-nothing proposition 414 ano gisi 183, 185, 186, 193, 195, 204, 240, 288, 291 ano gisi-igai 182, 183, 190–195, 199, 204, 240, 289 ano otoko 196, 240, 289, 290, 293 A-NP vs. so-NP test 241, 242, 244 Anti-locality condition 22, 23, 209–211, 215 Aristotle 409 Asoko 20, 25, 239–242, 244, 246, 247, 264–268, 271, 283, 285, 286, 291, 292, 294, 307, 317 Asymmetrical c-command relation 8 Attaining testability 58–67 https://doi.org/10.1515/9783110724790-011

Attempts at disconfirmation 15, 17, 18, 28, 189, 402, 415 Attentiveness 143, 242–244, 252, 284, 335–337, 340–343, 345, 361, 378, 379, 381, 382, 402 Auxiliary hypotheses 398, 410–412, 421 Average across a group 28–29, 225, 358–360 A-words 35–38, 44 Barish, Barry 406 Barker, Chris 348, 349 Basic CS patterns 423 Basic scientific method V, 3, 4, 57, 118–120, 124, 126, 127, 132, 138, 139, 141, 148, 149, 152–157, 159, 160, 165, 170, 175, 203, 206, 208, 212, 214, 217, 219, 220, 224, 231, 244, 246, 297, 298, 408, 434, 435 Beghelli, Filippo 91 Ben-Shalom, Dorit 91 Bruening, Benjamin 350 BVA (bound variable anaphora) 10, 15, 17, 19, 44, 45, 52, 69, 71, 89–91, 94–100, 166, 173–175, 181, 186–191, 196, 199, 204, 205, 207, 208, 215, 216, 219, 227, 232, 233, 238, 241, 253, 255, 258, 266, 274, 276–287, 289–295, 307–328, 332–337, 339–343, 345–347, 349–351, 353, 360–365, 367, 371–373, 375–378, 381, 432, 438 Categorical context 214, 290 Categorical data 414 Categorical experimental result 207, 231, 447 Categorical prediction 3, 6, 24, 25, 27–29, 159, 163, 180, 329, 433, 447 Categorical-thought 181, 213 C-command VIII, 4–11, 13, 15, 21–23, 33, 41, 53, 54, 70, 71, 75, 76, 85, 89, 91, 93, 95, 99, 111, 112, 120, 124–130, 132, 133, 142, 144, 150, 153, 167, 177, 178, 193, 203, 209, 210, 213, 215, 232, 236, 329–333, 335, 338–341, 343, 344, 348–354, 364, 394–397, 403–405,

450 

 Index

407, 411, 412, 414, 422, 430–432, 437–442, 444 C-command detection (=detection of c-command effects) VIII, IX, 4–5, 7, 14, 20, 22, 121, 124, 125, 127, 132, 133, 135, 139, 141, 144–152, 154–160, 163–220, 223, 227–230, 234, 236, 241–243, 247, 248, 252, 253, 258–261, 263, 264, 268, 270–285, 290, 297, 298, 301, 336, 403, 405, 407, 414, 420, 422, 432, 433, 438 C-command pattern 148, 150, 155, 275, 290, 422 Chierchia, Gennaro 45 Chomsky, Noam 3, 13, 58, 59, 75, 117, 119, 120, 126, 138, 153, 154, 211, 220, 330, 331, 348, 349, 389, 390, 392–396, 406, 407, 444 Co-argument 208, 209 Compatibility between the participant and the design of the experiment 244, 252 Compatibility-seeking research 7–19, 429, 436, 444 Computational System (CS) IX, 3, 58–60, 62–64, 67, 69, 73–75, 77, 79, 82–85, 92–94, 99, 105, 108–111, 120–121, 134, 138–141, 153, 154, 159, 160, 166, 217–219, 224, 297, 298, 330, 331, 353–354, 358, 359, 392–399, 401, 403, 406, 409–411, 414, 415, 418, 422–427, 431 Confirmed predicted correlations of schematic asymmetries 414 Confirmed predicted schematic asymmetries 74–80, 85, 110, 447 Conjoined NP 171, 228, 246 Considerations of logic 409 Constituent 170, 348, 394, 407 Contrapositive 132, 151, 152, 155, 157, 159, 166, 180, 197–198, 200, 203, 205, 207–208, 217, 219, 223, 226, 227, 230, 231, 257, 272–273 Control (in the sense of noise control) 81, 141, 144, 158, 160, 174, 177, 208, 210, 214, 217, 218, 244, 301, 337–339, 350, 351, 377, 378, 380, 383, 410, 418, 422, 439, 441–443 Coref (coreference) 10, 21, 40, 59, 147, 175, 177, 181, 183–186, 204, 205, 207–214,

216, 219, 227, 232, 241, 247, 258, 266, 273, 290, 293, 294, 297, 301, 334–337, 339–341, 343, 345, 347, 350–353, 360, 364, 365, 371–373, 375–377, 381, 407, 430, 432 Correlational/conditional prediction about c-command detection 150–152, 154, 156, 163, 208, 414, 432 – with BVA(X, Y) 176, 186, 190, 197, 203, 210, 211, 219, 223, 229, 230, 234, 245, 270, 274, 297, 298, 432, 433 – ✶Schema-based 415 Correlational methodology 6, 14, 22, 27–29, 135, 144–152, 154–156, 158–160, 165, 174, 177, 179, 180, 184, 191, 192, 194, 208, 209, 211, 214, 216–218, 220, 224, 231, 284, 298, 299, 301, 329, 330, 343, 344, 347, 351, 353, 354, 357, 360, 370, 373, 401–403, 406, 408, 409, 414, 418, 422, 432, 433, 435–444, 447 CS-based dependencies 414 CS-based explanation 410 CS-internal 134, 135, 154, 331 dareka 89, 247 daremo 13 DD (Distributive Dependency) 208, 210, 411 Deducing definite predictions 118, 119, 127, 152, 153, 163, 194, 217, 387, 434, 435, 447 Deep and surface anaphora 22 Deep OS 47, 48, 168, 169, 338 Definite and testable predictions 18, 119, 138, 427, 434 Definite/categorical prediction 3, 163, 180 Definite correlations between values 413 Definite experimental result 119, 152, 220, 408, 429, 434, 439, 442, 445 Definite prediction 6, 15, 18, 27, 29, 118–120, 153, 159, 164, 190, 333, 358, 401, 408, 411, 413, 429, 430, 435 Definite predictions about an individual’s judgments 29 Degree of effectiveness in the demonstration 236 Degree of effectiveness of demonstration 270

Index 

Degree of reliability of demonstration 270 Degree of reliability of the experimental result 236, 271 DeMerge 138, 202 Demonstrative 34–41, 95, 96, 98, 247, 294 Dependent term 33, 40, 45, 47 Design 5, 43, 77, 78, 134, 157, 177, 219, 224, 232, 237, 242, 244, 247, 250–252, 257, 296, 298–300, 337, 351, 354, 357–361, 368, 371, 376, 377, 380–383, 399, 402, 406, 447 Detection of gravitational waves 405, 406, 419, 420 Digital infinity 120, 170, 392, 393, 404, 444 Dillon, Brian 370 Dirac equation 405 Dirac, Paul 405 Direct object 8, 13, 198 Disconfirmable 28, 62, 64, 109, 150, 177 Disconfirmable prediction 142, 144, 148, 179, 188, 189, 227, 250, 261 Disconfirmation of the prediction 17, 18, 28, 189, 257, 259, 260, 271, 273, 284, 416 Disconfirmation of universal predictions 274 Discovery of Neptune 411 Distinction between observations and facts 409 Donkey anaphora 350 dono hito-mo 13 do-no N-cm 25 do-no N(-cm)-mo 25 DR (distributive reading) 10–13, 18, 19, 59, 70–72, 74, 75, 87–93, 100–104, 111, 145, 147, 174–183, 186, 187, 189, 190, 192–198, 203–205, 207, 208, 210–214, 216, 219, 220, 223, 226–236, 238, 241, 242, 245–248, 250, 252, 253, 257, 258, 263–267, 270–277, 279, 281–283, 285–288, 290, 294–298, 301–303, 334–337, 339–341, 343, 345, 347, 351, 353, 360, 364, 365, 371–373, 375–379, 381, 407, 409, 431–433, 438, 440, 442, 443 Duhem, Pierre 82–84, 89 Einstein, Albert 406–408, 420, 421, 445, 446 Empty nominal 12, 86

 451

English 10, 16, 17, 28, 34, 35, 53–54, 70, 74, 83, 85–87, 89, 91, 96, 119, 137, 146, 152, 156, 166, 171–172, 174, 232, 237, 238, 242, 329–354, 357, 360, 361, 367, 389–391, 396, 423, 434, 437, 443 Ergative 105, 202–206, 208, 211–213, 226, 227, 440–442 every 10, 13, 14, 22, 24–26, 38, 40, 44–46, 50, 51, 53, 70–72, 83, 85, 88, 89, 91, 96, 97, 110, 125, 135, 137, 146, 147, 150, 152, 166, 167, 171, 172, 175, 180–182, 187, 196, 209, 214, 228, 230, 233, 235, 238, 239, 241, 242, 246, 258–261, 266, 271, 272, 276, 277, 291–295, 302–306, 309, 320, 332–334, 336–339, 341, 344, 345, 348–351, 360, 363, 364, 366, 367, 369, 372–375, 377–379, 393–397, 437–438 Existential demonstration of the failure of the (predicted) entailment 268, 274, 277, 284, 419 Existential prediction 64, 178, 179, 219 Experimenter’s own judgment call 416 Experimenter’s purpose 366 Expert judgment 15 Externalized sentences 121, 430 External linguistic symbols 388 Fact identification (Facts in LFS) 403, 409, 411 Faculty of logic 409 Failure of the entailment 272, 274–278, 284–289, 291–301 Failure of the predicted entailment 257, 258, 268, 274, 275, 284, 286, 298–300, 419 FD (Formal Dependency) 208–210, 216, 234, 411, 434 FD(x, y) 208, 209, 211–213, 215 – interpretation of 215 – [-loc] requirement on 208–213, 215 Fedorenko, Evelina 220 Feynman, Richard 17, 421, 444, 446 Flowchart of conducting research 57, 81–82, 89, 100 Focus on self-experiment 6, 27, 29, 159 Formal system 387, 399, 400, 402, 405, 409, 424, 427

452 

 Index

FR (Formal Relation) 135, 137, 141, 142, 144, 150, 154, 166, 174, 176, 208, 210, 218, 219, 234–236, 431, 432 FR(x, y) 135, 136, 141–144, 154, 165, 177, 208 FR-✶S 142–144, 146 FR-based MR 169, 217, 431 FR-based-MR(X, Y) 135–137, 141, 142, 208 FR-okS 142–144, 146 Fukaya, Teruhiko 22 Fukui, Naoki 138, 202 Fundamental asymmetry between ✶ and ok 17, 64–67, 142 The Galilean challenge 392, 393 Geach, Peter 350 General Relativity 405, 406, 420 Gibson, Edward 220 Graphical presentation 361 Gravitational waves 405–407, 419, 420, 446 Guess-Compute-Compare 119, 434 Haider, Hubert 119 Hankamer, Jorge 22 Hard-core 84, 411 Hard-core hypotheses 410–412 Hayashishita, J.-R. 6, 19, 45, 88, 91, 103, 145, 171, 183, 213, 214, 228, 245, 295 Hierarchically nested sets 423 Higginbotham, James 348, 349 his 10, 11, 14, 17, 22, 36, 38, 59, 60, 69–72, 74, 82, 85, 87, 98, 125, 134, 137, 140, 146, 147, 156–158, 160, 164, 166, 167, 175, 208, 215, 223, 224, 236, 241, 242, 248, 250, 253, 261, 300, 305, 306, 332–334, 336–339, 343–345, 348–352, 360–363, 365–368, 372–374, 394–397, 401, 407, 410, 412, 425, 430, 440, 446 hituyoo da 204, 440, 441 hituyoo to site iru 204, 440 Hoji, Hajime 3–29, 34, 37, 39, 40, 53, 57, 59, 60, 65, 67, 69, 73–76, 83–86, 90, 91, 94, 97, 98, 105, 108, 117–160, 163–220, 223–338, 340, 343, 344, 347, 354, 357, 361–364, 367–371, 375, 376, 378, 380, 381, 387–447 Hornstein, Norbert 349

How we visually present experimental results 261 – Blue “x” 260, 261, 273, 279, 283, 286 – Green circle 260, 261, 270, 279, 283, 286, 296, 342–343, 353 – Red star 260, 261, 263, 265, 273, 286, 287 I-language 117, 118, 122, 127, 134, 153, 155, 157, 158, 165, 166, 191, 192, 195, 202, 207, 208, 215, 217, 224, 331, 381, 396, 397, 399, 408, 415, 418, 423, 424 I-language-particular hypotheses 118, 122, 133, 158, 166–175, 191, 217, 218 Implementation 138, 156, 231, 329, 357–383, 429, 434 Implicit variable binding 53–54 Increased rigor of testing 423 Indirect object 8, 13 Individual-denoting 42–44, 52 Individual – Individual’s linguistic intuitions 3, 6, 118, 153, 435 – Individual’s linguistic judgments 19, 58 – Level of the individual 331, 354, 431, 443 Interfering factors 397 Intermediate level 391, 392 Intermediate structural level 392 Internal representations of meaning 388 iru 12, 43, 105, 202, 204, 214, 239, 305, 440 Items (in the sense of “items” in “experiments”) 3, 58, 63, 84, 91, 93, 94, 104, 110, 111, 120–123, 138, 140, 153, 172, 215, 232, 294, 299, 301, 334, 336, 337, 357, 358, 360, 361, 365, 367–381, 438 Jackendoff, Ray 350 Japanese 5, 7, 13, 15–20, 28, 33–41, 44, 52, 83, 85–89, 92–96, 98, 100, 101, 104, 105–108, 110, 111, 119, 122, 131, 136, 137, 146, 154, 156, 164–167, 171, 198, 200, 201, 214, 215, 223–329, 335, 337, 338, 343, 353, 390, 391, 396, 423, 433, 436–439, 443, 444 John-mo 13 John-mo Bill-mo 13 John-sae 10, 13, 46 John to Bill 10, 12, 13 John ya Bill 13

Index 

Judgmental fluctuation with in a single speaker 159 Judgmental instability 27, 231 Judgmental variability 231 Judgmental variations among speakers 217 Judgment shifts within a speaker 217 K17s (Kyudai 2017 Spring) 228–238, 240–242, 244–298, 300–307 kanarino kazu-no 24 kare 13, 14, 21, 22, 34, 52, 171, 189, 216, 437, 438 Kayne, Richard 13, 348, 349, 351 Kinsui, Satoshi 34, 35, 44, 294 Kitagawa, Chisato 34, 89 Kitagawa, Yoshihisa 70, 88 Kochi dialect 214 Kuno, Susumu 34, 70, 88, 89, 105, 106, 438 Kuroda, S.-Y. 35, 44, 90, 96, 181, 202, 203, 213, 441 Lack of tolerance of deviation 414 Lakatos, Imre 9, 79, 84, 410 Langacker, Ronald 350, 407 Language faculty 3–6, 9, 19, 21, 27, 57–58, 60, 117–121, 124, 126, 127, 132–134, 136, 139, 141, 148, 149, 153, 154, 156, 158, 159, 163, 165, 175, 208, 212, 214, 217, 218, 220, 224, 297, 331, 359, 383, 388, 392–394, 396–401, 403, 408, 418, 424, 427, 430, 431, 434–436, 439, 444, 447 Language faculty, initial state of 117, 118, 134, 152–154, 158, 165, 208 Language faculty, non-initial state of 117 Language faculty, steady state of 117, 118, 153, 208 Language Faculty Science (LFS) 4, 57–112, 117–160, 163–166, 220, 359, 387–447 Language faculty scientist (LFStist) 134, 156, 158, 160, 164, 208, 224, 299, 424 Language phenomenon 410 Large-scale experiment 24, 225 Laser Interferometer Gravitational-Wave Observatory (LIGO) 406, 419, 420 Lasnik, Howard 332, 350 Learned systems for mapping between sentences and meanings 392

 453

The level of the individual 331, 354, 431, 443 LF(PS) 127–129, 131–133, 135, 136, 138, 141, 143, 164, 170, 235, 306 LF(X) 126–133, 135, 136, 140, 141, 146, 150, 154, 157, 164, 165, 169, 170, 178, 230, 235, 236, 306, 431 LF representation 44, 59–67, 70, 73, 74, 76, 84, 86, 90, 95, 104, 106–109, 111, 127–131, 135, 138, 140, 168–170, 178, 201, 203, 235, 236 Liu, Feng-hsi 91 Local anaphor 28 Local relation 120, 209 Locus of testability 430 Marr, David 387 Masuoka, Takashi 105 Maximizing chances of learning from errors 57, 70, 77, 80, 82, 83, 104, 105, 109 Maximizing testability 67–77 May, Robert 87, 126, 348 Mental Lexicon 58, 84, 138, 140 Mental systems outside the CS 422 Merge 3, 4, 9, 13, 75, 84, 120–124, 132, 138, 153, 392, 394, 407, 411, 412, 414, 423, 426, 444–445 – Merge-based CS 9, 120, 121, 394, 395, 406 – recursive/iterative application of 120–124, 127, 138, 153 Method of discriminating between supported and unsupported hypotheses 429 Methodological minimalism 119, 120, 434 Michelson-Morley experiment 420 mieta 202, 206, 207, 442 Mind-internal phenomena 413 Mind-internal structure 401 Minimal paradigm requirement 73, 74 mita 205 Model of judgment making 137, 140, 141, 154, 158, 331, 371 more than one 131, 164, 166, 168, 297, 336, 377 Most fundamental preliminary experiment 85, 87, 92–108, 110–112 MR1(X, Y) 147, 148, 150, 155, 156, 164, 165, 333, 334

454 

 Index

MR2(X, beta) 147, 148, 150, 151, 155, 165, 174 MR3(alpha, Y) 147, 148, 150, 151, 155, 165, 174 MR (meaning relation) 5, 6, 10, 12, 20–23, 29, 57, 59, 62, 63, 67–73, 85, 109, 110, 125–127, 131, 133–135, 136–137, 144–148, 154, 164–175, 207, 214, 219, 232, 236, 263, 332, 334, 335, 366, 367, 375, 376, 377, 401, 403–405, 407, 411, 414–416, 418, 430–432, 435, 438, 439, 441–443 MR(S, X, Y):J 132, 141–143, 145, 146, 150, 155, 164, 306, 307 Mukai, Emi 57–112, 202, 231, 232, 245, 338 Nakai, Satoru 34, 437 nan % izyoo-no 24 N-cm 3-tu 24 N-cm 3-tu izyoo 24 N-cm sukunakutomo 3-tu izyoo 24 Need for replication 423 Negative condition 10 Newmeyer, J. Fredrick 119 Newtonian physics 410–412 Newton, Isaac 401, 404, 410, 411 Newton’s Second Law 401 NFS (non-formal source) 59, 85, 173–174, 176–179, 184, 186, 192–198, 208, 210–218, 234–236, 244, 273, 276, 281, 287, 289, 290, 292, 297, 299–300, 333, 336, 371, 399, 411, 431–433, 436–440, 442 NFS1 212, 219, 290, 292, 294, 438, 439 NFS1 effects with BVA 290, 293–295 NFS1 effects with Coref 290, 293, 294 NFS1 effects with DR 290, 295 NFS2 212, 213, 217, 219, 243, 289, 291, 296, 438, 439 NFS2 effects with BVA 212, 213, 215, 216, 280, 287, 289, 293–295 NFS2 effects with Coref 212, 213, 215, 216, 292–294 NFS-based MR 431, 432 NFS-BVA 174, 178, 186, 195, 210 NFS-BVA/Coref 177, 179, 186, 210, 234, 297 N-head identity 198, 290, 292–297 Nishigauchi, Taisuke 20 Noise control 344, 380–382, 398–400, 403, 406, 418, 421–424, 427

Non-categorical context 213, 214, 290 Non-deictic so-word 38–41 Non-finiteness 389, 390 Non-FR-based-MR(X, Y) 135, 140, 142, 144, 146, 148, 151, 155, 156, 160 Non-individual-denoting 43–46, 51, 52 Non-individual-denoting so-word 33–54 Non-researcher-participant (=naïve participant) 24, 137, 173, 224, 237, 433 Non-self-experiment 6, 28, 29, 158, 160, 197, 203, 208, 223–245, 258, 266, 270, 274, 284, 290, 297–301, 417, 418, 443, 445 Non-self-experiment, large scale 152, 177, 180, 198, 220, 225 Non-variation-based conception of judgments 431 NP(-cm)-dake 24 NP(-cm)-sae 25 NP(-cm)-sika 25 NP-igai 171, 180, 183, 187, 189, 190, 194, 212, 228, 243, 277, 287, 290, 440, 441 NP-sae 95, 100, 104 Observable 58–59, 61, 124, 154, 398, 400, 404–406, 409, 411 Ok Example 63, 64, 68, 70–72, 430 Ok Schema 63, 64, 67–74, 93–95, 98–100, 106, 108–111, 142, 430–432 Ok Schema-based prediction 63, 64, 66–69, 71, 74, 76–78, 80, 92, 93, 109, 142, 430, 431 Orbit of Uranus 411 otagai 28, 83 Other types of entailments 177, 191–196, 225, 274–297 Paraphrase 361, 362, 366, 375 Partee, Barbara 53 Participant 24, 28–29, 137, 140, 143, 145, 164, 223, 225, 228, 232, 233, 235–245, 248, 250–253, 257, 259, 260, 263, 264, 268, 277, 291, 292, 295, 298–300, 335, 336, 340, 341, 345, 347, 353, 357, 358, 360–371, 375–382, 416 Passive 337, 341, 373 Penrose, Roger 406, 425 Penrose’s three worlds 427, 428

Index 

PF representation 60–65, 69, 109, 138, 140 Plato 390 Plato’s problem 390, 392 Plesniak, Daniel 3, 4, 14, 85, 91, 121, 126, 132, 136, 141, 146, 152, 163, 169, 174, 175, 177, 197, 225, 231, 232, 237, 238, 252, 297, 301, 329–354, 357–383, 387–447 Plural-denoting 14, 15, 19, 20, 21, 40, 46, 98, 240, 241, 243, 244 Poincaré, Henri 299, 412, 422, 423, 444 Popper, Karl 57 Positron 405, 407 Possessor 97, 330, 348–352, 354 Possessor-binding 330, 350 Possibility-seeking 225, 359, 367, 369, 381, 414 Postal, Paul 333 Potential c-command detection 259, 273 Potential disconfirmation of the prediction 259, 273 Predicted correlations of schematic asymmetries 409 Predicted schematic asymmetries 74–79, 92, 106, 108, 110, 447 Prediction deduction 5, 135, 399, 400, 409, 412 Prediction-testing experiment 420 Predicted with 100% accuracy 397 pro 14, 15, 18–21, 27 Probe 19, 330, 394–395, 397, 432–433, 436, 439, 442 PS (phonetic sequence) 123, 125–129, 131–135, 138, 139, 141, 142, 157, 164, 306, 442 Qualified speaker 25–27, 29, 251, 258, 264, 266, 268, 269, 272, 274, 284 Qualitative judgment 398 Qualitative judgment data 414 Quirky binding 33, 45–54, 59, 85, 88 Quirky MR 144 Randomization 369, 370 Reconstruction 330, 345–351, 353 Referential 35–41, 46–47, 189, 362, 363, 394, 395

 455

Reinhart, Tanya 22, 23, 33, 49, 75, 76, 332, 333, 348, 351, 394–396, 407 Reliably disconfirming incorrect hypotheses 429 Reliably measurable phenomena 387 Repeatability 390, 398, 402 Replication 29, 119, 121, 152, 155, 156, 158, 164, 223–328, 338, 353, 377, 381, 382, 398–400, 402, 408, 412, 415–417, 419, 420, 423, 434 Replication across (substantially) different I-languages 134, 399 Replicational demonstration 416, 433 Replication of correlations 399 Replication, need of 423 Research heuristics 57–112 Resourcefulness 225, 300, 379 Responses from an individual 29 Responses of the group of speakers 29 Rigorous attempt at disconfirmation  15, 415 Rigorously testable prediction 4, 138 Rigorous testability 17–19, 21, 60, 135, 148, 192, 287, 331, 354, 434, 435, 447 Ross, Claudia 89 Ruys, Eddy 91 Sag, Ivan 22 Saito, Mamoru 90 Saliency 49–51 Schematic asymmetries 57, 70, 74, 77, 82, 85, 86, 109, 110, 381, 409, 414, 427 Schütze, Carson 358, 360 Scientific integrity and honesty 422, 423 Self-experiment 5, 152, 158, 160, 176, 219, 223, 224, 237, 243, 274, 299, 300 Self-experiment as the most basic experiment in language faculty science 160 Sentence-MR pairing 404 Sentence template 430 Shared inborn intermediate level 392 Shibatani, Masayoshi 105, 106 Singular-denoting 24, 25, 27, 40, 97, 98, 166, 175, 215, 240, 243, 244, 284, 332 Singular-denoting test 240, 279, 283, 285, 286, 292 Sloppy identity 21–23, 442

456 

 Index

soitu 20, 25, 27, 98, 171, 183–191, 194, 195, 199, 212, 213, 215, 216, 228, 239, 240, 242–244, 246–248, 253, 258, 259, 263, 264, 266, 268, 271, 272, 277–281, 287, 291–296, 303, 304, 317, 318, 320, 322, 324, 326 soko 20, 21, 25–27, 171, 172, 183–185, 187–191, 199, 215, 216, 228, 239–244, 246, 258, 260, 264–268, 271, 272, 276, 277, 279–283, 285–287, 291–294, 296, 307, 309, 311, 313, 315, 317, 437 sono otoko 172, 183–191, 193–196, 199, 204, 212, 213, 215, 216, 228, 240, 243, 284, 288–291, 293–295, 440, 441 SO-type 33, 47, 48 Source of a given judgment 415 So-words 35, 38–44, 51, 53 Speaker judgments 16, 19, 23, 27, 58, 64, 69, 74, 93, 110, 111, 132, 134, 147, 174, 231, 300 Spec 348 Specific threshold chosen for determination of J=yes 252, 300 Specifier 348 Specifier-binding (also Spec-binding) 330, 348–353 Split antecedence 20, 21, 97, 240, 241, 243 Split antecedence test 20, 25, 27, 240–244, 264, 266–268, 284, 286 SR (semantic representation) 140, 141 Stages of Hoji’s I-language – Stage 1 184, 188–190, 194, 217, 243, 281, 294, 295 – Stage 2 184, 188–190, 193–196, 204, 212, 213, 217, 240, 243, 281, 288–291, 293–295, 440, 441 – Stage 3 185, 188–190, 213, 217, 243, 438 Stowell, Tim 91 Structure-building module 392 subete-no 24–26, 45, 46, 49–51, 88, 89, 91, 95, 100, 104, 172, 180, 187, 196, 228, 238, 241, 242, 246–248, 253, 258–261, 263–268, 271–273, 277–283, 285–295, 302–306, 320, 437, 438 subete-no gisi 182, 196, 288–290, 293–295 Sub-experiment 20, 24, 335–337, 340, 357, 363–365, 378–381, 402, 415, 418, 421

Sub-preliminary experiment 237–244, 250–252, 264–272, 281, 285, 286, 291, 296, 298, 299, 307–327 suisensita 45, 46, 50, 51, 88, 209 sukunakutomo 3-tu izyoo-no 24, 26 sukunakutomo numeral-cl-no 180 Supporting evidence for hypotheses 15 Surface OS 202 Systematic deviations from c-command patterns 422 Takano, Yuji 138, 202 Takubo, Yukinori 34, 35, 44 Telescope 406, 408, 436 Testability-seeking research 15, 18, 28–29, 192, 429, 434, 435, 439, 444, 445 Testable 4, 9, 15, 18, 19, 58, 109, 119, 120, 126, 129, 138, 149, 150, 152, 179, 194, 196, 227, 250, 257, 331, 338, 398, 402, 410, 427, 434, 445 Testable correlational/conditional prediction 150–152, 154, 156, 227, 242, 245, 270, 432 Testable correlational prediction 150, 217, 218 Theta-role 169, 337–339 Thorne, Kip 406 Tool for detecting syntactic structure 406 Transitive ni-marking (also, ni-marking transitive) 202, 204, 205, 208, 211–213, 226, 228, 234, 245, 247, 440, 441 Transitive o-marking (also, o-marking transitive) 204, 211–213, 226, 228, 234, 235, 245, 440 Typicality-seeking 225, 359, 360, 370, 414 Ueyama, Ayumi 6, 10, 19, 27, 33–54, 59, 60, 85, 88, 89–91, 94, 96, 131, 136–140, 144, 146, 154, 157, 164, 168, 171, 183, 202, 213–215, 217, 225, 228, 239, 245, 331–333, 350 V-tyuu 214 V-yuu 214 Wagers, Matthew 370 Wasow, Thomas 333

Index 

Weak crossover 70, 329, 330, 333, 335, 336, 337, 339–344, 347, 349, 352, 353, 357, 360, 364, 371, 372, 377, 381, 382 Weiss, Rainer 406 What counts as MR([+cc]S, X, Y):yes 259 What counts as MR([-cc]S, X, Y):no 259

Yoshimura, Noriko 20 zibun 13, 14, 26, 238, 437, 438 zibun-zisin 28, 83

 457