This page intentionally left blank Binding Theory Binding Theory seeks to explain how different kinds of nominal expressions such as names, noun phrases, and pronouns have anaphoric relations among one another, and how they come to have reference to things in the world. This textbook provides a thorough and comprehensive introduction to modern Binding Theory. Starting at a very basic level, it introduces the reader to a huge variety of nominal and especially pronominal expressions from the world's languages, the ways they can be used, and current theorizing about their grammatical properties and their interpretation. Daniel Buring discusses a wide range of cross-linguistic data and theoretical approaches, and, unlike in existing introductions, pairs the discussion of syntactic facts with a detailed introduction to the semantic interpretation of binding structures. Written in a clear and accessible style, and with numerous exercises and examples, this textbook will be invaluable to graduate and advanced undergraduate students of syntax and semantics. daniel buring teaches linguistics at the University of California, Los Angeles. He has published various influential articles in formal semantics, syntax, and pragmatics, in particular on intonational meaning, focus, and binding theory. He has previously published The Meaning of Topic of Focus: The 59th Bridge Street Accent (1997). CAMBRIDGE TEXTBOOKS IN LINGUISTICS General editors: p. austin, j. bresnan, b. comrie, s. crain, w. dressler, c. j. ewen, r. lass, d. w. lightfoot, k. rice, s. romaine, n. v. smith Binding Theory In this series: P. H. MATTHEWS Morphology second edition B. COMRIE Aspect R. M. KEMPSON Semantic Theory T. BYNON Historical Linguistics J. ALLWOOD, L.-G. ANDERSON, and O. DAHL Logic in Linguistics D. B. FRY The Physics of Speech R. A. HUDSON Sociolinguistics second edition A. J. ELLIOT Child Language P. H. MATTHEWS Syntax A. RADFORD Transformational Syntax L. BAUER English Word-formation S. C. LEVINS ON Pragmatics a. BROWN and G. YULE Discourse Analysis R. HUDDLESTON Introduction to the Grammar of English R. LASS Phonology A. COMRIE Tense W. KLEIN Second Language Acquisition A. J. WOODS, P. FLETCHER, and A. HUGHES Statistics in Language Studies D. A. CRUSE Lexical Semantics A. RADFORD Transformational Grammar M. G ARM AN Psycholinguistics G. G. CORBETT Gender H. J. GIEGERICH English Phonology R. CANN Formal Semantics J. LAVER Principles of Phonetics F. R. PALMER Grammatical Roles and Relations M. A. JONES Foundations of French Syntax A. RADFORD Syntactic Theory and the Structure of English: a Minimalist Approach R. D. VAN VALIN, JR, and R. J. LAPOLLA Syntax: Structure, Meaning and Function A. DURANTI Linguistic Anthropology A. CRUTTENDEN Intonation second edition J. K. CHAMBERS and P. TRUDGILL Dialectology second edition C. LYONS Definiteness R. KAGER Optimality Theory J. A. HOLM An Introduction to Pidgins and Creoles C. G. CORBETT Number C. J. EWEN and H. VAN DER HULST The Phonological Structure of Words F. R. PALMER Mood and Modality second edition B. J. BLAKE Case second edition E. GUSSMAN Phonology: Analysis and Theory M. YIP Tone W. CROFT Typology and Universals second edition F. COULMAS Writing Systems: an Introduction to their Linguistic Analysis P. J. HOPPER and E. C. TRAUGOTT Grammaticalization second edition L. WHITE Second Language Acquisition and Universal Grammar I . PL AG Word-formation in English W. CROFT and A. CRUSE Cognitive Linguistics A. SIEWIERSKA Person A. RADFORD Minimalist Syntax: Exploring the Structure of English D. BURING Binding Theory Binding Theory DANIEL BÜRING University of California, Los Angeles CAMBRIDGE UNIVERSITY PRESS cambridge university press Cambridge, New York, Melbourne, Madrid, Cape Town, Singapore, Sao Paulo Cambridge University Press The Edinburgh Building, Cambridge cb2 2ru, UK Published in the United States of America by Cambridge University Press, New York www.cambridge.org Information on this title: www.cambridge.org/9780521812801 © Daniel Buring 2005 This book is in copyright. Subject to statutory exception and to the provision of relevant collective licensing agreements, no reproduction of any part may take place without the written permission of Cambridge University Press. First published in print format iSBN-13 978-0-511-08112-5 eBook (NetLibrary) isbn-io 0-511-08112-x eBook (NetLibrary) iSBN-13 978-0-521-81280-1 hardback isbn-io 0-521-81280-1 hardback iSBN-13 978-0-521-01222-5 paperback isbn-io 0-521-01222-8 paperback Cambridge University Press has no responsibility for the persistence or accuracy of urls for external or third-party internet websites referred to in this book, and does not guarantee that any content on such websites is, or will remain, accurate or appropriate. Contents Preface page ix 1 The ABC of Binding Theory 1 1.1 Preliminaries 1 1.2 Binding 4 1.3 Command and precedence© 12 1.4 Reflexive verbs and reflexive phraseologisms© 21 2 Interpreting indexed structures 25 2.1 Basics of interpretation 25 2.2 Enter indexing 26 2.3 Compositional interpretation 31 2.4 Extensions and alternatives© 40 3 Domains and orientation 46 3.1 Binding domains in English: governing category 46 3.2 Orientation 58 3.3 Binding domains cross-linguistically 65 3.4 Long-distance reflexives 3.5 Some pronominal systems 75 4 Binding versus coreference 81 4.1 Quantified NPs and variable binding 83 4.2 The syntax of semantic binding 89 4.3 WTz-expressions 93 4.4 Summary 95 4.5 Semantic details© 96 5 Other cases of semantic binding 104 5.1 Overture 104 5.2 Focus constructions 105 5.3 Double indexing 109 5.4 A new system 110 5.5 Verb phrase ellipsis© 114 6 The Coreference rule 118 6.1 The proposal 118 6.2 Theoretical consequences 119 vii viii CONTENTS 6.3 Binding Theory obviations 126 6.4 Summary: the final system 128 6.5 Extensions 130 7 Descriptive pronouns and individual concepts 143 7.1 Anaphoric pronouns that don't corefer 143 7.2 Unknown and mistaken identity 151 7.3 Descriptive NPs and indexing© 156 7.4 Summary 159 7.5 An extension: unexpected sloppy identity 160 8 Semantic binding and c-command 163 8.1 The weak crossover phenomenon 163 8.2 Blocking weak crossover 166 8.3 A challenge: indirect binding 174 9 Plurals 188 9.1 The semantics of plural NPs 188 9.2 Anaphoric relations between plural NPs 191 9.3 Set indexation 193 9.4 More on overlapping reference© 199 10 Reciprocals 203 10.1 Plural preliminaries 204 10.2 Strong reciprocity 206 10.3 The syntax of reciprocal binding 209 10.4 Alternative meanings for reciprocal sentences 213 10.5 Reflexives and reciprocals 220 11 Exempt anaphora and reflexivity 222 11.1 Introducing exempt anaphora 222 11.2 Conditions on exempt anaphora 225 11.3 On the notion of higher coargument 227 11.4 Reflexivity Theory 235 11.5 Towards a cross-linguistic perspective 242 12 Binding and movement 244 12.1 Argument movement 244 12.2 W7z-movement 246 12.3 Analytical options 248 12.4 An apparent case of binding after w/z-movement 254 12.5 A real case of interaction of A-movement and BT? 256 12.6 Binding without binders 260 Bibliography 265 Index 277 Preface This book presents a comprehensive treatment of the syntax and semantics of binding. It is meant to fill the gap between existing introductory texts, both semantic and syntactic, and the rich primary research literature on the topic. If you work your way through this book, you should be able to read and understand almost any of the works mentioned in the references. There are at least two reasons why I thought such a book may be useful. First, Binding Theory figures prominently in a vast amount of works, either as the main research topic, or, perhaps even more frequently, as a diagnostic for constituency, derivational history, and other abstract aspects of grammatical analysis. I felt that an accessible survey of some of the more recent insights into the nature of binding would benefit both those who read those studies, as well as those who want to undertake them in the future. Second, by its very nature, Binding Theory involves an equal amount of syntax and semantics. As such, it recommends itself as the topic for an advanced level textbook. There is, I believe, no insightful syntactic analysis without a solid semantics to access its adequacy; in any event, there certainly can't be any insightful analysis of the syntax of binding without a semantics to accompany it. The present book, therefore, is an introduction to doing syntactic and semantic analysis side by side. It attempts to show you how to do semantically realistic (or responsible) analysis; it will also show you how, at least in some cases, figuring in the semantics carefully may solve some problems that would seem recalcitrant from a purely syntactic point of view. It's good old divide et impera. The book is organized as follows: the first six chapters develop, in incremental steps, the basic system of NP classification, indexing, and interpretation. They each crucially build and expand on the content of the preceding ones, and should be tackled in that order. Chapters seven through twelve then extend the basic system in various, sometimes opposite, directions, and can be accessed mostly independently of each other; this structure is schematized in the chart below. Within chapters, certain sections are marked as (f), for "extension"; these often contain more advanced and demanding material, and can be skipped without loss of coherence for later chapters (except possibly the ©-parts therein). ix x preface The ABC of Binding Theory Interpreting indexed structures Domains and orientation Binding versus coreference Other cases of semantic binding The Coreference rule Plurals Reciprocals Descriptive pronouns ... \ Semantic binding and c-command Exempt anaphora Binding and Movement I have attempted to introduce explicitly every piece of machinery used in the analysis, and make all assumptions explicit. I have also included a fair number of exercises, especially in the earlier chapters, that should help to master the material, but also to discover problems and open ends. Despite that, I think that a certain familiarity with linguistic argumentation, as well as with formal syntactic and semantic analysis is required to read this book. Most introductory textbooks should provide the necessary background. When Cambridge University Press invited me to write this book, I had taught 'The Syntax and Semantics of Binding Theory' at a couple of summer schools, and the plan was essentially to flesh out the existing course materials. In the process of writing the book, more and more literature made its way into these materials, and the scope of the book extended considerably. Still, this book is not a natural history of binding phenomena, especially not cross-linguistically, and makes no claim to do justice to the vast theoretical and especially descriptive literature, of which only a fraction is taken into consideration here. While I tried to use examples from many different languages, where I had sufficient sources, the primary language analyzed is English. And even there, I found that the reported judgments are often very subtle and highly controversial. I sincerely believe now that much more systematic primary work on establishing a firm data base needs to be done; as it is, I mostly report the data as given in the literature, pointing out points of controversy, and occasionally supplementing native speaker judgments I elicited. There are also some areas that are omitted altogether in this book, mostly for reasons of space, among them the diachronic changes in anaphoric systems (van Gelderen [2000]; Keenan [2002]), as well as their acquisition in young children (Wexler and Manzini [1987], a.o.). Furthermore, older theoretical approaches to Binding Theory are not discussed, though they might often Preface xi facilitate understanding more recent approaches (I recommend the first chapters of Kuno [1987] for an excellent overview). More people than I can mention here have helped me in the process of writing this book. I'd like to thank in particular Ed Keenan and Philippe Schlenker, my colleagues here at ULCA, for their input, and Daniel Hole and Chris Potts for their extremely detailed comments and suggestions; thanks also to Christina Kim for helping with the final proofs. Special thanks go to Summer Kern, my Herzallerliebste, for her support, encouragement, and patience, and for always (perhaps reluctantly) being willing to double-check yet another sentence or two. 1 The ABC of Binding Theory 1.1 Preliminaries 1.1.1 Reference, coreference, and indexing i What is Binding Theory (BT) about? To a first approximation, BT restricts the distribution of NPs (or DPs, if you prefer) that have the same referent (starting with chapter 4, we will add non-referential NPs to the picture, which will be ignored until then). We will indicate sameness of reference, coreference for short, by coindexing; that is, coreferent NPs carry the same index, for which we use integers throughout. Thus in (1.1), the NP the baroness and the NP she are coindexed, which signals that they are coreferent, which in turn means that they have the same referent - they refer to the same person or thing - namely the actual baroness in flesh and blood: (1.1) After [mp the baroness] i had visited the lord, [np she]i left the house. Note that on this understanding, BT is relevant for nominal categories only, and only for the maximal projections, i.e. NPs.1 As a convention we assume that two NPs corefer if and only if (iff) they are coindexed. Contra-indexing (or lack of an index on either NP) indicates non-coreference. This is illustrated in (1.2): (1.2) (a) After [np the baroness]2 had visited the lord, she2 left the house. (she=the baroness) (b) After p the baroness]i had visited the lord, p she]2 left the house, (sheathe baroness) It should be noted that the actual choice of integer is irrelevant; (1.1) expresses the same coreference pattern as (1.2a) (as would any sentence in which both occurrences of the index are replaced by the same integer). An NP marked 1 is in no sense prior, higher, or superior to one marked 2. All that matters is which NPs have the same index, and which do not. 1 The latter aspect I consider a genuine fact about Binding Theory. On the view pursued here, indexing on non-maximal projections (e.g. signalling specifier-head agreement or head-movement dependencies) simply is not subject to Binding Theory and should be kept separate from it. As for the former aspects, though there are sentential and adverbial (i.e. PP-) anaphors, little work on their distribution has been done, and we will ignore them here (see e.g. Hegarty et al. [2001] and the references therein). 2 THE ABC OF BINDING THEORY In traditional grammars, the NP the baroness in (1.1) is referred to as the antecedent of the pronoun she. We adopt the following: (1.3) Definition: Antecedent A is the antecedent of B iff (if and only if) (i) A precedes B, and (ii) A and B corefer. By our convention, an NP will be coindexed with its antecedent (if it has one). This holds for coreferring NPs within a single sentence, and across sentences. The latter, however, are usually not subject to Binding Conditions of the sort discussed here.2 1.1.2 The basic data i Restricting our attention to singular NPs for the time being, two NPs in a given sentence will show one of three logically possible coreference relations (Reinhart, 1983a: 29): (1.4) (a) obligatory coreference: Zelda bores herself (b) obligatory non-coreference: She adores Zelda's teachers. (c) optional coreference: Zelda adores her teachers. Given what was said before, grammatical representations for these will look like in (1.5): (1.5) (a) Zeldai bores herselfi. (b) She8 adores Zeldais's teachers. (c) Zelda4 adores her4 teachers. or Zelda4 adores hery teachers. Ungrammatical representations for (1.4a) and (1.4b) are given in (1.6): (1.6) (a) *Zeldai bores herself2. (b) *She8 adores Zeldas's teachers. It will be convenient to summarize patterns as in (1.5) and (1.6) as shown in (1.7), whose logic should be transparent: (1.7) (a) Zeldai bores herself 1/^2- (b) She8 adores Zeldais/^'s teachers. (c) Zelda4 adores her4/7 teachers. The key insight captured in BT is that the (un)availability of coreference between two NPs crucially depends on two factors: 2 See e.g. Grosz et al. (1995); Gundel et al. (1993); Walker et al. (1998) and the references therein for some discussion of trans-sentential anaphora. 1.1 Preliminaries 3 • the morphological shape of the NPs • the structural relation between the NPs This is not meant to exclude the possibility of additional factors that influence coreference options (which will be discussed especially in chapters 3 and 11). First, however, we will introduce the relevant NP-types of English and then, in turn, explore and characterize the syntactic configurations in which they require, allow, or disallow coreference. 1.1.3 Three types of NPs i Virtually all approaches to BT in English distinguish three types of NPs by (mostly) morphosyntactic criteria. These are illustrated in (1.8a-1.8c): (1.8) (a) reflexives and reciprocals ('anaphors'): himself, herself, itself, themselves, myself, yourself, ourselves, yourselves each other, one another (b) non-reflexive pronouns ('pronominals'): he, she, it, him, her, I, us, you, me, his, your, my, our (c) full NPs including names ('r-expressions'): the baroness, Peter, this, a disinherited Russian countess ... In parentheses I have given the terms for these categories as used in the influential work of Chomsky (e.g. 1981) and his school: anaphor, pronominal, and r-expression (with r reminiscent of 'referential'). For the first two, a cautionary remark is in order, because they unfortunately provide potential for confusion: traditionally the term anaphor (often with the plural anaphors rather than anaphora) is used for any NP, reflexive or not, that has an antecedent. Likewise, the term pronominal invites confusion with the traditional notion of pronoun, which applies to reflexive and non-reflexive pronouns alike. We will thus stick to the terms 'reflexive/reciprocal', 'non-reflexive pronoun', and 'full NP' in the remainder of this book. We will now motivate this tripartition, starting with reflexives versus the rest (reciprocals, being necessarily plural, will not be discussed until chapter 10). Consider the sentences in (1.9): (1.9) (a) That it rains bothers Peter. (b) That it rains bothers her/him. (c) * That it rains bothers himself/herself. All these sentences contain but one referential NP (the expletive it is of no interest to BT, since it lacks a referent - and perhaps semantic content in general). We can thus omit the indexing for expository convenience, given that no coreference is involved. We simply observe that reflexives cannot occur in this configuration, while both non-reflexive pronouns and full NPs can. 4 the abc of binding theory Table 1.1 Distribution of the three NP-types configuration ex. reflexive non-reflexive full NP no antecedent (1.9) * ok ok non-local antecedent (1.11) * ok * local antecedent (1.10) ok * * Inversely, only reflexives, but neither non-reflexives nor full NPs, are permitted in (1.10): (1.10) (a) * Peter3 watches Peter3 in the mirror. (b) * Peter3 watches him3 in the mirror. (c) Peter3 watches himself3 in the mirror. (Note that the two occurrences of Peter in [1.10a] are coindexed, indicating that we speak about the same Peter. The sentence is presumably acceptable if I point at a different Peter upon using the names, just as [ 1.10b] is of course grammatical if the pronoun is not coindexed with the name.) Let us finally turn to the difference between non-reflexive pronouns and the rest, illustrated by way of the sentences in (1.11): (1.11) (a)* Carla4 thinks that I hate Carla4. (b) Carla4 thinks that I hate her4. (c) *Carla4 thinks that I hate herself4. Here, reflexives pattern with full NPs, and in contradistinction to non-reflexive pronouns. Note that the difference between (1.10) and (1.11) is not the absence versus presence of an antecedent (there is one in each), but seems to be one of syntactic locality: the antecedent NP is within the same clause as the anaphor in (1.10), but in a higher clause in (1.11). We summarize these (preliminary) results in table 1.1. What is clear from this table is that at least this three-way distinction needs to be recognized to distinguish correctly the coreference options of NPs in English. Notice also that reflexive and non-reflexive pronouns seem to be in complementary distribution. We will now characterize the conditions for coreference for the three types of NPs in turn. 1.2 Binding 1.2.1 Reflexive and non-reflexive pronouns i We observed above that reflexive pronouns require an antecedent, and an antecedent within their local clause at that. This is illustrated in more detail in (1.12): 1.2 Binding 5 (1.12) (a) * That it rains bothers himself/herself. (no antecedent) (b) *Carla4 thinks that I hate herself^. (non-local antecedent) (c) Peter2 watches himself2 in the mirror. (local antecedent) Turning now to non-reflexive pronouns, recall that they can occur with or without a sentence-internal antecedent, cf. (1.13), as long as the antecedent is not in the same local clause, cf. (1.13c): (1.13) (a) That it rains bothers him/her. (no antecedent) (b) Carla4 thinks that I hate her4. (non-local antecedent) (c) * Peter3 watches him3 in the mirror. (local antecedent) Based on these data we formulate our first version of the Binding Conditions: (1.14) Binding Conditions (preliminary) (A) A reflexive pronoun must have an antecedent within its local clause. (B) A non-reflexive pronoun must not have an antecedent within its local clause. (1.15) Ancillary definition: a is within 0's local clause if a and 0 are dominated by the same set of clausal nodes (S, S, IP, CP, TP, AgrP...). Exercise 1.1 In the following sentences, designates an NP with the index given. For each sentence, determine by intuition what can/must be (there may be more than one option in some cases). Then give the local clause and the antecedent for and demonstrate that the Binding Conditions in (1.14) are met (example: O3 in [1.16a] must be himself, its local clause is the matrix S/IP, and its antecedent is Peter, which is, correctly, in the same local clause): (1.16) (a) Peter3 watches O3 in the mirror. (b) Mashas believes that the swamp elks admire O5. (c) Mashas believes that [the swamp elks]i6 admire Oi6. (d) Mashas introduced O5 to the swamp elks. (e) Hermann^ tried to be nice, and Gallia quite liked Now $8 and Gallia go out to see a mud wrestling show. (f) Mashas mentioned a swamp elk that was important to O5. (g) Oi's manager takes care of Ceciliai's business. (h) Oi takes care of Ceciliai's business. 1.2.2 Binding and binder 1 Before going on, we need to refine our previous treatment in one small but significant way. To see why, consider (1.17): (1.17) (a) Carlottan's dog accompanies hern/6 to kindergarten. (b) * Carlottai 1 's dog accompanies herself 11/6 to kindergarten. The judgments in (1.17) are the reverse of what the Binding Conditions lead us to expect: Carlotta is clearly in the same local clause as her/herself, yet we 6 the abc of binding theory have to choose a non-reflexive pronoun to express coreference. This is in marked contrast to our earlier example (1.10), repeated here, which led to the formulation of the Binding Conditions above: (1.18) Peter3 watches himself3/*hhxi3 in the mirror. One difference is that Peter and himself in (1.18) are clausemates, whereas Car-lotta and her(self) in (1.17) are not - Carlotta is the possessor to the subject, but only the subject and her(self) are clausemates. We can flesh out the notion 'clausemate' in various ways, e.g. as 'be arguments to the same predicate' (here: watch), or 'be immediate constituents of the same clause,' with subtly different results, as we will discuss immediately in sections 1.2.4 and 1.3. Postponing a precise definition of clausemate, let us say that only an antecedent which is a clausemate to an NP can be a binder for that NP: (1.19) Binding (preliminary): NPi binds NP2 if and only if (iff) (a) NPi and NP2 are coindexed (b) NPi precedes NP2 (c) NPi and NP2 are clausemates. Then NPi is the binder of NP2, and NP2 is bound (by NPi). (1.19a) and (1.19b) are the same as in the definition of antecedent in (1.3) above, but clause (1.19c) is added. A binder, then, is simply an antecedent that is a clausemate of the bindee. We now replace the notion of 'have an antecedent' with the notion of 'be bound' in the Binding Conditions: (1.20) Binding Conditions (still preliminary): (A) A reflexive pronoun must have a binder within its local clause. (B) A non-reflexive pronoun must not have a binder within its local clause. In (1.18), repeated in (1.21a) below, Peter qualifies as a binder with respect to the pronoun in the object position of watch - it is coindexed with it, precedes it, and, being the subject of watch, is a clausemate. Hence Binding Condition A licenses a reflexive in object position, and Binding Condition B prohibits a non-reflexive. All's well: (1.21) (a) Peter3 watches himself3/*him3 in the mirror. (b) Carlottan's dog accompanies fieri i//*herselfn to kindergarten. In the formerly problematic example (1.17), repeated in (1.21b) above, Carlotta is not a binder to the pronoun in the object position of accompany (though it is an antecedent); it is coindexed with it, and precedes it, but, being a modifier to dog rather than an argument to accompany, it fails on the clausemate condition in the definition of binder (1.19c). Binding Condition A thus prohibits a reflexive, and Binding Condition B allows a non-reflexive. 1.2 Binding 7 1.2.3 Full NPs I Turning now to full NPs, we observed that they cannot occur with a sentence internal antecedent at all, regardless of whether the antecedent occurs within the same local clause or not. The relevant data are repeated here: (1.22) (a) That it rains bothers Peter. (no antecedent) (b) *Carla4/she4 thinks that I hate Carla^ (non-local antecedent) (c) *Peter3/he4 watches Peter3 in the mirror. (local antecedent) The question that comes up is whether full NPs are allergic to antecedents, or just binders. To decide that question we have to look again at a case in which an NP antecedes a full NP without actually binding it, for example (1.23): (1.23) (a) Hern d°g accompanies Carlottan to kindergarten. (b) ?Carlotta'sn dog accompanies Carlottan to kindergarten. (c) Carlotta'sn dog accompanies the little darlingn to kindergarten. The pronoun in (1.23a) antecedes the full NP with no loss in acceptability. And even another full NP can, as in (1.23b), which is slightly degraded due to the repetition of the name, but head and shoulders above (1.22b); and (1.23c), which features an epithet, i.e. a definite NP which is coreferential with, though different in descriptive content from, its antecedent, is impeccable. We conclude that, just as in the principles governing the coreference options of pronouns, the principle responsible for full NPs must make reference to the notion of binding, rather than antecedence: (1.24) Binding Condition C: A full NP must not be bound. I should like to point out here that the judgments in (1.23), while widely accepted, are not uncontroversial. Generally, name-name cases (Peter■? .. .Peter?,) seem more acceptable than pronoun-name cases (he? .. .Peter?) and for many speakers approach the degree of acceptibility found in examples like (1.23b) (cf. e.g. Bach and Partee [1980], note 11; Evans [1980]:356 a.o.). This can be seen as a phenomenon outside of grammar (after all, in the double name cases, the coreferential reading is the only way to interpret the sentence at all, while in the pronoun-name cases, there is a host of grammatical non-coreferent readings) or as a fact about BT proper, suggesting that Binding Condition C should only ban full NPs from being bound by a pronoun Bach and Partee [1980]; Keenan [1974]; for further discussion see also Bresnan [2000], Lasnik [1986], as well as chapter 6. We will, for the time being, assume these cases to be unequivocally bad. 1.2.4 C-command 1 Before closing, we need to generalize the notion of binding slightly. As it stands, Binding Condition C does not exclude (1.22b), repeated here: (1.25) *Carla4/she4 thinks that I hate Carla4. The reason is that (the first occurrence of) Carlalshe in (1.25) doesn't bind the second in the technical sense defined in (1.19), because they are not clausemates: 8 the abc of binding theory they are not immediate constituents of the same clause, nor are they arguments to the same verb (think versus hate). We therefore replace the notion of clausemate by a more general, asymmetric, notion, that of c(onstituent)-command [Reinhart, 1976]: (1.26) Node A c-commands node B in a phrase marker iff (a) neither dominates the other, and (b) every (branching) node that dominates A also dominates B3 (1.27) Binding (revised, still preliminary): NPi binds NP2 iff (a) NPi and NP2 are coindexed (b) NPi precedes NP2 (c) NPi c-commands NP2 Then NPi is the binder of NP2, and NP2 is bound (by NPi). Let us first verify how these new definitions subsume the old ones. Take (1.21a), repeated here; a phrase structure tree for this sentence will have the essential constituency shown in (1.28): (1.28) Peter3 watches himself3/*hini3 in the mirror. s the The only (branching) node dominating [NP Peter]3 is S, which means that [NP Peter]3 c-commands VP and everything dominated by VP, including [NP himself/him]3. Thus [NPPeter]j, is a binder for [NPhimself/him]?,, and, given that it is in the same local clause, it is correctly predicted that the latter has to be a reflexive, rather than a full NP or a non-reflexive pronoun. Contrast this with (1.21b) repeated here along with a simple tree diagram: (1.29) Carlottan's dog accompanies hern//*herself 11 to kindergarten. 3 Definitions in the literature usually include the qualification 'branching', even though, as Barker and Pullum [1990] and Pullum [1986] note, this is rarely argued for, nor required, by the data in any obvious way. The cases discussed in this book provide no exceptions to that; indeed the notion of semantic binding to be introduced in chapter 4 directly embodies Pullum's stricter and arguably more natural notion of IDV-command, according to which a constituent's c-command domain simply consists of its sister constituent(s). 1.2 Binding 9 s the kindergarten Here, NPn, Carlotta, does not c-command VP or anything therein: nodes dominating NPn are NPs and S, which means that NPn merely c-commands the N dog; VP, and the pronominal NPn within it, though dominated by S, are not dominated by NPs, which means they are not dominated by every branching node dominating NPn, Carlotta, as is required for binding due to (1.26b). Accordingly, [NP her(self)]]] is not bound by [NP Carlotta ]n by the new definition of binding, especially (1.27c), so that the Binding Conditions correctly predict a non-reflexive (or a name) in that position. Crucially, the new definition of binding is 'downward unlimited', because an NP that c-commands a node A also c-commands every node dominated by A. This is the key to handling the Binding Condition C cases. Consider again (1.22b), repeated here: (1.30) * Carla4/she4 thinks that I hate Carlaz^ s hate Similar to (1.21a), the matrix subject NP4, Carla, c-commands the matrix VP, and everything dominated by the matrix VP, including the object NP4. Since the subject NP4 is also coindexed with the object NP4 and precedes it, it qualifies as a binder. Binding Condition C then excludes a name as the object NP4, while Binding Condition B allows a non-reflexive pronoun in that position. This completes our introduction to the ABC of Binding Theory for English. It should be stressed that the Binding Conditions as stated above are no longer 10 the abc of binding theory about the traditional, intuitive concept of antecedence, but about a more abstract concept, binding. Binding Theory, so construed, is then a theory only about a subset of anaphoric relations, excluding non-c-command anaphora, both across and within sentences. This embodies a strong and non-obvious hypothesis, namely that c-command, or some other command notion (more about which is discussed in section 1.3), is of utmost significance for BT, and that, accordingly, the data fall into two broad natural classes - binding versus non-c-command anaphora. We will continue to reflect upon the validity of these hypotheses in the course of this book. On the other hand, if Binding Conditions are indeed based on the notion of c-command, they can serve as a probe into the phrase structure of a sentence: if an NP blocks the occurrence of a coindexed pronoun or full NP', NP must c-command NP'. Binding Condition C in particular will be useful in this regard, since it applies across clause boundaries. It has been suggested, for example, that the pairs in (1.31) and (1.32) show that object clauses, but not temporal adverbial clauses, are c-commanded by the object, while both are c-commanded by the subject: (1.31) (a) * The dog told him i [that the horse i would fall], (b) The dog hit himi [while the horsei ate lunch]. (1.32) (a) *Sheg'll talk to me [when Sheilas gets back from lunch], (b) I'll talk to Sheilas [when sheg gets back from lunch]. While this method can be useful, it should be applied with care, for at least two reasons: first, as pointed out in section 1.2.3 above, the unacceptability of bound full NPs is itself not uncontroversial, and judgments seem to vary between speakers, but also in response to prosodic, stylistic, and discourse-pragmatic factors (see e.g. Carden and Dieterich [1981]; and Gerken and Bever [1986] for experimental results). Second, subordinated clauses are often found in displaced positions (e.g. through topicalization or extraposition), or at least could be for all we know, so that our conclusions from such examples rely in turn on our conclusions about the interaction of Binding Conditions with displacement (more on which in chapter 12). We will suggest that the phenomenon of semantic binding, to be introduced in chapter 4, may provide a more reliable diagnostic for c-command. Since we are presently concerned with demarcating the conditions on binding themselves (rather than presupposing them to figure out constituency), we will for the most part ignore constructions whose constituent structure is itself subject to debate. 1.2.5 Taking stock i It will be useful to separate several parts or components of the theory, as these will be subject to criticism, revision, or modification later, independent of each other: 1.2 Binding 11 • The classification of NPs according to their coreference and binding options. Here: three classes, reflexives and reciprocals ('anaphors'), non-reflexive pronouns ('pronominals'), non-pronominal or full NPs ('r-expressions'). • The identification of one or more domain(s) within which binding requirements apply. Here: the minimal clause. • The formulation of a proper notion of command or accessibility as prerequisite for, and source of, asymmetry in binding. Here: precedence and c-command. The general format of a Binding Condition can then be schematized as in (1.33): (1.33) An NP of class must (not) be coindexed with a commanding NP within its domain. Developing a general theory of binding is to formulate conditions of the general form in (1.33) for several languages, and in the process, to determine which exact values for the variables class, command, and domain are empirically most adequate, within a language, and cross-linguistically, and which are systematically irrelevant. In this book, we will be concerned with the notion of domain in chapter 3, the issue of classification in chapters 3 and 11, and the concept of command in section 1.3 of this chapter, as well as in chapter 12. Hopefully, the tripartition of the ingredients to the BT will prove useful in keeping track of the discussion.4 We have set up our system in such a way that each class of NPs may have (a) positive and/or negative Binding Condition(s) associated with it. Complementary distribution between two classes results if the negative binding domain for one element happens to be the same as the positive binding domain for the other. There is nothing in the formal system that accounts for the intuition one might have that one class of NPs is used because of the unavailability of the other, and vice versa. Intuitions of this kind have motivated so-called blocking approaches to Binding Theory, in which one form is used if (and perhaps only if) the other is excluded; examples of such approaches include Dowty (1980); Farmer and Harnish (1987); Huang (2000); and Levinson (1987, 1991, 2000); in light of the fact that non-complementary distribution between different pronoun classes is common across languages, including English (see chapters 3 and 11), we will not review these approaches further in this book; but see Burzio (1996, 1998); Kiparsky (2002) for refined, hybrid blocking approaches. Exercise 1.2 A naive approach to binding would be that the use of a pronoun is necessary to avoid repetition of full NPs, especially within a single sentence. 4 Most authors appear to assume, implicitly, that the notion of command, once defined properly, is invariant across all Binding Conditions across all languages (but see the remarks at the end of section 1.3); if so, the variable command shouldn't be treated on a par with the others; we will leave this issue open in this book. 12 the abc of binding theory Use the distinction between antecedence and binding established above to make an argument against this view. Construct and provide crucial examples. Exercise 1.3 Both sentences in (1.34) below could be used to illustrate that reflexives in English cannot be bound across a clause boundary. Explain how each does it. Can you think of reasons to prefer one mode of presentation over the other? (1.34) (a) Johni thinks that BUI2 likes himself2/*i. (b) * John thinks that I like himself. Exercise 1.4 Consider the English possessive form his/her I its own (pretend that it is a single form). What requirements, if any, does it impose on its binder or antecedent? Formulate an appropriate Binding Condition for it, and adduce (acceptable and unacceptable) examples to support your proposal. 1.3 Command and precedence © As discussed in 1.2.5 above, the notion of command is one key ingredient in the formulation of Binding Conditions. We have used c-command plus precedence in our definition of binding above, hence as our relevant notion of command in (1.33). In this section we will critically reexamine these two notions, suggesting that they should perhaps be replaced with an altogether different command notion. 1.3.1 Against precedence In our final definition of binding in (1.19) we have used two relations that give us an asymmetrical ordering among the NPs in a sentence, c-command and precedence. Is this necessary, or even tolerable? Consider the abstract phrase markers in (1.35): (1.35) (a) (b) In (1.35a), the standard kind of case we have been looking at in English, A could bind B because A c-commands and precedes B; and B could not bind A, because it neither c-commands nor precedes A. In other words, according to our present definition of binding, (1.27), there are two different reasons why A can bind B, but not the other way around; the account is thus somewhat redundant, though perhaps harmlessly so. 1.3 Command and precedence© 13 In (1.35b), on the other hand, A c-commands B asymmetrically, but B precedes A. If both factors are relevant, binding between A and B should be ineffable, i.e. impossible either way around, and regardless of the morphological shape of either NP. While English lacks clear instantiations of this configuration, other languages arguably provide them. Reinhart (1983a: 47) provides the following data from Malagasy (attributed to E. Keenan, p.c), an Austronesian language spoken in Madagascar (for convenience, I have set the pronouns in boldface and the antecedents in italics): (1.36) (a) namono azy ny anadahin- d- Rakoto (Malagasy) hitlkilled him the sister- of- Rakoto 'Rakoto's sister killed him.' (b) * namono ny anadahin- d- Rakoto izy hitl killed the sister- of- Rakoto he 'He killed Rakoto's sister.' Malagasy is a VOS language, but, as the translations make clear, behaves rather like English with respect to the BT. In particular, (1.36b) appears to be a Binding Condition C violation, with the subject pronoun izy, 'he', illicitly binding the full NP Rakoto within the object NP. No such effect is found in (1.36a), in which the full NP is the subject, and the pronoun is in the object. The data then suggest that the subject can bind the object, but not vice versa. Given that the object precedes the subject, this pattern straightforwardly prohibits a treatment in terms of precedence. C-command alone, however, would seem to provide the correct asymmetry, provided we assume that the basic clause structure of Malagasy is essentially as in (1.37) (cf. [1.35b] above): s (1.37) V OBJ We thus conclude that the inclusion of precedence in the definition of binding, while perhaps merely redundant for English, is actually harmful if we want to apply the notion cross-linguistically, and should be dropped.5 1.3.2 Limitations of C-command i Unfortunately, c-command is not unproblematic either. Languages that display flexible constituent ordering abound with examples in which a bound element precedes and c-commands its binder. Japanese, Korean, and German provide three examples. They are strictly or mostly left-branching languages with 5 I found exactly one case in the literature in which precedence does seem to play a role, namely Samoan, as discussed in Chapin (1970); and Keenan and Stabler (1995). Whether this requires inclusion of precedence in the BT for Samoan, or can be captured in any other way, will be left as an open question here. 14 the abc of binding theory free word order, in all of which we find, for example, (reflexive-like) pronouns preceding coreferring full NPs:6 (1.38) Zibun-o Hanako -ga utagatte iru (Japanese) self -acc Hanako -nom doubts 'Hanako doubts herself.' (1.39) Caki casin-eke Kim-un silmanghaossta (Korean) self -dat Kim -top disappointed 'Kim was disappointed in himself.' (1.40) Oft hat sich der Mann im Spiegel betrachtet. (German) often has self the man in the mirror watched 'Often the man watched himself in the mirror.' These data are problematic for our definition of binding, because, unlike in the case of Malagasy discussed above, there is agreement that none of Japanese, Korean, or German has a constituent structure in which the reflexive objects in these examples are c-commanded by their subject antecedents. It should be noted, however, that in all these examples, what could be called the 'logical subject' binds the 'logical object', regardless of order or c-command. In other words, one could conjecture that some more abstract level of representation displays an asymmetric ordering among constituents, at which all of the examples discussed in this section show the relative ordering in (1.41): (1.41) 'subject' >- 'indirect object' >- 'direct object' >- 'prepositional object' I have scare-quoted the notions 'subject', 'object', etc. because these are merely place-holders for whatever theoretical constructs a particular theory regards as relevant here. We will inspect several instantiations of such a hierarchy in turn. Before that, I want to bring up another set of data on which the c-command condition, even if it doesn't fail as spectacularly as on those in (1.38-1.40), seems deficient, namely so-called non-configurational or 'flat' structures. It is, for example, generally accepted that the objects in double object constructions in English show asymmetries in binding behavior:7 (1.42) (a) I showed John himself (in the mirror), (b) *I showed himself John (in the mirror). (1.43) (a) I showed Mary to herself, (b) * I showed herself to Mary. Suppose that the structure of the VP in double object constructions is as in (1.44): (1.44) 6 Korean data from Keenan (1988):131. 7 Reflexive examples from Barrs and Lasnik (1986):347 and Larson (1988):338. 1.3 Command and precedence© 15 Given this structure, no asymmetries in binding are expected, contrary to fact, since the two objects mutually c-command each other, at least in the double NP case. Similar remarks apply, for example, to nouns with more than one post-nominal attribute:8 (1.45) (Ja čital) staťju Tolstojao sebe. (Russian) / read article-ACC T.-gen about himself-hOC '(I read) an article of Tolstoj about himself.' Again, the immediately post-nominal phrase can bind the second one, but not vice versa. But, as in the case of verbal double-object constructions, the constituent structure appears to be flat.9 Both these examples would be treated correctly if we assumed c-command and precedence as prerequisites for binding, but, in the light of the problems with precedence encountered earlier, it is worth while to look for alternative treatments. A hierarchy along the lines of (1.41) above can provide such a treatment, as it provides a total ordering even among NPs that mutually c-command each other. To be sure, there have been attempts to reconcile the data in (1.42-1.43) with a c-command approach to binding, notably Larson (1988), which argues that the asymmetries rather show that these structures are more complex than (1.44) and have the essential properties of those in (1.46): (1.46) We do not have to elaborate on the details of this analysis (like the identity of the unlabelled nodes in [1.46]). The important thing is that a structure like (1.46) predicts the binding asymmetries as a function of asymmetrical c-command, without reference to anything else. On the other hand, we saw the need to reconsider the c-command condition anyway in the light of the Japanese/Korean/German cases, and the two alternatives to c-command presented below are intended to capture those and 'flat structure' cases by the same mechanism, namely an independent, non-phrase-structural hierarchy. This is obviously not the place to rule a final verdict on the (non-)existence of flat structures, but it bears mentioning that the binding facts alone can be treated by a refined command notion, which is presumably required independently, and that structures like in (1.46) are not necessitated by them. We will look at two such notions in what follows. 8 Data based on Rappaport (1986):106. 9 Note that the simplest binary branching structure would actually be // VIN XP] ] XP2 ], which would yield the opposite asymmetries in c-command from what we find with the binding data. 16 THE ABC OF BINDING THEORY 1.3.3 ©-command i Let us start with the perhaps most attractive version of a non-phrase-structural ordering among the NPs, namely the idea that this ordering is essentially semantic in nature, more precisely that it is based on the semantics of the thematic relations involved. An early and representative example of this approach is Jackendoff (1972): 148, where it is proposed that the relevant asymmetric ordering among elements is derived from the hierarchy of thematic roles (or ©-roles for short) in (1.47): (1.47) Agent >- Location, Source, Goal >- Theme The fact that der Mann can bind sich in the German example (1.40) above is then due to the fact that der Mann bears the ©-role 'Agent,' which outranks the role of sich, 'Theme' on the ©-hierarchy in (1.47). Likewise, Rakoto must not be coreferent with izy in the ungrammatical Malagasy example in (1.36), due to the fact that izy has the Agent' role in that sentence, which is higher than any other role in the ©-hierarchy. The initial appearance that binding asymmetries correlate with a hierarchy among phrase-structure positions (as established by c-command) is merely an epiphenomenon of the dependence of phrase structure upon this same thematic hierarchy. There are two more immediate advantages of ©-command over c-command. First, as it is defined over thematic roles, it captures certain parallelisms between NPs and PPs. Not only can binding proceed 'into' PPs - a fact captured by c-command and ©-command alike - sometimes it can also proceed 'out of PPs, if the pertinent thematic command obtains: (1.48) (a) We talked to John about himself. (b) * We talked to himself about John. (c) * We talked about John to himself. These examples (from Wilkins [1988b]:208) show that Goal (as realized by a complement of to) ©-commands Theme (as realized by a complement of about) and can therefore bind it, regardless of the fact that John doesn't c-command himself, because the first node dominating it is the PP. Second, ©-command allows a straightforward account of binding with so-called psych-verbs. With these verbs, a phrase-structurally lower argument appears to bind into a higher argument:10 (1.49) (a) Stories about herself generally please Mary, (b) Each other's health worried the students. While obviously unexpected from a phrase-structure point of view, these examples can be captured under the assumption that the role borne by the subject of 10 Examples from Pesetsky (1987):127. 1.3 Command and precedence© 17 these verbs, Theme, is lower on the ©-hierarchy than that of the object, Experi-encer.11 Attractive though the ©-command notion is for the purpose of BT, it also faces some serious problems. First, verbs that allow alternative realizations of their arguments, such as English double-object constructions, show alternations in binding possibilities as well, as discussed at length above, cf. the examples in (1.42) and (1.43) above; (1.50) reiterates the point:12 (1.50) (a) I sold the slave himself. (Goal binds Theme) A ©-command approach has to grab the bull by the horns and claim that, say, John bears a different ©-role in / sold the cabbage to John than in / sold John the cabbage, as is done in Wilkins (1988b):208ff., who claims that a Theme, when verb-adjacent, as in (1.50b), is also a Patient, which is a higher role than Goal, whereas it is only a Theme if it follows another object, as in (1.50a) and (1.50c), which makes it lower than Goal (a similar strategy is implemented in Jackendoff [1990]'s action tier). This strategy, while technically feasible, highlights a more general weak spot of ©-based approaches, namely the question of how to determine the role of a given argument in a given structure. For example, according to its proponents, the Theme/Patient distinction explains what is behind the dative alternation in English. According to the critics, it merely gives a semantically loaded name to what by all appearances is a purely syntactic distinction, namely that between direct and oblique object (alleged semantic distinctions between the two are indeed subtle, but see Krifka [2004]). According to critics, then, a sufficiently elaborated ©-command account is just a somewhat misleadingly labelled instantiation of an argument-structure based account, to be discussed in the next subsection. Before closing this section, let us briefly consider two other areas in which the ©-command idea has been argued to encounter difficulties. In certain cases of raising, reflexivization involves NPs that do not receive their thematic role from the same lexical element. Consider (1.51):13 (1.51) Max strikes himself as qualified for the job. It would appear that the ©-role of himself (Goal or Experiencer of strike) is higher than that of Max (Theme of qualified) as well as that of the clause Max has raised from (which is presumably the Theme of strike). Yet, binding can only proceed from the subject to the object position, in violation of the ©-command condition. 11 The facts are more complicated, as must therefore be the account; see, among many others, Belletti and Rizzi (1988). 12 From Pollard and Sag (1992:298); in contradistinction to Pollard and Sag, Larson (1988), and others, Jackendoff (1972:157) and Postal (1971:ch. 15) judge both variants of binding in double object constructions as ungrammatical. 13 cf. Pollard and Sag (1992):299. (b) I sold the slave to himself. (c) * I sold himself the slave. (Theme binds Goal) (Goal cannot bind Theme) 18 THE ABC OF BINDING THEORY The second argument asserts that morphological processes that change grammatical functions also change binding options. Similar to the case of verbs with alternating argument structures, no obvious change in thematic roles goes along with it:14 (1.52) (a) John shaved himself. (Agent binds Theme) (b) John was shaved by himself. (Theme binds Agent) (c) * We expect himself to be shaved by John (Agent cannot bind Theme) This argument has to be taken with a grain of salt, though. Many speakers find examples like (1.52b) rather marginal. Cross-linguistically, the diagnosis is actually much less clear, perhaps non-uniform. Toba Batak (Malayo-Polynesian; Sumatra) is a verb-initial language with no morphological case distinctions.15 Depending on the choice of verbal prefix, mang versus di, a sentence of the form V NP] NP2 is interpreted with NPi as Theme (or Patient) and NP2 as Agent (mang-prefix), or with NPi as Agent and NP2 as Theme (^/-prefix). In other words, mang- and di- determine different mappings between thematic roles and phrase-structure positions and are, as such, comparable to grammatical function-changing operations like passive in other languages. Strikingly, however, the binding always goes from Agent to Theme, regardless of prefix choice and, accordingly, linear order:16 (1.53) (a) Mang -ida si Ria si Torus MANG see the R. the T. 'Torus saw Ria.' (b) * Mang -ida si Torus dirina MANG see the Torus self 'Self saw Torus.' (c) Mang -ida dirina si Torus MANG see self the Torus 'Torus saw self.' (a') Di -ida si Torus si Ria Di see the T the R. 'Torus saw Ria.' (b') Di -ida si Torus dirina Di see the Torus self 'Torus saw self.' (c') * Di -ida dirina si Torus Di see self the Torus 'Self saw Torus.' It appears that neither the change in linear order (and the change in c-command relations presumably accompanying it) nor the change in grammatical function indicated by the verbal prefixes affects the binding options between the two NPs. A good candidate for what remains invariant in all structures are the ©-relations: Torus is the Agent and Ria is the Theme. We conclude that grammatical function-changing operations do not provide clear evidence for or against a ©-command approach to Binding Theory. Examples from Pollard and Sag (1992):298. See Keenan (1988) and Schachter (1974); Schachter argues that the basic phrase structure is [[V NP] NP], which would make it a VOS language, but nothing hinges on that, cf. also Sternefeld (1992). Data from Keenan (1988): 129f. and Keenan and Stabler (1995):4; very similar data are reported for Balinese in Wechsler and Arka (1998):406ff. 1.3 Command and precedence© 19 1.3.4 Obliqueness-command i The last command relation to be discussed in this chapter is that of obliqueness command, or o-command for short (sometimes also called a(rgument)-command for reasons to become clear shortly). O-command occupies a middle ground in between c-command and ©-command in that it claims that there is a non-thematic, syntactic ordering between the NPs in a sentence, but denies that that ordering is directly encoded in the phrase structure as c-command (or any other phrase-structure relation for that matter). I illustrate the idea using an argument-list notation;17 a verb like watch, for example, will have an a(rgument)-list like in (1.54): (1.54) a-list of watch: ( NP, NPacc) The order of elements on the a-list signals obliqueness. Thus the first NP on the a-list in (1.54) is less oblique than the second, NPflCC. Obliqueness does not necessarily correspond to linear order; if, for example, a language allows different linearizations of arguments, the obliqueness relations encoded on the a-list still remain the same. Thus the a-list associated with betrachten, 'watch,' in the German example (1.40) above is the same as in (1.54), despite the fact that the accusative NP, the reflexive sich, linearly precedes the subject NP (and presumably also c-commands it). Similarly for the Japanese and Korean examples discussed. Likewise, a flat constituent structure like in (1.44) is perfectly compatible with a total ordering in terms of obliqueness (in fact, by its very nature an a-list is always totally ordered), as, for example, in (1.55): (1.55) a-lists for ditransitive verbs like give: (a) Dt such that for all x in De, f(x) = 1 iff x smiled According to this, f (^CarolineJ8) equals 1 if Caroline smiled, 0 otherwise (since there are only two truth values, we can abbreviate this as '... 1 iff Caroline smiled', where 'iff stands for 'if and only if); and so on for every other individual. We can now introduce our most important semantic composition rule for non-terminal nodes, function application (Vn stands for any verb or projection thereof): (2.21) Function Application (FA): [V NPp = [NP V"F = [V"p(|[NPp) According to this rule, whenever we encounter an NP and a verb or verbal projection, which denotes a function, we apply that function to the denotation of the NP. Using lexical meanings and function application, we thus get: (2.22) (a) [[Caroline]g = Caroline (in flesh and blood) (b) [[smiled] # = the function / as defined in (2.20) above (c) [[Caroline smiled]* = [smiled]*([Caroline]*) (FA) =f (Caroline) = 1 iff Caroline smiled (0 otherwise) Functions, as in (2.20), are rather cumbersome to write, so it will be useful to introduce some abbreviatory conventions. To name a function like / in (2.20), we will alternatively write (2.23a), or (2.23b): (2.23) (a) Xx e De.x smiled (b) Xxe.x smiled (2.23a) and (2.23b) use the k(lambda)-prefix, borrowed again from formal logic. We employ the following convention (cf. Heim and Kratzer [1998] :37): (2.24) A-Convention: Read Xa.[(p ... a ... (a)...] as (a) 'that function which maps every a to 1 iff [^ ... a ... (a)...]', if 0 is a sentence, (b) 'that function which maps every a to [^ ... a ... (a)...]' otherwise 2.3 Compositional interpretation 33 We will return to clause (2.24b) below. Since lx smiled' in (2.23) is a sentence, clause (2.24a) applies: [kxe.x smiled] is that function which maps every x in De to 1 iff x smiled (brackets added for perspicuity). Note that the type of a A.-expression can be read off it straightforwardly: (2.23a)/(2.23b) for example has a k binding a variable of type (e), followed by a sentence, 'x smiled'. English sentences represent meanings in Dt, which means that the whole expression is of type (e,t) .5 Transitive verbs A typical transitive verb denotation is given in (2.25): (2.25) [likes]* = ke De.Xy e De.y likes x Here the sentence y likes x has two As in front of it, each binding an individual variable. Its semantic type is thus (e,(e,t)), or for short: (e,et). To disentangle this, we first apply (2.24b) with a = x and 0 = [ky <= De.y likes x], whereby (2.25) is a function which maps every individual x onto [A._y <= De.y likes x]. The latter in turn, according to (2.24a), is a function that maps an individual y to 1 iff y likes x. (2.25) as a whole, then, is a function from individuals to functions from individuals to truth values. Such a function subsequently takes two individuals and maps them onto a truth value: (2.26) (a) [Chris likes Caroline]-? = (b) [likes Caroline]*([Chris]*) = (c) [likes Caroline]*(Chris) = (d) [likes]*( [Caroline]*) (Chris) = (e) [likes] *(Caroline) (Chris) = (f) [Xx e De Xy e De.y likes x] (Caroline) (Chris) = 1 iff Chris likes Caroline Functions and the sets they characterize At the beginning of this section we said that an intransitive verb denotes a function from individuals to truth values (type (et)). In section 2.1, however, it was said that an intransitive verb denotes a set of individuals. While only the former is literally true, the latter almost is, because any function of type (r,t) uniquely characterizes a set (namely a subset of Dr), and vice versa. The function [kxe.x smiled] - call it fsmned - for example, characterizes the set of all individuals x which that function maps to 1, formally: {x e De \ fsmued(x) = 1}; informally: the set of all those who smiled. In the remainder of this book, I will often make use of this equivalence, which is rendered again in (2.27): 5 Note that individuals and functions are the meanings of object language expressions. We use names, English sentences, and A-expressions to name those denotations in our meta-language; accordingly, no italics are used here. I will, however, sometimes include expressions used to name a denotation in quotation marks to facilitate reading. 34 INTERPRETING INDEXED STRUCTURES (2.27) Set/function equivalence and convention: '[\x e DT... .x .. -](yY is equivalent to 'y is in {x e Dx will often write the latter instead of the former .. .x ...}'; we For example, l[Xx e De.x smiled](Caroline) = 1' is equivalent to 'Caroline is in {x e De | x smiled}'. The latter form is sometimes more convenient and more transparent, which is why I will use it. Also, I find it more intuitive to think of an intransitive verb denotation as a set of individuals, and of a sentence containing it as true if the denotation of the subject is in that set, than to think of an intransitive verb denotation as a function from individuals to truth values, and of a sentence containing it as true if that function applied to the subject denotation is 1 - though, as stated, both are equivalent. So, to reiterate, I will regularly describe a denotation by saying 'so-and-so is in set-x' where our official semantics assign it the denotation 'function-x applied to so-and-so =1'. The reason why we still officially use 'function talk' rather than 'set talk' is that the former is more general. Recall that, for example, a sentence with a transitive verb can be interpreted by repeated function application, as in (2.26), because a transitive verb denotes a function whose values are again functions. While, alternatively, we could think of a transitive verb as denoting pairs of individuals - those pairs (x,y) in which x loves, hates, knows, calls, etc. y - such a meaning would require an extra rule for combining transitive verb meanings with object NP meanings. Compositionality A theory of interpretation should be compositional, meaning: denotations of complex expressions should be composed of the denotations of the simpler expressions that make them up. For concreteness, let us adopt the following two conditions on interpretations: (2.28) (a) Every syntactic constituent C has an interpretation. (b) In interpreting a constituent C, no other information can be used for computing lCJg than that associated with C's daughters. The Function Application rule used so far is strictly compositional in this sense. Consider, for illustration, a different rule: (2.29) Non-Compositional S-rule: NPS VP V NP0 = [[Vp([[NP0p)([[NPsP) This rule will assign the same denotations to a transitive sentence as the repeated application of FA. It is, however, non-compositional. For one thing, VP doesn't 2.3 Compositional interpretation 35 Table 2.1 Types and their names Type Name Variable Categories e individual x,y,z names, pronouns t truth value p sentences et properties P VPs, Ns, APs e,et relations R transitive Vs and Ns et,t generalized quantifier q quantified NPs receive an interpretation, contra (2.28a). And, for another, the meaning rule for S directly uses the meanings of V and NP<9, which are not daughters of S, in violation of (2.28b). We will continuously check that all our interpretation rules to follow are compositional in nature. A first inventory of categories and types Different semantic types have common names in the semantic literature, which are summarized in table 2.1 (not all of them will look familiar to you yet). In table 2.11 also indicate the typical variables I use for the different semantic types. This is, of course, just a notational convention, but it will allow us to abbreviate the function notation further by leaving out the explicit domain specification. Instead of (2.25) we can now write (2.30), because it is understood that x and y are variables over De (upon demand, variables will be numbered, e.g. P\,P2, ■ ■ ■)■ (2.30) [likes]* = XxXy.y likes x Exercise 2.7 Why must the denotation for, say, likes be as in (2.25), rather than [Ay e De.kx e De.y likes x]l What is the difference? Argue! Exercise 2.8 Give the denotation for the verb danced both in set notation and A.-notation. Exercise 2.9 Give the denotation for the verb introduced. Exercise 2.10 Give denotations for all lexical elements in the following sentences, and provide derivations of their truth conditions a la (2.26); add indices where necessary: 36 INTERPRETING INDEXED STRUCTURES (2.31) (a) Sam danced. (b) Kim likes Jo. (c) She slept. (d) He admires him. (e) He admires himself. 2.3.2 Truth, truth conditions, and meanings i The denotation of a sentence is a truth value, 1 or 0. What is closer to the intuitive meaning of a sentence, however, is not its truth value, but the truth conditions the interpretation procedure assigns to it, for example a statement of the form '[[S] = 1 iff...'. The goal of semantic theory, after all, cannot be to determine if a sentence is true or false (unless it happens to be a tautologous or contradictory sentence), but to assign systematically to every sentence a description (in English or something resembling English) of the kind of situation in which it would be true. A slightly different way to think of the general meaning of a sentence involves the situations a sentence correctly describes. Generally, the following holds: (2.32) Given truth conditions of the form '[[S] = 1 iff 0,' we can say that S is true in a situation s iff 0 holds in s. Thus from the truth conditions '[[Caroline smiled]]5 = 1 iff Caroline smiled' we can go to 'Caroline smiled is true in a situation s if Caroline smiled in s.' The set of situations in which a sentence S is true is often called the proposition expressed by S. Instead of saying that a sentence denotes 1 iff 0, we will sometimes just say that it denotes the proposition that 0; thus Caroline smiled denotes the proposition that Caroline smiled. But sentences contain NPs, and NPs depend on the assignment function, so we have to refine this. Recall that the denotation of an NP„ is defined relative to an assignment g, if g(n) meets the presuppositions of the denotation of NP (e.g. that it be a female individual, or that it be called John, etc.); otherwise [NP„]]g is undefined. We can thus say the following: (2.33) Given truth conditions of the form '[[S]^ = 1 iff 0,' we can say that S is true in a situation s relative to assignment g if |[S]g is defined and 0 holds in s (it is false if |[S]g is defined and 0 doesn't hold in s). Whether or not [[S]]g is defined for a given S thus depends on what assignment g we choose. The proposition expressed by S relative to assignment g is the set of situations in which S is true relative to g. A given sentence can thus express different propositions relative to different assignments. This is plausible if, again, we think of an assignment as (part of) a context. Take as a concrete example the sentence She^ smiled. We derive the truth conditions [she6 smiled]]8 = 1 iff (g(6) is a female individual and) g(6) smiled. By (2.33), it then holds that: for any assignment g, [she6 smiled]]5 is defined iff g(6) is a single female individual; if defined, it denotes the proposition that 2.3 Compositional interpretation 37 g(6) smiled. Take a g such that g(6) = Caroline; then that means that 'Caroline smiled' is a possible meaning of she§ smiled. But given that there are other assignments g' such that g'(6) = Susan, 'Susan smiled' is also a possible meaning of shee smiled, and so on and so forth, for all and only those assignments according to which g(6) is single and female. Obviously, we haven't said anything about how g relates to the actual linguistic context, in other words: about the question which individual a given pronoun will, or is likely to, denote in a given linguistic context (see subsection 2.4.1 for a few remarks). Yet, by virtue of universally quantifying over all technically possible g, we can define the context-independent notion of a possible meaning: (2.34) 0 is a possible meaning of a sentence S iff there is an assignment function g such that [SF = 1 iff 0 For example, 'Caroline smiled' is a possible meaning of She^ smiled, as is 'Susan smiled'. 'Bill smiled' or 'Lake Balaton smiled' are not, assuming that neither Bill nor the popular Hungarian lake are female individuals. More interestingly, given PACO in (2.15), we derive for example that 'Caroline likes Susan' is a possible meaning of Caroline^ likes her3, but 'Caroline likes Caroline' is not. Note that the possible meanings for a sentence are independent of the actual choice of indices. Two sentences SI and S2 that only differ in indexing have the same set of possible meanings, provided all and only those NPs in S2 are coindexed that are coindexed in SI. For example, the set of possible meanings of Shen smiled is the same for any choice of index n.6 As a consequence, we can go one step further and define the notion possible meaning of an expression E, where an expression is a sentence (or rather: tree) without indices: if 0 is a possible meaning of E, that means that there is some indexing E1 of E which is grammatical (i.e. obeys the Binding Conditions A-C and whatever other syntactic conditions we'll impose on indexing), and 0 is a possible meaning of that indexed version E1 of E. For example, let E be the expression She likes her, which has an infinite number of possible indexings E1. By Binding Condition B, all of these in-dexings must assign different integers to she and her. By the PACO, such an indexing prohibits an interpretation by which she and her wind up coreferen-tial. Thus the set of possible meanings for She saw her (without indices) is the set of all propositions 'x saw _y' where x and y are two different female individuals. The expression She says Kim likes her, on the other hand, has infinitely many indexings on which she and her are coindexed, and infinitely many on which they are not (and all of those have Kim counter-indexed with both pronouns). The former all get interpretations of the form x says Kim likes x, the latter of the 6 This is not to say that there couldn't be syntactic rules or constraints that distinguish SI and S2, as argued recently in e.g. Heim (1997), or Sauerland (2000); we will discuss some relevant cases in sections 5.5.2 of chapter 5, and 6.5.1 of chapter 6. 38 INTERPRETING INDEXED STRUCTURES Table 2.2 Semantic objects related to sentences truth value truth conditions proposition expressed by S (relative to g) possible meanings of indexed S Oor 1 ISP = 1 iff # 4>l the set of situations s in which (p holds (assuming that [SP = 1 iff (p) the set of propositions S can express relative to any g (respecting PACO) the set of propositions any grammatical indexing of S can express relative to any g (respecting PACO) possible meaning of unindexed S form x says Kim likes y, with x / y / Kim. Hence all instances of these are possible meanings for the expression She says Kim likes her. In conclusion, the various semantic objects corresponding roughly to whole sentences introduced in this subsection are summarized in table 2.2. Since natural language doesn't come with indices, possible meanings of expres-sions/unindexed sentences are what our semantic interpretation procedure, in tandem with Binding Theory, should ultimately predict. Exercise 2.11 Give truth conditions, the proposition expressed, and, where you can, the possible meanings (regardless of indexing) for the sentences in (2.31) (again, add indices where necessary). Appendix: Meaning relations Given the above, we define, for the sake of completeness, some essential semantic notions for indexed sentences: (2.35) (a) SI entails S2 iff for any assignment g, every situation in which SI is true relative to g is one in which S2 is true relative to g. (b) S is a tautology iff it is true in any situation relative to any assignment (for which it is defined). (c) S1 and S2 are incompatible if for any assignment g, there is no situation in which S1 and S2 are both true relative to g. (d) S1 and S2 are synonymous if each entails the other. Type-driven interpretation In (2.26) above we used Function Application twice, once to combine the object meaning with the transitive-verb meaning, and then again to combine that VP meaning with the subject meaning. Note that in each case, FA is well defined, since the type of the argument, (e), 'matches' the type of the function, 2.3.3 More composition rules © 2.3 Compositional interpretation 39 (e,...). This is so because any element of type (e,...) denotes a function from individuals, i.e. elements of De, to something else (depending on what ...is). In a case like that we say that the argument, here an individual in De, is in the domain of the function.7 Among two given meanings M\ and M2, exactly one of three possible situations will obtain: • M\ is in the domain of M2, i.e. M2(M\) is defined (e.g. M\ is in De and M2 is in Det or De^et) • M2 is in the domain of M\, i.e. Mi (M2) is defined (e.g. M2 is in De and Mi is in Det or De^et) • neither is in the domain of the other, i.e. neither Mi (M2) nor M2 (Mi) are defined (e.g. both Mi and M2 are in Det) Given this, we can generalize our interpretation rule as follows: (2.36) Function Application, revised and final: [A Bp = [A]S([B]*) or [B]S([A]*), whichever is defined The reader may verify that this rule subsumes our earlier (2.21). An interpretation procedure along these lines, where the semantic type of the sister constituents (rather than their syntactic category and/or linear order) determines the mode of semantic composition, is called type-driven interpretation or, less respectfully, 'shake-and-bake semantics'; it was introduced into linguistic semantics in Klein and Sag (1985). We will add a few other composition rules that take care of situations in which neither Mi (M2) nor M2(M\) in the sense above are defined later on, but FA alone will bring us a long way. For concreteness, we will assume that a constituent to which no semantic composition rule applies is uninterpretable and hence un-grammatical. Interpreting flat structures In chapter 1 we encountered examples which appear to involve flat structures, i.e. nodes with more than two daughters, e.g. double object VPs, NPs with two post-nominal PPs, and generally clausal structures in non-configurational languages: (2.37) (a) I introduced the students to each other. (b) (Ja cital) stat'ju Tolstoja o sebe. (Russian) / read article-ACC T.-GEN about himself-hOC '(I read) an article of Tolstoj about himself.' 7 Confusion is lurking here: note that the denotation domain of an expression E is not identical to the domain of the denotation of E, even if E denotes a function. For example, the denotation domain of an intransitive verb is Det, the set of functions from individuals to truth values, but the domain of the denotation of an intransitive verb is De, the set of individuals. 40 INTERPRETING INDEXED STRUCTURES We suggested there that binding asymmetries among the daughters in such a multiply branching structure are encoded in a non-phrase structural aspect of syntactic representation, say the a(rgument)-list on which a command relation such as 0(bliqueness)-Command can be defined. But, independent of the proper formulation of the Binding Conditions, we also need to say something about the interpretation of these structures, as our Function Application rule (2.36) (or its earlier incarnation [2.21]) obviously doesn't apply to such a structure. This can be done by adding a rule like the following: (2.38) iV-ary Function Application: T x" 118 [X°F([A„F)([AM_iF)... «Aip) where A\, A2, ■ ■ ■, An-\, An is the order of A, A, A ...on X°'s argument-list For example, introduce Helena to Gwynn will be interpreted as [introduce]]8 ([(to) Gwynn]]5) ([Helena]]5), provided that the a-list of introduce looks like (NP, NPacc, PP).8 Note that this rule is strictly compositional, since it assigns a meaning to every node, based on the meanings of its daughters. 2.4 Extensions and alternatives© 2.4.1 A note on non-C-command anaphora 1 Above, we somewhat boldly called g (part of the representation of) the 'context'. However, nothing in our theory tells us how g relates to an actual context. Consider, for example, the following: (2.39) (a) A dog is sitting at the bar. (b) Suddenly, it sees a man come in. (c) The man has a dog too. (d) It is wearing a tie. Here's a selection of what our treatment so far tells us: (i) the bar cannot be bound by a dog in (a); (ii) it in (b) can refer to the dog from (a), or some other neuter individual; (iii) the man in (c) refers to the unique man, which can be the man from (b); etc. etc. Here's what our treatment doesn't tell us, though native speakers could: Notice that the FA notation is left-associative, i.e. X(A)(A)(A") is equivalent to ((X(A))(A))(A"). In other words, semantic composition still proceeds 'binary,' combining with one argument 'at a time.' We will justify this approach (rather than, say, one which combines the verb meaning with a tuple consisting of all its argument meanings at once) in chapter 4 below. Note, too, that (2.38) is built on the assumption that all but the least oblique argument Ao on the a-list are sisters to X°, though this could easily be refined. 2.4 Extensions and alternatives© 41 (2.40) (a) it in (b) is very likely to be taken to refer to the dog mentioned in (a), rather than to any other individual (b) it in (b) cannot possibly refer to the other dog (the one with the tie) (c) the man in (c) is very likely to be interpreted as the man mentioned in (b) (d) a dog in (c) cannot be the dog mentioned in (a) and (b) (e) it in (d) is very likely to be taken to refer to the second dog In short, what is missing is a theory about how a given sentence changes the context, including the referential options for sentences to follow. How do referents become available? How do speakers choose among different grammatically possible antecedents for a given pronoun? Informally, such a theory would start from the assumption that at the beginning of a discourse there are no anaphoric possibilities; technically, a discourse starts with an assignment function with an empty domain. NP occurrences can then be classified into those that introduce discourse referents, hence add their index to the domain of the assignment function, and those that don't, hence rely on their index being in the domain of the assignment function already; the latter are the anaphoric NPs. Whether or not an NP belongs to the introducing or the anaphoric type is not merely a matter of its morphological shape. Proper names, for example, can be used anaphorically, or to introduce a new discourse referent; so can definite NPs (the moon versus the guy), demonstratives, and first and second person pronouns (though the latter could be argued to be 'anaphoric' by definition). Even third person pronouns, though most commonly used anaphorically, can be used without an antecedent, for example when using, say, she to refer to a person who just walked in, or whom I point to, or whom we both watch on a TV monitor (though one could argue that their very appearance, or my pointing at them, actually introduces the discourse referent, which is then anaphorically picked up by the use of the pronoun). Indefinite NPs, including cardinals and certain quantifiers such as few, most, etc. are perhaps the only morphological class that is consistent, in that they always introduce new referents, and cannot, it seems, be used anaphorically. Discourse Representation Theory provides a framework in which notions such as 'having a discourse antecedent,' 'introducing a discourse referent,' or 'being incapable of having a discourse antecedent' can be formalized; representative examples are Heim's (1982, 1983) File Change Semantics; Kamp's (1981) Discourse Representation Theory; and Groenendijk and Stokhof's (1991) Dynamic Predicate Logic. Useful introductory texts include Kamp and Reyle (1993) and Chierchia (1995). In addition, a theory of discourse anaphora should address the question which type of NP will most likely be used anaphorically to pick up a particular referent (it seems that the choice between, say, a pronoun and a definite NP depends on roughly how long ago the pertinent discourse referent was introduced), and conversely, which one of a number of already introduced referents a given anaphoric NP is likely to pick out. Theories that address these questions include Centering 42 INTERPRETING INDEXED STRUCTURES Theory (Grosz et al., 1995; Walker et al., 1998), Accessibility Theory (Ariel, 1999, 2001), and Relevance Theory (Sperber and Wilson, 1995) among many others; see also Kehler (2002) for a hybrid proposal. I will not introduce a theory of context change, centering, or relevance in this book, which focuses mainly on binding within a sentence; that is, as far as our official theory goes, (b) in (2.39) for example, has as its possible meaning 'Fido sees a man come in,' where Fido is the dog mentioned in (a), but also 'Hasso sees a man come in,' where Hasso is the other dog mentioned in (c), and that it is up to a discourse theory to filter further among these. It is worth noting that the question when an NP can pick up the referent introduced by another NP; can sometimes influence our acceptability judgments on a simple sentence as well. For example, most speakers will find it very hard to interpret his and Edson as coreferential in (2.41a), while that interpretation is easy to get in (2.41b): (2.41) (a) How did other mothers cope with this? For example, what did his mother think, when she learned Edson got arrested? (b) How did other mothers cope with this? For example, what did Edson's mother think, when she learned he got arrested? The same sentence as in (2.41a), however, improves considerably on a coreferential reading if the preceding context is changed: (2.42) What did his family think of Edson! For example, what did his mother think, when she learned Edson got arrested? This pattern suggests that the problem with (2.41a) is that his cannot pick up the referent introduced by Edson, presumably because it precedes it (whence the improvement if the NPs are swapped as in [2.41b]). If the previous sentence already talks about Edson as in (2.42), no such 'backwards link' is necessary, and his is naturally understood to refer to Edson. It seems appropriate to say that while his and Edson corefer in the second clause in (2.42), his is actually anaphoric to the occurrence of Edson in the first sentence in (2.42) (as is the second occurance of Edson, presumably). To be sure, these remarks are not meant to provide a serious account of so-called backwards anaphora; it is quite possible that in some configurations, pronouns can be genuinely anaphoric to a full NP that follows them. If our remarks about the above examples are on the right track though, they do suggest that we were justified in separating out the purely structural Binding Conditions A-C, which have nothing to say about any of these cases, from a more general theory of introducing and anaphorically picking up discourse referents, which applies to intra- and inter-sentential anaphoric relations alike.9 9 This doesn't mean that there can't be controversy about which kind of effect belongs to which theory. Evans (1980) claims that Binding Condition C effects are not structural but result from 2.4 Extensions and alternatives© 43 2.4.2 Reflexives as reflexivizers i Above we assumed that reflexive and non-reflexive pronouns are se-mantically identical; they are interpreted like individual variables. Whatever differences we find between them are accounted for in the syntax, i.e. by indexing. An alternative approach is to view reflexives as semantically distinct from non-reflexive pronouns. As a starting-point, assume that the reflexive in a sentence like Natasha painted herself serves to map the relation denoted by painted onto the property of self-painting. Formally, this means that the meaning of herself represented as SELF, maps a relation R onto the property of being an individual that stands in the R relation to itself, kx.R(x)(x): (2.43) (a) SELF is that function from relations to properties such that SELF(R) = Xx.R(x)(x) for all R e Deet (b) fhimself/herself/itself] g = XRXx. (x is a male/female/neuter individual and) SELF(R)(x) Interestingly, a reflexive on this view does not need to be indexed, since it is not a referring expression. The fact that it is 'locally bound' simply follows from its lexical meaning: it maps the predicate it combines with onto what we might call a 'reflexive property.' Thus something close to Binding Condition A actually follows from this analysis of reflexives. Obviously, nothing about non-reflexives follows. One could formulate principles to re-establish complementary distribution again, however, e.g. (2.44): (2.44) (a) No pronominal can be used where a reflexive would yield the same meaning. (b) Unless a transitive verb V has a reflexive pronoun as its argument, interpret it as Xx.Xy.lYJg(x)(y) &x y. As just discussed, this approach is attractive in that it builds Binding Condition A into the lexical meaning of the reflexive: the reflexive will automatically be 'bound,' and it will be bound by a higher coargument. It is, however, well known that reflexives do not always combine with a relation denoting expression, cf. (2.45a) and (2.45b), and that in such cases the binder of a reflexive is not always the next higher coargument (in [2.45a] the 'highest' argument, the subject, binds the 'lowest,' the direct object, skipping, as it were, the indirect the fact that a pronoun is trying to pick up its referent from a full NP it c-commands (he claims, in line with the other authors quoted at the end of section 1.2.3 in chapter 1, that a full NP c-commanding a coreferent full NP isn't ungrammatical at all). He adduces the acceptability of (i) below (his [59]) as evidence for that position. I believe though that (i) instantiates a special case of what Heim (1993) calls 'when structured meanings matter', which we will discuss briefly at the end of section 6.3 in chapter 6. Generally, a pronoun c-commanding a coreferent NP, even in the presence of an independent antecedent as in (ii), appears to be rather bad (compare [ii] to [2.42]), and should be ruled out on structural grounds: (i) What do you mean John loves no one? He loves John. (ii) * I wonder how Edson felt. What did he think when Edson got arrested? 44 interpreting indexed structures object in the middle); indeed, the antecedent needn't be a coargument at all, cf. (2.45c) and (2.45d): (2.45) (a) Gabi introduced herself to John. (b) Gilberti told Spencer about himselfi/2. (c) The president saw [himself give a speech]. (d) The CEO despised [those recent articles about herself]. In none of these cases can the reflexive be interpreted as in (2.43b); furthermore, to capture the two coreference options in (2.45b), we would have to assume two different meanings for the reflexive (or two different structures for the sentence). While appropriate meanings can be devised, the challenge is to show that the resulting class of meanings is constrained in some natural way. Otherwise, the distribution of reflexives reduces to an idiosyncratic case of lexical ambiguities, hardly a satisfactory state of affairs.10 On the other hand, even if this treatment turns out to be inadequate for English himself, there might be other elements in other languages that are amenable to it. In particular, many languages have a verbal reflexive marker which essentially de-transitivizes a transitive verb to give you an intransitive verb. This marker might aptly be characterized as denoting SELF. Consider the following examples from Finnish and Chichewa (from Sells et al. [1987]:177,187): (2.46) (a) Jussi pese -yty -i. (Finnish) /. washed self past 'Jussi washed himself/self-washed.' (b) Jussi puolusta -utu -i. /. defended self past 'Jussi defended himself/self-defended.' (2.47) Alenje a- na- dzi- lum -a. (Chichewa) hunters SA past self bite indicative 'The hunters bit themselves/self-bit.' The affixes -ytyI-utu and dzi- are not syntactic arguments or clitics (they cannot, for example, be separated from the verb or be coordinated with other NPs; see Mcahombo [1993a] for more evidence), and the derived form is syntactically intransitive (e.g. the -ytyI-utu forms form causatives following the intransitive instead of the transitive pattern). It seems plausible that these affixes semantically function like SELF described above. Likewise, certain other reflexives with a more limited distribution than English himself are perhaps best treated this way (cf. Sells etal. [1987]). Exercise 2.12 The relation that is reflexivized in (2.45a) can be written as kx.ky.Gabi introduced y to x. (2.48) gives the appropriate lexical entry for 10 Another problem with this approach is what happens if the reflexive is coordinated, as in Polsen painted himself and Mary, especially on a non-distributive reading of the coordination, a matter we cannot go into here. 2.4 Extensions and alternatives© 45 herselfto derive this meaning (where R3 is a variable over three-place relations, i.e. a variable in De^e,et)): (2.48) [herself^]* = A./?^A.xi.Ajc2.rt3(*i)(*i)(*2) (= Ai?3 .Xx\.Xx2 .SELF(Xx3 .XX3.R3 (X3) (x*) (x2)) (*i)) Note that this rule assumes that the order of Function Application for V NPi [to NP2] is actually IVF(INP2FXI[NPiF. Characterize the relation that is reflexivized in (2.45b) in the same way. Then give a lexical meaning for herself that derives the correct meaning. Exercise 2.13 Consider the following Kannada data (from Lidz [1995]): (2.49) (a) shyaamui raamu2 tann-annui/* 2 hoDe -d -a 5. R. pron-acc hit -past-3SM anta heeL -id -a that say past 3sm 'Shyamu said that Raamu hit him.' (b) shyaamui raamu2 tann-annu2/* 1 hoDe -du -koND S. R. pron-acc hit -past REFL. -a anta heeL -id -a -3sm that say past 3sm 'Shyamu said that Raamu hit himself.' What unexpected property do these sentences show? Discuss! Does this suggest alternative treatments for the Finnish and Chichewa cases above? 3 Domains and orientation In chapter 1, section 1.2.5 we arrived at a general format for Binding Conditions, which I repeat here: (3.1) An NP of class must (not) be coindexed with a commanding NP within its domain. We saw that the grammar of English contains three conditions of this form, pertaining to three disjoint classes of NPs. In the discussion up until now we have assumed that the relevant binding domain in which reflexives need to be bound, and non-reflexive pronouns need to be free, is the local clause. In this chapter we are first going to refine this notion of binding domain. The discussion will start with English, again, showing that the correct description of the relevant domain should be something like 'smallest category containing a subject,' rather than just 'smallest clause.' Second, we will introduce an additional parameter into the Binding Conditions, orientation. Whereas orientation doesn't seem to be central for the description of English pronouns (but see chapter 11), it is very important in many other languages. We will introduce three kinds of orientation: subject-orientation, anti-subject orientation, and logophoricity. Third, we will then examine a range of data from other languages that requires Binding Conditions to make reference to different domains (both different from English, and different among the classes of NPs within a given language). Fourth, we will briefly discuss so-called long-distance reflexives, that is, cases in which reflexive pronouns lead a 'double-life' as either locally bound (like in English) or bound in the sentence domain, like subject oriented pronouns and logophors. With all these refinements in stock, then, we will fifth and finally look at a few complete pronominal systems in languages other than English. 3.1 Binding domains in English: governing category 3.1.1 Exceptional case marking (ECM) Consider first so-called exceptional case marking (ECM) constructions. The hallmark of these is that a certain class of ECM-verbs case-mark an NP 46 3.1 Binding domains in English: governing category 47 which is thematically the argument to an embedded verb. Examples are the verbs believe and want in English. The thematic subject of the lower verbs, Georgína, behaves like a grammatical object to the higher ECM-verbs: it is marked with accusative case, (3.2b), and gets promoted under passivization of the matrix verb, (3.2c). We will henceforth refer to these phrases, italicized in the following examples, as the ECM-subjects: (3.2) (a) O'Leary wants/believes Georgína to lie. (b) O'Leary wants/believes her to lie. (c) Georgína is believed to lie. Turning to binding, now, we observe that the ECM-subjects behave like they are clausemates to the matrix subject: coreference requires a reflexive and prohibits a pronominal: (3.3) 0'Leary6 believes himselfó/him^ to deserve the crown of England. At the same time the ECM-subjects behave like clausemates with respect to NPs in the embedded clause: they trigger reflexivization in the lower clause and do not tolerate coreference with a non-reflexive object pronoun: (3.4) O'Leary wants Georginas to protect herselfs/her^^. These two findings are yet unproblematic. They seem to suggest that all NPs in an ECM-construction, whether they thematically belong to the embedded verb or the matrix verb, populate the same binding domain in the sense relevant to Binding Conditions A and B. This, however, predicts that the matrix-subject should count as a clause-mate to the embedded non-subjects, too, which it patently does not: (3.5) 0'Learyi2 wants Georgína to protect himi2/*himselfi2. At this point we are in a bind: whatever constituent is the binding domain for the embedded object must include the ECM-subject (because of [3.4]), but exclude the matrix subject (because of [3.5]). But then that constituent, even though containing the ECM-subject, cannot be the binding domain for the ECM-subject (because of [3.3]). In other words, the pertinent binding domains must overlap as in (3.6):1 binding domain for SUBJe ,-A-, [s<«SUBJm ...V [Se SUBJe [ V OBJ ]]] (3.6) -•- binding domain for OBJ These domains are correctly computed by the definition of governing category (GC) in (3.7): 1 The schema in (3.6) assumes that the ECM-subject resides within the embedded clause. Essentially the same conclusion would hold, however, if we assumed that it raises to the matrix clause (i.e. if these sentences are to be analyzed as raising to object). 48 domains and orientation (3.7) y is the governing category (GC) for NP if and only if (iff) y is the smallest clausal category (S, S, IP, CP, TP...) which dominates (a) NP (b) NP's case assigner (3.8) Binding Conditions (preliminary): (A) A reflexive pronoun must be bound in its governing category. (B) A non-reflexive pronoun must be free in its governing category. The hypothesis expressed by (3.8), then, is that the binding domain for both reflexive and non-reflexive pronouns in English is their governing category. (3.7) can perhaps be understood best by looking at (3.6): the embedded object, OBJ, receives its case from the embedded V, hence its GC is the embedded clause. Accordingly, the embedded object must be reflexive if coreferent with the ECM-subject, but not if coreferent with the matrix subject. The ECM-subject, on the other hand, receives its case from the matrix-verb (exceptionally, as it were), which means that the embedded clause is not its GC: even though it contains the ECM-subject itself, it doesn't contain its case assigner. Indeed the smallest clausal category containing both the ECM-subject and its case assigner is the matrix clause. Accordingly, coreference of the ECM-subject with the matrix subject requires the former to be reflexive. 3.1.2 Infinitival clauses Next, let us briefly look at infinitival complements other than ECM. At first, pairs of examples like those in (3.9) seem to provide contradictory evidence. (3.9a) suggests that the binding domain for the embedded object is the matrix clause, but (3.9b) suggests that it is something smaller than that, e.g. the infinitival clause/VP: (3.9) (a) John3 tried to educate himself3/*him3. (b) Anai told John to educate heri/*herselfi. The mystery is resolved, however, once we recognize that try is a subject-control verb (the understood subject of the embedded verb is the matrix subject), whereas tell is an object-control verb (the understood subject of the embedded verb is the matrix object). We can then establish the following generalization: (3.10) An infinitival clause functions as a G(overning) C(ategory), with the understood subject acting as a binder. This generalization receives confirmation from sentences like (3.11) (compare to [3.9b]), and the structurally parallel Marathi example (3.12):2 2 From Dalrymple (1993):17. 3.1 Binding domains in English: governing category 49 (3.11) Ana told John3 to educate himself3 /*hini3. (3.12) Johni ne Jane2 laa swataahlaa2/* i maraaylaa saangitle (Marathi) John erg Jane dat self-acc hit told 'John told Jane to hit self (=Jane).' In both examples the embedded reflexive corefers with the controller, which happens to be the matrix object. What makes the Marathi case interesting is that the reflexive swataah is subject oriented; it cannot ever be bound by a non-subject argument (see section 3.2 below). If the matrix object Jane laa were the actual binder in (3.12), this sentence would constitute a mysterious exception to the subject-orientation of the reflexive. But assuming, as we did, that the logical subject counts as a binder in these cases provides an immediate explanation. How can we implement (3.10)? If we assume that phrase structural c-command is the relevant notion, we will have to assume an actual empty NP as the subject of the embedded infinitival clause, commonly called PRO. The index on PRO depends on the matrix verb (subject, object, or arbitrary control), but PRO is the inevitable binder for a reflexive or reciprocal in the embedded clause. Representative structures are given in (3.13): (3.13) (a) John4 tried [PRO4 to educate himself^. (subject control) (b) Ana2 told John4 [PRO^/^ to educate himself4/*herself2]. (object control) It bears emphasizing that none of the binding principles discussed so far is responsible for the indexing of PRO in (3.13), e.g. that it must be indexed 4 rather than 2 in (3.13b). BT as discussed here only governs the indexing of the embedded pronouns.3 Alternatively, if we pursue an argument-structure based account of binding, things are even more straightforward. All that needs to be assumed is that verbs like try, tell, and adjective-based predicates like be dangerous select for a VP, and encode that the index of the (unsaturated) subject-argument of that VP be unified with the pertinent element on the a(rgument)-list of that verb: (3.14) (a) a-list for try: (NPi, VPf0[SUBJ[(l)]]) (b) a-list for tell: (NP, NPacca, VPf0[SUBJ[(l)]]) (c) a-list for promise: (NPi, NPacc, VPf0[SUBJ[(l)]]) The indices indicate that NPi is on the a-list of both the matrix verb and the embedded verb (by virtue of being the SUB J of the embedded VP). Accordingly, more oblique elements on the a-list of the embedded verb can be bound by NPi, regardless of the position of NPi on the matrix a-list; for more details see Sag and Pollard (1991); Sag and Wasow (1999). 3 The interested reader is referred to Harbert (1995) and the references therein for attempts at subsuming the choice of (co)reference for PRO under BT as well; for a recent discussion of control see Landau (2001a,b). 50 DOMAINS AND ORIENTATION Exercise 3.1 The Marathi sentence in (3.12) made an argument that the reflexive is not bound by an NP in the matrix clause, but by the (invisible) subject of the infinitival clause. Make an argument to the same effect using the English sentences in (3.15): (3.15) (a) It is embarrassing to see yourself in the newspaper headlines like that. (b) It is dangerous to trust each other like that. (c) John agreed with Mary to bring each other's pictures to the meeting. 3.1.3 NPs as binding domains In this section we will briefly consider another modification of the definition of G(overning)C(ategory). NPs with and without subjects So far we have almost exclusively looked at NPs in verbal argument position (subject, object), when examining Binding Conditions. If we expand our data base, we find that NP positions in adnominal argument positions seem to behave as predicted: (3.16) Johns saw [np a picture of himself5 *him5 ] The pronoun, functioning as an argument to the noun picture, receives its case from the preposition of within the NP. But the only clausal node dominating the pronoun is the matrix S, which also dominates the subject. Therefore the subject is in the GC of the nominal argument, which therefore must be reflexive if coreferent with the subject. Things change, however, if we consider NPs with a prenominal argument (a possessor); now the non-reflexive is permitted, and, according to many authors, only the non-reflexive: (3.17) Johns saw [np Mary's picture of ?? himself5 him5 ] Assuming these judgments, it seems that an NP functions like a (finite) clause with regard to BT if and only if it has a possessor. Accordingly, the following revision of the definition of GC in (3.7) suggests itself: (3.18) y is the governing category for NP, iff y is the smallest category that has a Subject and dominates (a) NP (b) NP's case assigner (where a Subject is either a clausal subject or a possessive) Applied to all cases in sections 3.1.1 and 3.1.2, this definition will yield the same results as before, given that clauses contain subjects (and clausal subjects are Subjects in the technical sense). If a pronoun is contained in an NP, however, that 3.1 Binding domains in English: governing category 51 NP, rather than the clause minimally containing it, can be the GC, if it contains a possessor. This is the case in (3.17): the GC for him(self) is the NP Mary's picture of him(self), which implies that a reflexive has to be bound within that NP (which it evidently can't, given the gender mismatch). The NP in (3.16), on the other hand, is not the GC for him(self) because it contains no Subject, so the entire clause is. Accordingly, a pronoun must be free within the entire clause, and coreference with the clausal subject requires reflexivization. This is probably the right time to note that at least one part of the generalization that motivates the formulation in (3.18) is not borne out by the facts, namely that post-nominal reflexives in NPs with a possessor only allow binding to the possessor. In a magnitude estimation experiment with 52 English speakers, Keller and Asudeh (2001) found that reflexives and non-reflexives are judged equally acceptable in a sentence like (3.19a) (in contradistinction, a reflexive in a sentence like Joan's father respects herself was clearly rejected). Similarly, Runner et al. (2002) found in an eye-tracking experiment that almost 25 percent of their subjects interpreted the reflexive to denote Ken in a sentence like (3.19b); similarly, examples of reflexives bound from outside a possessive NP are widely attested, e.g. (3.19c):4 (3.19) (a) Hanna found Peter's picture of her(self). (b) Have Ken touch Harry's picture of himself. (c) "C.B.'s father had fared better in this respect than most of his forebears, but still resented his wife for her low opinion of himself, of the Whiting mansion, of Empire Falls, of the entire backward state of Maine..." Asudeh and Keller's experiment also revealed that post-nominal non-reflexives bound to a possessor or a local sentential subject in possessor-less examples (essentially the hims variant of [3.16]) - both of which are predicted to be un-grammatical by the account given here - while not fully acceptable, are significantly better than non-reflexive object pronouns bound to verbal coarguments. It is a well-acknowledged fact that the data in this area of BT are complex and hard to judge (see e.g. Kuno [1987]:section 4.3; Reinhart and Reuland [1993]:683, 690). Experiments like Runner et al. (2002) and Keller and Asudeh (2001) are of utmost importance in that they provide a way of establishing a reliable data base even where individual speakers' introspective judgments are insecure or inconsistent. For a theoretical interpretation of some of these findings, see Asudeh and Keller (2002). Non-complementary environments Let me finally discuss an additional complication regarding pronouns within NPs, and briefly sketch a way of addressing them that roughly follows the proposals in Huang (1983) and Chomsky (1986):164ff. Although these proposals 4 From Richard Russo, Empire Falls (Kampf), p. 4, found by C. Potts (italics added). 52 DOMAINS AND ORIENTATION have received a lot of attention in the literature, I believe they are superseded by the simpler and more accurate treatments afforded within the proposals discussed in chapter 11, so I will skip details wherever possible. The complementarity between reflexive and non-reflexive pronouns has been at the heart of the definitions we have provided so far. It turns out, however, that while complementarity is observed in the vast majority of cases, it isn't always. (3.20) is a case in question: (3.20) Johny believes [that [s? pictures of J1™7 ,r \ are on sale]] nimseli7 J According to our definition (3.18) the GC for him(self) in (3.20) should be the embedded clause Se: it contains the pronoun, its case assigner of, and a subject pictures ofhim(self). This result yields the correct prediction for the non-reflexive him, which is free in Se and can thus corefer with the matrix subject. By the same token, however, it blocks the reflexive, which is not bound in Se. It seems that reflexives and non-reflexives part company here: the GC for himself appears to be the matrix clause; that for him the embedded clause. Suppose this generalization is correct, then the question emerges if there is any way to define GC so as to get these two different domains for him and himself in (3.20). One ingenious attempt at that is found in Chomsky (1986). Perhaps the best way to illustrate the gist of Chomsky's proposal is this: the GC for the reflexive in (3.20) doesn't contain any c-commanding NP (note that the embedded subject itself doesn't c-command an NP it contains). It is thus in principle impossible for the reflexive to meet the Binding Condition pertinent to it, Binding Condition A, within that GC. This is different for the non-reflexive, because a GC without any c-commanding NP is just a special case of a GC without a binder, so a non-reflexive can, and in fact always will, meet its Binding Condition, Binding Condition B, within such a GC. The idea then is that the GC for a given NP must be chosen 'mercifully,' in such a way that NP can at least in principle meet its Binding Condition in that GC. Consider the following revision of (3.18) (cf. Chomsky 1986:171f.): (3.21) y is the governing category for NP iff y is the smallest category that has a Subject and dominates (a) NP (b) NP's case assigner (c) an NP c-commanding NP, if NP needs to be bound Take sentence (3.20) again, with a reflexive (John believes that pictures of himself are on sale). The GC for himself according to (3.21a-3.21b) would be the embedded clause Se, but that clause doesn't contain any c-commanding NP, which could function as the binder of himself, as required by Binding Condition A. Accordingly, (3.21c) mercifully 'broadens' the GC to the next clause up, which indeed contains a binder for the reflexive. 3.1 Binding domains in English: governing category 53 In the case of a non-reflexive (John believes that pictures of him are on sale), (3.21a-3.21b) again determine Se as the GC, but this time (3.21c) doesn't change anything about that because Binding Condition B, the Binding Condition pertinent for non-reflexive pronouns, doesn't require any binder at all. So Se is the ultimate GC for him, in which it is free, as required. It remains to verify that even the relaxed definition of GC in (3.21) doesn't rule in cases like (3.22): (3.22) (a) *Mary3 said that [^i John believes that [Se2 [pictures of herself^] are on sale]]. (b) *Johni believes that [[Mary's pictures of himselfi] are on sale]. The GC for herself in (3.22a) is Sel, not the matrix clause, because Sel already contains an NP which could bind the reflexive, meeting Binding Condition A; the fact that John cannot be the actual antecedent to herself is irrelevant to (3.21c). Likewise, himself in (3.22b) has the subject NP Mary's picture of himself as its GC, given that that NP contains a c-commanding NP, Mary's, that could serve as the binder (were it not for the gender mismatch). The definition of GC in (3.21) predicts non-complementarity for two more positions: possessives and clausal subjects, as both of these, being the Subject in the sense of (3.21), do not have a c-commanding NP in their 'original' GC. This prediction turns out to be correct for the former case, but incorrect for the latter (we use a reciprocal here, since English doesn't have reflexive possessives): their6 (3.23) (a) They6 love [ { , , > pictures]. [ each other6 s J (b) * They3 think [that each other3 will win].5 An independent reason why (3.23b) is unacceptable has been proposed in Rizzi (1989), namely that reflexives universally cannot occur in agreeing positions.6 But even if Rizzi's generalization is correct, reasonable doubts about the validity of the 'mercy-condition' on the definition of GC have been voiced, and alternative and more comprehensive accounts have been proposed (see chapter ll).7 5 Lebeaux (1983) and many following him have claimed that sentences like (3.23b) are slightly better than full-blown Binding Condition A violations, and become virtually acceptable in the context of w/z-extraction such as 11 They don't know what each other are doing. I am not aware of any coherent account of this contrast. 6 A very different attempt at explaining the ungrammaticality of (3.23b), involving movement of the reflexives, is found in Chomsky (1986). 7 It is also instructive to note that binding of reflexives and reciprocals in embedded finite subjects is by far not universally allowed. Languages as closely related as Dutch and German strictly prohibit this: (i) Martell hofft, dass eine Reportage über ihnj/^sichj im Radio gespielt M. hopes that a report about him/himself in-the radio played wird. (German) becomes 'Martell is hoping that a report about himself is going to be aired.' Likewise, Kannada, Italian (Yang, 1983); Polish (Reinders-Machowska, 1992); Russian (Rappa-port, 1986) do not allow this kind of binding for reciprocals and/or reflexives. 54 DOMAINS AND ORIENTATION Exercise 3.2 Assume we replace (3.21c) by a clause that says 'enough material for NP to meet its Binding Condition (ignoring mismatches in person, gender and number),' while maintaining the rest of definition (3.21) and the Binding Conditions in (3.8). This would seem to express the idea of 'be merciful where appropriate' even better. Yet it is haunted by a fatal formal problem. Which? 3.1.4 PPs as binding domains According to the definition of GC in (3.18), a GC needs to contain a Subject. Whatever the details of the technical Subject notion, it seems clear that PPs don't contain a subject and should therefore not constitute a GC for their complement NP. In other words, the prediction is that in [PP P NP], NP can be a reflexive bound from outside of the PP; by the same token, if NP is non-reflexive, it should have to be free within the next higher domain containing a subject (e.g. the clause of which PP is an immediate constituent). What are the data? Consider (3.24): (3.24) (a) Johni sent a letter to him*i/himselfi. (b) Johni always relies on him*i/himselfi. These sentences are as expected. But they contrast with the superficially parallel (3.25): (3.25) (a) Johni looked around himi/himselfi. (b) Johni pulled the blanket over himi/himselfi. (c) Muhammadi hid the book behind himi/himselfi. Here it seems as if the binding domain for him must be smaller than that for himself. For example Hestvik (1991), following unpublished work by Joan Bres-nan, proposes that the binding domain for him is the PP, while it is the clause for himself, accordingly, complementarity between him and himself breaks down, as the former is free within PP, while the latter is bound within S. If we adopt this kind of analysis, it means we give up on the assumption that there is one binding domain, the governing category, that is relevant for both Binding Condition A and Binding Condition B (we effectively gave up that assumption in section 3.1.3 above, but now it seems less likely that there could even be a uniform formulation of GC). While the binding domain for reflexives can remain what it was (the smallest category containing it, its case assigner, and a Subject), the binding domain for a non-reflexive pronoun should include the subject only if the pronoun is a complement of a verb (to block *He\ likes him\), but not if it is a complement of a preposition (as in [3.25]). Let us define NP's coargument domain as the smallest XP that contains NP, NPs case assigner C, and all other arguments of C (cf. section 3.3 below). Since 3.1 Binding domains in English: governing category 55 subject and all objects are arguments of a verb, the coargument domain for any verb argument is its minimal clause. Given that a preposition has only a complement, but no subject, the coargument domain of an NP selected by a P is the PP. This is the distinction we're after: (3.26) Binding Conditions (final) (A) A reflexive must be bound within the smallest category containing it, its case assigner, and a Subject (=its GC). (B) A non-reflexive must be free in its coargument domain. Given these definitions, we account for the non-complementarity in (3.25), but we lose our account of the complementary distribution in (3.24), because there, too, the non-reflexives would now be free within the PP. Now, there is arguably a difference between the PPs in (3.24) and those in (3.25). While they are all selected by the verb (or so we will assume), the prepositions in (3.25) make a clear semantic contribution to the sentences, while those in (3.24) seem semantically empty. To be sure, it is not easy to motivate this distinction in every specific case (surely on and perhaps even to can have semantic content in other sentences). It is, however, suggestive that the Ps in the complements to look, pull, and hide can be exchanged for others as in look behind NP, pull it around NP, or hide it next to NP, and the whole PPs for proforms as in look there, pull it up, and hide it away. No such variation is possible in the case of sent or rely (*sent it on him, *rely there, etc.). Let us assume for concreteness, then, that look, pull, and hide truly select a PP that denotes a path (or location), while sent to and rely on are really complex verbs that semantically combine with the NP denotation, so that the P has no semantic function on its own. Assuming this difference, how can it help us to explain the difference in the acceptability of non-reflexive pronouns? The intuition we are after is that the PPs in rely [PP on NP], sent X [PP to NP] and their likes are not the coargument domain for NP, because NP is 'really' an argument to V, whose arguments include the subject. Now, this would follow from our tentative definition of coargument domain above, if we could plausibly argue that V, not P, is the case assigner for NP in these cases. That, however, isn't obvious, and we will not pursue this option further here. Failing that, the non-domain status of these PPs would also follow if we replaced the notion of case assigner in the definition of coargument domain by the notion of thematic role (©-role) assigner, or semantic predicate. Consider the schemata in (3.27): ^-®-y^--0-— Johni relies [pp on [np himselfi/*himi]] coargument domain Johni looked [pp around [np himselfi/himi]] coargument domain 56 DOMAINS AND ORIENTATION Since him receives a ©-role from relies in (3.27a), all arguments to relies are within its coargument domain, in which it needs to be free. In (3.27b), on the other hand, him receives its ©-role from around (while the entire PP in turn gets a ©-role from looked), which doesn't have any further arguments and therefore makes PP the coargument domain, in which him must be, and is, free. This will be our account of the different behavior in these cases. The alert reader will recall that a non-reflexive pronoun cannot be bound to a local subject, even if the subject and the pronoun aren't thematic coar-guments. The crucial example involving exceptional case marking is repeated here: (3.28) 0'Leary6 believes himselfe/him^ to deserve the crown of England. Note that by a purely semantic definition of coargument domain in terms of ©-role assignment, /z/m's coargument domain in (3.28) should be the embedded clause, in which it is free, which wrongly predicts possible binding of him by O'Leary in (3.28). If, however, we defined the coargument domain to include the case assigner (rather than the © assigner) and its other arguments, we're back at predicting him to be possible in (3.24) (assuming that the case assigner is the semantically vacuous P). It seems we have to bite the bullet and adopt a disjunctive formulation of coargument domain as in (3.29): (3.29) NP's coargument domain is the smallest constituent X which contains (i) NP, (ii) NP's case assigner C, (iii) NP's 0-role assigner T, and (iv) every XP whose case or 0-role is assigned by C or T. This, I grant, is an ugly definition, but it copes with both kinds of PP cases, as well as the ECM cases (and raising cases, to be discussed in chapters 11 and 12). We will return to some of the issues involved here in chapter 11, especially section (11.4.3), but the basic disjunction will stay with us. In the discussion of binding domains in section 3.3 below, we will abstract away from these complications and use a simpler definition of coargument domain, ignoring the fact that similar data to the ones discussed in the present section are found in other languages as well (see e.g. Hestvik [1991] and the references there). Before leaving the issue of PPs as binding domains, let me briefly mention cases of true and uncontroversial adverbial PPs, as opposed to selected PPs, as in (3.30): (3.30) (a) Maxi saw a ghost next to himi/himselfi. (b) Johni found a dollar bill in front of himi/??himselfi. (c) John heard a strange noise behind him/*himself. The grammaticality marks in (3.30a), (3.30b), and (3.30c) are as given by Reinhart and Reuland (1993):686; Hestvik (1991); and Kuno (1987):65, 3.1 Binding domains in English: governing category 57 respectively; this variability of judgments is surprising, since the examples are parallel in all relevant respects, and no single author offers reasons to distinguish them. How does our current approach fare here? The acceptability of the non-reflexive is expected and predicted, since clearly the P heading the adverbial PP is semantically potent and thus defines its own coargument domain. However, the impossibility or degraded acceptability of a reflexive - as reported in (3.30) - is unexpected, since the smallest category containing a Subject is obviously the clause. We could introduce yet another quirk in the definition of binding domain for reflexives, if we wanted to rule out reflexives in adjunct PPs systematically; examples of such a strategy can be found in Hestvik (1991) and Kuno (1987):ch. 2. Generally, though, reflexive pronouns in adjunct PPs are found rather frequently in actual text:9 (3.31) (a) Mrs B. who hears the steps behind herself feels rottenly and goes faster. (b) He supplied the end of the cord near himself with a conductor in the shape of an iron key. (c) Mr. Meynell, finding him in better health but suffering from the loneliness of his life, brought him to London and established him near himself. (d) The seductress must be careful not to cast this spell near herself. Given this, we will assume that reflexives in adjunct PPs can generally be bound from outside, leaving open the question of what accounts for the degraded status of (some of) these examples; see chapter 11, especially section 11.3.2, for further discussion. This concludes our discussion of the distributional data from English. In the face of considerable unclarity in the data, we tentatively adopt the Binding Conditions A and B in (3.26), which fix the binding domains for reflexives and non-reflexives as the subject domain and coargument domain in the sense to be introduced in (3.3) below, respectively. In the following sections we will extend our perspective to a wider variety of languages. We will see that neither the tripartition into reflexives (and reciprocals), non-reflexive pronouns, and full NPs, nor the notion of GC developed in this section seem directly applicable cross-linguistically. Nonetheless I will attempt to make visible the outlines of a general format of Binding Conditions. Reinhart and Reuland (1993):687 remark, though, that the reflexive in (3.30a) is 'much more marked' than in sentences like (3.24), which they attribute to the competing possibility of using a non-reflexive (though they don't say why that should be preferred); Kuno on the other hand uses stars but describes the contrast by merely stating that '.. .for many speakers the following sentences are better with a [non-reflexive; DB] pronoun than with a reflexive' (p. 65). 9 From http://www.plew.de/english/Note.htm, a description of Benjamin Franklin's famous experiment at http://www.home.zonnet.nl/kitedude/VEK2.htm, the Catholic Source Book on Francis Thompson (http://www.newadvent.org/cathen/14703b.htm), and the Realm of the Dragons AD&D community and info, site, respectively. 58 DOMAINS AND ORIENTATION A word on the terminology to be used. I will use the term pronoun in the usual sense, including forms that are not inherently anaphoric, such as first and second person pronouns. The term reflexive will be used in the text and the glosses for pronouns that need to be bound, including forms that can be bound across clause boundaries, where this is done by the authors whose presentation I allude to (a more apt term for these pronouns might be command anaphor, which would leave the term reflexive for a morphologically denned class). I will use the general term binding domain instead of governing category. For one thing, it is unclear whether binding domains relevant in other languages are usefully described in those terms used in the definition of GC. For another, it is a particular property of English that the two relevant pronoun-classes, reflexives and non-reflexives, make reference to the same domain (if indeed they do) for their respective Binding Conditions. As we will see, different pronouns within the same language, and even different conditions pertaining to the same pronoun, can make reference to different domains. These domains will be designated by more informative names. The GC as denned in (3.18)/(3.26), for example, which is the binding domain for reflexives in English, will be called the subject domain. 3.2 Orientation In this section we will extend our blueprint for binding conditions by one parameter, orientation. Orientation is not relevant in English, but has an important role to play in many other languages. To a first approximation, orientation means that a certain anaphoric element must be bound to or free from NPs with a particular grammatical function, in most cases a subject. 3.2.1 Subject orientation and anti-subject orientation i An illustrative example of subject orientation is the Chinese reflexive ziji. As widely discussed in the literature, ziji must be bound within its root clause, but not necessarily within any local domain. However, the antecedent has to be a subject:10 Dalrymple (1993) (2.41); Tang (1989) (45); according to Huang (1982), ziji can only rind a nonlocal antecedent if it is itself in subject position. But as Tang (1989) argues using (i), this doesn't seem correct: (i) Zhangsan3 juede Lisis dui ziji3/8/*i mei xinxin. Zhangsan think Lisi to self no confidence 'Zhangsan thought that Lisi had no confidence in himself.' Also, ziji's antecedent can sometimes be an experiencer or a phrase contained in a subject (cf. Huang and Tang's [1992] notion of subcommand). 3.2 Orientation 59 (3.32) (a) Zhangsani shuo zijii hui lai. (Chinese) Zhangsan say self will come 'Zhangsan says he will come.' (b) Zhangsani renwei Lisi2 zhidao zijii/2/*3 de taitai shi yige Zhangsan think Lisi know self DE wife is one-Ch da hao ren. big good person 'Zhangsan thought that Lisi knew that his wife was a very good person.' Subject orientation and domain restriction are not mutually exclusive. The Finnish reflexive Use, 'self,' must be bound by a subject within its minimal finite clause:11 (3.33) (a) Pekkai naki etta Matti2 katsoi itseaan2/*i (Finnish) Pekka saw that Matti watched self-POSS 'Pekka saw that Matti watched himself.' (b) * Puhuin Pekallei itseaani/2 spoke-1SG Pekka self-POSS T spoke to Pekka about himself.' Subject orientation of pronouns is cross-linguistically very pervasive; apart from the cases mentioned here, it is reported to hold for Chinese taziji}2 Czech si, Danish (sig) selv, Dutch zich, French soi, Icelandic (sjdlfur) sig, Italian se, Japanese zibun, Kannada hon and ta-nu, Latin se, Malayalam swa,13 Marathi swataah, Norwegian seg self, Russian sebja and svoj(u),14 and Spanish se. Anti-subject orientation is exemplified by the weak pronoun form 6 in Yoruba (Kwa, Niger-Congo). This pronoun can be unbound, or bound to a non-subject; it must not, however, be anaphorically related to any commanding subject:15 Unless noted otherwise, the data and generalizations in this section and the next are taken from Tang (1989) (Chinese); Toman (1992) (Czech); Vikner (1985) (Danish); En? (1991) (Dogrib, Turkish); Koster (1984) (Dutch); van Steenbergen (1992) (Finnish); Pica (1983) (French); Ev-eraert (1992) (Frisian); Dalrymple (1993):157 (Fula); Iatridou (1986) (Greek); Mohanan (1990) and Yang (1983) (Hindi); E. Kiss (1992) (Hungarian); Koster and Reuland (1992b) (Icelandic, Italian); Katada (1991) (Japanese); Bhat (1978) (Kannada); Benedicto (1992) (Latin); Yang (1983) and Mohanan (1982) (Malayalam); Dalrymple (1993) (Marathi, Norwegian); Reinders-Machowska (1992) (Polish); Avrutin (1994) (Russian); Fontana and Moore (1992) (Spanish); and Pulleyblank (1986) (Yoruba). See also Koster and Reuland (1992b). Latin and Malayalam according to Yang (1983). Rappaport (1986). Cf. Pulleyblank (1986) and Pulleyblank (1990); data from Dalrymple (1993), (1.107) and (1.108). Pulleyblank (1990):987 actually provides an example in which 6, contained in an adverbial clause, is bound to the matrix subject. The purported generalization is that the anti-subject orientation only holds for pronouns embedded within complement clauses. Further investigation of this is beyond the scope of this book. 60 DOMAINS AND ORIENTATION (3.34) (a) Seguni so pe Tunde2 ro pe 03/*i/*2 sanra. (Yoruba) Segun say that Tunde think that he fat 'Seguni said that Tunde2 thought that he3/*i/*2 was fat.' (b) Tolui so fun Segun2 pe Dupe3 ro pe 62/* 17*3/4 sanra Tolu say to Segun that Dupe think that he fat 'Tolui told Segun2 that Dupe3 thought that he2/*i/*3/4 was fat.' Anti-subject orientation - sometimes within a delimited domain - is also found in Danish ham (selv), Dogrib ye, Kannada ova, Norwegian ham (self), and Russian ego. In the remainder, we will assume that the notion of 'subject' is available to us in formulating Binding Conditions. In theories in which this notion is not a primitive, it will have to be considered as a shorthand for some complex derived notion. Binding Conditions will then take the following form, with an extra parameter added: (3.35) An NP of class must (not) be coindexed with a commanding NP 1 Subject j within its domain. 3.2.2 Logophoricity 1 In the previous subsection, we have seen pronouns which restrict the class of their possible antecedents by reference to grammatical function (subject vs. non-subject). In this subsection we turn to pronouns, so-called logophors, which are oriented towards a semantically or pragmatically determined class of antecedents. Various languages have a special set of pronouns used to refer to the 'source' of an embedded statement. Consider the following examples from Ewe (Kwa, Niger-Kordofanian):16 (3.36) (a) kofii be yei/*2/*s -dzo Kofi say LOG left (b) kofii be e*i/2/*s -dzo Kofi say he left (c) kofii be me*i/*2/s -dzo Kofi say I left 'Kofi said that he/I left.' (Ewe) The distribution of e, 'he,' and me, 'I,' is more or less as expected: e refers to a non-speaker, non-addressee person (hence the subscript *s, for 'not the speaker'), while me must be the speaker; moreover, e, for reasons of no concern here, cannot anaphorically relate to the matrix subject. What is of interest here is the additional pronoun ye, which can only refer to the subject of be, 'say,' not 16 From Clements (1975) by way of Sells (1987):448 and Kuno (1987): 146. 3.2 Orientation 61 to any other person, speaker or not. Pronouns like that are called logophoric, glossed Log, following the terminology proposed in Hagege (1974). Logophoric pronouns are not restricted to verbs of saying. As the following Ewe examples show, the subject of be happy, know, or see in Ewe can antecede logophoric pronouns; similarly, in Tuburi, the experiencer of a psychological verb can bind a logophor:17 (3.37) (a) anai kpo dyidzo be yei/*2 -dyi vi (Ewe) Ana see happiness COMP LOG -bear child 'Anai was happy that shei/^ bore a child.' (b) kofi7 (me-) nya be me -kpo (o) Kofi (not) know COMP / see LOG 'Kofi7 knew/didn't know that I had seen himy/^.' (c) kofi7 kpo be yewo7+2/*2 -do go Kofi see COMP LOG-PL -come out 'Kofi saw that they (including Kofi) had come out.' (3.38) hl:nl d35 nei ga S£i/*2 II? tjlgi (Tuburi) fear make him COMP LOG fall illness 'He was afraid that he would fall ill.' Before going on, let us briefly consider alternative characterizations of these pronouns. Couldn't they just be reflexives that need to be bound by a higher subject? After all, the 'source of an embedded proposition' often is (the referent of the) subject of the verb embedding (the sentence expressing) that proposition. Indeed, logophoricity and long-distance subject orientation aren't always easy to tell apart. There are two straightforward ways to distinguish them, though; consider the following examples involving the Japanese logophoric pronoun zibun:1^ (3.39) (a) Takasii wa Taroo2 ni [Yosiko ga zibuni/*2 o nikundeiru Takasi TOP Taroo DAT Yosiko NOM self ACC be-hating koto] o hanasita (Japanese) COMP ACC told 'Takasi told Taroo that Yosiko hated him (Takasi).' (b) Taroo2 wa Takasii kara [Yosiko ga zibuni/*2 to nikundeiru Taroo TOP Takasi from Yosiko NOM self ACC be-hating to] kiita COMP heard 'Taroo heard from Takasi that Yosiko hated him (Takasi).' The only acceptable antecedent for zibun in both sentences in (3.39) is Takasi. Takasi is the subject and topic in (3.39a), but an oblique in (3.39b). If zibun were subject oriented, we would expect it to refer to Taroo in (3.39b), which has the same grammatical function and morphological marking as Takasi in (3.39a). It is clear from these examples that zibun is not subject oriented. Rather, it takes the Ewe data from Sells (1987):449; Tuburi from Hagege (1974). From Sells (1987):453f. 62 DOMAINS AND ORIENTATION source of the embedded proposition as its antecedent, which in both sentences in (3.39) is Takasi. A second hallmark of logophoric pronouns is that they can sometimes occur without a sentence internal antecedent at all. (3.40) illustrates this with an example from Icelandic, involving ser, the dative of the logophoric pronoun sig:19 (3.40) Forma5urinni var5 oskaplega rei5ur. Tillagan vseri avfvir5ileg. the chairman became furiously angry the proposal was-SVBJ outrageous Vseri henni beint gegn seri personulega? (Icelandic) was-SVBJ it aimed against self personally 'The chairman became furiously angry. The proposal was outrageous. Was it aimed at him personally?' Despite the sentence boundary, the second and third sentence are clearly reporting the chairman's thoughts (note also the subjunctive marking on the verbs in these sentences). As a consequence, the logophoric ser can be used here without any sentence internal binder. This again shows that we are not dealing with a reflexive that needs to be syntactically bound, however involved its syntactic binding domain. Having demonstrated the difference between logophoric and (subject-oriented) reflexive binding, note secondly that logophoric pronouns are different from English non-reflexives, too. They do require an antecedent, and moreover one with a special pragmatic property. While one could argue that pronouns need (discourse) antecedents as well, the difference is clear. Contrast the Ewe example (3.36) above and a random English sentence like Mary said that she left. Out of the blue, the latter might tempt speakers strongly towards an interpretation in which she is bound by Mary, but a disjoint reading, in which she is discourse related, is clearly available. This is very different from the case of a logophor, which absolutely needs to take the source of information as its antecedent. To get a better intuition about what counts as the 'source of information,' note the following paraphrases for the examples above: (3.41) (a) Kofi said: "/left." (b) Ana was happy thinking: "/ am bearing a child." (c) Kofi knew/didn't know: "X has seen me." (where X is the speaker of the sentence) (d) Kofi saw (something that triggered the mental representation): "We have come out." (e) He was afraid (of being in a state reflected by the mental representation): "/am falling ill." (f) Takasi told Taroo: "Yosiko hates me\" (g) Taroo heard from Takasi: "Yosiko hates me\" (h) The chairman, furiously angry, thinks: "The proposal is outrageous. Is it aimed at me personally?!" 19 From SigurSsson (1986), via Sells (1987):453. 3.2 Orientation 63 In all these cases, I have replaced the embedded clause by a direct quotation. Accordingly, the logophoric pronouns of the original sentences are replaced by first person pronouns (in italics). We formulate this as our rule of thumb for logophoric pronouns: (3.42) A logophoric pronoun can be used if it is embedded in a constituent c such that (i) c is embedded, (ii) c denotes a proposition p, which (iii) can be paraphrased as a mental state or reported utterance of the pronoun's antecedent such that the paraphrase contains a first person pronoun in place of the pronoun. What exactly qualifies as a logophoric antecedent, however, varies from language to language. Usually some lexical element indicates the presence of a 'logophoric environment,' e.g. a verb of saying, thinking, etc., or a special embedding complementizer. Further conditions may obtain. For example, in Ewe, only sentence embedding verbs license logophors, while verbs like remember, which selects an NP complement, or hear, with an ECM-type complement, do not (cf. [3.42ii]):20 (3.43) (a) * amai do rjku nyonuvi hi dze yei gbo dyi. (Ewe) Ama set eye girl rel stay log side on 'Ama remembered the girl who stayed with her.' (b) amai gbb be ye -do rjku nyonuvi hi dze yei gbo dyi. Ama say comp log -set eye girl rel stay log side on 'Ama said that she remembered the girl who stayed with her.' (3.44) (a) * Ama remembered: "The girl who stayed with me." (not propositional) (b) Ama said: "I remember the girl who stayed with me." Logophoricity is attested in many languages of the world. It is important to keep the option of logophoricity in mind when attempting to describe Binding Conditions in a given language, precisely because it can so easily be mistaken for something else, e.g. long-distance subject-oriented anaphors. 3.2.3 Towards a formal treatment of logophoricity i Sells (1987) provides a formal implementation of logophoricity within the framework of Discourse Representation Theory (Kamp and Reyle, 1993). We will follow the gist, though not the letter, of his proposal in what follows; many details, however, will be omitted. We take as our role model the treatment of first and second person pronouns from chapter 2, section 2.2.2: (3.45) (a) P/me/my/myselfn]g's'u = g(n) if g(n) = s, undefined otherwise (b) |[you/your„]^'^'" = g(n), if g(n) is the person s addresses in u Cf. Sells (1987):449f. 64 DOMAINS AND ORIENTATION Extending this treatment, we introduce another contextual parameter, which we call o for 'origo' (Latin for 'source', given that s is already in use). A logophoric pronoun will always refer to the individual o: (3.46) [pronoun^p'*'"'0 = o, if o = g(n) Note that we assume here that logophors, just like first and second person pronouns, are indexed and that their lexical content is just a presupposition. The origo-parameter can be shifted by verbs of saying, thinking, etc. to the sayer, thinker, etc. (all other aspects of these verbs' meanings are simplified as far as possible): (3.47) (a) [say (that) SJg's'u'° = Xx.x says something which entails lSJg's'u'x (b) [hear from NP (that) Sp'*'"'° = Xx.x hears y,y = [NPp'*'"'°, says something which entails [S]-^'"^ (c) [believe (that) SJg's'u'° = Xx. what x believes entails [Sp'*'"'x (d) [S frightens NPp'*'"'° = 1 iff x, x = [NPp'*'"'°, prefers a state of affairs in which [S]^'"'* is false to one in which it is true Note, in contrast, that the speaker parameter (just like the utterance-situation parameter) cannot be shifted, except for direct quotation. This can be seen, for example, from the fact that, unlike in a direct quotation, English first person pronouns cannot be used to refer to the speaker of an embedded sentence:21 (3.48) (a) *Gil6 said that 16 was happy. (b) Gihj said: "Ig am happy." (c) Gil6 said that she6 was happy. Rather, we appear to see ordinary coreference in these cases: she corefers with Gil, who happens to be the source or speaker of the embedded sentence. This is no different from the cases in (3.49), which don't involve reported speech or attitudes in the embedded clauses: (3.49) (a) Gil6 met a pilot who liked her6. (b) Gil6 arrived even though Roger had said that she6 wouldn't. According to the treatment in (3.47), re-setting the logophoric center is a lexical property of verbs like say, believe, etc. Other sentence-embedding verbs do not have this option, just as little as, say, sentence-embedding prepositions: (3.50) (a) [look as if SJg,s,u,° = Xx. the visual appearance of x makes it likely that lSJg's'u'° is true (not: ... that [Sp'*'"'x is true!) (b) [S unless S'P'*'"'° = 1 iff [Sp'*'"'° is true and will be as long as [S'P'*'"'° is It is rather clear why unless cannot, in principle, establish a new logophoric center: there is no individual argument to the function denoted by unless, therefore 21 The classical philosophical reference here is Kaplan (1977); see Schlenker (1999) for recent discussion. 3.3 Binding domains cross-linguistically 65 a fortiori none that could serve as the new origo. This is different for look like, which does have an individual argument. We could thus give it a meaning as indicated in the parentheses in (3.50a), which would wrongly shift the origo to the person whose looks are described. The intuition here is, of course, that say and think, but not look like, involve the report of an utterance or thought, and thus only they have a source to come along with it. This, however, is expressed nowhere in our formal treatment, and it is not easy to see how it could be.22 Exploring this issue further would be beyond the scope of this book. We will leave our formalization at this. Sells (1987) argues that we need in fact not just one origo parameter but three different ones, which he calls 'source,' 'self,' and 'pivot' (plus, of course, the familiar 'external speaker'). As these present nothing fundamentally new, I refer to Sells' work for further details. There are a number of essential points we leave unaddressed or unresolved: first, the formal counterpart of the 'origo' intuition (i.e. the point just mentioned above); second, cases such as (3.40) above, in which the re-set origo parameter extends beyond the scope of the embedding element (these should follow from a general treatment of modal subordination, cf. Roberts (1987, 1989, 1996)); and third, cases in which the origo is not introduced by a specific lexical element, but nonetheless seems to represent a 'source' in the intuitive sense (Sells' third person point of view cases; see Sells [1987] for details). 3.3 Binding domains cross-linguistically In section 3.1 we defined the notion of governing category and coar-gument domain, which could serve to identify the binding domain for reflexive (must be bound within) and non-reflexive (mustn't be bound within) pronouns, respectively. Other languages make reference to different domains. It seems, however, that the number of binding domains cross-linguistically might still be rather limited. For the purpose of the presentation to follow, I will focus on four different domains, listed in table 3.1.23 It is important to realize that these domains are collapsed in a great number of structures. In a simple transitive clause like (3.51), all four domains for the object NP are the same, namely the root clause: It may be tempting to try to give the logophoric pronouns a presupposition similar to the first and second person pronouns, e.g. [pronoun^]g,s'"'° = o, presupposition: o is the source of P. But the question here is: what is PI Unlike u in the definitions above, P is not a contextual parameter. We want it to refer to the reported utterance or thought, but where should this come from? These domains correspond, from bottom to top, to the domains Root S, Minimal Finite Domain, Minimal Complete Nucleus, and Coargument Domain in Dalrymple (1993); and to Domains 3, 2, and 1 in Koster and Reuland (1992b) (who deny the relevance of the C-Domain, a point we will return to later). Interestingly, it is noted in both works that these domains seem to be the only ones required for their quite comprehensive, though not typologically representative, sample of languages. 66 DOMAINS AND ORIENTATION Table 3.1 Cross-linguistic binding domains Domain Definition the... for A is the smallest constituent containing A, A's case assigner C and... Coargument domain, CD all arguments of C Subject domain, SD a Subject (within NP or S) Tense domain, TD a finite clause Root domain, RD the entire sentence (3.51) The Schindlers provoked the Flaunders. Below, on the other hand is a maximally complex example in which the four domains, relative to the NP Kim, are different: It was reported that they left upon Joe's telling jokes about Kim cd SD TD RD To find out which domain is crucial for a given pronoun type thus requires constructing intricate and often somewhat involved sentences. A lot of studies, especially older ones, on Binding Theory in particular languages do not distinguish between the different categories, which is why our cross-linguistic knowledge about these distinctions is very incomplete and likely to be in need of revision. The root domain The root domain is significant in English only as a negative domain for full NPs. There are languages, however, in which certain pronouns must be bound in the root domain, though not necessarily locally. Examples of such pronouns mentioned in the literature are Chinese ziji, Fula Dum, Greek o idhios and ton idhios, Icelandic sig, Italian se andproprio, Japanese zibun,24 Kannada ta-nu, Latin se, Malayalam swa and tanne, Marathi aapan, and Yoruba 6. It should be noted that many of these are also subject oriented, and also that the data provided are often not sufficient to rule out the possibility that some of these are actually logophoric.25 For the remaining core cases, however, the pertinent Binding Condition will take the general form in (3.52), with illustrative examples given in (3.53) and (3.54):26 Also Sportiche (1986). In the same vein, it is not clear in all of these cases that the antecedent actually needs to c-command, or command at all for that matter, the pronoun (this is, for example, not the case in Chinese and Fula). From Iatridou (1986):769 and Yang (1983): 183, respectively; o idhios must also be locally free, hence it cannot be bound by Maria in (3.53). 3.3 Binding domains cross-linguistically 67 (3.52) (3.53) (3.54) o idhios/caki (ziji/zibun/swa ...) must be coindexed with a commanding NP within its root domain. O Yanisi ipe ston Costa2 [oti i Maria3 aghapa ton idhioi/2/*3/*4 ] the Y. said to-the C. that the M. loves himself (Greek) 'Yanis told Costa that Maria loves him.' -ka [Tom4 -iy -NOM T. -'s -n -ta -ko] -ASP -DEC -COMP Johni -in [B1II2 -i [Mary3 /. -TOP B. -NOM M. tsehan thseto] -lil silhaha toward attitude -ACC hate -ta -ko] mit -nin -ta. -DEC -COMP believe -ASP -DEC 'Johni believes that BUI2 thinks that Mary3 hates Torre's attitude toward selfi/2/3/4/*5-' cakii/2/3/4/*s-e self ssengkakha -n think -ASP (Korean) The Korean example aptly illustrates how a Binding Condition like (3.52) is to be read: caki can be bound within any domain up to the root domain, but not from outside of the root domain (not: in the root domain, but not in any smaller domain). So, generally, if a Binding Condition says that an element needs to be bound in domain D, this means that its antecedent has to be somewhere within D, not necessarily at the 'top' of D. Note, on the other hand, that certain elements may be subject to more than one Binding Condition. On top of (3.52), Greek o idhios, for example, also needs to be free in its tense domain (which is why it can't be coindexed with Maria in [3.53]), so that effectively its binding domain 'starts' at the next higher clause and 'ends' at the root node. The tense domain The tense domain functions as the binding domain, for example, for Czech selsebe and svuj,21 Danish sig, Finnish han, Use, and han Use, French soi, Marathi swataah (in the less restrictive version), Norwegian seg and sin, Polish siebie and swoj, and Russian sebja and svoj(u)?^ Tense-domain-bound reflexives might look like subject-domain-bound reflexives (i.e. like English) at first glance, cf. (3.56a) and (3.57a). What is distinctive, however, is that the pronoun can be bound across subjects in NPs and non-finite complements such as gerunds or infinitival clauses (i.e. the subject domain, marked by parentheses):29 (3.55) swataahlseg (siglsoilsiebieluskii...) must be coindexed with a commanding NP within its tense domain. According to Toman (1992): 154f., the exact domain seems to be 'inflected clause' (including infinitival clauses), as Czech reflexives cannot be bound across infinite sentence boundaries; they can be bound across small clause and NP subjects though, which is why I group them here. Also Rappaport (1986). Examples are Dalrymple's (1993) (1.49), (1.42), (1.43), (1.87), and (1.86). 68 domains and orientation (3.56) (a) * Tomi mhanat hota [td=sd ki Sue ni swataahlaai maarle] Tom said that Sue erg self hit (Marathi) 'Tom said that Sue hit himself.' (b) [td Jane2 laa [sd Tom ne swataaci2 pustake phekun dilyaace] Jane dat Tom erg self books throw kalle] learned 'Jane learned about Tom throwing away (her)self's books.' (c) [td Jane2 [sd John ne swataahlaa2 maarlyaavar] rusun] Jane John erg self hitting angry 'Jane remained angry upon John hitting (her)self.' (3.57) (a) * Joni var ikke klar over [td=sd at vi hadde snakket om Jon was not aware of that we had talked about segi/2] (Norwegian) self 'Jon was not aware that we had talked about him.' (b) Joni likte [sd din artikkel om segi] Jon liked your article about self 'Jon likes your article about him.' The subject domain Ignoring the non-complementary cases (i.e. those where a reflexive or reciprocal is contained in a subject), the subject domain is the relevant one for English reflexives. It furthermore appears to be the binding domain for reciprocals in all languages I could find information about on this issue, including Danish hinanden30 Dutch elkaar, English each other31 Finnish toiset, Hungarian egymas, Icelandic hvor annar, Italian I'uno, I'altro, Norwegian hverandre, the reciprocal use of Polish siebie, and, shown here, Russian drug druga:32 (3.58) (a) [sd Pisatelii čitali [vospominanija drug o drug writers-NOM read reminiscences-acc each about other -ej] (Russian) -loc 'The writers read reminiscences about each other.' (b) * Pisatelii čitali [sd vospominanija Tolstoj a drug writers-NOM read reminiscences-acc Tolstoi-GW each o drug -ej about other -loc 'The writers read the reminiscences of Tolstoj about each other.' 30 Pica (1983). 31 But see Lebeaux (1983). 32 Rappaport (1986) (18). Yang (1983) actually speculates the SD is the binding domain for reciprocals universally, but see Huang (2000):n. 46, pp. lOlf. for possible exceptions. 3.3 Binding domains cross-linguistically 69 Cases of (non-reciprocal) pronouns that need to be bound in the subject domain alongside English herselji'himself'are Chinese taziji,33 all Danish -self pronouns, Dutch -zelf pronouns, French se, Frisian 'm(sels), Greek ton,34 Hungarian maga, Icelandic sjdlfur forms, Italian si and se stesso, Japanese -zibun forms, Marathi swataah (in the restricted dialect), Norwegian -self forms and possessive sin, and Turkish kendi:35 (3.59) zichzelf (hvor hannaltoisetleach other/ton/si/swataah ...) must be coin-dexed with a commanding NP within its subject domain. (3.60) (a) Peteri zag [$d Mary2's foto van zichzelf2/*i/*3]. (Dutch) P. saw M'.s pictures of self 'Peter saw Mary's pictures of him/herself.' (b) Maryi liet [so Peter2 op zichzelf2/* i/* 3 schieten] M. let P. at self shoot 'Mary let Peter shoot at himself.' The subject domain is also significant as a negative domain. Dogrib ye, Finnish han, Greek ton (idhios), Italian lui, and Turkish o need to be free in their subject domain. Moreover, Danish hamlhende and ham selv, Finnish han (Use), Icelandic hann sjdlfur, Malayalam taan, Norwegian ham (self), Polish nim and possessive jej, Russian ego must be subject-free within the subject domain: (3.61) nim (jej/ham (selv)/han (itse)/ taan ...) must not be coindexed with a commanding subject within its subject domain. (3.62) Piotri czytal [sd Janka2 artykuio nini7i/*2 ] (Polish) P. read J's article about him 'Piotr read Janek's article about him.' The coargument domain Turning finally to the coargument domain, it appears that this domain is only relevant as a negative domain. For example, Marathi to at first glance behaves like a non-reflexive pronoun in that it cannot be bound in its local clause, but across a finite sentence boundary, cf. (3.63a/3.63b); however, unlike a non-reflexive pronoun in English, it can occur with an antecedent within its subject domain, as long as the antecedent is not a coargument, cf. (3.63c/3.63d):36 (3.63) (a) Janei ne tilaa2/*i bockaarle. (Marathi) Jane ERG her-acc scratched 'Jane scratched her.' (b) Mary2 dukhi hoti. tilaa2 jaataa aale naahi. Mary sad was she-dat go could not 'Mary was sad. She could not go.' 33 Koster and Reuland (1992b); Tang (1989). 34 Also En? (1989). 35 Examples are Koster's (1984) (45) and (23). 36 Examples are Dalrymple's (1993) (1.63), (1.64), (1.60), and (1.61). 70 DOMAINS AND ORIENTATION (c) Janei ne ticyaakartaai saadi ghet li. Jane ERG her for sari bought 'Jane bought a sari for her (Jane).' (d) Janei ne John laa ticyaabaddali maathiti dili. Jane ERG John DAT her about information gave 'Jane gave John information about her (Jane).' Danish ham and hende, Dutch zich, Frisian 'm, Kannada ova-, Marathi to and aapan, and Norwegian seg all need to be free in their coargument domain: (3.64) to (hende/zich/ava-...) must not be coindexed with a commanding NP within its coargument domain. It is worth emphasizing that all these elements can be 'locally' bound (e.g. by a non-coargument subject within a finite clause), as in (3.63c/3.63d), and indeed some of them actually must be bound within the next higher tense domain (Norwegian), or subject domain (Frisian, Dutch, Danish). These pronouns have thus a very limited distribution, essentially as arguments to prepositions and inherent-reflexives (cf. chapter 1, section 1.4). In chapter 11 we will explore a possible rationale behind the peculiarities of this domain, namely that all elements that must be free in the coargument domain are simply incapable of marking a predicate as reflexive. This concludes our cross-linguistic discussion of domains. It is worth stressing that the choice of binding domain is not made once per language (say, tense domain for Marathi and Norwegian, subject domain for English), or even twice (say subject domain as negative domain for non-reflexives, and first clause with an accessible subject as positive domain for reflexives, as is sometime suggested for English). Rather, the choice of domain appears to be morpheme-specific, where within the same language different forms can have different binding domains (for example, tense domain as positive domain for Norwegian seg, but subject domain for ham selv and the reciprocal hveandre in the same language). This challenges attractive ideas about setting a single 'domain parameter,' and leaves us with what appears to be a less restrictive approach to the acquisition of binding domains (cf. Manzini and Wexler [1987]; and Wexler and Manzini [1987]).37 Exercise 3.3 Indicate all four domains relative to Y in the following schematized structures. Provide labeled bracketings for clarity: 37 There is, of course, the possibility that something about the form of a given pronoun, coupled with a universal theory of binding domains, yields the different domains (Dechaine and Wilschko [2000]; Reuland [2001, a.o.]). There are, however, to the best of my knowledge no cross-linguistically valid generalizations about form-domain correspondences that could serve as the basis for such a theory, and, accordingly, such approaches often remain stipulative. 3.3 Binding domains cross-linguistically 71 (3.65) (a) X loves Y (b) Z says that X loves Y (c) Z left upon X's hitting Y (d) X has information about Y Exercise 3.4 (i) Formulate, as far as the following data warrant, the positive and negative binding condition for Greek ton eafton tou and o idhios (cf. also [3.53]):38 (3.66) (a) O Yanisi aghapa ton eafton toui/^/ *ton idhioi/2-the Y. loves PRON (Greek) 'Yanis loves himself.' (b) O Yanisi theori oti o idhiosi/^ ine o kaliteros the Y. thinks COMP PRON is the best ipopsifios. candidate 'Yanis thinks that he is the best candidate.' (c) O Yanisi theli o Costas2 na voithisi ton idhioi/^/ton the Y. wants the C. helps PRON eafton tou2/*i. 'Yanis wants Costas to help him(self).' (d) O Yanis i theori ton eafton tou \/#2 ton kaliero the Y. considers PRON the best ipopsifio. candidate 'Yanis considers himself the best candidate.' (ii) Would you expect that ton eafton tou in (3.66d) can be replaced by o idhiosl Explain! Exercise 3.5 Complete, as precisely as the data allow, the following Binding Condition for the Hindi pronoun uskii:39 (3.67) uskii must_be coindexed with a commanding_in its Data from Iatridou (1986). Data from Mohanan (1990), via Dalrymple (1993):37f. 72 domains and orientation (3.68) (a) raviii uskii^i saikil -par baithaa (Hindi) Ravi his bicycle -LOC sit-PEKF 'Ravi sat on his bike.' (b) vijay -nei ravii -ko2 uskii2/*i saikil -par bithaayaa Vijay -erg Ravi -ACC his bicycle -LOC mY-CAUSE-perf 'Vijay seated Ravi on his (Ravi's) bike.' (c) raajaa-nei kahaa ki mantr^ uskei/^ ghar gayaa king -ERG say-PEKF that minister his house-hOC go-PERF 'The king said that the minister went to his (the king's) house.' (d) raajaa -nei mantrii -ko2 uske*i/*2 ghar jaanee -kii king -erg minister-ACC his hous-hOC go nonfin-gen aagyaa dii order give 'The king ordered the minister to go to his (someone else's) house.' 3.4 Long-distance reflexives (LDRs) There are many languages in which reflexives lead a 'double life': they can be locally bound, similar to English herself; or they can find an antecedent outside their minimal clause. In the latter case, that antecedent often needs to be a subject and/or a logophoric center. Such languages include Latin, Icelandic, and Japanese. We will follow the custom in the literature in this section and refer to such pronouns as long-distance reflexives (LDRs). For example, the Latin reflexive se (acc.)/sibi (dat.) can be bound to a non-local antecedent if that antecedent is a verb of saying, cf. (3.69a); we do not of course have un-grammatical examples, but note that in embedded clauses that are not embedded under verbs of saying, only plain pronominal forms are found, cf. (3.69b):40 (3.69) (a) Iccius ntintium mittit, nisi subsidium sibiy submittatur... Iccius message sends if-not relief refl furnished-PASSlVE (Latin) 'Iccius sends a message that unless relief be given to himself,...' (b) Ibi in proximiis villis ita bipartito fuerunt(i), ut there in nearest farmhouse so in two parts made-they that Tiberis inter eosi et pons interesset. Tiber between them and bridge lay between 'They set (themselves) up in farmhouses very nearby, divided in two, so that the Tiber and the bridge were in between them.' There are two general lines of analysis for LDRs: movement analyses, according to which LDRs are locally bound after covert movement into the local clause of their antecedent; and non-movement analyses. Among the latter some analyze 40 Examples from Kuno (1987):137. 3.4 Long-distance reflexives 73 LDRs as command anaphors bound within a certain (rather large) syntactic domain (e.g. Progovac (1992); Manzini and Wexler (1987)), while most claim that LDRs fall outside the domain of sentence grammar and are purely a matter of logophoricity (e.g. Hellan (1992); Kameyama (1984); Kuno (1987); Maling (1984); Thrainsson (1992)).41 We do not intend to review the rather extensive literature on the issue here (see e.g. Huang (2000):ch. 2.3 for a critical overview). Generally, proponents of movement accounts claim to offer a unified analysis of the short-distance and long-distance binding cases of LDRs, and stress the parallelism to languages with clitic climbing (it is noteworthy that LDRs are overwhelmingly prosodi-cally weak, monosyllabic forms; see below). Furthermore, these analyses often capture specific restrictions on LDRs connected to things like intervening subjects or mood. Proponents of non-movement accounts, on the other hand, often point out counter-examples to the restrictions dealt with on the movement analyses. Also, they argue that the existence of logophoric pronouns needs to be acknowledged anyway, and that LDRs can sometimes be found without any sentence-internal antecedent, which is, of course, typical for logophoric pronouns, but finds no natural account under movement approaches. Furthermore they point out that the purported movement of LDRs often needs to violate established restrictions on movement. It is probably fair to say that the problems for movement accounts are considerable and severe, but that logophoric approaches are only as restrictive as their underlying theory of logophoricity, an area where more work is required. Maling (1984), Sells (1987), and Thrainsson (1992), among others, make a convincing case that Icelandic LDRs are logophoric in nature. Their data also provide some nice examples of what does and what doesn't count as a logophoric antecedent. To emphasize this latter point, I have provided the direct speech paraphrases below the examples:42 (3.70) (a) Hanri2 sag5i [ a5 sig2 vanta5i hsefileika ]. (Icelandic) he said that self lacked ability 'He said that he lacked ability.' - 'He said: "I lack ability.'" (b) * Honuni2 var sagt [ a5 sig2 vanta5i hsefileika ]. he was said that self lacked ability 'He was told that he lacked ability.' - 'He was told: "You/# I lack abil-fry."' Among the movement analyses, we can furthermore distinguish between analyses that assume LDRs to move as heads (essentially parallel to clitics in clitic-climbing languages; cf. Cole et al. [1990]; Cole and Sung (1994); Cole and Wang (1996); Cole et al. [2000]; Hestvik (1992); Pica (1983, 1987)), and those that assume them to move as phrases (Huang and Tang (1992)), as well as mixed ones (e.g. Katada [1991]). (3.70) is Maling (1984)'s (37a/b), (3.71), and (3.72) are from SigurSsson (1986), both via Sells (1987):450ff. 74 DOMAINS AND ORIENTATION (3.71) (a) BarniSi lét ekki í ljós [ a5 pa5 hefSi veriS hugsaS vel child-thei let not in light that there had been though well um sigi] about self 'The child didn't reveal that it had been taken good care of.' 'The child didn't say: "I've been taken good care of" (b) * BarniSi bar bess ekki merki [ a5 pa5 hefSi veriS hugsaS child-thei bore it not signs that there had been though vel um sigi ] well about self 'The child didn't look as if it had been taken good care of #'The child didn't look: "I've been taken good care of" As mentioned before, logophoric pronouns can, under the right circumstances, appear without a sentence-internal antecedent at all. This is the case, too, for Icelandic LDRs: (3.72) Forma5urinri4 var5 óskaplega reiSur. Tillagan vseri avivirSileg. The chairman became furiously angry. The proposal was outrageous. Vseri henni beint gegn sér4 persónulega. (Icelandic) Was it aimed at self personally? Because LDRs usually occur in subjunctive, rather than indicative, subordinated sentences (where such a distinction exists), it has been suggested that their binding domain could be syntactically described as 'first dominating clause with independent (=indicative) inflection' (plus, possibly, subject-orientation). While this might be the case for some languages, Thráinsson (1992) shows that subjunctive mood is neither a necessary nor a sufficient condition for LDRs in Icelandic. Rather, the connection appears to be an indirect one: subjunctive mood marks certain embedded contexts (e.g. non-factual ones), which, in turn, coincide to a large degree with those created by verbs of saying and thinking. A striking generalization about LDRs is that they are morphologically simple. Languages that have only complex reflexives (like English) systematically lack LDRs, and in those that have simple and complex forms (e.g. Icelandic) only the simple ones are found to be LDRs. In some movement approaches to LDRs, this remarkable property has been taken to correlate with the head/phrase distinction: complex forms like himself are syntactically branching and thus inherently phrasal, confined to phrasal movement, while simple forms like Icelandic sig can act as heads and undergo head-movement (Pica, 1983, 1984). The success of such a story partly relies on how plausible it is to assume that head-movement is less local than phrasal movement (the natural expectation might be that it is the other way around), an issue we won't go into here. Non-movement accounts, on the other hand, have little to offer in the way of explaining the general morphological simplicity of LDRs (either) (see Hellan [1992] :29 for a few speculative remarks). 3.5 Some pronominal systems 75 Exercise 3.6 Explain the grammaticality contrast between the two Icelandic examples in (3.73) below (Maling [1984]'s [29a/b]) by paraphrasing them as direct speech: (3.73) (a) Jöri4 truir [aö hanri4 veröi alltaf froskur [nema John believes [that he will be forever a frog [unless konungsdöttir kysse sig4]] a kings daughter kisses self]] 'John believes that he will remain a frog forever unless a princess kisses him.' (b) * pvi er truaö [aö hanru; veröi alltaf froskur [nema it is believed [that he will be forever a frog [unless konungsdöttir kyssi sig4]] a king's daughter kisses self]] 'It is believed that he will remain a frog forever unless a princess kisses him.' 3.5 Some pronominal systems Now that we've seen a broader array of possibilities for domains and orientation, we can have a glance at some complete pronominal systems different from English. 3.5.1 Danish and Norwegian i Many Germanic languages other than English show a four-way split in their pronominal system, cf. table 3.2 (the labels SE- and P-form are taken over from the literature, reminiscent of the French reflexive clitic se and the generative term 'pronominal'). Table 3.2 Germanic pronoun systems SE-form: P-form: Danish sig ham bare Dutch zieh hem German sich ihn Norwegian seg ham Danish sig selv ham selv +'self Dutch zieh self hem self German sich selbst ihn selbst Norwegian seg selv ham selv 76 DOMAINS AND ORIENTATION Table 3.3 Danish and Norwegian pronoun system SE-form P-form bound to subject free from subject in tense domain in coargument domain bare D: sig D: ham, hende free in c-dom. N: seg N: ham +'self D: sig selv, hende selv D: ham selv bound in s-dom. N: seg selv N: ham selv Typically, the plain SE-form is found in all non-referring constructions, e.g. inherent reflexives and detransitivized forms such as middles. When referring, SE-forms tend to be locally subject oriented, while P-forms are often locally anti-subject oriented. Bare forms (SE- or P-) are usually locally free, whereas 'self forms must be locally bound. I will illustrate this using Danish and Norwegian, which provide particularly neat illustrations of such a system. As suggested in table 3.2, their pronominal systems can be thought of as arranged by two binary choices: SE-form vs. non-reflexive pronoun; and bare form vs. 'self form. Each value for these choices is associated with one binding requirement. As the choices cross-classify, so do the conditions, as shown in table 3.3.43 The examples given below illustrate these systems.44 (3.74) illustrates the simple case in which tense, subject, and coargument domain coincide. No bare forms can be used, given that these need to be free in the coargument domain, which includes all NPs here. Among the se/v-forms, siglseg selv must be chosen if the antecedent is a subject, but hende/ham selv, if it is a non-subject: * hendei/ *hende selvi/*sigi/ sig selvi] *hende2/ hende selv2/*sig2/*sig selv2j (3.74) (a) Susani fortalte Anne2 om Susan told Anne about (Danish) *hami/ *ham selvi/*segi/ seg selvi] (b) Harald i fortalde Jon2 om Harald told John about (Norwegian) *haixi2/ ham selv2/*seg2/*seg selv2j Note that both 'self and SE correspond to 'bound in domain D,' while both 'bare' and P correspond to 'free in domain D.' It is tempting to think that +/—SE corresponds to '+/— bound to subject in domain Dj,' and that +/—self corresponds to '+/— bound (at all) in domain D2.' But note that neither Dj nor D2 are the same across positive and negative conditions. Therefore, the formal similarity between the conditions associated with the opposing values cannot be captured in any straightforward way. I did not differentiate between tense domain and coargument domain for the sake of simplicity; cf. Dalrymple (1993) for evidence for the choice of the latter. Danish examples from Vikner (1985). 3.5 Some pronominal systems 77 Let us now turn to a case where coargument domain and tense domain are distinct. Since bad, 'asked,' is an object control verb, the understood subject of the embedded clause in (3.75) is Anne, not Susan. Therefore, coreference with Susan is a case of binding outside the coargument domain (and subject domain), but within the tense domain. In such a case, the se/v-forms are excluded because they require an antecedent within the subject domain. The bare SE-form sig/seg is possible, because it is free in its coargument domain and at the same time bound to a subject within the tense domain; the bare P-form hendelham is possible, too, because it is (subject-)free in the smaller coargument domain: (3.75) (a) Susani bad Anne om at ringe til hendei/ *hende selvi/ Susan asked Anne for to ring to sigi/ *sig selvi (Danish) 'Susan asked Anne to call her.' (b) Joni bad oss snakke om hami/ *ham selvi/ segi/ *seg Jon asked us to talk about selvi (Norwegian) As expected, sig/seg will no longer be available if the binding is to a non-subject (note that the understood subject of ringe is Susan, i.e. the pronoun is indeed bound by the matrix object): (3.76) Susani lovede Anne2 at ringe til hende^/ *hende selv2/*sig2/ Susan promised Anne to ring to *sig selv2 (Danish) Note finally that the bare SE-forms (and only those) are used in non-thematic positions such as with inherently reflexive verbs:45 (3.77) Peter sov over * ham/ *ham selv/ sig/ *sig selv (Danish) Peter slept over 'Peter overslept.' This contrasts with the reflexive object in true transitive constructions, which, as seen in (3.74) above, cannot be bare SE. It looks as if the subject qualifies as a binder in the sense of the SE-vs.-P distinction (otherwise we would expect to see ham), but not in the sense of the bare-vs.-'self' distinction (otherwise we would expect sig selv). We will return to some of these issues in chapter 11. 45 As Vikner (1985):8f. notes, certain verbs that allow either thematic or non-thematic objects give the misleading impression that sig and sig selv have a similar distribution, e.g. in (i): (i) Peter vaskede sig / sig selv. (Danish) Peter washed It turns out, however, that vaskede simply has two argument frames, one transitive and one inherently reflexive. Thus Peter vaskede sig selv corresponds to English Peter washed himself, while Peter vaskede sig corresponds to English Peter washed. As Vikner shows convincingly, the non-complementarity disappears as soon as one uses verbs that are either always transitive (sig selv), or always inherently reflexive (sig). 78 DOMAINS AND ORIENTATION Table 3.4 Marathi pronoun system bound to free in swataah subject in subject-domain — aapan (logical) subject in root-domain coargument-domain to — coargument-domain 3.5.2 Marathi i Marathi, as discussed in Dalrymple (1993), has three different pronoun forms, two of which require a binder; cf. table 3.4.46 As discussed in section 3.3 above, to, like English her/him, must be locally free, but unlike her/him (and like bare pronouns in Danish and Norwegian) only in its coargument domain (as opposed to subject domain). Swataah plays the role of a reflexive pronoun, except that its domain is slightly larger than that of English, Danish, or Norwegian 'self'-forms (tense domain in Marathi vs. subject domain in the Germanic languages), and it is subject oriented. In addition there is the long-distance reflexive aapan, which must be locally free, but bound within the root S;47 it can thus occur in embedded clauses, within NPs, and as the object of prepositions that assign their own #-role:48 (3.78) (a) Tomi mhanat hota ki Sue2 ni aaplyaalaai/!|<2/*3 maarle (Marathi) Tom said that Sue ERG self-ACC hit 'Tom said that Sue hit him.' (b) Janei ni aaplyei pustak phekun dile Jane ERG self-GW book threw give 'Jane threw away self's book.' (c) Janei ne aaplyaakartaai saadi ghet li Jane ERG self-for sari bought 'Jane bought a sari for herself.' Note once again that the distribution of aapan significantly overlaps with that of swataah (for all positions whose tense domain is bigger than the coargument domain), and is even a proper subset of that of to. This kind of non-complementarity is common among languages: the Turkish pronoun kendi is a domain reflexive -it needs to be bound within the root-domain. But Turkish also has a second form kendisi, which obeys no Binding Conditions at all. The distribution of kendi is thus a proper subset of that of kendisi:49 The line for swataah describes what Dalrymple calls the 'restrictive' dialect of Marathi. For speakers of the less-restricted dialect, the relevant domain is the tense domain. Dalrymple (1993):21-24 convincingly shows that aapan is not a logophor. Dalrymple (1993)'s (1.49), (1.35), (1.31). Keenan (1988):134, following En? (1989). 3.5 Some pronominal systems 79 (3.79) Herkesi ayna-da everyone mirror-hOC kendisi-(n)ii/2 kendi-(n)ii/*2 (him)self-ACC gördu (Turkish) saw 'Everyone saw themselves in the mirror.' Even more extreme, Fijian (Oceanic, Austronesian) has no reflexives at all. The Fijian pronoun koya occurs in all kinds of configurations, from locally bound, to non-locally bound, to free:50 (3.80) a mokuti koyai/2 oirakecei (Fijian) PAST hit him(self) PL all 'They all hit themselves.' These are but a few illustrative examples of pronominal systems found in the languages of the world. Our general schema for Binding Conditions at this point looks as in (3.81): (3.81) An NP of class must (not) be coindexed with a commanding coargument subject tense root NP 1 .,. . . . > withm its subject J domain. A class can be just one lexical item (with all its different case, person, and number forms, e.g. Marathi aapan, Turkish kendisi), or a set of (all forms of) stems that have a certain morphological shape in common ('self vs. bare, SE-based vs. P-based in Germanic), or, of course, their complement (all non-pronominal NPs). Any one of the parameters in (3.81) can serve to define natural superclasses of forms (e.g. all that need to be bound vs. all that need to be free, only the subject-oriented, etc.). Different languages distinguish different numbers of classes, and whether there is a minimal or a maximal number of classes, whether there are any impli-cational relations between the different classes cross-linguistically, and whether any super-classes (e.g. Chomsky [1981] 's anaphors versus pronominals) play a cross-linguistically privileged role, remains yet to be found out. What should have become clear, however, is that it is not useful to ask which elements in a given language are the counterparts of, say, reflexive and non-reflexive pronouns in English, or what the binding domain (or governing category) in a given language is. Exercise 3.7 Define the following classes using the schema in (3.81): 1. the class of domain reflexives, i.e. all pronouns that must be syntactically bound at all 2. the class of logophors 3. the class of bare SE anaphors in Danish and Norwegian 50 Keenan (1988): 132. 80 DOMAINS AND ORIENTATION Exercise 3.8 Using table 3.4, predict which of the three pronouns swataah, aapan, and to can occur in the position of PRONOUN in the Marathi sentence frames in (3.82a) and (3.82b) (Dalrymple [1993] [1.60/61]): (3.82) (a) Janei ne PRONOUNi-FOR saadi ghet li (Marathi) Jane ERG self-for sari bought 'Jane bought a sari for herself.' (b) Janei ne John laa PRONOUNi-ABOUT maahiti Jane ERG John DAT self-about information dili gave 'Jane gave John information about herself.' 4 Binding versus coreference In this chapter we will refine the interpretation procedure developed in chapter 2. An important distinction - that between coreference and binding - will be introduced, motivated, and technically implemented. An early and very lucid explication of the distinction is found in Bach and Partee (1980), so lucid, in fact, that I'll simply quote it: Let's summarize the places where something like coindexing is used in the literature: (1) The same pronoun appears in several places in a sentence: He said he was OK. (2) A pronoun appears together with a referring NP: John said that he was OK. (3) A pronoun appears together with a quantificational NP: No woman doubts that she is OK. (4) A pronoun occurs in a relative clause: ... the woman who said that she had found the answer. (5) A reflexive or other obligatorily bound pronoun appears in a sentence: John loves himself Oscar is out of his head. It is really only in situation (1) (in some sentences), and (2) that it seems appropriate to talk about coreference. In every other case (...) coindexing a pronoun with some other expression is a shorthand way of saying that the pronoun in question is being interpreted as a bound-variable... Other authors have emphasized this point, too, in particular Tanya Reinhart (Reinhart, 1982, 1983a, b). Up to now we have uniformly interpreted coindexing to mark coreference. If, as Bach and Partee, and Reinhart in the aforementioned works, point out, coreference is only one of two semantic concepts that fall under the pre-theoretic concept of 'binding,' the other being variable binding, we should explore how this second concept can be implemented in our little grammar. To get a better understanding of what is behind this distinction it is perhaps best to look at prototypical examples: (4.1) Coreference: , ,5 Johns (4.2) Variable binding: No womans doubts that shes is okay. said that hes was okay. 81 82 BINDING VERSUS COREFERENCE As the name coreference suggests, we can think of the embedded pronoun in (4.1) as a referring expression of its own, which just picks out the same individual as its antecedent. Accordingly, apart from the awkwardness resulting from a Binding Condition C violation, the sentence in (4.3) is a semantically accurate paraphrase of (4.1): (4.3) John said that John was okay. This is as expected. If John and he in (4.1) are coreferential, that means that both expressions refer to the same individual, John. In other words, they have the same denotation (under the given indexing, that is). Therefore the two expressions should be exchangeable salva veritate (preserving the truth conditions), which they turn out to be. In striking contrast to that, no paraphrase similar to (4.3) can be given for the example that illustrates variable binding. Consider (4.4): (4.4) No woman doubts that no woman is okay. (4.4) patently differs in meaning from (4.2): if (4.2) is true, women in general have a high opinion of themselves (and possibly other women, though nothing is said about that); if (4.4) is true, women in general think lowly of women. Also, (4.4) doesn't feel like a Binding Condition C violation in the way that (4.3) does. It is, in fact, just unclear what it is supposed to mean to interpret the two occurrences of no woman as coreferent. This is very much unlike Binding Condition C violations with names as in (4.3), where we have a perfectly sensible and easily expressible meaning, albeit expressed in a way that is felt to be syntactically deviant. The conclusion is that she in (4.2) does not denote the same as no woman, and can therefore not be analyzed as coreferent with its antecedent. What we are going to see next is that no woman does not refer in the first place, which is why no pronoun can corefer with it. It turns out that there is an entire class of NPs which are, by their very semantic nature, incapable of reference, and hence coreference. These are quantified noun phrases (QNPs) such as no woman or each of the women, and w/z-phrases such as which man or who. It will be useful to start with an exploration of these (in particular, to abstract away from unnecessary complications: the QNPs) in order to find out more about what variable binding is. We will return to the question of whether ordinary, non-quantified NPs like John, the soprano, or even pronouns like she can function as variable binders in chapter 5, and to the question of whether (some) QNPs can refer in chapter 7. A terminological remark: in chapter 1 we defined the notion of a binder in syntactic terms (a c-commanding, coindexed NP). In this chapter we focus on a semantic distinction between variable binding and coreference. To keep matters clear, we will therefore distinguish between syntactic binding or syn-binding for short, and variable binding, which we will mostly refer to as semantic binding, or simply sem-binding. As a cover term for both, and for 4.1 Quantified NPs and variable binding 83 coreference without syntactic binding, we will henceforth use the term anaphoric relation. To exemplify, there is no doubt that both antecedent NPs in (4.1) and (4.2) -John/he and no woman - syntactically bind the pronouns, but what is at issue here is which anaphoric relation that syntactic binding expresses semantically: semantic binding (a.k.a. variable binding) or coreference. 4.1 Quantified NPs and variable binding Above we saw that we cannot simply replace a bound pronoun by a copy of its antecedent, if the antecedent is a QNP. The result is not a Binding Condition C violation, as predicted, but, curiously, a well-formed sentence that means something completely different. To get an idea of what is going on in these sentences, let us start with a simple example like (4.5). We assume, as before, that VPs such as is happy denote functions from individuals to truth values, here the function that maps all and only the happy creatures to 1. Additionally, we assume that common nouns such as manager also denote such functions, here: that function which maps all and only the managers to 1. (4.5) is true, then, if every individual that is a manager is a happy creature, (4.5b): (4.5) Every manager is happy. (a) [[is happy]g = [[happy] # = Xy.y is happy (the characteristic function of the set of happy creatures, {y | y is happy}) (b) [[Every manager is happy]g = 1 iff for every x, if x a manager, (i) [happy]8(x) = l or (ii) x is in {y \ y is happy} The two alternative renderings in (4.5b-i) and (4.5b-ii) relate back to the set/function equivalence and the convention related to it we introduced in chapter 2, section 2.3.1. Each function of type (et) (such as the denotation of an intransitive verb, or of a common noun such as manager) characterizes a set of individuals (cf. [4.5a]). Instead of saying that the function maps an individual to 1, as done in (4.5b-i), we can, and often will in the discussion to follow, say that the individual is in the corresponding set (cf. [4.5b-ii]). Other cases of QNPs can be interpreted in an analogous manner: (4.6) [[One manager is happy] g = 1 iff for (at least) one x such that [manager] #(x) = 1, [[happy]g(x) = 1 (4.7) [[No manager is happy] g = 1 iff for no x such that [manager] # (x) = 1, [[happy]g (x) = 1 The interested reader will find more about the meaning of QNPs in subsection 4.5.2 below. For the moment, all we need to know is that sentences of the form QNP VP are interpreted as in (4.8), where the VP meaning is a regular 84 BINDING VERSUS COREFERENCE property/set of individuals, but the QNP meaning is a more complicated logical construct, not just an individual or plurality of individuals: (4.8) QNP VP for every some (at least) n no x such that x is an N, x is in {y \ y is happy sleeps There is something curious about this schema, as well as the examples in (4.5-4.7), to wit that it doesn't mention any indices on the QNP. There is a reason for this, namely that the meaning of the QNP isn't dependent on the assignment function at all, and doesn't even relate to it. That is, while pronouns rely on the assignment functions to be interpreted in the first place, and names and definite NPs can at least be related to the assignment function because they denote individuals after all (see chapter 2), QNPs are simply not sensitive to assignments, period. This makes sense semantically, given that an assignment assigns referents to NPs, and that QNPs don't refer. It raises the embarrassing question, however, of whether the index on a QNP does anything at all. It is this question we will turn to next. Let us thus look at a case that involves coindexing. What do our interpretation rules predict as the denotation for an example like (4.9)? By the interpretations given in chapter 2, the denotation of the matrix VP will depend on the assignment g - (4.9a) - and so will the interpretation of the entire sentence - (4.9b) ([(4.9)]]g stands for the meaning of the entire sentence in (4.9); we'll ignore the gender and number presuppositions of pronouns throughout this chapter): (4.9) Every tenor2 thinks that he2 is competent. (a) |[thinks that he2 is competent Jg = Xx.x thinks that g(2) is competent (~ the set of people who think that g(2) is competent) (b) |[(4.9)]j# = 1 iff for every x such that x is a tenor, x is in the set of people who think that g (2) is competent Under the interpretation (4.9b), hei is a free pronoun, which receives its value from the context. This reading is clearly available for the sentence per se, but it shouldn't be under the indexing given in (4.9). What we rather want the example to denote is (4.10a). We will get this denotation if we can make the matrix VP have the denotation in (4.10b): (4.10) (a) wanted: [[(4.9)] # = 1 iff for every x, x a tenor, x is in the set of people who think that x is competent (b) required: |[thinks thathe2 is competent Jg = Xx.x thinks that* is competent (~ the set of people x who think that x is competent) How can we get this denotation for the VP? Suppose we treat the presence of a quantified NP with index n as an indicator that, within the c-command domain of that NP, pronouns bearing the index n are no longer referring pronouns, but bound pronouns. That means that their value is no longer determined by the 4.1 Quantified NPs and variable binding 85 contextual assignment, but by the argument slot filled by the NP. Our official implementation of this will have two parts: one syntactic; one semantic: (4.11) Index Transfer (preliminary): For any quantified noun phrase QNP with index n, adjoin fin to QNP's sister constituent: (4.12) Binder Index Evaluation rule (BIER) (final): For any natural number n, Index Transfer in (4.11) implements the observation made above that an index on a QNP doesn't play a role for the interpretation of QNP, but rather expresses that any coindexed pronoun within QNP's sister (its c-command domain) is to be interpreted as a bound variable. The BIER (4.12) consecutively interprets such a configuration. It states that /3„'s sister constituent Y is not to be interpreted relative to the original assignment g, but to a changed assignment g[n -> x], which is just like g except that the index n is mapped to the individual x. Since x is also the individual argument to [[Y]]^'2-^, this means in effect that any pronoun bearing the index n in Y is bound by the open argument slot of Y (if you are unfamiliar with the g[n -> x] notation for assignment modifications, you might want to skip to subsection 4.5.1 below before reading on). Let us see this in a sample derivation: (4.13) Every tenor4 thinks that he4 is competent, after Index Transfer: s that he4 is competent |[[yo thinks ]Jg = XptXze z thinks (that) p |[[^that he4 is competent]= the proposition that g(4) is competent l^[yp thinks that he4 is competent] = Xz.z thinks that g(4) is competent ~ the set of those who think that g(4) is competent ßn = Xx.lYJ (x) (a) (b) (c) 86 BINDING VERSUS COREFERENCE (d) Hyp* Pa [yP thinks that he4 is competent]]Jg = . Hyp thinks that he4 is competent ]] «^x^ (x) (by BIER) = Xx.[Xz.z thinks that g[4 -> x](4) is competent ](x)1 = Xx.x thinks that g[4 -> x](4) is competent2 = Xx [x thinks that x is competent ] ~ the set of those x who think that x is competent ~ the set of people who consider themselves competent It is very important to understand the difference between (4.13c) and (4.13d): (4.13c) denotes the set of people who have a belief about a particular individual g(4). (4.13d) denotes the set of people who have a belief about themselves. It is precisely the transition to the latter that accounts for binding. What remains to be done is to combine (4.13d) with the meaning of the subject every tenor in the standard way (the standard way being [4.8], or some reasonable generalization thereof), which gives us the desired truth conditions: (4.14) fevery tenor4 [^4 [thinks thathe4 is competent]]]# = for every y such that y is a tenor, [Ax.|[[vp thinks that he4 is competent]F[4^x](x)](y) = 1 = ... y is in {x I |[[yp thinks that he4 is competent]]#[4^x] (x)} = ... y is in {x I x thinks that x is competent} = every tenor is in the set of those who consider themselves competent Note in particular that the denotation of the entire sentence is not dependent on the assignment g at all, despite the fact that it contains indexed NPs. The reason is that the index on every tenor is interpreted by the BIER, rather than as a referential index on an NP, and that the pronoun is semantically bound within the clause (by the QNP). In what follows, I will call the adjoined fin a binder prefix, and the NP minimally c-commanding it the (semantic) binder NP. The index on a binder NP (before Index Transfer) or on a binder prefix (after Index Transfer) will be called a (semantic) binder index. I will use the term semantic binder both for binder prefixes and, derivatively, binder NPs. Semantic binding (sem-binding) is defined as follows: (4.15) Semantic binding (final): A binder prefix sem(antically)-binds an NP if and only if (a) and NP are coindexed (b) c-commands NP (c) there is no binder prefix fi' which is c-commanded by and meets (a) and (b) If an NP is not bound by any binder prefix in a phrase marker P, we say that NP is semantically free in P. 1 [Az.z thinks that g[4 —► x](4) is competent] is copied from (4.13c), with g[4 —► x] replacing g. 2 Note that 'x thinks that g[4 —► x](4) is competent' is the result of applying the function [Az.z thinks that g[4 —► x](4) is competent] to the variable x. 4.1 Quantified NPs and variable binding 87 In the example (4.13) above, the adjoined fa is a binder prefix with binder index 4; it sem-binds the pronoun he\, which is therefore semantically bound. Derivatively, the NP every tenor sem-binds he4, and is thus a semantic binder as well. Applying Index Transfer (4.11) is a necessary condition for semantic binding, but not a sufficient one. If a pronoun is in the c-command domain of a QNP, but not coindexed with it, it will still be interpreted as a free pronoun whose denotation is determined by the global assignment.3 This is demonstrated in (4.16) below: (4.16) Every tenor3 thinks that he2 is competent. (a) after Index Transfer: Every tenor [yP* fa [vp thinks that he2 is competent.]] (b) l_[yP thinks that he2 is competent]]g = Xz.z thinks that g(2) is competent (~ the set of those who think that g(2) is competent) (c) Hyp* fa [vp thinks that he2 is competent ]]Jg = XxllVP thinks that he2 is competent ]]^^3^x^(x)] = Xx x thinks that g[3 -> x](2) is competent = Xx .x thinks that g (2) is competent ~ the set of those x who think that g(2) is competent (d) for every y such that y is a tenor, [Axft^, thinks that he2 is competent ]J8[3^x](x)]](y) = 1 ~ every tenor is in the set of those who think that g(2) is competent Crucially, the BIER manipulates the assignment in (4.16c) to g[3 -> x], but this change doesn't have any effect on the VP-denotation, which doesn't contain an NP indexed 3. In particular, the denotation of the pronoun he2, g[3 -> x](2), remains unaffected by the change, which is why we could simply write g{2) instead (recall that g[3 -> x] is just like g except perhaps which value it assigns to 3). That is, hei remains a free pronoun that receives its value from the context. Note that even though there isn't any sem-bound NP in (4.16), the prefix fa and the NP every tenor qualify as semantic binders. In such a situation we speak of vacuous binding: the index 3 is null and void. There is no immediate reason to prohibit indexings that lead to vacuous binding; the indexing in (4.16) leads to an interpretation in which he and every tenor are independent of one another, which is intuitively satisfactory. One last terminological remark: indices which are not binder indices (i.e. not on a QNP or a binder prefix fa will be referred to as referential indices. This makes sense, since these indices - unlike binder indices - determine the denotation of the NP they are on, which is an individual. Confusion might arise, though, given that not every NP that bears a referential index refers to one particular individual. For example, the index 2 on he in (4.16) is a referential index, as is 3 The interpretation of a bound pronoun is, of course, dependent on the assignment, too, but only locally. 88 BINDING VERSUS COREFERENCE the index 4 on he in (4.13). However, the latter is semantically bound in the sentence, which means that it will not refer to a particular individual or referent, but rather covary with its semantic binder. So only semantically free pronouns are referential in the intuitive sense, while semantically bound pronouns are not. However, both types carry referential indices. Put in other words, the distinction between referential indices and binder indices is a local one (it only regards the NP and possibly its neighboring binder prefix), while the distinction between bound and referential NPs is a global one (an NP is bound or free only within a larger syntactic domain). We observed earlier that since QNPs do not refer, we don't really have any use for an index on them. Given the workings of the Index Transfer rule in (4.11) we no longer need to worry about that, since it doesn't just adjoin an index to the QNP's sister constituent, but strips that index off the QNP at the same time, leaving a 'bare' QNP. Having two notions of binding, semantic and syntactic, the question arises which of these is relevant for our Binding Conditions. We will return to this question in chapters 5 and 6. For the moment, we will assume that all Binding Conditions regard syntactic binding. This implies that Binding Conditions have to apply before Index Transfer, while all NPs still bear their indices. Let us henceforth refer to the result of applying Index Transfer as the Logical Form or LF of a sentence. We then require: (4.17) The Binding Conditions apply before LF. (to be revised) Also, we must make sure that Index Transfer applies obligatorily to all QNPs (and w/z-NPs, see below); otherwise we could have a pronoun be syntactically bound by a QNP, without being semantically bound by that QNP at LF. To ensure this, we demand: (4.18) No QNP (or wh-NP) bears an index at LF. Not too much should be read into the LF terminology. As will be discussed in chapters 5 and 6, the LF/pre-LF distinction is entirely reducable in the present context and merely serves expository purposes. An equivalent non-derivational formalism can easily be formulated. By the same token, the introduction of the binder prefix as a separate syntactic constituent is merely for expository convenience; alternative systems that don't involve an actual extra node in the tree can be defined, but are considerably less transparent. Exercise 4.1 Which of the following statements are true? Which are false and why? Add qualifications where necessary. 1. If a pronoun 0 is semantically free in S, its meaning in S will depend on the contextual assignment g. 2. If a pronoun 0 is semantically bound in S, its meaning in S will depend on the contextual assignment g. 4.2 The syntax of semantic binding 89 3. If a pronoun cpn is semantically bound by an NP in a configuration like [sNP[a 0! [y... 0!...]]], 0 is free in S. 4. If a pronoun (pn is semantically bound by an NP in a configuration [sNP \aPi[Y...\ •••]]], 0 is free in y. 5. If a pronoun (pn is coindexed with an NP in a configuration [sNP$ [oily - ■ ■ 03 • • •]]], 0 is semantically bound in S. 6. If a pronoun 0„ is coindexed with an NP in a configuration isNP3 [a[y ■ .03 • • •]]], 0 is free in S. 7. If a pronoun is not c-commanded by any coindexed NP in S, that pronoun is semantically free in S. 8. If a pronoun is syntactically bound by an NP in S, it cannot be semantically free in S. 9. If a pronoun is syntactically bound by a QNP in S, it cannot be semantically free in S. 10. If a pronoun is syntactically bound by a QNP in S, it cannot be semantically free in any subtree of S. 11. If an LF contains an adjoined binder prefix /3n, it also contains a semantically bound pronoun indexed n. 12. If a structure S contains an NP indexed n at s-structure, and NP is not semantically bound, the denotation of S will depend on the contextual assignment g. 13. If all indices within a structure S are on QNPs or NPs that are semantically bound in S, the denotation of S will not depend on the contextual assignment g. Exercise 4.2 Suppose that structures of the form NP[fin Y] are base-generated (rather than derived by Index Transfer). 1. Formulate a definition of syntactic binding that will derive the familiar Binding Condition A-C effects both for quantificational and non-quantificational antecedents under this assumption. 2. Provide an indexing and prefixing procedure which generates all and only the interpretable structures under this assumption. 4.2 The syntax of semantic binding Let us now turn to the syntactic conditions that need to hold for semantic binding to occur. I will demonstrate that two central aspects of the binding behavior of QNPs follow directly from the semantic assumptions we made in the previous two sections. 4.2.1 The C-command requirement on semantic binding Consider the examples in (4.19): 90 BINDING VERSUS COREFERENCE (4.19) (a) *He2 exploits the secretary that Felix2 hired. (b) He2 exploits the secretary that he2 hired. (c) * He2 exploits the secretary that each of the tenors2 hired. The fact that he and Felix in (4.19a) cannot corefer is due to a Binding Condition C violation: the name is syntactically bound (whence the contrast with [4.19b]). The fact that he cannot be coindexed with each of the tenors in (4.19c) can likewise be attributed to Binding Condition C. Quantified noun phrases such as each of the tenors are full NPs and are therefore subject to Binding Condition C. Consider next the examples in (4.20): (4.20) (a) The secretary he2 hired thinks that Siegfried2 is despotic. (b) *The secretary he2 hired thinks that each of the tenors2 is despotic. Intuitively, (4.20b) does not allow for a bound variable reading for he, i.e. it cannot mean 'each of the tenors is considered despotic by the secretary he hired'; this is why the sentence is starred under the given indexing. The unavailability of this reading cannot be attributed to Binding Condition C, given that he does not c-command each of the tenors. This is also confirmed by the fact that the otherwise parallel (4.20a) is perfectly well-formed on a reading where he and Siegfried are coreferent. It seems that - unlike names - QNPs don't allow for coindexing without c-command.4 One way to block cases like (4.20b) is by a stipulation like (4.21) (I'll return to the w/z-trace part below): (4.21) Bound Anaphora Condition (BAC, Reinhart, 1983a: 122/137) Quantified NPs and w/Vtraces can have anaphoric relations only with pronouns in their c-command syntactic domain. Since each of the tenors is a quantified noun phrase, (4.21) requires it to c-command the pronoun him in order to be coindexed ('have an anaphoric relation') with it. This is not the case in (4.20b), which is therefore correctly ruled out. It is interesting to note that (4.21) also renders ungrammatical our earlier example (4.19c), repeated here as (4.22): (4.22) *He2 exploits the secretary that each of the tenors2 hired. On top of violating Binding Condition C (because the pronoun syn-binds the QNP), this sentence is now ruled out because it violates the BAC (since the QNP doesn't c-command the pronoun it is coindexed with). As Reinhart argues, however, it is dubious that something like the BAC in (4.21) is needed as an independent principle of grammar. We have already established that QNPs (and w/z-expressions) do not refer and therefore don't allow for coreference. The only option to interpret the coindexing is thus via semantic binding, as discussed in section 4.1. Given this much, the BAC can be reduced to the restriction in (4.23) (cf. also Heim and Kratzer [1998]:264): 4 See e.g. Reinhart (1983a): 112f. for more examples. 4.2 The syntax of semantic binding 91 (4.23) Semantic binding requires syntactic binding (i.e. that the binder be coin-dexed with and c-command the pronoun). In neither one of the examples (4.20a) and (4.20b) does the full NP Siegfried and each of the tenors, respectively, c-command - and hence syn-bind - the pronoun. In the case of (4.20a) this doesn't matter since the coindexing is simply interpreted as coreference, which doesn't require c-command. The QNP each of the tenors in (4.20b), however, can neither corefer with the pronoun (since it doesn't refer), nor can it semantically bind it (since it doesn't c-command, hence syntactically bind it, as required by [4.23]). Therefore the indexing can neither be interpreted as coreference nor as semantic binding. The same dilemma is found for the QNP in (4.22). We can go further than reducing (4.21) to (4.23); the effect of the restriction in (4.23) actually follows from the way we defined semantic binding, in particular two aspects of it: first, the binder prefix is adjoined immediately c-commanded by the QNPs; its c-command domain is the same as that of the QNP (prior to Index Transfer); second, the BIER manipulates the assignment function and thereby binds all free coindexed pronouns within 's sister constituent; a pronoun outside of that constituent will remain unaffected by it, regardless of its index. In other words, if a coindexed pronoun is outside of the c-command domain of a QNP, coindexing between the two will be semantically vacuous. Exactly this case is found in our sentence (4.20b), repeated here. Its pertinent representation is (4.24a). Applying the BIER, the pronoun he& will be contextu-ally interpreted as g(6), while each of the tenors^ will bind vacuously, as there are no free occurrences of the index 6 within its sister constituent, the VP is despotic, cf. (4.24b): (4.24) The secretary he6 hired thinks that each of the tenors6 is despotic. (a) LF (i.e. after Index Transfer): The secretary he6 hired thinks that each of the tenors [^[is despotic]] (b) |[ (4.24a)]g = the secretary g(6) hired thinks that for each x, x a tenor, x is in {z | |[despoticp[6^z](z)} = the secretary g(6) hired thinks that for each x, x a tenor, x e {z | z is despotic} The important fact in (4.24) is that the pronoun he§ will end up as a free pronoun with the interpretation g(6), while the binder index 6 doesn't get to bind anything. This result is more general and thus bears restating: semantic binding involves adjunction of a binder prefix /3. Coindexed pronouns are interpreted as bound variables only within the c-command domain of that prefix. And the c-command domain of /3 after Index Transfer equals that of the QNP before. Taken together this means that: (4.25) Theorem: Coindexing between a QNP and a pronoun results in semantic binding only if the QNP c-commands the pronoun. 92 BINDING VERSUS COREFERENCE It should be easy to see that this captures the effect of Reinhart's BAC and Heim and Kratzer's (4.23) above. It is derived that the indexing in (4.24) does not result in a bound variable reading for the pronoun him. We have not, however, excluded the coindexing itself. To do that, we would still have to evoke a principle like the Bound Anaphora Condition in (4.21). The question we should ask ourselves at this point is whether we should do this or not. If our task is to generate intuitively correct indexing patterns, we obviously still need the Bound Anaphora Condition. But, as pointed out at the end of section 2.3 in chapter 2, indices are not part of the linguistic data, but merely ancillary devices to predict the data, in particular the range of possible interpretations. If that is the task, then there is no need to ban the indexings in (4.22) and (4.24), as long as our interpretation procedure correctly predicts that no semantic binding (let alone coreference) results from it. Consequentially we will not adopt the BAC or (4.23). The indexing in (4.24) (The secretary he§ hired thinks that each of the managers^, is despotic.) is thus grammatical (though misleading), but won't result in any anaphoric relation. 4.2.2 Why QNPs cannot be bound i Let us now return to a case in which a QNP is illicitly c-commanded by another coindexed NP, e.g. (4.19c), repeated here: (4.26) *He2 exploits the secretary that each of the tenors2 hired. LF: he2 exploits the secretary that [each of the tenors] [ ^2 [ hired ]]. As discussed above, the indexing in (4.26) is ruled out by Binding Condition C, since each of the tenors is a full NP. But interestingly, the coindexing will also be semantically vacuous: the pronoun is not semantically bound by the QNP, due to lack of c-command (cf. last subsection). The index on the QNP has been stripped off it and acts as a binder index, which binds vacuously (since there are no coindexed pronouns in the binder prefix's c-command domain); since binder prefixes don't refer, there is no coreference here either. The interpretation we assign to (4.26) is thus the same we would assign to it if he and each of the managers were contra-indexed. The situation in which a NP c-commands a coindexed QNP is thus technically no different from that in which neither c-commands the other: the coindexing is simply vacuous. So while Binding Condition C blocks the indexing in (4.26) (as would the BAC), even without Binding Condition C this sentence would receive only its intuitively available interpretation, in which hei is free. Exercise 4.3 Consider the following sentence with the indexing given: (4.27) Every girk; told her4 mother that every boy4 forgot his4 books. 4.3 l/l/f?-expressions 93 1. Does this structure violate any Binding Principles before Index Transfer? If so, which and how? 2. List all pairs of syntactic binders and bindees in this structure. 3. Give the LF representation(s) for this sentence. 4. List all syntactic binder/bindee pairs for this LF. 5. List all semantic binder/bindee pairs for this LF. 6. Does this LF representation violate any Binding Principles? If crucial, specify which elements in your LF fall under which Binding Principle(s). 7. Calculate the interpretation for this LF. Be sure to give the meanings of all VPs and VP-segments as intermediary steps. 8. Does your interpretation match with your intuitions about the meaning of this sentence? Comment! Exercise 4.4 Contrast (4.27) with the following: (4.28) Every girLt told \1er4 mother that every boy4 stole her4 books. 1. What interpretation will this sentence get, and why? 2. Give at least one indexing which represents an intuitively available reading for this sentence. Exercise 4.5 Give an LF and calculate the denotation for the following sentence: (4.29) Every girLt told her4 mother that every boy2 showed her4 his2 books. 4.3 l/l//7-expressions Before closing this chapter, a few words should be said about wh-expressions, in particular those that are NPs such as which crocodile or what. Our intuition, presumably even stronger than with QNPs, is that w/z-expressions do not refer: after all, it is the point of asking a w/z-question to learn about the referent of the pertinent NP.5 So thinking of another NP, say a pronoun, as coref-erent with a w/z-expression makes no sense. Things are different with regard to semantic binding. Consider the sentence in (4.30): (4.30) Which girl told Suzie that she had detention? This sentence has two plausible interpretations. One where we ask which girl said to Suzie: 'You have detention'; and one where we ask which girl said to 5 The term 'referential w/z-expression' is sometimes applied in the syntactic literature to phrases like which of the boys (or even just which boy) and is presumably meant in the sense of 'wh-phrase with lexical content,' or 'w/z-phrase containing a referential expression' (as opposed to e.g. how or who the hell). It should be clear that the w/z-phrases themselves are in no sense referring. 94 BINDING VERSUS COREFERENCE Suzie: 'I have detention.' In addition there is a less salient, but nonetheless plausible, reading where the girl said: 'She has detention.' We can represent these different readings by different coindexings: (4.31) Which girl\ told Suzie2 that shei/2/3 had detention? The different indices represent different anaphoric relations: she\ is the teller (the 'I have dentention' reading), shei is Suzie ('you have detention'), and shej, is some third party ('she has detention'). It is straightforward to get the second and third interpretations: she gets its referent from the contextual assignment g, which is the same as that of Suzie (index 2) or some other antecedent NP (index 3). The 1-indexing, however, encodes a bound variable reading. The denotation for the question should be something along the lines of (4.32): (4.32) for which girl x is the following the case: x told Suzie that x has detention This contrasts with the two other readings, which we might represent as in (4.33): (4.33) for which girl x is the following the case: x told Suzie that detention (where g(2) = Suzie, and g(3) = some contextually given female person) To derive an interpretation along these lines, we fortunately don't have to go into the semantics of questions at all. All we need to concern ourselves with is the part after the colon in (4.32) and (4.33). Assume the following LF for (4.31) under the 1-indexing: (4.34) which girl [yp*Pi[Vp told Suzie2 that shei had detention ]] Here, Index Transfer has adjoined a binder index /3i to the sister of which girl, VP (this is assuming that which girl has not been displaced, which we will for the sake of the exposition; I will return to the issue of w/z-movement in chapter 8). By the BIER, the interpretation of VP* will be (4.35): (4.35) I[yp*^i [yp told Suzie2 that shei had detention ]]Jg = Xx.[[VP told Suzie2 that shei had detention ]]K'-1_>x-'(x) = Xx.x told Suzie that g[l -> x](l) had detention = Xx .x told Suzie that x had detention This is precisely the meaning we assumed in (4.32). We see thus that while wh-expressions must have a denotation very different from QNPs (roughly something to get us from [4.35] to [4.32]), their behavior with regard to the Binding Theory is parallel: they do not corefer, but can act as semantic binders. The derivation of readings where they do behaves parallel to that of examples with QNPs in all relevant respects. We can close by noting that this parallelism extends to the structural prerequisites for semantic binding. For example, neither sentence in (4.36) can get a reading in which which girl can have anaphoric relations with the pronoun: 4.4 Summary 95 (4.36) (a) She\ wondered which girl\ had detention. (b) Her\ mother wondered which girl\ had detention. These examples display violations of the BAC and, in the case of (4.36a), Binding Condition C. Independent of that, and in full parallelism to the case of QNP-binding discussed in 4.2, they also fail to get an interpretation on which the w/z-phrase and the pronoun are anaphorically related (in fact, it is unclear what such a reading would be). This again falls out from our semantics, given that the w/z-expressions fail to c-command the pronouns, which would be a prerequisite to interpret the coindexing as semantic binding. 4.4 Summary This concludes our main discussion of coreference vs. variable binding. Let me summarize the main points: • Coindexing between two NPs can be interpreted in two essentially different ways: as coreference (both NPs denote the same individual); or as semantic binding (the bound NP covaries with its binder). • Quantified NPs and w/z-expressions do not refer, hence a fortiori do not corefer. The only available interpretation of coindexation with such an expression is thus semantic binding. • Semantic binding is restricted to configurations in which the binder (the QNP) c-commands the bound pronoun. In a slogan, semantic binding requires syntactic binding. Coindexing without c-command can only be interpreted as coreference (if no QNPs are involved) or not at all (if QNPs are involved). • This restriction can very plausibly be attributed to the way the interpretation of binding dependencies works. In other words, the c-command restriction on coindexing with a QNP doesn't have to be stipulated, since it follows from the interpretive semantics that any other coindexing simply doesn't result in a bound variable interpretation of the pronoun at all. We distinguished different syntactic and semantic relations between NPs: coindexing and coreference are symmetric relations; if NPi is coindexed and/or co-rereferent with NP2, then NP2 is coindexed/coreferent with NPi. Syntactic and semantic binding are asymmetric notions: among two NPs in a binding relation, one is the binder, the other one the bindee. As things stand, the two pairs are not in a one-to-one correspondence: coindexing encompasses all types of anaphoric dependencies (not just coreference), and syntactic binding may or may not coincide with semantic binding (though semantic binding always presupposes syntactic binding, as discussed). This situation is schematically summarized in table 4.1. 96 BINDING VERSUS COREFERENCE Table 4.1 Syntactic and semantic relations between NPs SYNTAX coindexing non-coindexing syntactic binding (one NP c-commands the other) no syntactic binding (no c-command) semantic binding (binder is a QNP) coreference (no QNP involved) SEMANTICS non-coreference 4.5 Semantic details 4.5.1 Assignment modification i The BIER in (4.12) above makes crucial use of assignment modification, using the notation g[n -> x], which is defined as follows: (4.37) Assignment modification: For any assignment g, g[n -> a] is that assignment which is like g, except that g{n) = a. Modifying an assignment is thus to exchange one individual in the sequence for another. To illustrate, let's take our assignment g\ from chapter 2. (4.38) gi= 1 2 3 4 5 6 7 \ 4 4 4 4 4 4 Bo Kim Sven Ana Kurt Eszter Tanya We can now modify g\ to, say, g\[4 —> Marge]:6 (4.39) gi [4 -> Marge] = 1 2 3 4 5 6 7 4 4 I I 1 4 Bo Kim Sven Marge Kurt Eszter Tanya Modification can be done iteratively ... (4.40) (a) gi[4 -> Marge\\\ -+ Ana] = 1 2 3 4 5 6 7 Ana Kim Sven Marge Kurt Eszter Tanya 6 We could also 'modify' g\ to g\[4 —► Ana], i.e. assignment modification can be vacuous. 4.5 Semantic details© 97 (b) gi[4 -> Marge][l -> Ana][5 -> Car/] = 1 2 3 4 5 6 7 ... Ana Kim Sven Marge Carl Eszter Tanya ... and even to the same argument slot: (4.41) (a) gl[4 -> Marge][5 -> Car/][4 -> Sy/v/e] = 1 2 3 4 5 6 7 ... Bo Kim Sven Sylvie Carl Eszter Tanya (b) gi[4 -> Marge] [5 -> Carl][4 -> Ana] = 1 2 3 4 5 6 7 ... Bo Kim Sven Ana Carl Eszter Tanya Note that in (4.41a), the change brought about by [4 -> Marge] is 'overwritten' by the later modification [4 -> Sylvie]. In (4.41b) the later modification [4 -> Ana] even changes the value of g(4) back to the original value, 'cancelling' the first modification [4 -> Marge]. Modifications are thus carried out left-to-right. Obviously, if an XP contains a free pronoun, changing the assignment can change the denotation of the XP as a whole: (4.42) (a) [like hirrnjp = the set of all those who like g(4), for any g (b) [like \lim4J8[4^Abraham] = the set of all those who like g[4 -> Abraham](4), for any g = the set of all those who like Abraham, for any g = the set of all those who like Abraham, regardless of g Accordingly, if an XP denotes a proposition, changing the assignment can change the proposition denoted by XP: (4.43) (a) |[she2 is asleep] # = 1 iff g(2) is asleep. (b) |[she2 is asleep] g\-2^clara\ = l iff Clara is asleep. (c) |[she2 is asleep] 8i2^Bet^y] = \ iff Betsy is asleep. (d) |[she2 is asleep] 8l2^x^ = 1 iff ^ is asleep. In (4.43d), the second slot in the assignment is replaced by a variable x. Since our meta-language - English - doesn't contain free variables, this expression doesn't describe any denotation. We can, however, use this technique to form a predicate out of an expression by 're-opening' an argument slot. Thus (4.44a) denotes the set of all individuals who, if replaced for slot 2 in the assignment g, would make the sentence shei is asleep true. That, of course, is the set of all (female) sleepers (or rather: the characteristic function thereof): (4.44) (a) Xx.|[she2 is asleep]8l2^x^ the set of those who are asleep (b) [Xx.lshe2 is asleep]«[2^x]](Betsy) Accordingly, (4.44b) expresses the same proposition as (4.43c). 98 BINDING VERSUS COREFERENCE Exercise 4.6 Which sets of people do the following functions characterize? (4.45) (a) [likeher9F (b) [like hergjSl^Arabella] (c) [like her9p[4^4raM/fl] (d) [like her4P[4^4raM/fl] (e) [introduce her4 to hini9p (f) [introduce her4 to him9] gV^Kim^ (g) [introduce her4 to him9] sl^Kim] (h) [introduce her4 to him9p[4^&m][9^w>] (i) [introduce her4 to him9F[4^w?][9^'m] (j) [introduce her4 to him9p[9^'m][4^w^] (k) Ax [[introduce her4 to him9p[9^'m][4^w%;0 (1) Xx [[introduce her4 to him9] «[9^x]l4^Sandyl ] (x) (m) Xx[[introduce her4 to him9F[9^Km][4^x]](x) (n) Xx [[introduce her4 to him9 ] g\-9^Kim\ ] (Sandy) (o) Xx [[introduce her4 to him9 ] -?[9^x] [4^x] ] (Sandy) 4.5.2 The semantics of quantified NPs i Above we never concerned ourselves with what QNPs actually denote. Rather, we concentrated on logical paraphrases of entire sentences containing QNPs such as in (4.46a): (4.46) Every manager is happy. (a) [Every manager is happy] 8 = 1 iff for every x, if [manager] g(x) = 1, then [happy]8(x) = 1 (b) {x | x is a manager} is a subset of {x \ x is happy} (4.46a) is equivalent to (4.46b) (recall that {x \ x a manager} and [x \ x is happy} are the sets characterized by [manager]]8 and [happyp, respectively). Similar renderings can be given for the other examples discussed in 4.1: (4.47) One manager is happy. there is some x such that [manager] 8 (x) = l and [happy]8(x) = 1 {x | x is a manager} and {x \ x is happy} have a common element (4.48) No manager is happy. there is no x such that [manager]8(x)=\ and [happy]8(x) = 1 [x | x is a manager} and [x | x is happy} do not have a common element We will not go into the meaning of quantificational determiners like every or no in isolation here (see e.g. Heim and Kratzer [1998]:ch. 6:4); we will, however, give meanings for selected QNPs such as every manager in (4.49): (4.49) [every manager]8 = XP e Detlor all x, if [manager]g(x) = 1, then P(x) = 1 (the set of all properties P which maps every manager to 1; i.e. the set of all properties that characterize a superset of the manager-set) 4.5 Semantic details© 99 (4.50) Every manager is happy. (a) [every manager is happy]g = 1 iff [XP e Def.for all x, if [manager]g{x) = 1, then P{x) = l](|[is happy]g) (b) = 1 iff for all x, if [manager]g(x) = 1, then [[is happy]g(x) = 1 (c) = 1 iff for all x, if [manager]g(x) = 1, then [Ajc.;t is happy](x) = 1 (d) = 1 iff the set of managers is a subset of the set of happy creatures The sample calculation in (4.50) shows how such a meaning combines with a VP meaning. Three things are particularly important to note: first, the semantic argument of the QNP meaning, corresponding to the variable P in (4.49), is not an individual, but a property; it is of type (e,t). This is explicitly written in (4.49), but henceforth we will just stick to the convention introduced in table 2.1 of chapter 2 and use variables P, P\, P2 ■ ■ ■ Second, the denotation of an NP headed by a quantificational determiner like every in (4.49) is not an individual, but something more complicated: a function from properties (which themselves are functions from individuals to truth values) to truth values (which corresponds, in set talk, to a set of sets of individuals); this kind of semantic object is called a generalized quantifier, following Barwise and Cooper (1981). It denotes a property of properties (e.g. that the property of being a happy creature has the property of mapping every element of the manager-set to 1), and crucially not an individual, even broadly construed as including pluralities of individuals. This is why it was said earlier that QNPs do not refer, given that reference is the subcase of denotation in which the denotation is an individual. Third, the Function Application rule applies the NP denotation to the VP denotation in (4.50a), not the other way around, i.e. [NP VPF = ENPF(EVPF) instead of [VPF(ENPF), as we used to do. This is a direct consequence of the fact that NP denotes a generalized quantifier, as just discussed. The reader may verify that the choice of functor and argument for Function Application follows from the types of the NP and VP denotations, according to the type-driven interpretation rule given in chapter 2, section 2.3.3. Readers with further interest in the compositional interpretation of QNPs are invited to consult a textbook on formal semantics such as Heim and Kratzer (1998):ch. 6, at this point. Exercise 4.7 Define analogous meanings for one of the managers and none of the managers. 4.5.3 Object QNPs 1 The discussion in the main text concerned only QNPs in subject position. The reason for that is not specific to Binding Theory, but pertains to the general question of how to interpret QNPs in object positions. Consider, for example, (4.51): 100 BINDING VERSUS COREFERENCE (4.51) met Jenny every gymnast [met] 5 is a function in De et, it corresponds to a relation. Jenny denotes an individual, so she is in the domain of [met]]g, so that [met]]8([Jenny]]5) yields a well-formed interpretation by our rule of Function Application (FA) from chapter 2. Every gymnast denotes a generalized quantifier (type (et,t)). The problem is that neither [met]]8([every gymnast]8) nor [every gymnast]]8([met]8) are defined; since neither function is in the domain of the other, Function Application is simply not defined for types (e,et) and (et,t). One way of remedying this is by a syntactic transformation, quantifier raising, which we will briefly discuss in chapter 8 (see e.g. Heim and Kratzer [1998]:ch. 7.3). In our official grammar, we will use a semantic combinator instead. The idea is this: a standard generalized quantifier q can combine by FA with a property P to yield q(P). Now, if we want to combine q with a relation R, we want the result to be kx.Q([ky.R(y)(x)]). In other words, q 'targets' the innermost argument slot of R, y, while the higher one, x, is 'transferred to the outside.' If q combines with a three-place relation R3, the two outermost arguments are 'transferred,' as in kxky.Q([kz.R3 (z)(x)(y)]), and so forth. To implement this idea, we first define a two-place composition function C (for combine), which can interpret QNPs and verbal meanings (among others) of any arity: (4.52) C() if 0 € Det, (b) Xf\C{Xy{(f){y){f)\, q)] otherwise (a) Predicate denotation: joinable type 4.5 Semantic details© 101 (4.54) NP* NP is that function which maps any 0 onto C(0,|[NP]g) Since C can only combine generalized quantifiers with predicate meanings, it follows that NP must be a quantified NP (or in any case, denote a generalized quantifier), and that [k NP] can only combine with expressions that denote predicate denotations. The workings of k and C are demonstrated below, where EC abbreviates [every cakeP (in Det,t) and ATE abbreviates [[ate]]g (in De^et): (4.55) Muni ate every cake. VP (a) (b) (c) (d) every cake C(ATE,EC) C(ATE,EC) = Xx[C(Xy[ ATE(y)(x)]], EC)] (by [4.52b]) (i) C(Xy[ATE(y)(x)],EC) = EC(Xy[ATE(y)(x)]) (by [4.52a]) C(EC,ATE) = Xx[EC(Xy[ATE(y)(x)])] (from [4.55b], [4.55b-i]) [Muni ate every cakep = Xx[ EC(Ay[ATE(y)(x)])](Muni) = 1 iff EC(Xy [ATE (y) (Muni) ]) As the reader is invited to verify for her/himself, these rules work fully generally, i.e. generalized quantifiers can combine with ditransitive verbs, verbs that take an additional sentential object, and so forth. Exercise 4.8 Calculate the meaning for no student solved every problem in the way shown in (4.55). Exercise 4.9 Assume that [[her„ desk]]5 = the desk of g(n) (type (e)). Give a complete LF based on the following unindexed tree for a reading on which every girl binds her. Make sure to include all necessary indices, prefixes, etc. Then interpret that LF (I chose a verb-final structure here to abstract away from complications with the English double object construction; think of this structure as German with English words): (4.56) 102 BINDING VERSUS COREFERENCE 4.5.4 Semantic binding in flat structures In section 2.3 of chapter 2 we introduced a rule to interpret n-ary branching structures as found, by assumption, in English VPs, Russian N's, and any sort of 'flat' structure. This rule is repeated in (4.57): (4.57) Af-ary Function Application: x" [X°F([A„F)([AM_iF)... «Aip) where A\, A2, ■ ■ ■, An-\, An is the order of A, A, A ...on X°'s argument-list This rule only works if all arguments A, A ... denote individuals. As a first step, then, we change this rule so as to apply [NP]g to [X°P instead, in case [[NP]]8 is a generalized quantifier; to facilitate readability, we define an operation '+': (4.58) P+A =def P(A) if defined, C(P, A) otherwise (cf. [4.52]) Like Function Application, we understand + to be left-associative, i.e. P + A + A' = (P + A) + A'. The + operation has the combinator C built into it. It will deliver a well-formed denotation for a generalized quantifier or an individual combined with any predicate denotation. Thus k prefixing is superfluous when using our new n-ary FA rule: (4.59) Af-ary Function Application with generalized quantifiers: -\-\S [X°F + [An]* + [An_iF + ... + [Ail* where A\, A2, ..., A„_i, An is the order of A, A', A"...on X argument-list 0' (4.59) still does not allow for binding. In fact, the Index Transfer rule (4.11) is undefined for NPs in an n-ary branching structure, given that it wants to adjoin to the sister of the QNP bearing a binder index. The correct generalization about semantic binding in flat structures is, not surprisingly, that it is possible wherever syntactic binding is. That is, an NP can bind (into) another XP if and only if NP commands XP. Descriptively speaking, NP can semantically bind (into) XP iff NP can, under the right circumstances, bind a reflexive (in) XP or block a non-reflexive (in) XP: (4.60) (a) I showed every boy3 (b) *I showed himself3 his3 friend himself^ (in the mirror) MS3 desk every boy3 (in the mirror). 4.5 Semantic details© 103 (4.61) (a) I showed every girls to (b) *I showed herself^ hers sister herself^ hers sister to every girls • To get this to follow from the semantic interpretation requires a little work. The reason is that (4.59) above gets the meanings of the argument NPs A, A , ... relative to the assignment g as its input. But if NPi were to bind NP2, it would have to change the assignment for NP2. In order for that to work composition-ally, we must input the composition rule not with [NP2P, but with A.g.[[NP2]]'?, a function from assignments to ordinary NP-denotations. We then define a semantic binding operator £>, the counterpart to the syntactic binder prefix fi: (4.62) let 0 be a function from assignments to elements in De T (where r can be any type), then Bn{) = kgkx. x])(x) There is nothing essentially new in this definition. Our BIER from above, for example, is equivalent to the following: (4.63) Binder Index Evaluation rule, notational variant: We now assume that the binder prefix is inserted on the a(rgument)-list of the head of an n-ary branching structure. The a-list of introduce in introduce every soprano^ to heri partner, for example, looks like (4.64); no /3 is adjoined in the syntactic tree: (4.64) (NPnom, NPacc, /32, [pp to NP]> We then replace (4.57) with the following rule: (4.65) n-ary Function Application with Binding: Let X°'s argument-list be (Aq, ..., A;, j3n, Aj, x then A"... (Bn(lX°J + lAmJ +...+ |[Aj]| )+[Ai] +... +[Ai] )(g), where P +A =def Xg.P(g)(A(g)) if defined, Xg.C(P(g), A(g)) otherwise (to replace [4.58] above) Exercise 4.10 Convince yourself of the equivalence between (4.63) and the original (4.12) by calculating through a simple example such as every girl painted her desk. Exercise 4.11 Give a complete calculation for (4.60a) and (4.61a). 7 Where 0 x](8)'s room} = ... is in {x | x should stay in x's room} Crucially, since this ordinary meaning in (5.11a) has the pronoun bound by the local subject, the meaning of the embedded VP is [kx.x should stay in x's room], the characteristic function of {x \ x should stay in x's room}. Given that the VP doesn't contain anything focused, the alternatives in the P-Set of VP and any constituent containing VP will be built around that property, i.e. the P-Set will look something like (5.12): (5.12) P-SET(7A7yA/VA8 fi% should stay in her% room): { Claudia is in {x \ x should stay in x 's room}, Marijana is in {x \ x should stay in x 's room}, Steven is in {x | x should stay in x's room}, Tatjana is in {x \ x should stay in x 's room}, • ••} The meaning of the matrix VP will now be the property of saying that Tatjana should stay in her room and not saying that anyone else should stay in their room. It is false if I also said 'Marijana, stay in your room,' but true even if I also said 1 Note that the alternatives to [TATJANAsP are individuals, not names (i.e. they are meanings, not words). Accordingly they don't bear indices and cannot enter into any systematic coreference patterns. 108 OTHER CASES OF SEMANTIC BINDING 'Mirko, stay in Tatjana's room' - the exact mirror image of the situation with LFj in (5.10). We see, thus, that the two readings of this sentence correspond to two different LFs, one of which has Tatjana semantically bind the pronoun her. For that to be possible, we have to allow the non-quantificational NP Tatjana to undergo Index Transfer, which is what we set out to show. Exercise 5.2 Calculate ordinary meanings and P-SETs for VP, SI, and S2 in (5.13), once under the assumption that Bill and his are coreferent, once under the assumption that the former binds the latter. Then give the truth conditions for the whole sentence in each reading (you don't need to give an LF): (5.13) We only [yp know that [si she told BILL where [32 his bicycle is parked]]]. Exercise 5.3 Above we assumed that Index Transfer needs to retain the index on the binder NP itself. This was argued for regarding example (5.1)/(5.3a), where Index Transfer applied to a pronoun. Show that a pronoun can indeed semantically bind another pronoun by constructing and discussing an example parallel to (5.6). Exercise 5.4 Use (5.14) to make an argument that Index Transfer needs to retain the index even on a full NP: (5.14) She only said that TATJANA would stay in her room. Exercise 5.5 Which of the following variants of (5.13) are ambiguous in the same way as (5.13); discuss why (not): (5.15) (a) We only know that SHE told Bill where his bicycle is parked. (b) We only know that SHE told BILL where his bicycle is parked. (c) We only know that she told Bill where HIS bicycle is parked. (d) We only know that SHE told Bill where her bicycle is parked. (e) We only know that she told BILL where HER bicycle is parked. Exercise 5.6 Consider the sentence / only knew that GANDALF had lost his mind. Give all possible LFs for this sentence and explain why these and only these are possible. Exercise 5.7 Consider a case of bare contrastive focus such as (5.16): 5.3 Double indexing 109 (5.16) I told KATHRIN to write down her name (not SABINA). (a) Give paraphrases that indicate the three different readings of (5.16) analogous to the different readings of (5.6) above. (b) Give the LFs for the non-parenthesized part on the different readings and classify them as coreferent/semantically bound, where applicable. (c) Sketch a treatment to interpret the parenthesized part that interacts in the desired way with your LFs above. 5.3 Double indexing We have established in section 5.2 that the Index Transfer rule, originally designed for use with QNPs, can apply to referring NPs, too. As noted in passing, the latter case requires us to amend the rule so that it copies the index on NP onto the binder prefix (rather than literally transferring it); otherwise the binder NP would wind up with no index whatsoever, which, in the case of a referential NP (unlike a QNP), leaves us with no way to interpret it (this is true in particular if the NP is a pronoun). In this section, we will formulate the final version of this rule, which will function yet differently from that: it will freely introduce an index on the binder prefix, retaining the index on the NP, if there is one: (5.17) Binder rule (final version): fim X The major difference to a rule that copies the index onto the /3 -prefix and then optionally deletes it on the NP is that (5.17) allows the index m on the binder prefix to be different from the index n on the binder NP. Why would this be useful? Consider example (5.18) below: (5.18) We only know that MARY parked her car in her garage. Consider this example in the context of a question like 'Did Mary and Sue both park their cars in Mary's garage?' In such a context, (5.18) as an answer would entail that (5.19) (a) we know that Mary parked Mary's car in Mary's garage, but (b) we don't know that anyone else parked their car in Mary's garage (in particular, we don't know if Sue parked Sue's car in Mary's garage) And, intuitively, (5.18) can serve to express just that. Let us now ask ourselves what an LF for (5.18) that expresses this reading would have to look like. A straightforward guess would be (5.20): (5.20) we only know that MARYi [fi \ parked heri car in heri garage] 110 OTHER CASES OF SEMANTIC BINDING In (5.20), Mary semantically binds the pronouns, which appears to be correct, given that we are not interested in a purely referential reading, in which we don't know if anyone else parked Mary's car in Mary's garage (this reading, to be sure, is generally available, but it is pragmatically odd in this particular example). But despite that, (5.20) is not what we are after, because it says (5.21) instead of (5.19): (5.21) (a) we know that Mary parked Mary's car in Mary's garage, but (b) we don't know that anyone else parked their car in their garage (in particular, we don't know if Sue parked Sue's car in Sue's garage) Again, this reading is available for (5.18), but it is not the one we are after. It turns out that in order to get the interpretation in (5.19), we need to have one pronoun be referential, but the other one be semantically bound. (5.22) is an LF that expresses that reading: (5.22) we only know that MARYi [f$2 parked her2 car in heri garage] Here the car, but not the garage, covaries with the focus alternatives to Mary, as desired. But this LF crucially assumes that the binder prefix next to Mary has an index different from the referential index on Mary. That is, (5.22) cannot be derived by copying the index in Mary onto the binder prefix; rather, we have to allow the binder prefix to introduce its own index. In this particular example, Mary sem-binds the first her, but corefers with the second. But the same effect can be seen in the more complex variant of this example in (5.23) below: (5.23) Almost every woman only admitted that SHE parked her car in her garage. Suppose that we are trying to get confessions of the form: 'Sue parked her car in my garage, too.' (5.23) can be used to report the futility of that endeavor. In order for that reading to arise, the LF must be as in (5.24): (5.24) almost every woman [fii only admitted that SHEi [^2 parked her2 car in heri garage]] Here, both occurrences of her are bound; extending our terminology we will say that shei and herj are co-bound (by almost every woman), while shej, mediated through the binder prefix /?2, binds the pronoun heri, as in (5.22). Formal definitions of these notions and the reformulated Binding Conditions will be provided in section 5.4. 5.4 A new system In this chapter we have established so far that any NP, quantifica-tional or not, can serve as a semantic binder, as witnessed by the interpretive options in focusing constructions (a second construction that leads to the same 5.4 A new system 111 conclusion will be discussed in section 5.5 below). This situation in some sense represents the zero-hypothesis. Index Transfer, or its successor, the Binder rule, is a syntactic operation available to any NP by the rules of our grammar. The opposite situation - that only QNPs can act as semantic binders - would have required us to make its applicability sensitive to the semantic type of the NP in question (referring vs. quantificational).2 The way things are stated now, all NPs are treated as equal, although the effect of semantic binding with referential NPs can be detected only in special contexts like the ones involving focus explored in this chapter. We formulated a new Binder rule, to replace our earlier Index Transfer rule, repeated here: (5.25) Binder rule (final version): fim X The BIER from chapter 4 remains the same. In chapter 4 we saw that the index on a QNP cannot be interpreted as a referential index, and we used the Index Transfer rule to strip it off such an NP. Since its successor, the Binder rule, doesn't remove indices any more, we need to amend our indexing conventions: (5.26) Indexing convention (preliminary): All and only non-quantiricational NPs bear an index.3 Where a QNP acts as a semantic binder, the Binder rule will introduce the prerequisite index; in all other cases, QNPs are indexless. As for the Binding Conditions, we will need a couple of reformulations. Note first that LFs derived using (5.25) do not correspond to any 'normal' surface indexings, the obvious reason being that the (indices on) binder prefixes no longer correspond to indices on NPs. Heim (1993) approaches this problem by assuming that NPs can bear two indices at s-structure: an inner index, corresponding to our referential indices, i.e. those found on NPs at LF; and outer indices, corresponding to our binder indices, i.e. those found on binder prefixes at LF. She then goes on to formulate the pertinent Binding Conditions on s-structures thus enriched. We will follow the opposite strategy here and formulate our binding conditions so as to apply at LF ([5.28] repeated from chapter 4): 2 A similar issue is discussed in the transformationalist literature regarding the question whether only quantificational NPs can undergo Quantifier Raising. Since there, Quantifier Raising is considered a prerequisite for semantic binding, our conclusion about semantic binding carries over to this question, contra e.g. May (1985). 3 Since no rules delete or add indices on NPs, nothing needs to be said about where in the derivation (5.26) applies. 112 OTHER CASES OF SEMANTIC BINDING (5.27) Syntactic binding (final): NP syn(tactically)-binds NP' if and only if (iff) (a) NP and NP' are coindexed (b) NP c/o/... -commands NP' If an NP is not syntactically bound by any NP in a phrase marker P, we say that NP is syntactically free in P. (5.28) Semantic binding (final): A binder prefix sem(antically)-binds an NP iff (a) and NP are coindexed (b) c-commands NP x (c) there is no binder prefix fi' which is c-commanded by and meets (a) and (b) If an NP is not bound by any binder prefix in a phrase marker P, we say that NP is semantically free in P. Binding Conditions A and C for English then say: (5.29) Binding Conditions (A) A reflexive pronoun must be semantically or syntactically bound in its domain. (C) A full NP must be semantically and syntactically free in the root domain. The (implicit) disjunctions4 in (5.29) are the price we pay for our new Binder rule, given that semantic binding no longer presupposes syntactic binding (in those cases where the binder is a QNP). We will address the issue of unifying these definitions in chapter 6. For the moment, we will leave it at this, and turn to Binding Condition B. By our new Binder rule, we get LFs in which a pronoun is (locally) semantically bound without being syntactically bound, (5.30a), and vice versa, (5.30b); furthermore it can end up being anaphorically related to a local NP without being locally sem- or syn-bound, as in (5.30c); all of these express intuitively unavailable readings and need to be blocked: (5.30) (a) *Jeannei ^2 saw her2. (b) *Jeannei sawheri. (c) *Jeannei ^2 thought that she2 saw heri. To predict all these stars by Binding Condition B, we need a notion that encompasses all these possibilities, a notion we will call codetermination:5 4 Note that 'semantically and syntactically free' is the same as 'not semantically or syntactically bound.' 5 A term from Heim (1993):233f.; the original insights behind double indexing and codetermination seem to go back to Higginbotham (1983):404 and 406. 5.4 A new system 113 (5.31) Codetermination: NP and NP' codetermined if any of the following holds: (a) they are coindexed (b) one semantically binds the other (c) there is an NP" such that NP and NP" are codetermined and NP" and NP' are codetermined We will revisit (and ultimately render obsolete) this rather motley set of sub-conditions in chapter 6. For the moment it allows us to formulate Binding Condition B as follows: (5.32) Binding Condition B: A non-reflexive pronoun must not be codetermined with a c-commanding NP in its domain. To see how this works, reconsider the unacceptable bindings in (5.30): Jeanne and her are illicitly codetermined in (5.30a) by (5.31b) because the former sem-binds the latter, and in (5.30b) by (5.31a) because they are coindexed (in this particular case, coindexing corresponds to coreference, but the same would apply to an example like Jeannei ft i thinks that shei saw her! in which it corresponds to being co-bound). Finally, in (5.30c), Jeanne and she are codetermined because the former binds the latter, Jeanne and her are codetermined because they are coindexed, and therefore she and her are codetermined by (5.31c), with NP" = Jeanne, triggering a Binding Condition B violation, as desired. This concludes the introduction of our new system. The novel elements, in a nutshell, are that referential NPs can act as semantic binders, and that we have double indexing, i.e. an NP with referential index n binding a different index m. Since, by the latter innovation, an NP can semantically bind a pronoun without actually syntactically binding it, we had to reformulate our Binding Conditions so as to apply at LF, after the Binder rule, and take into consideration both syntactic and semantic binding, as well as any combination thereof (codetermination). It may seem that these reformulations are a high price to pay for double indexing, and that we should have made more of an effort to avoid it. Readers who feel that way might be consoled to know that the system as set up in this section is only a transitional one, and will be replaced by a more parsimonious one in chapter 6. In the discussion to follow, it will often be useful to present an example and mark two NPs as anaphorically related without committing to the exact nature of that relation (coreference, semantic binding, or co-binding). Where this is intended, I will henceforth simply italicize the NPs in question, without adding any indices or binder prefixes. Exercise 5.8 Reformulate Binding Conditions A and C using the notion of co-determination. Are there empirical differences? Argue! 114 OTHER CASES OF SEMANTIC BINDING 5.5 Verb phrase ellipsis 5.5.1 Strict and sloppy identity i A famous argument for non-quantificational NPs as semantic binders comes from Verb Phrase Ellipsis (VPE). Observe that (5.33) can have two kinds of interpretations: (5.33) Every woman in Culver City hates her neighbor, but no woman in Los Feliz does. On the first, her refers to some contextually given female person, call her Ms. Jones, whose neighbor is universally hated by Culver City women, but not Los Feliz women. Call this a strict reading. On the second reading, every woman in Culver City hates her own neighbor, whereas no Los Feliz woman hates the people she herself lives next door to; call this reading sloppy. Obviously, on the first, strict, reading the pronouns in the overt VP and in the elided VP are both referential (since no NP referring to Ms. Jones syntactically binds them), whereas on the second, sloppy, one, they are each sem-bound (since their binder is a QNP). It is, however, impossible to 'mix and match' these two options. (5.33) cannot mean that all Culver City woman hate Ms. Jones' neighbor, while no Los Feliz woman hates her own neighbor, nor vice versa. Likewise, if neither pronoun is bound, they actually need to corefer. There is no reading for (5.33) on which Ms. Jones' neighbor is hated in Culver City, while, say, Ms. Smith's neighbor is well liked in Los Feliz. We can capture this in the following generalization: (5.34) VP-Ellipsis generalization: If a pronoun p in the antecedent VP is semantically bound, the corresponding pronoun p' in the elided VP must be semantically bound in parallel. If p is referential, p' must corefer with it. Crucially, (5.34) applies to non-quantificational NPs as well. Consider (5.35): (5.35) Felix hates his neighbor, and Max does, too. Here again, we find a strict reading, in which Felix and Max both hate some third guy's, say John's, neighbor, and a sloppy reading, in which Felix hates Felix's neighbor, and Max hates Max's. But, again, no mixed readings are possible (e.g. Felix hates John's neighbor, and Max Bill's; or Felix hates Max's neighbors, and Max Felix's; or Felix Felix's and Max Bill's ...). There is one special reading, though, on which Felix's neighbor is hated by Felix and Max. This, however, turns out to be just another instance of the strict reading on closer inspection, one where Felix and his happen to corefer. The three readings can be represented as follows (where strikeout marks the elided VP): 5.5 Verb phrase ellipsis© 115 (5.36) (a) Felixi hates MS3 neighbors and Max2 does hate hiS3 neighbors, too (b) Felixi hates hisi neighbors and Max2 does hale hisi neighbors, too (c) Felixi fi\ hates hisi neighbors and Max2 P2 does hate MS2 neighbors, too (strict, 3rd party) (strict, coreference) (sloppy, sem-binding) This pattern exactly follows the generalization in (5.34): either both pronouns are bound, or both are referential. The generalization only captures (5.35) though -and this is the heart of the argument - if it is assumed that Felix and Max can function as semantic binders, as in (5.36c). Suppose we assumed instead that non-quantificational NPs canned act as semantic binders; then sloppy identity in (5.35), unlike in (5.33), would have to be captured by allowing the elided pronoun to have a different referent from the one in the overt VP if and only if they are both coreferent with an antecedent in their respective conjunct. While such a condition is not unimaginable, it seems preferable to avoid it, if the alternative hypothesis - that sloppy identity always involves semantic binding - is viable. Exercise 5.9 Why doesn't example (5.33) have a strict reading analogous to the 'Felix hates Felix's neighbors, and Max hates Felix's neighbors too' reading for (5.35)? Exercise 5.10 One plausible idea to capture the generalization in (5.34) is to say that the elided VP has to be syntactically identical to the overt one. This identity is warranted in the strict reading (5.36b) for VP = hate(s) hisi neighbors, and in the sloppy reading if we assume the LF in (5.37) instead of (5.36c) above, so that the VP in both conjuncts is ft1 hates hisi neighbors: (5.37) Felixi fi\ hates hisi neighbors and Max2 fi\ hales hisi neighbors, too In other words, the strict/sloppy alternation is a function of whether the elided VP contains a ft or not (this is, in fact, more or less the solution put forth in Sag [1976]). 1. Discuss the details of the LF in (5.37). In particular, address issues regarding (i) auxiliary verbs, and (ii) the choice between Index Transfer and the Binder rule. 2. Discuss the implication that examples like (5.38) (from Jacobson [1992]:206f.) have for this idea. (5.38) Tom wanted Sue to water his plants, while John wanted Mary to water his plants. 116 OTHER CASES OF SEMANTIC BINDING 5.5.2 Deriving the generalization In the discussion so far we implicitly assumed that pronouns in VP ellipsis can be freely indexed and bound, with the resulting structures subject to the following filter, repeated from above: (5.39) VP-Ellipsis generalization: If a pronoun p in the antecedent VP is semantically bound, the corresponding pronoun p' in the elided VP must be semantically bound in parallel. If p is referential, p' must corefer with it. Let us ask now whether (5.39) can be derived in a systematic way. Let us start with the second half of the generalization. Can we derive the requirement that referential pronouns in the two VPs must be coreferent? An obvious idea is to have a syntactic condition on VP ellipsis that requires that the two VPs must be syntactically identical, including indices. From this it will follow that the pronouns are coindexed. This, however, only guarantees that they corefer if neither of them is bound. As pointed out above, we need to exclude the possibility that one of them is semantically bound, while the other is referential. That is, we have to exclude LFs as in (5.40) below: (5.40) (a) * every woman in Culver City $2 hates her2 neighbors, but no woman in Los Feliz does hate her2 neighbors (b) * Felix 1 $2 hates his2 neighbors, and Max3 does hale his2 neighbors, too In both of these LFs, the pronouns are coindexed, but the elided pronoun is referential, referring to some contextually given g(2), while the overt one is bound. This is excluded by the original generalization (5.39), but would wrongly be allowed if syntactic identity was the only condition on VP ellipsis. Let us call such LFs pseudo-strict. We can get rid of pseudo-strict readings, and thus derive the second half of (5.39) from the strict syntactic identity, if we add the following condition:6 (5.41) No Spurious Coindexing: No LF representation may contain both semantically bound and free occurrences of the same pronoun. (5.41) effectively requires that indices used to express semantic binding are a disjoint set from those used to express (co)reference. It straightforwardly rules out the pseudo-strict readings in (5.40) since the index 2 is bound in the first conjuncts, but free in the second. For the general case it requires that the overt pronoun and the elided pronoun - since they bear the same index - be either both free, i.e. (co)referential, or both bound, yielding sloppy readings. 6 From Heim and Kratzer (1998):254; a constraint along these lines is also proposed in Heim (1997) :202, for more or less the same reasons; see the discussion in Heim and Kratzer (1998) :9.3.2. 5.5 Verb phrase ellipsis© 117 Let us turn to the first half of the generalization in (5.39), bound readings, then. Note that even if we require strict syntactic identity, we can still derive sloppy readings, given that we are allowed by the new Binder rule to use the same index on different binders. So we can derive sloppy readings of our standard examples, including the one discussed in section 5.5.1 above, by LFs such as (5.37), repeated here, and (5.42b): (5.42) (a) Felixi ^3 hates MS3 neighbors, and Max2 fi 3 does hate MS3 neighbors, too (b) Tonii ^2 wanted Sue to water his2 plants, while John3 ^2 wanted Mary to water his2 plants. So strict syntactic identity is compatible with sloppy readings, and, together with (5.41), No Spurious Coindexing, it will derive (5.34). We will return to these issues in section 6.5.1 of chapter 6, where we will weaken the strict syntactic identity condition on VP ellipsis. Exercise 5.11 In section 5.3 above, the possibility of double indexing was motivated using examples (5.18) and (5.23). 1. Make a parallel argument from the realm of ellipsis using (5.43) instead of (5.18): (5.43) Mary parked her car in her garage, and Lee did, too. 2. Find a variant of (5.43) which is parallel to (5.23) and discuss those aspects of it in which that variant differs from (5.43). 6 The Coreference rule 6.1 The proposal In chapters 4 and 5 we introduced the central distinction between coreference and binding. We have formally implemented the notion of a semantic binder by introducing binder prefixes, i.e. indexed elements adjoined not to the actual NP but to the sister of the binding NP. We saw that, in a number of constructions, the question of whether coindexing among two NPs (with one c-commanding the other) is to be interpreted as coreference or as semantic binding can be answered on semantic grounds: • If the higher NP is quantified (or a w/z-expression), it semantically binds the lower NP. • Embedded under focus-sensitive particles like only, the corefer-ence/semantic binding distinction leads to different truth conditions. • In the context of VP ellipsis, the coreference/semantic binding distinction leads to different resolutions for the elided constituent (strict versus sloppy). Nothing in our system, however, allows us to decide in a run-of-the-mill case, like (6.1), whether the anaphoric relation between the italicized NPs is to be interpreted as coreference, as in LF (6.1a), or as semantic binding, as in LF (6.1b): (6.1) John thinks he is sick. (a) Johni thinks hei is sick (b) John i thinks hei is sick In Reinhart (1983a):ch. 7 it is proposed that in such a configuration, coindexing has to be interpreted as semantic binding. Following Grodzinsky and Rein-hart (1993):79 and Heim (1993):209 we can render the pertinent condition as follows:1 1 Grodzinsky and Reinhart (1993) call this rule Rule I. The original condition in Reinhart (1983a):167 is: "Where a syntactic structure you are using allows bound-anaphora interpretation, then use it if you intend your expression to corefer, unless you have some reason to avoid bound anaphora." We will return to the particulars of this formulation below. 118 6.2 Theoretical consequences 119 (6.2) Coreference rule (CR) a cannot corefer with if an indistinguishable interpretation can be generated by replacing a with a variable bound by fi. Applied to (6.1), a is he and /3 is John. In (6.1a), a and /3 corefer, in (6.1b), a is semantically bound by fa Since both sentences have the same interpretation, only (6.1b), the LF with semantic binding, accords with the CR (6.2); (6.1a) violates it, and since the CR is a principle of grammar, (6.1a) is ungrammatical. We say that the bound variable construal (6. lb) blocks the coreferent construal (6. la) by virtue of the CR. At first glance, the CR in (6.2) rules out coreference in all those cases in which we couldn't tell the difference from a bound variable construal anyway. In other words, it removes certain spurious ambiguities, but never seems to make any novel predictions. Is this really all there is to it? In the following sections we will discuss cases in which the CR yields significant simplifications in our theory, and moreover yields different empirical predictions. 6.2 Theoretical consequences 6.2.1 Eliminating codetermination i In chapter 5 we saw that our new device of double indexing forces certain complications in the formulation of the binding principles. In particular, we saw that an additional notion, codetermination, needed to be defined to block anaphoric relations between the subject and the object in a simple clause like (6.3), given that any of the LFs in (6.3a)-(6.3c) could yield such a reading and hence needs to be excluded by Binding Condition B: (6.3) she saw her (a) *shei fa saw her2 (b) * shei saw her] (c) * Jeannei fa thought that she2 saw her] Using the CR, we can actually alleviate the need for this notion. To see how, note first that (6.3b) is ruled out by the CR: she (=fa and her (=a) corefer, where an indistinguishable reading could be achieved by semantic binding as in (6.3a); therefore, (6.3a) blocks (6.3b) by virtue of the CR. The star on (6.3b) is thus explained independent of Binding Condition B. Next consider (6.3c): here, Jeanne (=fa and her (=a) corefer, instead of the former binding the latter. A structure in which a is replaced by a variable bound by p is (6.4): (6.4) Jeannei fa thought that she2 saw her2 120 THE COREFERENCE RULE (6.3c) above and (6.4) have the same interpretation, so the latter blocks the former, which means that (6.3c), too, is rendered ungrammatical by the CR, independent of Binding Condition B. So far so good, but note that she (=/3) and her (=a) in (6.4) are coindexed, and neither sem-binds the other. Let's assume for the moment that this configuration counts as 'coreference' in the sense of the CR, too, and must therefore be avoided, preserving meaning, in favor of the LF in (6.5): (6.5) Jeannei ^2 thought that she2 ^3 saw her3 Here, Jeanne sem-binds she, which in turn sem-binds her (the choice of actual indices is irrelevant, of course). None of the NPs corefers, but the interpretation is still the same. Therefore (6.5) blocks (6.4) and (6.3c) by the CR. The only LFs allowed by the CR are thus (6.3a) and (6.5); all others are blocked. But these two involve bona fide local violations of Binding Condition B: a pronoun is bound within its local domain. Moreover, it is semantically bound in both cases. We can thus do away with the notion of codetermination and simply state Binding Condition B as in (6.6): (6.6) Binding Condition B (final version) A non-reflexive pronoun must be semantically free in its domain. We thus see that the CR, apart from blocking spurious binding/coreference ambiguities, yields a significant simplification of our Binding Condition B. Before going on, we have to amend the Coreference rule in (6.2) slightly. The intended effect of the CR in an example like Jeanne thought that she saw her (=[6.3c] above) was to force local semantic binding wherever possible. There is an LF for this sentence, however, that has no coreference in it but still involves non-local binding: (6.7) Jeannei f$2 @3 thought that she2 saw her3 She and her are co-bound, but aren't coindexed. But, with neither coreference nor coindexing, the CR, even on its most favorable interpretation, cannot block this LF. Still, (6.7) has an interpretation in which she and her have the same referent, i.e. one that is intuitively unavailable. This was blocked under our previous formulation of Binding Condition B, since she2 and her^ are codetermined in (6.7), but it is no longer under the simplified (6.6), since herj, is not semantically bound in (6.7). How can we block this LF, then? One possibility is to replace '... cannot corefer...' in the formulation of the CR by '... cannot corefer or be co-bound by the same NP...'. Then (6.7) is blocked by (6.5) in a way analogous to the way that (6.4) was. Another possibility is to disallow by stipulation one NP from binding two indices at once (i.e. two /3 s next to one another), as Jeanne does in (6.7). Then 6.2 Theoretical consequences 121 the only way to get her to be sem-bound by Jeanne is again (6.4), which we already saw is blocked by (6.5). Both these amendments, however, would fail to block the more complex example in (6.8), which involves neither coreference nor co-binding nor double /3 s, but poses the exact same problem in the lowest clause as (6.5): (6.8) Jeannei /J 2 said that she2 ^3 thought that she2 saw her3 Let us, then, invoke an additional constraint that prohibits non-minimal binding of a pronoun. The idea is that Jeannei'shei mustn't sem-bind herj, in (6.7)/(6.8) (or [6.4] for that matter) because she could minimally bind her as in (6.5) with the same interpretation. Such a principle is proposed as Rule H in Fox (2000):ch. 4:2 (6.9) Rule H A pronoun, a, can be bound by an antecedent, 8, only if there is no closer antecedent, y, such that it is possible to bind a to y and get the same semantic interpretation. The CR and Rule H in tandem enforce something like a 'closest antecedent requirement'; the former forces semantic binding among c-commanding NPs that are supposed to receive the same interpretation, and the latter forces semantic binding among more than two c-commanding NPs to be strictly transitive, local, and non-overlapping. Since the closest binder structure is automatically the one most likely to violate Binding Condition B, we can thus stick with the much simpler version of it given in (6.6). Despite this successful elimination of the notion of codetermination (and, effectively, co-binding), we are left with the somewhat unsatisfactory division of labor between the Coreference Rule and Rule H, which moreover formally resemble each other. A rule which subsumes them both under one roof is (6.10): (6.10) Have Local Binding! For any two NPs a and fi, if a could bind (i.e. if it c-commands and is not bound in a's c-command domain already), a must bind fi, unless that changes the interpretation In a sentence like Jeanne thought that she saw her, the only way in which (6.10) can be met for all pairs of (relevant) NPs is indeed the binding pattern in (6.5), i.e. the one that violates Binding Condition B, as desired. In (6.7), for example, her is free within the c-command domain of she and could thus be bound by it. Since this will yield an indistinguishable interpretation, it must be bound by it. The resulting LF has her locally bound by she, and Jeanne vacuously binding the index 3, which is equivalent to (6.5). Similarly in (6.4) and (6.3c), where she fails to bind her, the necessary insertion of ft 2 (Pi m [6.3c]) right next to she will yield another LF equivalent to (6.5). 2 Fox (2000):11 Iff. attributes his Rule H (though not by this name) to Heim (1993), but it is unclear to me that Heim's (1993) proposal, which, unlike Fox's, distinguished between semantic and syntactic binding, actually entails something like Rule H (unfortunately, Fox doesn't provide any page references). 122 THE COREFERENCE RULE Officially, we will adopt (6.10), which subsumes the CR and Rule H. I will, however, continue to refer to the CR in the arguments that follow, where the minimality of pronoun binding isn't at issue. Exercise 6.1 Give a step-by-step treatment of examples (6.3), (6.4), and (6.8) using the rule in (6.10). 6.2.2 Eliminating Binding Condition C i In Reinhart (1983a), the CR (or its counterpart there) is used to eliminate yet another part of BT, namely Binding Condition C. To see how, note first that the CR blocks the following LFs for sentences like He/John likes John or He/John likes John's mother. (6.11) (a) *hei/Johni likes John] (b) *hei/Johni likes Johni's mother The reason is that we could replace John by a variable sem-bound by he, in accordance with the CR: (6.12) (a) hei/Johni ^2 likes himself2 (b) hei/Johni ^2 likes his2 mother Obviously, these LFs receive the same interpretations as those in (6.11), but avoid coreference in favor of sem-binding (note that the choice between him and himself is determined by Binding Conditions A and B, as before). We have thus successfully ruled out the Binding Condition C violations in (6.11) without invoking Binding Condition C, but just the CR. One question left open by this is whether we couldn't also sem-bind John, as in (6.13): (6.13) (a) hei/Johni ^2 likes John2 (b) hei/Johni ^2 likes Johns's mother Obviously, if we want to do away with Binding Condition C, we have to block this possibility. There is now a rather radical way to do so, by adopting the following: (6.14) Full NPs don't carry referential indices. We have already adopted something like this for quantified NPs in chapter 4. How will the generalization proposed here work? First, note that we never used referential indices on full NPs to determine their reference. That, by assumption, was done by the lexical content of the NP. Indeed, we had to go somewhat out of our way actually to make the indices on full NPs do something at all. Adopting (6.14), then, does away with all of that: since there are no indices on full NPs, there is no need to have anything but the lexical content enter into their interpretation. 6.2 Theoretical consequences 123 Second, indices on full NPs were never used to signal that they were bindees, since full NPs couldn't be bound due to Binding Condition C. In other words, although full NPs were technically bindable in the same way that pronouns are, this option was effectively blocked in the syntax. Doing away with referential indices on full NPs altogether resolves this situation: full NPs cannot be bound because they don't have a bindable index in the first place. Third, indices on full NPs were required to induce Binding Condition B violations in a sentence like Peter shaved him. But this is taken care of by a conspiracy now: the LF in (6.15a) is ruled out by the CR, because the same interpretation could be gotten by Peter binding him, as in (6.15b). But that LF is ruled out because it constitutes a Binding Condition B violation: him is bound in its local domain, not syntactically by a referential index, but semantically by ft1: (6.15) (a) * Peter shaved himi (violates CR) (b) * Peter fi\ shaved himi (violates Binding Condition B) In sum, we see that the CR allows us to do away with referential indices on full NPs, and, accordingly, with Binding Condition C. In closing this section, let me point out that it wasn't claimed that it makes no sense to bind a full NP. On the old account, Kareni was interpreted as g(2), with the presupposition that g(2) was Karen. Thus a sentence like every girl ft 2 voted for Karen2 would have received an interpretation like 'for every x, if x is a girl, then x is Karen and voted for herself.' Granted, this sentence can only be uttered in a situation where there is only one girl, Karen; but it certainly is an imaginable proposition that would be expressed here.3 The case seems even clearer for descriptive NPs or epithets like the idiot: there is no reason whatsoever why / told every boy ft 5 that the idiots should stay home wouldn't denote the perfectly reasonable proposition that I told every boy that he is an idiot and should stay home. That none of these readings exists is thus a fact about the way full NPs happen to function in natural language. Unlike with QNPs, there is no logical reason why definites and names couldn't be bound, but it seems that they just can't. If we assume that they do not bear referential indices, we have a natural way of capturing this fact (more natural, arguably, than imposing an extra condition such as Binding Condition C), but I wouldn't claim that we have explained it in any deeper sense. It has been reported in passing in the literature that some languages like Thai or Vietnamese, and possibly even English (see Lasnik [1986], and the references at the end of chapter 1, section 1.2.3), don't display Binding Condition C effects:4 3 Note that an NP like her^ bicycle is usually interpreted as 'the bicycle of g(2)' with the presupposition that g(2) has a bicycle. Every girl P2 rode her^ bicycle is a perfectly well-formed sentence, and it is usually understood to quantify only over girls who own bicycles. So it wouldn't be unreasonable even to expect that every girl P2 voted for Karen^ would generally denote something like 'every girl who is Karen voted for herself,' which is an even more sensible proposition than the one mentioned in the main text. 4 Data from Lasnik (1986):153. 124 THE COREFERENCE RULE (6.16) (a) coon kbit waa coon chalaat (Thai) John thinks that John is smart (b) John tin John s e tharig (Vietnamese) John thinks John will win Does the reduction of Binding Condition C to the CR allow for such languages? It does, and in several ways. One simple possibility is that the CR is language particular (I deliberately refrained from claiming above that the CR follows from any general maxims of cooperative communication, as is sometimes done); if a languages doesn't have it, it shouldn't show Binding Condition C effects (but it also, crucially, should lack the other empirical effects of the CR discussed in the present chapter). Another possibility is that the CR in such languages lacks, as it were, the 'replaced by a pronoun' part of our formulation in (6.2), that the CR in such languages prohibits a from coreferring with only if a could itself be a bound variable instead (rather than be replaced by one), i.e. it blocks pronouns coreferential with a c-commanding NP, but never full NPs. A third possibility is that some languages do have indices on full NPs, so that full NPs can, in fact, be semantically bound. In all these cases, there might additionally be independent restrictions on full NPs (it seems, for example, that speakers of Thai and Vietnamese, as well as those of English who allow sentences of the 'full NP... full NP' variety, still reject 'pronoun... full NP' under c-command). There hasn't been enough research on languages that appear to lack Binding Condition C effects to determine which (if any) of these options may indeed occur. Exercise 6.2 Explain step by step how anaphoric relations are blocked between pronoun and full NP in The dog howls at it and It howls at the dog. 6.2.3 Upshot: BT regards semantic binding only i Let me point out once more the arguably most interesting consequence of adopting the CR, with or without indices on full NPs: all Binding Conditions now regard semantic binding only. This might not be obvious at first, but the reasoning is quite simple: first, the Binding Conditions refer to pairs of c-commanding NPs only (where by 'c-commanding NPs' I mean NPs one of which c-commands the other[s]); second, Binding Conditions aside, c-command is a sufficient condition for semantic binding; third, by the CR, coreference is blocked by semantic binding, wherever the latter is possible, which means: among c-commanding NPs, which means, among all pairs of NPs for which Binding Conditions are relevant. Another way of saying this is that only binding relations of the form fin ... NPn are relevant to BT. Or, put yet the other way around: (6.17) Consequence of the CR: (semantically) unbound indices are irrelevant to BT. To illustrate, consider (6.18): 6.2 Theoretical consequences 125 (6.18) (a) Nina's mother visited her2. (b) Heri/2 mother visited her2. Both sentences allow coreference between the possessive Nina's/her and the object pronoun (note that binding is not an option, since the possessive doesn't c-command the object). However, in the pronoun case (6.18b) this is indicated by coindexing, while in the name case (6.18a) it isn't, and can't be, given that names no longer bear indices. But, even in the case of (6.18b), the indices on the two occurrences of her are irrelevant to the BT, as claimed in (6.17): since neither pronoun c-commands the other, Binding Conditions won't apply to them. Reinhart (1983a) proposes therefore that the pronouns in (6.18) shouldn't bear indices either. Rather, being referring pronouns, their reference should be determined by principles outside of core grammar, i.e. pragmatic principles. The only anaphoric relation that is encoded in the grammar proper, by indexing, is that of semantic binding. Reinhart thus advances the following radical hypothesis, which subsumes our earlier (6.14): (6.19) Reinhart's (1983a) hypothesis: Only semantic binding is represented in the syntax. Coreference is not. Implementing this hypothesis into our current system amounts to adopting the following conditions: (6.20) (a) NPs may, but need not, bear an index. (b) All indices on NPs must be sem-bound at LF. Unsurprisingly, adopting (6.20) doesn't change the predictions made about the sentence in (6.18). With or without indexing, the two NPs may, but need not, corefer. Indeed, it doesn't change the predictions about the core BT cases either, again thanks to the CR. Consider a simple case like (6.21): (6.21) (a) she likes her (no coreference possible by CR) (b) * she ^2 likes her2 (out by Binding Condition B) While it is now possible to have a structure devoid of any indexing, as in (6.21a), the two pronouns, unlike those in (6.18b), are not allowed to corefer, due to the CR: since semantic binding is possible in this configuration, any coreference is blocked. True semantic binding, on the other hand, introduces indices as in (6.21b), which in turn betray the Binding Condition B violation. So, even if we adopt (6.19)/(6.20), we still block pronouns with local antecedents by a conspiracy of the CT and Binding Condition B. Reinhart's conjecture that, once we adopt the CR, indices on referential NPs are irrelevant for deriving BT-violations and thus superfluous syntactic objects is certainly correct. It leaves us with the task of interpreting referential NPs without the help of indices, however. This, as noted above, seems simple enough for full NPs, but far less obviously so for referential pronouns. We will add some comments about this in section 6.5.2 below. For the moment, let us register that all Binding Conditions will refer to semantic binding only, and that the choice 126 THE COREFERENCE RULE to maintain referential indices on some (or perhaps all) referential NPs would merely be motivated by semantic considerations. 6.3 Binding Theory obviations So far we have reviewed a number of theory-internal changes that result from the CR. But there are empirical differences as well. Note that all our formulations of the CR invoke the notion of 'indistinguishable interpretation.' Why haven't we simply said that an NP cannot corefer with another NP if it can be bound by it? The reason is that, in a few cases, coreference and binding do not yield indistinguishable interpretations. An example are the focus constructions discussed in chapter 5: (6.22) I only said that TATJANA should stay in her room. (a) coreference: I only said that TATJANA should stay in heri room entailment: I didn't say anything else of the form 'x should stay in g(l)'s room' = I didn't say anything else of the form 'x should stay in Tatjana's room'5 (b) binding: I only said that TATJANA i should stay in heri room entailment: I didn't say anything else of the form 'x should stay in x's room' Here, her is a and Tatjana is f> in the sense of the CR; f> corefers with a in (6.22a), and binds a in (6.22b). But, as we discussed at length in chapter 5, the two LFs yield distinct interpretations: (6.22a) is true if I also said 'Marijana, stay in your room', but (6.22b) is false then. (6.22b) is true even if I said 'Mirko, stay in Tatjana's room,' but (6.22a) is false then. Since the two LFs don't yield the same interpretation, CR doesn't apply here and neither LF blocks the other, i.e. both are grammatical, correctly predicting the ambiguity. This ambiguity, however, would have been predicted, too, if we had never adopted the CR in the first place, so this case doesn't provide an argument in favor of 'the CR. But, as Reinhart (1983a) points out, in exactly those cases where the CR allows for exceptional coreference among c-commanding NPs, Binding Conditions seem to be suspended. Consider for example, the following cases of what might be thought Binding Condition C violations: (6.23) (a) Only JOHN thinks John is smart, (b) Even LARA voted against Lara. The argument here is straightforward: a grammatical LF for e.g. (6.23a) is (6.24): (6.24) only John? thinks John is smart interpretation: no one other than John thinks that John is smart (6.24) does not violate any Binding Conditions, since it doesn't contain any semantic binding (this holds even if we index names, since this will be mere 5 Assuming that g(l)= Tatjana. 6.3 Binding Theory obviations 127 coreference). But shouldn't it be ruled out by the CR? The CR wants us to compare this representation to (6.25): (6.25) only John? ^2 thinks he2 is smart interpretation: no one other than John considers himself/herself smart Here the lower NP has been replaced by a bound pronoun. But crucially, as indicated below the representations, the interpretations aren't identical, so the CR doesn't apply; (6.25) does not block (6.24), and both (6.25) and (6.24) are predicted to be grammatical. Generally, an NP can occur c-commanded by a coreferring NP if coreference yields a different interpretation from sem-binding. If this is the case, Binding Conditions do not apply, and we find obviations (though not literally violations) of Binding Conditions B and C. As Heim (1993) points out, the same argument can be made for cases of exceptional co-binding. According to her, examples like (6.26) are slightly more complex than, but generally as acceptable as, (6.24): (6.26) Everyone feared that only he voted for him. Why doesn't it have to be ... that only he voted for himself! The LF for (6.26) is (6.27): (6.27) everyone 1 feared that only hei voted for himi This LF doesn't violate Binding Condition B because him, although coindexed with hei, is not semantically bound by it. If we bind it locally, we get (6.28) instead: (6.28) everyone 1 feared that only hei P2 voted for himself2 (Note that ft 2 voted for him2 would be a sem-bound reading, too, but blocked by Binding Condition B.) This LF, however, expresses a different meaning from that of (6.27). The former says that everyone feared 'No one but me voted for me!'; the latter expresses the fear that 'No one but me voted for themselves.' Given the non-synonymy of these sentences, the CR remains silent here and both structures are allowed, each with its respective interpretation. These cases present strong empirical evidence for the CR or something like it.6 A similar class of examples is constituted by what Heim calls 'when structured meanings matter':7 (6.29) (a) I know what John and Bill have in common. John thinks that Bill is terrific and Bill thinks that Bill is terrific, (b) Look, fathead. If everyone loves Oscar's mother, then certainly Oscar must love Oscar's mother. 6 Heim's (1993) proposal doesn't, in fact, use the CR but rather allows for violation of the Binding Conditions in exactly those cases in which a non-distinguishable reading results. 7 Both examples from Evans (1980):356. 128 THE COREFERENCE RULE The reasoning here is, informally, that the rhetorical effect of these examples rests on ascribing the same property to John and Bill, and to Oscar and the rest. While, say, Oscar loves Oscar's mother and Oscar loves his (own) mother denote the same proposition, loves Oscar's mother and fii loves hisi mother do not denote the same property. If we understand these two as 'distinguishable interpretations' in the sense of the CR, we correctly predict cases like (6.29) to be acceptable (for closer examination of these cases cf. Heim [1993], especially pp. 216ff.). Reinhart's and Heim's works contain more examples of exceptional anaphora, which we will not go into here. One case that is often subsumed here, too, is that of unknown or mistaken identity, which we will discuss separately in 7.2. What we saw in this section is that the CR doesn't generally block corefer-ence among c-commanding NPs, but only if the same reading can be expressed via binding. While this justifies the particular way the rule is formulated, and provides a strong empirical argument in its favor, it also highlights one of its noteworthy, and to some worrisome, properties: The CR, Rule H, and Have Local Binding are all genuinely transderivational constraints. They don't just (dis)allow a particular structure by checking certain properties of that structure, but by comparing different structures within a certain set of 'similar' structures. None of the other rules and conditions employed in this book is transderivational, and some researchers regard the inclusion of transderivational constraints as a problematic step (see e.g. Potts [2001] for recent discussion). Exercise 6.3 Construct and discuss an example of exceptional coreference involving a focus construction with an element other than only. Exercise 6.4 What is remarkable about the example Only MILLI brought her own lunch. What is the reason? Speculate! Exercise 6.5 Do we find exceptional coreference in Only MILLI talked to her teacher! Exercise 6.6 The following statement is false: 'Binding Condition obviations only occur if there is no way to have a coreferent reading without violating a Binding Condition.' Show why! (Hint: look at exceptionally coreferring full NPs.) 6.4 Summary: the final system In this chapter we have discussed the consequences of adding Rein-hart's CR, or its generalization in (6.10), repeated in (6.32) below, to our theory. 6.4 Summary: the final system 129 This constraint, which encodes a general preference for local binding over coref-erence and codetermination, immediately yields two consequences: one theory-internal, one empirical:8 • Binding Conditions can be formulated using the notion of semantic binding only. There is no need to make reference to the more complex notion of codetermination, or to syntactic binding, in addition. • We get an account of exceptional coreference and exceptional co-binding, including the Binding Condition obviations that can go along with them. In addition, adopting (6.10) allows for further simplifications in the theory: • We can eliminate indices on full NPs, and, consequentially, Binding Condition C. • We can eliminate unbound indices altogether, simplifying the syntactic representations and - as we will see momentarily in section 6.5.1 -alleviating the need for the ban against spurious coindexing ([5.41] in chapter 5, section 5.5) in the treatment of VP ellipsis. These last two steps are logically independent of the first two, and one might choose to refrain from them on independent grounds (see e.g. the arguments in Lasnik [1986] against dropping Binding Condition C, and the discussion in 6.5.2 regarding referential indices on pronouns). For the remainder of this book, we will adopt (6.32) and the assumption that full NPs are indexless. We will stop short of implementing Reinhart's full proposal, though, and continue to index referential pronouns, for reasons of perspicuity (see the remarks in 6.5.2 below). A summary of our conventions and definitions is given below: (6.30) Indexing Convention (final) All and only pronouns (and binder prefixes) bear an index. (6.31) Binding Conditions (final) (A) A reflexive pronoun must be semantically bound in its domain. (B) A non-reflexive pronoun must be semantically free in its domain. (6.32) Have Local Binding! For any two NPs a and fi, if a could bind (i.e. if it c-commands and is not bound in a's c-command domain already), a must bind fi, unless that changes the interpretation Semantic binding is defined as per chapter 4, section 4.2 (and syntactic binding, if required, as per chapter 5, section 5.4). In addition to these definitions, we continue to use the Binder rule from chapter 5, section 5.3, the Binder Index Evaluation Rule (BIER) from chapter 4, section 4.1, and the Prohibition against Accidental Coreference (PACO) from chapter 2, section 2.2.2: 8 In addition, (6.32) has welcome consequences for the analysis of various 'many pronoun puzzles' in VP ellipsis, as discussed in Fox (2000):ch. 4. 130 THE COREFERENCE RULE (6.33) (6.34) (6.35) (6.36) Semantic binding (final): A binder prefix sem(antically)-binds an NP if and only if (a) and NP are coindexed (b) c-commands NP (c) there is no binder prefix fi' which is c-commanded by and meets (a) and (b) If an NP is not bound by any binder prefix in a phrase marker P, we say that NP is semantically free in P. Binder rule (final version): NP, (") Binder Index Evaluation rule (BIER) (final): (x) For any natural number n, LL Pn Y Prohibition against Accidental Coreference (PACO): l_SJg is a possible interpretation of sentence S only if g(n) ^ g(m) if m ^ n, for any natural numbers n,m. In the remainder of this book I will often use the term 'binding' without specification. Assuming that our conjecture about the irrelevance of syntactic binding is correct, this is a shorthand for 'semantic binding'; for the most part, however, little in the exposition hinges on this distinction. 6.5 Extensions 6.5.1 VP ellipsis revisited © ■ Before concluding our discussion of the CR, we need to revisit VP ellipsis (VPE), as discussed in chapter 5. There we saw that a sentence like (6.37) is ambiguous between a strict reading, on which the two pronouns corefer, as in LF (6.37a), and a sloppy reading, on which they are each bound in their own conjunct, as in LF (6.37b); to make coreference assumptions more perspicuous I will index names throughout this discussion, contrary to our official treatment, but nothing in the argument hinges on this: (6.37) John saw his sister, and Bill did, too. (a) Johni saw hisi sister, and B1II3 did see hisi sister, too. (b) Johni f>2 saw his2 sister, and Bill3 f>2 did see hiS2 sister, too. In section 5.5 of chapter 5 we discussed how to derive these two readings and, in subsection 5.5.2, how to allow for the prerequisite LFs while blocking any LF that leads to 'mixed' readings. It was proposed that two conditions are needed to capture the correct generalization, namely a strict syntactic identity condition on VP ellipsis, requiring that the indexing in the elided VP be identical to that in the overt one, and a condition No Spurious Coindexing, which prohibits the 6.5 Extensions 131 same index to be used referentially in one conjunct, and semantically bound in the other. As the alert reader might have noticed, however, (6.37a) violates the CR in the first conjunct, since John corefers with, yet c-commands, his. What if John binds his instead? Two possible LFs for this option would be: (6.38) (a) Johni fi\ saw hisi sister, and Bill3 did see hisi sister, too (b) Johni ^2 saw his2 sister, and B11I3 did see hisi sister, too Both LFs express a strict reading, as intended. But (6.38a) violates the prohibition against spurious coindexing: the index 1 is used both as a referential index (on John and the elided his) and as a bound index (on /3 and the first his), which is precisely what that condition prohibits. (6.38b), on the other hand, doesn't violate this prohibition (index 1 is only used referentially), but fails the VP identity condition instead: since the indices on the two pronouns aren't identical, the VPs aren't either. Given that none of these sem-binding LFs seems to work, then, one might try to grab the other end of the dilemma. Maybe (6.37a) is the correct LF after all, and coreference in the first conjunct is exceptionally possible precisely because the only well-formed sem-bound structure, (6.37b), leads to a different interpretation, namely the sloppy one. This requires us to think of the CR as applying at the higher level: while bound and coreferent construal yield the same reading for the first conjunct alone, they yield distinguishable interpretations for the coordinate structure as a whole.9 While this is certainly a plausible idea, it is undermined by the facts. To see how, simply recall that exceptional coreference structures are generally immune to Binding Condition violations (since Binding Conditions by assumption only 'see' semantic binding). The consequence of this for the VP-ellipsis case would be (6.39) below: (6.39) Consequence of assuming that strict identity in VP ellipsis always involves binding theory obviations in the first conjuncts: All first conjuncts in strict identity VP ellipsis should be exempt from Binding Conditions B and C. This prediction, however, is clearly wrong, as can be seen in (6.40) below. To interpret John and him/John as anaphorically related in the examples in (6.40) is as hard or impossible as it is if you consider their first conjuncts in isolation, regardless of whether construed as strict or sloppy; but in contrast to that, both examples should, according to (6.39), have a grammatical strict construal as given in (6.41): (6.40) (a) *John saw him and Bill did, too. (b) *John saw John's sister, and Bill did, too. 9 Note that this 'higher level' isn't always the sentence level, since the same ambiguity shows up in cross-sentence VP ellipsis such as John saw his sister. Bill did, too. See Fox (2000) for much more discussion of these issues. 132 THE COREFERENCE RULE (6.41) (a) Johni saw himi and B1II2 did see hinii, too. (b) Johni saw hisi sister, and B1II2 did see hisi sister, too. We conclude that strict identity in VP ellipsis cannot involve exceptional coref-erence in the first conjunct. To the contrary, to render the sentences in (6.40) ungrammatical, the CR must apply to these, forcing semantic binding and thus provoking a Binding Condition violation. The correct LF for the strict reading of (6.37) must thus be either (6.38a) or (6.38b) above. Let us assume that (6.38b), repeated here, is the correct LF. We are then forced to give up the strict identity condition on VP ellipsis, replacing it by something like (6.43): (6.42) Johni P2 saw his2 sister, and Bill3 did see hisi sister, too (6.43) VP Identity Condition (final): An elided VP must find a syntactic antecedent that is identical except perhaps for indices. This condition obviously allows for (6.42), but along with it for a whole new set of unwanted LFs in which the elided pronoun has a completely new, referential index, e.g.: (6.44) * Johni P2 saw his2 sister and B1II3 did see MS4 sister, too. The correct generalization about pronouns in VP ellipsis seems to be the following (cf. e.g. Fox [2000]: 116f.): (6.45) NP Parallelism: NPs in the antecedent and elided VPs must either (a) have the same referential value, or (b) be bound (in parallel) in their respective conjuncts (6.46) Ancillary definition: individual a is the referential value of NP in sentence Sif (a) NP refers to a, or (b) NP is semantically bound in S to an NP' with referential value a The innovation with respect to the generalization (5.34) discussed earlier in section 5.5 of chapter 5 (roughly: pronouns must either corefer or be bound in their respective conjuncts) lies in clause (6.45a): The pronouns now don't need actually to corefer, as long as they end up denoting the same individual. The prerequisite notion of (same) referential value is defined in (6.46). Unfortunately, (6.45), unlike our earlier stricter condition, doesn't follow at all from the theory of binding and ellipsis. That is, an intuitively clearly unavailable reading such as that expressed by LF (6.44) isn't ruled out by conditions on VP ellipsis or the prohibition No Spurious Coindexing, but only by explicitly stipulating (6.45) in addition. A way to improve on this unsatisfactory situation might be provided by the VP-ellipsis condition proposed in Rooth (1992a). Simplifying considerably, Rooth argues that the relation between the first conjunct and the second in VP ellipsis is one of focus-matching, along the following lines: 6.5 Extensions 133 (6.47) VPe can be elided if there is an antecedent constituent Ca that is focus-matched hy some Ce dominating VPe- (6.48) CEf(ocus)-matches CA if [CAF is an element of P-SET(CE). As indicated, the formal rendering of focus-matching uses the P-SETs introduced in chapter 5, section 5.2. For all intents and purposes, we can think of Ce focus-matching Ca if some substitution for the focused elements in Ce yields a C^ that is synonymous to Ca under any assignment g. Consider third-party readings first: (6.49) (Everybody liked Amandai.) Tirri2 asked heri out for dinner. ROBERTA3 did ask heri out for dinner, too. What is crucial for VP ellipsis according to (6.47) is not the relation between the VPs themselves, but between constituents dominating them. Let's look at the I-level: we find asked herj out for dinner (=Ca) and did ask herj out for dinner (=Ce), which are synonymous under any assignment g; this is a trivial case of focus-matching (no foci, hence no replacements), meeting (6.47). Thus VP ellipsis is predicted to be possible. Suppose now, for illustration, that we looked at the IP level instead: although we find Tim2 asked herj out for dinner and ROBERT A3 asked herj out for dinner, which are not synonymous, we still meet (6.47), given that ROBERTA is focused, and therefore its P-SET contains the proposition 'Tim asked her out for dinner.' Informally, we replace the focused ROBERTA3 by Timi, notated as in (6.50): (6.50) Tirri2 asked heri out for dinner <£> Tim2 roberta3 asked heri out for dinner IP e - Tirri2 asked her] out for dinner - is synonymous with IPA, thus IPE -ROBERTA3 asked her] out for dinner - focus-matches IPA, and ellipsis is, again, predicted possible. The gist of this proposal is thus that ellipsis is possible if you can replace focused material around the elided VP so as to match the antecedent site. Turning now to strict identity, note that if the pronoun in the antecedent VP is semantically bound (as required by the CR if it is c-commanded by its antecedent), the pronoun in the ellipsis site cannot be coindexed with it, unless it is bound, too. This is a consequence of No Spurious Coindexing. The LF for (6.51) must thus be (6.51a) (= [6.42]), which meets Rooth's condition (6.47) by virtue of focus marking on Bill, cf. (6.51b): (6.51) John saw his sister and BILL did, too. (a) Johni ^2 saw his2 sister, and B1II3 did see hisi sister, too (b) Johni ^2 saw his2 sister <£> Johni BILL3 did see hisi sister Note that, in this example, the fact that the matching takes place at a higher constituent is crucial; did see hisi sister doesn't focus-match either saw his2 sister or ^2 saw his2 sister (it isn't synonymous - thanks to the different indices and it doesn't contain any focused items). But BILL3 did see hisi sister f-matches Johni P2 saw his2 sister, as shown in (6.51b). 134 THE COREFERENCE RULE Note, too, that focusing Bill in (6.51) is a necessary condition for this to go through. Presumably a theory of focus will derive this fact independently,10 but for the moment we can be content with the observation that in all the cases of VP ellipsis we looked at, such focusing is indeed found. In this case of strict identity, then, the two pronouns aren't coreferring, although they have the same referential value in the sense of (6.46) above. This comes out to be a case of semantic focus-matching in the sense of (6.47), since 'viewed from the IP level,' both pronouns denote John for every assignment g. Turning now to sloppy identity, (6.47) licenses parallel binding in the familiar way, but regardless of the actual choice of indexing:11 (6.52) John saw his sister, and BILL did, too. (a) Johni ^2 saw his2 sister, and BILL3 ft 4 did see MS4 sister (b) Johni ^2 saw ms2 sister <£> Johni Bill3 ^4 didseehis4 sister (6.53) Tom wanted Sue to water his plants, while JOHN wanted MARY to. (a) Tomi ^2 wanted Sues to water his2 plants, while JOHN3 ^4 wanted MARY6 to water MS4 plants (b) Tomi ^2 wanted Sues to water his2 plants ^4 wanted Tornx john3 Sues MARYg to water his4 plants Finally, what of the possibility of arbitrarily changing the indexing in the elided VP, as allowed by (6.43)? Note that this option is now severely constrained by (6.47): if both pronouns are referential, they have to be coreferential, and hence - by PACO - coindexed, in order to ever meet (6.47). If one of them is bound and the other one isn't, (6.47) can only be met if the bound one is eventually bound by (an NP bound by) an NP that is coreferent with the referential one, as in (6.51) (i.e. same referential value). If both are bound, their binding has to be parallel to meet (6.47), as in (6.52) and (6.53). Thus, while (6.44), repeated below, is possible as far as the identity condition (6.43) is concerned, it violates (6.47) because his^ is not coreferent with Johni, nor ever bound by something that is, so focus-matching will fail at every node in the structure, which is therefore correctly ruled out. (6.54) * Johni P2 saw his2 sister and B1II3 did see his4 sister, too. In sum, adopting (6.48) together with the more lenient VP identity condition (6.47) gives us a way of deriving sloppy readings and strict readings with c-commanding antecedents without actually having to stipulate NP Parallelism as in (6.45) and without using the notion of referential value. Coupled with No Cf. Rooth (1992a,b); Schwarzschild (1999); Tomioka (1999) among many others. If one chooses to use the same index on both binder prefixes (e.g. replace 4 by 2 in these examples), (6.47) is actually met at the I level here as well. While this doesn't do any immediate harm in this example, it effectively provides a loophole to circumvent the parallelism requirement otherwise enforced by (6.47). To evaluate whether this is problematic is beyond the scope of this book, though. 6.5 Extensions 135 Spurious Coindexing, it derives the complex range of strict/sloppy facts in VP ellipsis in full compatibility with the CR and our earlier treatment of exceptional coreference. An interesting consequence of this setup is that the actual name of an index never plays a role in the theory. That is, while in the treatment of VP ellipsis of chapter 5, identity of indices on the overt and elided pronoun was a crucial ingredient to derive strict readings, this is no longer the case in the present, revised, system, nor could it be, given that we allowed mere sameness of referential values as in (6.51) in strict identity cases. Note furthermore that, by the same token, this treatment of VP ellipsis is compatible with Reinhart's suggestion, mentioned in subsection 6.2.3, to omit indices on referring NPs, including referential pronouns, altogether. Focus-matching, as defined in (6.47), checks sameness of denotation in the semantics, not by looking at indices. It therefore doesn't matter how the theory actually assigns a denotation to referential pronouns. Adopting Reinhart's proposal can actually lead to a simplification of our theory, namely the elimination of No Spurious Coindexing. Consider again a potentially problematic case like (6.55) below: (6.55) every boy \ loves hisi mother and almost every man ^2 d°es love hisi mother, too Under full indexing, (6.55) above meets (6.47) at the I-level (loves hisi mother is synonymous with does love hisi mother for any assignment g), predicting an unattested reading. This had to be blocked by No Spurious Coindexing, which bans the index 1 from being bound in the first conjunct, but referential in the second. If we dispense with referential indices altogether, the second conjunct is ungrammatical independently, because it shows an unbound index, in violation of (6.20b) above. Dropping the index on the elided his altogether yields a grammatical LF, but one which doesn't yield synonymous expressions at the I (or any other) level. 6.5.2 Indices on referential pronouns© ■ In this section I will briefly explore what it would take to adopt Reinhart's idea of eliminating unbound indices completely, even on referential pronouns. As remarked earlier, given our reformulations of the Binding Conditions, there is clearly no syntactic reason to maintain referential indices, once we have followed Reinhart to the point of eliminating both indices on full NPs, and Binding Condition C. Our concern in this section, then, will be with the semantic consequences of this move. For the purpose of the discussion, we will refer to a theory as outlined in the summary above (i.e. one that uses the CR/(6.32), doesn't index full NPs, but does index referential pronouns) as the official theory. Its Reinhartian extension, in which all unbound indices are eliminated, will be called the R-theory. 136 THE COREFERENCE RULE How is a non-indexed element interpreted on the R-theory? Full NPs, including in particular names, don't pose a problem (remember that we only made them assignment dependent in order to derive Binding Condition C effects in the first place). But what about referential pronouns? Technically, we could be content assuming that the pronoun gets assigned a random individual (of the right sex and number, that is). But this would run counter to our intuitions that pronouns are context-dependent.12 So far we have assumed that the assignment g is all the context we need. But if pronoun meanings are context-dependent, but not assigned by the assignment function g, it is obviously time to revise that assumption. Let us assume therefore that expressions are evaluated with respect to an assignment g and a context, c. It is the context that provides referents for indexless expressions, in the way illustrated in (6.56): (6.56) |[she]c'^ = the most salient female person in c if there is one, undefined otherwise We don't really need to decide just what kind of object c is, as long as it allows us to define 'maximally salient individual.' Let us consider an example: (6.57) gives the interpretation of a sentence relative to an assignment g and a context c in the official theory and the R-theory (parentheses mark presuppositions): (6.57) She(6) is writing a book. (a) official: (g(6) is a female individual and) g(6) is writing a book (b) R-theory: (there is a maximally salient female person in c and) the most salient female person in c is writing a book Both denotations capture the context-dependency of the pronoun meaning as well as its lexical content. Moreover, the R-theory definition captures an aspect of the pronoun meaning that isn't mentioned in the official one, namely salience. It opens the door for an explanation of why she in (6.58) is understood to be Sally, rather than any other female individual: using a name creates a context in which the referent of that name is maximally salient; therefore, a following pronoun will pick up that referent: (6.58) I saw Sally this morning. She was riding a bicycle. No such explanation is available in the official theory (though see the remarks in chapter 2, section 2.4.1). On the other hand, the official theory could easily represent the two most salient interpretations of a text like (6.59) as an ambiguity; she can be coindexed with either Norma or Sally: (6.59) Norma saw Sally this morning. She was riding a bicycle. Within the R-theory, these different interpretations are not discriminated in the grammar proper. Furthermore, it is necessary to assume that there exist at least two different contexts which can result from an utterance of Norma met Sally 12 It is also not trivial to implement. After all, [pronoun]g should be a semantic object of type (e). But what object would this be? 6.5 Extensions 137 this morning, which differ in who is the more salient woman in them. More precisely, while for any given context, the coreference pattern in (6.59) is uniquely determined, because the utterance of Norma met Sally this morning will make one woman maximally salient, to which she in the next sentence will then refer; in two different initial contexts, the utterance of Norma met Sally this morning might yield a different maximally salient woman, resulting in a different coreference pattern. Since a number of facts outside of the scope of this book become relevant here (for example intonation of the pronoun and its antecedent), we will leave the issue at these remarks. Note finally that contexts in the R-theory must be able to change sentence internally. Otherwise, two agreeing pronouns within the same sentence would inevitably end up with the same referent. Consider e.g. (6.60): (6.60) Norma saw Sally and her son this morning. Her son invited her for tea. The first and the second her in the second sentence of this example can (and, in fact, are most likely to) be interpreted as referring to two different women: Sally and Norma, respectively. To derive this possibility, we need to assume that the context for the two is not the same. In other words, we need to acknowledge that uttering her son invited changes the saliency in the context (note that coreference between the two her?, is grammatically possible since neither c-commands the other; this is also evident from the possibility of a coreferent reading in the structurally parallel example Her son was helping her with the groceries). Once again, implementing this properly is beyond the scope of this book, but the idea should be clear by now. Likewise, we cannot explore here what facilitates a change in contextual saliency and what does not (e.g. invite for dinner vs. help with groceries). Note in passing, though, that for a sentence like She greeted her to get an interpretation at all, the saliency change must be available. In sum, making the R-theory semantically adequate requires us to develop a complete theory of context, and reference assignment to pronouns by contexts. While perhaps such a theory would be beneficial in other respects, in particular if it could say more about the preferred interpretation of referential pronouns in particular contexts, it is beyond the scope of this book. Mostly for this reason, we will continue to index all pronouns, bound or free, in our official theory. It should be borne in mind, though, that, apart from technical questions, the choice between the two theories seems to reflect in part our opinion about the more substantial question whether referential pronouns are more akin to bound pronouns (official theory), or to other deictic and, in particular, demonstrative expressions (the R-theory, as sketched here). Exercise 6.7 English, like many other languages, uses the same elements as bound and free pronouns. Suppose this was universally so; would that provide evidence 138 THE COREFERENCE RULE for/against the R-theory? Suppose some languages did use different forms for the two. Which position would that argue in favor of? 6.5.3 More on reflexives© ■ The arguments presented in this chapter all argued in favor of a version of Binding Theory which regards semantic binding only, and is blind to syntactic binding. We were thus able to simplify the Binding Conditions and avoid disjunctive formulations such as given in chapter 5. For Binding Condition A, which we haven't discussed in this chapter so far, this yields a formulation as in (6.61) (argued for in particular in Grodzinsky and Reinhart [1993]): (6.61) Binding Condition A A reflexive pronoun needs to be semantically bound in its domain. The reasoning here is transparent: we know that reflexives are always anaphori-cally related to a c-commanding NP; but, according to the CR, two NPs cannot, ceteris paribus, corefer if one c-commands the other. Therefore, reflexives will have to be sem-bound. The question we may ask now is if there are any empirical effects of (6.61), similar to the obviation effects observed above. For example, (6.61) is often argued for on the grounds that it allows one to derive the following (purported) generalization about VPE: (6.62) Reflexives do not allow strict identity readings in VPE. Let us ask, first, whether (6.62) is empirically correct; and, second, whether it actually follows from (6.61). For English, it has been argued e.g. in Hestvik (1992): 1 that a '... strict reading is only weakly acceptable or impossible' in a sentence like (6.63) (a strict reading, recall, would have the second conjunct interpreted as 'Bill defended John'): (6.63) John defended himself well, and Bill did, too. (?*strict) On the other hand, Sag (1976) claims that (6.64a) is acceptable on a strict reading (i.e. Sandy could imagine Betsy dating Bernie), and speakers report similar judgments on sentences like (6.64b) and (6.64c): (6.64) (a) Betsy couldn't imagine herself dating Bernie, but Sandy could. (b) Bill was scared to introduce himself to Monica. Thankfully, the officer had already agreed to. (c) I could see myself having a romantic dinner with Winona Ryder, but my girlfriend couldn't. Cross-linguistically, strict readings with reflexives are attested, e.g. in Icelandic, Norwegian, Finnish, and Swedish:13 13 See Thrainsson (1992):60; and Huang (2000):99f. 6.5 Extensions 139 (6.65) Joni sag5i a5 pu hef5ir sviki5 sigi og Petur ger5i pa5 lika. J. said that you had betrayed self and Peter did so too (Icelandic) 'Jon said that you had betrayed yourself and Peter said so (that you had betrayed Jon/Peter), too.' (strict or sloppy) Supposing that sentences like these are the rule, rather than the exceptions (contrary to e.g. Hestvik [1992]), would (6.61) rule them out? No! As we have seen in section 6.5.1 above, strict identity readings can, and usually do, involve sem-binding in the overt VP (cf. e.g. [6.42] and the discussion around it). Therefore, the problematic generalization (6.62) doesn't follow, even if (6.61) is assumed. A licit LF for e.g. (6.64a) is given in (6.66) (as in section 6.5.1, I continue to index full NPs for ease of discussion): (6.66) Betsy2 f> i couldn't imagine herselfi dating Bernie, but SANDY3 \ could? imagine her2 dating Bernie. SANDYF 3 f}j could imagine her^ dating Bernie f-matches Betsy2 fii couldn't imagine herselfi dating Bernie ('replace' SANDY by Betsy and COULD by couldn't), making (6.66) a well-formed ellipsis structure given the assumptions made in section 6.5.1 above. In other words, (6.61) can be adopted despite the fact that (6.62) is too strong, because it doesn't entail it anyway. It should be mentioned, though, that this state of affairs is not without problems either. For one thing, it is true that strict readings involving reflexives seem much harder to get than with non-reflexives, a fact that remains to be captured given the present proposal. For another, as Hestvik (1992) shows convincingly, strict readings with reflexives are much more readily acceptable in subordinate ellipsis such as (6.67): (6.67) John defends himself better than Peter. (sloppy or strict) Speakers immediately accept a reading of (6.67) on which John defended himself better than Peter defended John, i.e. a strict one. This contrast is reported to be even stronger with the Dutch reflexive zichzelf, which doesn't allow strict readings in coordinations like (6.68a) at all, but is fine, just like its English counterpart, in subordinated ellipsis like (6.68b):14 (6.68) (a) Jan wast zichzelf en Piet 00k. (Dutch) J. washes himself and P. also 'Jan washes himself, and Piet does, too.' (sloppy/*strict) 14 Everaert (1986):254; these examples, as well as those in n. 15, are all found in the discussion of VP ellipsis. It should be noted, though, that they actually resemble the English stripping or bare argument ellipsis construction (John defends himself better than Peter.) more than English VP ellipsis. To the extent that these constructions behave differently (see e.g. Kennedy and Lidz [2001]), conclusions based on these examples have to be taken with a grain of salt. 140 THE COREFERENCE RULE (b) Zij verdedigde zichzelf beter dan Peter. she defended herself better than P. ' She defended herself better than Peter.' (sloppy/strict) As Hestvik points out, the crucial element in subordinate ellipsis that is lacking from coordinate ellipsis is that the antecedent of the overt reflexive actually c-commands the ellipsis site, and can thus bind the pronoun in the elided VP. In present terms, the elided pronoun can be semantically bound by the antecedent of the overt pronoun, as is the case in the LF in (6.69): (6.69) Johni f> \ [defends himselfi better than Peter2 defends him(self)i] It seems that there is at least a preference for an elided pronoun to be semantically bound, if its overt counterpart is a reflexive, be it in its own conjunct, as in sloppy identity coordinate ellipsis, or across the board, as in (6.69). Again, this preference is not captured by the rules we have so far.15 Finally, it should be noted that (6.61) does make very clear predictions for the case of focus constructions, namely that focus constructions involving reflexives - unlike those involving non-reflexive pronouns - are unambiguous: (6.70) (a) Only IDI voted for his proposal. (i) strict: only IDIi p voted for hisi proposal (ii) sloppy: only IDI^j? f>\ voted for hisi proposal 15 To make things worse, a whole number of reflexive markers cross-linguistically don't allow for strict identity with reflexives at all, not even in subordinate ellipsis, among them the verbal reflexive suffixes -utu and koll in Finnish and Kannada (Sells et al. [1987]: 178; Lidz [2001]: 129), the pronouns zich and sejsebe in Dutch and Serbo-Croatian (Sells et al. [1987] [123]), and even the long-distance reflexive zibun in Japanese (Sells et al. [1987]:186): (i) Jussi puolusta -utu -i paremmin kuin Pekka. /. defended self PAST better than P.-NOM (ii) Rashmi Siita-ginta cheenage tann-annu rakshisi -koll -utt -aale. R. S.-COMP better self-ACC defend -selfNONPAST3SGFEM (iii) Zij verdedigde zich beter dan Peter. she defended self better than P. (iv) Petar se branio bolje nego Ana. P. self defended better than A. (v) Petar je branio sebe bolje nego Ana. P. AUX defended self better than A. (vi) Taroo wa Jiroo yori zibun -o umaku bengosi -ta. T. TOPIC Jiroo than himself-ACC better defend -PAST 'Jussi/ Rashmi/ she/ Petar/ Taroo defends/defended self better than Pekka/ Sita/ Peter/ Ana/ Jiroo does/did (defend self/*her/*him).' (According to Huang [2000]:99f., 141, Hindi/Urdu, Marafhi, and Telugu don't allow strict identity either, but all his examples involve coordinate ellipsis.) It is interesting to note that verbal reflexive suffixes do not generally block strict identity, as they do in Kannada and Finnish: at least in subordinate ellipsis, the reflexive affix -dzi in Chichewa (Sells et al. [1987]:187; Mchombo [1993a]: 195) allows for strict identity. 6.5 Extensions 141 (b) Only IDI voted for himself. (i) * strict: only IDIi p voted for himselfi (ii) sloppy: only IDI^j? ß 1 voted for himselfi As discussed in chapter 5, section 5.2, semantic binding as in the (ii) examples results in a 'sloppy' reading, according to which no one else voted for themselves/their own proposal. Without semantic binding the reflexive and its antecedent simply corefer, resulting in a 'sloppy' reading: no one else voted for Idi/Tdi's proposal. If reflexives needed to be sem-bound, this latter construal should be out for the reflexive case, given that the reflexive in (6.70bi) is syntactically, but not semantically, bound. In other words, (6.70b), unlike (6.70a), is predicted to be unambiguous. This, however, does not accord with speakers' intuitions. While the sloppy reading for (6.70b) is generally preferred, the strict one is clearly judged possible. As far as I know, this wrong prediction has not been addressed in the pertinent literature. The only immediate way to capture this behavior would seem to be to reformulate Binding Condition A so as to require that reflexives be either semantically or syntactically bound within their local domain, accepting the fact that Binding Conditions A and B are simply not on a par. In sum, then, whether (6.61) is the correct way to state Binding Condition A seems unclear. At least under the generalizations about VPE put forward in section 6.5.1, it predicts that strict and sloppy readings should generally be possible in VPE, which is a good approximation to the facts, though not obviously correct. Its predictions for focus constructions, on the other hand, are arguably wrong. It is, of course, conceivable that the assumptions in section 6.5.1 are in need of changing, and that strict identity does require coreference in the antecedent clause as well as the ellipsis site. In that case, strict readings are predicted to be generally impossible in coordinated ellipsis, but possible in subordinated ellipsis (the position advocated in Hestvik [1992]). But that assumption doesn't help with the focus constructions either, and is faced with the problem of acceptable strict readings in coordinate ellipsis such as (6.64a) and (6.65). It seems that, in either case, something is missing from the picture as of yet. Exercise 6.8 Throughout the discussion in this section we have assumed that the pronoun in the ellipsis site doesn't itself impose any syntactic Binding Conditions. A conceivable alternative, explored e.g. in Kitagawa (1991), is that the elided pronoun itself must be a reflexive if its overt counterpart is, and that Binding Condition A directly applies to the elided pronoun in its local clause. (1) Discuss the predictions of such a proposal for the cases discussed here (coordinate ellipsis, subordinate ellipsis, focus constructions), and show where it is advantageous and where it is problematic. 142 THE COREFERENCE RULE (2) Try to construct examples that could help you to determine whether similar assumptions would be reasonable for Binding Conditions B and C. (For much more on this see the discussion of 'vehicle change' in Fiengo and May [1994].) Exercise 6.9 What are the possible interpretations for the elided pronoun in the examples in (6.71)? Discuss these data in the light of the various proposals floated around in this section. What conclusions, if any, can you draw? (6.71) (a) John wanted to introduce himself to Mary, but Carl hoped that Bill would. (b) John wanted to introduce his plans to Mary, but Carl hoped that Bill would. 7 Descriptive pronouns and individual concepts 7.1 Anaphoric pronouns that don't corefer 7.1.1 Introduction So far, we have assumed that a pronoun can be either semantically bound or free. In the latter case, it might or might not have the same referent as some other pronoun, name, or definite NP. Consider now (7.1) on a reading where he is anaphorically related to the direct object of the preceding sentence: Evidently, he in (7.1) cannot be semantically bound by its antecedent, given that they occur in different sentences (remember that c-command is a precondition for binding). If the antecedent is him (which, in turn, is likely to have a full NP antecedent in the context), the two pronouns corefer, and are coindexed. If it is Fido or her dog, which we take to be referring expressions, too, the anaphoric dependency involved is again coreference (which could be marked in the grammar by coindexing, though we saw in chapter 6 that indexing full NPs, even referential ones, can be omitted). However, neither a dog nor exactly one dog refer -they are quantified noun phrases (QNPs) - and thus cannot, a fortiori, corefer with he, given that they are genuinely quantificational NPs. What we see, then, is that, as things stand, all these anaphoric dependencies, with the exception of the pronoun-pronoun case, are not represented in the grammar at all.1 The pronoun in the second sentence is a referential pronoun, which, depending on its index, can refer to any individual, including Fido the dog.2 1 As discussed in chapter 6, sections 6.2.3 and 6.5.2, even pronoun-pronoun cases need not be, as far as Binding Conditions are concerned. 2 It is worth while to stress that the question of the grammatical representation of anaphoric relations is independent of the question of preferred reference. That is, even if we did express the anaphoric dependencies above by means of coindexing, we would still need a theory of indexing that explains why coindexing is so strongly preferred, i.e. why he almost inevitably has to refer to Fido rather than anyone else. him Fido (7.1) Janet brought her dog a dog exactly one dog to the party. But he had to wait outside. 143 144 DESCRIPTIVE PRONOUNS AND INDIVIDUAL CONCEPTS While there is nothing wrong with this formally, there is at least a persistent intuition that the pronoun he in some sense stands for a repetition of the description her dog, or goes proxy for a description like the (one) dog Janet brought to the party. But, honorable though such intuitions are, is there any way to show that there is more to the anaphoric dependency in (7.1) than so far assumed? Put the other way around: assume, for the sake of the argument, that he does in fact 'expand' to a full-fledged definite NP at LF. What empirical difference would that make? None for the case of (7.1) - not at least until we have a restrictive theory of which description a pronoun expands into in a given context. But we can construe cases in which a definite should allow for more readings than a plain referential pronoun; we will turn to these cases now. Exercise 7.1 Suppose, contrary to what was assumed above, that a dog and exactly one dog can somehow bind the pronoun in the second sentence. That is, assume an LF for these cases which essentially looks like (1) (more realistically, the first him] should be a trace of sorts, but since we haven't discussed the binding of traces yet, and nothing hinges on this here, we will go with this representation): (7.2) a dog exactly one dog wait outside]] fi\ [[Janet brought himi to the party] but [hei had to Derive the truth conditions for these sentences, assuming the following NP meanings. Then argue whether the derived truth conditions are intuitively adequate. They shouldn't be for at least one case; show this using concrete scenarios about dogs brought to the party. (7.3) (a) |[a dog]# = XP. there is at least one dog x s.t. P(x) = 1 (b) [[exactly one dog]g = XP. there is one and only one dog x s.t. P (x) = 1 7.1.2 Pronouns as descriptions i Evans (1980):342 points out that the conditional in (7.4) has a natural reading on which a man or a man in the garden antecedes him: (7.4) If there is a man in the garden, John will tell him to leave. Note, however, that the antecedent clause of the conditional does not talk about a particular man in the garden. And, accordingly, there is no particular referent for him that would yield the intuitively correct truth conditions for this sentence. If we interpret him as a full-fledged NP, on the other hand, we get the intuited reading: (7.5) If there is a man in the garden, John will tell the man who is in the garden to leave. 7.1 Anaphoric pronouns that don't corefer 145 The same point can be illustrated using the following variant of our original example: (7.6) Every woman brought her dog/exactly one dog to the party, but left him outside. Here we have embedded sentences of the type in (7.1) under another QNP. Accurate paraphrases for these sentences seem to be those in (7.7): (7.7) (a) every woman brought her dog to the party but left her dog outside, (b) every woman brought exactly one dog to the party but left the dog she brought to the party outside. Starting with Geach (1962), pronouns like in (7.7a), which can be replaced by a literal repetition of their antecedent NPs, have been called pronouns of laziness; pronouns like those in (7.7b), whose paraphrase involves a description that has to be 'distilled' from the preceding context, were christened e-type pronouns in Evans (1977). In the discussion to follow we will give the same analysis to pronouns of laziness and e-type pronouns, and use the term e-type pronoun indiscriminately to refer to pronouns of laziness and 'proper' e-type pronouns alike. Can we get meanings as in (7.7) by assuming that him is a referential pronoun? Clearly not, because no particular dog can serve as the referent of that pronoun. Rather, dogs need to covary with women, which is what makes the paraphrases in (7.7) work so well, in which the pronouns are replaced by definite descriptions containing a variable bound by every woman. Can we assume instead that him is a bound pronoun? No again, since for one thing, her dog/exactly one dog doesn't c-command him and, hence, cannot bind it. For another, even if it somehow could, the resulting truth conditions would be wrong at least in the case of (7.7b), namely something like (7.8): (7.8) for every woman x, there is one and only one dog y such that x brought y to the party and left y outside In (7.8), exactly one dog takes scope over both conjuncts, which it has to to bind the pronoun in the second one. But, as a result of that, the sentence is true as long as the number of dogs brought to the party and left outside was exactly one. In other words, a woman could have brought any number of dogs to the party, as long as she left exactly one of them outside. But these are not the truth conditions this sentence intuitively has (rather, they would correspond to a sentence like Every woman has exactly one dog she brought to the party and left outside). Rather, there must be exactly one dog per woman that she brought to the party, and that dog must have had to wait outside. In other words, exactly one dog must crucially not scope over the second conjunct and, hence, him cannot be a bound pronoun (cf. section 7.1.1 above). Examples like these prompted Evans (1977, 1980) to conclude that it must be possible actually to interpret pronouns as definite descriptions. That is, the paraphrases in (7.7) and (7.5) - or something very similar to them - are the correct representations for (7.6) and (7.4). 146 DESCRIPTIVE PRONOUNS AND INDIVIDUAL CONCEPTS There are two extreme ways of formalizing this idea in our grammar: either we assume that these pronouns transform into full-fledged definite NPs at LF, i.e. (7.7a) and (7.7b) essentially are the LFs for the sentences in (7.6); or we do this in the semantics, assuming that a pronoun such as himn is, or at least can be interpreted not as g(n), but as f(g(n)), where / is some contextually given function from individuals to individuals, e.g. one that maps individuals to the dogs they brought. The specific proposal we will adopt here is somewhere in the middle between these two. Following Cooper (1979), we assume that the pronouns are represented as full NPs at LF, but that their terminal nodes are filled by special, phonologically inert, elements. Concretely, the LF representation for an e-type pronoun will look as in (7.9) (cf. also Heim and Kratzer [1998]:ch. 11): NP (7.9) Pronoun expansion: pron =^lf the^^^^^N (a) lRmp = g(Rm), where g(Rj) e De