This page intentionally left blank
Topics in Ellipsis
Ellipsis occurs when certain portions of a sentence are not sp...
43 downloads
1036 Views
1MB Size
Report
This content was uploaded by our users and we assume good faith they have the permission to share this book. If you own the copyright to this book and it is wrongfully on our website, we offer a simple DMCA procedure to remove your content from our site. Start by pressing the button below!
Report copyright / DMCA form
This page intentionally left blank
Topics in Ellipsis
Ellipsis occurs when certain portions of a sentence are not spoken – for example ‘‘Mary has read more books than Bill has’’ and ‘‘Jack called, but I don’t know where from.’’ These constructions interest linguists because the meaning of the sentence cannot be traced directly to the words it contains. This volume brings together a team of leading syntacticians to propose new and original solutions to some key questions in the study of ellipsis: What characterizes ellipsis? Under what conditions is it possible? What kinds of meanings are allowed to go unspoken? Drawing on a variety of authentic constructions, they examine ellipsis in the context of a range of syntactic phenomena such as binding, raising, anaphora, movement, and Scrambling. Making significant progress towards solving some central problems in syntactic theory, this cutting-edge volume will be of key interest to anyone working on theoretical syntax, semantics, and psycholinguistics. is Professor in the Linguistics Department, University of Massachusetts, Amherst. He has previously published Beyond Principles and Parameters (co-edited with Ian Roberts, 1999), and a variety of articles on aspects of ellipsis.
KYLE JOHNSON
Topics in Ellipsis Edited by
Kyle Johnson University of Massachusetts, Amherst
CAMBRIDGE UNIVERSITY PRESS
Cambridge, New York, Melbourne, Madrid, Cape Town, Singapore, São Paulo Cambridge University Press The Edinburgh Building, Cambridge CB2 8RU, UK Published in the United States of America by Cambridge University Press, New York www.cambridge.org Information on this title: www.cambridge.org/9780521815086 © Cambridge University Press 2008 This publication 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 2008
ISBN-13 978-0-511-39449-2
eBook (NetLibrary)
ISBN-13
hardback
978-0-521-81508-6
Cambridge University Press has no responsibility for the persistence or accuracy of urls for external or third-party internet websites referred to in this publication, and does not guarantee that any content on such websites is, or will remain, accurate or appropriate.
Contents
1
Introduction KYLE JOHNSON
2
VP Ellipsis and constraints on interpretation DANIEL HARDT
3
183
A step-by-step guide to ellipsis resolution SATOSHI TOMIOKA
10
154
The silent content of bound variable pronouns ULI SAUERLAND
9
132
On binding scope and ellipsis scope WINFRIED LECHNER
8
95
Variable island repair under ellipsis JASON MERCHANT
7
69
Argument Contained Ellipsis CHRISTOPHER KENNEDY
6
30
The view of QR from ellipsis KYLE JOHNSON
5
15
Direct compositionality and variable-free semantics: the case of Antecedent Contained Deletion PAULINE JACOBSON
4
page 1
210
Shared constituents and Linearization CHRIS WILDER
229
Notes References Index
259 285 300
v
1
Introduction Kyle Johnson University of Massachusetts at Amherst
Ellipsis is the consummate crowd-pleaser. To the anaphora theorist, it represents an extreme case of context sensitivity. Like a pronoun, an ellipsis gets its meaning from the context it’s in. But unlike a run-of-the-mill pronoun such as her in (1a), an elided phrase, like the VP in (1b), provides very little guidance in navigating that context to resolve its meaning.1 (1)
a. Mary kissed John, and then Fred kissed her. b. Mary might kiss John or marry Fred, but Sam won’t [VP].
The phonological form of the pronoun in (1a) signals that her should get an interpretation that involves just singular females, and this is sufficient to make it clear that the utterer of (1a) means her to refer to Mary. The elided VP in (1b), by contrast, has no phonological form and provides, therefore, no help in gathering a meaning for it. (1b) is consequently ambiguous in a way that (1a) is not. An ellipsis provides at most information about the syntactic and semantic type of its antecedent. The rest is up to how contexts furnish the information necessary to complete the messages that sentences convey. If you’re interested in investigating how the meanings of sentences are made complete by the contexts they are in, then ellipsis provides a dramatically pure laboratory. To the syntactician, ellipses give an interesting view of the processes that structure sentences. That ellipses seem to be defined over phrases of certain sorts – the ellipsis in (1b), for instance, appears to be restricted to VPs – suggests that the strings they affect should correspond to the standard diagnostics for constituency. Interestingly, they don’t seem to. And they have other syntactic abnormalities. Certain ellipses appear to defeat the normal constraints that hold of long-distance dependencies. ‘‘Sluicing,’’ a process that elides clauses, seems to create grammatical sentences from what should be constructions that violate these constraints. Ross (1969) made this discovery, and the contrast in (2) illustrates it. 1
2
Kyle Johnson
(2)
a. *We will discover something that solves a well-known problem, but I won’t divulge which we will discover [something that solves]. b. We will discover something that solves a well-known problem, but I won’t divulge which [IP ].
The bracketed string in (2a) constitutes one of Ross’s (1967) famous islands, into which the relative pronoun which may not bind. Although the ellipsis in (2b) has the same semantic content as this bracketed string, and Merchant (2001) provides compelling evidence that it also has a syntactic form not too different, it nonetheless does not contain an island. There is something about the ellipsis, perhaps its lack of phonological form, or maybe some subtle attribute of its syntax or semantics, that removes islands.2 Ellipses confront the syntactician with unexpected deviations from normal behavior, and thereby provide the seeds for the next generation of syntactic theory. To the semanticist, ellipses provide an intriguing testing ground for hypotheses concerning the syntax – semantics interface, for there are cases in which an ellipsis seems to demand a certain sort of syntax. Perhaps the most famous of these are instances of ‘‘Antecedent Contained Ellipsis,’’ first discussed by Bouton (1970). In these scenarios, an ellipsis seems to be contained within a string from which its meaning derives; (3) is an example. (3)
She has [VP worked on every ellipsis problem that you have [VP ]].
The elided VP means something like ‘‘worked on’’ or ‘‘worked on x,’’ where x is bound to every ellipsis. Because the only sentence in the context that provides this meaning is (3) itself, this would seem to demand that the VP worked on every ellipsis that you have be given this interpretation. This somewhat surprising conclusion seems to be related to the fact that the DP containing the ellipsis in (3) can scope out of the VP it’s in. When an ellipsis is within a DP that cannot scope out of the antecedent, as in (4), the result is ill-formed. (4)
*She has [VP decided that no one knows about every ellipsis problem that you have [VP ]].
Ellipsis, then, promises to help work out how quantificational expressions like every ellipsis problem make their semantic contribution. And finally, to the ellipsis theorist, ellipses are everything. It is to them that we turn for an account of where ellipses are possible, and where their properties derive from. Sadly, the ellipsis theorist depends on all the others for clear answers, and consequently brings up the rear. So little is known
Introduction
3
about ellipsis that even its taxonomy is up for grabs. Some of the cases that need classification, each with the moniker they inherited from the construction-specific days of their discovery, are in (5).3 (5)
a. Sluicing: She read something, but she won’t say what [IP ]. b. Verb Phrase Ellipsis: She read something and he did [VP ] too. c. Pseudogapping: She’ll read something to Sam, but she won’t [VP ] to Billy d. Gapping: Some read something to Sam and others [VP ] to Billy. e. Right Node Raising: She deliberately [VP ], and he accidentally, read something. f. Comparative Deletion: Mary has read more books than Bill has [VP ].
In all of these examples, a VP (and maybe a bit more in the case of (5a)) is missing, and so it might seem reasonable to think of them as particularized results of the same process. And yet, none of them seem to have identical properties. For instance, Verb Phrase Ellipsis does not seem to have the island alleviating property that Sluicing does; compare (2b) to (6). (6)
*We will discover something that solves a well-known problem, but I won’t divulge which we will [VP ].
Nor may Sluicing exist in the antecedent contained environments that VP Ellipsis can; compare (3) with (7). (7)
*She met a guy that you asked who [IP ]. compare: ‘‘She met a guy that you asked who she met.’’
Is this because Sluicing and VP Ellipsis are different things? Or could it be that the mechanisms which allow ellipses to overcome island constraints or find antecedents from the strings they are part of are sensitive enough to the differences in (6)/(2) and (7)/(3) to distinguish them? Answering these questions is the ellipsis theorist’s first order of business. Ellipsis is generous, then, helping each advance on the goals of their specializations. But it also focuses these potentially divergent specializations onto a single, coherent, goal and in so doing enforces collaboration. There will be no solution to the problems ellipsis presents that comes from just one area of linguistic theory, and recognizing this has lured those working on ellipsis outside the borders of their disciplines. Ellipsis forces
4
Kyle Johnson
us to calibrate findings across the syntax – semantics – pragmatics divides, and this alone has generated interesting results. The chapters in this volume spring from this borderless orientation. Dan Hardt’s chapter offers an account for a particularly puzzling instance of sloppy identity that Dahl (1974) discovered. Sloppy identity refers to the ability that certain anaphoric items have to gain new denotations in ellipsis contexts. If the pronoun in the first clause of (8) is understood to refer to the same individual that Sam does, then the elided VP in the second clause can be understood as if it had a pronoun in a parallel position that refers to George. (8)
Sam likes his shoes and George does [VP ] too.
The elided pronoun is said to be sloppily identical to the first under such readings. Dahl discovered that when there are two pronouns involved in these constructions, the sloppy interpretation of the second depends on whether the first is understood sloppily. In (9), for example, the elided VP can be understood to mean what (9a) paraphrases, but not what (9b) does. (9)
Sam said he likes his shoes and George did [VP ] too. a. [VP ] ¼ say George likes George’s shoes. b. [VP ] 6¼ say Sam likes George’s shoes.
Analyses of this phenomenon range from placing syntactic constraints on the dependencies that sloppy pronouns may have on antecedents (see Fiengo and May 1994) to exploiting locality conditions on Binding Theory (see Fox 1998). Hardt demonstrates how this, and related effects, can be derived from the conditions that determine how ellipses are resolved by the discourse contexts they are in. The effect in (9) reflects constraints on what constitutes the relevant discourse contexts. The proposal Hardt makes leaves open what an ellipsis consists of. Among ellipsis theorists there are two popular ideas about what an ellipsis is made of. On the one hand, it is a garden-variety phrase, similar in all respects to an overt phrase, but not interpreted phonologically.4 On the other, ellipses are a specialized sort of silent word, maybe along the lines of the silent pronouns that languages like Japanese or Italian are thought to have.5 This second ‘‘proform’’ approach is appealing as it makes sense of the similarities ellipses have with overt pronominals in finding antecedents from their contexts. If proforms are defined as having denotations that make use of contextual information, then analyzing ellipses as proforms explains this fact about them. And to the extent that the way in which pronouns access contextual information is like the way that ellipses do, we have evidence for equating them. Hardt’s proposal is consistent with this view of ellipses.
Introduction
5
But it is also consistent with the first view: that ellipses are unpronounced strings. And that is fortunate, for there are obstacles to overcome if ellipses are proforms. Maybe the most serious is that ellipses can be bound into by interrogative pronouns, as (2b) and (10) illustrate, while unambiguous proforms can’t, as the contrast with (11) shows. (2b)
We will discover something that solves a well-known problem, but I won’t divulge which [IP ].
(10)
I know which problems you’ve solved but I don’t know which you haven’t [VP ].
(11)
*I know which problems you’ve solved but I don’t know which you haven’t done so.
This contrast makes sense if ellipses have internal syntactic structure and proforms don’t. On the assumption that proforms are lexical items, it could be reduced to whatever makes words ‘‘islands’’ for anaphoric processes (see the ‘‘anaphoric island’’ constraint on words discussed in Postal 1969a, but also Sproat 1988). That unpronounced strings derive their meanings from contexts, just as pronouns do, could be related to the fact that deaccented material is also sensitive to contexts (see Rooth 1985 and Schwarzschild 1999, for example) and, of course, unpronounced strings are necessarily deaccented. The anaphoric nature of deaccented material can be exemplified in connected discourses like those in (12). (12)
a. James ate the yellow banana. No, he ate the BLACK banana. b. James ate the yellow banana. #No, he MASHED the black banana.
In (12a), everything in the second sentence is deaccented except black and this partition corresponds perfectly to what is new to this sentence and what has already been introduced in the first sentence. That is, everything that is deaccented in the second sentence of (12a) can be found in the first sentence of (12a). This isn’t true for the second sentence of (12b), however, and this results in an ill-formed discourse. In general, deaccented material must convey information that has already been introduced in the discourse.6 Ellipsis, then, could be thought of as the limiting case of deaccenting – a suggestion floated in Sag (1976) and pursued in earnest in Tancredi (1992) – and the many properties it shares with deaccentuation could be removed from the list of things that the ellipsis theorist must explain. To
6
Kyle Johnson
the extent that the anaphoric behavior of ellipsis can be matched to the anaphoric properties of deaccented material, we have evidence in support of equating them.7 Here, then, are two opposing views of ellipsis. In one, ellipses are silent in the same way that silent pronouns are, and their anaphoric behavior derives from their being proforms. In the other, because elided strings are necessarily deaccented ones, their anaphoric properties derive from the principles that govern deaccentuation. Interestingly, there is a way of reading the facts concerning binding into ellipses that suggests both these approaches are correct. The ability of interrogative pronouns to bind into ellipses seems to differ depending on whether the elided phrase is a clause or a VP. The contrast in (2b) and (6) already indicates that there is some difference between these cases. Extraction from an elided clause does not invoke island effects, but extraction from elided VPs does. In fact, extraction from some elided VPs is ungrammatical even when no island is present. These examples, ones like (13), led Sag (1976) to the conclusion that elided VPs are not transparent for binding.8 (13)
*What Sandy carried was the baseball bat, and what Betsy did [VP ] was the catcher’s mitt. (Sag, 1976, (1.3.23): 43)
Perhaps there is a way of seeing cases such as (10), in which an interrogative pronoun does apparently bind into an elided VP, as exceptional. Maybe, for instance, a more cautious examination of (10) (repeated below) would reveal that the ellipsis site does not contain a variable that the interrogative pronoun binds.9 Maybe, in fact, binding into elided VPs is not grammatical, and this contrasts with binding into elided IPs, which are. The proform analysis could be correct for elided VPs, and the radical deaccenting account could be correct for Sluices. Perhaps these two approaches to ellipsis provide a framework for modeling the taxonomy of ellipsis. Jason Merchant’s chapter in this volume leans against this direction. Building on observations in Schuyler (2001), he shows that a key ingredient in the grammaticality of (10), a grammatical example of movement out of an elided VP, is that it introduces a contrastive element in the right position. To see this, consider how (10) compares to the minimally different (14). (10)
I know which problems you’ve solved but I don’t know which you haven’t [VP ].
(14)
?? I know which problems you’ve solved, but Sam doesn’t know which you have [VP ].
Introduction
7
The presence of the contrasting haven’t seems to license extraction from the elided VP. When the contrast in these two clauses is placed elsewhere, as it is in (14), the result is degraded. Merchant shows that an explanation for contrasts of this sort can be credited to a competition between Sluicing and VP Ellipsis. The availability of Sluicing is another thing that distinguishes (14) from (10), as the contrast in (15) shows. (15)
a. *I know which problems you’ve solved but I don’t know which you HAVEN’T solved. b. I know which problems you’ve solved but SAM doesn’t know which you have solved.
Merchant formulates the competition as a constraint that favors larger elisions over smaller ones, and shows how it captures a range of interesting cases. His idea is explored in Lasnik (2001), where some limiting cases are offered, and extended to cases of pronoun binding in Takahashi and Fox (2006). If what is relevant to making examples like (10) grammatical is that they are immune to Sluicing, then there is no reason to believe that what makes them grammatical is that they are not transparent to an outside binder. They still stand as counterexamples to the proform analysis of VP Ellipsis, therefore.10 If (10) is problematic for the proform analysis, there are examples which are problematic for both the proform account and the radical deaccenting proposal. Satoshi Tomioka’s chapter examines a class of such examples discovered by Dan Hardt and Bernhard Schwarz.11 In these examples an elided phrase behaves like a pronoun in being able to receive a sloppy interpretation. This is possible even in contexts where an interrogative pronoun binds into the ellipsis, as in (16).12 (16)
A: When John cooks something, he won’t acknowledge what he can’t [VP]. B: And when he BAKES something, he won’t [VP ] either.
The elided VP in (16A) has the same meaning as cook. This makes the main VP have a meaning equivalent to acknowledge what he can’t cook, and it’s this meaning that serves as the antecedent for the elided VP in (16B). And yet, the elided VP in (16B) can have a meaning equivalent to acknowledge what he can’t bake. It’s as if the antecedent for the elided VP in (16B) has the form in (17), and the pro gets sloppily bound to bake. (17)
[VP acknowledge what he can’t [VP pro] ].
That’s what would be expected under the proform account, of course, but it would seem to be incompatible with the ability what has to bind into
8
Kyle Johnson
[VP pro]. Tomioka provides an account for these examples that preserves the radical deaccenting analysis of ellipsis. His proposal makes crucial use of the idea that syntactic representations are interpreted semantically in ‘‘phases,’’ and that these interpretations have access to the information relevant for fixing sloppy interpretations. Part of resolving the problem posed by examples such as (16) requires understanding the mechanisms which give rise to sloppy interpretations. In a wide range of cases, a pronoun gains a sloppy interpretation just when it is interpreted as a bound variable. In (18a), for instance, his can be interpreted sloppily, but it cannot be in (18b). (18)
a. Every father admires his nephew and every uncle does [VP ] too. ¼ ‘‘every uncle1 admires his1 nephew.’’ b. The woman every father marries admires his nephew and the woman every uncle marries does [VP ] too. 6¼ ‘‘the woman every uncle1 marries admires his1 nephew.’’
To get a sloppy interpretation seems to require being interpreted as a variable. Pronouns can be interpreted as variables for some term when they are c-commanded by that term. Because every father c-commands his in (18a), this pronoun can pick up a new c-commanding binder in the ellipsis site. But because every father does not c-command his in (18b), this pronoun cannot pick up a new binder in the ellipsis site. This restriction on sloppy interpretations presents something of a problem for the cases of sloppy interpretations for ellipses in (16), since the antecedent VP does not c-command the ellipsis site. Tomioka solves this problem by using a semantics for focus that creates a binding relation between a focused item and the terms that get an interpretation that is dependent on that focused item.13 Because the antecedent VPs for the ellipses in cases like (16) are focused, this manufactures the necessary denotations for the sloppy interpretations. Sloppy interpretations of pronouns also arise in contexts where they are not c-commanded by antecedents, and so the exceptionality of the sloppy readings in (16) is not unexpected. In the case of non-c-commanded sloppy pronouns, a famous example of which is (19), one suggestion14 about how they gain a variable-like interpretation is to use Evans’s (1977) proposal that pronouns can be interpreted as hidden definite descriptions. This suggestion would give a way of characterizing the sloppy interpretation of (19). (19)
The police officer who arrested John insulted him, and the one who arrested Bill did [VP ] too. [VP ] ¼ ‘‘insulted Bill’’
Introduction
9
On Evans’s view, the pronoun him contains a hidden description that makes it equivalent to, say, the expression that guy, and this creates the illusion that there is a bound pronoun in the ellipsis. The status of this hidden predicate has been the focus of some attention, and one popular account has it that ellipsis is responsible for hiding it. On this view, personal pronouns like him are definite determiners that come with an elided NP. The elided NP expresses the hidden predicate. If the antecedent conditions on ellipsis are amenable, and the meaning of him is roughly that of the, then this proposal would allow the first clause of (19) to get a representation like that in (20). (20)
The police officer who arrested John insulted [DP him [NP guy the police officer arrested]].
The sloppy reading in (19) could then result from the fact that this elided NP has a meaning that manufactures the illusion of a bound variable. Postal (1969b) was the first to propose this account of personal pronouns, and Elbourne (2001) has a fuller working out of the idea in contexts like those of (19). The existence of sloppy interpretations in examples like (19) has led some to include personal pronouns in the roster of ellipses. In Uli Sauerland’s contribution to this volume, he examines the conditions under which pronouns get a variable-like interpretation because of a hidden predicate, and uses properties of focus to investigate the nature of this hidden predicate. He shows that the hidden predicate has to be credited with playing a role in the sloppy interpretations even of c-commanded pronouns, like those in (8). And he argues that an ellipsis account is not correct, at least not always. One of his arguments, interestingly, turns on the difference in how focus works in cases of personal pronouns and the instances of sloppy interpretations discussed by Tomioka. Like (10), instances of Antecedent Contained Deletion (ACDs) are also cases where an elided VP is bound into, this time by a relative pronoun. In (3), for example, the clause with the ellipsis has a representation like that in (21), where that is to be understood as binding the variable x in something like the way that a l operator does. (3) (21)
She has [VP worked on every ellipsis problem that you have [VP ].] every ellipsis problem that1 you have worked on x1.
The fact that the argument containing this ellipsis must scope out of the VP that serves as antecedent has led some to the conclusion that this antecedent VP must, like the elided VP in (21), have a variable in it. Thus, the clause containing the antecedent VP in (3) should have a representation like that in (22).
10
Kyle Johnson
(22)
[every ellipsis problem that you have]1 she has worked on x1.
This demands a syntax in which the object is outside the antecedent VP, and binds a variable in the object position of this VP. On most accounts, this requisite representation is achieved by covertly moving the object.15 If the antecedent conditions on VP Ellipsis require that the antecedent VP match the elided VP at least to the degree that they contain variables in the same places, then instances of ACD will require that the argument containing the ellipsis get interpreted outside of the antecedent VP. This, in turn, will guarantee that this argument’s scope always includes the antecedent VP, thereby accounting for contrasts like those between (3) and (4). Chris Kennedy examines this hypothesis in his contribution to this volume. He puts together cases like (23a), first discussed in Wasow (1972), with cases like (23b), mentioned in Fiengo and May (1994), and observes that these share a trait that distinguishes them from (3). (3) (23)
She has [VP worked on every ellipsis problem that you have [VP ]]. a. * A proof that God exists does [VP ]. b. * Polly visited a town located in every country that Erik did [VP ].
This trait, he argues, is responsible for their ungrammaticality. To see what that trait is, consider the representations for the clauses containing the elided VPs in (23), shown in (24), and compare them to the representation posited for the antecedent VPs in (23), shown in (25). (24)
a. [a proof that God exists]1 does [VP x1 exist]. b. every country that1 Erik did [VP visit x1].
(25)
a. God1 [VP x1 exists] b. [a town located in every country that Erik did [VP ]]1 Polly [VP visited x1].
Note that (24a) assumes that the elided VP contains a variable bound by the subject; this is necessary to bring the two cases in (23) together. It could be enforced by adopting the now-popular Derived Subjects Hypothesis, which claims that subjects bind a variable within a VP from their surface position,16 and something that requires elided VPs to be large enough to include this variable.17 What makes (23) different from (3) is that the variables in the antecedent and elided VPs of (3) are bound by expressions that refer to the ‘‘same’’ objects, whereas the variables in the antecedent and elided VPs of (23) are bound by expressions that don’t. In (23a), for instance, the elided VP (¼(24a)) contains a
Introduction
11
variable bound to a proof that God exists whereas the antecedent VP (=(25a)) contains a variable bound to God. Similarly, in (23b) the variable in the elided VP is bound by every country and the variable in the antecedent VP is bound by a town. By contrast, the variables in both antecedent and elided VP in (3) are bound by every ellipsis problem. Kennedy suggests that a generalization about ACD is that binders of the variables in antecedent and elided VP must be the ‘‘same’’ in some sense to be made precise. This chapter was first circulated in 1994, and this generalization – now known as ‘‘Kennedy’s Generalization’’ – has generated considerable interest. Deriving it has become something of a quest. For some notable recent attempts, see Heim (1997), Jacobson (2004), Sauerland (2004), and a forthcoming paper by Kennedy in Natural Language Semantics. Cases of ACD also arise in cases of Comparative Deletion, such as (26), and these are examined in Winfried Lechner’s chapter. (26)
John wanted to write more plays than Sam D.
The ellipsis in this sentence can have one of two interpretations, and these are tied to the scope of the object containing D. If the object scopes out of the want VP, then the ellipsis can have the interpretation given in (27a), and if it doesn’t, then it can have the interpretation in (27b). (27)
a. D ¼ ‘‘wanted to write’’ b. D ¼ ‘‘wrote’’
This is just the sort of behavior that we witness with ACDs in relative clauses. In (26), however, it is not run-of-the-mill VP Ellipsis that is responsible for what has elided because VP Ellipsis always leaves a finite auxiliary verb behind and there is none in (26). Lechner argues that Gapping is responsible for the ellipsis in (26). There are certain properties that Gapping alone displays, and an important one of these is found in ACD comparatives. Gapping, unlike VP Ellipsis, for instance, cannot elide an embedded VP, as (28a) indicates, and the same is true for the ellipsis in comparatives, as Lechner illustrates with (28b). (28)
a. * Some wrote plays and others wanted to D novels. b. * John wrote more plays in a month than Sam wanted to D in a year.
After formulating this condition on Gapping exactly, Lechner goes on to show how it interacts with extraposition to derive the fact that in cases such as (29), where the comparative is related to a subject, the ellipsis can only take the root VP as its antecedent.
12
Kyle Johnson
(29)
More people wanted to write plays than D poems. D ¼ ‘‘wanted to write’’ D 6¼ ‘‘writes’’
The use of Gapping to derive the ellipses in comparative constructions offers a new method of capturing some of the constraints on the scope of the comparative operator, and this allows Lechner to cut through some of the problems that presently make treating these constraints so difficult.18 The correspondence between the scope an argument takes and what an ellipsis inside it finds as antecedent has been used to support the view that the syntax of quantificational expressions involves bound variables. So too has Kennedy’s Generalization, since it seems to make direct reference to variables in antecedent and elided VPs. In my description of these phenomena, I’ve relied on this characterization of the relationship between quantificational expressions and the VPs that they are in. On this characterization, the surface representation is not what the semantic interpretation analyzes. Instead, there are hidden syntactic operations which form representations in which the quantificational expressions are placed in the positions from which their scopes are computed, and they bind a variable that resides in their surface position.19 But this is not the only way of conceiving of these relationships, and Pauline Jacobson has famously defended a different direction. On this alternative, the surface syntactic representations are what the semantic procedures analyze. In place of enriching the syntactic representations, so that a single surface string is associated with a family of parses, Jacobson’s view would enrich the semantic processes, so that a single string of words is associated with a family of denotations. In her contribution to this volume, she gives a clear and concise explication of this alternative and takes up many of the cases adduced on behalf of the variable-based view and shows how they could be characterized without variables. While she concentrates her discussion on cases involving VP Ellipsis, the techniques she describes might extend in interesting ways to cases like those that Lechner describes. Jacobson’s work brings out how the proper treatment of these matters can be leveraged into answering foundational questions about linguistics. The difference that Lechner describes in (28) convinces me that Gapping is not the result of the same mechanism that produces VP Ellipsis and Sluicing, neither of which are sensitive in the same way that Gapping is to embedding. In my contribution to this volume I explore this difference, and others, that have led me to suggest that Gapping is not an ellipsis at all, but is instead a perverse instance of movement. What Gapping does share with VP Ellipsis is the ability to form unorthodox constituents. In the
Introduction
13
case of VP Ellipsis this gives rise to those cases called Pseudogapping.20 Both Pseudogapping and Gapping treat the bracketed strings in (30) as constituents. (30)
a. Pseudogapping: Jeri might [want to eat] popcorn, but she doesn’t want to eat kumquats. b. Gapping: Some [want to eat] popcorn, and others want to eat kumquats.
This is one illustration of where ellipsis does not seem to give the same image of constituency that other syntactic phenomena do. I explore some of the problems raised by the standard proposals for forming these constituents, and offer a solution to those problems. My proposal would put Sluicing, VP Ellipsis and Pseudogapping into one class, and Gapping into another. Chris Wilder, in his chapter, argues that Right Node Raising is also not an instance of ellipsis in the sense that Sluicing, Pseudogapping, and VP Ellipsis are. Early accounts of Right Node Raising held that they involved rightward movement ‘‘across the board’’ out of the conjuncts. So that an example like (31a) was the rightward analogue of the cleft in (31b). (31)
a. John read, and Mary admired, this book on ellipsis. b. It’s this book on ellipsis that John read and Mary admired.
But there are some difficulties with this view. The material that undergoes Right Node Raising need not be a string that is otherwise susceptible to movement, as Abbott (1976) noted. (32)
a. * It’s twenty cakes in less than an hour that Mary baked and George frosted. b. Mary baked, and George frosted, twenty cakes in less than an hour. (Abbott 1976, (4): 639)
And Right Node Raising does not seem to be sensitive to (certain) islands, as the contrast in (33) shows. (33)
a. * It’s this nifty book on ellipsis that I met the man who’d read and you met the man who hadn’t read. b. I met the man who’d read, and you met the man who hadn’t read, this nifty book on ellipsis.
For these reasons, many working on Right Node Raising have suggested that it is an instance of ellipsis, rather than an instance of rightward
14
Kyle Johnson
movement.21 What no ellipsis-based account has been able to derive, however, is the fact that the material that is Right Node Raised must be found at the far right edges of both conjuncts. So, for instance, the first object of a double object construction cannot Right Node Raise, as (34) illustrates, because this object cannot stand at the right edge of a VP, as (35) demonstrates. (34)
*John gave a book, and Mary gave a magazine, our friends at the coffeehouse.
(35)
*John gave a book our friends at the coffeehouse.
Wilder’s proposal derives this feature of Right Node Raising. His proposal is that Right Node Raising arises as the result of linearizing coordinations.22 In particular, he uses Kayne’s Linear Correspondence Axiom23 to derive not only the ‘‘right edge’’ effect of Right Node Raising, but several of its other peculiar properties. Kayne’s Linear Correspondence Axiom is a Linearization scheme that is very sensitive to the geometry of phrase markers and, as a consequence, imposes tight restrictions on what phrase markers can look like. Wilder recrafts the Linear Correspondence Axiom so that it allows multidominance, and then shows how Right Node Raising is a special instance of linearizing a string that has two (or more) mothers. The proposal is ingenious and derives not only the availability of Right Node Raising but also its central properties. If he and I are correct, then there are three separate processes responsible for the list of ‘‘ellipses’’ in (5). There is ‘‘true’’ ellipsis, which forms VP Ellipsis, Sluicing, and (perhaps) also NP Ellipsis. These ellipses have many of the properties of pronominal anaphora or deaccentuation; they are a kind of discourse anaphora. There is movement which, in its across-theboard form, is responsible for producing Gapping and the ellipses in the comparative construction. And there is Linearization, which produces Right Node Raising. These last two processes have their ellipses producing consequences in the very specialized environment of coordinations and comparatives. They are not part of the family of discourse anaphora, and the conditions that govern how they find antecedents are therefore dramatically different. While this is a very modest start on the project of forming a taxonomy from the list in (5), it is a start. I hope that in these chapters, the reader will find material from which answers to the many remaining questions can be fashioned.
2
VP Ellipsis and constraints on interpretation Daniel Hardt Copenhagen Business School
2.1
Introduction
Beginning with early work such as Ross (1967) and Sag (1976), the interpretation of ellipsis has been an enduring preoccupation of linguistics research. Many researchers have attempted to use ellipsis as a tool to gain insight into otherwise hidden structures and mechanisms underlying interpretation. According to Sag’s account of VP Ellipsis (VPE), the facts of ellipsis reflect logical aspects of the representation of pronouns, which are ambiguous between a lambda-bound and referential reading. On this influential view, ellipsis reveals an intricate machinery of lambda binding, pronoun indexing, and scope relations. Subsequent research has shed doubt on this pleasing picture – it has been shown that ellipsis interpretation is more flexible than Sag’s theory permits, and this has lead to the proposal of numerous alternative theories. In Hardt (1993), I proposed that VPE resolution be completely free, at least with respect to pronoun interpretation. This makes it possible to represent pronouns in a uniform way, without indexing or ambiguity. If VPE resolution is indeed free in this sense, then any constraints that appear to arise in ellipsis must instead be imposed by more general mechanisms. Perhaps ellipsis is revealing in a rather different sense than suggested by Sag: instead of providing insight into the intricacies of the logical form of sentences, ellipsis emerges as a window onto the general mechanisms governing the interpretation of multi-sentence discourse. In this chapter I will use ellipsis to investigate two potential mechanisms for general, intersentential constraints on interpretation: Background Matching, based on Rooth (1992a); Dalrymple et al. (1991), and Parallel Dependencies, based on Fiengo and May (1994). The Background Matching approach attempts to determine a common semantic Background to the ellipsis and antecedent clause, while the Parallel Dependencies approach places rather involved constraints on the structural material connecting dependent expressions with antecedents. 15
16
Daniel Hardt
Background Matching was originally formulated by Rooth (1985) as an account of focus, and it thus has clear motivation independent of its application to ellipsis. Parallel Dependencies, on the other hand, appears to be tailor-made to account for complex cases of ellipsis, and it is not clear whether it can be given independent, general motivation. Furthermore, Background Matching is conceptually clearer and simpler than the Parallel Dependencies approach. Unfortunately, it appears that the complex Parallel Dependencies approach is necessary to account for certain complex ellipsis examples, involving multiple occurrences of pronouns, as has been argued by Fiengo and May (1994) and Fox (2000). No convincing account of these cases has been given using Background Matching. I will argue however, that a simple version of Background Matching can indeed account for such cases, with proper attention to aspects of their logical structure that have previously been ignored. In section 2.2, I will present a free approach to VPE resolution, and show how this fits together with a somewhat modified version of the Background Matching and Parallel Dependencies approaches. In section 2.3, I consider the complex cases that have proved problematic for previous version of Background Matching, including the so-called two-pronoun puzzle. I review two accounts, Fiengo and May (1994) and Fox (2000), in which versions of the Parallel Dependencies approach are used to analyze the two-pronoun puzzle. I show that the Background Matching approach as presented here can also correctly analyze these examples. I show that this modified Background Matching approach can also correctly analyze a complex example involving embedded ellipsis occurrences. Finally, I suggest that the Background Matching operation is best viewed as the reflex of a particular discourse structure. As evidence for this, I show that Background Matching does not necessarily apply when the discourse structure involves a relation other than parallelism. 2.2
Two general constraints: Background Matching and Parallel Dependencies
2.2.1
Free VPE resolution
As in Hardt (1993), I assume that the VPE resolution operation itself imposes no constraints on the interpretation of pronouns within the antecedent. Pronouns are represented in a simple uniform way, as instructions to retrieve an appropriate semantic object from the current context. In Hardt (1993), this idea was implemented using a processing model of Store and Retrieve instructions. Here I will implement the same idea using LF representations. Pronouns will have no index at LF, and VPE can be
VP Ellipsis and constraints on interpretation
17
resolved by a VP-copy operation.1 Since pronoun – antecedent relations are not determined prior to VPE resolution, the only constraints on interpretation must arise from external mechanisms. In what follows, we consider Background Matching and Parallel Dependencies. These mechanisms will apply to structures after VP-copy has applied and after pronoun – antecedent relations have been determined. 2.2.2
Background Matching
Several authors (Dalrymple et al. 1991; Rooth 1992a; Tancredi 1992; Asher 1993; Prust ¨ et al. 1994; Hobbs and Kehler 1997; Asher et al. 2001) have argued that interpretation is constrained by an operation that attempts to match Background material between parallel clauses in a discourse. Essential to this view is Background Construction, which can be thought of as follows: given two related clauses C1 and C2, construct a Background from these clauses by substituting variables for any material that differs between C1 and C2. Here we focus on the case where C2 contains a VPE occurrence, while C1 contains the antecedent VP, which is copied by the VP-copy operation described above. Consider the following example: (1)
John saw his mother. Bill did too. (saw his mother)
Following Dalrymple et al. (1991), we produce a Background from the antecedent clause (ANT) by means of lambda abstraction, and the Background is then applied to the parallel element(s) in the ellipsis clause. A given reading for an ellipsis occurrence is only permitted if it can be derived in this way. The following is a simplified version of the Background Construction algorithm from Dalrymple et al. (1991) (henceforth the DSP algorithm):2
(STEP 1) prefix ANT with l x (STEP 2) replace primary occurrence s with x (STEP 3) for each pronoun p in ANT, optionally replace p with x if p is coindexed with s For the above example, we have: (STEP 1) l x.John1 saw his1 mother (STEP 2) l x.x saw his1 mother (STEP 3 – optional) l x.x saw x’s mother If the optional STEP 3 is performed the sloppy reading is derived. Otherwise, we derive the strict reading. This is very similar to the approach in Rooth (1992a), where Rooth suggests that general constraints on focus might also be applied as a constraint on ellipsis interpretation. The alternative semantics theory of
18
Daniel Hardt
focus is that a focused element is permitted if it is contained in a clause that matches an antecedent clause. Rooth defines the focus value of an expression as the set of alternative expressions, where each alternative expression is constructed by replacing the focused element with some alternative (i.e., an element of the same semantic type.) Applied to ellipsis we have the following definition (Rooth 1992b):
Ellipsis (E) requires that there be some phrase P containing E and some antecedent phrase ANT in the discourse, such that [[ANT]] FðPÞ.
FðPÞ ¼ fP0 j9x:P0 ¼ P½e=xg, e focused in P. The set of all P0 where occurrences of the focused element e is replaced by some element of the appropriate type. According to Rooth’s approach, we assume there is focus on Bill. For the strict reading (where Bill and John both see John’s mother), the focus value of VPE is then the set of propositions of the form: x saw John’s mother. The ANT meaning John saw John’s mother is clearly an element of this set, so the reading is permitted. For the sloppy reading, the VPE focus value is the set of propositions of this form: x saw x’s mother. Again, the ANT meaning John saw John’s mother is clearly an element of this set, so the reading is permitted. The empirical content of the Dalrymple et al. (1991) and Rooth approaches appears quite similar. I will ignore several differences in presentation which may or may not have empirical content – for example, Rooth uses focus to determine the element(s) which are to be replaced by variables, while Dalrymple et al. (1991) leave this to a general processing of determining parallel elements, which would presumably take into account focus as well as general contextual reasoning. For presentational purposes, I will define the Background with respect to LF structures, although Rooth explicitly defines his approach with respect to semantic denotations. Although this difference undoubtedly has empirical significance, it is not relevant to my purposes in this chapter. Another difference is that Rooth constructs the Background from the ellipsis clause, while Dalrymple et al. (1991) construct it from the antecedent clause. Again, the empirical significance of this is unclear. However, I would like to propose a constrained version of Background Construction, which is more conveniently described using Background Construction from the antecedent clause, as is done by Dalrymple et al. (1991). Recall that the DSP algorithm permits a choice with respect to each coreferential pronoun. In example (1), this leads to two possible readings, since there is one coreferential pronoun. In general, the algorithm will generate 2n distinct Backgrounds, where n is the number of coreferential pronouns. I will propose a more restrictive variant, where either all or none
VP Ellipsis and constraints on interpretation
19
of the coreferential pronouns are replaced with the lambda-bound variable. Background Construction (constrained version):
(STEP 1) prefix ANT with lambda x (STEP 2) replace primary occurrence s with x (STEP 3 – optional) replace all pronouns p in ANT with x (for p coindexed with s) To illustrate the constrained version of Background Construction, consider the following example: (2)
John said he saw his mother. Bill did too.
The original DSP algorithm produces the following four Backgrounds:
l x.x said x saw x’s mother l x.x said John saw John’s mother l x.x said x saw John’s mother l x.x said John saw x’s mother
According to the more restrictive variant, we produce only two Backgrounds:
l x.x said x saw x’s mother l x.x said John saw John’s mother We will return to this issue in section 2.2.3, when we consider examples with multiple pronoun occurrences. 2.2.3
Parallel Dependencies
The intuition behind Parallel Dependencies can be stated like this: ellipsis resolution always permits strict identity of pronouns, while sloppy identity is only permitted if the sloppy pronouns are resolved in a parallel way to each other. While this intuition can be found with many authors, I will focus on the Parallel Dependencies approach due to Fiengo and May (1994). In this theory, elided material must be a syntactic copy of the antecedent, apart from index change.3 Index change, or sloppy identity, is only permitted if there is a parallel dependency, defined as follows (Fiengo and May 1994: 51): (i) A dependency is a sequence of syntactic categories connecting a dependent category with its antecedent. It is represented as follows: 5S; I; SD4, where S is a sequence of category labels bearing the index I, and SD is a sequence of category labels of the material connecting the indexed elements.
Completeness Requirement:4 The sequence S must include all elements of the SD which bear the index I.
20
Daniel Hardt
(ii) Two dependencies are identical if ‘‘they vary at most in the value of I’’ (the index) (Fiengo and May 1994: 95). In other words, they must have the same sequence of elements and the same structural description. Note that the relevant notion of identity here is identity of category labels, not of lexical items; although the sequence is typically written as a sequence of lexical items, it is their category labels which are significant. To demonstrate the analysis, let us consider again example (1), repeated here.5 (3)
John1 saw his1 mother. Bill2 did too. (saw his2 mother)
The strict reading is always permitted. The sloppy reading is portrayed by the above indexing. To permit this reading, we must show that the two occurrences of his are in parallel dependencies. The two dependencies are the following: 5ð½JohnNP ; ½hisNP Þ; 1; 5½JohnNP ½sawV ½hisNP 44 5ð½BillNP ; ½hisNP Þ; 2; 5½BillNP ½sawV ½hisNP 44
Recall the conditions on i-copy: both the sequence and the structural description must be identical in terms of category labels. (I include the lexical items only for the sake of readability.) Clearly this is the case above: the two sequences are NP NP, and the two SDs are NP V NP. Thus sloppy identity is correctly permitted. 2.3
Complex examples
In this section, we consider complex examples which have previously been accounted for by Parallel Dependencies, but not by Background Matching. I will show that Background Matching can in fact successfully account for these well-known cases. 2.3.1
The two-pronoun puzzle
We turn now to the two-pronoun puzzle, where the antecedent VP contains two pronouns (Dahl 1974). (4)
John said he saw his mother and Bill did too.
One might expect there to be four possible readings, since each of the two pronouns could be either strict or sloppy. In fact, only three of these four readings are possible, as shown by the following list.
Readings: 1. JJ 2. BB
VP Ellipsis and constraints on interpretation
21
3. BJ 4. *JB Note that Sag’s theory incorrectly permits the fourth reading, since it permits a representation where the first pronoun is referential and the second is lambda bound. (5)
[John1, l x.x said he1 saw x’s mother]
(6)
[Bill2, l x.x said he1 saw x’s mother]
For similar reasons, both DSP and Rooth’s theory also incorrectly permit the fourth reading. 2.3.1.1 The Parallel Dependencies account As shown by Fiengo and May (1994), the Parallel Dependency approach correctly rules out the fourth reading, while permitting readings 1–3. To see this, let us examine readings 3 and 4.
Reading 3: John1 said he1 saw his1 mother, Bill2 [said he2 saw his1 mother] too. Consider the two pronoun occurrences at the ellipsis site. The first pronoun, he2, has changed its index, and thus we must determine whether it participates in a parallel dependency. The following two dependencies are i-copies, differing only in the value of the index – both the sequences and the structural descriptions are identical categories. 5ð½JohnNP ; ½heNP Þ; 1; 5½JohnNP ½saidV ½heNP 44 5ð½BillNP ; ½heNP Þ; 2; 5½BillNP ½saidV ½heNP 44
Consider now Reading 4: (7)
John1 said he1 saw his1 mother, Bill2 [said he1 saw his2 mother] too.
Here it is the second pronoun, his2, which has changed index, and therefore must participate in a parallel dependency. Thus we must find in the first clause a dependency that is parallel to: 5ð½BillNP ; ½hisNP Þ; 2; 5½BillNP ½saidV ½heNP ½sawV ½hisNP 44
There is no such dependency. It might appear that we have the following dependency: 5ð½JohnNP ; ½hisNP Þ; 1; 5½JohnNP ½saidV ½heNP ½sawV ½hisNP 44
But the structural description contains an element indexed with I that is not included in the sequence – namely, he1. This is not permitted, as it violates the Completeness constraint. If we include the omitted element, we get: 5ð½JohnNP ; ½heNP ; ½hisNP Þ; 1; 5½JohnNP ½saidV ½heNP ½sawV ½hisNP 44
22
Daniel Hardt
But this dependency doesn’t match – it has a sequence of three elements, while the ‘‘Bill – his’’ dependency has a two-element sequence. 2.3.1.2 Fox’s account Fox (2000) adopts Fiengo and May’s claim that index change requires Parallel Dependencies, but combines this with an economy-based preference for local links. Interestingly, Fox’s Locality constraint plays exactly the same role as Completeness does in Fiengo and May’s account. Recall that, for Reading 4, the dependency in the ellipsis clause is the following: 5ð½BillNP ; ½hisNP Þ; 2; 5½BillNP ½saidV ½heNP ½sawV ½hisNP 44
To satisfy parallelism, we need the following dependency from the antecedent clause: 5ð½JohnNP ; ½hisNP Þ; 1; 5½JohnNP ½saidV ½heNP ½sawV ½hisNP 44
This dependency is not permitted for Fox, because the link between his and John is not local, and the following dependency is therefore preferred: 5ð½JohnNP ; ½heNP ; ½hisNP Þ; 1; 5½JohnNP ½saidV ½heNP ½sawV ½hisNP 44
But with this dependency, parallelism fails, and thus Reading 4 is not permitted. Thus, for Fox, Reading 4 would either violate parallelism or would have a non-local link. Note that the non-local link ruled out by Fox would be ruled out by Fiengo and May’s Completeness constraint, for essentially the same reason: the existence of a coindexed pronoun ‘‘he’’ intervening between ‘‘John’’ and ‘‘him.’’ 2.3.1.3 Solving the two-pronoun puzzle with Background Matching Here I will show that Background Matching can indeed account for the twopronoun puzzle. The key to this result is that I will permit any non-atomic indexed expression [. . .]i to be Collapsed, i.e., to be replaced with an atomic expression ei. (I assume that maximal projections NP, VP, and S can receive indices.) This makes it possible to ignore the internal content of indexed expressions in ellipsis resolution. It seems to me that this is an entirely natural option to permit prior to ellipsis resolution. Indeed there are a variety of ways that this effect can be achieved.6 For simplicity, I will simply allow Collapsing as an option in the construction of the Background. Recall also that we have modified the Background Matching approach so that it only generates two readings in the twopronoun example; either both pronouns are replaced with bound variables, or neither are.
VP Ellipsis and constraints on interpretation
23
Consider now Reading 3: (8)
John1 said he1 saw [his1 mother]2, and Bill3 did (said he3 saw [his1 mother]2) too.
We make use of Collapsing to derive the following representation for the antecedent clause: ‘‘John1 said he1 saw [e]2.’’ We can then derive the following Background: l x. x said x saw e2.
Applying P to Bill, we derive the following representation, showing that Reading 3 is permitted: (9)
John1 said he1 saw [his1 mother]2, and Bill3 (said he3 saw [e]2) too.
Consider now Reading 4: (10)
John1 said he1 saw [his1 mother]2, and Bill3 did (said he1 saw [his3 mother]4) too.
It is easy to see that Reading 4 cannot be produced. First, note that the Collapsing option is not relevant here; if we collapse [his1 mother]2 to [ ]2, we would only be able to produce readings referring to John’s mother. Thus we must produce a Background in which the pronoun his from the antecedent clause is lambda-bound. However, because of our restriction on Background Matching, this means that he must also be lambda-bound, meaning we will produce Reading 2. Thus, Reading 4 is correctly ruled out on our version of Background Matching. As several authors have noted (see Fiengo and May, 1994), the twopronoun puzzle is sensitive to the particular positions of the two pronouns; embedding of the first pronoun causes the fourth reading to reemerge, as in the following variant of (4): (11)
John1 said [his1 mother]2 [saw him1]3, and Bill4 did too.
To derive Reading 3 (‘‘Bill said Bill’s mother saw John’’), we produce the Background as follows:
(ANT Clause) John1 said [his1 mother]2 [saw him1]3 (Collapse VP 3) John1 said [his1 mother]2 (did) [e]3 (Lambda Prefix) l x.x said [his1 mother]2 (did) [e]3 (Replace pronoun) l x.x said [x’s mother] (did) [e]3
When applied to ‘‘Bill,’’ the Background produces the following representation: (12)
Bill4 said [Bill4’s mother] (did) [e]3
24
Daniel Hardt
Since e3 represents the VP [saw him1], this captures Reading 3. This example illustrates that our Collapsing option can apply to VPs as well as NPs. Consider now Reading 4 (‘‘Bill said John’s mother saw Bill’’). The derivation involves the Collapsing of [his1 mother]2, as follows:
(ANT Clause) John1 said [his1 mother]2 [saw him1] (Collapse NP 2) John1 said [e]2 [saw him1] (Lambda Prefix) l x.x said [e]2 [saw him1] (Replace pronoun)l x.x said [e]2 [saw x]
When applied to ‘‘Bill,’’ the Background produces the following representation: (13)
Bill4 said [e]2 [saw Bill4]
Since e2 represents the NP [his1 mother]2, this captures Reading 4. 2.3.2
A symmetry effect
There are additional examples, due to Fox (2000, page 119), where Fox argues that a similar constraint is at work. (14)
Every boy1 said that Mary2 liked her2 dog. Well Mary2 did too. (said that she2 liked her2 dog)
(15)
Every boy1 said that he1 liked his1 dog. Well Mary2 did too. (said that she2 liked her2 dog)
(16)
Every boy1 said that he1 liked Mary2’s dog. Well Mary2 did too. (said that she2 liked her2 dog)
(17)
*Every boy1 said that Mary2 liked his1 dog. Well Mary2 did too. (said that she2 liked her2 dog)
Fox points out that reading (17) is ruled out by Locality in his Parallel Dependencies account. To see this, note that her in the ellipsis clause must participate in a parallel dependency. There are two choices: (i) 5ð½sheNP ; ½herNP Þ; 2; 5½sheNP ½likedV ½herNP 44 (ii) 5ð½MaryNP ; ½herNP Þ; 2; 5½MaryNP ½saidV ½sheNP ½likedV ½herNP 44
It is clear that choice (i) is not permitted; the structurally parallel material in the antecedent clause is <Mary2 liked his1>, but his is not dependent on Mary. For choice (ii), we can find the following parallel dependency in the antecedent clause: 5ð½Every boyNP ; ½hisNP Þ; 1; 5½Every boyNP ½saidV ½MaryNP ½likedV ½hisNP 44
VP Ellipsis and constraints on interpretation
25
However, note that choice 2 contains an element in its SD ([she]NP) that bears index 2. Thus it is ruled out for familiar reasons: Completeness in Fiengo and May’s version, and Locality in Fox’s account. On the proposed Background Matching account, it appears that (17) would incorrectly be permitted. The following Background can be produced from the antecedent: (18)
l x.x said that Mary2 liked x’s dog
Applied to Mary, this gives the reading in question. These examples point up the fact that the Parallel Dependencies account is completely symmetrical between antecedent and ellipsis clause. On the other hand, Background Matching (as formulated here) is not symmetrical. This can be demonstrated by applying Background Matching in the opposite direction. That is, we can extract the Background from the ellipsis clause instead of from the antecedent clause. We consider the case where the ellipsis clause is resolved as follows: (19)
Mary2 said that she2 liked her2 dog
There are the following two possible Backgrounds: (i) l x.x said that x liked x’s dog (ii) l x.x said that Mary2 liked [her2 dog]3 These readings, applied to every boy, give the readings in (14) and (15). The reading in (16) is derived using the outer index, so that the Background property is l x.x said that x liked e3. There is no way to derive Reading 4, given that we are constructing Backgrounds from the ellipsis clause. Thus, a symmetrical version of Background Matching would correctly capture the above judgments. In other words, we would require that a given reading must be derivable both using a Background extracted from the ellipsis clause, and from the antecedent clause. We will leave for future research the question of whether this symmetrical version should be used in all cases, or perhaps a subset of cases. 2.3.3
A six-reading example
The following example, studied by many researchers (Gawron and Peters 1990, Dalrymple et al. 1991, Asher et al. 2001), motivates an additional application of the Collapsing option. I show here that our proposed version of Background Matching successfully accounts for constraints that are not captured by previous versions of Background Matching or Parallel Dependencies. Consider (20).
26
Daniel Hardt
(20)
John revised his paper before the teacher did (revise his paper), and Bill did (revised his paper before the teacher did revise his paper) too.
As pointed out by Dalrymple et al. (1991), there are six conceivable readings, which can be represented as follows: Readings: (i) JJJJ (ii) JTBT (iii) JJBB (iv) ?JTJT (v) ?JJBJ (vi) *JJJB Note that there are three copies of the pronoun his produced by ellipsis resolution. Thus JJJJ represents the reading where all copies are interpreted as John, and so on. While there is disagreement in the literature about the acceptability of these six readings, it is generally agreed that (i)–(iii) are completely acceptable, while (vi) is not acceptable. I will give (iv) and (v) an intermediate status. We will set aside Readings (i) and (ii), which receive straightforward derivations in all the approaches we are considering. Consider Reading (iii): (21)
[[John1 revised his1 paper]S1a before [the teacher2 did revise his1 paper]S1b ]S1, and [[Bill3 did revise his3 paper]S2a before [the teacher2 did revise his3 paper]S2b]S2 too.
We assume that the following constraints are imposed by Background Matching: S1a must match S1b, S2a must match S2b, and S1 must match S2.
S1a, S1b: Background ¼ l x.x revised his1 paper S2a, S2b: Background ¼ l x.x revised his3 paper S1, S2: Background ¼ l x.x revised x’s paper before the teacher2 revised x’s paper We can see that a common Background can be identified for each of the required clause pairs. We turn now to Reading (iv): (22)
[[John1 revised his1 paper]S1a before [the teacher2 did revise his2 paper]S1b ]S1 , and [[Bill3 did revise his1 paper]S2a before [the teacher2 revise his2 paper]S2b ]S2 too.
Again we attempt to match three pairs of clauses:
VP Ellipsis and constraints on interpretation
27
S1a, S1b: Background ¼ l x.x revised x’s paper S2a, S2b: Background ¼ ?? S1, S2: Background ¼ l x.x revised his1 paper before the teacher2 revised his2 paper Here, matching of S2a and S2b fails. To see this, consider the options in producing the Background. First we prefix S2a with the lambda binder, resulting in: (23)
P ¼ l x.x revised his1 paper.
We cannot replace his1 with x, since his1 is not coreferential with the subject Bill. Thus P applied to the teacher will produce the reading ‘‘the teacher revised his1 paper.’’ However, there is one way to produce the desired reading, if we include time variables in the logical representation, and allow the Collapsing option to apply here as well. (24)
[[John1 revised his1 paper]S1a before [t5 [the teacher2 did revise his2 paper]S1b ]S1 , and [[Bill3 did revise his1 paper]S2a before t5 ] too.
This representation is meant to be analogous to the following variant: (25)
[[John1 revised his1 paper]S1a before the time t5 that [the teacher2 did revise his2 paper]S1b]S1, and [[Bill3 did revise his1 paper]S2a before (then)t5] too.
If this option is taken, there is no longer an issue of matching S2a and S2b, so the reading is now permitted. Reading (v) can be similarly derived: (26)
[[John1 revised his1 paper]S1a before [t5 [the teacher2 did revise his1 paper]S1b ]S1 , and [[Bill3 did revise his3 paper]S2a before (then)t5 ] too.
Note that Reading (vi) can still not be derived. This is because the temporal index associated with S1b cannot be identified with that associated with S2b. Thus, all three pairs of clauses must match, and there is no Background that can be constructed for S1b and S2b. The Parallel Dependencies approach will accept Readings (i)–(iii), and rule out (iv)–(vi). To see this, consider again Reading (iii): (27)
[[John1 revised his1 paper]S1a before [the teacher2 did revise his1 paper]S1b ]S1 , and [[Bill3 did revise his3 paper]S2a before [the teacher2 did revise his3 paper]S2b ]S2 too.
We assume Parallel Dependencies must be determined for the same three pairs of clauses as with Background Matching. For the pairs 5S1a, S1b4
28
Daniel Hardt
and 5S2a, S2b4 there are only strict pronouns, so there are no dependencies to compare. For the pair 5S1, S24, there are two sloppy occurrences of his in S2 that must be licensed by Parallel Dependencies. Both occurrences can be included in a single dependency, that has a corresponding identical dependency in S1: 5ð½JohnNP ; ½hisNP ; ½hisNP Þ; 1; 5½JohnNP ½revisedV ½hisNP ½paperNP ½beforeADV ½theDET ½teacherN ½revisedV ½hisNP 44 5ð½BillNP ; ½hisNP ; ½hisNP Þ; 3; 5½BillNP ½revisedV ½hisNP ½paperNP ½beforeADV ½theDET ½teacherN ½revisedV ½hisNP 44
Readings (iv)–(vi) are ruled out. Consider Reading (iv): (28)
[[John1 revised his1 paper]S1a before [the teacher2 did revise his2 paper]S1b ]S1 , and [[Bill3 did revise his1 paper]S2a before [the teacher2 did revise his2 paper]S2b ]S2 too.
Consider the pair 5S2a, S2b4. Here, we have to match the dependencies involving his1 and his2. But this clearly cannot be done. These are the corresponding dependencies: 5ð½JohnNP ; ½hisNP ; ½hisNP Þ; 1; 5½JohnNP ½revisedV ½hisNP ½paperNP ½beforeADV ½theDET ½teacherN ½revisedV ½hisNP ½paperNP ½andCC ½BillNP ½revisedV ½hisNP 44 5ð½teacherNP ; ½hisNP Þ; 2; 5½teacherNP ½revisedV ½hisNP 44
The reader can verify that Readings (v) and (vi) are similarly ruled out. 2.4
Motivation for general constraints: discourse structure
I have shown that a free approach to VPE resolution, together with a general constraint on interpretation, can account for important puzzles in interpretation. We examined two candidates for general constraints: Parallel Dependencies and Background Matching. While it has been previously argued that Parallel Dependencies is superior to Background Matching in accounting for constraints on complex cases of ellipsis, I have presented a simple form of Background Matching that successfully accounts for these well-known cases. The question arises: what is the ultimate source of the Background Matching operation? The account of Dalrymple et al. (1991) suggests that Background Matching is tied specifically to the process of ellipsis
VP Ellipsis and constraints on interpretation
29
resolution.7 Rooth’s account would suggest that it applies more generally to the interpretation of focus and Background, with ellipsis being a special case of this. An alternative view, put forward by many authors (Asher 1993; Prust ¨ et al. 1994; Hobbs and Kehler 1997; Asher et al. 1997; 2001) is that the constraints on ellipsis and focus in fact arise from discourse relations such as Contrast and Parallel. It has often been suggested that particles like too signal a Parallel relation, while negation signals Contrast. This would suggest that other discourse relations might show rather different requirements on interpretation than we have seen up to this point, where all the examples involved parallel or contrastive relations. The following example (Hardt 1992a) shows that this is indeed the case: (29)
Every student1 expected Professor Jones2 to like his1 work. In John3’s case, I’m sure she will. (like his3 work)
Here, there is a felicitous sloppy reading, although Background Matching fails (note that Parallel Dependencies also fails). The preposed clause In John’s case places a requirement on the matrix clause, that it be in some sense ‘‘about’’ John. It also seems clear that the relation between the VPE and antecedent clause is not Parallel or Contrast, as has been the case in the other examples we have examined. This suggests that the relevant constraints on interpretation in VPE do not arise from the occurrence of VPE, but rather, they are tied to the existence of particular discourse relations. The nature of these relations is a topic of ongoing research. I have suggested here that ellipsis might be used as a tool to bring the structure of discourse into sharper focus. As a preliminary step in that direction, I have shown that ellipsis facts can be used to show that there is a general discourse process better described as Background Matching than as Parallel Dependencies.
3
Direct compositionality and variable-free semantics: the case of Antecedent Contained Deletion Pauline Jacobson Department of Cognitive and Linguistic Sciences, Brown University, Providence, Rhode Island, USA
Since at least as early as Sag (1976), the phenomenon of Antecedent Contained Deletion (hereafter, ACD) has generally been taken as strong evidence for a level of Logical Form or some other abstract syntactic representation1 mediating between the surface syntax and the modeltheoretic interpretation of a sentence. In particular, it is often taken as a settled matter that ACD shows that at least sometimes a quantified NP in object position is not interpreted ‘‘in situ.’’ Rather, the semantic composition of a sentence like John read every book involves a level of representation in which the object is just a simple pronoun or trace interpreted as a variable. The purpose of this paper is to show that this received wisdom is mistaken.ACD is perfectly compatible with the view that quantified NPs are interpreted ‘‘in situ’’ – i.e. the meaning of read directly combines with the meaning of every book to give a VP meaning.2 What is at stake here is more important than just the question of how to interpret quantified NPs in object position: my main concern is to provide support for two broader hypotheses. The first is the hypothesis of direct compositionality: the syntactic combinatory operations ‘‘build’’ (i.e. define as well-formed) surface syntactic expressions while the semantics works in tandem to supply a model-theoretic interpretation for each expression as it is ‘‘built’’ in the syntax. The second is the stronger hypothesis that the syntax makes no use of abstract representations which can be transformed into other representations by things like substitution Acknowledgments: Some of the material in this chapter appears in my papers in SALT 2, SALT 8, and in Jacobson (2003). I have benefited from discussion of these issues at those conferences, at the Formal Grammar conference in Utrecht (ESSLLI 1999), at the Santa Cruz Ellipsis workshop (2003) and at ESSLLI 2001, UCLA, Northwestern University, UCSD, and Harvard. I would also like to thank Anna Szabolcsi, David Dowty, Jason Merchant, Danny Fox, and Chris Barker for various discussions which have impacted on this chapter.
30
Direct compositionality and variable-free semantics
31
operations. This means that there can be no representations at which positions occupied by quantified NPs on the surface are occupied by just pronouns and/or variables which are then ‘‘converted’’ into quantified NPs. (Throughout I will use the term ‘‘NP’’ rather than ‘‘DP’’ for things like every book (which Bill read)). Let me refer to this stronger view as the hypothesis of Strong Direct Compositionality. (I distinguish this from ‘‘Weak Direct Compositionality’’ for the reasons detailed in note 1 (see also Jacobson 2002).) That is, one can maintain a version of direct compositionality which nonetheless makes use of abstract representations in this syntax, as was done in Montague 1974. I will argue below that Strong Direct Compositionality – coupled with a variable-free semantics as proposed in Jacobson (1992b; 1999; 2000a) – can straightforwardly accommodate ACD. This not only removes one of the most hallowed arguments for a level of LF or other abstract representation, but in fact ACD teaches us more: we will see that the interaction of ACD with Pied Piping (originally discussed in Jacobson 1998) is puzzling for the traditional view, but that the puzzle evaporates under the LF-free and variable-free approach.3 The paper will proceed as follows. Section 3.1 briefly discusses the basic program of and rationale for direct compositionality. Section 3.2 is a stepby-step walkthrough of the standard argument for LF (or other abstract level) based on ACD. This may seem to be a rather elementary exercise, but there is a reason for this: I wish to highlight the fact that the entire argument has as its foundation a particular assumption about how semantic composition proceeds. Although this assumption has enjoyed a good deal of sentimental attachment within the linguistics community, there is as far as I know no a priori reason to accept it. And once it is abandoned, it turns out that the basic cases of ACD are entirely compatible with direct compositionality. Of course the story does not end here, for over the years there has been an incredibly rich body of facts and puzzles uncovered concerning the distribution of ACD. A serious theory, then, will want to cover far more than just the ‘‘basic’’ cases. Obviously it is beyond the scope of this chapter to account for all the puzzles and generalizations – indeed, I know of no single theory that does this (at least not in a natural and non-stipulative way). But section 3.3 will address a few of the interesting cases – enough, hopefully, as to make the direct compositional program a plausible alternative to the (more complex) standard view. Section 3.4 turns to a slightly more elaborate apparent argument for LF based on the interaction of ACD with pronominal binding. But this argument also immediately disappears once one adopts a variable-free semantics. Thus section 3.4.2 outlines the variable-free approach which is advocated in a number of other papers of mine, and section 3.4.3 returns to show how this solves the ACD puzzle. Section 3.5 develops a more
32
Pauline Jacobson
explicit account of the phenomenon of ACD (and VP Ellipsis in general) in the variable-free and LF-free program, which will allow us to fill in some of the details left open in sections 3.2–3.4. Section 3.6 turns to the interaction of ACD with Pied Piping discussed in Jacobson (1998). First, we will note (section 3.6.1) that the semantics of Pied Piping is perfectly straightforward under the direct compositional, variable-free approach. But its interaction with ACD tells us more: it turns out that this interaction is mysterious under the standard view, but is perfectly straightforward under the view advocated here. Finally, section 3.7 contains some concluding remarks, including a very brief discussion of the facts discussed in Kennedy (1994, this volume chapter 5), and Heim (1997) (see also Sauerland 2004) and their implications for the claims made here. 3.1
Direct compositionality
The basic program which I will be advocating in this paper is that of ‘‘direct compositionality’’ (see, e.g., much work in the 1970s and 80s within classical ‘‘Montague Grammar,’’ and more recent work within Categorial Grammar and the Type Logical tradition). Under this view the combinatory syntactic operations define a set of well-formed expressions in terms of the well-formedness of other (generally subpart) expressions, while the compositional semantics works in tandem with this to directly supply a model-theoretic interpretation to each expression as it is ‘‘built’’ in the syntax. (‘‘Built’’ is just a metaphor for ‘‘defined as well-formed.’’) Note that the semantics assigns not some kind of representation (LF) which itself has to subsequently be interpreted. Rather, the semantics directly assigns a model-theoretic interpretation. Thus assume (rather uncontroversially) that each linguistic expression is a triple . We can then think of the rules of grammar as rules which take one or more triples as input and give back a new triple as output. While such rules generally combine two or more expressions, I assume also the existence of unary rules: rules which take a single triple as input and map it into a new triple. Moreover, a rule need not alter all three parts of the input. For example, rules which have gone under the rubric of ‘‘type shift’’ and/or ‘‘category changing’’ rules are a special case of unary rules and happen not to alter the phonology.4 It seems obvious that this general picture is the simplest and most elegant conception of how the grammar works. After all, any theory needs a system of rules which ‘‘build’’ complex syntactic expressions ‘‘from the bottom up.’’ (Again, the terms ‘‘build’’ and ‘‘from the bottom up’’ are metaphoric – these mean only that the well-formedness of larger expressions is specified in terms of the well-formedness of smaller
Direct compositionality and variable-free semantics
33
expressions.) Of course a theory with transformations contains additional rules which map representations into other representations, but nonetheless all theories contain some group of rules (possibly in highly generalized form) which do specify the well-formedness of expressions ‘‘from the bottom up.’’ And any theory needs a compositional semantics which also works ‘‘from the bottom up’’ – where the meaning of larger expressions is given in terms of the meaning of smaller ones. The simplest hypothesis is that the two systems work together. And, as discussed in much more detail in Jacobson (2002), there are other suspicious consequences of divorcing the syntax and the semantics. For example, if the two systems are divorced it is a complete accident that they refer to essentially the same type of objects. That is, just about all theories agree that the compositional semantics interprets strictly local trees (by which I mean a mother and its daughters). This is completely unsurprising if the syntax ‘‘builds’’ expressions in tandem with the semantics interpreting such expressions – for here a local tree is nothing more than a representation of how the rule worked to prove an expression well-formed. (In this view, then, it is actually not quite right to say that the compositional semantics ‘‘interprets’’ trees – for the grammar actually never ‘‘sees’’ trees. Rather, it provides an interpretation of the output (mother) expression on the basis of the interpretation of the input (daughter) expression(s). But a tree is a reasonable pictorial representation of the syntactic and semantic combinatorics.) But if the syntax does its work first – building some sort of representation which is then sent to the semantics – why should the compositional semantics work on local trees? There would be no reason not to expect semantic rules whose inputs were, e.g., large chunks of trees. Of course these remarks are rather abstract – and the ultimate viability of the direct compositionality program depends on just how simple the rules are which build syntactic expressions and assign them an interpretation. If direct compositionality required extremely complex rules to build syntactic expressions then it would not be a particularly interesting alternative to a (hypothetical) theory without direct compositionality but with a relatively small and simple set of rules (and/or principles). But I will argue here and have argued elsewhere that the machinery needed to maintain it is actually quite simple. As to the binary rules, I will be assuming that the rules which combine two expressions either concatenate them (such rules are equivalent to context free phrase structure rules) or allow one expression to be infixed into another by a ‘‘Wrap’’ operation (see, e.g., Bach 1979; 1980; Dowty 1982; Pollard 1984; Jacobson 1987). We will also make use of a small set of unary rules but, as will be documented below, the required rules here are quite simple and very natural (and, as pointed out in note 4, one can always trade in a unary rule for a ‘‘silent operator’’).
34
Pauline Jacobson
3.2
The received wisdom on ACD
Consider a run-of-the-mill ACD example: (1)
John will read every book which Bill will.
My purpose in this section is to walk step-by-step through the reasoning which leads to the received position that ACD necessitates a level of representation (e.g. LF) at which quantified NPs (or DPs) are ‘‘pulled out’’ of the main sentence and where these bind a variable (and/or trace) in the argument position in the sentence. This will allow us to pinpoint exactly where the fallacy lies. But first, we need three background pieces: the analysis of VP Ellipsis in general, the analysis of auxiliaries like will, and the semantic composition of relative clauses. 3.2.1
Background pieces
Turning to the nature of VP Ellipsis, there are any number of accounts of this phenomenon. Some involve deletion and some interpretation, and some require LF identity of the ellipsis site and its antecedent while others require meaning (not LF) identity between the two. We will make a more specific commitment to one way to think about VP Ellipsis in section 3.5, but for the moment we need not make any decision about these matters. All that is immediately relevant is the standardly made assumption that there is a condition to the effect that the ellipsis site (or, the ‘‘missing’’ material which is supplied by interpretation) is identical in meaning and/or LF to some other linguistically overt expression (I will, in fact, reject this assumption in section 3.5, but maintaining it until then will do no harm). Next, we need to say something about the semantic composition of sentences involving auxiliaries in general, as in (2): (2)
John will read Crime and Punishment.
There are two tacks which one could take here. In the ‘‘LF program,’’ it is usually assumed that the meaning of will takes a proposition as argument, and so will is ‘‘raised’’ at the level which inputs the semantics so as to take scope over the entire sentence. Under this view, its type is (hence maps a set of times into a set of times). More particularly, we can take (3) as a reasonable first approximation of its meaning (I use the notation a0 to indicate the meaning – in model-theoretic terms – of a linguistic expression a; I do not intend the prime-notation to indicate some translation into another language like IL): (3)
will0 ¼ lp[li[9i0 [p(i0 ) & i0 > i]]]
Direct compositionality and variable-free semantics
35
(Note that here and throughout I will suppress the world argument. Note also that here I do make explicit the time argument – since it is essential to the meaning of will – but I will suppress it wherever not essential. This means that I will go back and forth between talking about, for example, VP-type meanings as <e,t> and <e,>.) However, one could also treat will0 as mapping a property into a new property – hence of type ,<e,>>. Under that view, its meaning is simply a ‘‘fancied-up’’ version of (3), which is given in (4): (4)
lP[lx[li[9i0 [P(x)(i0 ) & i0 > i]]]]
Incidentally, it is worth pointing out that functions which take propositional arguments (such as (3)) can always be converted into functions taking property arguments (such as (4)) by what is known in the Categorial Grammar literature as the ‘‘Geach rule’’ (Geach 1972). This operation maps any function f of type to a function h of type where h ¼ lX[lC[f(X(C))]] (for X of type and C of type c). The interested reader can verify that (4) is just the result of applying this operation to (3). The relevance of the Geach operation here is twofold. First, we will make use of it in the variable-free program. Second, note that this kind of an operation is often useful in the program of direct compositionality: if one has the intuition that some expression is an S modifier yet English syntax treats it as combining on the surface with a VP, one can simply assume that its lexical meaning is the ‘‘Geached’’ version of the intuitive S modifier. Before leaving this, I should point out that under a certain set of assumptions about semantic composition which will be detailed below, the meaning of will would actually have to be (3) (the propositional operator) and not (4) (the property operator). These assumptions, however, are exactly what I will be challenging here, and so I will continue to exposit the rest of the semantic composition using both versions of will0 . To conclude the background remarks, we turn to the semantic composition of an ordinary relative clause, as in (5): (5)
John read every book which Bill will read.
What I am particularly interested in here is the conventional wisdom about the meaning of the expression which follows will (in the surface syntax) which is that the meaning of this expression is the property read0 (xi) and not just the 2-place relation read0 . (Of course the usual thinking on this is that the material following will in the syntax is not just an ordinary transitive verb read but is rather the VP read ti) Under the view that will0 takes VP-type meanings as arguments, this VP meaning is the argument of will. Under the ‘‘LF Raising’’ view of will the VP meaning read0 (xi) first combines with the subject to give the proposition read0 (xi)(b), and this then occurs as argument of will0 .
36
Pauline Jacobson
But from whence comes this conventional wisdom about the semantic composition? As far as I can tell, the motivation behind this is the (often unstated) assumption that – with a few exceptions to be noted below – functions must be ‘‘saturated’’ before they can combine with anything else – i.e. they must get their arguments in order for semantic composition to proceed. A related assumption found in much of the syntactic literature is that if some linguistic expression ‘‘wants’’ something – either in the syntax or in the semantics – it must get it. The assumption that a function must be saturated entails two things: functions cannot combine with other functions by function composition (we return to this below), nor can a function take another function as argument (since the argument itself would not be saturated). (The analogous thinking in syntax gives rise to, e.g., the Projection Principle, and to the view that there is a principle ensuring that all ‘‘theta roles’’ in the argument structure of some verbs do indeed get assigned to something.) Note this view leads to the conclusion that the meaning of the phonological string [rid] in (5) has to, at some point in the semantic composition, be read0 (xi) rather than just the ordinary 2-place relation read0 . The latter is of course the lexical meaning of read – but it will have to get a variable as argument in object position in order for the semantic composition to proceed. As to the rest of the composition of the relative clause, the precise details depend to some extent on how one treats which – but it is assumed to work roughly as follows. The meaning of the material following which is an open proposition, i.e., a proposition with an unbound variable within it. The open variable is then l-abstracted over (possibly as a result of the meaning of which, possibly because of a unary shift rule, or possibly (as in Heim and Kratzer 1998) because the syntactic representation contains an indexed ‘‘silent operator’’ whose meaning is the operation of l-abstraction over the variable indicated by the index). Thus the meaning of Bill will read is – at some point in the semantic composition – will0 (read0 (xi))(b) (under the assumption that will0 takes property argument) or will0 (read0 (xi)(b)) (under the assumption that it takes propositions). This is then mapped into the property lx[will0 (read0 (x))(b)] or, lx[will0 (read0 (x)(b))]. And finally, this property is intersected with the meaning of book.5 Of course an astute reader will notice that if (which) Bill will read and book are both of type <e,t> (or, <e,>) then we already have here an exception to the purported rule that functions must be saturated – here, neither gets their e-argument. Modification is, in fact, one of the cases which is usually assumed to be an exception to this (see Heim and Kratzer 1998 for explicit discussion), where modification itself can be seen as a separate semantic operation (whose set analogue is intersection) which allows two functions to directly combine.
Direct compositionality and variable-free semantics
3.2.2
37
The implications for ACD
We are now ready to review the familiar implications of these assumptions for the analysis of ACD as in (1) (John will read every book which Bill will). Although we have already seen that given the view that functions must get their arguments requires an LF Raising view of will, let me nonetheless for expository convenience proceed under the assumption that will0 takes properties as arguments. (The outcome of the discussion will be the same either way, as the reader can verify.) Then in order for the semantic composition to proceed, will0 is going to have to find a VP-type of meaning as its complement. And in order for the entire relative clause to get put together properly, that ‘‘missing’’ meaning is going to have to contain a variable – that is, it is going to have to be read0 (xi) (under the relevant reading). Hence there must be some linguistic expression whose meaning, at some level, is the VP-type meaning read0 (xi). But the problem is that on the surface there is no VP which can supply the missing meaning: the only candidate here would be the VP read every book which Bill will. But this can’t be the VP which supplies the missing meaning because this would lead to the infinite regress – i.e. to the ‘‘antecedent containment’’ paradox. To the rescue comes LF, or some other abstract level at which the quantified NPs are not ‘‘in the sentence’’ but instead are occupied by material (such as a pronoun or trace) whose meaning is a variable. Notice that this can be implemented in a variety of ways. One, for example, is Montague’s Quantifying-In solution (which performs a substitution operation in the syntax but which is weakly direct compositional and makes no use of LF); a second strategy posits an LF which is interpreted and then subsequently mapped into surface structure by rules including Quantifier Lowering (McCawley 1970; Lakoff 1971); and a third is the mirror image descendant of this whereby surface structure is mapped to LF by QR (May 1977). I will refer to any of these as a Binders Out approach – what they have in common is the view that the matrix VP (at some level) has as its meaning read0 (xi). And so this can be the antecedent for the ellipsis site – this supplies the missing complement of will0 . Again, if will0 is actually taken to be of type and hence raised at LF, the same basic point will hold. 3.2.3
The fallacy: doing without LF
The fallacy here is evident: there is no motivation for the initial assumption that functions (except when they combine by modification) must be ‘‘saturated’’ (and the concomitant assumptions that a function therefore cannot be an argument, nor can it combine with another function by an
38
Pauline Jacobson
operation such as function composition). Not only is there no motivation for this assumption, but it actually would be nothing more than an extra stipulation in the theory of grammar. One might argue that the theory is simpler if there is only one mode of combining two meanings in the semantics (functional application), but even so – nothing would prevent the case of one function taking another as argument. Moreover – even though the assumption that all functions must be ‘‘saturated’’ (or, the syntactic analogue which is that all items must get what they are subcategorized for) has driven much of the research in syntax and semantics for the last forty years – no well-articulated theory of semantics really does maintain this assumption. First, as noted above, modification is already generally seen as an exception to this. Second, just about all theories of semantics maintain that expressions have world and time arguments – and these are not always filled as semantic composition proceeds. Third, consider the role of ‘‘lambda abstraction’’ in the semantic composition. Thus take a sentence like Every man loves his mother. Under the standard view, there is some stage in the semantic composition where either just the VP or the whole sentence (with the subject ‘‘pulled out’’) is a function from individuals to propositions – i.e. (ignoring world and time arguments) it is the function lx[x loves x’s mother]. This then occurs as an argument of every man of type ,t> . In that case, the meaning of the expression which has ‘‘undergone l-abstraction’’ is itself a function which does not get its argument; rather it occurs as argument of (in this case) a generalized quantifier. So there would seem to be no actual or even desirable theory from which the assumptions which led to the above story are a happy set of assumptions. And so, let us drop them. Then a variety of new possibilities opens up. First, as detailed in Steedman (1987), Dowty (1988) and many others, the meaning of a relative clause can be put together directly without the use of any trace or variable in object position by allowing function composition. Using this method, an ‘‘incomplete’’ expression like read0 – which hasn’t found its object – can directly combine by function composition with the auxiliary; the result can then function compose with the subject (assuming that the subject can have a generalized quantifier type of meaning), and so on ‘‘up the tree.’’ To illustrate, let us pick the version where will0 is of type ,<e,>> Thus to derive the meaning of Bill will read in (5), the first step is as follows: (6)
will read; will0 o read0 ¼ lx[will0 (read0 (x))]
Before continuing, a word about the meaning of the subject. We will assume that Bill has its generalized quantifier meaning lP[P(b)]. Following Partee and Rooth (1983), I will assume that the lexical meaning of Bill is of type e, but that any expression of type e can lift to a generalized quantifier. Notice
Direct compositionality and variable-free semantics
39
that there is considerable independent motivation for thinking that proper names can denote generalized quantifiers – they can, for example, conjoin with run-of-the-mill quantified NPs. Given this, the semantic composition of Bill will read proceeds as follows: (7)
Bill0 will-read0 ¼ lP[P(b)] lx[will0 (read0 (x))] ¼ lx[will0 (read0 (x))(b)] ¼ lx[li[9i0 [read0 (x)(b)(i0 ) & i0 > i]
What happens from here on in depends in part on what meaning one assigns for which but – as in the standard account – eventually (7) combines with the meaning of book by intersection. Of course we have dealt here only with the semantics: the actual proposals of Dowty, Steedman, and others making use of function composition often embed this within a Categorial Grammar syntax, and this automatically allows an expression like will – which syntactically subcategorizes for a VP complement – to directly combine in the syntax with a transitive verb. (See also the work in GPSG and HPSG for similar proposals.) There is another closely related analysis. Suppose it is true that when two expressions combine, their meanings always do so by functional application. It still does not follow that the material following will in (5) must have a VP-type meaning as opposed to a transitive-verb-type meaning. This is because one can disallow function composition but instead make use of a unary rule: the Geach rule. Notice that if one applies the Geach rule to some function f and then applies the result to a function h, this is equivalent to directly function composing f with h. In other words, the Geach rule is simply a ‘‘Curryed’’ (i.e., unary) version of the function composition operation.6 This fact will become significant later, and I will actually adopt this analysis of relative clauses rather than the direct function composition analysis. (Until section 3.5, however, I will generally continue to illustrate the point with the direct function composition analysis, as it is undoubtedly more familiar to many readers.) Of course the Geach rule itself would be precluded by the stronger assumption noted above that all functions must be saturated, since its output is an expression which wants a function as argument. There is one further point about this analysis worth stressing before continuing. Note that it is a variable-free treatment of extraction. What I mean by this is the following. Under the traditional view, the meaning of read ti is read0 (xi) – this contains an unbound variable. To make sense of what this actually means in model-theoretic terms, we need to invoke the notion of assignment functions: let every linguistic expression actually have as its semantic value a function from assignment functions to something else. An expression with no unbound variables within it will be a constant function from assignment functions; one with an unbound
40
Pauline Jacobson
variable within it will, in general, be a non-constant function. (It can happen to be a constant function – either due to accidental facts about a particular world or because it is a tautology, but it need not be a constant function.) In the Categorial Grammar (CG) analysis of (5), there is no expression whose meaning at any point in the semantic composition is a non-constant function. In the case of (5), then, the assignment functions are playing no role, and can be dispensed with. Of course they would still be needed if there were other types of expressions whose value was a nonconstant function from assignment functions – which is the case in any theory which treats the semantic value of pronouns as variables. Note, though, that this treatment of extraction provides a first step towards a fully variable-free view: there is no need for a ‘‘trace’’ and/or variable in the position of the extraction site which later becomes l-abstracted over. We will push this approach to its logical conclusion in section 3.4, and arrive at a fully variable-free semantics. 3.2.4
The consequences for the analysis of ACD
Once we realize that the meaning of a relative clause like (5) can be put together without positing a variable in object position, ACD immediately lends itself to a strong direct compositional account. This point was made originally in Cormack (1984) and developed also in Evans (1988) and Jacobson (1992a; 1992b). Consider again (1) (John will read every book which Bill will). The idea in a nutshell is that the ‘‘missing meaning’’ here is simply the two-place relation read0 – and this is supplied by the meaning of the overt expression read within the matrix VP. This meaning combines with the meaning of the auxiliary just as it does in the case where read is overt. For the moment, let us continue to use the direct function composition analysis of extraction. Then the idea here is that the ‘‘missing’’ meaning read0 is picked up (from the meaning of the matrix transitive verb), and function composes with the meaning of will; the rest of the semantic composition is exactly as in (7).7 This way of putting it is a bit mysterious, for one might ask: what does it mean to ‘‘pick up’’ a meaning during the semantic composition of which Bill will when that meaning is supplied by the semantic composition of something higher up (the matrix)? Is this not quite non-compositional? Indeed it is, and so I will revisit this in section 3.5, where the analysis will be refined/recast in such a way as to make it fully (direct) compositional. But for now, I believe that the informal way of looking at it will do. Under the view here, there is nothing special about ‘‘ACD’’ as some sort of category – it is simply a particular instance of the more general phenomenon which I will refer to as ‘‘TVP Ellipsis.’’ Like with VP Ellipsis, we
Direct compositionality and variable-free semantics
41
would thus expect it to, for example, occur across sentences. And Evans (1988) pointed out that indeed it does: (8)
Bagels, I like. Donuts, I don’t.
Of course cases like this are exactly the kind of case that was predicted to be impossible by Sag’s (1976) condition on ‘‘alphabetic variants,’’ but I think it is by now fairly well established that Sag’s condition was simply incorrect.
3.3
Some additional cases of ACD: a quick survey of some past results
The above provides a direct compositional sketch of a simple case of ACD, but of course there are any number of complex facts regarding this construction which have been discussed in the literature, and the ‘‘TVP Ellipsis’’ analysis is a serious alternative only if its coverage goes beyond a simple case like (1). It is obviously beyond the scope of this paper to provide analyses of all of the ACD facts which have been discussed in the literature, but it is worth briefly commenting on a few of these cases to show that they pose no threat to the analysis. I begin with the argument in Diesing (1992b) to the effect that ACD provides additional evidence for the Binders Out approach on the basis of the claim that those items in object position which do not exhibit wide scope effects (and so presumably cannot scope out) also do not allow for ACD. The paradigm case would be bare plurals denoting kinds – which always have narrow scope (see Carlson 1977) and thus can be argued to resist an analysis in which they are treated like quantified NPs. Hence, Diesing notes the ungrammaticality of cases like (9): (9)
*John reads books that Bill does.
But this argument is easily answered, for Diesing has failed to provide a minimal pair. In fact, (9) is bad with or without ellipsis (probably for pragmatic reasons): (10)
*John reads books that Bill reads.
So to test the claim, we need a case where we can have a relative clause on a bare plural: (11)
John reads books that no one else would read.
And, lo and behold, ACD is impeccable here: (12)
John reads books that no one else would.8
42
Pauline Jacobson
A second problem which is often raised with this analysis is the observation that the ‘‘missing’’ meaning can be more than just a single transitive verb, as in (13) (I use the strikethrough notation not to mean that there is some actual silent material here, but simply to notate the intended understanding): (13)
John said that Mary knew everyone that Bill also did say Mary knew.
But as pointed out in Cormack (1984), this is easily accommodated in Categorial Grammar since this theory provides a simple means for said that Mary knew to ‘‘compose up’’ to be a constituent with a meaning of type <e,<e,t>>. (For discussion, see, e.g. Cormack 1984; Steedman 1987; Dowty 1988; Jacobson 1992a; 1992b). In other words, the claim here is not that this is transitive verb ellipsis, but transitive verb phrase ellipsis. Along these lines, one of the most compelling arguments for the Binders Out approach centers on the facts and analysis in Sag (1976) and related facts in Williams (1977) to the effect that (14) allows only a de re reading on the object: (14)
John said that his son should read every book which his teacher did say his son should read.
This is discussed in detail in Jacobson (1992a) (see also Cormack 1984); let me informally sketch the basic idea here, although I will actually depart from the details of Jacobson (1992a) in that I will treat object quantifiers in a different way. Thus, in order to provide any kind of analysis for (14), we need to settle first on a way to allow for quantified NPs in object position. Two possibilities are sketched in note 2; let us here adopt the view that phrases like every book can either be of type ,t> (when they occur in subject position) or of type ,<e,t>> (when they occur in object position). Thus more generally, they are argument reducers.9 If we now look at the full non-elided version of (14), the reader can verify that the de dicto reading of every book which his teacher said his son should read will arise if this combines with the simple embedded transitive verb read. The de re reading, on the other hand, will arise if we first form (by, for example, function composition) the complex transitive verb said that his son should read and then take this 2-place relation as the argument of the object. But then, in the ellipsis version we will get the same result as in the standard theory: since we are assuming here that the ‘‘supplied’’ meaning must be the meaning of some other expression, we can supply the 2-place relation say-his-son-should-read0 only if that has been composed up in the syntax. And that will yield (only) the de re reading. (It is interesting to note that this is quite parallel to the story under the Binders Out approach.)
Direct compositionality and variable-free semantics
43
Turning to another apparent challenge: Larson and May (1990) (in response to Baltin, 1987) discuss cases where the gap site is internal: (15)
John put every book that Mary also did put on the table on the table.
These are also discussed in detail in Jacobson (1992a), where I advocate a ‘‘Wrap analysis’’ of direct object NPs (as developed earlier in Bach 1979; 1980; Dowty 1982; Pollard 1984; and Jacobson 1987 among others). According to this, put on the table first composes to give a complex transitive verb, into which the object NP is subsequently ‘‘wrapped’’ (i.e. infixed). This means that the transitive verb phrase put on the table is available to supply the ‘‘missing’’ 2-place relation. Additionally, Larson and May note the following contrasts: (16)
John believes everyone that Bill also does believe to be a fool to be a fool.
(17)
*John believes that everyone that Bill also does believe is a fool is a fool.
The grammaticality of (16) is unsurprising assuming that believe to be a fool is also a complex TVP into which Wrap applies. The grammaticality of (17) is also unsurprising given certain assumptions about where and how Wrap can apply; see Jacobson (1992a) for full details. In addition, that paper points out some new contrasts which have received little discussion in the literature. Thus I show there that under one way to formalize Wrap, the following contrasts are predicted: (18)
*John said that Mary put every book that Bill also did say that Mary put on the shelf on the shelf.
(19)
John said that Mary put on the shelf every book that Bill also did say that Mary put on the shelf.
This is not to say that I have a ready analysis of everything known about ACD. A striking fact of immediate relevance for which I have no explanation is the ‘‘extraposition generalization’’ noted in Tiedeman (1995) and discussed in Wilder (1995) and Fox (2002). Thus, the following is also good: (20)
John said that Mary put every book on the shelf that Bill also did say that Mary put on the shelf.
This looks suspiciously similar to the fact in (19), enough so that one might suspect that there should be a unified explanation for both of these.
44
Pauline Jacobson
(See Wilder (1995) and Fox (2002) for relevant discussion.) I do not have such an explanation to offer here, and I thus have to leave this for future research. But I hope in this section to have convinced the reader that this view accounts for enough of the preliminary data as to be a reasonable competitor to the standard Binders Out story. 3.4
Pronominal binding, ACD and variable-free semantics
3.4.1
An apparent problem for the TVP Ellipsis analysis
But we are not yet home free, since there is another well-known case of ACD, noted again originally in Bouton (1970) – I will refer to this type henceforth as Case 2 ACD: (21)
John voted for every girl who hoped he would.
(For expository ease, I will treat vote for as a single transitive verb.) What appears to be ‘‘missing’’ here is a full VP with a pronoun/variable in object position. Note that the full paraphrase requires a full VP, as in (22a), and not just a transitive VP (TVP) as in (22b): (22)
a. John voted for every girl who thought he would vote for her. b. *John voted for every girl who thought he would vote for.
Hence it would seem that the missing meaning is vote-for0 (xi) where xi is ultimately bound by the subject of think. But if so, we have come full circle – and we are led back to the original argument for a Binders Out level at which the object NP of the main clause is just a pronoun or a trace which is interpreted as a variable. For once again there is no surface VP whose meaning is vote-for0 (xi). And so, once again it looks like we must invoke LF or some other abstract level to rescue us – under the Binders Out view we can happily assume that the meaning of the matrix VP at some level is vote-for0 (xi) and so this can supply the missing meaning in the ellipsis site.
3.4.2
The fallacy: variable-free semantics
Again there is a fallacy in the reasoning above. This time, the fallacy lies in the belief that the only way to get the effect of ‘‘binding’’ is to assume a theory with variables, which would entail that the meaning picked up here is votefor0 (xi). It turns out, though, that the entire problem disappears once we adopt a variable-free semantics as advocated in, e.g., Jacobson 1992b; 1999; 2000a. In fact, as noted above, we can already see that the Steedman – Dowty
Direct compositionality and variable-free semantics
45
CG approach to relative clauses developed in the late 1980s already provides the first step towards the elimination of assignment functions as an essential part of the semantic apparatus – and so eliminating them entirely thus becomes a very logical next step. The full program and independent motivation for it are detailed in the papers above, but for readers unfamilair with this program, I will briefly review the basic ideas here. 3.4.2.1 Two accounts of binding 3.4.2.1.1 The standard account In order to elucidate the variable-free approach, it is worth first reviewing the machinery which is needed in the standard approach to pronominal binding (that is, an approach making use of assignment functions and variables). Thus take: (23)
Every mani loves hisi mother.
(24)
Every mani thinks that hei lost.
Under the standard story, the semantic value of any expression which contains a pronoun unbound within that expression is a (generally nonconstant) function from assignment functions to something else (where each assignment function is a function from the set of variable names to ‘‘normal’’ model-theoretic objects such as individuals10). Thus take his mother in (23). We usually represent the meaning of this as something like xi’s mother, but what do we mean by such a representation? Its value can only be determined with respect to an assignment function – it is thus a function from assignment functions to individuals. Thus xi’s mother will assign to each assignment function g the individual who is the mother of g(xi). In fact, under the standard account all expressions have as their value functions from assignment functions to something else. Thus in something like Every man loves Bill’s mother, Bill’s mother also has as its value a function from assignment functions to individuals, but since it contains no pronouns unbound it is a constant function from assignment functions to individuals. Similarly, the value of that hei lost in (24) is a function from assignment functions to propositions. How does ‘‘binding’’ take place? There are actually two major ways to do this which have been explored in the literature. One – which is implementable within the program of direct compositionality without any abstract level – is by means of the ‘‘Derived VP Rule’’ of Partee (1975) (and used later in, e.g., Sag 1976; and Williams 1977; and revived in Buring ¨ 2005). Basically what happens here is this. thinks hei lost is a non-constant function from assignment functions to VP type meanings. Similarly for loves his mother. The Derived VP rule is a unary rule as follows. Take a VP with some phonology and whose value is a function from assignment
46
Pauline Jacobson
functions to functions of type <e,t>; call that value [[VP]]g for any g. Then the shift rule will leave the phonology and the syntactic category intact (and also the semantic type), and will return a new function F from assignment functions to functions of type <e,t> such that for any assignment function g and any individual a, F(g)(a) is ½½VPg½xi !a (a) (where ½½VPg½xi !a means the value of the input VP on the assignment function just like g except that the variable xi is assigned to a). Thus the rule ‘‘binds’’ a variable which was unbound in the input (note that ‘‘binding’’ really means shifting a value which is non-constant on some set of assignment functions into something which is constant on the set of assignment functions which differ only in the value that they assign to the relevant variable). Incidentally, I have phrased this as a unary shift rule but, as should be evident by now, we can instead introduce a silent operator into the syntactic structure to do the work; see, e.g., Buring (2005). ¨ The more usual view is rooted in the Binders Out approach. This also makes use of a unary rule (or silent operator) but the binding is done at the sentential rather than the VP level. Thus in the case of (23) we have a representation like ti/hei loves hisi mother whose meaning – relative to some assignment function g – is love0 (the-mother-of 0 (g(xi)))(g(xi)). Hence, for any variable xi, the requisite shift allows any expression whose value is a function a from assignment functions g to truth values to shift to a new function b from assignment functions to <e,t>-functions, such that for any assignment function g and any individual a, bðgÞðaÞ ¼ ½½ag½xi !a . I will refer to this ‘‘binding rule’’ as l-abstraction, since this is indeed the semantics of the operation that we refer to in this way. Thus in the case at hand here, we shift the meaning of ti/hei loves hisi mother into a constant function which, for all assignment functions, denotes the set of self’s mother lovers, and this is taken as argument of the generalized quantifier. An important point to note here is that either view requires a unary shift rule (or, silent operator). 3.4.2.1.2 Variable-free view Let us now turn to a variable-free semantics of the sort developed in, e.g., Jacobson (1999) and a number of other papers of mine. (There are other versions of a variable-free semantics, but space precludes considering here how/if these will handle the cases at hand; some relevant discussion, however, is provided in Jacobson (2003).) The basic idea is that an expression containing a pronoun unbound within it is not a function from assignment functions to something else, but instead a function from individuals to something else. (An expression with two open pronouns within it is a function from two individuals to something else.) Thus the meaning of his mother in (23) is a function from individuals to individuals. In particular, it is that function mapping each
Direct compositionality and variable-free semantics
47
individual into its mother – call this the-mother-of0 . In (24), the meaning of he lost is a function from individuals to propositions. More specifically, it is the function mapping each individual into the proposition that that individual lost, hence he lost has the same meaning as lost. (This is an oversimplification; I assume that pronominal gender has a semantic effect, and hence he lost is actually a partial function defined only for the domain of male individuals. But we will for the most part ignore gender here.) As to a pronoun itself: this too is a function from individuals to individuals – in particular, the identity function on individuals. (Again, he is presumably actually the identity function on male individuals.) This means that there are no assignment functions as part of the semantic machinery. Moreover, there is no need for any indexing conventions in the syntax. Note that in the standard theory there are actually an infinite number of pronouns pronounced ‘‘he’’: he1 he2, etc. Here on the other hand there is just one, whose meaning is the identity function. At first glance, this approach would seem to wreak havoc with the semantic combinatorics. After all, the appeal of the standard approach is that the combinatorics are uniform: an open expression (one with a pronoun unbound within it) combines with other things in exactly the same way as does a closed expression. All expressions have as their meanings functions from assignment functions to something else, so the meanings of John lost and he lost are composed in the same way; viz: [[S]]g ¼ [[VP]]g ([[NP]]g). But in the variable-free approach, the composition of John-lost0 involves combining a function of type <e,t> (lost0 ) with an expression of type e (John0 ). But in he lost, the subject is of type <e,e>. How, then, do the combinatorics work here? And how can we account for the generalization that any environment which can support an ordinary NP like John also allows for a pronoun like he – even though these are not of the same type? There is an immediately obvious answer: let lost combine with he by function composition. As it turns out, though, this is a slightly oversimplified answer: there are several advantages to actually breaking function composition down into two steps (see Jacobson 1999). In other words, rather than directly using function composition, we will make use of the ‘‘Geach rule’’ discussed in section 3.2.1. Thus assume that there is a very general type-shift rule allowing any expression whose meaning is a function of type to type-shift into one whose meaning is of type . Lost is born with a meaning of type <e,t> , but it can therefore shift into a function of type ,<e,t>> and hence combine with a pronominal subject (or, with any subject that contains a pronoun which is unbound within the subject). The result will, moreover, be not a proposition, but a function from individuals to propositions.
48
Pauline Jacobson
In order to really see the system at work, it will help to introduce some syntactic machinery at this point. I will embed the proposal into a Categorial Grammar syntax. Hence assume that each linguistic expression is a triple: . ‘‘Phrase structure’’ rules are binary combinatory rules taking two such triples and yielding a third (where the phonological parts of each input are concatenated in the output). I assume that there are also binary combinatory rules beyond just concatenation – in particular ‘‘Wrap’’ rules (Bach 1979; 1980) which infix one string into another, although these will play little role here. And again, I will make use of unary rules which take a single triple as input and yield a new one as output. Assume further a small set of basic categories (the set{S, NP, N, CP, PP} will suffice here), while other categories are of the form A/LB, and A/RB. The former means something which wants to combine with a B to its left to give an A; the latter wants to combine with a B to its right to give an A. (Additional subscripts will be needed for Wrap operations.) Let us assume further that the syntactic category of some expression records the fact that that expression contains within it an unbound pronoun. We will notate this with a superscript feature. Thus for any two categories A and C, there is a category AC, and any expression of category AC has as its meaning a function of type (i.e., it is a function from the kind of thing that C expressions denote to the kind of thing that A expressions denote). Finally, we will have a unary rule shifting both the syntactic category and the semantic type of an expression: thus this rule will allow an A/B to shift to an AC/BC. In other words, the resulting category is one which says: give me a B with an unbound C within it, and I’ll pass up the information that I contain an unbound C. The associated semantic shift is the ‘‘Geach rule.’’ The full rule is given in (25); we will call this rule g: (25)
Let a be an expression of the form: . Then there is an expression b of the form: ): (26)
lost; S/LNP; lost0 —> g lost; SNP/LNPNP; lf <e,e> [lxe [lost0 (f(x))] he; NPNP; ly[y] he lost SNP; lf [lx [lost0 (f(x))](ly[y]) ¼ lx[lost0 (ly[y](x))] ¼ lx [lost0 (x)] ¼ lost0
We will not show the full composition of the meaning of his mother, but note that any reasonable way to get John’s mother to mean ix[mother-of 0 (j)(x)] will predict that his mother means ly[ix[mother-of0 (y)(x)]] which is the function that we will notate as the-mother-of 0 . Continuing on with the analysis of (23), the next question to address is how to effect the ‘‘binding’’ of his by the subject? Recall that the standard account uses some kind of unary rule – or some analogue – to do this. One possibility is the Binders Out approach – in which case the shift applies on the domain of the whole sentence (with just an indexed pronoun or trace in subject position). The second possibility is the Derived VP Rule – in which case the shift happens at the VP domain. Thus all accounts make use of some sort of shift rule. The account here will too – it is simply a different shift rule. The most interesting difference is that the rule is much more local: in (23), for example, it will apply to the meaning of love (rather than love hisi mother or ti love hisi mother). And, of course, the other difference is that this shift has nothing to do with assignment functions or indices, since neither of these are part of the machinery here. In fact, the reason that the variable-free account can apply the shift so locally while the standard account cannot is exactly because in the standard acccount, the shift has to refer to the index on the material being ‘‘bound.’’ And it is precisely this locality which gives the variable-free approach a number of its advantages – see Jacobson 1999; 2003 for discussion. I call the requisite shift rule z; this operation takes a function a of type > and maps it into a new function of type ,<e,b>> such that z(a) ¼ lf[lx[a(f(x))(x)]]. In prose, we start with a function which wants an argument of type a and then a later individual argument, and we map this into a function that wants as its first argument a function from
50
Pauline Jacobson
individuals into a-type meanings. Moreover, it ‘‘merges’’ its own higher individual argument slot with the individual argument slot of its first argument. (This generalizes straightforwardly to the case of 3-place verbs, see Jacobson 1999.) The full rule, complete with syntax, is spelled out in (27): (27)
Let a be an expression of the form . Then there is an expression b of the form: [lx[a0 (f(x))(x)]]>.
The intuition of the z rule is easy to grasp by example. Take an ordinary relation between two individuals (i.e., a function of type <e,<e,t>> such as love0 . This can shift by z and the result is now a function between individuals and functions of type <e,e> – i.e., z(love0 ) is of type ,<e,t>>, where to z(love0 ) a function f is to be an x who ordinary-loves f(x). In the case of (23), then, z(love0 ) combines with his mother whose meaning is the-mother-of0 and the resulting VP has the meaning lx[love0 (the-motherof 0 (x))(x)] – it denotes the set of self’s mother lovers. This then combines with the subject generalized quantifier every man in the normal way. (Syntactically, I assume that subject generalized quantifiers are of category S/R(S/LNP). I also assume that ordinary NPs – such as John – can freely lift both syntactically and semantically; thus an NP can be shifted so that its meaning involves the usual lift from e to ,t> and its syntactic category involves the corresponding lift from NP to S/R(S/LNP).) One would, of course, ultimately want to be convinced that the account generalizes to the case of the binding of pronouns in adjuncts, the case of multiple pronouns ‘‘bound’’ by a single ‘‘binder’’; the case of multiple pronouns with different ‘‘binders,’’ etc. With some small (and simple) generalizations of the above apparatus it indeed does; I refer the reader to Jacobson (1999) for details. This view of pronominal binding removes the need for assignment functions (and indices in the syntax). But – is there a high cost to this? I would argue not. As discussed above, any theory needs a shift rule to effect binding – the rule here is simply a different one from the standard one. So the only really extra apparatus here is the use of the g rule. But this is both simple and very natural – and, moreover, it has considerable additional benefits. For example, Jacobson (2000a) documents that the existence of this rule means that it becomes automatic that pronouns have additional ‘‘paycheck’’ readings (most theories simply have ordinary individual pronouns and paycheck pronouns as separate things). This, I believe, is quite a striking payoff of the general program here. Moreover, in section 3.6 I will show that the existence of the g rule allows for a simple account of Pied Piping semantics. Because the variable-free approach plays a large role in the direct compositional analysis of ACD, it is important to stress that its motivation does not come just (or even primarily) from ACD. Indeed I have argued in
Direct compositionality and variable-free semantics
51
a series of papers that it simplifes the analysis of a wide range of phenomena. I obviously do not have space to detail this here and so will point out just a few (in the style of a late-night television ad.): The existence of functional readings for questions (somewhat of a surprise under the standard view) comes for free, as does the existence of a variety of other functional phenomena. As noted above, paycheck pronouns come for free. Jacobson (2000a) shows that i-within-i effects – along with some puzzling exceptions to them – are also an automatic consequence of this approach combined with a Categorial Grammar syntax, and Jacobson (1999) discusses the distribution of across-the-board binding readings in extraction constructions. Moreover, Jacobson (2000b) shows that not only can contrastively stressed bound pronouns be accounted for in this framework, but in fact some subtle interactions between contrastively stressed pronouns with paycheck pronouns are also automatically accounted for.11 3.4.2.2 Free pronouns A quick comment is in order about free pronouns, as in: (28)
He lost.
These are no different than ‘‘bound’’ pronouns, and so the analysis here predicts that – at the end of the day – (28) does not denote a proposition, but rather a function from individuals to propositions (and its syntactic category is not S, but SNP). Presumably, then, listeners compute propositional information from something like (28) by applying this function to some contextually salient individual. It may strike the reader as somewhat peculiar to treat (28) as not being a proposition. But notice that (28) is not actually a proposition in the standard view either – for it is a proposition only relative to some assignment function. (Thus, it is really a function from assignment functions to propositions.) How then, do we extract propositional information from this? The context must supply a salient assignment function to which (28) can be applied. Actually, all sentences are functions from assignment functions to propositions, but in the case of closed expressions these are constant functions, so it makes no difference which assignment function is picked. But in the case of an open sentence, we need a mechanism for supplying an assignment function. Surely this is no simpler than having a mechanism whereby an open sentence is applied to a contextually salient individual. 3.4.3
Implications for ACD Case 2
Armed with these tools, we return to the apparent problem for the direct compositional treatment of Case 2 ACD:
52
Pauline Jacobson
(29)
John voted for every girl who thought (that) he would vote for her.
Recall that the difficulty for the TVP Ellipsis analysis was that what appears to be missing here is a full VP, whose meaning is vote-for0 (xi) where xi eventually gets bound by the subject (or, subject position) of thought. But in order to find a constituent whose meaning is vote-for0 (xi) we needed to assume a Binders Out level in order to supply the variable in object position. The answer to this under the variable-free view should be obvious. Here vote for her means exactly the same thing (modulo the gender contribution) as does vote for – and so once again this is simply an instance of TVP Ellipsis. All that needs to be supplied, then, is again the 2-place relation vote-for0 , and the object ‘‘slot’’ of this will be bound ultimately by application of z on thought. The fact that a pronoun is required in the full paraphrase in (22) is a matter of syntax not semantics – vote for her and vote for mean essentially the same thing (and the object slot is bound in both in the same way) but have different syntactic categories. That the syntax correctly requires the full paraphrase to be a VP and not a TVP will be documented in section 3.5.4. Before continuing, there is one potential confound which should be cleared up. Notice that (29) also contains a pronoun as subject of the lowest auxiliary, and Ha ¨ık (1987) pointed out that this seems to be required in that (29) contrasts sharply with (30) without a pronoun: (30)
*John voted for every girl who thought (that) Bill would.
She thus proposes an analysis which requires that these contain a bound pronoun as subject of would – a requirement which is obviously not mirrored in this analysis. But indeed, I would argue that there really is no such requirement; the contrast between (29) and (30) seems, instead, to have to do with considerations of focus. It has been known since at least as early as Rooth (1992a) that ellipsis interacts with focus, and these interactions become especially vivid and interesting in the cases of TVP Ellipsis. Unfortunately, to properly investigate this would take an entirely separate paper and so for the most part I am ignoring these interactions here (see, however, the brief discussion in section 3.7). Nonetheless, even without a full analysis, we can convince ourselves that the difference between (29) and (30) has nothing to do with having a bound pronoun as subject of the ellipsis site but rather has to do with finding the requisite contrast (or, in this case, lack thereof). For notice first that (31) is not particularly bad (at worst, it involves a mild kind of ‘‘Principle C’’ violation – it is somewhat awkward to repeat the full NP in this configuration – but it is no worse than (32) which contains the same kind of violation (one needs in both cases to destress the second occurrence of John):
Direct compositionality and variable-free semantics
53
(31)
?John voted for every girl who thought that John would.
(32)
?John voted for every girl who thought that John would vote for her.
And second, we can greatly improve cases like (30) by setting up the appropriate focus conditions. Consider, for example, (34) in light of the following scenario: (33)
The Dating Game Scenario: The Dating Game is a television show in which two men and two women are contestants each evening. Each woman writes down the name of one man that she would like to date. Meanwhile, each man writes down the name of one woman who he thinks will pick him. If woman A picks man B and B succeeds in guessing that A will pick B, then A and B get to go out on a fabulous date, all expenses paid by the show. Last night, though, no one won, because:
(34)
?Unfortunately, Mary picked the man who thought that Sue would, and Sue picked the man who thought that Mary would.
While perhaps not the world’s happiest sentence, (34) is still considerably improved over (30) (taken out of context); this judgment has been confirmed by several informants. Finally, notice that since we are claiming that the phenomenon illustrated in (21) is a case of TVP Ellipsis, we would once again expect this to occur across sentences (and in other sorts of environments). Indeed it does. Jacobson (1992b) and Hardt (1992a) both notice cases like: (35)
John kissed every girl. But Mary is the only one who wanted him to.
(What the ‘‘missing’’ meaning is here is the two-place relation kiss and the appropriate paraphrase would – for syntactic, not semantic – reasons – be kiss her and not kiss. ‘‘Binding’’ is effected by z on want.) Similarly, Jacobson (1992b) notes cases like (36) (see also Fiengo and May 1994): (36)
John asked Mary to water his plants and Bill asked Sue to.
Cases like (35) and (36) have – in some of the recent literature (see Merchant 2001; Takahashi and Fox 2006) gone under the rubric of ‘‘rebinding’’ – a rubric which might suggest that something special is going on here. On the contrary: I claim that these are just perfectly expected instances of TVP Ellipsis; the surprise would come if these were bad. (Once again these show some interesting interactions with focus, which I unfortunately do not have space to discuss.)
54
Pauline Jacobson
3.5
Refining the account
We have seen the broad outlines of a view under which ACD is perfectly compatible with Strong Direct Compositionality. But the informal sketch above leaves at least three questions unanswered. The first is the question raised in section 3.2.4: what does it mean to say that the meaning of some constituent is ‘‘supplied’’ during the semantic composition of some other expression? In particular, how can we reconcile direct compositionality with a view that the meaning of some lower expression (in this case, the expression containing the ellipsis site) is dependent on the meaning of some higher expression (in this case, the matrix TVP)? I have phrased this in terms of what happens in TVP Ellipsis, but the same questions arise for run-of-the-mill VP Ellipsis. Second, how exactly does the syntax work – what licenses ACD (or TVP Ellipsis) in general? This is a rather pressing question since under the standard view, ACD is just a special case of VP Ellipsis, and so nothing new needs to be said in order to account for its existence (beyond what accounts for VP Ellipsis in general). Does the present account do just as well? Third, we have left it as a promissory note that the syntax will predict that the full paraphrase of (21) is (22a) (with an overt pronoun), rather than (22b) (with a gap in object position) – even though the two mean essentially the same thing. 3.5.1
VP Ellipsis
To address these questions, let us first turn to the account of VP Ellipsis more generally. Rather than survey the numerous proposals in the literature, I will simply begin the discussion with an account roughly along the lines of (although not identical to) that of Merchant (2001). Like the account in Sag (1976) (and other earlier accounts), ‘‘ellipsis’’ here is taken to be a deletion of material in the syntax or – if one prefers – a ‘‘silencing’’ of syntactic material. Unlike the account in Sag, the conditions here are taken to be semantic identity (rather than formal identity at the level of LF) between the elided material and some antecedent. (Strictly speaking, Merchant’s actual account is not semantic identity but ‘‘e-givenness,’’ but this difference will not affect the main point here.) Merchant implements this by positing a silent operator (E) which suppresses the phonology of its sister and which can successfully combine in the semantics only with expressions whose meaning satisfies the identity condition. We can recast this into the formalism here by introducing a unary rule as follows: (37)
Let a be an expression of the form . Then there is an expression b of the form provided that 0 is the meaning of some other expression in the discourse context.
Direct compositionality and variable-free semantics
55
(The syntax would of course need refining to restrict this to those VPs that occur as complements of auxiliaries or of to.) But the problem with this formulation – or any conceivable variant thereof – is that the condition in italics is not something which can be determined by local inspection of the input to the rule: it requires reference to material which can be in an entirely different sentence. We might change the italics to say something like provided that 0 is salient in the discourse context – but this is of no help, for this too is not a local property of some expression. In other words, any sort of identity condition of this type would seem to pose a serious challenge to direct compositionality, or indeed to any kind of compositional analysis of the elided material.12 But of course the inability of some material to be completely ‘‘determined’’ locally is nothing new and is just what anaphora is in general. Put differently, there is a perfectly ordinary precedent for this situation in the behavior of free pronouns. We have already seen how these can be dealt with (in either theory of binding). In the standard theory the ultimate value of a sentence with a free pronoun is assignment-function dependent, and the context must supply an assignment function. In the variable-free theory, a sentence with a free pronoun is a function from individuals to propositions, and the context supplies a salient individual to take as argument of this function. So why not just say (as many have in fact said) that VP ‘‘Ellipsis’’ is really ‘‘VP Anaphora’’? Indeed, this has been proposed by a number of authors; see especially Hardt (1993). In variablefree terms, a sentence with an ‘‘ellipsis’’ site is a function from properties to propositions, and it is applied to some property made salient by the discourse context. (Bowing to tradition, I continue to use the term VP Ellipsis, but VP Anaphora is more felicitous.) Of course since at least as early as Hankamer and Sag (1976), it has been known that there is a major difference between ordinary anaphora and the situation with VP Ellipsis. In the case of ordinary personal pronouns, no linguistic antecedent is necessary: an expression containing a free personal pronoun can ‘‘pick up’’ any contextually salient individual. If VP Ellipsis were just like the case of ordinary free pronouns, then this too should be able to pick up any contextually salient property. But Hankamer and Sag argued that this is false: they claimed that an elided VP requires a linguistically overt antecedent. In fact, though, the claim that a VP Ellipsis site can be pragmatically controlled has been the subject of years’ worth of debate, and there is a considerable body of work showing that – albeit rare and difficult – this is possible. Examples of pragmatic control abound in the literature, see, e.g., Schachter (1977), Webber (1978), Dalrymple et al. (1991), Hardt (1993) among many others. But Hankamer and Sag were not entirely wrong – a ‘‘pragmatic control’’ account does need to explain the fact that pure pragmatic control is very
56
Pauline Jacobson
difficult, and that the property which is supplied is one which very much likes to be the meaning of some linguistically overt expression. My hope is that this can be explained in something like the following way. Suppose that a function of type <e,t> is a kind of ‘‘fragile’’ object – not one that robustly lives in the discourse context. This is quite different from the situation with individuals (including abstract individuals). If this is true, it makes sense to think that a function of type <e,t> is most easily made salient by being the meaning of some overt linguistic expression (which we will continue to refer to as the ‘‘antecedent’’). This, moreover, is compatible with another wellknown fact (see, e.g., Hankamer and Sag 1977), which is that the function likes to be the meaning of some relatively close linguistic expressions. We can assume that such functions do not stay as stable, salient objects, and so can only be easily accessed if they have recently been named. Now of course there is more to say. First, we need to explain why such objects are ‘‘fragile’’ (not to mention supplying a serious account of what is meant by ‘‘fragile’’), but this I suspect is a job for my cognitive psychologist colleagues. Second, we need to explain the flip side of the Hankamer – Sag observations: which is that a free pronoun like it – even when it does appear to be picking up something roughly like a property – is perfectly happy with a non-linguistic antecedent. This fact is exemplified by Hankamer – Sag contrasts like the following: (38)
Context: We are walking down the street and I see someone trying to scale the walls of the Sciences Library. I turn to you and say: a. He’ll never do it. b. ?*He never will.
But it’s not really clear exactly what (or, more to the point – what type of thing) it really is ‘‘picking up’’ in (38a), and it’s certainly not clear that it’s picking up the same kind of function of type <e,t> as is required as complement of will0 . The literature has assumed that it in (38a) and the missing complement of will in (38b) are the same here simply because He never will scale the walls seems to be a reasonable description of what the speaker might have meant in both cases. But there are many other possible ways to ‘‘paraphrase’’ the understanding of (38a) here (He’ll never do that thing, etc.), and so I think it is an entirely open question as to just what sort of model-theoretic object is accessed by it. Let us now turn to the consequences of this for the assumption made up until this section: that there is some kind of ‘‘identity condition’’ between the meaning supplied at the ellipsis site and the meaning of some other expression, and that VP Ellipsis (and ACD) can therefore be used as a diagnostic for the meaning of the ‘‘antecedent.’’ Strictly speaking, these assumptions and this terminology are not correct. But since it is the case
Direct compositionality and variable-free semantics
57
that the understood property is in general the meaning of some overt linguistic expression, I will continue to speak in these terms. That is, we will continue to assume that in at least most cases – and especially in the cases that go through completely smoothly – that the property is supplied via some other overt expression. We can, then, continue with the usual strategy of using ellipsis as a diagnostic for the meaning of the ‘‘antecedent.’’ In other words, the grammar itself does not impose an identity condition, but the effect of such a condition is a by-product of the (general) need for naming a function of type < e,t> in order to make it sufficiently salient. 3.5.2
Formal details
We can now formalize this a bit more. Using VP as an abbreviation for S/LNP, assume that auxiliaries (and to) are listed in the lexicon as being of category VP/RVP – that is, they are VP modifiers. (Presumably each auxiliary is also specified as taking a complement with a given feature; this is needed to get the morphology right.) Thus, for example, the auxiliary will is of this category and has the meaning (roughly) lP[lx[li[9i0 [P(x)(i0 ) & i0 > i]]]]. Assume further that there is a rule mapping all auxiliary items of category VP/VP into members of category VPVP with the same meaning. In other words, such auxiliaries now are not functions which syntactically want a VP complement, but are rather essentially pro-VPs. (Unlike pronouns they have more complex meanings than just the identity function – although this may well be the correct meaning for do – but they have the same syntax.) They thus combine in the same way that pronouns do. (Note, then, that we are not positing a silent proform in the ellipsis site; the auxiliary itself is the ‘‘proform.’’) The information that these are proforms is passed up in the syntax and – just like the case of free pronouns – the final expression can contain an unbound pro-VP and hence be of category SVP with meaning of type ,t>. It will thus be applied to some contextually salient property (again, with the caveat that the relevant types of properties are difficult to access by inference and hence like to be made available by being the meaning of an overt and a close VP). There is, moreover, an additional role played by focus parallelism (see Rooth, 1992a) – a role which can hopefully be explained in terms of its ability to make the relevant property salient. However, I continue to ignore this here. 3.5.3
TVP Ellipsis
An obvious feature of the standard account of ACD is that ACD is simply a special case of VP Ellipsis. Whatever the account of VP Ellipsis, it is
58
Pauline Jacobson
unsurprising in the standard view (combined with Binders Out) that ACD exists, and nothing extra is needed. Before considering whether the TVP Ellipsis analysis also needs nothing extra, let us first reassure ourselves that we do want a unified account of the two. Indeed, it seems that we do – for TVP Ellipsis has essentially the same basic properties as VP Ellipsis. First – like VP Ellipsis – the antecedent can follow, as in (39), and it can be in a different sentence within the discourse, as in (40) ((39c) is best with stress on every):13 (39)
a. VP Ellipsis case: Before John did, Bill climbed the mountain. b. TVP Ellipsis – Case 1: ? Everything which Bill does, I’ll also read. c. TVP Ellipsis – Case 2: Although only Mary wanted him to, John kissed every girl.
(40)
a. VP Ellipsis: I ran. Bill did too. b. TVP Ellipsis – Case 1: Bagels, I like. Donuts, I don’t. (Evans 1988) c. TVP Ellipsis – Case 2: John kissed every girl. But only Mary wanted him to.
We should also reassure ourselves that – like ordinary VP Ellipsis – TVP Ellipsis also allows the missing 2-place relation to be contextually supplied (i.e., this too can be a case of deep anaphora). Actually, this is quite difficult (especially for examples of Case 2), but it does seems marginally possible: (41)
a. Case 1: Context: I see you about to reach for some cookies that just came out of the oven. Pointing first to a batch on a different plate, and then to the hot batch: These, you may. Those you mustn’t, at least not until they cool down. b. Case 2: Context: Bill is at a dog park, and is trying to pet each dog. Most of the dogs are quite unsocialized, and don’t like to be petted – with the exception of Kolya. So, as he tries to pet Mitka, I turn to him and say: ??‘‘Better not – I think Kolya is the only one who would actually like you to.’’
The fact that these are so much more difficult to construct than VP Ellipsis cases remains to be explained, but if the general line taken above is correct then it seems quite reasonable to assume that 2-place relations – i.e. functions of type <e,<e,t>> – are even more ‘‘fragile’’ and hence harder to access from context than are functions of type <e,t> .
Direct compositionality and variable-free semantics
3.5.4
59
Formal details
The account of VP Ellipsis developed in section 3.5.3 generalizes to the case of TVP Ellipsis (and hence ‘‘ACD’’) with very little effort. For full disclosure, I should point out that the mechanism for VP Ellipsis extends to the case of TVP Ellipsis only if we generalize some of the unary rules (we will note the requisite generalizations below). Thus the claim that we get TVP Ellipsis ‘‘for free’’ is true only if we can be sure that these generalizations are correct. At the moment, I have independent motivation for only one of them – but the kind of generalizations needed seem to me to be in keeping with the general program here and we would probably discover that we want these generalizations (either for formal or empirical reasons) independently of this domain. Admittedly this remains to be demonstrated, but we can certainly note that only very simple extensions of the mechanisms serve to capture the TVP Ellipsis cases. To work out the precise syntactic details, we need some more concrete formalism for extraction. Within the Categorial Grammar literature, there have been two main proposals. One (Steedman 1987) is that the italicized material in (42) is just an ordinary transitive verb which has not yet found its object, and is thus of category (S/LNP)/RNP (I abbreviate this as VP/RNP). (42)
John read every book which Bill read.
The other is that extraction ‘‘gaps’’ are not really the same as ordinary argument positions; this is argued for in, e.g., Oehrle (1991). Although I believe that there is actually good evidence for the latter, I will continue in this discussion to adopt the former – both because it is more familiar and because it simplifies the exposition. In the end, nothing crucial hinges on this. I will, however, depart from Steedman (1987) in assuming that extraction constructions are not put together by direct function composition, but rather once again by the two-step process of the Geach rule followed by application. So this means that the syntax of g as discussed earlier is generalized so that anything of category A/B can shift syntactic category to become (A/C)/(B/C) (as well as being able to map to Ac/Bc). The semantics is the same. I assume further that the input possibilities are generalized too: not only does A/B shift but so can AB, where the outc ðB Þ (B/C) or ðAC Þ The last of these generalizations was puts would be (A/C) already motivated on independent grounds in Jacobson (2000a). In the case of extraction, then, an item of category A/B shifts to (A/C)/(B/C), and then takes an expression of category B/C as argument. With these generalizations, consider again the auxiliary will in a classic ACD case: (43)
John will read every book which Bill will.
60
Pauline Jacobson
We have seen that it shifts to a VP proform; that is, it can be a VPVP. In a normal VP Ellipsis case, it combines with a subject – the information that there is a VP proform is passed up by the superscript feature, and a sentence with a VP Ellipsis site is – at the end of the day – of category SVP, and is of type ,t> . As with any free pronoun, the listener applies this to a contextually salient property. In the case of TVP Ellipsis, matters are similar. Since will can be a VPVP, the above conventions allow it to also be a (VP/NP)(VP/NP). In other words, it is a proform over VPs with extraction gaps. The superscript feature – indicating that there is an unbound proform of category VP/NP – will be passed up, and a sentence like (42) will ultimately be of category S(VP/NP). Hence its type is >,t> – that is, it is a function from 2-place relations to propositions. In the case of (43), it can pick up the contextually salient relation read0 . What about Case 2 examples like (44)? (44)
John will vote for every girl who thinks that he will.
Again, we begin with the proform will of category VPVP, but – assuming that g is given in its NP most general form – it follows that will can also be of category ðVPNP ÞðVP Þ – a proform over VPs with NP proforms within them. (This, of course, has exactly the same meaning as the will in (43) – it is only its syntactic distribution which differs.) Once again this feature will be ‘‘passed’’ up so that the final category of (44) is S(VPNP) – which once again has as its meaning a function from 2-place relations to propositions. (Again, z on think ensures that the ‘‘object’’ argument slot of this 2-place relation is merged with (bound to) the subject position of thinks – which gives the correct semantics.) I have glossed over one complication. In the composition in general of nouns modified by relative clauses (such as book which John read) I have been making the standard assumption that (which)-John-read0 ends up characterizing a set of individuals which intersects with the meaning of book. I have, however, not made any commitment as to how this happens: it could be done by a special semantic rule directly allowing intersection in certain cases, or it could be done by a unary rule (or, silent operator) which shifts the meaning either of the relative clause or of the head from P to lQ[P ˙ Q].14 In the case of TVP Ellipsis, however, the meaning of the relative clause is not of type <e,t> , but is a function from two-place relations to set-denoting expressions. What this means is that when combining, e.g., which John will with read, the semantics needs to be set up in such a way as to ‘‘hold off’’ on the missing 2-place relation, and intersect ‘‘inside’’ of this. To clarify – suppose that the system is such that the composition of book which John read involves directly intersecting
Direct compositionality and variable-free semantics
61
book0 with which-John0 read0 (rather than first shifting one of these by the unary rule above). Then, in the case of book which John will, the semantics needs to be set up so as to allow the meanings of the two to be put together as follows: lR[book0 ˙ lx[will0 (R(x))(j)]. This might appear problematic, but again generalizations of this type are needed throughout the system (for discussion of a related case, see Jacobson 1999). We can now make good on an earlier promise: answering the question as to why the full paraphrase of something like (21) contains a transitive verb with an overt pronominal object and not just a simple intransitive verb. In other words, we are concerned with the contrast in (22): (22)
a. *John voted for every girl who thought he would vote for. b. John voted for every girl who thought he would vote for her.
The answer is purely a matter of the syntax. If there is a missing object or extraction gap in the syntax after vote-for then ultimately there must be something higher up which expects to co-occur with material containing a gap. But there is no such expression. We do, of course, find the pronoun who which occurs as subject of thinks, but this has its desires filled by combining with the VP headed by thinks. (Alternatively, one can think of this as ‘‘extracted’’ such that it combines with a sentence that has an extraction gap in subject position; either way it does not want to find an extraction gap lower down.)15 It is worth noting that this is completely analogous to the answer that would be given in the standard theory. After all, there too vote for heri and vote for ti have the same meaning. But vote for ti will not be licensed in a construction in which there is no ‘‘displaced’’ or extracted element to bind the trace.
3.6
ACD and Pied Piping
I have argued throughout this chapter that the conventional wisdom is simply mistaken: ACD is perfectly compatible with (strong) direct compositionality. Given that this is a much simpler view of the organization of the grammar than one positing abstract levels and/or a dissociation between the syntactic composition and the semantic composition, this should be seen as a welcome result. But, in fact, there is more to be learned from the case of ACD: when we consider its interaction with Pied Piping we see an unexpected problem for the ‘‘standard’’ view (at least one which maintains an identity condition – whether it be semantic or formal – between an antecedent VP and the elided VP). But the problem disappears under the variable-free and direct compositional view of Pied Piping combined with the TVP Ellipsis analysis.
62
Pauline Jacobson
3.6.1
Pied Piping in general
Consider first a run-of-the-mill Pied Piping case such as: (45)
every candidate the father of whom Mary voted for
Before I turn to a variable-free account of this, let me give (essentially) the account of this in Sharvit (1998) – which is a direct compositional but nonvariable-free account. (I will then give its variable-free translation and show how it provides some simplifications.) The point of departure for Sharvit’s analysis is the analysis of functional questions in Groenendijk and Stokhof (1983) and Engdahl (1986), whereby the functional reading of a question like (46) involves a gap whose meaning is a complex variable: a variable over functions of type <e,e> applied to an individual variable. Thus the ultimate meaning of (46) is shown (informally) in (47): (46)
Who does every man love the most? (His mother)
(47)
what is the function f such that every man0 ? (lx[x loves f(x)])
Sharvit notes that if ‘‘traces’’ can have complex meanings of this type (i.e. to correspond to variables like fi(xj)), then there is nothing to preclude the possibility of the variable xj simply not happening to be bound by the subject. It should, then, be able to stay free within the meaning of the lower clause. Additionally, there is nothing to preclude the possibility of functional gaps in relative clauses as well as questions; indeed these have been discussed in detail in Jacobson (1994) and Sharvit (1999). So consider again the NP in (45) (every candidate the father of whom Mary voted for). Continuing with the standard terminology, suppose that the gap after vote for is a functional gap, with meaning fi(xj), and suppose that xj is not bound within the lowest clause. Then one can imagine a stage in the semantic composition of Mary voted for in which the meaning of that expression is lf[vote-for0 (f(xj))(m)]. Moreover, in a framework with variables, it is reasonable to suppose that the meaning of the father of whom is the ‘‘open’’ individual the-father-of0 (xj). Now assume that something allows ‘‘open’’ variables to be l-abstracted over (hence, bound) at various points. That is, the input the-father-of0 (xj) (which is an individual but which varies according to assignment functions) can shift into a function of type <e,e> (which is constant across assignment functions), in particular it shifts to the function lx[the-father-of0 (x)] (which is the-father-of0 function). If this is taken as argument of the expression above, then the meaning of the father of whom John voted for ends up being vote-for0 (thefather-of0 (xj))(m). We need one further step (possibly licensed as a consequence of the assumption made just above that variables can be bound by
Direct compositionality and variable-free semantics
63
l-abstraction as needed). This step is to l-abstract over xj; the result of this gives an ordinary set of individuals (the set of individuals x such that Mary voted for the father of x), which is now ready to intersect with the head noun candidate. This is a direct compositional analysis of Pied Piping semantics without any kind of reconstruction, although some of these details are awkward. But in fact, these awkwardnesses disappear once this is recast into the variable-free program. First – let us return to the basic case of functional gaps – as in functional questions such as (46). The existence of functional readings for questions (and many other constructions) is one of the most striking arguments for variable-free semantics (see Jacobson 1999). For their existence is predicted given the mechanism here – they come as no surprise and are just part and parcel of the system for binding in general. Note first that in the case of a functional question like (46), the ‘‘gap’’ (which is nothing more than a missing argument slot) does not have a complex meaning – it is simply a missing argument of type <e,e> . The fact that an argument of this type can be expected is simply a consequence of the fact that loves has undergone z. Moreover, the argument ‘‘slot’’ of this function is merged with (or ‘‘bound to’’) the subject slot of loves – and so we end up with the Groenendijk and Stokhof – Engdahl semantics without further ado. While the existence of functional readings for questions (and relative clauses) has always seemed like somewhat of a surprise under the standard view, the surprise here would be if these readings did not exist. As to a Pied Piping case like (45) we can simply apply g to vote-for rather than z. This is analogous to Sharvit’s observation that the individual ‘‘variable’’ in a functional gap can remain free within the lowest clause. The translation of this into the variable-free setting is that an expected function of type <e,e> is introduced by g rather than z – and so the argument slot of that function is not bound within the lowest clause (it is, rather, ‘‘passed up’’ for higher binding). In fact, though, there is another derivation which gives exactly the same result but which will be more useful for our purposes below. This is to compose up the meaning of (ordinary) vote-for with the subject Mary to give a function of type <e,t> (an ordinary extraction construction), and then apply g at that point. This is spelled out in (48); as an abbreviation I use direct function composition in the first step: (48)
lP[P(m)] love0 ¼ lx[vote-for0 (x)(m)] (type <e,t> )!g lf[ly[lx[vote-for0 (x)(m)](f(y))]] ¼ lf[ly[vote-for0 (f(y))(m)]] (type ,<e,t>>)
Notice that we require no extra apparatus to ‘‘l-abstract’’ over either the function ‘‘variable’’ or its argument ‘‘variable’’ – these argument slots
64
Pauline Jacobson
are always, as it were, bound in virtue of the basic apparatus. Thus nothing new is needed; the meanings are directly put together by the apparatus already introduced for pronouns in general. The same point holds with respect to the meaning of the father of whom. Assuming again that whom is an ordinary pronoun, it follows automatically that this denotes the-fatherof0 function. And – as in Sharvit’s derivation – this can be argument of Mary vote for so that the meaning of the entire relative clause the father of whom Mary voted for is ly[vote-for0 (the-father-of0 (y))(m)]. And then this set intersects with the head noun. 3.6.2
Pied Piping and ACD
The interest in Pied Piping for the discussion here is that we also have grammatical cases like (49), discussed in Jacobson (1998): (49)
John voted for every candidate the father of whom Mary had.
Interestingly, these are perfectly smooth for the analysis here – the only piece of this which might not be obvious is that once again we will have to apply g ‘‘inside’’ the argument slot for the missing two-place relation. (But again this type of generalization is required throughout the system). Thus we can compose the meaning of Mary had so that it is: lR[lx[had0 (R(x))(m)]]; this is just like the composition of the run-of-the-mill ACD case in (1). Now applying g ‘‘inside’’ of the missing R-argument slot, we shift this to mean: lR[lf[lx[had0 (R(f(x)))(m)]]]. Still holding off on the R slot, this takes as argument the meaning of the father of whom (which is the fatherof0 function) to give (50) as meaning of the father of whom Mary had: (50)
lR[lx[had0 (R(the-father-of0 (x))(m)]]
which (again holding off R) intersects with candidate0 . Nothing special needs to be said about this case. The important point to note is that the material is looking for an ordinary 2-place relation (such as vote-for0 ). The fact that the ultimately understood object of this relation happens to be complex and involve a function such as the-father-of0 function is built into the fact that the g operation applies to the expression Mary had (combined with the fact that this takes the father of whom as argument). What needs to be contextually supplied, then, is not a VP meaning but rather the very ‘‘small’’ meaning which is just a 2-place relation. But things are not so smooth for the conventional account – at least not one which requires identity between the elided material and the meaning of (or form of) some other expression. Of course in the account that I am ultimately advocating, the grammar itself does not contain an ‘‘identity condition’’ as a part of its machinery. But I am assuming that such a
Direct compositionality and variable-free semantics
65
condition follows (in general) as a by-product of the fact that the missing meaning likes to be made salient by being the meaning of some actual linguistic expression. (There is of course the ‘‘in general’’ caveat here; I return to this below.) In the TVP Ellipsis analysis above, the expression that supplies the ‘‘missing meaning’’ is just the simple transitive verb vote-for0 . But in the standard view, there is no expression which can do the trick. What sort of meaning needs to be supplied (or, be the meaning of the elided material) under the standard view? This depends on whether or not one adopts the standard reconstruction analysis of Pied Piping or, instead, Sharvit’s analysis. Let us take first the former. Then the meaning of the elided material (or, the meaning which is supplied as complement of had0 must be the meaning vote-for0 (the-father-of0 (xi)). This will combine with the rest to give the open proposition which we can represent informally as Mary had vote for the father of xi, xi is l-abstracted over so that this shifts into a set, and that intersects with the head. But there is no linguistic expression which can possibly supply this meaning. QR to the matrix VP gives only the meaning vote-for0 (xi)(j) which is not remotely the same. Since Sharvit’s analysis is quite similar to the variable-free analysis that we gave above, one might think that adopting it would solve the problem. Indeed, at first glance things look more promising, but this is an illusion. Under this analysis, the missing (or, elided) meaning must be vote-for0 (fi(xi)). But this is not the meaning of the matrix VP either. Of course it has often been suggested that ‘‘identity’’ means identity up to the difference in variable names, so that (under some accounts) vote-for0 (xi) and vote-for0 (xj) would count as identical for the purposes of VP Ellipsis. (Given that these both vary according to the assignment function, they are not actually identical, but many accounts require identity up to alphabetic variance, and these are alphabetic variants under any usual definition of that notion.) But the case here is not analogous to this: votefor0 (fi(xi)) and vote-for0 (xi) are quite different and are not just a matter of using different variable names. One could, of course, try to come up with some looser definition of ‘‘identity’’ according to which these are identical, but I don’t know that there would be any independent motivation for this. Moreover, if it is correct that the identity ‘‘condition’’ is really just a by-product of the need to make the antecedent function salient by naming it, then we really do need to have the meaning of the antecedent be the relevant function; we are simply not free to tinker with the definition of ‘‘identity.’’ Let me return to the caveat noted above. It is the case that what is understood at the ellipsis site is not always identical to the meaning of some other linguistic expression. After all, this fact is exactly what has led many authors to conclude that pragmatic control is indeed possible.
66
Pauline Jacobson
So one might be very tempted to put this case into the pragmatic control basket thereby circumventing the problem that the standard analysis has no way to supply the missing meaning in virtue of the meaning of some other expression. While it is difficult to definitively dismiss this possibility, it seems unlikely. For those cases where we do see pragmatic control, we can generally tell a fairly simple story about how the function which is supplied comes to be contextually salient. In this case, though, there is no obvious source for the salience of vote-for0 (the-father-of0 (xi) or votefor0 (f(xi). (As discussed in note 3, Fox (1999a) notes that this kind of case can be handled under the standard view if one abandons any kind of condition on identity between the elided material and some other expression in favor of only Rooth’s focus condition which applies on a bigger domain (see Rooth 1992a). I will not discuss Fox’s proposal here, except to say again that if ‘‘identity’’ is indeed a by-product of the need to make a function of type <e,t> salient, then this could not be dropped.) Why does the analysis here have no difficulty with this case while the standard view does? It is because in this case we do not need to find an entire VP-type meaning, but need only the simple 2-place relation votefor0 . It is also worth pointing out that if we needed to find the function g(vote-for0 ) we would have the analogous problem here. But we don’t need this; the fact that the argument position is a function of type <e,e> is a consequence of the fact that g can apply to the bigger domain which is the meaning of Mary had.16
3.7
Final remarks and looking ahead
Kennedy (1994, this volume chapter 5) points out that ACD is impossible in cases like (51b) and (51c) (I have modified Kennedy’s particular examples to avoid irrelevant complications centering on the complements of relational nouns): (51)
a. Mary invited every boy that BILL did invite. b. *Mary invited every boy who lives next door to a girl that BILL did invite. c. *Mary invited every boy who lives next door to a girl who’s a penpal of the girl that BILL did invite.
These facts – which have come to be known as Kennedy’s puzzle – are of key concern to the analysis here, and, in fact, to the variable-free program in general. For Heim (1997) – in a well-known analysis of these contrasts – argues that they can be accounted for by Rooth’s focus condition (Rooth, 1992a) combined with the Binders Out (standard) view of
Direct compositionality and variable-free semantics
67
ACD and combined a theory that makes crucial use of variable names. Thus Rooth proposed that (in addition to an identity condition), ellipsis is licensed only if the elided material is contained within some constituent CELL which has focused material in it, and that the antecedent VP is contained within a constituent CANT such that the meaning of CANT is a member of the focus value (i.e. is an alternative to) CELL. Thus, Heim observes that in (51a), Rooth’s focus condition will have to be satisfied by finding a member of the focus value of BILL did invite (ti), which is the set of propositions {invite0 (xi)(m); invite0 (xi)(s); . . . }. And, the meaning of the matrix here (with the object scoped out) is invite0 (xi)(m) – which is in fact a member of the above set. But in the case of (51b) and (51c), the focus condition cannot be satisfied – essentially because the object variable names are not the same. For (51b), for example, we will be looking for a member of the focus value of BILL did invite (tj), which is the set of propositions {invite0 (xj)(m); invite0 (xj)(s); . . .}. But there is no expression whose meaning is a member of this set of alternatives; the trace in the matrix is not coindexed with the lowest trace, and so the meaning of the matrix here (with the object scoped out) is again invite0 (xi)(m). Hence Rooth’s focus condition cannot be satisfied. Thus this analysis makes crucial use of two pieces of machinery which we are rejecting here: the use of variable names, and the Binders Out approach to ACD. But, while I suspect that Heim is correct in pinning the problem on focus, there are enough problems with her particular analysis as to make one suspect that the precise tools that she uses are not correct. First, the analysis requires the ‘‘No Meaningless Coindexation’’ condition – a condition which prevents two occurrences of the same variable name from being bound by different binders. (This is crucial in order to keep the object variable in the matrix in, e.g., (51c), from accidentally being the same as the variable name for the lowest trace.) But this condition is just a stipulation, and is not really expected under the standard apparatus of variables. Second, the analysis requires the ‘‘formulas’’ hypothesis – whereby quantification is quantification over assignment functions, and relative clauses are never of type <e,t> . Rather, they are always open sentences (sets of assignment functions). The same must be said for the head nouns. Thus book in every book which Bill read is an open proposition of the form book0 (xi). This then is also a set of assignment functions, which intersects with the head. (So, modification here involves intersection of sets of assignment functions rather than intersection of sets of individuals. The reader is referred to Heim’s paper for discussion of why the formulas hypothesis is necessary.) But the formulas hypothesis in turn leads to the consequence that ellipsis should be impossible in (52); a patently incorrect prediction:
68
Pauline Jacobson
(52)
Every girl that the teacher praised envied every girl/BOY that the PRINCIPAL did praise.
The details of the problem are spelled out in Jacobson (1998) and Jacobson (2004), and so I will simply refer the reader to those papers. These problems do considerably lessen the temptation to accept the precise machinery behind Heim’s implementation. But, as noted above, the basic idea – that the problem lies in focus – seems quite likely. A proponent of the variable-free (and direct compositional) program would, then, hope that the basic insight can be recast in these terms and that – hopefully – this recasting will avoid the problems with Heim’s implementation. In Jacobson (2004) I explored exactly this route, although the analysis there is quite preliminary. To fully flesh out this case (along with an account of focus and ellipsis in general), one needs to develop an account of focus in the variable-free setting. Moreover, Rooth’s focus condition would need to be recast: the way it is stated above is obviously not compatible with direct compositionality. I believe, though, that any theory would want to recast this in some way, for it should follow from something else and not just be a stipulation. Hopefully, it will ultimately have something to do with what it takes to make the antecedent function salient. At this point, though, I have to leave this for future research, hoping to have at least convinced the reader that the particular implementation of Heim’s analysis which makes use of variable names and Binders Out has enough open questions/problems as to make it reasonable to search for an alternative set of tools.
4
The view of QR from ellipsis Kyle Johnson University of Massachusetts
Ellipsis is one of the operations commonly used to determine syntactic constituents. The set of strings that VP Ellipsis picks out in English, for example, correlates very well with the set of strings that other constituentsensitive processes, such as coordination and movement, pick out. From correlations such as these, we arrive at the conclusion that the italicized strings in (1) – antecedents to VP Ellipsis – constitute constituents. (1)
a. Carrie eats kale on Tuesdays because Will does D. b. Carrie eats kale on Tuesdays because Will does D on Wednesdays. c. Carrie likes kale tremendously because Will does D. d. Carrie served Mr. White skittles before Will did D. e. Carrie considers Max beautiful because Will does D.
And by the same measure, the italicized strings in (2) do not form constituents. (2)
a. *Carrie ate every kumquat after Will did D bean. b. *Carrie eats kale on Tuesdays because Will does D Wednesdays. c. *Carrie eats kale before Tuesday because Will does D after. d. *Carrie ate each vegetable slowly because Will had D legume.
That strings partition into these sets – sets that hold constant across construction types – forms the basis for modern notions of constituency. But the correlation is not perfect. There are contexts in which ellipsis singles out strings that, unlike those in (1), do not correlate with those selected by coordination and movement. One such famous case is Levin’s (1986) Pseudogapping. Examples of Pseudogapping and its mismatch with movement and coordination, are illustrated in (3) and (4). (3)
a. Carrie will eat kale on Tuesday, but she won’t D asparagus. b. *It’s eat on Tuesday, that Carrie will kale. compare: [It’s eat kale, that Carrie will on Tuesday.] 69
70
Kyle Johnson
c. *Carrie will eat kale on Tuesday and store on Wednesday. compare: Carrie will eat kale on Tuesday and store asparagus. (4)
a. Carrie might speak to Max loudly, but she won’t D to Skitty. b. *It’s spoken loudly that Carrie has to Max. compare: It’s spoken to Max, that Carrie has loudly. c. *Carrie has spoken to Max loudly and petted carefully. compare: Carrie has spoken to Max loudly and petted Skitty.
Let’s call the set of strings that ellipsis can affect, but that topicalization, coordination, and other processes cannot, ‘‘ellipsis specific constituents.’’ There is a way of explaining the existence and properties of ellipsis specific constituents that builds on Jayaseelan (1990) and Lasnik (1999a; 1999b), with antecedents in Sag (1980). Suppose that the VP in (3b), to take the first example, has both of the parses in (5). (5)
a. VP VP V
PP
DP
on Tuesday
eat kale b. VP DP
VP* VP V
PP
kale
on Tuesday
eat Let the parse in (5a) correspond to the linear order of the terms involved as they are pronounced, and (5b) correspond to another ordering of these same terms created by moving the object. If, in this way, (5b) can be projected from the VP of (3a), then we have a representation that provides the needed constituency for the ellipsis, namely the VP with ‘‘*’’ appended to it. On this approach, Pseudogapping is just a special instance of VP Ellipsis, and I will consequently use either label in what follows.
The view of QR from ellipsis
71
While the details of this account are worked out in different ways by different authors, the general direction has support from the fact that certain constraints on movement seem to determine which strings are ellipsis specific constituents. There is a match between the strings that movement can traverse and the strings that constitute ellipsis specific constituents. The contrast in grammaticality that ellipsis produces in (6), for example, matches the contrast in movement in (7). (6)
a. Will might try to buy kale, but he won’t D asparagus. b. *Will might decide when to buy kale, but he won’t D asparagus.
(7)
a. What might Will try to buy? b. *What might Will decide when to buy?
Similarly, just as extraction out of the ‘‘complex NP’’ in (8) is blocked, so is eliding a string that leaves a remnant inside a complex NP, as in (9). (8)
*What did Carrie read a book that described?
(9)
*Carrie read a book that described kale, but she didn’t D asparagus.
And just as something cannot extract from an adjunct clause, as in (10), so also may ellipsis not strand a remnant within an adjunct clause, as in (11). (10)
*What did Carrie read a book after she ate?
(11)
*Carrie might read a book after eating kale, but she won’t D asparagus.
Ellipsis specific constituents are, however, more constrained than movement seems to be. They cannot, it seems, contain a portion of a finite clause, while eliding only a portion of a non-finite clause is possible. This is illustrated by the contrast between (12a) and (12b). (12)
a. *Will might decide that Carrie should eat kale, but he won’t D asparagus. b. Will might want to eat kale, but he won’t D asparagus.
While there is no overt movement operation in English that is constrained in this way, the covert process of Quantifier Raising (QR), which assigns quantificational DPs their scope, obeys a similar restriction. Thus, the embedded object in (13b), but not in (13a), can be understood as having the root subject in its scope. (13)
a. A different instructor might decide that Carrie should read every book. b. A different instructor might want to read every book.
72
Kyle Johnson
QR is capable of bringing the object out of the nonfinite clause in (13b) to a position higher than the root subject, a position from which its wider scope can be fixed. But there is, apparently, a constraint preventing QR from moving the object out of the finite clause in (13a) to produce a parallel result. This constraint matches what is seen in ellipsis specific constituents. These matches make sense on a movement-based account of ellipsis specific constituents. If ellipsis specific constituents are formed by moving the remnant out of a standard constituent, constraints on movement will (partly) determine what these constituents can be. This account allows the ellipsis specific constituents in (3) and (4) to elide, but does not spread perniciously to allow the strings in (2) to elide. In these examples, there is no known movement operation that would create these strings. To form the constituents that have elided in (2a) and (2d), for instance, it would be necessary to move a preposition out of the PP it heads and out of the VP this PP is embedded within. These are possibilities not attested in English. The case in (2c) is somewhat less transparent as it would require the, apparently commonplace, process of moving a DP out of a PP. Here we must speculate that the movement operation by which these constituents are formed is, like Scrambling in the Germanic languages, incapable of stranding prepositions. Indeed, like QR, Scrambling in the Germanic languages is blocked out of finite clauses, and so fits the constraint on Pseudogapping that (12) illustrates. For the purposes of this chapter, we may equate Scrambling and QR. This is the first reason for believing the account (5) illustrates: it captures the match that the exceptional constituents in (3) and (4) have with constraints on movement. The second reason is a correspondence between constraints on scope-taking and constraints on what may elide in these contexts. This can be made sense of if the movement operation responsible for forming ellipsis specific constituents is the very operation which fixes the scope of arguments. This is in line with the locality constraints we have seen evidence for above: they are identical to those that are thought to limit QR. An illustration of the correlation with scope is found by examining the pairs in (14) and (15). (14)
a. Dumbledore handed a wand to every student. b. Dumbledore handed a student every wand.
(15)
a. While Dumbledore might hand it to every student, he surely wouldn’t D to every muggle. b. ?*While Dumbledore might hand him every wand, he surely wouldn’t D every broom.
The view of QR from ellipsis
73
As is well known, the second object of a double object construction cannot include the first object in its scope. As a consequence, (14b) unambiguously places every wand within the scope of a student. In this respect, the double object construction is different from other two-complement constructions. It’s possible, for instance, to understand a wand to fall within the scope of to every student in (14a). This suggests that QR is prevented from putting the second object of a double object construction higher than the first object, but that it is not similarly constrained in other twocomplement constructions. This matches what we see in the ellipsis examples in (15). The second object manages to escape the elided VP in (15a). But in (15b), where a double object construction is involved, stranding the second object with ellipsis is considerably worse. Putting these observations together, and bracketing various complications, we can form the hypothesis in (16). (16)
Ellipsis can be fed by QR (aka Scrambling).
For the purposes of this chapter, we can think of QR as being a movement operation that adjoins DPs and PPs to some phrase outside VP. The thesis, then, is that ellipsis specific constituents are just those formed by QR. What remains is to explain why QR does not feed coordination and VP topicalization in the same way that it feeds ellipsis. Why, in other words, does ellipsis have this privileged view of QR? One approach would explain the difference between ellipsis and topicalization/coordination in terms of interactions between these phenomena and QR. For instance, we might imagine that though QR can in principle feed topicalization and coordination, there are properties of these processes that happen to prevent them from applying to the output of QR. This is easiest to see in the case of coordination in (3c). If QR forms a representation like that in (5b), it would not put the constituents in the right linear order to form the basis for the coordination in (3c). To prevent QR from feeding the instance of VP topicalization in (3b), we might exploit the fact that the variable left by moving a QRd term would have to be moved along with the topicalized VP to a position where it is not c-commanded by its binder. This seems to be prohibited in English. This is the type of account of ellipsis specific constituents that I will advocate. But, as we shall see, it immediately presents problems that must be overcome. These problems, I will try to show, are related to another hallmark of this general account: that QR can be a surface process in English. This has not been the traditional view, and for good reason. Allowing QR to occur overtly in English raises a cluster of problems. I will introduce these problems by way of a detour through Gapping, a process which also has access to ellipsis specific constituents. But Gapping
74
Kyle Johnson
and Pseudogapping are not, at their heart, the same phenomenon. I will argue for that position here on the basis of a small difference in the class of strings that Gapping and Pseudogapping can affect. This, in turn, permits a view of the problems we are faced with that affords a solution. I’ll begin by establishing that Gapping targets ellipsis specific constituents. 4.1
Similarities between VP Ellipsis and Gapping
Gapping is similar to ellipsis in that it too is capable of eliding strings which include a verb. Some examples, matching those in (1), are (17). (17)
a. Some eat kale on Tuesdays and others D on Wednesday. b. Some like kale tremendously and others D only partially. c. Some like kale and others D asparagus.
Also like VP Ellipsis, Gapping is capable of eliding what, by the lights of coordination and movement, are not constituents. (18)
a. Some ate kale on Tuesday and others D asparagus. b. Some spoke to Max loudly and others D to Skitty.
And interestingly, Gapping and VP Ellipsis are similarly constrained with respect to the strings they cannot affect. Thus, like the ungrammatical (2) and (6)–(11), Gapping products like (19) and (20) are also ungrammatical. (19)
a. *Some ate every kumquat and others D bean. b. *Some ate kale on Tuesdays and others D Wednesdays. c. *Some ate kale before Tuesday and others D after. d. *Some ate each vegetable slowly and others D legume.
(20)
a. Some tried to buy kale and others D asparagus. b. *Some decided when to buy kale and others D asparagus. c. *Some read a book that described kale and others D asparagus. d. *Some read a book after eating kale and others D asparagus. e. *Some decided that Carrie should eat kale and others D asparagus.
This similarity suggests that we should let Gapping and VP Ellipsis both be capable of affecting the constituents formed from QR. Just as in the context of VP Ellipsis, the limits on QR – that it is subject to the island constraints, and that it may not strand prepositions, determiners or
The view of QR from ellipsis
75
move prepositions – will account for the pattern of strings that Gapping may elide. There is one superficial difference between Gapping and VP Ellipsis that we should set aside before examining more important contrasts. This is that Gapping may elide all of the verbs involved, and VP Ellipsis may not. Thus there are examples of Gapping, like (21), which have no VP Ellipsis counterpart, as in (22). (21)
Some have eaten kumquats and others D apples.
(22)
*Carrie has eaten kumquats because Will D apples.
Although it may not be completely understood what this difference traces back to, there is no reason to think it bears on the issue of which constituents ellipsis may affect. On the currently standard conception of the constraints on ellipsis, it is subject to a requirement that forces the ellipsis site to be governed by a lexical head of a certain sort. This requirement is supposed to have the effect of forcing the VPs that elide to be governed by an overt verb or I0. Gapping is, for some reason, not subject to this constraint, and is therefore capable of producing (21). If QR is capable of producing the strings that ellipsis can affect, then (21) teaches us that QR is capable of moving the object out of a constituent that contains all the verbs. This is an unexceptional hypothesis, as QR is commonly taken to be able to adjoin arguments to clausal-level constituents.
4.2
Differences between VP Ellipsis and Gapping
There are many differences between VP Ellipsis and Gapping, and these have pushed some into the conclusion that they are, at heart, not the same phenomenon. From a typological standpoint, Gapping is considerably more common than VP Ellipsis. Whether VP Ellipsis is even found outside of English is somewhat controversial, since in no other language has the construction been found in a form identical to the English version.1 So far as I know, there is nothing in the rest of Germanic like VP Ellipsis, for example. Gapping, on the other hand, is extremely widespread, found in all the Germanic languages, and widely reported in the rest of Indo-European too. From a distributional standpoint, then, Gapping and VP Ellipsis seem to be distinct processes. Moreover, there are descriptive properties which appear to distinguish Gapping from VP Ellipsis. Gapping, for instance, is typically found just in coordinations. Unlike VP Ellipsis, it does not easily sit within adjunct clauses (as the contrast in (23) illustrates).
76
Kyle Johnson
(23)
a. Will went to Holland because Carrie had D to Spain. b. *Some went to Holland because others D to Spain.
And though VP Ellipsis can elide a VP that precedes its antecedent, this isn’t possible for Gapping, as the contrast in (24) indicates. (24)
a. George might D but many Americans don’t endorse torture. b. *Some D Rumsfeld and others blame Cheney.
And finally, although VP Ellipsis can occur in embedded contexts, the same is not possible for Gapping. In (25a), for example, an elided VP is found within a relative clause, but the parallel instance of Gapping in (25b) is impossible. (25)
a. Carrie hasn’t eaten kale, but I met a man who has D asparagus. b. *Carrie hasn’t eaten kale, but I met a man who D asparagus.
These sorts of differences have led many of the original researchers on ellipsis to classify Gapping as a different phenomenon than VP Ellipsis. The modern response to catalogues of constructions is to launch a search for underlying primitive operations whose combinations yield the differences. Impressed with the contrasts listed above, and others, I have suggested elsewhere that Gapping uses a fundamentally different mechanism than is employed by VP Ellipsis and Pseudogapping. The current best guess is that VP Ellipsis and Pseudogapping arise when strings of deaccented words go unpronounced. Unpronounced phrases are then held to a recoverability condition that, in many recent treatments, is a strengthening of the condition that holds of deaccented material. One might think of these sorts of ellipses as being extreme cases of deaccenting, then, related to their antecedent material in the same kind of way that deaccented material is.2 I suggested that Gapping is, by contrast, a kind of movement operation. I argue that the phrase which Gaps has actually moved into the position that its antecedent is in. This, I argued, is what is responsible for the differences between Gapping and other forms of ellipsis. More particularly, because the movement operation is necessarily leftwards, it will always generate the Gap to the right of its antecedent. Because the movement is an across-theboard type, it will be restricted to coordinations. And, finally, the locality conditions on movement are responsible for the constraint against embedding a Gap. My project was therefore to derive the extra constraints on Gapping from constraints known to hold of movement, by caching out the mechanism that produces Gaps as a kind of movement operation. But recently, some have argued that this proposal exaggerates the differences between Gapping and VP Ellipsis.3 The fact that the cross-linguistic
The view of QR from ellipsis
77
distribution of VP Ellipsis and Gapping do not coincide, for instance, is far too weak to license conclusions about whether they are, at heart, the same operation. Certainly there is nothing about the proposal that Gapping is a kind of movement operation that would explain this difference. And the claim that Gapping is restricted to coordinations is empirically clouded in contexts of comparative ellipsis, like (26), in which it has been argued that Gapping plays a role in creating the ellipsis. (26)
Carrie ate more rutabagas than Will D kumquats.
In his chapter in this volume, and in Lechner (1998; 2000), Winnie Lechner makes a forceful case that it is Gapping which is responsible for removing the verb ate in the than-clause of (26). If this is correct, then reducing Gapping to across-the-board movement will not give an account of the distribution of Gapping, since across-the-board movement is not known to be able to involve than-clauses. Further, the fact that (24) makes it appear that ‘‘backwards’’ VP Ellipsis but not ‘‘backwards’’ Gapping is possible could be due to an effect attendant to the existence of an object remnant in Pseudogapping and Gapping. When a pseudogap precedes its antecedent, as it does in (27a), the result is ungrammatical in the same sort of way that a gap preceding its antecedent is. (27)
a. * Johnny might D kumquats, but Jerry will never eat rutabagas. b. * Some D kumquats and others will eat rutabagas.
Whatever is responsible for the ungrammaticality of (27a) could be blamed for the badness of (27b), and there is no need to resort to my movement hypothesis. This leaves only the embedding constraint on Gapping as evidence on behalf of the thesis that Gapping is movement. Despite these muddy waters, I will argue here that the movement account of Gapping should not yet be jettisoned, as it makes sense of another difference between Gapping and VP Ellipsis. Gapping can elide the string that VP Ellipsis could not in (15b) (repeated here). As (28) indicates, Gapping can delete a verb and the first object of a double object construction, and leave the second object behind. (15b) (28)
*While Dumbledore might hand him every wand, he surely wouldn’t D every broom. Some gave you a wand and others D a broomstick.
For what I will argue are similar reasons, VP Ellipsis cannot elide a verb þ small clause combination, and leave the predicate of that small clause behind. Gapping can. Compare (29a) to (29b).
78
Kyle Johnson
(29)
a. * Carrie finds kumquats edible, but she doesn’t D delicious. b. Some find kumquats (merely) edible, and others delicious.
The explanation for the ungrammaticality of (28a) sketched above is that QR cannot bring the second object of a double object construction higher than the first to form the necessary constituent. This kind of account can be extended to (29a). To form this ellipsis, it would have to be possible for QR to move the predicate of a small clause beyond the small clause’s subject. There is no evidence I know of which suggests that this is possible, and so it fits the account to assume that it isn’t. That Gapping can affect these strings follows from the across-the-board movement account just sketched. 4.3
Deriving the differences
To see this, we need first to dive into the mechanics of the across-the-board movement account. As we’ve seen above, one of its consequences should be to preserve the fact that, in general, Gapping has access to ellipsis specific constituents. Thus, for example, we want QR to be able to produce a representation like (30), from which the Gap in (31) can be produced. (30)
IP DP
I
others I
VP DP1 kumquats
VP V
t1
like (31)
Some like rutabagas and others D kumquats
If Gapping a phrase means moving that phrase, then the VP in (30) must move. In general, assume that VPs move leftwards in English, and let QR feed this movement. The hypothesis, then, is that English sentences can have a surface representation like that in (32) on the following page. I have arbitrarily called the phrase to which moved VPs front: FP. The precise identity, and role, of FP will not matter here. That surface representations can involve leftward movement of this sort has been examined in Kayne (1994, 1998), Zwart (1997), Koopman and Szabolcsi (2000) among others,
The view of QR from ellipsis
79
where it is argued that a variety of scope and word-order effects can be derived from it. Thus, to the extent that these works are successful, there is independent support for VP-to-FP movement. (32)
IP DP
I
others I
FP VP
FP VP
likes t1 F
DP1 kumquats This hypothesis is almost sufficient to derive Gapping as an epiphenomenon. Imagine what will happen if a phrase lower than FP is coordinated. When the VP fronts, it will move across-the-board, giving rise to a representation like (33). (33)
IP I
DP
FP
Sam I VP2 V
FP IP
wants PRO
I I to
VP
F VP VP V
DP
eat
t1
or
VP
DP1
VP
DP1 VP
beans
t2
rice
t2
In this case, the VP wants to eat beans has been disjoined with wants to eat rice, and QR has lifted the objects of each embedded infinitival into the
80
Kyle Johnson
higher clause. VP fronting, now acting across-the-board, produces (33), forming from a disjunction of VPs what appears to be a disjunction of objects. Indeed, Sam wants to eat beans or rice can have an interpretation equivalent to Sam wants to eat beans or wants to eat rice,4 and Schwarz (1999) makes a good case that this interpretation arises as a function of Gapping. The hypothesis that there is leftward VP movement, fed by QR, then, can generate these instances of Gapping. To generate the more traditional variety – ones, like (17), in which the second coordinate holds a subject – we must adopt the derived subjects hypothesis. On many versions of this hypothesis, the subject has an underlying position in the Specifier of a phrase specifically charged with determining the external y-role, often represented as P. From this position, the subject is, under normal circumstances, forced to surface in a higher position, one I’ve represented as Specifier of IP in the graphs above. If we assume that P is embedded within FP, then the hypothesis that English houses VP-to-FP fronting will allow for representations like (34). (34)
IP I I
FP VP1 V like
FP
νP
DP F
νP
t2 DP
some ν
νP
and
ν
ν
DP VP
others
DP2 VP rice
t1
ν
VP DP2
VP
beans
t1
And if we assume that the force which is responsible for compelling subjects to surface in the Specifier of IP is partially independent from Case assignment, then it is conceivable that the subject of the first P could satisfy this force by moving into the Specifier of IP, leaving the subject of the second P still able to receive Case. The result would be (35) on the next page, which has the surface form of a garden-variety Gapping structure.
The view of QR from ellipsis
81
To produce these instances of Gapping, then, the hypothesis that there is VP fronting will have to be wedded to a theory that allows a subject to move out of the first, and only the first, of a series of coordinates in apparent violation of the Coordinate Structure Constraint. Lin (2002) provides an account of the Coordinate Structure Constraint that permits this, and she provides further evidence that this derivation is what lies behind Gapping sentences. (35)
IP DP3
I FP
some I VP1
F
νP
V DP F
νP
like t2 DP
t3 ν
νP
and
ν
DP VP
DP2 VP rice t1
others ν
ν VP DP2 VP beans t1
Not only do these two conjectures derive Gapping, they also derive some of its properties as mentioned in the previous section. Its apparent restriction to coordination, for instance, is a simple consequence of the contexts in which across-the-board movement is tolerated. Similarly, the fact that Gaps cannot appear in the coordinate that precedes its antecedent is derived from the fact that VPs move leftward only. And finally, again as noted above, the fact that Gaps are not found in embedded contexts derives from locality constraints on VP movement itself. It’s clear that countenancing VP fronting of the sort described here must, like all such instances of movement, be constrained in appropriate ways. In the case of VP movement, it must be limited so as not to allow a VP to move out of an embedded clause.5 This constraint, whatever its ultimate source, will spread to predict that Gapping cannot occur in embedded contexts. Finally, this account of Gapping explains why it is able to delete the strings that we saw at the end of section 4.2 that VP Ellipsis can’t. Because
82
Kyle Johnson
this treatment of Gapping allows anything that normally moves to Gap, it applies to a wider class of strings than does ellipsis. In fact, the strings that can Gap, under this account, need not form a constituent at all. All that is required is that the strings can be partitioned into phrases, each of which move. Ellipsis, by contrast, can affect strings only if they can be fit into one constituent. Let’s consider in detail how these cases work. Because QR is prevented from lifting the second object higher than the first in a double object construction, it cannot form a structure like that in (36). (36)
XP XP
DP1 every broom
X
VP hand him t1
This is what is required to form the constituent that has elided in (15b), repeated below, and so this example is blocked. (15b)
*While Dumbledore might hand him every wand, he surely wouldn’t D every broom.
Interestingly, what prevents (36) is only that the second object has the first in its scope. There is nothing that prevents the second object from moving, just so long as it doesn’t end up higher than the first object. This can be seen by observing that the second object may QR high enough to have the subject within its scope, as in (37).6 (37)
A different guide showed me every sight.
It can also be deduced from the observation that Pseudogapping can strand the second object if it also strands the first, as in (38). (38)
? While Dumbledore might give the witch a broom, he surely wouldn’t D the dementor a wand.
Both of the objects in (38) have moved out of the elided VP, with the first necessarily ending up in a position to the left of (and higher than) the second. It’s this mobility of the second object that allows Gapping to affect the string that ellipsis can’t in (15b). The derivation makes use of two acrossthe-board movements. One brings the first object out of coordinated VPs, as shown in (39).
The view of QR from ellipsis
(39)
83
IP
DP1
I FP
some I
νP
F
νP
DP2
νP
you
νP
and
ν
DP
ν
t1
ν
DP VP
others
ν
gave t2 a wand
VP gave t2 a broomstick
The other brings the remnant VPs, from which the objects have moved, into FP. This yields the representation in (40). (40)
IP
DP1
I
some I
FP FP
VP4
νP
gave t2 t3 F
νP
DP2
νP
you
ν
DP t1
ν
νP
and
ν
DP VP
DP3
VP
a wand t4
others
ν
VP DP3
VP
a broomstick t4
84
Kyle Johnson
Nothing more than the movement operations already adopted have been used here, and the condition that the second object remain lower than the first has been honored. A similar set of circumstances accounts for the second of our examples. From the ungrammaticality of (29a), repeated below, we can conclude that the predicates of small clauses may not move higher than their subjects, as in (41). (29a)
*Carrie finds kumquats edible, but she doesn’t D delicious.
(41)
XP AP1 delicious X
XP VP find kumquats t1
The conditions on movement apparently prevent the constituency shown in (41) from forming. Parallel to what we’ve just seen with double objects, this condition does not prevent predicates from moving, it just prevents them from moving to a position higher than their subjects. This, at any rate, is what the relative grammaticality of (42) indicates, under the assumption that Pseudogapping is VP Ellipsis.7 (42)
?? Carrie might find kumquats edible, but she won’t D oranges attractive.
Both oranges and attractive have moved out of the elided VP in (42), but the predicate remains lower than the subject. As with double objects, then, the relevant constraint is one that judges unfavorably an output that changes the underlying order of subject and predicate, but does not otherwise prevent their movement. The derivation that allows (42) is what is responsible for Gapping’s ability (see (43)) to affect the string that ellipsis can’t in (29a). (43)
Some find kumquats (merely) edible, and others D delicious.
As with the double object example, there are two across-the-board movements. One brings the subject of the small clause out of the conjoined Ps, as in (44) on the following page. And the second brings the VPs into FP after the predicates of the small clauses have moved out of them. The result is (45).
The view of QR from ellipsis
(44)
85
IP I
DP1 some I
FP
νP
F
νP
DP2
νP
kumquats
νP
and
ν
DP
ν
DP
t1 ν
others ν
VP
find t2 delicious
find t2 edible (45)
VP
IP
DP1
I FP
some I VP4
FP
νP
find t2 t3 F
νP
DP2 kumquats
νP
DP t1 ν
νP
and
ν
DP VP
AP3 VP edible t4
others ν
ν VP AP3
VP
delicious t4
86
Kyle Johnson
The small clause’s predicate remains lower than its subject in (45), conforming to the constraint that is violated in (29a)/(41). Therefore, the across-the-board movement account of Gapping not only makes sense of the several differences Gapping has with ellipsis, it also captures why Gapping shares with ellipsis the ability to affect ellipsis specific constituents and why it goes a bit further. For these reasons, let’s adopt the view that Gapping is a special instance of across-the-board movement. 4.4
Why does QR only sometimes feed overt representations?
We are now ready to address the question that begins the chapter: why are ellipsis specific constituents not able to topicalize or coordinate? The intervening sections have added to this that Gapping, a consequence of short VP movement, is also capable of affecting ellipsis specific constituents. On the solution to this problem sketched at the outset, the inability of topicalization and coordination to access these constituents is a consequence of the mortal interaction these processes have with QR. For example, in the case of the illicit coordination in (3c), repeated below, there is no way of combining QR and short VP movement (now embracing this consequence of the Gapping account) so that the correct word order arises. (3c)
*Carrie will eat kale on Tuesday and store on Wednesday.
As desired, once the VP has moved to FP in each of the conjuncts, there is no means of coordinating eat on Tuesday with store on Wednesday so that they share the object kale.8 But the suggested means of blocking VP topicalization from acting on the result of QR is now in jeopardy because it no longer correctly distinguishes VP topicalization from Gapping. I suggested that topicalization cannot relocate a VP from which an object has QRd (as in the ungrammatical (3b), repeated below) because there is a constraint that prevents a trace of QR from being put into a surface position in which it’s not c-commanded by the argument which binds it. (3b)
*It’s eat on Tuesday, that Carrie will kale.
But precisely this is allowed under the approach to Gapping outlined above, as the representations in (40) and (45) indicate. So how, then, do Gapping and VP topicalization differ? The constraint operative here has its antecedents in Fiengo’s (1980) Proper Binding Condition, which we can reframe as (46). (46)
The trace of a moved phrase, , must be c-commanded by at Spell Out.
The view of QR from ellipsis
87
This condition has many variants, mostly distinguished by virtue of whether it holds of Spell Out, LF, or all levels of representation.9 That it cannot hold of Spell Out, at least if it is taken to be a linguistic universal, now seems clear given instances of remnant topicalization, like that in (47), first examined in den Besten and Webelhuth (1990). (47)
[VPt1 gelesen]2 hat das Buch1 keiner t2 read has the book no one ‘‘No one has read the book.’’ (Muller 1996, (9a): 360) ¨
In (47), the object das Buch has moved out of the VP before VP topicalization. I will therefore assume that the Proper Binding Condition is just (48), which can be viewed as a theorem of the mechanisms of semantic interpretation. (48)
A variable must be c-commanded by its binder.
If traces of movement are (typically) interpreted as variables, then (48) can be regarded as (46) holding of LFs. An example such as (47) will satisfy the Proper Binding Condition if its LF representation puts the topicalized VP within the scope of the object. But if the Proper Binding Condition is just an LF condition on variables, how are we to guarantee that QR cannot feed VP topicalization to produce the illicit surface representations in (3) and (4)? Once upon a time, this would have been blocked by restricting QR to the LF portion of the derivation, preventing it from ever forming surface representations. This, then, is one of the reasons why QR was originally taken to be covert in that sense in English. What is needed, it seems, is something that distinguishes VP topicalization from the VP fronting used in the account of Gapping, allowing QR to feed only the former. There is independent support that constraints which make this sort of distinction exist. Muller and Sternefeld (1993) argue, for instance, that a ¨ direct object can scramble out of topicalized constituents in German – including the topicalized VP in (47) – but that they cannot scramble out of scrambled constituents, as in (49), where the VP has scrambled to the left of the subject.10 (49)
*daß [VP ti gelesen]k das Buchi keiner tk hat. that read the book no one has ‘‘No one has read the book.’’ (Muller 1996, (10a): 361) ¨
Muller, and also Takano (1996) working from Japanese data, proposes ¨ conditions that will block certain combinations of movement, depending
88
Kyle Johnson
on the ‘‘types’’ of movement involved. For example, the difference between (47) and (49) on this view is that object Scrambling can feed a different type of movement – namely topicalization – but not a movement of the same type, such as VP Scrambling. In general, the description of this constraint is something like (50). (50)
If Move A affects , and Move B affects and contains , then the types of A and B must be different. and The ‘‘types’’ of movement that Muller’s study distinguishes are A, A ¨ Scrambling. (50) has a certain similarity to Chomsky’s A-over-A condition, and various people have proposed deriving it, and the classical A-over-A effects, from more elementary features of movement theory.11 The formulation in (50) will be sufficient for our purposes. Having in view the German data in (47) and (49) highlights another dimension to the problem. The contrast in these examples makes it appear that German behaves in a way that is nearly opposite to what we need. Scrambling and QR obey what appear to be the same locality conditions, and are in other ways rather similar. This is why we equated them. Moreover, the instance of VP Scrambling in (49) is superficially similar to the VP-to-FP movement used to derive Gapping. If we equate these two, then the contrast between (47) and (49) makes it appear that German behaves precisely opposite to what is needed in English. The typical presumption, one that explains why English has examples that are neither like (47) or (49), is, again, simply that English lacks Scrambling. This is another good reason for thinking that QR does not feed overt structures in English. Indeed, the comparison with German is instructive, for German is a language in which QR is assumed to be overt and the consequences are thought to explain other differences in these languages. One is, as we’ve seen, the availability of remnant topicalization in German. Another is that German lacks Superiority effects of the sort that English has. (51) illustrates. (51)
a. *What did you persuade whom to buy? b. Was hast du wen uberredet zu kaufen? ¨ what have you whom persuaded to buy? ‘‘What did you persuade whom to buy?’’ (Pesetsky 2000, (146b): 76)
This contrast can be explained by letting German feed the step which brings an interrogative pronoun into the specifier of CP with Scrambling, and depriving English of this option. This is because there are ways in which local Scrambling appears to be like A movement, and in this regard it has the ability to overcome the Superiority effect that would otherwise block (51b). Like (50), Superiority is a condition sensitive
The view of QR from ellipsis
89
movement past a to the kinds of movement involved. It prevents A c-commanding term that will itself undergo A movement. Thus, (51a) is moved past whom, which will itself undergo prevented because what has A covert A movement. This compares with (52), in which who first A moves moves into specifier of CP. past whom, and then A (52)
Who seems to whom to have left?
The relevant difference between (51a) and (52), then, is that (52) allows a -movement step to precede the A -movement step that brings who into non-A specifier of CP. A parallel derivation is argued to rescue (51b). The object move into specifier of may first scramble past wen and then from there, A CP. This account makes the presence of overt QR (¼ Scrambling) in German responsible for derivations which allow a wh-phrase to get past moving past it, circumventing Superiority. another without A There is some prima facie evidence for this story. Fanselow, for instance, argues that this would explain why Superiority effects reappear in longdistance environments, where Scrambling is not an option. The example in (53), reported in Pesetsky, illustrates. (53)
*Wen glaubte wer, dass der Peter ihr vorstellte? whom thought who that Peter to her introduced ‘‘Who did who think that Peter introduced her?’’ (Pesetsky 2000, (147b): 77)
Because Scrambling out of finite clauses is blocked in German, there is no moving. derivation of (53) in which wen first scrambles past wer before A Thus, the difference between (51b) and (53) can be seen as a function of the constraints on Scrambling. Promising though this account seems to be, it is not problem-free. Pesetsky (2000), in particular, has questioned each of the arguments on its behalf (as well as proposed an alternative); and there are facts which do not immediately fit the account. Wiltschko (1997), for instance, notes that when the wh-in-situ is a subject, the Superiority effect seems to be determined by whether the subject is within the VP or not. One of her examples illustrating this is (54). (54)
a. Wen hat wer oft gesehen? whom has who often seen ‘‘Whom did who often see?’’ b. ?* Wen hat oft wer gesehen? whom has often who seen ‘‘Whom did who often see?’’ (Wiltschko 1997, (20b), (21b): 440)
90
Kyle Johnson
She suggests that the absence of a Superiority effect in German is a consequence of the conditions under which wh-in-situ may be D-linked. Superiority effects in English are known to weaken when the wh-phrases involved are D-linked,12 and Wiltschko suggests that wh-pronouns in German are able to be D-linked in a wider set of contexts than they are in English. Because VP internal positions coerce pronouns into their nonD-linked readings (see Diesing (1992a)), the contrast in (54) can be seen as a consequence of this effect. Despite the controversy, let’s adopt the view that the absence of Superiority effects in German is a consequence of German’s ability to exploit Scrambling in wh-movement contexts. Because from our viewpoint this is the problematic account of German, it’s the safer position to take. We have, then, three correlative differences between English and German which pose problems. First, German can topicalize VPs from which objects have Scrambled, and English can’t. Second, the method we’ve adopted for producing constituents that Gap in English involves Scrambling objects out of VPs which have undergone a short movement. But precisely this derivation – object Scrambling plus short VP movement – is blocked in German. And, finally, Superiority effects found in English seem to be absent in German in just those cases where wh-movement could be fed by short Scrambling. All of this would seem to speak against Scrambling being able to feed overt representations in English because only, it would seem, in that case could all these differences, and their correlation, be explained. But, on the other hand, allowing Scrambling to feed overt representations is the only method known for explaining how the constituents that Gapping and VP Ellipsis target manage to differ from those that coordination and topicalization see. I suggest that the resolution of this conflict resides in the conditions properties. In particular, I speculate under which Scrambling has A or A that these conditions, whatever they may turn out to be, class the instances of Scrambling that have A movement qualities in German as instances of A movement in English. This proposal exploits the fact that the differences between English and German we have just reviewed all involve processes distinction, whereas the need we have to form that are sensitive to the A–A constituents from Scrambling in English are blind to this distinction. Thus, movement, we will still be able to if we let Scrambling in English be A fashion the constituents needed for VP Ellipsis and Gapping. But, because movement cannot feed VP topicalization, according to (50), this will be A movement does not bleed Superiority blocked in English.13 And because A effects, the means German has for avoiding Superiority will not be avail movement it will able to English. Finally, because English Scrambling is A
The view of QR from ellipsis
91
be able to feed short VP movement, again in accordance with (50). Thus, this supposition would capture all of the differences between English and German reviewed here, while maintaining a movement-based account of the exceptional constituents that ellipsis is able to affect. There is evidence from English consistent with this supposition. Another movement is disjoint phenomenon that distinguishes A movement from A reference effects under reconstruction. Under certain circumstances, names movement must be disjoint in reference within phrases that have undergone A with terms that c-command the position from which they moved. Thus, for instance, there is a disjoint reference effect between she and Carrie in (55a). (55)
a. Which pictures of Carrie did she like best? b. She likes pictures of Carrie best.
Presumably this is because, although which pictures of Carrie has moved from the object position of like, it is still (at least partly) interpreted in this position.14 Therefore, the disjoint reference effect in (55a) emerges because it is enforced by the Binding Theory in (55b), where the pictures argument is visibly in the position that it is interpreted in (55a). Instances of A movement do not have this property; compare (55) to (56). (56)
a. Pictures of Carrie seem to her to be the best. b. It seems to her that pictures of Carrie are the best.
Just as in (55), the pictures of Carrie phrase in (56a) has moved from the position it is shown in (56b). And yet the disjoint reference effect between her and Carrie that exists in (56b) is not inherited by (56a) in the same way that it is in (55a). Inheriting disjoint reference effects from the positions a movement. phrase moves from, then, is diagnostic of A With this in mind, consider now how disjoint reference effects operate in these contexts in which the account we have adopted countenances object movement. Sauerland (1998: 143–144) discovered that in these contexts, a disjoint reference effect is manufactured. (57)
a. Jerry can’t introduce her to Frank’s teachers but of course he can D to Sally’s teachers. b. Some introduced her to Frank’s teachers and others D to Sally’s teachers.
There is a clear tendency to understand the referent of her and Sally to be different in these examples. In both cases, the constituent that has elided is formed by moving the object out of VP, creating a representation like that in (58).
92
Kyle Johnson
(58)
VP VP
PP to Sally’s teachers
V introduce
DP PP her
t
If this representation is to trigger the required disjoint reference effect, the to phrase will have to be interpreted in its underlying position, in the same way that (55a) illustrates. Only then will Sally be within the scope of her in the way required. That the disjoint reference effect in (58) is in fact due to a so-called ‘‘reconstruction’’ effect of the sort that (55a) illustrates, and not some property of focus or one of the many other factors that play a role in ellipsis constructions, can be seen by comparing it to (59). (59)
a. Jerry can’t introduce her mom to Frank’s teachers but of course he can D Sally’s. b. Some introduced her mom to Frank’s teachers and others D to Sally’s teachers.
There is no disjoint reference effect between her and Sally in these examples, and this can be traced back to the fact that her is buried within the object in (59). As a consequence, the PP containing Sally will not be within the scope of her, even in its reconstructed position, and the Binding Theory will not enforce a disjoint reference requirement. distinction. Parasitic Parasitic gaps provide another diagnostic for A/A gaps are licensed when an A movement occurs near them. They cannot live on a neighboring A movement, as the contrast in (60) indicates.15 (60)
a. What did John recognize to be inflammatory drivel even before reading e? b. *This book was can be recognized to be inflammatory drivel even before reading e?
Takahashi (2004) discovered that the remnant of a pseudogap is able to license a parasitic gap (see (61)), indicating again that the remnant escapes movement. the elision by an A (61)
Although John didn’t file a recent article about HNPS, he did D [without reading e] [a recent article about Object Shift]. (Takahashi 2004, 577: (22))
The view of QR from ellipsis
93
A parallel example involving Gapping is (62). (62)
4.5
Some read without understanding e an article about HNPS, and others without acknowledging e a recent article about Object Shift.
Summary and qualifications
What distinguishes German from English, and gives the illusion that English does not have overt Scrambling, is how these languages class Scrambling. German considers it a form of A movement, while English movement. This not only accounts for the differconsiders it a type of A ences between English and German, but gives an explanation for why the constituents formed by Scrambling are visible to ellipsis and Gapping, but not topicalization and coordination. It completes our account of ellipsis specific constituents. This removes the problems attendant with letting English resemble German in having an overt form of Scrambling, but it does so by pushing the differences between English and German back to the deeply mysterious distinction. This distinction is, to my knowledge, grounded in nothA–A ing. To the extent that I have been successful, then, these problems have been hidden within a groundless distinction in movement types. There are other problems. The parasitic gap criterion does not always indicate that the movement responsible for forming ellipsis specific con type. Takahashi (2004) notes that if the remnant stituents is of the A precedes the parasitic gap, the result is not grammatical, as indicated by the Pseudogapping example in (63).16 (63)
*Although John didn’t kiss Mary, he did Sally without looking e. (Takahashi 2004: 580 (41))
Takahashi argues that this is because phrases may move out of ellipses movement, or ‘‘Object either by way of ‘‘Heavy NP Shift,’’ a kind of A Shift,’’ a kind of A movement. Heavy NP Shift is rightwards and Object Shift is leftwards. Object Shift has produced (63), moving the object leftwards across the adjunct, and therefore a parasitic gap is not licensed. By contrast, (61) is produced by Heavy NP Shift, which moves the object movement, a rightwards across the adjunct, and since this is a kind of A parasitic gap is licensed. , If movement out of an elided or gapped VP can be A, as well as A movement in English, then all of the consequences credited to the A nature
94
Kyle Johnson
of this movement will be lost when the A movement option is permitted or forced. If Object Shift is the only type of A movement permitted, then this will restrict the class of relevant cases to just those that involve structurally Case-marked DPs, as only these DPs can Object Shift. This generates the following predictions. (64)
a. Structurally Case-marked DPs will not show disjoint reference effects with respect to material they have moved past in escaping an elided VP. b. Structurally Case-marked DPs will not show Superiority effects with respect to material they have passed in vacating VP. c. If the movement of VP-to-FP operation employed in the analysis of Gapping is classed as a kind of A movement, then Gapping cannot leave remnant DPs that could only escape the VP by way of Object Shift. d. Structurally Case-marked DPs should be able to escape topicalized VPs.
Object Shift does not move structurally Case-marked DPs far enough to be certain that there is any material that these DPs have passed, and so (64a) and (64b) are difficult to test. To test (64c), we can use the fact that the first object of a double object construction cannot Heavy NP Shift, and so it can escape the VP only through Object Shift. Thus, Gapping should not be able to strand the first object of a double object construction. There should be a contrast, then, between (65a), in which this has happened, and (65b), in which an object that can Heavy NP Shift has been stranded. (65)
a. ?? Some gave Bill apples and others Sally. b. Some gave apples to Bill and others oranges.
While there is a contrast in these examples, and it goes in the right direction, I have not been able to reliably reproduce it with consultants. It’s clear that (64d) is incorrect; (66) illustrates. (66)
*[VP read today], no one has this book.
The proposals here, then, do not quite knit together all the pieces of our puzzle. We must either abandon Takahashi’s explanation for the contrast , or we must between (61) and (63), and class all movement out of VP as A find another explanation for the absence of remnant VP topicalization.
5
Argument Contained Ellipsis Christopher Kennedy University of Chicago
Abstract This chapter investigates an unusual identity constraint on English Verb Phrase Ellipsis which imposes the following requirement: when an elliptical relation holds between two verb phrases A and B such that A is contained in an argument b of B, then the corresponding argument a of A must be identical to b. The paper argues that this is due to two factors: (i) an identity constraint on ellipsis in general that requires indices in related VPs to be the same, and (ii) constraints on the distribution of indices that rule out referentially circular constructions. 5.1
Introduction
Wasow (1972) notes the impossibility of Verb Phrase Ellipsis in sentences such as (1): (1)
*A proof that God exists does.
Wasow accounts for (1) and similar sentences with a constraint that prohibits an elided VP from finding an antecedent within its subject. There is evidence that this constraint is too strong, however. Configurations of the type specified by Wasow are grammatical just in case the subject of the elided VP is identical to the subject of its antecedent, as shown by (1), a contrast that was brought to my attention by Jorge Hankamer. (2)
a. Every man who said he would buy some salmon did. b. Every man who said George would buy some salmon did.
In fact, the issue is even more general than these examples indicate: in any structure in which one of two elliptically related VPs is embedded in an Acknowledgments: This chapter is an essentially unmodified version of Kennedy (1994): for the sake of historical (in)accuracy, I have only corrected typos and inlined a few notes; all other aspects of the original remain the same. I remain very grateful to Sandra Chung, Jorge Hankamer, William Ladusaw, and James McCloskey for their guidance on this work.
95
96
Christopher Kennedy
argument of the other, the arguments must be identical. The examples in (3), which involve Antecedent Contained Deletion (ACD), show that the internal arguments of the related VPs must be identical. (3)
a. Polly visited every town Eric did. b. *Polly visited every town located in a country Eric did.
(3a) can be paraphrased as follows: for every town Eric visited, Polly visited it too. A similar interpretation of (3b) – for every town located in a country that Eric visited, Polly visited it – is unavailable. The crucial difference between (3a) and (3b) is that in (3a), the direct objects of both the overtly expressed and elided verbs are understood to be the same, whereas in (3b) they differ. If ellipsis has not occurred, a sentence with the meaning of (3b) is perfectly grammatical: (4)
Polly visited every town located in a country Eric visited.
The following examples indicate that similar facts hold of indirect objects. (5)
a. Eric sent letters to every senator Polly did. b. *Eric sent letters to every aide who worked for a senator Polly did. c. Eric sent letters to every aide who worked for a senator Polly sent letters to.
The generalization that arises from this data, which I will refer to as Argument Contained Ellipsis (ACE), can be stated as in (6): (6)
Argument Contained Ellipsis Ellipsis between VP1 and VP2, VP1 contained in an argument A2 of VP2, is licensed only if A2 is identical to the parallel argument A1 of VP1.
(6) is a descriptive statement reflecting a law of English grammar, but it is unclear why argument identity should hold in just these configurations. Ideally, this constraint should follow directly from the architecture of a theory of VP Ellipsis. The goal of this paper is to demonstrate that the argument identity constraint described in (6) follows from a fundamental property of anaphoric relations: dependence. A dependence relation is the relation that holds between some element of linguistic structure (possibly silent) that cannot be independently interpreted and some other expression that provides the dependent element with a full interpretation. Specifically, I will show that the interaction of two types of dependence relations is responsible for the argument identity constraint in Argument Contained Ellipsis. The first is
Argument Contained Ellipsis
97
the relation that holds between an anaphoric element in an argument position and the linguistic element from which it receives an interpretation (its antecedent, a relation that has been referred to as referential dependence, see Higginbotham 1983; Ha ¨ık 1984; 1987). Intuitively, the interpretation of an anaphoric element is determined by the composition of its antecedent; hence an anaphoric element is dependent on anything contained in its antecedent. For example, in (7), the pronoun he, when interpreted as coreferential with the matrix subject, is dependent on the DP those dogs. In order to determine the value of the antecedent of the pronoun, and hence the value of the pronoun, it is necessary to first determine the value of those dogs. (7)
[The man who owns those dogs]i says hei takes good care of them.
The second relevant dependence relation is the one that holds between an elided verb phrase and the overt VP from which it receives its interpretation. I will refer to this relation as elliptical dependence. The status of elliptical dependence as a relation distinct from other types of anaphoric dependencies has been discussed at various points in the past (for example, in Grinder and Postal 1971; Hankamer and Sag 1976; Sag and Hankamer 1984), but it is not directly represented in recent theories of ellipsis. Since Sag (1976) and Williams (1977), it has generally been accepted that ellipsis is licensed by an identity relation between VPs which holds at some level of semantic representation. Subsequent analyses of VP Ellipsis within the Extended Standard Theory have pursued the idea that identity can be stated in terms of a syntactic level of Logical Form (LF) (e.g. May 1985; Kitagawa 1991; Fiengo and May 1994), while more purely interpretive approaches, in which the resolution of ellipsis is carried out completely in the semantic component, have followed Sag in stating identity in terms of logical expressions (e.g. Rooth 1981; Dalrymple et al. 1991). In each of these analyses, the extent to which an elided VP can be said to be dependent on an antecedent VP is the identity relation that governs the construction of the interpretation of the elided VP (whether a syntactic or semantic representation). This relation is loose, because a common aspect of the identity requirement in these models is that it permits variation of indexical values (in the syntactic accounts) or variable values (in the semantic accounts) across related VPs. Once the representation of an elided VP has been constructed, there is no fact about its structure to indicate that its interpretation is dependent on some other VP in the discourse. Elliptical dependence is not directly encoded into the representation of ellipsis. In this paper, I will claim that the elliptical dependence relation should be explicitly represented in a theory of ellipsis interpretation, and I will demonstrate that only in such a model can the Argument Contained Ellipsis data be given a uniform analysis. I will propose that elliptical
98
Christopher Kennedy
dependence be incorporated into a theory of ellipsis by treating elliptically related VPs as identical occurrences of a single VP. That is, similarly to the approaches listed above, I will state elliptical dependence in terms of an identity relation. In contrast to the above approaches, however, I will argue that the identity relation that holds between elliptically related VPs must be strict, where strict identity forbids variation of indexical or variable value across elliptically related VPs. The reason behind this move is to incorporate into the architecture of a theory of ellipsis the following claim: ellipsis involves the reuse of linguistic material (a notion suggested by Bill Ladusaw, p.c.), material which is treated by other components of the grammar as identical to the original. This, I claim, is the nature of elliptical dependence. I will show that it is this aspect of ellipsis that is responsible for the argument identity constraint in Argument Contained Ellipsis. Dependence relations are determined with respect to semantic representations, and are subject to certain restrictions. In particular, referentially circular constructions must be ruled out. This is standardly accomplished by stipulating that referential dependence relation is irreflexive: no element may be referentially dependent on itself (Higginbotham 1983; Ha ¨ık 1987; cf. Chomsky’s 1981 i-within-i condition). I will demonstrate that by treating elliptically related VPs as identical occurrences of the same VP, configurations in which one of two elliptically related VPs is contained in an argument of another, give rise to circular constructions unless argument identity obtains. In this way, the descriptive constraint (6) will be explained in terms of a general semantic constraint that rules out structures that contain circular dependencies. The representations of referential and elliptical dependencies, and the model of ellipsis in which these will be cast, will be stated in terms of the principles and assumptions of the Principles and Parameters framework, circa Chomsky (1981; 1986a, 1986b). I will state conditions on referential dependence as conditions on Logical Forms; the basic claims, however, are not theory-dependent. The strict identity hypothesis and its interaction with referential dependence are theory-independent, and could in principle be recast in other syntactic frameworks. The chapter is organized as follows. Section 5.2 gives a more detailed presentation of the data which lead to the generalization given above in (6) and outlines the problems it presents for a number of current theories of VP Ellipsis. Section 5.3 presents the strict identity hypothesis and describes its implementation in the grammar. In particular, I argue that the strict identity hypothesis requires adopting a theory of binding and coreference such as that described in Reinhart (1983b) and extended in Heim (1993). In section 5.4 I discuss referential dependence, and, drawing on the work
Argument Contained Ellipsis
99
of Higginbotham (1983) and Ha ¨ık (1985; 1987), I formulate conditions on referential dependence. Section 5.5 lays out crucial assumptions regarding the semantics of relative clauses, and finally, section 5.6 presents an account of Argument Contained Ellipsis. 5.2
Argument Contained Ellipsis
VP Ellipsis is characterized by the apparent deletion of a full VP under identity with some VP in the discourse. This phenomenon is illustrated by the examples given in (8), in which the missing VP is represented by the symbol D. (8)
a. Fedka loved Catherine the Great but Lyosha didn’t D. b. Lena wants to move to Rome but her husband doesn’t expect to be able to D. c. Phyllis should have brought some artichokes, and you should have D, too.
VP Ellipsis is not limited to conjoined structures, as shown by (9): (9)
a. If John plays well, Lou does D too. b. Sterling will try the punch after Maureen has D. c. Mo likes every song that Lou does D.
The data that I will be investigating in this chapter form a subset of non-conjoined elliptical structures, in which one of two elliptically related VPs is contained in an argument of the other. As noted in (6), if this relation obtains, the containing argument must be identical to the parallel argument of the contained VP. The sentences in (8)–(9) illustrate that this identity requirement does not hold in general. Section 5.2.1 is devoted to demonstrating that the identity requirement is enforced in a variety of structures in which the structural configuration described in (6) does hold. 5.2.1
The data
5.2.1.1 Matrix subjects As shown by (1) above, sentences in which an elliptical relation holds between a matrix VP and a VP embedded in a clausal complement of the head of the subject are ungrammatical. (10a) and (10b) further illustrate this paradigm. (10)
a. *Every claim that her theory is incorrect is D. b. *The belief that we are mistaken is D.
100
Christopher Kennedy
There are similar constructions in which an elliptical relation may hold between a matrix VP and a VP embedded in its subject, however. Examples in which the subject contains a restrictive relative clause are grammatical just in case the subjects of the related VPs are identical.1 Consider the following sentences: (11)
a. Every man who wants to buy some salmon should D. b. *Every man who wants George to buy some salmon should D.
(12)
a. At least one woman who thought she would catch a tuna did D. b. *At least one woman who thought the man from Portland would catch a tuna did D.
(13)
a. The man who promised Horace to leave town did D. b. *The man who persuaded Horace to leave town did D.
In each pair, when the subject of the non-elided VP is distinct from the subject of the elided VP, the result is an ungrammatical sentence.2 The examples in (13) are a particularly interesting minimal pair. In (13a), which contains the subject control verb promise, the subjects of the related VPs are identical. (13b) on the other hand contains the object control verb persuade. In this case, the subject of the embedded VP is interpreted as coreferential with the object of persuade: the DP Horace. The subjects of the related VPs are not identical, and the sentence is ungrammatical. Note that the judgments are the same regardless of whether the elided VP is the embedded or the matrix VP: (14)
a. Every man who wants to D should buy some salmon. b. *Every man who wants George to D should buy some salmon.
(15)
a. The man who promised Horace to D left town. b. *The man who persuaded Horace to D left town.
The non-elided counterparts of all of these sentences are grammatical, as shown by (16a)–(16c). (16)
a. Every man who wants George to buy some salmon should buy some salmon. b. The woman who thought the fisherman would catch a tuna caught a tuna. c. The man who persuaded Horace to clean the fish cleaned the fish.
Argument Contained Ellipsis
101
These generalizations hold for other types of subjects as well. An elliptical relation between a matrix VP and a VP embedded in a sentential subject is impossible, as shown by (17). (17)
*That you think that their actions are surprising isn’t D.
An elliptical relation is possible, though, when the subject is a free relative clause: (18)
What you think is surprising isn’t D.
Examples like (18) are grammatical only when the free relative is interpreted as the subject of both the matrix and embedded VPs, however. (19), in which the embedded VP has a distinct subject, is ungrammatical. (19)
*What I thought would persuade Stanley to change his mind didn’t D.
In (18), the entity referred to by the free relative is the subject of both the matrix and embedded VPs. In (19), however, the free relative is interpreted as the subject of the elided matrix VP, but the subject of the embedded VP is Stanley, and the sentence is ungrammatical. 5.2.1.2 Embedded subjects The identity requirement holds of structures in which one of two elliptically related VPs is contained in the subject of a clausal complement of the other, for example, in cases of Antecedent Contained Deletion (ACD) in embedded nonfinite clauses (20) and small clause structures (21). (20)
a. Mona wants the candidates that Jack does D to be successful. b. *Mona wants the rivals of the candidates that Jack does D to be successful.
(21)
a. Max considers every senator Hector does D corrupt. b. *Max considers the aides of every senator Hector does D corrupt.
These examples are somewhat more complex than those involving identity of matrix subjects, because they involve an elliptical relation between VPs that contain embedded clauses. (20) and (21) do not involve identity of arguments of the elliptically related higher VPs, but rather identity of arguments of embedded VPs. The basic configuration is shown in (22), where VP1 and VP2 are the elliptically related VPs, and XP is some clausal constituent.
102
Christopher Kennedy
(22)
IP VP1
Subj V
XP DP
VP
... DP VP2 ... As (20)–(21) demonstrate, the interpretation of the boxed DPs must be the same in both the elided and non-elided VPs. The (b) sentences, in which the embedded subject contained in the elided VP is not identical to the embedded subject contained in the matrix VP, are ungrammatical. As in the case of matrix subjects, the non-elided counterparts of these sentences are grammatical. (23)
a. Mona wants the rivals of every candidate that Jack wants to be successful to be successful. b. Max considers the aides of every senator Hector considers corrupt corrupt.
5.2.1.3 Internal arguments The identity requirement holds when the elided VP is contained in an internal argument of the antecedent VP, as illustrated by the following ACD examples. (24)
a. Polly visited every town Erik did D. b. Polly visited every town in every country Erik did D.
(25)
a. Henry read each book that Fedja did D. b. *Henry read each chapter of every book that Fedja did D.
(26)
a. Max fooled none of the senators that Hector will D. b. *Max fooled none of the aides of the senators that Hector will D.
The interpretation of (24a) is given by the semantic expression in (27a): for every town that Erik visited, Polly visited it, too (similarly for the other (a) sentences). The ungrammatical reading of (24b) is the one represented by the semantic formula in (27b): for every country that Erik visited, Polly visited all of its towns (similarly for the other (b) sentences).
Argument Contained Ellipsis
103
a. 8x½townðxÞ ^ visitedðErik; xÞ½visitedðPolly; xÞ b. 8x8y½townðxÞ ^ countryðyÞ ^ inðx; yÞ ^ visitedðErik; yÞ ½visitedðPolly; xÞ As noted above, in the ungrammatical examples, the elliptically related VPs have nonidentical internal arguments. (I only consider examples of direct objects here, but as shown in the introduction, similar facts hold of indirect objects as well.) (28a)–(28c), which do not involve ellipsis, are grammatical. (27)
(28)
a. Sally visited every town in every country Jerome visited. b. Henry read each chapter of every book that Fedja read. c. Max fooled none of the aides of the senators that Hector will fool.
One could argue that these data are not surprising: since it is true that ellipsis is licensed by some sort of identity relation between VPs, and internal arguments are contained within the VP, internal arguments of elliptically related VPs must always be identical. Although this claim is true in general, it is not an absolute fact. An example like (29), first noted by Hirschbuhler (1982), has an interpretation in which the phrase most houses in the first conjunct refers to a different set of houses from its counterpart in the interpretation of the second conjunct, thus the internal arguments of the related VPs are not identical. (29)
A Canadian flag flew over most houses and an American one did too.
Similarly, in (30a), which involves wh-movement of a direct object out of the VP, and (30b), in which the related VPs are in parallel relative clauses, the internal argument of the elided VP is not identical to the internal argument of the antecedent VP. (30)
a. I know which senators Polly interviewed and which congressmen Eric did. b. I’ll interview everyone Max robbed if you interview some of the people that Molly did.
Examples like these demonstrate that it is not an absolute property of VP Ellipsis that internal arguments are identical, though it may be a common one.3 5.2.2
The intractability of Argument Contained Ellipsis (ACE)
Clearly, the most interesting and enlightening analysis of the data presented above is one that provides a general account of the entire array of
104
Christopher Kennedy
facts – that is, one that accounts uniformly for the fact that both external and internal arguments of elliptically related VPs are subject to an identity requirement when one of the related VPs is contained in the parallel argument of the other. Ideally, this constraint should follow from the structural characteristics of the sentences presented above and from the interaction of a theory of ellipsis with other components of the grammar. The argument identity constraint in ACE structures is not generally explained within current models of ellipsis, however. Although syntactic analyses such as Kitagawa (1991) and Fiengo and May (1994) are able to account for a subset of the data – the cases involving Antecedent Contained Deletion, in particular – these models predict that sentences such as (1) and (2b) – those in which an elliptical relation holds between a VP and a second VP embedded in the subject of the first – should be grammatical. The same generalization holds of Sag’s (1976) deletion analysis, and current interpretive approaches such as Dalrymple et al. (1991): these models do not predict the identity requirement in the case of subject containment. These analyses of ellipsis involve use of the lambda calculus. In Sag (1976), deletion is licensed iff the lambda expression corresponding to the meaning of a deleted VP is an alphabetic variant of the lambda expression corresponding to the meaning of some other VP in the discourse. In Dalrymple et al. (1991), a sentence containing an elided VP is interpreted by determining the property (a lambda expression) predicated of its subject. Crucially, there is no connection (other than the predicate–argument relation) between the subject and the lambda expression, and therefore no way to place constraints on subject identity across related VPs. There is a potential explanation for the ACE facts in the theory developed in Dalrymple et al. (1991), however. In this theory, the interpretation of ellipsis involves two separate tasks. First, parallelism between the structurally expressed arguments of an elided VP and the arguments of an overt VP is determined; second, an equation that determines what property is predicated of the subject of an elided VP is solved. The second part of this process is crucially dependent on the first: in order to find an appropriate interpretation for a sentence containing an elided VP, it is necessary to determine which arguments in the discourse are parallel to the arguments of this VP. Building on this aspect of the theory, one could devise an explanation for the facts presented in the previous section along the following lines: it is impossible to determine parallelism between the arguments of two elliptically related VPs when the structural configuration described in (6) obtains, unless the parallel arguments were identical. The burden placed on such an analysis would be to show that it was not a
Argument Contained Ellipsis
105
restatement of (6); i.e., to avoid being a descriptive generalization, not an explanation. The question asked in the introduction could be rephrased: why is argument identity required in ACE constructions in order to determine parallelism? If the above characterization of the issues facing a parallelism-based approach is correct, however, then the analysis that I will propose in this chapter can be thought of as an answer to this question as well as the one posed in the introduction. In attempting to explain the argument identity requirement in terms of the interaction of dependence relations in elliptical constructions, I take dependence, as defined in the introduction, as a fundamental, theory-independent aspect of anaphoric relations. If this claim is correct, then the results of this work should be able to be implemented in any model of ellipsis. As noted above, two types of dependence relations play a crucial role in ACE configurations: elliptical dependence – the relation between an elided VP and the VP from which it receives its interpretation, and referential dependence – the relation between a nominal anaphoric element and its antecedent. In section 5.4, I will discuss those aspects of referential dependence that are crucially involved in Argument Contained Ellipsis. In the next section, I will propose a means of directly representing elliptical dependence in a syntactic model of VP Ellipsis. 5.3
Elliptical dependence
Since Sag (1976) and Williams (1977) it has generally been accepted that the licensing and interpretation of elliptical structures is controlled at some level of semantic representation. The theories of ellipsis articulated in Sag (1976) and Williams (1977) were conceptually very similar, differing primarily in that Sag’s approach defined the conditions licensing VP deletion, whereas Williams assumed that ellipsis involved the interpretation of empty syntactic structure corresponding to a null VP by copying the semantic interpretation of an antecedent VP (cf. the Empty Structures Hypothesis of Wasow 1972). Both approaches shared the hypothesis that the conditions governing ellipsis were stated in semantic terms. Subsequent work in Government and Binding (e.g. May 1985; Chao 1987; Kitagawa 1991; Fiengo and May 1994) has assumed that the interface between the semantic component and the syntactic component occurs at a syntactic level of Logical Form (LF), which is derived transformationally from S-structure. These analyses claim that, given the assumption that certain well-formedness constraints apply at LF, the phenomena exhibited by elliptical structures can be shown to follow from the assumption that ellipsis involves the ‘‘reconstruction’’ of a null VP at LF, a process
106
Christopher Kennedy
that is controlled by the syntactic form of the antecedent VP. The resulting syntactic structure is then subject to LF well-formedness constraints. Reconstruction-based accounts of ellipsis differ from deletion-based accounts on issues related to the organization of the grammar, but agree on the basic hypothesis that ellipsis involves syntactic representation of an elided VP at some level of structure.4 This is the position that I will adopt in this chapter, though I will remain agnostic as to whether ellipsis involves reconstruction of syntactic material at LF or deletion of syntactic material at Phonological Form (PF).5 The crucial difference between previous syntactic accounts of ellipsis and the one to be proposed here is that these accounts license ellipsis through a ‘‘loose identity’’ relation between the semantic representations of related VPs. That is, although all syntactic accounts of ellipsis agree that the related VPs must have the same basic syntactic structure, loose identity accounts permit variation – within defined parameters – in the value of syntactic indices or variables across related VPs. The net result of loose identity is that the elliptical dependence relation – the connection between the elided VP and the antecedent – is not explicitly represented at LF. At the syntax–semantics interface, the input to the interpretive component is a syntactic structure in which the surface-null VP and the VP from which it receives its interpretation are distinct. Although the related VPs have the same basic structure (up to permitted indexical variation), there is no fact about them indicative of the elliptical relation which could interact in a crucial way with some other component of the grammar. My claim is that the ACE data can be uniformly explained only if there is an explicit representation of elliptical dependence at LF, allowing it to interact with other components of the grammar; specifically, with other dependence relations. In the following section, I will propose a way of incorporating elliptical dependence into a syntactic model of ellipsis. 5.3.1
Strict identity in VP Ellipsis
In section 5.1, I claimed that ellipsis involves the reuse of linguistic material, and that the elliptical dependence relation reflects this ‘‘recycling.’’ I propose that this relation be incorporated into a syntactic theory of ellipsis by imposing the following licensing condition on ellipsis, where strict identity means that indexical values are constant across elliptically related VPs.6 (31)
Deletion of VP1 is licensed at PF only if there is some VP2 in the discourse such that VP1’s LF is identical to VP2’s LF.
Consider the implications of this statement, which I will refer to as the strict identity hypothesis. The claim that elliptically related VPs are strictly
Argument Contained Ellipsis
107
identical entails that all syntactic indices that appear in the LF representation of an antecedent VP must be reproduced exactly in the LF representation of an elided VP. Given these assumptions, it follows that a sentence like (32a) has the LF shown in (32b).7 (In what follows, I will indicate ellipsis in LFs by striking through the structure that is omitted from the surface form.) (32)
a. Leo drives a Dart and Felix does too. b. Leo1 [VP t1 drives a Dart] and Felix1 does [VP t1drive a Dart] too.
A structure like (32b) would not arise in loose identity models of ellipsis for two reasons. First, in e.g. Kitagawa (1991) and Fiengo and May (1994), the VP internal subject trace is just the type of anaphoric element whose indexical value may vary across elliptically related VPs. Second, in these models, coindexation entails coreference, so a structure like (32b) could only have an interpretation in which the names Leo and Felix refer to the same individual. The more normal reading, in which the names refer to distinct individuals, would be unavailable. (32a) is a specific example of a general result of the proposed model of ellipsis: in order to satisfy the LF requirement that the trace in the specifier of an elided VP must have a coindexed antecedent, the subjects of elliptically related VPs must bear identical syntactic indices. A Logical Form in which this relation does not obtain, such as (33), would be ruled out by independent principles of the grammar (cf. note 7). (33)
Leo1 [VP t1 drives a Dart] and Felix2 does [VP t1drive a Dart] too.
Clearly, if the strict identity model is to be accepted, a distinction must be made between coindexation, coreference, and the type of relation that holds between traces and their antecedents, in order to ensure that structures such as (32b) are both well-formed with respect to syntactic constraints and interpretable in a normal way. Precisely this distinction is made in the model of anaphoric relations developed in Reinhart (1983b). A fundamental claim of this theory is that coindexation does not entail coreference, nor does non-coindexation entail non-coreference. Indices – numerical subscripts on DPs – represent only syntactic binding relations. The central claim of Reinhart’s approach to anaphora is that bound variable anaphora and coreference are fundamentally different: whereas bound variable anaphora is dependent on structural relations between constituents and should be represented in the syntax, coreference is essentially a discourse-dependent relation that is controlled by pragmatic factors. In this framework, the role played by syntactic indices in the determination of reference is substantially
108
Christopher Kennedy
different from that in other models of anaphoric relations, such as Fiengo and May (1994), in which coindexation entails coreference. For Reinhart, indices represent only binding relations; they do not indicate coreference. In this system, then, there is no requirement that the like-indexed subjects of the elliptically related VPs in (32b) corefer. (32b) is a perfectly wellformed LF. The indices on the subjects of the elliptically related VPs indicate which elements within their respective c-command domains are to be interpreted as variables bound by them. That is, the coindexed subjects bind all and only like-indexed anaphoric elements within their respective scopes; the interpretation of these anaphoric elements is governed solely by their binders, not by the numerical values of syntactic indices (Reinhart 1983b; Heim 1993; Grodzinsky and Reinhart 1993). This fact can be clearly demonstrated if the semantic translation of (32b) is represented using lambda notation, substituting the same letter, x, for the numerical index 1, as in (34): (34)
Leo lx(x drives a Dart) and Felix lx(x drives a Dart).
The interpretation of the variable bound by the lambda operator in each conjunct is governed solely by the scope of the lambda operator that binds it, not by the alphabetic value of the variable itself. Although Reinhart’s system provides a framework of anaphoric relations in which the strict identity model of ellipsis can be implemented, it does not provide a formal means of indicating coreference. As noted above, coreference in Reinhart’s model is not encoded in the formal representation of a sentence, rather it is pragmatically determined. There is no way to state grammatical constraints on coreference relations in terms of syntactic indexing configurations, because these types of indexing configurations do not exist. Only binding relations are indicated by syntactic indexing. In section 5.4, we will see that both coreference and bound variable anaphora are involved in referential dependence relations, and that constraints on possible referential dependence relations can be clearly stated in terms of constraints on indexing configurations at Logical Form. As I intend to show that the Argument Controlled Ellipsis facts follow from the interaction of elliptical dependence and referential dependence, it would be to my advantage to employ a system of anaphoric relations that formally encodes both binding and coreference. Heim (1993) develops a revision of Reinhart (1983a) which does exactly this. Heim’s system captures the formal distinction between bound variable anaphora and coreference, but also allows coreference relations to be represented in the syntax. In order to ensure maximal clarity of the presentation of the analysis in section 5.6, I will adopt Heim’s revision of Reinhart in this chapter. In section 5.3.2, I will give a brief introduction to Heim (1993), then show how
Argument Contained Ellipsis
109
the Logical Form of a sentence involving VP Ellipsis is constructed within the more articulated theory of indexing proposed by Heim.
5.3.2
Indices, binding, and coreference
Heim (1993) makes the distinction between bound variable anaphora and coreference formally precise by encoding both types of relations into the structural representation of sentences via syntactic indexing configurations. This task is accomplished by allowing DPs to receive two indices, ‘‘an inner index that encodes what they are bound by, and an additional index that encodes what they in turn bind . . . The inner and outer index need not be the same’’ (Heim 1993: 37). (As Heim notes, such a system has a precedent in the linking framework of Higginbotham (1983), a connection that I will return to below.) Indexing relations are defined as in (35a) and (35b) where linking is the relation relevant to bound variable anaphora, and colinking is the relation relevant to coreference. For reasons that will become clear in section 5.4.1, having to do with issues of referential dependence and referential circularity, I have modified Heim’s definitions in order to explicitly state that these relations are irreflexive.8 (35)
a. X is linkeddef to Y iff Y’s outer index equals X’s inner index and X 6¼ Y. b. X and Y are colinkeddef iff X’s inner index equals Y’s inner index and X 6¼ Y.
As noted by Bill Ladusaw (personal communication), the effect of this indexing system is to represent in the syntactic structure the distinction between a variable and the name of a variable. Intuitively, a variable is an anaphoric element whose interpretation is set by some outside operator or function, while a name of a variable is a diacritic on an operator that indicates which variables within a certain domain (e.g. the operator’s scope) that operator may bind. This distinction is illustrated by the predicate logic formula in (36). (36)
9x½PðxÞ ^ QðxÞ
In (36), there are two occurrences of the variable x: as arguments of the predicates P and Q. The occurrence of x to the immediate right of the universal quantifier is not a variable, but rather a diacritic indicating which variables may be bound by the quantifier. Heim’s indexing system directly parallels this example. Inner indices correspond to variables in the sense described above, and outer indices are diacritics indicating the ‘‘binding potential’’ of a particular DP.
110
Christopher Kennedy
Crucially, outer indices do not play a role in determining the referential value of the DP to which they are associated, a fact which will be important to the implementation of the strict identity hypothesis in section 5.3.3. The function of an outer index is to indicate that a particular DP has operator status and to indicate which indices within that DP’s scope are to be interpreted as variables bound by it. Outer-indexed DPs in Heim’s system are precisely those DPs in Reinhart (1983a) that are antecedents in cases of bound variable anaphora; that is, those DPs that c-command a coindexed pronoun. In Reinhart (1983a), both the variable status of a pronoun and the operator status of a DP are represented by the same type of indexing notation. The effect of Heim’s notation is to distinguish between indices that correspond to variables (inner indices), and indices that indicate the operator status of a DP (outer indices). With this understanding of the formal notation, we can make the following claims about the interpretation of linking and colinking configurations.9 (37)
a. A node X is bounddef by a node Y iff X is linked to Y and Y c-commands X. b. Translation Definition Inner indices are translated as variables at LF.
These relations are illustrated by the Logical Form in (38). (38)
Leo2:1[VP t1 likes his2 Dart]
In (38), the VP internal trace is linked to and c-commanded by the subject Leo, and so is bound by this DP. The pronoun his is not linked to the subject, but these DPs are colinked. In Heim’s system, colinking – identity of inner indices – is the indexing relation relevant to coreference. Given the translation definition (37b), it follows that an occurrence of an inner index that is not c-commanded and linked to some outer index must be a free variable. For an LF to be interpretable, all free variables must receive referents. This task is accomplished by a rule of reference assignment which requires all occurrences of a particular free variable to denote the same individual (see Heim 1993: 53). According to this rule, the LF in (38) is interpretable just in case the colinked DPs Leo and his corefer. A detailed explication of the colinking/linking system is beyond the scope of this chapter; I refer the interested reader to Heim (1993) for a discussion of how incorporating these relations into the syntax allows her to account for a range of traditionally problematic data (e.g. apparent Condition B violations). It is important to note, however, that this system
Argument Contained Ellipsis
111
departs from the concept of syntactic indexing assumed in Reinhart (1983a). In particular, it departs from Reinhart’s claim that coreference is not formally represented in the syntax by syntactic indexing. Heim’s rule for interpreting colinked structures means that coindexation entails coreference: every occurrence of a particular free inner index must be associated with the same referent. Although there is nothing to prevent different inner indices from being associated with the same referent (as in Reinhart 1983a), it cannot be the case that the same free index is associated with different referents. To summarize, all inner indices correspond to free variables. Outer indices are diacritics that indicate which indices within a particular DP’s c-command domain are to be construed as bound by that DP. This system provides a formal means of representing both binding (via linking), and coreference (via colinking). This more complex formal apparatus permits us to maintain the conceptual division between coreference and binding at the heart of Reinhart (1983a), but, at the same time, allows us to impose specific grammatical constraints on each relation. This fact will be of crucial importance in the discussion of referential dependence in section 5.4. But before we address those issues, I will discuss the representation of ellipsis under the strict identity hypothesis, given the assumptions made in this section. 5.3.3
The representation of ellipsis
As stated in (31), I claim that an elided VP is licensed in the phonological representation only if its Logical Form is identical to the Logical Form of some VP in the discourse.10 Following Heim (1993), I will assume that LFs are related to surface structures by an optional operation of quantifier raising (QR), which takes any DP that bears an outer index and adjoins it to a dominating clausal node, leaving behind a trace which bears an inner index of the same value as the outer index of the moved DP.11 For the purposes of this chapter, we may assume that QR does not apply unless it has to (e.g. in Antecedent Contained Deletion structures; see May 1985 and section 5.6 below). In particular, I will assume that a DP in SpecIP that bears an outer index is interpreted as an operator, and the trace in SpecVP as the variable that it binds (see Diesing 1992b). As noted in section 5.3.1, given the claim that the LF of an elided VP is identical to the LF of its antecedent, this requirement would seem to force coindexation of the subjects of elliptically related VPs. The situation is somewhat different now, given the assumptions regarding the representation of anaphoric relations that were made in
112
Christopher Kennedy
the last section. In Heim’s system, the Logical Form of a sentence like (39a) is (39b).12 (39)
a. Leo drives a Dart and Felix does too. b. Leo2:1 [VP t1 drives a Dart] and Felix3:1 does [VP t1 drive a Dart] too
(31) states that the LF of an elided VP must be identical to the LF of some VP in the discourse. This condition is true of (39b). Moreover, this LF is well-formed with respect to operator-variable relations: in both conjuncts, the subjects bind a trace in SpecVP. This binding relation in the second conjunct holds because the subject of the elided VP bears the same outer index as the subject of the antecedent VP. (39b) should be contrasted with (40), in which the subjects of the elliptically related VPs do not bear the same outer index. (40)
*Leo2:1 [VP t1 drives a Dart] and Felix3:4 does [VP t1 drive a Dart] too.
Although an elliptical relation is licensed in a structure like (40) because the VPs are strictly identical, (40) is not a well-formed LF. Assuming a general condition on LFs that subjects must bind a trace in SpecVP (see note 7, cf. also Full Interpretation, Chomsky 1986b), (40) is ill-formed because the subject of the second conjunct does not bind the trace in SpecVP. Finally, the analysis claims that a structure like (41) is not a possible LF of (39a). (41)
*Leo2:1 [VP t1 drives a Dart] and Felix3:4 does [VP t4 drive a Dart] too.
Since the VPs in (41) differ in indexical value of the traces in SpecVP, they do not satisfy the identity constraint in (31), and ellipsis is not licensed. These examples illustrate a specific result of the proposed analysis of ellipsis: the subjects of elliptically related VPs must bear the same outer indices; elliptical structures in which this indexing does not hold are uninterpretable. Such indexing configurations are perfectly compatible with Heim’s system, the crucial point being that identity of outer indices does not indicate an anaphoric relation between two DPs. It is this particular aspect of Heim (1993) that allows us to maintain the claim that ellipsis involves strict identity of related VPs, as stated in (31). The most important difference between a model of VP Ellipsis based on strict identity and one based on loose identity is that in the former, elliptical dependence is explicitly represented at LF by the identity of indexical values across related VPs. This identity should be understood as a formal means of representing the proposal that the Logical Form of an
Argument Contained Ellipsis
113
elided VP is treated by the grammar as identical to the LF of the VP from which it receives its interpretation; that is, that elliptical dependence represents the reuse of linguistic material. In the loose identity model, in which elliptical dependence is not directly represented, the representations of elliptically related VPs are distinct. In most cases of ellipsis, the hypothesis that an elided VP is treated as identical to its antecedent does not have unexpected results, given a model of anaphoric relations such as Heim (1993) (or Reinhart 1983a). Because two distinct DPs may bind like-valued variables (by virtue of bearing identical outer indices) – modulo configurational restrictions such as c-command – the strict identity analysis is indistinguishable from an analysis in which the indexical values of anaphoric elements within the VPs are allowed to vary. There is a set of configurations in which the strict identity and loose identity models make different predictions, however. These are configurations in which one of two elliptically related VPs is contained in an argument of the other; that is, Argument Contained Ellipsis configurations. The two models diverge in their predictions as the result of the interaction between elliptical dependence, as formalized in this section, and referential dependence, which I defined in the introduction as the relation between an anaphoric element in an argument position and its antecedent. In section 5.4 I will present a theory of referential dependence based on the work of Higginbotham (1983) and Ha ¨ık (1987), relativized to the system of anaphoric relations adopted here. In section 5.6 we will see that the interaction of referential and elliptical dependence accounts for the Argument Contained Ellipsis facts. 5.4
Referential dependence
5.4.1
Coreference and binding
As noted by Heim, her double indexing system bears a strong resemblance to the linking framework of Higginbotham (1983), a central principle of which is referential dependence. Stated roughly, an anaphoric element is dependent on its antecedent and anything contained in its antecedent. The exact formulation is given in (42) (Higginbotham 1983: 404). (42)
X is dependentdef on Y if: i. Y is contained in an antecedent of X or ii. for some Z, X is dependent on Z, and Z is dependent on Y.
(43)
X is an antecedentdef of Y if Y is linked to X or, for some Z, Y is linked to Z and X is an antecedent of Z.
114
Christopher Kennedy
This relation is irreflexive: an element must never be construed as dependent on itself. Higginbotham formalizes this condition as the following constraint on LF representations: (44)
*dependent(X,X)
This condition is designed to prohibit referential circularity: it should not be the case that some element’s interpretation must be known in order to determine its interpretation. (45) is a typical example of referential circularity.13 (45) *[DP His wife] loves [DP her husband] By (42), the pronoun his is dependent on the pronoun her, which is contained in the antecedent of his.14 At the same time, the pronoun her is dependent on his, which is contained in the antecedent of her. By transitivity each pronoun is dependent on itself, in violation of (44). In addition, (44) rules out so-called i-within-i violations (Chomsky 1981), which are exemplified by (46).
(46)
*[DP Every picture of its frame] was taken on Wednesday.
The antecedent of the pronoun its is the DP containing it, therefore this pronoun is dependent on itself, in violation of (44). In Higginbotham’s system, in which anaphoric relations are uniformly represented via linking, (44) is a satisfactory definition of dependence. In a system such as Reinhart/Heim’s that assumes two types of anaphoric relations – bound variable anaphora and coreference – giving a specific statement of dependence becomes somewhat more complex. Some constraint against referential circularity is required in order to rule out sentences like (45) and (46); the question is, how should this relation be formulated? Let us begin with the assumption that Higginbotham’s (42) can be directly applied to Heim’s system without modification. Recall that X is linked to Y if X’s inner index equals Y’s outer index. With antecedence defined as in Higginbotham (1983), the LF of (46) will be ruled uninterpretable in Heim’s system for the same reason that it is uninterpretable in Higginbotham’s system: the pronoun is both linked to and contained in its antecedent, as shown in (47).
Argument Contained Ellipsis
(47)
115
[DP every picture of its2 frame]:2 was taken on Wednesday.
What about (45)? In Heim’s system, a sentence like (45) may have an LF as in (45), in which no DP bears an outer index (these are optional). (48)
[his1 wife]3 loves [her3 husband]1
His wife and her are colinked, therefore coreferential; her husband and his are colinked, therefore coreferential. Because dependency is defined in terms of antecedence, and antecedence is defined in terms of linking, according to (42), the pronouns are dependent neither on these DPs nor on anything contained within them, and (48) should not involve a violation of (44). This is clearly an undesirable result, as the intuition is very strong that (48) is uninterpretable because it involves referential circularity. How should this intuition be formally stated in a system that assumes both linking and colinking? Clearly, my initial assumption that Higginbotham’s linking directly parallels Heim’s notion of linking must be modified. This is understandable since Higginbotham’s linking represents both binding and coreference, whereas Heim’s linking represents only binding. The fact that Heim’s system represents anaphoric relations in two ways – via linking and via colinking – suggests a solution to the question raised above: if there are two types of anaphoric relations, then it follows that there are two types of ‘‘referential’’ dependence.15 In the following paragraphs I will present a reformulation of dependence in terms of Heim (1993). I propose that a system of anaphoric relations that assumes both binding and coreference must in turn assume two types of dependence. The first, which I will call r-dependence, arises in colinked structures, i.e., in cases of discourse-controlled coreference. R-dependence is defined in (49) (following Higginbotham, I assume the ‘‘contained in’’ relation to be reflexive): (49)
r-dependence X (X has no descriptive content) is r-dependentdef on Y if: i. for some Z, X and Z are colinked and Y is contained in Z, or ii. X is r-dependent on Z and Z is r-dependent on Y.
The second type of dependence arises in binding relations. Intuitively, this type of dependence is a relation between a bound variable and the restriction of its binder: if the domain over which the variable may range is determined by the restriction, then the possible value of the variable is dependent on (the elements within) the restriction. I will call this type of dependence b-dependence, and define it as in (50):
116
(50)
Christopher Kennedy
b-dependence X (X has no descriptive content) is b-dependentdef on Y if Y is contained in the binder of X.
The crucial question is whether both of these relations, like Higginbotham’s statement of dependence, are irreflexive. We can answer this question by reexamining the two examples of dependence violations given above. First consider (51), which was not ruled out by (44). (51)
[his1 wife]3 loves [her3 husband]1
In (51), the pronoun his is colinked with the DP her husband. According to (49i), his is dependent on both this DP and the pronoun her, which is contained in the DP. Similarly, her, which is colinked with his wife, is dependent on his. By (49ii), each pronoun is dependent on itself. The fact that (51) is ungrammatical indicates that r-dependence is irreflexive. We may now turn to linking structures. In (52) the pronoun is linked to the DP that contains it. (52)
[DP every picture of its2 frame]:2 was taken on Wednesday]
According to (50), because the pronoun its is contained in its binder, it is b-dependent on itself. In general terms, the pronoun is part of the restriction of its binder, which, as discussed above, is a configuration that must be ruled out. (52) and other i-within-i violations are straightforwardly explained if b-dependence, like r-dependence, is irreflexive. These conclusions allow us to reformulate Higginbotham’s (44) into the following two constraints, which rule out circular structures at LF: (53)
a. *r-dependent(X,X) b. *b-dependent(X,X)
(53a) and (53b) can be implemented in the grammar as constraints on syntactic indexing configurations. These constraints reject Logical Forms that contain indexing configurations that correspond to reflexive dependence relations, with respect to the definitions stated in (49) and (50). 5.4.2
Dependence relations within DP
An immediate worry, pointed out by Ha ¨ık (1987), is that (something like) (53b) is too strong a constraint. Crucially, it must not rule out structures such as (54a), which are discussed in Ha ¨ık (1987), or cases of ordinary relative clauses, such as (54b).
Argument Contained Ellipsis
(54)
117
a. [Every man near his2 Dart]:2 looks happy. b. [Every man who cleans his2 Dart]:2 is happy.
(54a) and (54b) should be contrasted with the ungrammatical structures (55a) and (55b), both of which involve b-dependence violations. (55)
a. *[every picture of its2 frame]:2 was taken on Wednesday. b. *[every proof that it2 is correct]:2 is invalid.
Ha ¨ık (1987) observes that the difference in grammaticality between the examples in (54) on the one hand and those in (55) on the other follows directly from structural differences between the grammatical and ungrammatical examples. Let us assume the structure of DP to be as in (56) (ignoring bar-level categories).16 (56)
DP D
NP N
(XP)
The examples in (55) contain selected complements of N – a PP in (55a) and a CP in (55b) – which occupy the structural position of XP in (56). In either case, regardless of whether NP or DP is the binder of the pronoun in the selected complement, the pronoun is contained in its binder and therefore b-dependent on itself, in violation of (53b). In the examples in (54), however, the phrase containing the bound pronoun is an unselected modifier. Assuming that modifiers are adjoined to the phrases they modify, they do not occur in the position of XP in (56). The adjunction prohibition (Chomsky 1986a) prohibits adjunction to an argument, therefore the PP in (54a) and the CP in (54b) must be adjoined to the NP complement of D, which is the complement of a functional head and not an argument. That is, the PP and CP occupy the position represented by YP in the DP structure (57). (57)
DP D
NP NP N
YP
118
Christopher Kennedy
In (57), YP is contained in DP, so if DP binds an element within YP, the structure is ruled out by (53b). YP is not, however, contained in the lower NP. If this NP is the binder of an element within the adjoined YP, (57) does not incur a violation of (53b): the bound element is not contained in its binder, so it is not dependent on itself.17 To illustrate, the structure of the subject in (54a) is given in (58): (58)
DP:2 D every
NP NP2:2 man
PP P
DP
near
his2 Dart
The pronoun his is bound by the circled NP – not by DP, so this structure does not violate (53b). The analysis of relative clauses is similar to that of PP modifiers, but some details must be discussed, as an account of binding relations within relative clauses will be of crucial importance to the analysis of the Argument Contained Ellipsis structures. The (abbreviated) structure of the subject relative clause in (54b) is shown in (59) below. In (59), the relative operator bears an inner and outer index. This structure formally represents the binding relation that, according to Safir (1986), holds between the head of the relative clause and the relative operator. The outer index binds the traces within the CP; the inner index is linked to the head of the relative clause, which c-commands the relative operator and therefore binds it, according to the assumptions made above. (59) does not violate (53b): nothing in the relative clause is bound by an element that contains it, therefore nothing in the relative clause is b-dependent on itself. The analysis of b-dependence presented here and its relation to i-within-i condition violations builds on that of Ha ¨ık (1987), who notices that referential circularity is blocked if an element which has the same interpretation as the XP that contains it is bound within that XP. Ha ¨ık formalizes this observation as the following principle (Ha ¨ık 1987: 506):
Argument Contained Ellipsis
(59)
119
DP:2 D every
NP NP2:2 man
CP who2:2
IP t2
(60)
VP V
DP
cleans
his2 Dart
Principle of Referential Dependency In [XP . . . a . . .], if a referentially depends upon some antecedent A, then XP is annotated as referentially dependent upon A, unless a is bound inside XP.
Ha ¨ık assumes a constraint on circularity identical to Higginbotham’s (44); the combination of this constraint and (60) rules out the ungrammatical structures discussed above, but allows the grammatical sentences in (54). The crucial difference between the analysis presented here and Ha ¨ık’s analysis is that, as noted in Ha ¨ık (1985), (60) is a descriptive constraint – in a system that assumes a single referential dependence relation (as in Higginbotham 1983 and Ha ¨ık 1987), it does not follow that binding within XP should save structures like those in (54) from ill-formedness with respect to (44). In the current analysis, which distinguishes between dependence in binding relations and dependence in coreference relations, the wellformedness of the examples in (54) is expected. Because the anaphoric elements in these structures are bound, the relevant dependence relation is b-dependence. As illustrated by the structural representations (58) and (59), the pronouns are not contained within their binders, so are not b-dependent on themselves. The same cannot be said of the sentences in (55), in which, as shown in (56), the pronouns are contained within their binders. The contrast between the examples in (55) on the one hand and (54) on the other is straightforwardly explained. To summarize, this section claimed that within the system of anaphoric relations developed in Heim (1993), the general notion of ‘‘referential’’ dependence should be stated both in terms of coreference (colinking), and in terms of binding (linking). These two types of dependence were defined
120
Christopher Kennedy
as r-dependence and b-dependence, respectively. By assuming that these relations are irreflexive, and irreflexivity is enforced in the form of constraints on indexing configurations at LF, we were able to account for the uninterpretability of a variety of circular constructions. In addition, it was shown that the grammaticality of a set of structures that are superficially similar to reflexive b-dependence structures, which were previously accounted for by Ha ¨ık (1987) with a descriptive constraint, follows directly from structural characteristics and the definition of b-dependence. 5.5
Relative clauses
A generalization common to the grammatical examples of Argument Contained Ellipsis presented in section 5.2.1 (the ones in which argument identity obtains) is that the VP contained within the argument of the other VP is contained in a relative clause. Before moving to an analysis of ACE, I will first lay out some assumptions regarding the syntax and semantics of relative clauses. The structure given in (59), repeated below with more syntactic detail, can be used to illustrate the discussion. (61)
DP:2 D every
NP NP2:2 man
CP who2:2
C′ IP
C t2
I′ VP
I t2
V′ V
DP
cleans
his2 Dart
In his discussion of the relation between dependence and crossover effects, Higginbotham (1983: 409) defines a formal variable as ‘‘an empty category
Argument Contained Ellipsis
121
that occurs in an argument position and is linked to a nonargument; the operator to which a formal variable is linked will be called its binder.’’ I will assume that this definition is directly transposable into the Heim’s model of anaphoric relations with the following modifications: a formal variable is an empty category that is linked to a c-commanding DP at LF. This modification reflects the fact that I assume QR to be optional, and ensures that if a trace in SpecVP is linked to a subject in SpecIP, it is a formal variable. In (61), both the trace of the relative operator in SpecIP and the internal subject trace in SpecVP are formal variables, as both are linked to the operator in SpecCP. Higginbotham later notes that ‘‘the semantics of relatives justifies regarding [the trace of the moved subject and the trace of the relative operator] as distinct occurrences of the same formal variable’’ (1983: 414). Applying this statement to the LF shown in (62), we can conclude that the VP internal traces (circled) are ‘‘distinct occurrences of the same formal variable.’’ (62)
IP
DP2:2
I′ I
every
VP
NP t2 NP2:2 man
CP who2:2
IP t2
I′ I
VP t2
V′ V
DP
cleans
his2 Dart
V′ V
AP
is
happy
122
Christopher Kennedy
What is implicit in Higginbotham’s analysis is that each distinct occurrence of the same formal variable is bound by a distinct occurrence of the same binder. In other words, the subject DP in (62) and the relative operator within this DP are distinct occurrences of the same binder, since the variables they bind are distinct occurrences of the same formal variable.18 The crucial point is that because the relative operator and the DP in a structure like (62) can be considered distinct occurrences of the same binder, we may conclude that the distinct occurrences of a particular formal variable in this and similar examples are cobound. This relation will play a crucial role in the analysis of Argument Contained Ellipsis to be presented in the next section. 5.6
Argument Contained Ellipsis reanalyzed
5.6.1
External arguments
The data presented in section 5.2.1 were the basis for the generalization stated in (64a). (63)
Argument Contained Ellipsis Ellipsis between VP1 and VP2, VP1 contained in an argument A2 of VP2, is licensed only if A2 is identical to the parallel argument A1 of VP1.
(64a) and (64b) exemplify this paradigm in the subcase of subject containment. (64)
a. Every man who said he would buy some salmon did. b. *Every man who said George would buy some salmon did.
(64b) is ungrammatical because the subjects of the related VPs are not identical. In the following paragraphs, I will show that the interaction of elliptical dependence, as implemented in section 5.3, with referential dependence, as described in 5.4, forces argument identity in ACE structures. In section 5.3, I showed that the strict identity hypothesis requires the subjects of elliptically related VPs to bear identical outer indices. In the proposed model, the LFs of (64a) and (64b) are (65a) and (65b), respectively (omitting irrelevant details). (65)
a. [every man who1:1 t1 said he1 would [VP t1 buy some salmon]]:1 did [VP t1 buy some salmon] b. [every man who1:1 t1 said George3:1 would [VP t1 buy some salmon]]:1 did [VP t1 buy some salmon]
Argument Contained Ellipsis
123
Both (65a) and (65b) appear to be well-formed LFs – in both examples, each argument is part of a chain that receives a theta role, satisfying Full Interpretation, and the VPs are strictly identical, in line with (31). What, then, is the crucial difference between (65a) and (65b) that makes the latter ungrammatical? I claim that (65b), but not (65a), violates the LF well-formedness constraint (53b) (*b-dependent(X,X)). According to the characterization of elliptical dependence that I have argued for in this chapter, every syntactic entity in the LF representation of an elided VP is treated by the grammar as an occurrence of the syntactic entity that occupies the parallel position of an antecedent VP. The importance of this claim with respect to the data under discussion is that all VP internal traces in the LF of an elided VP are to be considered identical to their counterparts in the antecedent’s LF. For example, in (65b), the subject traces in the embedded and matrix VPs are to be considered as two occurrences of the same linguistic entity. The claim is that the relation between the traces in an elided VP and the traces in its antecedent is similar to but subtly different from the relation between the traces in the relative clause discussed in section 5.5: whereas the latter are distinct occurrences of the same formal variable, I claim that the former are identical occurrences of the same formal variable. With these assumptions in mind, we may now examine the b-dependence relations in the LFs (65a) and (65b). In both cases the trace in the specifier of the matrix VP is bound by the subject. By the definition of b-dependence, this trace is b-dependent on everything that is contained in the subject. It follows that the trace in the specifier of the matrix VP is dependent on the trace in the specifier of the embedded VP. According to the proposed characterization of elliptical dependence, these traces are identical, therefore the trace in the specifier of the matrix VP is b-dependent on itself, in violation of (53b). By this logic, however, both LFs should violate (53b). In both cases, the trace in the matrix SpecVP is b-dependent on itself (the identical occurrence in the embedded SpecVP). To understand how identity of arguments ensures that (65a) is grammatical despite its apparent ill-formedness with respect to (53b), we must return again to the discussion of formal variables. Recall that in section 5.5 I suggested that if the trace of a moved DP with a relative clause modifier and the trace of the relative operator were to be considered distinct occurrences of the same formal variable, following Higginbotham (1983), then the binders of these variables should be considered distinct occurrences of the same binder, and the formal variables themselves should be considered to be cobound. This means that all occurrences of t1 in (65a) – including the identical traces in the matrix and embedded SpecVP – are cobound. This is not true of (65b), because the
124
Christopher Kennedy
trace in the embedded SpecVP is within the scope of a closer potential binder, the DP George. In other words, the occurrences of the identical traces in (65a) are cobound, but the occurrences of the identical traces in (65b) are not cobound. This is the crucial difference between structures in which argument identity obtains and structures in which it does not. To understand how this difference ensures the well-formedness of examples manifesting argument identity, it is necessary to consider the basic idea behind b-dependence. B-dependence states that ‘‘X (X has no descriptive content) is b-dependent on Y if Y is contained in the binder of X.’’ The processes that generate LFs from S-structures, which are outlined in Heim (1993) and which I have assumed in this chapter, create structures in which all potential syntactic binders (i.e. DPs with outer indices at S-structure) become operators at LF (see Higginbotham 1983). B-dependence is a formal statement of the relation between the variables that are bound in an operator’s nuclear scope (by that operator) and the elements in the operator’s restriction. It is the operator’s restriction that determines the domain over which the variables in the nuclear scope may range; therefore, the bound variables in the nuclear scope are dependent for their interpretations on their binder’s restriction. The restriction set must be established before the nuclear scope can be evaluated; similarly, it should never be necessary to know the value of some variable in the nuclear scope in order to establish the restriction. It follows that b-dependence must be irreflexive, and the data discussed in section 5.4.1 support this conclusion. The role of b-dependence in regulating the syntax–semantics interface can be most clearly stated in terms of indices (which, in Heim 1993, are what ultimately get translated as variables). B-dependence states that (the semantic translation of) any given LF must be such that the values of all indices contained in the binder of some index i can be established before the value of i itself is determined. Therefore, i must not itself be contained in the binder of i, otherwise it would be necessary to know its value before its value could be determined, resulting in circularity. If it were the case, however, that i occurred both in an operator’s restriction and in its nuclear scope and the value of both occurrences of i were set simultaneously, then, in effect, no b-dependence relation would hold between the two occurrences of i. In other words, if both occurrences of i were cobound, it would not be the case that the value of one occurrence of i would have to be ascertained in order to compute the value of the other occurrence of i. Neither would be dependent on the other. As noted above, this type of configuration is precisely what occurs in the LF of the grammatical sentence (65a), but does not occur in the LF of the
Argument Contained Ellipsis
125
ungrammatical (65b), both of which are repeated below. (In each example, the relevant traces are in boxes.) (66)
a. [every man who 1:1 t1 said he1 would [ VP t1 buy some salmon]]:1 did [VP t1 buy some salmon] b. [every man who 1:1 t1 said George3:1 would [ VP t1 buy some salmon]] :1 did [VP t1 buy some salmon]
The crucial difference between the two structures is that in (66a), the trace in the matrix SpecVP is bound by the QRed DP and the trace in the embedded SpecVP is bound by the relative operator, therefore both occurrences of t1 are bound by (distinct occurrences of) the same binder: they are cobound. In (66b) however, t1 in the matrix SpecVP is bound by the subject of the matrix clause, while t1 in the embedded SpecVP is bound by the DP George; these DPs are different binders. Because the identical occurrences of t1 in the restriction and nuclear scope in (66a) are cobound, no b-dependence relation holds between them. If no b-dependence relation holds between the traces, then (66a) is vacuously well-formed with respect to the LF constraint (53b). (66b), in which there is a b-dependence relation between the identical traces, violates (53b). The analysis of Wasow’s sentence (1), repeated below, is exactly the same as the analysis of (66b). (67)
*A proof that God exists does.
According to the strict identity hypothesis, the LF of (67) must be as in (68): (68)
[DP a proof that God2:1 [VP t1 exists]]3:1 does [VP t1 exist]
(68) is ill-formed for the same reason as (66b): the trace in the matrix VP is b-dependent on an identical occurrence of that trace in the subject-contained VP, in violation of (53b). The difference between Wasow’s example and e.g. (66b) is that for (1), there is no corresponding grammatical sentence in which subject identity holds. The explanation of this fact parallels the discussion of i-within-i violations in section 5.4.1 (see (45) and (46)). In sentences like (64a) and (64b), the argument-contained VP is inside a relative clause modifier of the subject. Binding of the subject of this VP by the relative operator prevents a reflexive b-dependence relation from obtaining within the subject DP. In (1), however, the argument-contained VP is in a selected complement of N0. In order for the reflexive b-dependence relation between the VP internal traces to be eliminated, the subject of the contained VP must itself be bound by the subject of the matrix VP. This configuration would be
126
Christopher Kennedy
ill-formed with respect to (53b), however, because the embedded subject would be contained in its binder, the matrix subject. This structure is shown in (69). (69)
[DP a proof that it1 [VPt1 is correct]]2:1 [VPt1 is correct]
Although the b-dependence relation between the identical traces has been eliminated, a new one has been created: the subject of the embedded VP it1 is contained in its binder, the subject of the matrix VP. Therefore it1 is b-dependent on itself, in violation of (53b).
5.6.2
Internal arguments
The other cases of Argument Contained Ellipsis can be analyzed in exactly the same manner as the subject cases. The other configurations discussed were examples of Antecedent Contained Deletion (ACD) structures. Examples like those in (70) involve sentences in which one of the VPs is contained in an internal argument of the other; (71) shows sentences in which one of the elliptically related VPs is contained in the external argument of an embedded VP (see section 5.2.1.2). (70)
a.
Polly visited every town Erik did.
b. *Polly visited every town in every country Erik did. (71)
a.
Sydney believes every proposal Evelyn does to be insane.
b. *Sydney believes every aspect of every proposal Evelyn does to be insane. Following May (1985), I assume that in order for an ACD structure to be interpreted, the phrase containing the elided VP must undergo QR at LF (in order to eliminate the antecedent containment configuration). The LFs corresponding to the examples above are given in (72) and (73).19 (72)
a. [DP every town Op1:1 Erik did [VP visit t1]]:1 [IP Polly [VP visited t1 ]] b. [DP every country Op1:1 Eric did [VP visit t1]]:1 [DP every town in t1 ]:1 [IP Polly [VP visited t1]]
(73)
a. [DP every proposal Op1:1 that Evelyn does [VP believe [IP t1 to be insane]]]:1 [IP Sydney [VP believes [IP t1 to be insane]]] b. [DP every proposal Op1:1 that Evelyn does [VP believe [IP t1 to be insane]]]:1 [DP every aspect of t1 ]:1 [IP Sydney [VP believes [IP t1 to be insane]]]
Argument Contained Ellipsis
127
In each example, the DP containing the surface null VP has been QRd, leaving a trace which is contained in the antecedent VP. The resulting LFs are the structures with respect to which the identity relation described in (31) must hold. These must be evaluated for a potential b-dependence relation between the trace in the LF representation of the elided VP and the trace in the antecedent VP. These examples contain exactly the same configuration as the examples of subject containment discussed above: a trace in the matrix VP is b-dependent on a trace in a VP contained in a relative clause. According to the strict identity hypothesis, these traces are identical, so these structures should violate (53b). The explanation of the contrast between the examples in which argument identity holds and those in which it does not parallels the explanation of the subject identity examples. In the grammatical sentences (72a) and (73a), the relative operator is linked to the DP that has undergone QR, therefore, according to the assumptions made above, both occurrences of the identical trace (italicized) are bound by distinct occurrences of the same binder. As in the subject examples discussed above, if the identical traces are cobound, no b-dependence relation holds between them. The LF constraint (53b) does not apply to (72a) and (73a). In (72b) and (73b), however, the relative operator is not linked to the DP that has undergone QR, but to a different DP contained within the QRd argument. In these cases, it cannot be said that the relative operator and the QRd DP are distinct occurrences of the same binder; they are completely different binders. The identical traces in these examples are not cobound, therefore the trace in the antecedent VP is b-dependent on the trace within the elided VP. Because these traces are identical, (72b) and (73b) violate (53b). 5.6.3
Summary
To summarize, I have argued that structural configurations in which one of two elliptically related VPs is contained in an argument of the other (what I have called Argument Contained Ellipsis) violate the LF wellformedness constraint *b-dependent(X,X). This constraint regulates the syntax – semantics interface by ruling out circular constructions on the basis of syntactic indexing configurations. If, as I have argued, variables (inner indices on traces) in an elided VP’s LF are actually identical to the parallel variables in an antecedent VP’s LF, then Argument Contained Ellipsis configurations will inevitably violate this constraint, because one of the identical variables will be contained in the binder of the other. This
128
Christopher Kennedy
relation is eliminated, however, just in case both occurrences of the identical variable are cobound. Basing my argument on Higginbotham’s (1983) discussion of relative clauses, I showed that a cobinding relation will hold when the parallel arguments of the related VPs are identical. If the b-dependence relation between the identical variables is eliminated, then the structure will vacuously satisfy *b-dependent(X,X). The interaction of the revised theory of referential dependence presented in section 5.4 and a theory of elliptical dependence, as implemented by the strict identity hypothesis, thus forces argument identity in cases of Argument Contained Ellipsis. 5.7
Elliptical dependence and alphabetic variance
Two basic claims of this chapter are that (53b) reflects a general semantic constraint against referential circularity and that sentences that contain indexing configurations that would lead to a violation of this constraint are ruled out at LF. Argument Contained Ellipsis structures in which the argument identity constraint described in (63) does not hold have been argued to contain such ill-formed LFs. The analysis of ACE presented in this chapter has implications beyond the explanation of a puzzling set of data, however. In particular, it may provide a principled explanation of a well-established but nevertheless stipulative constraint on VP identity in elliptical structures. The basic structure of the LF of an example of ACE is shown in (74), where the occurrences of t1 are identical: (74)
[DP . . . [VP1 . . . ti . . . ] . . . ]:i [ . . . [VP2 . . . ti . . . ] . . . ]
(74) is the LF syntactic representation of a tripartite quantification structure in the semantic component (see section 5.5, note 7). The hypothesis advanced in this chapter is that an elliptical relation between VP1 and VP2 is licensed in these configurations only if the identical occurrences of the VP internal variable are cobound; otherwise the LF is ruled out by (53b) as a circular structure. This constraint has a precedent in the literature on VP Ellipsis. In Sag (1976), two VPs may stand in an elliptical relation iff the lambdaexpressions that correspond to their logical forms are alphabetic variants. There are conditions on this relation, though, one of which is stated in (75) (Sag 1976: 104). (75)
For two l-expressions lx.A and ly.B to be alphabetic variants, if some operator Op outside lx.A binds a variable in A, then the corresponding variable in ly.B must also be bound by Op.
Argument Contained Ellipsis
129
Arguably, (74) – and therefore all cases of ACE – is a configuration that fits the structural description in (75). The VPs correspond to lx.A and ly.B, the VP internal traces correspond to variables within the lambda expressions, and DP corresponds to D. Therefore, according to (75), if the trace in the matrix VP is bound by DP, the trace in the embedded VP must also be bound by DP. This will be true just in case argument identity holds (cf. the discussion of relative clauses in section 5.5). If ACE structures are construed in this way, then in a model of ellipsis along the lines of Sag (1976), the argument identity constraint follows directly from (75). In structures in which argument identity holds the related VPs are alphabetic variants; in structures in which argument identity does not hold they are not alphabetic variants. Although (75) gets the right results, it is a stipulation. This constraint states that in configurations such as (74), the variables in the related VPs must be cobound. It does not, however, explain why. Moreover, there is reason to believe that this constraint is too strong. On the one hand, (75) correctly rules out ellipsis in the second sentence in (76): (76)
Who did Martin see? *Who did Bill?
It is not clear that (75) is the correct explanation of such sentences, however.20 Jacobson (1992) presents a number of sentences that are problematic for the alphabetic variance approach (see also examples (30a)–(30b) in section 5.2.1.3; examples (77b) and (77c) are from Evans 1988): (77)
a. I know which student Al likes, and I know which student Mary doesn’t. b. Bagels, I like. Donuts, I don’t. c. John was supposed to read several books this semester. But the only one that he actually did was The Brothers Karamazov.
According to (75), each of these examples should be ungrammatical because each sentence contains variables in elliptically related VPs that are bound by distinct operators. This is clearly the wrong result. The important question raised by the examples in (77) is whether (75) is active only in syntactic configurations that correspond to tripartite structures in the semantics. If so, then (75) takes on the status of (63), the original description of the conditions licensing Argument Contained Ellipsis. As with (63), we have the intuition that if (75) is true only of certain configurations, it must reflect some deeper fact about ellipsis. If the analysis of ACE can be generalized to all tripartite quantification
130
Christopher Kennedy
structures (where one VP is in the restriction and the other in the nuclear scope), then the analysis developed in this chapter claims that the fact reflected by (75) is referential circularity. (75) need not be stated, except as a descriptive law of VP Ellipsis, because the requirement that the variables in the VPs in structures like (74) must be cobound follows from the interaction of elliptical dependence and a very general claim about the syntax – semantics interface: the binding relations represented in the syntax by indexing configurations must be such that they do not give rise to referential circularity in the semantics. 5.8
Concluding remarks
This chapter has argued that a straightforward account of the argument identity constraint that holds in Argument Contained Ellipsis structures follows from the interaction of two types of dependence relations: elliptical dependence, which characterizes the relation between an elided VP and the VP from which it receives its interpretation, and referential dependence, which characterizes the relation between a nominal anaphoric element and its antecedent. Several important results came out of the discussion. First, the chapter supports the theory of anaphoric relations developed in Heim (1993). Heim (1993: 3) notes that her revision of Reinhart (1983a) may share with Reinhart the ‘‘wrong descriptive generalizations’’ about ellipsis. Although I have avoided discussion of some of the problems associated with syntactic accounts of ellipsis (several of which are discussed in Wescoat (1989) and Dalrymple (1991), I have shown that Heim’s system allows us to implement an explicit representation of elliptical dependence via the strict identity hypothesis, and thereby give a precise account of a complex array of facts. Second, this work relativizes the concept of referential dependence developed in Higginbotham (1983) and Ha ¨ık (1987) to a system of anaphoric relations in which binding and coreference are represented separately. This relativization of dependence was achieved by dividing it into the r-dependence and b-dependence, which are active in coreference and binding relations, respectively. The formulation of b-dependence as a distinct dependence relation provided a direct explanation of the grammaticality of certain types of apparent i-within-i violations, which were previously accounted for stipulatively, and was of crucial importance to the analysis of ACE. Future work should be directed towards exploring these relations (e.g. with respect to Bach – Peters sentences). In particular, it should be ascertained at what level, if at all, r-dependence and b-dependence interact.
Argument Contained Ellipsis
131
Finally, this work makes a fundamental claim about ellipsis, namely that the relation between an elided VP and the VP from which it receives its interpretation, which I have referred to as elliptical dependence, should be directly encoded in a theory of ellipsis. I suggested that this relation can be implemented in a syntactic model of ellipsis by the strict identity hypothesis, which requires elliptically related VPs to be identical in terms of indexical or variable value at Logical Form, then showed that the interaction of elliptical dependence and referential dependence accounts for the Argument Contained Ellipsis facts.
6
Variable island repair under ellipsis Jason Merchant University of Chicago
One of the most startling, and hence theoretically challenging, properties of wh-movement in Sluicing is that it can move wh-phrases out of islands, an important observation which goes back to Ross (1969). Equally challenging is the fact that similar wh-movement out of VP Ellipsis sites remains for the most part illicit. Briefly put, it seems that for a wide range of cases, deletion of an IP containing an island voids the effect of that island for wh-movement, while deletion of a VP containing an island does not. This chapter investigates one aspect of this puzzling dichotomy with respect to island repair, and attempts to show that an interesting and partly novel range of data follow if island deviancies come about due to illicit traces of intermediate movement, working in tandem with a constraint on ellipsis operative in structures that host wh-movement. I will argue that a wide range of islands are indeed active at PF, but not in the way that this claim has usually been understood thus far. Instead of the island node itself being responsible for the degradation in acceptability, I will show that the data support the idea that it is the traces of wh-movement outside the island itself which trigger a PF-crash. I begin with some relevant Background on Sluicing, given in section 6.1, before taking up the Sluicing data in section 6.2 and the VP Ellipsis facts in section 6.3. 6.1
Background on Sluicing
Sluicing is ellipsis of the sentential complement to an interrogative complementizer hosting a wh-phrase, as in (1): (1)
a. Jack bought something, but I don’t know what. b. A: Someone called. B: Really? Who? c. Beth was there, but you’ll never guess who else. d. Jack called, but I don’t know when/how/why/where from.
132
Variable island repair under ellipsis
133
e. Sally’s out hunting – guess what! f. A car is parked on the lawn – find out whose. These structures have been extensively investigated in the literature (the most detailed studies being Ross 1969; Levin 1982; Riemsdijk 1982; Chao 1987; Chung et al. 1995; Lobeck 1995; Romero 1998; Lasnik 2001; and Merchant 2001), and the most common approach to generating these structures is to assume that the wh-phrase has been moved to specCP out of the missing IP by the usual mechanisms of wh-movement in questions, and that the IP has undergone deletion (ellipsis). While neither of these two components has been universally accepted, I will follow most recent work in assuming that these are the two operations that generate the sentences in (1) (see Merchant 2001 for extensive justification). The relevant derivation is given schematically in (2). (2)
CP C′
XP[+wh] C0[+wh,+Q]
IP . . . t. . .
Given this analysis, one question that immediately arises is the question of what licenses the ellipsis of IP (in the sense of licensing developed in Lobeck 1991). Lobeck (1995) claims that the null IP (a base-generated empty category in her approach) is licensed only by the null [þwh, –pred] C0 of interrogatives (following Rizzi’s 1990 typology). I will here recast her approach as a featural matching requirement in a head-head (or feature of a head) relation. Assume that PF deletion is triggered by the presence of a feature on a head. Let us call this triggering feature E. Ideally, E will have exactly those syntactic, phonological, and semantic effects that yield all the attested properties of the elliptical construction at hand, with nothing further needing to be said. The syntax of E must encode the checking requirements, in order to capture Lobeck’s licensing. In Frampton and Gutmann’s (1999) notation, the syntactic featural makeup of E is [[þwh], [þQ] ], that is, an element which must be checked by a þwh, þQ head. Such checking heads are limited in languages like English at least to the complementizers that occur in constituent questions, as desired. (Whether E is freely available to be merged with any head, or whether it must be on C or on I is an independent
134
Jason Merchant
question: presumably locality considerations will rule out this featural combination on E from occurring anywhere where it cannot be checked by a þwh, þQ C0. This restricts E to C or I. If E is on C, the effects follow directly. If it is on I, then we must posit head-to-head or feature-to-head movement to check E. For present purposes, the decision between these two options is immaterial.) The phonology of E, in broad terms, should be such that E instructs PF not to parse its complement. Assuming E to be located on C at the relevant point in the derivation (when the syntactic structure is parsed by the mechanisms operative at PF), we can view E to be something like a suprasegmental, but with the unusual effect of parsing its complement IP into a prosodically unrealized category. In SPE terms, E has the effect of triggering a context sensitive rule of the following sort, however we may wish to conceptualize this in more recent phonological theories: (3)
½=IP ! ø=E
PF-‘‘deletion,’’ in this view, is the result of a feature in the syntax, not of a freely operating ‘‘deletion transformation.’’ Finally, the semantics of E must capture the traditional identification of the ellipsis site, ideally encoding all and only those requirements that regulate under what conditions an XP can be deleted. Two general approaches to this question have been pursued: one, that the elided XP must be syntactically (LF-structurally, in current formulations) isomorphic to an antecedent, and two, that the elided XP must be semantically equivalent to an antecedent. Both views have weaknesses: generally, the syntactic isomorphism approach undergenerates, while the semantic identity approach overgenerates. I will follow here the approach advocated in Merchant (2001), where the semantic identity condition is defined as e-givenness: an XP a can be deleted only if a is e-given. Once we have defined such a predicate that applies to (the meanings of) XPs, we can use it to define a semantics for the E feature. (By a similar token, we could easily define a predicate over LF structures and impose a syntactic isomorphism condition in the same way, assuming presumably a structured meaning approach.) The semantics of E will therefore be a partial identity function on propositions, a semantic filter which allows the composition to proceed only if its argument is e-given. In the notation of Heim and Kratzer (1998), the semantics of E is the following: (4)
[[E]] ¼ lp: p is e-given. p
A sluice like (5), then, will have the structure in (6). (5)
Abby was reading, but I don’t know what.
Variable island repair under ellipsis
(6)
135
but I don’t know
CP C′
DP2 what
C[E]
IP Abby was reading t2
At the relevant point in the computation, E will take IP as its argument, as in (7). This expression will return a value (namely Abby was reading x2) only if Abby was reading x2 is e-given, otherwise it returns no value. Computation up the tree therefore proceeds only if E’s IP complement is e-given, as desired. (7)
[[E]]([[IP]]) ¼ lp: p is e-given. p(Abby was reading x2)
The greatest advantage of using E, encoded as a partial identity function, to impose the identity requirement is that it localizes ellipsis identification, and allows us to dispense with the more usual formulations of the requirement on ellipsis which essentially postulate a separate ‘‘ellipsis module’’ in the grammar (i.e. a global, late, well-formedness condition imposed just on the structures containing ellipsis) parallel to the Binding Theory module (cf. Giannakidou’s 1998; 2001 elimination of a ‘‘polarity’’ module by encoding polarity requirements as local, lexical semantic well-formedness conditions, using type-combinatorics). A second important advantage is that the licensing (the local featural requirements of E) and identification (the semantic condition E imposes on its complement) requirements on ellipsis can be directly linked. Most theories posit no direct link between these requirements at all. In sum, I will assume that a unified theory of PF-deletion based on semantic identity is possible (indeed, desirable), and that Sluicing instantiates PF-deletion of an IP out of which wh-movement has occurred. 6.2
Sluicing and wh-extraction out of islands
Assuming that wh-movement of the usual kind occurs in Sluicing brings us directly to the puzzle that has inspired much of the work on the topic since Ross (1969) first discovered it: the wh-movement found in Sluicing (at least with certain kinds of correlates) is insensitive to syntactic islands (see Baker and Brame 1972; Chomsky 1972a; Lakoff 1972; Chung et al. 1995; Lasnik 2001; Merchant 2001).1 Examples for the major kinds of
136
Jason Merchant
syntactic islands are given (occasionally with nonelliptical controls) in (8)–(16).2 (8)
Relative Clause island: a. They want to hire someone who speaks a Balkan language, but I don’t remember which. b. *I don’t remember which (Balkan language) they want to hire someone [who speaks ].
(9)
Left-branch (attributive adjective case): a. She bought a big car, but I don’t know how big. b. *I don’t know how big she bought [a car].
(10)
Derived position islands (subjects, topicalizations): a. A biography of one of the Marx brothers is going to be published this year guess which! b. *Guess which (Marx brother) [a biography of ] is going to be published this year.
(11)
COMP-trace effects: (Chung et al. 1995 (90), (91a); Perlmutter 1971: 112)) a. It appears that someone will resign, but it’s not yet clear who. b. Sally asked if somebody was going to fail Syntax One, but I can’t remember who.
(12)
Coordinate Structure Constraint: a. They persuaded Kennedy and some other Senator to jointly sponsor the legislation, but I can’t remember which one. (Chung et al. 1995: (88b)) b. Bob ate dinner and saw a movie that night, but he didn’t say which.
(13)
Adjuncts: a. Ben will be mad if Abby talks to one of the teachers, but she couldn’t remember which. b. *Ben will be mad if Abby talks to one of the teachers, but she couldn’t remember which (of the teachers) Ben will be mad [if she talks to ]. c. Ben left the party because one of the guests insulted him, but he wouldn’t tell me which.
(14)
Complement to nouns: (Chung et al. 1995: (84c)) The administration has issued a statement that it is willing to meet with one of the student groups, but I’m not sure which one.
Variable island repair under ellipsis
137
(15)
Sentential subject: (Chung et al. 1995: (84b)) That certain countries would vote against the resolution has been widely reported, but I’m not sure which ones.
(16)
Embedded question: (Chung et al. 1995: (84a)) Sandy was trying to work out which students would be able to solve a certain problem, but she wouldn’t tell us which one.
In (8a), for example, the wh-phrase which has moved out of the relative clause, interpretationally parallel to its unelided but ungrammatical counterpart in (8b). Similar remarks apply to the remaining islands. One possibility for accounting for the fact that deletion of the island rescues the sluice from ungrammaticality is to posit that the PF interface cannot parse crossed island nodes. One way of formalizing this, following in essence Chomsky (1972b), is to assume that crossed island nodes are marked with some PF-uninterpretable feature, call it *. This general approach, ‘‘* as a feature of island nodes,’’ has been pursued in one form or another by Ross (1969), Lasnik (2001), and Kennedy and Merchant (2000) (the latter for the Left Branch Condition only). For the example in (8a), repeated in (17a) on the facing page with the structure in (17b), the account works as follows. Wh-movement (cyclic or otherwise – I suppress here for simplicity possible intermediate traces) extracts the DP which from its base position (marked by t1), moving it to the highest specCP. In doing so, the relative clause island is crossed. As such, it is marked with a *. In (17b), the island node is assumed to be the CP adjoined to the NP, and this CP is marked with *. In nonelliptical cases, when this *CP reaches PF, it will cause a PF-crash, since * is by hypothesis PF uninterpretable. Under this formulation, ellipsis will have the desired effect: deletion of the boxed IP in (17b) eliminates the *CP as well, preventing the * from triggering a PF-crash. The structure is therefore saved, and surfaces as the grammatical sluice in (17a). This general solution applies mutatis mutandis to the other islands in (9)–(16) as well.
6.3
VP Ellipsis and wh-extraction out of islands
As appealing as the solution sketched above is, it faces a serious problem when we turn our attention to parallel extractions out of elided VPs. In the same environments, with the same correlates and the same islands, extraction out of a VP Ellipsis site is no more grammatical than in nonelliptical cases. The relevant data from VP Ellipsis, contrasting with the sluices in (8)–(16) above, are given in (18)–(25).
138
(17)
Jason Merchant
a. They want to hire someone who speaks a Balkan language, but I don’t remember which. b. IP-Deletion eliminates *CP
...CP C′
[DP which]1 C
IP I
they I0
VP want
IP I′
PRO to
VP hire
NP NP
*CP
someone who speaks t1 (18)
Relative Clause island: *They want to hire someone who speaks a Balkan language, but I don’t remember which they do.
(19)
Left-branch (attributive adjective case): *She bought a big car, but I don’t know how big she did.
(20)
Derived position islands (subjects, topicalizations): *He said that a biography of one of the Marx brothers is going to be published this year – guess which he did!
(21)
COMP-trace effects: a. *It appears that someone will resign; it’s just not clear how it does. b. *Sally asked if somebody was going to fail Syntax One, but I can’t remember who she did.
Variable island repair under ellipsis
139
(22)
Coordinate Structure Constraint: a. *They persuaded Kennedy and some other Senator to jointly sponsor the legislation, but I can’t remember which one they did. b. *Bob ate dinner and saw a movie that night, but he didn’t say which he did.
(23)
Adjuncts: a. *Ben will be mad if Abby talks to one of the teachers, but she couldn’t remember which he will. b. *Ben left the party because one of the guests insulted him, but he wouldn’t tell me which he did.
(24)
Complement to nouns: *The administration has issued a statement that it is willing to meet with one of the student groups, but I’m not sure which one it has.
(25)
Embedded question: *Sandy was trying to work out which students would be able to solve a certain problem, but she wouldn’t tell us which one she was.
In fact, the problem of wh-extracting out of VP Ellipsis sites is even more severe, as highlighted in Lasnik (2001), where the data in (26)–(28) are given. Movement of a wh-phrase whose correlate is an indefinite out of an elided VP is degraded even when no island is involved. (26)
They said they heard about a Balkan language, but I don’t know a. which they said they heard about. No Ellipsis b. which. c. *which they did.
(27)
(28)
Sluicing VP Ellipsis
They attended a lecture on a Balkan language, but I don’t know a. which they attended a lecture about. No Ellipsis b. which. Sluicing c. *which they did.
VP Ellipsis
They studied a Balkan language, but I don’t know a. which they studied.
No Ellipsis
b. which. c. ??which they did.
Sluicing VP Ellipsis
140
Jason Merchant
One possibility for accounting for this range of data, including the data in (18)–(25), would be to claim simply that there is a general ban on whextraction out of VP Ellipsis sites, a` la Sag (1976) and Williams (1977). Unfortunately, such a claim is too strong, as the following examples demonstrate. (29)
a. I know what I LIKE and what I DON’T. b. I know which books she READ, and which she DIDN’T. c. What VP Ellipsis CAN do, and what it CAN’T. (Johnson 2001)
(30)
a. GREEK, you should take; DUTCH, you shouldn’t. b. I know which books ABBY read, and which ones BEN did.
(31)
a. I think YOU should ride the TALLEST camel, but I don’t know which one PHIL should. (Schuyler 2001: (48)) b. I think you SHOULD adopt one of these puppies, but I can’t predict which one you actually WILL. (Schuyler 2001: (49)) c. ABBY took GREEK, but I don’t know what language BEN did. d. We know that Abby DOES speak [Greek, Albanian, and Serbian]F – we need to find out which languages she DOESN’T <speak t>! (Merchant 2001: 115 fn. 5 (ii)) e. (I know) ABBY wants to take GREEK, but I don’t know what language BEN does <want to take>. f. ABBY1 said she1 took GREEK, but I don’t remember what language BETH2 did <say she2 took>.3 g. ABBY attended a lecture on KEATS, but I don’t know what poet BEN did.
What distinguishes the examples in (29)–(31) from (18)–(25), (26c), (27c), and (28c) is the presence in the former of an element in the elliptical clause which contrasts with some element in the antecedent clause. The observation is simple: it appears that some kind of contrast is required in the cases where VP Ellipsis is licit (see Schuyler 2001 for one formulation). When such contrast is absent, as in (18)–(28), VP Ellipsis is disallowed. Perhaps, as Merchant (2001) and Lasnik (2001) suggest, there is a ban on eliding less than possible under wh-extraction (whose ultimate source
Variable island repair under ellipsis
141
remains obscure). For the present, let us capture this ban in the form of an inviolable constraint, MaxElide (this constraint may be in part derivable from economy, since putting the E feature higher in a given structure allows for less pronunciation). Roughly put, it states that if ellipsis applies in a structure with a wh-trace, ellipsis should target the largest constituent possible. More accurately, it requires that if ellipsis targets an XP containing an A0 -trace, XP must not be properly contained in any YP that is a possible target for deletion. (32)
MaxElide [Definition] Let XP be an elided constituent containing an A0 -trace. Let YP be a possible target for deletion. YP must not properly contain XP (XP ¸ YP ).
The VP Ellipsis in the (c) examples of (26)–(28) violates MaxElide; in (26c), for example, the VP <say they heard about t> contains a wh-trace and the VP is properly contained in the IP , which, as (26b) shows, is itself a possible target for deletion. The possibility for deleting the containing IP, then, blocks deletion of any contained VP (likewise for the more deeply embedded VP: *They said they heard about a Balkan language, but I don’t know which they said they did.). Parallel reasoning applies to the examples in (18)–(25), in which the deleted VP happens to contain an island: in each case no contrasting material is present, and the IP containing the deleted VP is a possible target for deletion (as witnessed by the Sluicing counterparts in (8)–(16)). The examples in (29)–(31) differ in precisely this regard. Consider (29a): the deleted VP is , which is properly contained in the IP . But this containing IP is not a possible target for deletion: there is no antecedent which would license deletion of the sentential negation (technically, in the theory assumed here, the IP is not e-given). Hence the containing IP is irrelevant to the deletion of the VP, MaxElide doesn’t apply, and VP Ellipsis is not blocked in this case. In other words, if the material outside the VP Ellipsis site contrasts in some way with the antecedent clause, the contrasting material cannot be deleted (since it is not e-given), and hence no larger constituent will be a possible target for deletion. This contrasting material can be in the auxiliary domain (negation as in (29), (30a), or modals as in (31b)), or the subject (as in (30b), (31a, c–f)), or elsewhere external to the VP but internal to the IP (see Schuyler 2001 for further examples and discussion). MaxElide seems also to be responsible for the unexpected oddity of examples like (33b,d) (from Merchant 2001: 58). While VP Ellipsis targeting the highest VP is fine, as in (33a), and while ellipsis is not required, as in
142
Jason Merchant
(33c), ellipsis targeting either the most deeply embedded IP (33b) or VP (33d) is distinctly degraded. Both these latter examples are in violation of MaxElide: taking XP in (33b) to be the matrix VP whose subject is Charlie, MaxElide is violated by the ellipsis of the embedded IP, properly contained in the matrix VP. Likewise in (33d): MaxElide prohibits deletion of the embedded VP since the matrix VP is a possible target of deletion itself. (33)
a. Ben knows who she invited, but Charlie doesn’t. b. ??Ben knows who she invited, but Charlie doesn’t know who. c. Ben knows who she invited, but Charlie doesn’t know who she invited. d. ??Ben knows who she invited, but Charlie doesn’t know who she did.
Similar reasoning applies to the example in (34), from Williams (1986), where the possible deletion of the higher IP blocks deletion of the lower one: (34)
John knows how to do something, but I don’t know what (*he knows how).
It is crucial to note that MaxElide applies only to XPs that contain a wh-trace, since no similar blocking effect is found in the absence of wh-movement: (35)
a. Ben knows that she invited Klaus, but her father doesn’t. b. Ben knows that she invited Klaus, but her father doesn’t know that she did.
The contrast between XPs containing wh-movement and those that do not appears in the data in (36) as well, from Merchant (2001: 58 note 9): (36)
a. Abby knew that he had quit, but Beth didn’t know that he had. b. Abby asked if he had quit, but Beth didn’t ask if he had. c. ?? Abby knew when he had quit, but Beth didn’t know when he had. d. ?? Abby asked when he had quit, but Beth didn’t ask when he had.
The fact that MaxElide applies only to XPs containing A0 -traces also allows us to set aside the possible objection raised in Lasnik (2001) on the basis of an example with subject extraction: as he points out, the possibility for Sluicing in (37a) does not preclude the VP Ellipsis variant in (37b).
Variable island repair under ellipsis
(37)
143
a. Someone solved the problem. b. i. Who? ii. Who did?
The reason that Sluicing doesn’t block VP Ellipsis here follows from the definition of MaxElide. In particular, in (37b) MaxElide has no provenance, since the elided VP does not contain a wh-trace, assuming the structure in (38) with the type of the two traces of the subject notated as superscripts. (38)
0
[CP who [IP tA [VP tA solved the problem]]]
Constraints similar to MaxElide are discussed in Hirschbuhler (1978) ¨ and Tancredi (1992: 123) for a related set of data; nevertheless, at this point, a satisfying theoretical reduction of MaxElide remains elusive. For present purposes, I will be concerned only with the obvious effects it has. At this point, we have constructed a coherent account of some surprising differences between Sluicing and VP Ellipsis, and it might appear that the ungrammaticality of the examples in (18)–(25) with which this section began do not, after all, pose a difficulty for the *-as-a-feature-of-islandnodes view that accounted so elegantly for the lack of island effects in Sluicing in section 6.2. Unfortunately, this view makes the following prediction: if the island node is internal to a deleted VP and if MaxElide is satisfied (by the presence of contrasting material external to the VP and internal to the IP), then we should find that ellipsis has the same islandameliorating effect that we find in Sluicing. This prediction is incorrect, as the following examples show (see also the examples and discussion in Merchant (2001: 114–115). (39)
Relative Clause island: *Abby DOES want to hire someone who speaks GREEK/a certain Balkan language, but I don’t remember what kind of language she DOESN’T.
(40)
Left-branch (attributive adjective case): *ABBY bought a big car, but I don’t know how big BEN did.
(41)
Derived Position islands (subjects, topicalizations) *Abby DID say that a biography of HARPO is going to be published this year – guess which Marx brother she DIDN’T.
(42)
COMP-trace effects: a. *It appears to ME that SEN. HATCH will resign, but I don’t know which senator it does to YOU. b. *ABBY asked if I was going to fail Syntax One, but I can’t remember who BEN did.
144
Jason Merchant
(43)
Coordinate Structure Constraint: a. *They got the president and thirty-seven Democratic Senators to agree to revise the budget, but I can’t remember how many Republican ones they DIDN’T. b. BOB ate dinner and saw five movies that night, but he didn’t say how many ABBY did.
(44)
Adjuncts: a. *BEN will be mad if Abby talks to Mr. Ryberg, and guess who CHUCK will. (Merchant 2001: 115 (15)) b. *BEN left the party because Charlene/some guest insulted him, but God only knows which guest ABBY did.
(45)
Complement to nouns: *The dean’s office has issued a statement that it is willing to meet with Students for a Democratic Society/a certain student group, but I’m not sure which student group the provost’s office has.
(46)
Embedded question: *Sandy was trying to work out how many students would be able to solve problem #4/a certain problem, but she wouldn’t tell us which problem she wasn’t.
To see how the account fails on these examples, consider (39). The presence of negation in the IP beneath the wh-phrase means that MaxElide is satisfied (vacuously). The island node, as in (17b) above, is the *CP of the relative clause internal to the deleted VP <want to hire someone who speaks t>. Hence the deletion of the VP prevents the *CP from reaching the PF-interface, and this example should have the same status as its Sluicing counterpart in (17a), contrary to fact. Instead of thinking of * as a feature of island-nodes, I suggest we think of * as a feature of traces. In particular, let intermediate traces of islandescaping XPs be marked with the * feature. One way of implementing this is to say that each link in a chain of wh-movement must be licensed either by locality or by being in a spec-head relation with a C (or perhaps simply by being pronounced). As each new copy of an XP is generated (via ‘‘re-merge,’’ i.e. Move), the relevant locality restrictions are checked (subjacency, etc.) and if locality is not respected (e.g. if an island node is crossed), the new copy is given the feature * (compare the g-marking of Lasnik and Saito 1984; 1992 and Chomsky and Lasnik’s 1993 *-marking; see Kitahara 1999 for an alternative viewpoint). All later copies of this *XP will themselves also be *-marked. Finally, I assume that the *-feature can be erased
Variable island repair under ellipsis
145
(checked) in the final spec-head relation that a þwh XP comes to be in with a þwh C (perhaps * should in fact be thought of as feature of [wh], in some conceptions); in other words, the [þwh, þQ] C0 that licenses the movement of a wh-XP to its specifier also checks the *-feature on that copy, eliminating the * from the PF-representation of the highest copy. (Perhaps even the E feature itself checks the *-feature.) Under this conception, a standard island effect will come about whenever intermediate *-traces survive until PF. I assume, following Chomsky (1986a), Fox (2000), and Lo´pez and Winkler (2003) (see also Koster 1978 for a precursor), that wh-movement proceeds by adjunction to intervening maximal projections (VPs and IPs at the least; other intermediate landing sites will not be crucial here). I illustrate the workings of this system with the schematic derivation of (47) in (48), giving only the steps of interest here: (47)
*What language do they want to hire someone who speaks?
(48)
1 Merge wh-DP [what language]: speaks [what language] 2 Move DP out of relative clause, marking it with *: [*what language] [CP who speaks [what language] ] 3 Move DP to intermediate landing site, adjoined to matrix VP: [*what language] [VP want to hire someone [*what language] [CP who speaks [what language]]]. 4 Move DP to intermediate landing site, adjoined to matrix IP: [*what language] [IP they [*what language] [VP want to hire someone [*what language] [CP who speaks [what language] ] ] ]. 5 Move DP to specCP, erasing * on final (namely, the highest) copy of DP: [CP [what language] [IP [*what language] [IP they [*what language] [VP want to hire someone [*what language] [ who speaks [what language] ] ] ] ] ].
In the final representation, all the intermediate traces between the CP island node and the final landing site are *-marked. Since the *-feature is by hypothesis PF-uninterpretable, the final structure will cause a PF-crash, yielding the ungrammaticality of (47) as desired. Conceiving of the *-feature as a feature of traces, however, allows us to make the necessary distinction between IP and VP Ellipsis. In a Sluicing example like (49a), deletion of the highest IP eliminates all *-traces from the PF-representation, yielding the attested amelioration of island effects under Sluicing. In VP Ellipsis, on the other hand, the elided VP does not
146
Jason Merchant
Table 6.1. Overview of the data with respect to MAX ELIDE and *t Examples Sluicing
(8)–(16) (26b), (27b), (28b)
VP Ellipsis
*(18)–(25) *(26c), (27c), (28c) (29)–(31) *(39)–(46)
MaxElide
*t
p p
p p
p p
p p *
Island-containing Non-island-containing Island-containing Non-island-containing Non-island-containing Island-containing
contain all the offending traces: in particular, the trace of adjunction to the highest IP (*t00 2 in (49b)) remains in the PF-representation, causing a crash just as in (47). (49)
a. They want to hire someone who speaks a Balkan language, but I don’t remember which. b. ...CP
C′
[DP which]2
IP
C t′′2
IP I′
they do
VP t2′
VP want to hire [NPsomeone[who speaks t2]]
The proposed system, consisting of MaxElide and of * as a feature of traces, captures a wide range of data, both in Sluicing and VP Ellipsis, and both in structures containing islands and those without. All the relevant data presented here is tabulated by example number in table 6.1. The top two rows include all the Sluicing data examined. These examples satisfy MaxElide and no *-traces survive to PF (indicated by check marks in the columns labeled
Variable island repair under ellipsis
147
MaxElide and *t respectively). The VP Ellipsis examples fall into four groups, given in the final four rows of table 6.1. The only grammatical examples are those in (29)–(31), which satisfy both MaxElide and do not present any *-traces. All other kinds of examples fail on one or both counts, as indicated by the * in the respective column. Interestingly, the kinds of examples which have been taken as most indicative of the inability of VP Ellipsis to repair islands, namely (18)–(25), are doubly bad: not only do their island-violations survive the deletion, but they also violate MaxElide. Notice that this account locates the deviancy of the island-containing VP Ellipsis examples in the clause that hosts the wh-movement. An intermediate *-trace of the successive cyclic wh-movement survives VP Ellipsis, but not IP Ellipsis. Wh-movement in the elliptical clause in both cases is successive cyclic. In this regard, the account thus differs from Fox and Lasnik (2003), who account for the difference between Sluicing and VP Ellipsis with respect to island repair by assuming that successive cyclic wh-movement occurs only in the VP Ellipsis case, but not in the Sluicing case; they trace the different status of the examples to this posited difference, and its effects on the resulting structures for satisfying an LF-isomorphism constraint (in the cases they examine, the correlates are all indefinites, which Fox and Lasnik assume are interpreted as choice functions and do not move). Up to this point, the correlate to the moved wh-phrase has played no role. In Sluicing, we have seen mostly examples where the correlate was an indefinite, though it is known that other kinds of correlates are possible. Both names and quantifiers can be correlates in Sluicing (see Chung et al. 1995 and Romero 1998 for relevant discussion): (50)
a. Abby speaks GREEK, but I don’t remember what OTHER languages. b. She met RINGO, but I don’t know who else. c. He said he talked to ABBY, but I don’t know who else he said he talked to. d. John met most applicants, but I can’t remember exactly which ones.
The same holds for VP Ellipsis, as the examples in (31) above demonstrated, three of which are repeated here in (51a–c): (51)
a. ABBY took GREEK, but I don’t know what language BEN did. b. We know that Abby DOES speak [Greek, Albanian, and Serbian]F – we need to find out which languages she DOESN’T! speak! (Merchant 2001:115 fn. 5 (ii))
148
Jason Merchant
c. (I know) ABBY wants to take GREEK, but I don’t know what language BEN does want to take. d. ABBY interviewed two-thirds of the applicants, but I don’t remember exactly how many of them BEN did interview. This state of affairs is expected under the present account, since the wh-movement in the elliptical clause violates no islands, and the semantic identity condition based on e-givenness is satisfied (since e-givenness is defined using F-closure, the focused correlates will be replaced by variables; see Merchant 2001: 35–37). The data can also be accommodated on Fox and Lasnik’s account, assuming that the focused correlates scope at LF. For them, the antecedent clause in e.g. (50a) and (51a) must have at least the following structure (assuming scoping to IP, and an intermediate landing site at VP): GREEK lx[IP Abby [VP x lx0 [speaks x0 ] ] ]. This structure will license, via their posited LF-identity requirement, deletion of the IP or VP as needed, assuming that the wh-movement in the elliptical clauses is likewise successive cyclic. Now we are in a position to appreciate the puzzle that arises with the examples in (52). In these cases, the focus correlate occurs inside an island, and the resulting sluice is ungrammatical. We have already seen that this holds for VP Ellipsis in some of the sentences in (39)–(46) above, two of which are repeated here in (53). (52)
a. *Abby wants to hire someone who speaks GREEK, but I don’t remember what OTHER languages she wants to hire someone who speaks. b. *The radio played a song that RINGO wrote, but I don’t know who else.
(53)
a. *Abby DOES want to hire someone who speaks GREEK, but I don’t remember what kind of language she DOESN’T. b. *BEN will be mad if Abby talks to Mr. RYBERG, and guess who CHUCK will. (Merchant 2001:115 (15))
In short, Sluicing with indefinite correlates repairs islands, but Sluicing with focused correlates does not. Given the account presented so far, this contrast is puzzling. We cannot assimilate the deviance of the sluices in (52) to that of the VP Ellipsis cases in (53), since only in the latter does the *-trace remain after ellipsis has applied. Nor does the Fox and Lasnik (2003) account fare any better: their account, like the present one, is based on the idea that *XPs (intermediate VPs and IPs, for example) are PF-uninterpretable (which is why PF-deletion repairs the islands). In their
Variable island repair under ellipsis
149
presentation, ‘‘avoiding an intermediate landing site . . . yield[s] an island violation . . . [unless] the island is deleted’’ (p. 12); I take this to be a member of the *-as-a-feature-of-island-nodes family of analyses. Consider now (52a): the correlate must scope out of the island in one fell swoop: GREEK lx[Abby wants to hire someone who speaks x]. This structure licenses the deletion of the IP beneath what OTHER languages, which has similarly undergone non-successive-cyclic wh-movement, by hypothesis. Under their account, therefore, we again expect the sluices in (52) to be grammatical. These sluices’ ungrammaticality cannot be assimilated to that of the VP Ellipsis examples in (53) on Fox and Lasnik’s account either (longdistance wh-movement in the latter is ruled out by virtue of the surviving *IP in the elliptical clause). Alternatively, Fox and Lasnik could assume that the focus movement must be successive-cyclic, while the movement in the Sluicing cases is not; in this case, the Sluicing examples would be ruled out (as violations of LF-parallelism), but the VP Ellipsis cases should then be grammatical. One possibility that would rescue both the present account and Fox and Lasnik’s is that focus movement is island-sensitive for other reasons. If so, then the movement needed to generate the LF structures (on Fox and Lasnik’s account) or to provide the semantic antecedent (on the e-givenness account) would be illicit in any case. If this were so, the examples in (52) and (53) would be ruled out because ellipsis parallelism could not be satisfied (and the examples in (53) redundantly also because of the *t in the elliptical clause). The problem with this idea is the usual assumption that focus is not, in fact, island respecting (Chomsky 1972b; Rooth 1985; Kratzer 1991), whether this is accomplished by island-insensitive scopal movement or in situ. Kratzer (1991a) in particular provides examples involving VP Ellipsis which seem to indicate that a focused item can be ‘scoped out’ of an antecedent VP for purposes of ellipsis as well. (54)
I only talked to the woman who chaired the ZONING BOARD because you did.
As Kratzer points out, (54) has a reading as follows: the only x such that I talked to the woman that chaired x because you talked to the woman who chaired x is the zoning board. This reading necessitates island-violating scoping of the focused element zoning board in order to allow the bound reading in the elided VP. The example in (54) suffers from a slight defect, however: since the pitch accent falls on the final DP in the relative clause, it is difficult for some speakers to distinguish this from focus on the entire DP the woman who chaired the zoning board (cf. Drubig 1994 and Winkler 1996). Kratzer’s claim can be seen perhaps more clearly in the following
150
Jason Merchant
example, where no focus percolation from within the relative clause to the containing DP is possible: (55)
I only played a song that RINGO wrote because you did.
This example, parallel to (54), has a reading that can be paraphrased as follows: the only x such that I played a song that x wrote because you played a song that x wrote is Ringo. These latter facts are unsurprising under the view of islands that emerges from the study of Sluicing especially: islands are essentially PF phenomena, so the movement necessary for the focus in (54) and (55), since it has no PF consequences, is not expected to result in island violations (see Rooth, 1996). Likewise for wh-in-situ inside islands, if these require longdistance movement (see Simpson 2000 for a recent overview). We seem to have reached an impasse: the evidence from ‘contrast’ sluices in (52) indicates that focus movement is island-sensitive, while the data in (54) and (55) seems to indicate that it is not. But this is, luckily, not the only difference between (52) and (54)–(55): in the former, but not the latter, we also have an instance of wh-movement out of the ellipsis site. This wh-movement has scopal properties of its own, and I would suggest we can capitalize on these properties to rule out (52) while allowing (54)–(55). Wh-movement out of an ellipsis site forces its correlate to take scope over the entire antecedent clause, just as the whphrase itself does in its own clause (see Chung et al. 1995 and Romero 1998 for extensive analysis of this fact; see Fox 2000 for general discussion of scopal parallelism in ellipsis). All the data examined so far indicate that when focus movement extracts a focused XP out of an island, the focus movement can no longer target the highest clause node, but is limited to the VP. Metaphorically speaking, it is as though escaping from an island cripples or hobbles further focus movement; it can only limp along up to VP, not to IP. I thus suggest that the differences seen above be traced to these differing possibilities for satisfying the identity condition on ellipsis. In the non-island cases, focus movement of the correlate in the antecedent can scope the focused XP to IP, as we saw above (yielding an LFparallel structure or a structure which satisfies e-givenness, assuming that the existential binder is inserted at the locus of the moved focused XP). For these cases, the antecedent after focus movement will be the following (for (50a); mutatis mutandis for (51a)): (56)
GREEKF lx[ Abby [ x lx0 [speaks x0 ] ] ]
This structure has the correlate in a parallel position (clause-external) to the moved wh-phrase in the elliptical clause, satisfying parallelism:
Variable island repair under ellipsis
(57)
151
what OTHER languages lx[IP Abby [VP x lx0 [speaks x0 ] ] ]
Consider now the example in (55). Here, focus movement must scope RINGO to the highest VP (above the VP-adjoined adjunct because clause), satisfying the parallelism requirement for the deletion of the VP in the because clause: (58)
I only *RINGOF lx[VP [VP played a song that x wrote] because you did] play a song that x wrote]
The movement in (58), though island-violating, targets the matrix VP, not IP. Finally, in the puzzling cases of (52) and (53), the stipulation that islandescaping focus movement cannot target the highest IP will prevent the correlate from attaining the necessary scopal parallelism with the whphrase (clause-external), and hence these clauses can never satisfy the identity requirement needed to license deletion. The highest the focus movement can go is VP, yielding the following as the antecedent clause, by hypothesis (intermediate traces suppressed): (59)
[ IP Abby [VP *GREEKF lx [VP wants to hire someone who speaks x ] ] ]
A structure like (59), unlike (56), does not have a scopal element in a position parallel to the moved wh-phrase, and hence will license neither deletion of IP nor of VP. These considerations will not apply in cases where the correlate is an indefinite, since indefinites are known to be able to freely take wide scope, even out of islands (see Farkas 1981). Hence such indefinites will always be able to provide scopally parallel antecedents for wh-phrases moved out of ellipsis sites, as Chung et al. (1995) discuss (whether or not these indefinites are interpreted as choice functions or not). So scopal parallelism is satisfied in the Sluicing cases we began with in (8)–(16) as well as in the VP Ellipsis cases in (8)–(25) and those with indefinites in (39)–(46). The latter, therefore, still necessitate the theory of islands developed above, since the focusbased restriction that rules out (53) will not apply to those cases with indefinite correlates. A related point comes from correlates that are interpreted as generalized quantifiers, forcing them to scope via QR, known to be even more local than merely island-respecting (though not quite, as often assumed, clausebound: see Farkas and Giannakidou 1995). These provide possible antecedents for Sluicing and VP Ellipsis, as in (50d) and (51d) above (repeated here as (60a,b) respectively). (60)
a. Abby met most applicants, but I can’t remember exactly which ones.
152
Jason Merchant
b. ABBY interviewed exactly two thirds of the applicants, but I don’t remember how many of them BEN did. The grammaticality of these examples is expected, since movement of the correlates via QR will provide appropriate antecedents for the elliptical clauses. This holds both for an e-givenness-based account and an LFisomorphism account of Fox and Lasnik (2003). The fact that such correlates are island-bound is also expected, given the local nature of QR: (61)
a. *If most senators resign, Abby will stop her hunger strike, but I can’t remember exactly which ones. b. *If exactly two thirds of the senators resign, ABBY will stop protesting, but I can’t remember how many of them BEN will.
But notice again that something more must be said to rule out examples like (62), which differs from (60b) only in not having a contrasting element in the clause hosting the VP Ellipsis. (62)
?? Abby met most applicants, but I can’t remember exactly which ones she did.
Recall that Fox and Lasnik rule out examples similar to this one in which however the correlate is an indefinite – such as (28c) above – by positing that the indefinite, unlike the wh-phrase, does not move, and hence the elliptical clause does not satisfy LF-parallelism (since it, unlike the antecedent clause, contains traces of successive cyclic movement, by hypothesis). But as we have just seen, when the correlate is a quantified or focused DP, not an indefinite, successive cyclic movement must be posited, in order to rule in (60b) and (51a–c). The current account, employing MaxElide, rules out (62) and (28c) on a par, while one based on the postulated presence vs. absence of successive cyclic movement apparently must be supplemented by MaxElide in any case. 6.4
Conclusions
This chapter has investigated a number of surprising asymmetries in island repair between Sluicing (IP Ellipsis) and VP Ellipsis, and has argued that these fall out from taking certain island effects to be due to ill-formed intermediate traces at the PF interface. The conclusion, then, is that a number of deviancies that have been ascribed to other parts of the grammar (derivational constraints, LF output constraints) may best be located at the PF interface (at least part of the effects of island-violating extraction). Prima facie paradoxically, it seems that the nature of PF and the constraints that operate there can be illuminated by investigating structures that have no PF exponence.
Variable island repair under ellipsis
153
Certain elements of the analysis presented here remain at present theoretically unsatisfying, in that the effects encoded in the various constraints (especially MaxElide and the restriction on island-escaping focus movement) have yet to be reduced in an insightful way to the theoretical primitives they presumably derive from. This project is ongoing, and I mention here a number of other areas in which repair effects of ellipsis seem to be indicated: 1 lack of complementizer agreement in Bavarian Sluicing 2 lack of Wackernagel clitics in S. Slavic Sluicing 3 multiple Sluicing in Germanic, Greek, and Turkish (and perhaps in Bulgarian, Japanese, Russian, and Serbo-Croatian as well) 4 remnant movements in Gapping (Johnson 2003, Richards 1998) 5 remnant movements in Pseudogapping (Johnson 2001) 6 lack of verb movement in Pseudogapping (Lasnik 1995, 2001) 7 swiping in English, Norwegian, Danish (Merchant 2002) 8 ‘‘vehicle change’’ effects in anti-pronominal contexts (Potts 1999) 9 long-distance reflexives in English (Kennedy and Lidz 2001) 10 wh-movement in wh-in-situ languages 11 lack of I-to-C movement in matrix sluices in Germanic (Lasnik 1999b and Merchant 2001) 12 lack of the otherwise obligatory complementizer in Irish sluices (Merchant 2001). This potpourri of effects has emerged mostly recently from investigating elliptical structures from the perspective of repair, and it is not surprising that they have yet to be made to follow from primitives of any theory. It appears that we have just begun to uncover a new domain that may help us shed light on phenomena that have been traditionally investigated only with respect to their pronounced manifestations.
7
On binding scope and ellipsis scope Winfried Lechner University of Athens
7.1
Introduction
In contexts of Antecedent Contained Deletion (ACD) such as (1), the size of VP Ellipsis (VPE) and the semantic scope of the quantifier containing the silent node are systematically related.1 Simplifying somewhat,2 the quantifier hosting the ellipsis site has to take scope at least as high as the node containing the antecedent. For (1), this generalization rules out the wide ellipsis reading (1b) if the object (some play) is construed de dicto, because such a combination would require the hosting quantifier to take scope below the ellipsis node. (1)
John wanted to read some play Sam did D. a. D ¼ read b. *D ¼ wanted to read (relative to de dicto reading for the object).
The correlation between noun phrase interpretation and ellipsis in (1) represents a strong argument for the hypothesis that ellipsis resolution and quantifier scope are both encoded at a level of representation that also incorporates information about the syntactic organization of the expression (LF). Similar conclusions have been drawn from the interaction of ellipsis scope with binding scope, which tracks admissible coreference relations of NPs inside the elliptical clause (see e.g. Sag 1976; Larson and May 1990; Fox 1999b, 2000). For instance, the r-expression John in (2) can be construed coreferentially with the preceding pronoun only if ACD is resolved with wide ellipsis, as in (2b): (2)
I expected himj to buy everything that Johnj thought I did D. a. *D ¼ bought b. D ¼ expected himj to buy
(Fox, 1999b: 187 (74)) Acknowledgments: I am grateful to Kyle Johnson, Chris Kennedy, Jason Merchant, Cecile Meyer, Alan Munn, Maribel Romero, Wolfgang Sternefeld, Arnim Von Stechow, Uli Sauerland, Hubert Truckenbrodt, Chris Wilder, and the audiences at the University of Pennsylvania, University of Utrecht, and Zentrum fur ¨ Allgemeine Sprachwissenschaft, Berlin, for suggestions and comments.
154
On binding scope and ellipsis scope
155
In this chapter, I will present further evidence indicating that ellipsis and scope are closely linked by considering the behavior of a related group of elliptical constructions, viz. phrasal comparatives (PCs; see (3)). Descriptively, PCs are comparative clauses in which the comparative complement (than-XP) only contains a single, nominal remnant (Sam in (3)): (3)
John wanted to write more plays [than-clause than Samremnant D].
Although PCs resemble ACD in many respects, the former possess interesting properties that set them sufficiently apart from the latter in order to permit qualitatively new insights into the relation between ellipsis, scope, and binding. In particular, it will be seen below that relocating the focus of attention from ACD to PCs entails that the triad (4a), which defines properties of VPE, is substituted by the triple in (4b), which marks corresponding characteristics of elliptical comparatives. (4)
a. Quantifier scope – Size of VP Ellipsis – binding scope b. Scope of than-XP extraposition – Comparative Ellipsis – binding scope
This change in turn makes it possible to establish three results, two of which are of immediate relevance for the theory of ellipsis more generally. First, it will be seen that PCs and ACD differ in that in PCs, the correspondence between ellipsis size and the scope of the category hosting the silent node can be described by a biconditional, whereas the implication holds only in one direction in contexts of ACD (section 7.2.1). This contrast will be attributed to differences in the types of ellipses involved in the two constructions. In addition, the operation marking the scope of ellipsis can be observed in overt syntax only in PCs. Thus, PCs elicit further evidence strengthening the hypothesis gained from the analysis of ACD that the size of ellipsis and the scope of the containing categories correlate. The specific analysis for the formation of PCs to be advocated will moreover be shown to support a new empirical generalization about syntactic restrictions on referentially transparent readings of verbal predicates (section 7.2.2). Second, I will present data demonstrating that ellipsis scope and binding scope in PCs are systematically related, emulating closely the pattern that can be observed in ACD (section 7.3.1). However, the evidence in support of this view will be seen to manifest itself even more transparently in comparatives. This is so as in PCs, the diagnostics for binding scope do not test for properties of NPs embedded inside a relative clause, as in ACD (see (2)), but can be directly applied to the remnant. As a consequence, it becomes possible to control for the potentially aberrant influence of
156
Winfried Lechner
relative clauses (anti-reconstruction effects and counter-cyclic merger; see Fox 2002; Lebeaux 1990; among others). Third, the distribution of possible coreference dependencies between NPs inside the elliptical than-XP and potential antecedents in the matrix clause vindicates a specific perspective on the derivation of PCs, according to which PCs are not base-generated, but contain empty elliptical structure inside the than-XP that is accessible to the principles of Binding Theory. In this way, PCs provide additional evidence that ellipsis is structurally encoded at the same abstract level – LF – which also serves as the basis for the computation of referential dependencies (section 7.3.2).
7.2
The size of ellipsis in PCs
7.2.1
Comparatives and Conjunctions
In Lechner (1999; 2001; 2004), it is argued that the syntactic processes responsible for phonological reduction in comparatives are defined by the same set of restrictions which characterize ellipsis operations that target conjunctions (Gapping, Right Node Raising, and ATBmovement). As a result, the effects of Comparative Ellipsis (Bresnan 1975) can be subsumed under independent operations, and need no longer be listed separately in the grammar. What is of particular significance for present concerns is the fact that this hypothesis also leads to a new strategy for generating PCs, which treats them as Gapped clausal comparatives. The surface string in (5a) can for instance be assigned the underlying representation in (5b): (5)
a. John wrote more plays than Sam. b. [IP [IP-First conjunct John wrote [DegP more plays tk]] [than-XP than [IP-Second conjunct Sam wroteGappingd-many playsCD]]k c. [IP John wrote [DegP more plays [than-XP than [IP Sam wrote d-many playsCD]]]]
The configurational constraints on Gapping are usually held to include the requirement that the antecedent clause and the elliptical clause form part of a coordinate structure. But in the example above, the than-XP, which contains the Gap, originates in a subordinate position, inside the degree phrase (DegP) of (5c) (see Abney 1987; Corver 1990). In order to resolve this conflict, it will be assumed that comparatives can satisfy the coordination criterion by extraposition of the than-XP (henceforth than-XP Raising or TR), resulting in a derived comparative coordination (for details see Lechner 2004). Thus, in an initial step, (5b) is generated from
On binding scope and ellipsis scope
157
the underlying representation (5c) by TR of the than-XP to the right periphery of the clause. The derived comparative coordination (5b), in which than mimics the syntax of a coordinator,3 is then sufficiently similar to regular coordinate structures in order to supply a suitable context for the application of Gapping, yielding the PC (5a). (In addition, the surface appearance of comparatives is shaped by Comparative Deletion, signaled by the subscript CD in (5); see Bresnan 1973, among others.) On this conception, Gapped comparatives and Gapped coordinate structures are expected to behave alike in all respects. But, as will be explicated below, the two constructions also differ in important aspects, the identification of which will subsequently prepare the way for a refinement of the analysis. The relevant asymmetry is empirically reflected by the varying degree to which comparatives and conjunctions, respectively, observe a syntactic condition on the structural relation between an elided string and its antecedent. This Isomorphism requirement demands that Gapping only affect environments in which the antecedent and the Gap are embedded at the same depth inside their conjuncts (see Hankamer 1979; Hudson 1976; Sag 1980; Johnson 2003; among others). In (6), for instance, which is unambiguously associated with the wide ellipsis interpretation (6b), Isomorphism is responsible for excluding the narrow reading (6a). The antecedent in the first conjunct is embedded under two IP nodes, while the Gap is part of the matrix clause in the second conjunct. (6)
[IP John wanted [IP to write plays]] and [IP Sam D poems]. a. *D ¼ wrote/writes b. D ¼ wanted to write
It is well known that unlike Gapping, VPE is not regulated by Isomorphism, but may choose its antecedent freely from inside the antecedent clause (see Hankamer and Sag 1976; Hudson 1976; Sag 1980; Hardt 1993; Johnson 2003; among others). The missing VP in (7) can e.g. be recovered either from the higher or from the lower VP in the first conjunct: (7)
John wanted to write plays, and Sam did D, too. a. D ¼ write plays b. D ¼ wanted to write plays
The same observation has been made for ACD (Larson and May 1990; Fiengo and May 1994: 254). As illustrated by (8), a broad reading for a host quantifier (some city) is compatible with reconstruction of an embedded VP, although the elided VP,4 headed by visited, is
158
Winfried Lechner
non-isomorphic to its antecedent, which is embedded under an additional IP layer: (8)
John wants to visit some city that you did D. want > some city a. D ¼ visited b. *D ¼ wanted to visit
PCs appear to instantiate a third, inhomogeneous group of constructions with respect to Isomorphism. To begin with, subject comparatives strictly observe Isomorphism, as documented by the observation that (9) lacks the narrow ellipsis construal (9a). (9)
[IP More people wanted [IP to write plays]] than [IP D poems]]. a. *D ¼ d-many write/wrote b.
D ¼ d-many people wanted to write
Presupposing, as before, a Gapping analysis for PCs, the underlying source of (9) can be rendered as in (10). The representation in (10), which simultaneously captures the output of TR (i.e. extraposition of the thanXP) and the effects of Gapping, clearly respects the dictate of Isomorphism, as it locates the ellipsis site and the antecedent in structurally identical positions: (10)
[IP [IP More people [IP wanted to write plays]] than [IP d-many peopleCD [IP wanted to writeGapping poems]]
Interestingly, object PCs tolerate a wider range of interpretations than their subject counterparts. In addition to the isomorphic, wide ellipsis reading (11b), the expression (11) can also be assigned the narrow, and apparently non-isomorphic parse in (11a).5 The contrast between (9) and (11) is puzzling inasmuch as it suggests that the formation of subject and object PCs implicates two distinct types of ellipsis, which display the Isomorphism properties of Gapping and VPE, respectively.6 (11)
[IP John wanted [IP to write more plays]] than [IP Sam D] a. D ¼ write/wrote d-many plays b. D ¼ wanted to write d-many plays
The impression that PCs do not lend themselves to a unified analysis is reinforced by a second class of contexts which discriminate between Gapping and VPE. In all of the Isomorphism violations considered so far, the antecedent was more deeply embedded than the elided predicate. But VPE may also operate on strings which reverse the structural relations,
On binding scope and ellipsis scope
159
such that it is the elliptical VP – and not the antecedent VP – which is located farther away from the root node (see Johnson 2003 and references therein): (12)
[IP John wrote plays], and [IP Sally wanted [IP to D]], too. D ¼ write plays
Corresponding examples of Gapping, in which the antecedent is less deeply embedded than the Gap are still excluded by Isomorphism: (13)
*[IP John wrote plays] and [IP Sam wanted [IP to write poems]].
Crucially, object comparatives behave just like Gapped coordinations, they equally fail to allow ellipsis to reach into an embedded clause: (14)
*John [VP wrote more plays] in a month than [IP Sam wanted [IP to D in a year]]. D ¼ write d-many plays
In sum, the ellipsis process which feeds PC-formation appears to share properties of Gapping as well as VPE. On the one hand, deletion displays the same restrictiveness as Gapping in subject PCs (see (9)) and object PCs that construe the Gap lower than the antecedent (see (14)). On the other hand, ellipsis was also seen to be licit in selected contexts which appear to violate Isomorphism (see (11)). The objective of an analysis which uniformly attributes PCs to the workings of Gapping accordingly has to consist in finding a principled explanation for why (11) also possesses the non-isomorphic reading (11a). In what follows, it will be shown that the current approach is not only compatible with, but already implicitly contains, such an account. Above, it was assumed that the coordinate parse of comparatives is sponsored by the movement process of TR, which shifts the than-XP to the right periphery of the clause. If TR extraposes the than-XP into the higher clause, as suggested by the bracketing in (11) and the tree diagram (15), the narrow ellipsis interpretation (11a) can evidently only be obtained at the cost of an Isomorphism violation (16). But the TR analysis also provides an alternative derivation, which proceeds in terms of local extraposition, and which parses the than-XP into a coordinate structure with the subordinate, instead of the matrix, clause as illustrated by (16). In (16), low attachment to IP2 creates now an appropriate context for Gapping of the embedded predicate (write), generating the narrow interpretation (11a). On this view, ellipsis ambiguity in PCs is resolved derivationally by mapping the two readings to two distinct syntactic representations. As a
160
Winfried Lechner
corollary, the analysis restores the homogeneity of PCs with respect to Isomorphism, removing an obstacle for a unified Gapping account of PCs. As already pointed out above, subject PCs such as (9) contrast with object comparatives in that they fail to exhibit ellipsis ambiguity: (15)
Wide Reading (¼ (11b))
IP IP1 John
IP
than Sam
VP wanted
CP
VP wanted
CP
IP2 PRO
IP
to write more plays
(16)
VP
PRO
VP
to write d-many plays
Narrow Reading (¼ (11a))
IP1 VP
John
IP
wanted IP2 PRO
IP
than VP
Sam
to write more plays (9)
More people wanted to write plays than D poems. a. *D ¼ d-many people write/wrote b. D ¼ d-many people wanted to write
VP write/wrote d-many plays
On binding scope and ellipsis scope
161
On current assumptions, this follows directly from the fact that in (9), the than-XP originates inside the matrix subject, as detailed in (17a). Subsequent to TR in the transition from (17a) to (17b), the than-XP enters a comparative coordination with the matrix IP node in (17b): (17)
a. [IP [More people [than d-many peopleCD want to write poems]] want to write plays] b. [[IP [More people tk] want to write plays] than [IP d-many peopleCD want to write poems]k ] c. [[IP More people want to write plays] than [IP d-many peopleCD want to writeGapping poems]]
Isomorphism now ensures that Gapping in (17b) results in the wide ellipsis interpretation (17c). The narrow reading (9a) cannot be produced, because Gapping of the embedded predicate to the exclusion of the higher verb would require the than-XP to lower into the subordinate clause, in violation of the ban on downward movement: (18)
a. [IP [More people [than d-many peopleCD write/wrote poems]] want to write plays] b. *[IP [More people tk] want [ [IP to write plays] than [IP d-many peopleCD write/wrote poems]k]]
Thus, the TR analysis reduces an at first sight puzzling inconsistency in the behavior of PCs with respect to Isomorphism (subject vs. object PCs) to a general property of syntactic movement operations. The account is furthermore supported by a third class of PCs, represented by adjunct comparatives, which pattern along with object PCs in that the than-XP can be read with narrow or broad ellipsis:7 (19)
The boys wanted to go to the movies more often than the girls D. a. D ¼ went to the movies d-often b. D ¼ the girls wanted to go to the movies d-often
Just as with object PCs, ambiguity derives from variation in the scope of TR. The adjunct originates inside the subordinate clause, and the than-XP may therefore extrapose locally, coordinating with the embedded IP, or move into the higher clause, leading to matrix coordination and wide ellipsis. Before integrating additional data into the analysis, a remark is in order regarding the scope of the than-XP in (11)/(16) relative to the intensional predicate want. The following section will address this issue, proceeding from there to further empirical ramifications.
162
Winfried Lechner
7.2.2
Binding of world variables
The analysis for the object PC (11) advanced so far, which was made explicit in (15) and (16), encounters a complication. (Relevant portions of (11) are repeated below.) (11)
Narrow reading: John wanted [[PRO to write more plays] than [Sam write/wrote]]
Even though in the narrow reading of (11) the than-XP resides within the syntactic scope of want (see also (16)), the content of the than-XP is not part of John’s wishes. On this specific interpretation, sentence (11) compares the number of plays John wants to write with the number of plays Sam actually wrote and not with the number of plays which John wants Sam to complete. Thus, it must be ensured that the Gapped predicate inside the than-XP is interpreted at the evaluation index w0, as in (20), and not at the bouletic alternatives provided by want, as in (21): (20)
a. John wantedw0 to writew more plays than Sam wrotew0 (evaluated at w0) b. 8w[Rbouletic alternative for John(w0)(w) ! 9d[John_ writes_ d-many_ plays_ in_ w & d > max{d0 j Sam_ wrote_ d0 -many_ plays_ in_ w0}]] for all worlds w which are consistent with John’s wishes in w0, John writes more books in w than Sam wrote in w0
(21)
a. *John wantedw0 to writew more plays than Sam wrotew (evaluated at w0) b. 8w[Rbouletic alternative for John(w0)(w) ! 9d[John_ writes_ d-many_ plays_ in_ w & d > max{d0 jSam_ wrote_ d0 -many_ plays_ in_ w}]] for all worlds w which are consistent with John’s wishes in w0, John writes more books in w than Sam wrote in w
This aspect of the analysis proves largely unproblematic, though. It has been observed by various authors that world variables associated with predicates inside comparative complements can be anchored to the actual world by non-local binding of world variables, or double indexing,8 resulting in a de re reading without scoping of the than-XP (see Postal 1974; Hoeksema 1984; Heim 1985; 2001; Kennedy 1995; Rullmann 1995). What needs to be added is the proviso that if the than-XP contains a Gap – to be precise, a finite Gap (see (25) below) – the double indexing strategy (20) is not optional, but obligatory.9
On binding scope and ellipsis scope
163
The present account also naturally extends to a slightly more complex class of examples, in which nominal comparatives serve as subjects of ECM-infinitivals, illustrated in (22). The paradigm in (22) is of specific interest inasmuch as it brings to light another factor which is instrumental in determining the interpretation of elliptical comparatives. (22)
John wanted more people to write plays than D poems. a. *D ¼ d-many people write/wrote b. D ¼ John wanted d-many people to write
Observationally, the ECM-PC in (22) licenses the wide ellipsis interpretation (22b), but lacks the narrow construal (22a).10 Evaluated at some point in the actual world, (22) does not express the proposition that John wanted the number of playwrights to exceed the number of people who are actual poets under current circumstances. Rather, (22) conveys information about the relation between poets and playwrights in worlds that accord to John’s desires. On current views, (22) is parsed in the tree in (23) below, a representation similar to the one assigned to the narrow reading for the object PC in (11a)/ (16) (repeated below). But in contrast to the object PC (16), where the world variable inside the than-XP is bound by the highest evaluation index w0 (see also note 9), the world variable in (23) must be identified locally. If the world variable were bound at a distance, one would be incorrectly led to expect that the surface string (22) feeds the LF (24a) and its pertaining semantic translation in (24b), both of which represent the absent narrow interpretation (22a). (23)
ECM-PC (¼ (22b))
IP VP
John wantedw0
IP IP more people
than VP to writew plays
IP d-many people
VP to writew poems
164
Winfried Lechner
(16) Object PC (=11a)
IP1 VP
John wantedw0
IP IP2 PRO
VP to writew more plays
(24)
IP
than Sam
VP wrote/writew0 d-many plays
a. *John wantedw0 [[IP more people to writew plays] than [IP D poems]] D ¼ d-many people writew0 poems narrow reading of (22) b. 8w[Rbouletic alternative for John(w0)(w) ! 9d[d-many_ people_ write_ plays_ in_ w & d > max{d0 jd0 -many_ people_ write_ plays_ in_ w0}]]
Hence, the question emerges why non-local binding of world variables is attested in object PCs ((11a)/(16)), but not in ECM-PC ((22)/(23)). A possible answer to the question above can be devised by combining a descriptive condition on double indexing with an independent difference between object and ECM-PCs. Turning to the latter ingredient of the analysis first, observe that object PCs may embed nominative remnants (John wants to write more books than Sam/heNom), which is diagnostic of the presence of finiteness features inside the than-XP.11 In contrast, the than-XP of ECM-PCs by definition lacks such a nominative term. In the ECM-PC (23), the topmost subject position of the than-XP is the accusative ECM-subject,12 and the predicate is necessarily realized as an infinitival. This contrast suggests the hypothesis on double indexing in (25): (25)
Conjecture about Double Indexing (CDI) a. A finite gapped predicate inside a than-XP must be double indexed (but see note 9). b. A non-finite gapped predicate inside a than-XP must be locally bound.
On binding scope and ellipsis scope
165
On the one side, clause (25a) of the CDI licenses double indexing in object PCs, such as (11a)/(16), because the nominative remnant signals that the Gapped predicate is finite.13 Clause (25b) on the other side prohibits longdistance binding of world variables if the predicate is located inside nonfinite ECM-complements, accounting for the absence of the narrow ellipsis construal (24). In the latter case, the world variable of the infinitival must be identified by the next available binder, which is provided by the matrix predicate want, as documented by the LF (26a). The translation (26b) finally reveals that the content of the than-XP is now correctly interpreted de dicto, as being part of John’s desires: (26)
a. John wantedw0 [[IP more people to writew plays] than [IP D poems]] D ¼ d-many people writew poems Wide reading of (22) b. 8w[Rbouletic alternative for John(w0)(w) ! 9d[d-many_ people_ write_ plays_ in_w & d > max{d0 jd0 -many_ people_ write_ plays_ in_ w}]]
To summarize the analysis of ECM-PCs, the embedded comparative coordination (23) is – somewhat surprisingly – associated with the wide ellipsis reading (22b)/(26), while the narrow reading (22a) is blocked by the CDI (in case the string is parsed into a comparative coordination at the level of the embedded IP-node) and by the general isomorphism constraint on Gapping (if the string is parsed into a comparative coordination at the level of the matrix IP-node).14 While the two clauses of the CDI in (25) only express descriptive generalizations in their present form, at least (25b) can be reduced to a more general property of verbs. Percus (2000: 201) observes that verbal predicates differ from their nominal counterparts in that the world variables of the former are indiscriminately bound by the closest available binder. Percus does not consider evidence from ellipsis (or PCs, for that matter), though, and one might therefore speculate that this principle applies in all generality only to overt verbs, but that a proper subset of silent verbal predicates behaves more liberally in being able to skip the most local binder. As was seen above, the members of this set must minimally fulfill the additional requirement of finiteness. From this it directly follows that Gapped verbs inside object comparatives can be bound at a distance, while all other verbal predicates – including Gapped verbs in ECM-PCs – do not have access to the double indexing strategy. Finally, the fact that the world variables of nominal predicates are subject to looser binding conditions does not pose further complications (on transparent and opaque readings
166
Winfried Lechner
of narrow scope NPs (see e.g. Fintel and Heim 2002 and references). As Gapping is restricted to verbal predicates, NPs do not fall under (25) and a further qualification in the definition of (25) is therefore unnecessary. At least three open questions remain, two of which are directly related to the CDI. First, any analysis of comparatives – elliptical or not – needs to explain why verbal predicates inside than-XPs are exempt from the locality condition on world-variable binding identified by Percus (see note 8). Second, why does the absence of a phonological matrix enforce double indexing (again, modulo note 9)? Finally, which role does finiteness play in licensing long-distance binding? Plausibly the ability of finite predicates to introduce independent eventualities contributes to their exceptional status (see e.g. Wurmbrand 2001). At the moment, I have to delegate a more solid theoretical grounding of these restrictions to further research. As will be shown below, there is some suggestive evidence, though, that at least the finiteness distinction should indeed be a component of the analysis, supporting the general direction of the present account. Some independent confirmation for parts of the CDI in (25), in particular the assumption that only finite Gapped predicates can escape the verdict of Percus’s generalization, comes from the distribution of reflexives inside the comparative complement and their interaction with de dicto/de re ambiguities. The argument has two steps to it. To begin with, note that PCs differ from clausal comparatives in that they license long distance reflexives (see e.g. Brame 1983): (27)
a.
John couldn’t possible be taller than himself.
b. *John couldn’t possibly be taller than himself is. This contrast correlates with a further observation. The subject position of (predicative) PCs is not necessarily assigned nominative case, but may be marked by default accusative (Brame 1983): (28)
a.
John is older than me.
b. *John is older than me am. These two properties discriminating PCs from clausal comparatives lend themselves to a unified analysis in terms of finiteness (Lechner 2001, 2004). On the one hand, PC remnants may bear accusative case because the thanXP of PCs does not (necessarily) contain finiteness features. Provided that binding domains are defined in terms of finiteness (Chomsky 1981; 1986b), the than-XP of PCs does, on the other hand, not constitute a binding domain, and the anaphor in (27a) can therefore be licensed by an external antecedent.15 Finally, what is relevant for present concerns is the observation that reflexive remnants signal the absence of finiteness.
On binding scope and ellipsis scope
167
The second step of the argument builds on Russell’s (1905) observation that predicates inside certain comparatives embedded under propositional attitude verbs can be bound by non-local world indices, as exemplified by (29). (29) possesses a contradictory de dicto reading, which reports an inconsistent belief, as well as a sensible de re interpretation (see von Stechow 1984; Rullmann 1995; Kennedy 1995): (29)
Bill believesw0 that Mary isw taller than she is D. D ¼ d-tallw0/w
de dicto/de re
Proceeding to the synthesis, paradigm (27) demonstrated that long distance reflexives in the subject position of PCs are licensed only if the than-XP lacks finiteness features. Moreover, the CDI (25) precluded double indexing of non-finite predicates. Taken together, these two hypotheses entail the prediction that PCs with reflexive remnants should exclusively license a contradictory, de dicto reading. And in fact, this expectation is borne out. Substituting a reflexive for the name bleeds the consistent de re reading, in support of the CDI (see Hellan 1981; von Stechow 1984; Heim 1985; McCawley 1998): (30)
Bill believesw0 that Mary isw taller than herself D. a. *D ¼ d-tallw0 b. D ¼ d-tallw
de dicto/*de re
To recapitulate the findings of the current section, extraposition by TR was seen to mark the size of ellipsis for Gapped comparatives. PCs resemble in this respect ACD. However, contrary to ACD, which employs QR (or some variant thereof), the movement process which determines the size of ellipsis in PCs also feeds PF. Thus, its presence or absence can be directly verified. Moreover, unlike ACD, PCs are subject to strict Isomorphism.16 These two properties make it possible to infer a one-to-one mapping between the overt syntactic position of the than-XP, and the scope of ellipsis. Analyses of ACD, on the other hand, arrive at a structural correlation (at LF) only by way of the additional hypothesis that not all movement can be detected at PF. Even though this latter assumption is well established, the direct empirical evidence for the correspondence between movement and ellipsis scope which is elicited by PCs provides a missing link in the pattern, and thereby contributes substantially in consolidating the standard account of this subclass of ellipsis phenomena. Object PCs and ECM-PCs were moreover seen to differ in that only the former display the effects typically associated with double indexing. While the content of the than-XP can be construed outside the scope of the matrix predicate in object PCs which are parsed into a narrow comparative
168
Winfried Lechner
coordination (20a), ECM-PCs do not tolerate long distance anchoring of the word variable (26a): (20a)
John wantedw0 [[IP to writew more plays] than [IP Sam D]]. D ¼ writes/wrotew0 d-many plays
(26a)
John wantedw0 [[IP more people to writew plays] than [IP D poems]]. D ¼ d-many people writew/*writew0 poems
The analysis proposed above reduces this contrast to a difference in finiteness and the empirical generalization in (25), which could be shown to receive additional empirical justification from the distribution of reflexive remnants in Russell sentences. Section 7.2.2 put forward arguments for the hypothesis that ellipsis ambiguity in PCs correlates with ambiguous syntactic representations. In section 7.3, I will examine certain empirical ramifications of this view, presenting evidence from correlations between ellipsis scope and the binding scope of the remnant. The results will be seen to vindicate the TR analysis as well as the assumption that PCs embed empty syntactic structure and can therefore not be base-generated.
7.3
Binding scope in PCs
In the previous section, it was argued that Isomorphism ensures that in PCs, the height of comparative coordination directly determines the size of the Gap. This leads one to expect that the height of attachment of the thanXP (i.e. its scope) should also have repercussions on other c-command and scope sensitive phenomena. In the current section, I will address this prediction by means of an investigation of the interaction between ellipsis scope and the syntactic binding scope of the remnant embedded inside the than-XP. Since quantifier scope and anaphoric dependencies in comparatives are subject to various interfering factors,17 the discussion will be exclusively based on the distribution of disjoint reference effects. 7.3.1
The size of ellipsis and binding scope
On current assumptions, PCs are derived by extraposition by TR, followed by Gapping. Gapping is in turn restricted by Isomorphism, and the size of the ellipsis inside PCs therefore directly matches the scope of TR: narrow TR correlates with local Gaps, and wide TR with long-distance Gapping. (As was seen in section 7.2, PCs differ in this respect from ACD.) In addition, extraposition also interacts with Binding Theory, in that
On binding scope and ellipsis scope
169
extraposed clauses do not reconstruct for the computation of Principle C (see Williams 1974; Taraldsen 1979; Fox and Nissenbaum 1999): (31)
a. ?? I gave himj a book that Johnj liked yesterday. b. I gave himj a book yesterday that Johnj liked.
Thus, comparatives which combine Gapping, extraposition, and disjoint reference effects are shaped by two intersecting constraints: first, TR determines the size of ellipsis and second, TR marks the LF position of the extraposed category. In conjunction, these two generalizations also imply that the size of ellipsis in PCs should directly correspond to the syntactic binding scope of remnants inside the than-XP. In what follows, I will substantiate the TR analysis by evaluating the correctness of a number of corollaries which follow from this hypothesized correlation between ellipsis scope and binding scope. The tests to be employed use disjoint reference effects as a diagnostic for the binding scope of the remnants. To anticipate, it will turn out that the facts corroborate the TR analysis, but pose a challenge for base-generation analyses of PCs, such as Brame (1983) and Napoli (1983). Ambiguous object PCs such as (32), in which the remnant contains an embedded r-expression and the matrix clause contains a pronoun to be construed coreferentially with that r-expression, provide the first diagnostic context for the correctness of the TR analysis. (32)
[IP1 Mary promised himj [IP2 PRO to invite more people than [Johnj’s sister]]]
On the Gapping analysis of PCs, variation in the scope of TR is predicted to affect the shape of ellipsis as well as the size of the c-command domain of the pronoun. As illustrated by (33) and (34), this prediction is correct. The narrow ellipsis interpretation (33a), which correlates with comparative coordination at the embedded node IP2 (see (34a) on the next page), induces a Principle C violation, as the name remains within the syntactic scope of the pronoun.18 Wide ellipsis, as in (33b), on the other side involves comparative coordination at the matrix IP-node IP1 (see (34b) on the next page). Thus, the than-XP has been raised to a position outside the c-command domain of the pronoun, leading to obviation of a Principle C effect (see (2) for a related example involving covert movement and ACD). (33)
Mary promised himj PRO to invite more people than [Johnj’s sister] D. a. *D ¼ invited d-many people b. D ¼ promised himj to invite d-many people
170
Winfried Lechner
(34)
a. Narrow reading *IP VP
Mary promised
VP himj
IP than
IP VP
PRO
IP Johnj’s sister
to invite more people
VP invited d-many people
b. Wide reading IP IP
Johnj’s sister
VP
Mary
IP
than
VP
promised himj
VP
promised IP
PRO
VP
himj VP
to invite more people
IP PRO
VP to invite d-many people
The same contrast can also be replicated for comparatives embedded under object control verbs: (35)
We convinced himj PROj to donate more money than [Johnj’s sister] D. a. *D ¼ donated d-much money b. D ¼ convinced himj to donate d-much money
On binding scope and ellipsis scope
171
In (33) above, the remnant contained the name, while the matrix clause embedded the trigger for the Principle C violation. But reflexes of the systematic covariation between the scope of TR – i.e. the height of coordination – and ellipsis scope can also be detected by examining the behavior of r-expressions which have been reconstructed into the than-XP via ellipsis resolution (see Fiengo and May 1994 for ACD). In particular, if the positions of binder and bindee are reversed, as in (36), only the narrow ellipsis reading (36a), which is represented by (38a), is compatible with a coreferential construal, while the wide ellipsis interpretation (36b) (tree (38b)) induces a disjoint reference effect: (36)
Mary promised [Johnj’s sister] to invite more people than hej. a. D ¼ invited d-many people b. *D ¼ promised [Johnj’s sister] to invite d-many people
Since pronouns do not make good remnants in English for independent reasons, this second prediction can be better evaluated on the basis of the German example in (37). (37)
Maria hat [der Schwester von Hansj] versprochen mehr Leute Maria has [the sister of Hansj] promised more people einzuladen als erj to invite than he (Mary promised John’s sister to invite more people than he.) a. D ¼ promised to invite d-many people b. * D ¼ promised Johnj’s sister to invite d-many people
(38)
a. Narrow reading IP VP
Mary promised
VP Johnj’s sister
IP IP PRO
IP
than VP
to invite more people
hej
VP invited d-many people
172
Winfried Lechner
b. Wide reading *IP IP
than VP
Mary
hej VP
promised
IP
Johnj’s sister
promised IP
PRO
VP VP Johnj’s sister
VP
IP VP
PRO
to invite more people
to invite d-many people
Again, it is possible to find examples involving object control which attest to the same point: (39)
We convinced [Bill Gatesj’s sister] to donate more money than hej D. a. D ¼ donated d-much money b. *D ¼ convinced [Bill Gatesj’s sister] to donate d-much money
(40)
mehr Leute Maria hat [die Schwester von Hansj] uberredet ¨ Mary has [the sister of Hansj] convinced more people einzuladen als erj. to invite than he (Mary has convinced John’s sister to invite more people than he.) a. D ¼ convinced to invite d-many people b. *D ¼ convinced [Johnj’s sister] to invite d-many people
The first two predictions were based on tests for the c-command relations which obtain between the remnant and some other NP in the matrix clause. However, it is also possible to design a more fine-grained diagnostic which measures the distance between the remnant and the NP inside the matrix clause, and which derives from the inspection of Principle B effects in contexts such as (41). (41)
Mary convinced us to send himj more money than Johnj. a. *D ¼ sent himj d-much money b. D ¼ convinced usk PROk to send himj d-much money
If (41) is read with narrow ellipsis, as in (41a) (diagrammed in (42a) on the facing page), the pronoun is reconstructed into the local binding domain of the name, and Principle B accordingly prohibits coreference. Broad ellipsis
On binding scope and ellipsis scope
173
as in (41b) (diagrammed in (42b) below) leads on the other side to the reconstruction of an additional binding domain which separates the name from the pronoun. Since the pronoun is now free in its minimal clause, the coreferential interpretation becomes available. (42)
a. Narrow reading *IP Mary
VP VP
convinced
IP
us IP PRO
than VP
IP Johnj
VP
to send himj more money
sent himj d-much money
b. Wide reading IP
Mary
IP
than
IP
Johnj
VP VP
convinced
convinced IP
us PRO
VP VP us
VP to send himj more money
IP PRO
VP to send himj d-much money
To summarize, it was illustrated that the height of coordination interacts with the interpretive component in a systematic way, furnishing strong support in favor of generating PCs by TR and the assumption that ellipsis
174
Winfried Lechner
resolution is subject to Isomorphism. This finding is also of broader theoretical interest, as it consolidates the insight gained from the study of ACD that ellipsis and binding scope are systematically related, in support of the view that each interpretation is assigned a distinct abstract syntactic representation, which is the output of a derivational process. But PCs and ACD also differ along two parameters. First, extraposition by TR proceeds in overt syntax, while ACD relies on covert movement (at least of the ACD-host19). This disparity might be a welcome result, as a conflation of properties is represented overtly in PCs, which can be inferred to exist only at an abstract level in ACD. Second, the ellipsis in PCs is contained inside an argument (the than-XP leaves an interpretable trace inside the DegP), while it is embedded inside a restrictive relative clause in ACD. This property strongly indicates that PCs cannot be treated as a special instance of ACD. In particular, it poses an obstacle for a common analysis of the two phenomena along the lines of a late merger account, as it has been developed for ACD (see Fox and Nissenbaum 1999, Fox 2002); because only adjuncts can be merged counter-cyclically. Next, I will turn to a comparison between the TR analysis with (aspects of) Heim (1985)’s direct analyses of PCs, which will revolve around the binding theoretic properties of remnants in simple root comparatives.
7.3.2
Evidence for syntactic ellipsis
Before proceeding to further empirical ramifications of the TR analysis, section 7.3.2.1 outlines the essentials of the semantics for base-generated PCs developed in Heim (1985), and some predictions of the direct account. The remaining part of this chapter presents a further set of binding effects in PCs, which will be seen to support the TR approach over the competing direct analysis. 7.3.2.1 Semantics for direct analysis: Heim (1985) According to the direct analysis of PCs, the than-XP does not syntactically embed any other covert categories apart from the remnant. Heim (1985) demonstrates that such non-elliptical LF representations can be given a compositional interpretation.20 More specifically, Heim assumes that the comparative morpheme -er denotes the function from pairs of individuals to individual properties in (43). (43) distributes the members of the pair over the predicate P, and then compares the degree to which individual a is P to the degree to which individual b is P. The translation of the predicative PC (44) accordingly proceeds as in (45): (43)
[[-er]](< a,b >)(f) ¼ 1 iff f (a) > f(b)
(44)
Mary is taller than Bill.
On binding scope and ellipsis scope
(45)
175
[[-er]]()(lxiy[[x is y-tall]]) ¼ lxiy[x is y-tall](Mary) > lxiy[x is y-tall](Bill) ¼ iy[Mary is y-tall] > iy [Bill is y-tall]
Three covert movement operations are implicated in the derivation of the transparent LF which provides the input for the semantic representation (45). The comparative morpheme -er covertly raises to an IP-adjoined position, as illustrated by (46), followed by movement of the correlate and the remnant, which in turn adjoin to -er (ignoring the internal structure of the adjunction complex): (46)
IP XP, t > -erj
YP Maryi Billk
IP< e, t > ti
. . .AP tall-tj than tk
l-abstraction over the trace of the correlate Bill yields the relation lxiy[x is y-tall] which serves as the second argument of the translation of -er. Moreover, movement of the remnant is semantically vacuous. The account generates two predictions which constitute independent evidence in favor of Heim’s claim that the correlate and the comparative morpheme undergo semantically motivated movement. First, Heim points out that the movement operation which raises the correlate to the comparative morpheme is subject to island constraints. Separating the correlate and the comparative by a complex NP, as in (47a), leads to a deviant output, since LF-movement of the correlate (see (47b)) fails to observe syntactic locality: (47)
a. ?* I spent more time with [NP a woman [CP that played the clarinet than the lute]]. b. . . . [[[more time] the clarineti] the lutek] with [NP a woman [CP that played ti] than tk] (Heim 1985: 49, (52))
Second, Heim demonstrates that movement of the comparative morpheme is equally subject to locality conditions. Structures in which only -er – but not the correlate and the remnant – is trapped inside an island, such as (48a), are strongly ill-formed: (48)
a. *[NP Someone [CP who could answer fewer questions]] made a good impression on Bill than on Fred. b. [IP [-erj [Billi Fredk]] [someone [CP who could answer littletj-many questions] made a good impression on ti than on tk]] (Heim, 1985: 47, (38))
176
Winfried Lechner
But the examples under (47a) and (48a) are also successfully handled by the TR account, although they are excluded for a different reason. To begin with, the TR analysis assigns to (47a) the parse in (49), in which all but one category of the than-XP have been removed by Gapping: (49)
*I spent more time with [NP a woman [CP that played the clarinet]] than I spent D with [NP a woman [CP that played the lute]]
The representation (49) reveals that the Gap includes an island as well as (parts of) a finite CP-node. Such instances of long-distance Gapping always result in ill-formed expressions, though, as illustrated by the reduced coordinate structure (50) (see e.g. Johnson 2003): (50)
*Pat spent two hours with [NP a woman [CP that played the clarinet]] and Pat spent three hours with [NP a woman [CP that played the lute]]
Thus, the TR analysis relates the deviance of (47) to a general property of Gapping, and not to a violation of the bounding conditions on movement. Similar and additional considerations apply to (48). A derivation of (48) in terms of Gapping, as in (51), would once again imply an illicit instance of deletion of a finite CP-node: (51)
*Someone [CP who could answer fewer questions tk] made a good impression on Bill than [someone [CP who could answer D] made a good impression on Fred]k
In addition, the extraposition process which supplies the context for the application of Gapping in (51) violates syntactic locality,21 as extraposition out of complex NPs in subject position is generally prohibited (Culicover and Rochemont 1990). (52)
*Someone [CP who could answer the question ti] made a good impression on Bill [which had been asked in class]i.
In sum, locality effects measuring the distance between the remnant and the comparative morpheme fall out both from the direct analysis and the TR account. But the two hypotheses can be empirically distinguished on the basis of the predictions they entail for the distribution of coreference relations in PCs. 7.3.2.2 Binding and the direct analysis Returning to disjoint reference effects in PCs, observe that the examples discussed in section 7.3.1 (at least potentially) exhibited ambiguity with respect to the interpretation of the ellipsis. Interestingly, though, it can be shown that legitimate binding relations are governed by structural restrictions also in simple, unambiguous PCs, whose interpretation is not a function of the scope of TR. In particular, these contexts illustrate that the correct determination of possible referential
On binding scope and ellipsis scope
177
dependencies presupposes the assumption that the ellipsis in PCs replicates the structural relations of its antecedent clause. Thus, these findings provide straightforward evidence for the claim that PCs project structure at LF, but contradict the premise of the direct analysis, according to which the than-XP does not embed a syntactically represented ellipsis site. The environments relevant for present purposes test for the effects of varying the structural positions of the remnant and its correlate in the matrix clause, while keeping constant an indexical dependency between (parts of) the remnant and a term distinct from the correlate in the matrix clause. To begin with, (53a) shows that a pronoun which resides in a position lower than the correlate can be construed coreferentially with a name inside the remnant. As documented by the tree in (54a), the TR analysis correctly captures this generalization, because it locates the reconstructed pronoun below the (non-c-commanding) name: (53)
a. Sally introduced himj to more people than [Peterj’s sister]SUB b. *Hej introduced Sally to more people than [Peterj’s sister]DO
(54)
a. (=(53a)) IP IP
Sally
IP
than Peterj’s sister
VP introduced himj to more people
VP introduced himj to d-many people
b. (=(53b)) *IP IP Hej
IP
than VP
introduced Sally to more people
hej
VP introduced Peterj’s sister to d-many people
Moreover, once the content of the subject and the direct object is exchanged, as in (53b)/(54b), a Principle C violation ensues. This is so as promotion of the pronoun to the subject position in the matrix clause leads to a corresponding reversal of the terms in the than-XP, and the remnant therefore ends up within the c-command domain of the silent pronoun.22 Thus, the strong disjoint reference effect in (53b), and the sharp contrast between (53a) and (53b),
178
Winfried Lechner
provides solid evidence that Principle C has access to information about the structural organization of the than-XP. Such a result is in line with the ellipsis analysis, but can be shown to pose a problem for the direct analysis of PCs, which employs covert raising of the remnant and the correlate. More specifically, the direct analysis predicts (53a) and (53b) to be on a par, because LF-movement of the correlate and the remnant in (55a) and (55b), respectively, generates indistinguishable LF representations as far as the structural relations determining binding are concerned. (55a)
Sally introduced himj to more people than [Peterj’s sister]SUB (¼ (53a)) IP
XP -er
IP YP
ti
Sallyi [Peterj’s sister]k
VP introduced himj to t-many people than tk
(55b)
*Hej introduced Sally to more people than [Peterj’s sister]DO (¼ (53b)) *IP
XP -er
IP YP
Sallyi [Peterj’s sister]k
Hej
VP introduced ti to t-many people than tk
On binding scope and ellipsis scope
179
In both LF representations under (55), the remnant Peter’s sister crosses over the pronoun and thereby escapes its c-command domain, regardless whether the pronoun originates in a position lower ((55a)) or higher ((55b)) than the correlate (Sally).23 Depending on the particular perspective one takes on which positions of the remnant chain are visible to Binding Theory, the pronoun should therefore either trigger a Principle C effect in both structures (by reconstruction; but see below for a qualification), or in none of them. Hence, without the aid of additional assumptions, the direct analysis cannot account for the generalization that the structure of the ellipsis site plays an important factor in determining legitimate referential dependencies. Before concluding, I will briefly consider a possible amendment to the direct analysis, which – at least at first sight – seems to restore the tie between the two competing accounts. Assume that the representations in (55) misconstrue the position of the than-XP, and that the than-XP does not always originate low in the right periphery, but is base-generated as an adjunct to the node immediately dominating the correlate.24 For (53a), this yields the alternative parse in (56) below, whereas the structure of (53b) remains unaffected: (56)
Sally introduced himj to more people than [Peterj’s sister]SUB (¼ (53a)) IP
XP -er
IP YP
than tk
IP
Sallyi [Peterj’s sister]k ti
VP introduced himj to t-many people
Crucially, the analysis (56) succeeds now in capturing the contrast (53), because the r-expression inside the remnant (Peter’s sister, index k) is generated above – and not below – the pronoun.25 Thus, a Principle C effect can be avoided. But this modified version of the direct analysis equally faces complications.
180
Winfried Lechner
First, there is strong evidence that the than-XP forms a syntactic constituent together with the comparative XP to the exclusion of the correlate, and should therefore originate in a position inside the DegP.26 For instance, the than-XP and the comparative XP may move as a unit to the topic position in languages such as German, stranding the correlate (jemand/‘somebody’). (57)
als ShakespeareSU]DO hat noch nie [CP [ Bessere Stucke ¨ [ [ better plays than ShakespeareSU]DO has not yet [IP jemandSU [VP tDO verfasst]]] [ somebody [ written]]] (Nobody has ever written better plays than Shakespeare.)
However, (57) cannot be derived from a structure such as (56). As demonstrated by (58a), the subject remnant Shakespeare (by the assumptions adopted above) originates as an IP-adjunct, the node immediately dominating the correlate jemand/‘nobody’. Hence, the than-XP would have to lower from its base position to the comparative NP in object position (58b) prior to topicalization (58c), in violation of the ban on downward movement (in German, SpecCP may hold only a single XP, and adjunction to CP, as well as to SpecCP, is prohibited). (58)
a. [IP [IP nobody [VP has written better plays]] than Shakespeare] b. [IP [IP nobody [VP has written [NP [NP better plays] than Shakespearek]]] tk] c. [CP [NP [NP better plays] than Shakespearek]j . . . [IP [IP nobody [VP has written tj]] tk]]
Similar considerations carry over to VP-fronting in English, which can pied pipe the than-XP to the exclusion of the correlate nobody.27 (59)
. . . and [VP write a better play than Shakespeare], nobody did
Second, neither the revised nor the original version of the direct analysis is compatible with current assumptions about the nature of the LF interface to semantics. Going back to the representation (55b), which tracks the derivation of an ill-formed expression, notice that Principle C has to be computed on the basis of the reconstructed position of the remnant, whereas only the non-reconstructed, higher occurrence of the remnant is compositionally interpretable. (Note that the remnant had to QR for reasons of interpretability.) But this distribution of labor between the pre- and post-movement occurrences of the remnant is at conflict with the widely accepted hypothesis that binding and scopal properties of an NP are encoded in a single LF-copy (cf. ‘‘trapping effects,’’ as e.g. discussed
On binding scope and ellipsis scope
181
in Lebeaux 1990 and Fox 1999b). Moreover, it is not evident how reconstruction in (55b) can be effectuated in the first place, given that movement of the remnant does not leave an interpretable trace (recall from section 7.3.2.1 that the raised remnant does not bind a variable). In sum, the alternative implementation outlined above fails to remove the obstacles for the direct analysis. 7.4
Conclusion
In this chapter, I have defended a Gapping analysis of PCs which succeeds in capturing (i) restrictions on the size of ellipsis inside the than-XP, (ii) co-dependencies between the size of ellipsis and the syntactic domain in which binding principles are evaluated, and (iii) the impact of the structural organization of the matrix clause on possible referential dependencies between the remnant and the matrix clause. In addition, a descriptive condition was identified (Conjecture on Double Indexing), which restricted non-local binding of world variables to contexts in which the world variable is supported by a finite, silent predicate. Direct analysis of PCs were finally seen to be challenged by Principle C effects in PCs. PC-formation by TR shares properties with ACD, but the two constructions also differ in important ways. As for the commonalities, both ACD-QR and TR feed ellipsis resolution and bleed Principle C. In both constructions, the respective types of ellipsis (Gapping and VPE) presuppose a parse which relocates the container of the silent node. Moreover, ACD as well as PCs are characterized by a clustering of properties: the size of ellipsis matches the position in which binding scope is computed. However, the two operations also differ in three, partially related, aspects. First, PCs mark the position of the node embedding the ellipsis in overt syntax, while in contexts of ACD, the location of the ellipsis host can only be inferred indirectly, as QR proceeds covertly. Second, in PCs, which by assumption implicate Gapping, but not in ACD, the match between the size of ellipsis and the syntactic position of the ellipsis host is conditioned by Isomorphism (for conditions on VPE in ACD see Fox 2000, among others). Third, the category which embeds the ellipsis is an adjunct relative clause in the case of ACD, but an argument of the Deg0head in the comparative construction. These differences motivate the choice of diverging analytical approaches – although in the end, the strategies clearly reinforce each other. To begin with, in order to account for the correlation between ellipsis and binding scope, most studies of ACD hypothesize movement at a post-syntactic level (QR), or, more recently, overt, phonologically vacuous QR and late merger of the relative clause (see Fox and Nissenbaum 1999 and Fox 2002;
182
Winfried Lechner
and on covert overt movement also Pesetsky 2000). The latter view exploits the parallelism between extraposition and ACD-QR in their ability to obviate disjoint reference effects by adopting the assumption that the (extraposed) relative clause is combined with the hosting quantifier counter-cyclically, just like adjuncts of overtly fronted wh-phrases (Lebeaux 1990). Since on this account, only the hosting quantifier moves in overt syntax (by QR), marking the position in which the relative clause is merged, it follows that the correlation between size of ellipsis and binding scope is only indirectly mediated by movement in ACD. In contrast, PCs provide direct evidence for a ‘‘traditional’’ version of ellipsis resolution in terms of movement. Given that the than-XP in comparatives does not originate as an adjunct, a post-cyclic merger analysis is not viable for this type of construction. Moreover, movement by TR proceeds overtly. Hence, the current analysis of PCs unifies all properties which by hypothesis collaborate in ACD: overt (and in the case of ACD silent) movement and a strict correlation between ellipsis and binding scope. In this way, PCs demonstrate that ellipsis interacts with movement in the same way in which coreference relations are influenced by dislocation, supplying a missing link in the evolution of the analysis of ellipsis phenomena.
8
The silent content of bound variable pronouns Uli Sauerland Zentrum fur ¨ allgemeine Sprachwissenschaft, Berlin
Abstract Bound variable pronouns are usually taken to be interpreted solely as bound variables. This chapter argues that bound variable pronouns have an internal property argument P and that P is interpreted as a presupposition on the value of the bound variable pronoun. The argument leading to these conclusions proceeds through the following intermediate theses, which are of independent interest: 1) bound variable pronouns can differ in interpretation 2) indexation alone is insufficient to explain the way bound variable pronouns differ 3) bound variable pronouns can have a presupposition 4) in general, the silent content must be a silent property variable internal to the bound variable pronoun, rather than being an elided syntactic representation.
When I confronted a certain non-linguist I know with sentence (1), she voiced the opinion that the pronoun his refers to the set of boys. Of course, I quickly showed her that this cannot be the case because (1) does not mean the same as every boy likes the mother of the boys, and went on to tell her that a genuinely new concept like that of a bound variable is required to analyze the contribution of his to the meaning of (1). (1)
Every boy likes his mother.
My non-linguist friend, however, may easily be forgiven, given that it took Frege’s (1879) ingenuity to come up with the concept of a bound Acknowledgments: Thanks to Elena Anagnostopoulou, Danny Fox, Irene Heim, Hans Kamp, Graham Katz, Winfried Lechner, Sonja Lohrengel, Pauline Jacobson, Peter Krause, Orin Percus, Wolfgang Sternefeld, and Kazuko Yatsushiro for their comments. I presented an earlier version of this chapter at SALT 10 at Cornell University, and parts of this version are based on my paper in the proceedings of that conference (Sauerland 2000), while others substantially deviate from that version. This chapter was written at Tsukuba University in April 2002 and I would like to express my thanks to Hidekazu Suzuki and Tsutomu Yatsushiro for their hospitality. The chapter was revised in July and November 2005. I thankfully acknowledge the financial support of the German Research Council (Grant SA 925/1–2).
183
184
Uli Sauerland
variable which led to a successful analysis of sentences like (1) and that Frege’s insights are not well known in the general population. In this chapter, I address a puzzling phenomenon that will give us another reason to forgive the mistake of my non-linguist friend. This phenomenon, focus on bound pronouns, came to my attention some years ago, and I will argue that its analysis requires an appeal to precisely what my non-linguist friend took to be the reference of the pronoun in (1). I will argue that this pronoun can contain in it a silent reference to the set of boys. Once I’ve defended my non-linguist friend in this way, I will go on make this chapter relevant to this volume by considering the question of whether this silent content is silenced by ellipsis or silent for some other reason. The structure of my chapter is the following: In Section 8.1, I introduce the phenomenon of focus on bound pronouns and some background assumptions of focus semantics to argue that bound pronouns differ in meaning. In section 8.2, I consider and reject the possibility of accounting for these facts solely within the standard analysis of bound variable pronouns as indexed variables. In section 8.3, I argue for an analysis of the phenomena in question, making use of the idea that the bound variable pronoun is a bound definite description with its range as presuppositional content. In section 8.4, I consider the possibilities of accounting for the silence of this presuppositional content, namely an ellipsis analysis or silent property variable. Section 8.5 is the conclusion.
8.1
Contrast between bound variable pronouns
Consider the two sentences in (2). The salient interpretation is one where each occurrence of the pronoun his is interpreted as a bound variable pronoun bound by the subject quantifier of its respective sentence. (2)
Every boy called his mother. Every TEACHER, on the other hand, called HIS mother.
Note that the second occurrence of his in (2) must be focused. Here and in the following I indicate pitch accents, which show focus in English, by capitalization. Focus on a bound variable pronoun as in (2) is my primary interest in this chapter. In most cases, this focus is optional and I use a trick to make it obligatory in (2): it is easy to see that the focus in (2) is obligatory because of the presence of on the other hand, and, if we leave this out, the focus becomes optional. It is well known that focus is intimately connected with the meaning of constituents. Indeed, focus is as important as it is in current semantic
The silent content of bound variable pronouns
185
theorizing because it provides a way to test for the meaning of constituents that is independent of sentence meaning and a theory of composition. The examples in (3) are just a simple illustration of the generalization that focus is placed on those parts of a sentence that plausibly differ in meaning from a relevant antecedent sentence. In (3a) with focus on the pronoun her, the pronoun cannot refer to Mary, while unfocused her in (3b) must refer to Mary. (3)
a. On Monday, Bob called Mary. On TUESday, JIM called HER. b. On Monday, Bob called Mary. On TUESday, JIM called her.
From this perspective, example (2) suggests that two occurrences of a bound variable pronoun can differ in meaning. To argue that this is indeed true, I adopt the theory of focus licensing of Schwarzschild (1999). I shall briefly summarize the consequences of this theory that my research relies on, and then return to the question of focus on bound variable pronouns.
8.1.1
Focus licensing
The question of how focus is licensed has been discussed by Rooth (1992b) and Schwarzschild (1999). Here, I adopt an account that makes use of ideas of both (cf. Truckenbrodt 1995). From Schwarzschild’s (1999) proposal, I adopt the idea that focus is generally something to be avoided. Therefore focus is only licensed if none of the alternatives that are accessible has less focus. For example, consider the fact that the focused her in (4) cannot refer to Mary. Schwarzschild provides the following explanation: if her did refer to Mary, it would be possible to not focus her as (3b) shows. Here, we must assume that (3b) is an accessible alternative. (4)
On Monday, Bob called Mary. On TUESday, JIM called HER.
If her does not refer to Mary, but to someone else, the focus is obligatory. Since the focus cannot be omitted in that case, the focus is licensed in (4). In this way, Schwarzschild’s general idea provides an account for the facts in (3) and (4). From Rooth’s (1992b) analysis, I adopt the idea that focus licensing involves a relationship between two phrases: the focus domain and the focus antecedent. In example (4), the first focus domain is the second sentence and the focus antecedent is the first sentence. Intuitively the focused constituents in the focus domain cause the interpretation of the focus domain to differ from that of the antecedent. The formal condition I assume is (5) (cf. Rooth’s (1992b) Focus Interpretation Principle or
186
Uli Sauerland
Schwarzschild’s (1999) Givenness with (5a), Schwarzschild’s (1999) Avoid F with (5b)). (5)
A focus on the phrase XP is licensed only if there is a Focus Domain constituent FD dominating XP such that for its Focus Antecedent FA in the preceding discourse (or entailed by it) the following two conditions are satisfied: a. Givenness: [[FA]] 2 [[FD]]f. (i.e. there is a Focus-Alternative FD0 of FD with [[FA]] ¼ [[FD0 ]]) b. Contrastiveness: [[FA]] 62 [[FD]]f, where FD is identical to FD, except that XP is not focused in FD.
Consider again the examples in (3) and (4): The informal reasoning above can now be given as in (6). FD is the entire second sentence, while FA is the preceding sentence. Condition (5a) is satisfied because the FD0 in (6c) has the same interpretation as FA. Furthermore, (5b) is satisfied. Specifically, it is easy to see that FA is not a focus alternative of the FD in (6), where the pronoun is not focused. (6)
a. FD ¼ On [TUESday]F, [JIM]F called [HERj]F b. FA ¼ On Monday, Bob called Maryi c. FD0 ¼ On Monday, Bob called Maryi d. FD ¼ On Tuesday, [Jim]F called herj
Note, furthermore, that the reasoning we just went through relied on the fact that her and Mary were not coreferent. If they were coreferent, the meaning of the focus antecedent (6b) would be a focus alternative of FD, and therefore this choice of focus domain would not be sufficient to license the focus on the pronoun. One difference between the accounts of Rooth and Schwarzschild concerns the choice of focus domains and corresponding focus antecedents. In Schwarzschild’s account, the notion corresponding to a focus domain is a constituent that must be given. He assumes that all non-focused constituents must be given. Rooth, on the other hand, assumes that only certain non-focused constituents are marked as focus domains and leaves it open how these are determined. I follow Rooth’s account. 8.1.2
Focus on bound pronouns
Now consider again examples with focus on a bound pronoun. Specifically, focus on the bound variable pronoun is acceptable in example (7). In the introduction, I considered an example containing the adversative focus particle however, namely example (2), which made the focus
The silent content of bound variable pronouns
187
obligatory. I will return to the analysis of adversative focus particles as in (2) in section 8.1.3, and consider example (7) here. Intuitively, the focus seems to involve a contrast between his in the discourse and the focused his. (7)
Discourse: On Monday, every boy called his mother. a. On TUESday, every TEAcher called HIS mother.
How, then, is the focus in (7a) licensed? I will now show that it follows from the focus licensing condition adopted in the previous chapter, that the contributions the two bound pronouns in (7) make to the meaning of their sentences must be different. To show this, all possible choices of FA and FD must be looked at to verify that focus on HIS is only licensed if HIS and his differ in interpretation. I consider only two cases of FA and FD in (8). First, I look at FA and FD that do not include the binders, then at FA and FD that include the binders of the two pronouns in (9). (8)
a. FA ¼ his mother b. FD ¼ [HIS]F mother c. FD ¼ his mother
For the FA and FD in (8), contrastiveness requires that [[FA]] 62 [[FD]]f, and this entails that [[his mother]] 6¼ [[his mother]], which is a contradiction unless the representations contain some silent content that is not represented. Contrastiveness can only be satisfied if the occurrence of his in FA and the occurrence in FD have different interpretations. Second, consider FA and FD in (9), which include the binder. (9)
a. FA ¼ Every boy called his mother. b. FD ¼ Every [TEAcher]F called [HIS]F mother. c. FD0 ¼ Every boy called his mother. d. FD ¼ Every [teacher]F called his mother.
Contrastiveness requires that the focus alternative FD0 of FD shown in (10) differ in interpretation from FA. Again, the occurrences of his in (10) must somehow have different interpretations for (10) to be true. (10)
[[FD0 ]] ¼ [[Every boy called his mother]] 6¼ [[Every boy called his mother]] ¼ [[FA]]
In this way, Schwarzschild’s proposal for the licensing of focus implies the conclusion that bound pronouns are ambiguous in a way not evident from their segmental phonology.
188
Uli Sauerland
One might be tempted to propose a different explanation for the above facts. Namely, one might conclude that Schwarzschild’s theory of focus licensing should be dismissed. However, the difficulties in formulating such an approach seem to me to be substantial. Consider two ideas that come to mind here, but which I believe are quite immediately seen to be flawed: one idea takes feature agreement in variable binding to be the source of the focus features on bound variables, while the other idea assumes that the focus licensing takes into account a representation where a quantified statement is reduced to one not containing any quantifiers.1 Consider first the idea that feature agreement in variable binding is the source of the focus in (7). It is well known that bound variable pronouns agree in number, person, and gender with their antecedents (Heim 1994; Pollard and Sag 1994; Kratzer 1998). Example (11) illustrates number agreement on the pronoun they. (11)
The participants of the chess tournament all believe that they will win.
The natural interpretation of (11) is that each participant of the chess tournament believes that there’ll only be a single winner, namely himself or herself. However, this must be obligatorily expressed as in (11) by the use of the plural pronoun. This follows from the assumption that the bound variable pronoun must exhibit agreement in number with the matrix subject. For (7), the idea would then be to say that focus on the pronoun can be inherited by some form of agreement with the antecedent similar to what is assumed in (11). This intuition is drawn from the observation that at least parts of the antecedent of the pronoun in (7a) are focused. However, there are numerous problems with this idea. Consider only the case in (12). Here, the agreement idea predicts too many antecedents to be possible for the focused pronoun: both QPs, every boy and every teacher in (12), are equal in their focus structure. Nevertheless, only the latter QP can be the antecedent of the focused pronoun HIS in (12). (12)
*Every BOYi called hisi mother before every TEAcherj called HISi mother.
A second problem with the agreement idea is that the focus is only optional while agreement is obligatory: if they in (11) is replaced with he the bound interpretation is impossible. And, a third problem is that the focus on the antecedent can be deeply embedded inside of the binder. For example, only Mary and Sue bear focus in (13), but the focused pronouns are bound by the subject DPs.
The silent content of bound variable pronouns
(13)
189
The boy MARY was dancing with called HIS mother, while the boy SUE was dancing with called HIS mother.
Now consider the second idea for an alternative explanation; that focus licensing in (7) makes reference to representations that do not contain any quantifiers. The idea would be that representations like (14) would exist for the sentences of (7) at this level. (14)
a. Student A called student A’s mother, and student B called student B’s mother, and . . . b. Teacher A called teacher A’s mother, and teacher B called teacher B’s mother, and . . .
While it may be possible to work out a proposal along these lines, one problem that would need to be addressed and which I do not know how to address, arises with sentences like (15) containing indefinites. (15)
John ate a banana, and Bill ate a banana.
Example (15) evokes a scenario where two different bananas, banana A and banana B, are involved. Hence, a quantifier-free representation of (15) in this scenario is (16). (16)
John ate banana A, and Bill ate banana B.
Since in (16) the objects of ate in the two conjuncts have different referents, this would seem to predict that focus on the indefinites in (15) should be just as acceptable as focus of the bound pronouns in examples like (7). However, focusing the indefinite a banana is very marginal, and has a special effect. Namely with such a focus, one is forced to assume that there are two different concepts of banana. This is a fairly marked interpretation in (15), but it does have natural uses, as in (17). (17)
There are books, and there are BOOKS.
One natural interpretation of (17) can be paraphrased as stating that there are (1) groupings of printed pages with a cover around and (2) objects that will change your, the reader’s, life with the wealth of wisdom they reveal. This interpretation seems to involve two concepts of book, and this is licensing the focus on the second occurrence of books in (17). If we put aside this phenomenon, (15) does not allow focus, and the idea of licensing focus by reduction to some quantifier-free form would need to explain why this is the case. Unless these problems can somehow be overcome, what we are left with seems to be the conclusion that occurrences of bound variable pronouns must be allowed to differ in their interpretation. In contrast to most other
190
Uli Sauerland
cases of focus, this difference in interpretation is not evident from their segmental phonology. 8.1.3
Finding bound elements
The result of section 8.1.2 has an interesting application: it can be used to test whether an expression is a bound variable or not. This is of particular interest in the case of complex anaphoric expressions the internal composition of which is opaque. Consider the contrast in (18) (Jacobson 2000 gives similar examples). The optional focus on a bound variable pronoun, in the case of the English reflexive himself, must be located on the pronominal part. (18)
a. John likes himself and Bill likes HIMself. b. *John likes himself and Bill likes himSELF.
Observe that in principle the self part can also bear focus, as shown in (19). Therefore, the contrast in (18) corroborates the idea that the English reflexive is semantically a complex expression consisting of a bound variable pronoun and a reflexivizer self that underlies, for example, the analysis by Reinhart and Reuland (1993). (19)
John likes Mary, but Bill likes himSELF.
Consider, furthermore, the reflexive sich of German. Sich can be optionally focused in (20). Since there is an antecedent with the same lexical VP available in (19), I conclude that the focus on sich is licensed only because sich is interpreted as a bound variable. (20)
Die the die the
Michaela hat sich fur ¨ den Posten vorgeschlagen, und Michaela has self for the office proposed, and TANJA hat SICH vorgeschlagen. Tanja has self proposed
Furthermore, there’s a contrast between (20) and example (21), where sich is the internal argument of the inherently reflexive verb ergeben (to surrender). (21)
# Die Michaela hat sich ergeben, und die TANJA hat the Michaela has self surrendered, and the Tanja has SICH ergeben. self surrendered
These observations corroborate the claim that sich is interpreted as a bound variable in (20), but not in the inherently reflexive (21) (Reinhart and Reuland 1993).
The silent content of bound variable pronouns
191
Finally, take a look at reciprocals. Neither the English examples in (22) nor the German examples in (23) allow the reciprocal to bear focus. (22)
*John and Mary like each other. Sue and Bill, however, like EACH other/each OTHER/ONE another/one ANOTHER.
(23)
*Michaela und Tanja haben einander vorgeschlagen. Peter und Michaela and Tanja have each other proposed. Peter and Dieter (hingegen) haben EINander/einANDER vorgeschlagen. Dieter (however) have EACH other/each OTHER proposed
Therefore, we can conclude that neither English each other and one another nor German einander have overt parts that are interpreted as bound variables. This seems to mesh well with analyses like Roberts (1991), where the reciprocal is analyzed as having a complex semantic structure that contains only silent bound variables. A nice contrast to (23) is (24), which does not contain the reciprocal, but rather the reflexive sich that is compatible with a reciprocal interpretation. In fact, the reciprocal interpretation is forced in (24) by use of the adverbial gegenseitig (mutually). (24)
8.1.4
Michaela und Tanja haben sich gegenseitig vorgeschlagen. Peter Michaela and Tanja have self mutually proposed. Peter und Dieter (hingegen) haben SICH gegenseitig vorgeschlagen. and Dieter (however) have SELF mutually proposed
Disambiguating strict and sloppy
A second application of the main result of this section is that it provides an explanation of facts observed by Hirschberg and Ward (1991). Their experimental finding is that the strict/sloppy ambiguity of pronouns in ellipsis is disambiguated by the presence or absence of a focus on the pronoun in the antecedent of ellipsis. Specifically, they observe a preference for the sloppy interpretation in examples like (25a) and a preference for the strict interpretation in examples like (25b). (25)
a. John likes HIS mother and Bill does too. [sloppy reading preferred] b. John likes his mother and Bill does too. [strict reading preferred]
Hirschberg and Ward’s (1991) observation can be seen to largely follow from the observation that bound pronouns can differ in interpretation. Note that this generalization predicts that, while (26) does not allow
192
Uli Sauerland
interpretation (26a) with coreferent pronouns, it does allow interpretation (26b) with variable binding. (26)
John likes his mother and Bill likes HIS mother. a. *Johnj likes hisj mother and Bill likes hisj mother. b. Johnj li i likes hisi mother and Bill li i likes i mother.
Therefore, I conclude that in (25a), too, the focus on the pronoun can only be licensed if both conjuncts contain a bound variable pronoun. The only difference between (25a) and (26) would then be that the order of conjuncts is the opposite, but we can assume that in this case the antecedent for focus licensing is accommodated. It follows that (25a) in isolation allows only the sloppy interpretation. This account still predicts that both the strict and the sloppy interpretation should be available for (25b), though Hirschberg and Ward (1991) observe that it is overwhelmingly used with the strict interpretation in mind. I propose that the observed effect in (25b) results from the setup of Hirschberg and Ward’s (1991) experiment, where subjects were specifically asked to disambiguate between the strict and sloppy interpretation. It is natural to speculate that in this scenario the absence of a cue disambiguating in one direction is taken as evidence for the opposite disambiguation. In this way, the disambiguation observed in (25) could be entirely explained as resulting from the observation that bound variables can be contrasted.
8.2
Indices
Standardly, occurrences of bound variable pronouns are taken to differ only in the indices they bear. It seems natural to make use of this difference to explain the focusability of bound variable pronouns (cf. Sauerland 1998; 1999). Consider the representation of (7) with indices, given in (27). (27)
Discourse: On Monday, every boy l1 called his1 mother. a. On TUESday, every TEAcher l2 called HIS2 mother.
The indexation shown in (27) could be sufficient to license focus if focus licensing applies to constituents in which the bound variable pronouns are not bound. Recall that Rooth (1992b) has already shown that the focus licensing conditions can apply sentence internally. Rooth uses examples like An AMERICAN farmer talked to a CANADIAN farmer to demonstrate this point. Consider now how this can apply to (27). For the choices of focus domain and antecedent in (28), the contrastiveness condition is
The silent content of bound variable pronouns
193
satisfied. The requirement imposed by contrastiveness for (28) is that [[FA]] 6¼ [[FD]], which indeed holds for certain assignments. (28)
a. FA ¼ his1 mother b. FD ¼ [HIS2]F mother c. FD0 ¼ his1 mother d. FD ¼ his2 mother ¼)9 g: [[FA]]g 6¼ [[FD]]g
Contrastiveness as stated in (5b) above cannot be applied to (28) since FA and FD both contain an unbound variable. A restatement of contrastiveness could in fact be done in such a way as either to make the resulting condition sensitive to differences in indexation, as in (29), or to make it not sensitive to such differences. Since the focus in (7) is licensed (see the account presented in Sauerland (1998; 1999)), the statement (29a) must be adopted. Then contrastiveness is satisfied in (28) because, for any g with different results for 1 and 2, [[FA]]g 6¼ [[FDs]]g . (29)
a. Contrastiveness (index-sensitive): for all Focus Alternatives FD0 of FD there is an assignment g such that [[FA]]g 6¼ [[FD0 ]]g b. Contrastiveness (index-insensitive): for all Focus Alternatives FD0 of FD and for every assignment g: [[FA]]g 6¼ [[FD0 ]]g
Independent evidence for favoring condition (29a) over (29b) comes from cases with focus on unbound pronouns (cf. Rooth (1992a)). In example (30), focus on the second occurrence of him is required unless it refers to the same individual as the first occurrence of him. In (30), noncoreference is indicated by contra-indexation. (30)
Isabelle knows himi. But she does not know HIMj.
The index-insensitive condition (29b) would wrongly predict that focus on the second occurrence of him should not be licensed, because any focus domain containing this pronoun will be identical in meaning to the corresponding antecedent under an assignment that assigns the same value to i and j. The analysis based on the index-sensitive contrastiveness condition (29a) successfully explains the examples considered up to now, and I present some further consequences of this account in Sauerland (1998; 1999). However, the account does not yet fully explain all the relevant facts, and needs to be amended. Section 8.2.1 shows that bound pronouns can even be contrasted if, in the domains that are considered, both are bound.
194
Uli Sauerland
8.2.1
Contrastiveness above the binder
There are problems in trying to reduce all cases of focus on bound pronouns to indexation. The account just presented depends on the application of contrastiveness to a focus domain that contains a bound pronoun but not its binder. However, focus particles like however and too can be used to determine the distribution of focus domains. The problem shown in this section is related to a basic property of indexed variables – the fact that two representations that are identical except for the index of a variable binder and all variables bound by it do not differ in interpretation. In this situation, the two representations are called alphabetic variants of one another. The effect of alphabetic variants can be illustrated for (27) by considering the FA and FD in (31), which include the binders of the two pronouns. For (31), the focus licensing conditions are not satisfied (cf. Rooth 1992a) because FA and FD are alphabetic variants and therefore do not differ in interpretation. (31)
a. FA ¼ l1 called his1 mother b. FD ¼ l2 called [HIS2]F mother c. FD0 ¼ l2 called his2 mother d. FD ¼ l2 called his2 mother
The equivalence of alphabetic variants predicts, therefore, that two occurrences of bound variables should not be able to contrast when the compared constituents include their binders. This prediction can be tested if there are means to control for the size of the compared constituents. I think that the focus sensitive particles however and too (as well as many similar expressions) provide these means, in particular the adnominal variants of these. As far as I know, no descriptive work on however has been done, but its essential properties can be captured quite easily. Consider the paradigm in (32): Adnominal however construed with the subject presupposes that both the subject and the VP differ in meaning with an antecedent.2,3 (32)
Discourse: Carl called Mary. a. JOHN, however, WROTE Mary. b. JOHN, however, called BERta. c. # JOHN, however, called Mary. d. # Carl, however, called BERta.
The contrasts of acceptability in (32) can be captured as a presupposition of however. For example, (32b) has the presupposition that there are X and
The silent content of bound variable pronouns
195
Y such that (a) X called Y, (b) X did not call Berta, and (c) John did not call Y. Given the discourse in (32), the presupposition is fulfilled with X ¼ Carl and Y ¼ Mary. Generalizing this picture a little, I propose that the presupposition for adnominal however is the following: (33)
[[[NP however] VP]] presupposes that there are focus alternatives NP0 and VP0 of NP and VP respectively such that: a. [[NP0 VP0 ]] ¼ 1, b. [[NP0 VP]] ¼ 0, and c. [[NP VP0 ]] ¼ 0.
The argument in the following is now based on the observation that (34a) is acceptable, where however is attached to the antecedent of the bound pronoun and the bound pronoun intuitively is contrasted with another occurrence of a bound pronoun in the discourse. As the unacceptability of (34b) without focus on the bound pronoun shows, the focus on the bound pronoun satisfies the presupposition of however. (34)
Discourse: Every teacher believes that she’ll win. a. Every GIRL, however, believes that SHE’ll win. b. # Every GIRL, however, believes that she’ll win.
It turns out that the observation in (34) is not predicted on the index based account of focus on bound pronouns. The indexed representation of (34a) is (35a). The focus alternatives licensing however should be the NP0 in (35b). For the VP, I consider the two possibilities in (35c) and (35d) – it is not clear whether VP00 should be available as a focus alternative, but I consider it here just for argument’s sake. (35)
a. [NP every GIRL] however [VP l2 believes that SHE2’ll win.] b. NP0 ¼ every teacher c. VP0 ¼ l2 believes that she1’ll win. d. VP00 ¼ l1 believes that she1’ll win.
For NP0 and VP0 , none of the focus licensing conditions are satisfied in the discourse in (34): Since she1 is an unbound variable, in (35c) the satisfaction of the three licensing conditions is not affected by the presence of the discourse antecedent in (34), but rather depends on what the assignment assigns to index 1. This is clearly not the desired result. The focus licensing conditions are not satisfied for VP00 in (35d) either: Since VP00 and VP are alphabetic variants, the presuppositions (33b) and (33c) of however are not satisfied for NP0 and VP00 . Therefore, the incorrect prediction made by the index based account is that (34a) should be just as unacceptable as (34b).
196
Uli Sauerland
An argument similar to the one with however can be made with adnominal too. The adnominal use of too is illustrated in (36). Adnominal too seems to presuppose that for a focus alternative to the subject, the VP is true. (36)
Discourse: Carl visited Mary. a. JOHN, too, visited Mary. b. *JOHN, too, visited BERta.
Applying Soames’s (1989) insights on the semantics of too to the adnominal case, I propose the semantics in (37). (37)
[[NP too VP]] presupposes that there is a focus alternative NP0 of NP such that a. [[NP0 VP]] ¼ 1.
Consider now the example in (38), which combines adnominal too with variable binding. As (38b) shows, the presence of too does not allow the bound variable to be focused. (38)
Discourse: Every teacher believes that she’ll win. a. Every GIRL, too, believes that she’ll win. b. # Every GIRL, too, believes that SHE’ll win.
The oddness of (38b) is not predicted by the index-based account of focus in such examples. Specifically, the indexed representation in (39) shows that the VP here is predicted to be true of the NP0 every teacher, and therefore the presupposition of too should be fulfilled in (38b). (39)
[NP every GIRL] too [VP l2 believes that SHE2’ll win]
In summary, this section has argued the following points. The main point has been to argue that indices can explain many cases of focus on bound pronouns, but that indices alone are insufficient to explain the focusability of bound variables. In addition to this negative result, we have also improved our working description of the phenomena over that of section 8.1. In section 8.1, I showed that two occurrences of a bound pronoun can differ in meaning. In this section, I have shown that the difference in meaning between two bound pronouns can be such that even otherwise identical constituents in which the bound variables are bound can differ in interpretation. Because of this, indexation alone cannot explain the focusability of bound pronouns, and in section 8.3 I argue for a new account based on the idea that bound pronouns have presuppositional content.
The silent content of bound variable pronouns
8.3
197
Bound pronouns have content
The account I argue for in this section assumes that bound variable pronouns have actual semantic content in addition to the variable index. This content must be interpreted as a presupposition of the bound variable. This claim is novel for bound variable pronouns in this form, however, there are two well-known cases where pronouns have descriptive content beyond their index. For one, bound variable pronouns actually can have some overt semantic content, namely person, gender, and number marking. This content is actually interpreted (at least in many cases), as example (40) illustrates. The feminine marking on she in (40) has the effect that (40) is understood as a generalization about female teachers. (40)
Every teacher thinks shefeminine is brilliant.
Following Cooper (1979) and others, I assume that the content of the bound pronoun is interpreted as yielding a presupposition on the denotation of the bound variable. An interpretation rule for the feature [feminine] to this effect is (41) (see also Sauerland 2003). (41)
[[proi [feminine]]]g presupposes that gðiÞ is female. When defined: [[proi P]]g ¼ gðiÞ
Following what is known about presupposition projection from the scope of universal quantifiers (Heim 1983; and others) this predicts that (40) has a presupposition that all teachers are female. My claim is that not only overtly realized features such as feminine in (40), but also other, silent presuppositional content can occur with bound variables. Silent content has been observed already for pronouns other than bound pronouns: Evans (1977) first claimed that some occurrences of pronouns can be hidden definite descriptions. These are the so-called e-type pronouns. Therefore, it is natural to assume that even bound variable pronouns may be e-type pronouns. However, bound pronouns also bear an index. Therefore, this requires the interpretation rule for bound definite descriptions in (42). (42)
[[thei P]]g presupposes that P(gðiÞ)=1. When defined [[thei P]]g ¼ gðiÞ
Obviously, an important question is to characterize what exactly the unpronounced presupposition can be. It is quite apparent that the bound presupposition must always be satisfied, since it could never be accommodated successfully. For the discussion in the following, I start with the assumption that the restrictor of the antecedent is identical to the content
198
Uli Sauerland
of the bound pronoun. As we’ll see, however, it will be necessary to adjust this assumption and to allow any presupposition that is satisfied in the interpretation of the sentence within the current discourse context. 8.3.1
Account of focus licensing
How does the assumption that bound pronouns can have semantic content explain the possibility of focus on a bound pronoun? Consider again (43) (repeated from (7)) with focus on the bound pronoun. (43)
Discourse: On Monday, every boy called his mother. a. On TUESday, every TEAcher called HIS mother.
The new account allows the representations in (44), where both bound pronouns – that in the discourse antecedent and that in the target sentence – are bound definite descriptions. (44)
Discourse: every boy l1 called [the1 boy]’s mother a. every TEAcher l2 called [the2 teacher]’sF mother
One selection of FA and FD from (44) that satisfies the focus licensing conditions is given in (45). (45)
a. FA ¼ l1 called the1 boy’s mother b. FD ¼ l2 called [the2 teacher’s]F mother c. FD0 ¼ l2 called [the2 boy’s]F mother d. FD ¼ l2 called the2 teacher’s mother
In particular, contrastiveness is satisfied because FA and FD denote functions with different domains: (46)
a. [[FA]] ¼f where f: fx: teacher(xÞ ¼ 1g7!f0; 1g, . . . b. [[FD]] ¼g where g: fx: boy(xÞ ¼ 1g7!f0; 1g, . . .
Since the FA and FD considered in (45) contain the binders of the bound pronouns, it has been shown here that the new account predicts a difference in meaning between these constituents. This also explains why the use of adnominal however, which was a problem for the index-based account in the previous section, is licit. 8.3.2
A-traces and pronouns
One prediction of my proposal arises from what is known about traces. It has been argued that traces are syntactically and semantically definite
The silent content of bound variable pronouns
199
descriptions, with unpronounced parts (Chomsky 1993; Sauerland 1998; Fox 1999b; Sauerland 2004). -traces should be able to license Therefore, my proposal predicts that A destressing of pronouns, as Danny Fox (p.c.) first pointed out to me. -bar trace: Moreover, this should depend on the lexical material in the A the destressing of the pronoun should be possible if and only if the antecedents are lexically the same (see also section 8.3.4). This prediction is borne out by (47) (after Fox, p.c.) and (48). (47)
a. I saw [every picture of every man who wanted me to see #him/ HIM today]DP1. b. I saw [every picture of every man who wanted me to see it/#IT today]DP1.
(48)
a. Every studenti beat every teacherj who expected that shei would beat herj. b. Every studenti beat every teacherj who expected that SHEj would beat HERi.
Consider the LF-representation of (48b) in (49). Since the example exhibits antecedent contained destressing, I assume that QR of the object is required. Therefore, the FA in (49) contains two traces, the trace of the subject and the QR trace of the object. These traces contrast with the two focused pronouns in the FD. (49)
[every student] λ1 [every teacher [λ 2 t2 expected that [the1 teacher]F beat [the2 student]F]] FD
λ 2 [the1 student] beat [the2 teacher] FA 8.3.3
Non-conservative quantifiers
Does the e-type representation of bound pronouns ever have an effect on the semantics? Consider again the two alternative representations proposed for bound pronouns in (50), where Q may be any quantifier. (50a) and (50b) do not in general have the same interpretation – the scope of Q in (50b) presupposes teacherhood.
200
(50)
Uli Sauerland
Q teacher called his mother. a. Q teacher l1 called x1 ’s mother b. Q teacher l1 called the1 teacher’s mother
For conservative quantifiers Q, the presupposition of (50b) is always satisfied -traces). With non-conservative quantifiers, however, (cf. Fox 1999b on A representation (50b) should result in a presupposition failure. I claim that the predicted effect can be demonstrated in examples with bound indexical pronouns. Heim (1994) first observed that examples with only show that indexicals in English can be bound. Consider the ambiguity of example (51): the two readings arise depending on whether the second occurrence of you is bound by the first or whether it is coreferent with it, but not bound. (51)
Only you brought something you like. a. Only you l1t1 brought something x1 liked (bound) entails: Nobody else brought something he likes. b. Only you l1t1 brought something you like (coreferent) entails: Nobody else brought something you like.
A similar ambiguity is found in (52) with the indexical I. (52)
Only I know when I arrived.
My argument is based on the new observation that (51) and (52) are disambiguated by focus. A focused indexical cannot be bound, as shown by (53a). A destressed indexical, on the other hand, prefers the bound interpretation as (53b) shows. The same observations hold for (52), except that the coreferent interpretation seems to be even more easily available when the second occurrence of I is not focused than in (53b). (53)
a. Only YOU brought something YOU like. (coreferent, *bound) b. Only YOU brought something you like. (bound, ??coreferent)
Further evidence is provided in (54): In (54a), the bound reading is blocked by focus on you. (54)
Discourse: Everybody else likes all his colleagues. a. Only YOU have colleagues you/#YOU can’t stand.
Consider the representation for (53) in (55) with a bound e-type you, which is required for focus licensing4 – I assume here that you has an interpretation as a predicate true only of the person addressed. (55)
Only you l1t1 brought something [the1 you] like
The silent content of bound variable pronouns
201
Since [the1 you] presupposes that gð1Þ ¼ [[you]], the scope of the quantifier only you in (56) is a function with a singleton set as its domain – therefore, (55) is either trivially true or a presupposition failure. Intuitively, the sentence (53a) with focus on you is paraphrasable as the tautologous: Only you are you and brought something you like. Plausibly, this tautologous interpretation is not considered to be available when judging (53a) and therefore only the coreferent interpretation is available.5 (56)
8.3.4
[[l1 t1 brought something [the1 you] like]] ¼ f with domainðf Þ ¼ fthe addresseeg
The antecedent effect
More support for the claim that bound pronouns may have hidden content comes from the following observation: If the antecedents of the two bound pronouns are identical, the second pronoun cannot be focused. Irene Heim (p.c.) first hinted at this observation. This is illustrated in (57) and (58). (57)
Discourse: I did not expect every teacher to get what she wanted. a. # But every teacher GOT what SHE wanted. b. In the end, every GIRL got what SHE wanted.
(58)
Discourse: Did every boy call his mother? a. # Yes, every boy called HIS mother. b. No, every TEAcher called HIS mother.
This antecedent effect follows from the main proposal of this section. Consider the representations for (58) in (59). (59)
Every boy l1t1 called the1 boy’s mother a. Every boy l2t2 called the2 boy’s mother b. Every teacher l2t2 called the2 teacher’s mother
For the FA and FD in (60), which are analogous to the domains considered in the analysis of (44) above, contrastiveness is violated, because FD and FA are identical in meaning. (60)
a. FA ¼ l1t1 called the1 boy’s mother b. FD ¼ l2t2 called [the2 boy’s]F mother c. FD0 ¼ l2t2 called [the2 boy’s]F mother d. FD ¼ l2t2 called the2 boy’s mother
The demonstration that the focus licensing condition is not satisfied for a particular choice of FA and FD is of course not sufficient to explain the
202
Uli Sauerland
impossibility of focus: it needs to be shown that for every permissible choice of FA and FD, the focus licensing condition is not satisfied. In particular, section 8.2 showed in detail that differences in indexation license focus if a focus domain that does not include the binder is considered. This prediction would not arise if the index-insensitive focus licensing condition (29b) is adopted, but I argued above that the index-sensitive focus licensing condition should be adopted. Therefore, I conclude that a focus domain lower than the l-operator binding the pronoun is not possible in (57) and (58). Then the analysis of bound pronouns as e-type pronouns accounts for the antecedent effect. The antecedent effect allows us to determine more precisely the question of what the content of the bound pronoun may be. Consider the following cases. 8.3.4.1 Different extensions As Orin Percus (p.c.) first pointed out to me, sometimes it is sufficient that the extensions of the two antecedent restrictors differ to license focus on a bound pronoun. (61)
Discourse: Did every flight leave at the time it was scheduled for on Tuesday? a. All I know is that, on Wednesday, every flight left at the time IT was scheduled for.
This observation shows that in these cases the bound variable pronoun cannot just have as its content the restrictor of its antecedent since these are identical in (61) and therefore could not license the focus. Rather, in these cases the content of the bound pronouns seems to be flight on Tuesday contrasting with flight on Wednesday. A possibly related observation was made independently by Orin Percus (p.c.) and Dimitriadis (2001). They note that examples like (62) are acceptable only if John and his brother have different mothers. (62)
John called his mother and John’s BROTHER called HIS mother.
This effect seems to be specific to possessives. An explanation of it could be to assume that the apparent focus on his in (62) is actually a focus on the entire DP his mother, and that all referential DPs, not just pronouns may have presuppositional content (see also example (70) below). 8.3.4.2 Different quantifiers Does the quantifier of the antecedent matter for the focusability of a bound pronoun? In the following examples, at least the quantifier seems to be not relevant, as predicted.6
The silent content of bound variable pronouns
(63)
203
Discourse: I expected no student to call his mother. a. But EVERY student called his/#HIS mother. b. But at least one student called his/#HIS mother.
(64)
Discourse: Did every student call his mother? a. No, NO student called his/#HIS mother. b. All I know is that at least one student called his/#HIS mother.
8.3.4.3 Overlap Is the antecedent effect observed if the restrictors of the two antecedent quantifiers overlap? It seems that focus is licit in case of overlap ((65b)) unless a sub- or superset relation ((65a) and (66a)) holds. However, the judgments are quite subtle. (65)
Discourse: Did every young student call his mother? a. In fact/No, EVery student called his/#HIS mother. b. All I know is that every BLOND student called his/HIS mother.
(66)
Discourse: Did every student call his mother? a. All I know is that every YOUNG student called his/#HIS mother.
Definitely the effect in (66) is expected because it is well established that entailments from the preceding discourse can license focus and destressing (Tancredi 1992). Sometimes this is called Implicational Bridging. In (66), the discourse entails the question Did every young student call his mother, and therefore the example is expected to exhibit the antecedent effect.
8.3.5
Classical e-type pronouns and focus
The proposal that pronouns have descriptive content that is interpreted as a presupposition is, of course, not new: Evans (1977) proposes that there is a class of pronouns, e-type pronouns, that are semantically definite descriptions. New, however, is my proposal that even bound variable pronouns can be e-type pronouns. I believe that the two proposals are closely related, though the lines of argumentation are different: the classical evidence for the e-type analysis of pronouns comes solely from the available interpretations of sentences with pronouns. For example, Evans and others argue that on the salient interpretations of (67a) and (67b), they must be analyzed as a definite description: The pronoun they in (67a) might be understood as the congressmen that voted for the bill and the pronoun it in (67b) as the donkey he owns.
204
(67)
Uli Sauerland
a. Few congressmen voted for the bill. They were very junior. b. Every farmer who owns a donkey beats it.
Since I have argued that focus provides evidence for an e-type analysis of bound variable pronouns, it is natural to ask the following: does focus also provide evidence for the e-type analysis of such bona fide e-type pronouns as those in (67)? Indeed, this is the case, as (68) shows. (68)
a. Few congressmen voted for the bill and they were very junior, but most SENATORS voted for the bill and THEY were all SENIOR. b. Every farmer who owns a donkey beats it, but every farmer who owns a HORSE beats IT.
In both examples, there is a second pronoun that receives contrastive focus. I propose that the two e-type pronouns are contrasted with each other and that the contrast is due to the content of the two pronouns. The second occurrence of they in (68a) would be analyzed as the senators who voted for the bill and therefore contrast with the first occurrence of they, which is understood as the congressmen who voted for the bill. Similarly, the focused it in (68b), if understood as the horse he owns, contrasts with the donkey he owns. At this point, a further, more ambitious prediction my analysis makes comes to mind. Namely, the prediction that contrastive focus on an e-type pronoun should be obligatory in examples like (68). Indeed, this prediction is borne out in (68a). Donkey anaphora, as in (68b), however, do not seem to bear out this further prediction. On closer consideration, though, the behavior of donkey anaphora comes as no surprise. Over the last twenty-five years, various semantic mechanisms have been developed and independently motivated that account for donkey anaphora with less content in the relevant pronoun than Evans’s original proposal (Kadmon 1987; Heim 1990; Chierchia 1995; Lin 1996; Elbourne 2001). For example, the pronouns in (68b) could possibly be understood as the same definite description, the animal he owns, as in the paraphrase (69). (69)
Every farmer who owns a donkey beats the animal he owns, and every farmer who owns a HORSE beats the animal he owns.
I believe that the study of focus in this context can help to determine what precisely the content of the pronoun is. In Sauerland (2000), I presented several examples where, in my judgment and those of my informants, the focus does seem to be obligatory. However, a full discussion of further predictions concerning donkey anaphora is beyond the scope of the present paper.
The silent content of bound variable pronouns
8.4
205
NP ellipsis or NP-pro?
In section 8.3, I argued that in addition to the classical cases of e-type pronouns, bound variable pronouns too can be e-type pronouns. Semantically, e-type pronouns are analyzed as having some silent content that is presupposed of the referent of the pronoun. I have shown that focus can indicate what this silent content may be. In this section, I want to address the question of how this silent content of the pronoun is represented. For classical e-type pronouns, two proposals have been made for the representation of the silent content. Evans (1977; 1980) proposes that the content of the pronoun is syntactically represented and, in fact, develops LF-copying rules to account for this. While Evans himself does not explicitly draw a parallel between these rules and VP Ellipsis, Heim (1990) and Elbourne (2001) develop his proposal in this direction. Cooper (1979), on the other hand, proposes that the content of e-type pronouns does not have such an explicit syntactic representation, but rather is a silent relation variable inherent to the pronoun (see also Heim and Kratzer 1998). This proposal is, at least superficially, more flexible than the ellipsis proposal since the mechanism establishing the value of the relation variable remains unspecified.7 Prima facie, the NP Ellipsis analysis is preferable as it is reductionist – reducing the question of licit pronoun content to general ellipsis licensing – and more restrictive. Furthermore, the arguments for the e-type analysis offered by Heim (1990) and Elbourne (2001) have to my knowledge not been repudiated. Therefore, one would hope that bound e-type pronouns too were amenable to an ellipsis analysis. However, in this section, I show that the facts with e-type pronouns occurring as bound variables provide some evidence in favor of a more flexible proposal. I present two arguments, the first of which is that focus licensing does not follow in a straightforward way from the ellipsis analysis since in corresponding examples with VP Ellipsis focus is not licensed. The second argument is that in some cases an analysis in terms of ellipsis seems to be impossible since there is no available antecedent. Neither of the two arguments rules out the possibility that there is also an elided NP in pronouns in some cases. In fact, we find cases similar to those with bound pronouns with definite determiners in German, which must receive an analysis involving NP Ellipsis, as Wiltschko (1998) argues. For cases like (70), I therefore assume that the definite determiners have the elided NP complements Jungen and Mann, but also have Cooper’s inherent relation argument and that it is the latter that accounts for the possibility of focus.
206
(70)
Uli Sauerland
Eine Schwester jedes Jungen hat den angerufen und eine a sister every boy has the called and a Schwester jedes MANNES hat DEN/? den angerufen. sister every man has the called. ‘‘A sister of every boy called the boy and a sister of every man called the man.’’
The judgment is in fact, if anything, sharper with bare definite determiners than with pronouns: In (70), it is not fully acceptable to destress the second occurrence of den while in the examples with pronouns considered in previous sections there was often true optionality. Example (71) adds to this picture the observation that den can be destressed – (71) is also an illustration of the antecedent effect with definites. (71)
8.4.1
Eine Schwester jedes Jungen sollte den anrufen. Also WIRD a sister every boy should the called hence will eine Schwester jedes Jungen den angerufen haben. a sister every boy the called have. ‘‘A sister of every boy was supposed to call the boy. Hence, a sister of every boy will have called the boy.’’
Focus placement
This section develops an argument from the placement of focus. Schwarzschild (1999) argues that focus must be placed on the smallest constituent possible. For example, focus on part A of a phrase [A B] is preferred over focus on the entire phrase. This, however, raises a problem for the proposal that the content of an e-type pronoun is represented syntactically: on this analysis, the structure of the pronoun would be [pronoun NP], where NP is different from the antecedent. Therefore, focus should be placed on the elided NP rather than on the pronoun or the entire phrase. One might think of slight modifications to Schwarzschild’s proposal, to rescue the ellipsis proposal. However, Schwarzschild’s proposal makes exactly the right prediction for VP Ellipsis, as shown by the contrast in (72). The elided material is shown using strikeout. (72)
a. When I talk, you say I should not talk, and when I keep quiet, you also say I should not keep quiet. b. # When I talk, you say I should not talk, and when I keep quiet, on the other hand, you say I SHOULD not keep quiet.
The examples in (72) are of a type first discussed by Schwarz (2000): an elided VP appears in a corresponding position to another elided VP
The silent content of bound variable pronouns
207
with different content. Hence, the examples in (72) are exactly analogous in structure to the ellipsis analysis of pronouns, which is exemplified in (73). (73)
Every boy called his boy mother, every TEACHER called HIS teacher mother.
However, VP Ellipsis in (72) does not allow focus on the head that takes the elided complement. This is predicted by Schwarzschild’s analysis because the elided VP is assumed to have internal constituent structure, and a focus on the constituent should not keep quiet is, therefore, not placed on the smallest constituent possible. Rather, the focus should be placed on the constituent keep quiet, and indeed we find this focus obligatorily when the elided VP is pronounced: (74)
When I talk, you say I should not talk, and when I keep QUIET, on the other hand, you say I should not keep QUIET.
In fact, examples with bona fide NP Ellipsis behave just like VP Ellipsis. Consider, for example, (75). (75)
When many boys play, one boy ends up crying. a. *When many GIRLS play, however, ONE girl ends up crying.
The ellipsis structure (73) could not explain the focus found on bound variable pronouns unless some explanation was found for the difference between this case and VP Ellipsis in (72) and NP Ellipsis in (75). 8.4.2
Bound pronouns in the restrictor
A second problem for the ellipsis analysis was brought to my attention by Pauline Jacobson (p.c.) (see also Jacobson 2000). She provided me with the example (76), which allows the bound variable pronoun his in the second relative clause to receive contrastive focus.8 (76)
Every man who loves his mother talked to every man who HATES HIS mother.
The focus in (76) must be licensed by a contrast with the bound pronoun in the first relative clause. For example, we might analyze the two pronouns as having the content the man who loves his mother contrasting with the man who hates his mother to explain the focus in (76). In contrast to the examples in section 8.1 and elsewhere, however, the bound pronouns in (76) both occur in the restrictor of a quantifier. Moreover, the contrast between the two is established only by other material in the relative clauses that the pronouns themselves occur in.
208
Uli Sauerland
Therefore, an ellipsis account of (76) faces the problem of antecedent containment: the plausible antecedent for ellipsis licensing man who hates his mother contains the putative ellipsis site, the pronoun his. Antecedent containment is known to occur in the case of VP Ellipsis as well (Sag 1976 and others), and there it has been seen that covert movement (e.g. Larson and May 1990) or string-vacuous movement (e.g. Fox and Nissenbaum 1999) resolves antecedent containment. An analogous treatment of (76), however, seems hard to swallow: the proposal would be that the pronoun his moves to adjoin to the NP man who loves/hates x’s mother. In example (77), a similar movement analysis would even have to cross two relative clause boundaries. (77)
Every man who found a girl who loved his mother talked to every man who found a girl who HATED HIS mother.
By contrast, it is well established that the resolution of antecedent containment in VP Ellipsis is subject to strong locality constraints (e.g. Jacobson chapter 3 in this volume and references there). This further difference between VP Ellipsis and the silent content of pronouns hence provides another argument against an ellipsis analysis of the latter. In sum, then, I conclude that the NP Ellipsis analysis cannot straightforwardly apply to bound e-type pronouns as things stand. These facts could be taken to argue in favor of a revival of the NP-pro analysis of e-type pronouns. However, the direction I would prefer to see pursued in the future is to develop a better understanding of focus placement to deal with the first placement and ellipsis, where ellipsis is a little more flexible to deal with the second problem. 8.5
Conclusion
In this chapter, I have postulated three new theses concerning pronouns, in particular bound variable pronouns, and raised one open issue. The structure of the chapter has been layered in the following sense: I started with the most basic and least controversial thesis. I then argued, based on the already established thesis in the following section for a more specific thesis. One benefit of this structure is that even if the reasoning that has led me to a more specific thesis in a later section were shown to be wrong, still the more basic theses established up to that point would still be supported. Consider now each of the theses of my argument sequence. Thesis one is that bound variable pronouns can differ in their interpretation. My evidence for this assumption has come from examples like (78), where the bound pronoun seems to bear contrastive focus.
The silent content of bound variable pronouns
(78)
209
Every boy called his mother. Every teacher, however, called HIS mother.
This result is interesting because it argues against the proposal of Jacobson (1999) and others that bound variable pronouns always denote the identity function. Furthermore, the result can be applied as a test for whether a phrase is interpreted as a bound variable. Thesis two is that bound variable pronouns can cause a difference in meaning even for constituents in which they are bound. The argument I presented for this thesis in section 8.2 was based on the licensing of however in (78). Assuming that however takes two arguments, as shown in (79), and is licit only if both arguments differ in meaning from the relevant antecedent, the thesis is supported. (79)
however (every teacher) (lx x called HISx mother)
This result is important because it establishes that differences in indexation alone are insufficient to account for the contrastiveness of bound variable pronouns. Thesis three is that bound variable pronouns can have presuppositional content. This proposal accounts for the licensing of however in (79) because its argument would be analyzed as (80), which is a function that has as its domain the set of teachers, while the antecedent denotes a function that has the set of boys as its domain. (80)
lx x called [the teacher x]’s mother.
In section 8.3, I gave three additional arguments for this analysis: cases -traces mean the same, the analysis of where bound pronouns and A variables bound by non-conservative quantifiers, and the effect the antecedent has on whether focus on a bound variable is possible or not. The remaining problem I discussed in section 8.4 concerns the origin of the new-found content of pronouns. I argued that it is desirable to view the result I achieved in line with analyses of classical e-type pronouns. However, while an NP Ellipsis analysis for classical e-type pronouns has been widely accepted, the NP Ellipsis analysis faces some new challenges with bound e-type pronouns. I concluded that the NP Ellipsis analysis needs to probably treat NP and VP Ellipsis in slightly different fashions, but offered no proposal as to how this is to be spelled out.
9
A step-by-step guide to ellipsis resolution Satoshi Tomioka University of Delaware, Newark, Delaware, USA
9.1
Outline of the problem
The topic of this chapter is how VP Ellipsis sentences, such as (1), are represented in syntax. (1)
When John had to cook, he didn’t want to [VP].
It is undeniable that an elided VP demands the presence of the ‘‘same’’ VP or VP meaning, but defining the ‘‘sameness’’ is no easy task. One of the two main approaches to this issue is the proform approach, according to which the missing VP is represented as some kind of silent pronoun whose semantic value is retrieved via a meaning recovery strategy similar to nominal anaphora resolution. The advocates for this thesis include Bach and Partee (1980), Rooth (1981), Klein (1986), and most recently Hardt (1993; 1999). The other approach, which I call the LF Structure approach, posits full-fledged syntactic structure for the missing VP at the level of Logical Form (LF). Within this approach, there are two variants. In the ‘‘PF-Deletion’’ analysis, a missing constituent begins its life in the derivation with fully represented structure, which will be deleted in the phonological component but remain intact at LF (cf. Chomsky and Lasnik 1993; Fox 2000). The other variant, including Fiengo and May (1994), treats the missing material to be truly missing at the beginning but copied from its antecedent at LF. The debate between the Proform and the LF Structure camps has a long history and is still ongoing. The purpose of this chapter is not to provide the full review of this debate but to concentrate on a Acknowledgments: Many thanks to Benjamin Bruening, Danny Fox, Hajime Hoji, Kyle Johnson, Chris Kennedy, Yuki Takubo, and Ayumi Ueyama, and most of all, Jason Merchant for comments and discussions. Earlier versions of this chapter were presented at The Third Annual Kaken Workshop on Ellipsis at Kyushu University, Ellipsis Workshop at Tubingen University, and GLOW-Asia at Seoul National University. I would like to thank ¨ the participants in those events for comments and suggestions, in particular Akira Watanabe and Nina Zhang. None of these individuals are responsible for all the remaining errors. A shorter version of this chapter appeared as Tomioka (2004), and I hope that the current chapter overcomes some of the problems and shortcomings that the previous version had.
210
A step-by-step guide to ellipsis resolution
211
particular phenomenon which is considered by many to be one of the most convincing arguments in favor of the proform approach. Hardt (1999) and Schwarz (2000) independently noted that, when an elided VP is contained within another elided VP, the first VP can get the ‘‘sloppy’’ interpretation. Here are the examples that Hardt and Schwarz present. (2)
A. When John had to cook, he didn’t want to D. B. When he had to clean, he didn’t D either. (Schwarz 2000: Chapter 4, (35))
(3)
I’ll help you if you want me to D. I will kiss you even if you don’t D. (Hardt 1999: (17))
The elided VP in B’s utterance in (2) is understood to mean ‘‘want to clean,’’ and similarly, the missing VP in (3) is interpreted as ‘‘want me to kiss you.’’ For a theory of ellipsis which assumes full-fledged LF structure for elided material, these examples are problematic. The second VP Ellipsis which contains another elided VP fails to satisfy the LF identity condition. Take (2), for instance. Its LF representation would look like (4), in which the material which has been striked out corresponds to the unpronounced VPs. (4)
A. When John had to cook, he didn’t want to cook. B. When he had to clean, he didn’t want to clean, either.
The first VP Ellipsis (i.e. in A’s utterance) can be successfully represented as [VP cook] since there is an identical VP in the preposed adjunct clause. However, the second VP Ellipsis creates a problem. The sloppy interpretation requires that the elided VP be of the form [VP want to clean]. However, the preceding sentence furnishes no such VP. Intuitively, the second VP in A’s utterance, [VP want to D], should count as an antecedent, but its LF representation, [VP want to cook], would not be considered as identical to the missing VP under any theory of LF identity. The second difficulty, pointed out by Schwarz (2000) is that the sloppy interpretation is not possible when the first sentence does not have ellipsis. In (5), for instance, the missing VP is only understood to mean ‘‘want to cook’’ although this ‘‘strict’’ reading is rather odd pragmatically. (5)
A. When John had to cook, he didn’t want to cook. B. When he had to clean, he didn’t, either. (Schwarz 2000: (37))
As far as the LF representations are concerned, (4) and (5) should be identical. Therefore, the LF structure approach fails to distinguish them, leaving the contrast completely unaccounted for.
212
Satoshi Tomioka
What can the proform approach say about this? The answer depends on what kind of pronominal element is assumed for VP Ellipsis. Hardt (1999) argues, extending Muskens’s (1996) dynamic logic system to VP anaphora, that all instances of VP Ellipsis be treated as silent dynamic property anaphora. Schwarz (2000) is considered as a mix of the proform and the LF structure approaches. While he maintains the possibility that elided material has full structure, he argues that those instances that yield sloppy interpretations must be represented as silent VP pronouns. Hardt’s theory must be evaluated in the light of all the phenomena which arguably suggest that VP Ellipsis requires internal syntactic structure (e.g. an elided VP’s capability of hosting a wh-trace, subject agreement in VP Ellipsis in the there-insertion construction, etc.). On the other hand, Schwarz’s idea may be more modest but seems to have the best of the two approaches. For this reason, I will focus on Schwarz’s analysis and use it as the starting point of this chapter. In section 9.2, I will first review Schwarz’s analysis and demonstrate how it accounts for the presence and absence of sloppy interpretation of ellipsis within ellipsis. Despite its success, however, Schwarz’s account encounters several problems of its own, some of which were already anticipated by Schwarz himself. In section 9.3, I offer an LF structure account, in which the LF identity of an elided VP is checked in a ‘step-by-step’ fashion. I will make a crucial use of e-feature marking which is derivationally interpreted at PF and at LF. A reinterpretation of Merchant’s (2001) theory, to which I owe the idea of e-feature marking, is presented as a comparison. Section 9.4 begins with another problem. Even if the LF identity problem of sloppy ellipsis is successfully circumvented, it appears that the new LF identity condition makes a condition on focus unworkable unless sloppy ellipsis is treated as an instance of a bound variable. I argue that the problem can be avoided by adopting Kratzer’s (1991) analysis of focus which employs focus indexing and designated variables. 9.2
VP ellipsis as silent bound variables
Sloppy readings of nominal anaphora have been regarded as the indication that they are construed as bound variables (cf. Lasnik 1976; Reinhart 1976; Sag 1976; Williams 1977), and Schwarz analogizes the sloppy VP Ellipsis cases to nominal anaphora. Specifically, he suggests that a missing VP can, but not necessarily must, be represented as a phonologically silent VP pronoun, and that such a pronoun can be bound by an overt VP which undergoes LF fronting. Let us illustrate his analysis with the example (2).
A step-by-step guide to ellipsis resolution
(2)
213
A. When John had to cook, he didn’t want to. B. When he had to clean, he didn’t either.
Under the sloppy interpretation, the sentence has the following LF. (6)
A. [VP cook]3 [ When John had to t3, he didn’t want to D3] D4], either. B. [VP clean]4 [When he had to t4, he didn’t want to 7
What we see in the second instance of VP Ellipsis is a mixture of the proform and the LF structure approaches, in which the smaller VP is represented as a bound variable VP and the larger VP has structure associated with the lexical verb want. With the LF representation (6), the two VPs, want to D3 and want to D4, can be considered LF identical under a certain theory of ellipsis (such as Rooth 1992a; Heim 1997; Tomioka 1997; Romero 1998) which states that the LF identity condition demands the lexical items and the structure to be identical between an elided VP and its antecedent but is oblivious to indexical differences. Want to D3 and want to D4 use the same lexical items and have the identical structure with the only difference being the indices on the silent VPs. Hence, they satisfy the LF identity condition. The proposal also provides an account for the lack of sloppy interpretation in (5). (5)
A. When John had to cook, he didn’t want to cook. B. When he had to CLEAN, he didn’t, either.
Imagine that the sentences in (5) undergo LF VP movements comparable to the ones in (6). The resulting LF configuration is (7). (7)
A. [VP cook]3 [ When John had to t3, he didn’t want to cook] B. [VP clean]4 [When he had to t4, he didn’t want to 7 D4], either.
Unlike in (6), the elided VP [want to D4] fails to find an appropriate antecedent. One of the VPs in A’s utterance, want to cook, cannot be considered LF equivalent to the elided VP. Hence, the sloppy reading is blocked. While the Bound Variable VP theory of Schwarz successfully accounts for the sloppy reading of VP Ellipsis as well as the absence of such a reading in the non-ellipsis contexts, it faces several challenges. First of all, this analysis presupposes that the antecedent of elided material can scope out to a position high enough to c-command the ellipsis site where a bound variable sits. As Schwarz himself admits, however, such a movement seems oblivious to the stricter-than-expected constraint (i.e., clauseboundedness) on QR. In the LF (6), for instance, the overt VP must scope out a tensed clause. Moreover, this movement would violate an island
214
Satoshi Tomioka
constraint since it comes out of an adjunct CP. Apart from this mystery of a locality condition (or the lack thereof), the LF raising of a VP is not completely unintuitive. A VP can be fronted overtly in English, as shown in (8a), and Johnson (2001) indeed suggests that VP Ellipsis is an instance of VP topicalization where the topic VP becomes unpronounced; a process akin to the null topic construction in German, exemplified in (8b), where we find the V1 order instead of the canonical V2. (8)
a. Mark always said he would move to LA, and [move to LA]1, he finally did t1. b. D habe ich schon gehort. ¨ D have I already heard ‘‘I’ve already heard (about) it.’’
However, the availability of movement and that of sloppy interpretation of ellipsis do not always coincide. Sloppy interpretations are also available with NP Ellipsis (cf. Elbourne 2001) and with Sluicing, as shown below. (9)
NP Ellipsis within VP Ellipsis a. Everyone who arrested some murderers insulted a few, and everyone who arrested some burglars did D, too. (D = insult a few (of the) burglars) (adapted from Elbourne 2001: (109)) b. Speaking of syntacticians, of course Ken knows many, but semanticists . . . I don’t know. My guess is that he doesn’t D. (D = know many semanticists)
(10)
Sluicing within VP Ellipsis If Fred IS marrying someone, we want to know who, but if he (just) MIGHT be marrying someone, we don’t D. (D = want to know who he might be marrying) We have had too many false alarms.
If these sloppy ellipsis cases were to be analyzed in the way analogous to the VP Ellipsis cases, it would be necessary to posit LF movement that has been unattested, either in the overt syntax or at LF; movement of an NP out of a DP and movement of an IP out of a CP.1,2 The bound variable analysis also predicts that the occurrence of sloppy ellipsis should be regulated by the same structural constraints that are imposed on nominal variable binding. One of such constraints is Weak Crossover. It turns out, however, that sloppy ellipsis is not subject to Weak Crossover. Consider (11). (11)
If you tell me to, I will gladly quit drinking, but even if the Queen did D, I would never quit smoking! (D = tell me to quit smoking)
A step-by-step guide to ellipsis resolution
215
To obtain the bound variable interpretation, this sentence must have the following LF. (12)
[[VP quit drinking]1 [[if you tell me to]1] I will gladly t1]]], but [[VP quit smoking]2 [[if the Queen did {tell me to D2}] I would never t2]]]
This is, however, a kind of configuration that causes a Weak Crossover violation for variable binding, as witnessed in (13).3 (13)
*If he reads it1, Bobby criticizes [every paper]1.
The availability of a sloppy reading in (13), therefore, indicates that its source is not the bound variable nature of the elided VP. The conclusion we can draw at this point is that if a sloppy reading inducing VP is pronominal, it is unlikely to be a bound variable. Schwarz indeed points to the fact that the sloppy reading of a personal pronoun doesn’t require the usual c-command requirement, either (e.g. the well-known Wescoat example, the subject of many subsequent papers, such as Dalrymple et al. 1991; Hardt 1993; Fiengo and May 1994 and Tomioka 1997). With this possibility in mind, let us come to the final problem, which I believe is applicable to a proform hypothesis in general. It is predicted that the sloppy VP cannot host a trace because a proform is believed to be incapable of doing so. Once again, Schwarz already anticipated this problem with the following example. (14)
a. When John cooks something, he doesn’t say what he does. b. When he BAKES something, he doesn’t, either. (Schwarz 2000: 169; (44))
He says: Unfortunately, this case and similar ones are not fully significant since the (a) sentence is degraded in isolation. However, it seems that to the extent that (a) is acceptable, the above prediction is not borne out, that is the discourse has a sloppy interpretation. (Schwarz 2000: 169)
I believe that the reason why (14a) above does not sound good is that it violates Schuyler’s (2001) generalization that wh-movement out of an elided VP must be accompanied by a contrastively focused constituent that is c-commanded by the moved wh.4 Unlike (14), the example (15) has focus on the negative AUX, doesn’t, which is helpful in overcoming the degradation problem that (14) has. (15)
a. John has a very indirect way of telling what he thinks. For instance, when he likes someone, he tells you who1 he DOESN’T D. (D = like t1)
216
Satoshi Tomioka
b. Wait a minute. But, when he HATES someone, he doesn’t D. Instead, he tells you exactly who he hates. (D = tell you who2 he doesn’t hate t2) Although the judgment is subtle, perhaps due to the complexity of the sentence, (15) seems to have the sloppy reading, contrary to what a proform analysis would predict. Jason Merchant (personal communication) warned me, however, that the issue of a wh-trace may be far more complex. He points out that Antecedent Contained Deletion (ACD) and Comparative Deletion (CD) cases, which also involve A-bar traces in elided VPs, are not acceptable. (16)
ACD Fred read the book that he was supposed to D, and *he also REVIEWED the one that he was D.
(17)
CD Fred read more books than he was supposed to D, and *he also REVIEWED more books than he was D.
What is striking about these examples is that they are simply ungrammatical, regardless of the interpretations of the second empty VPs. In other words, they don’t even have the strict readings. That indicates that the ungrammaticality of (17) is unrelated to the question of whether sloppy VP Ellipsis can host a wh-trace or not. As a matter of fact, the examples in (18) and (19) are also unacceptable despite the fact that they do not involve sloppy VP Ellipsis. (18)
ACD a. Fred read the book that he was supposed to D, and *ERIC also read the one that he was D. b. Fred read the book that he was supposed to D, and *he also read the JOURNALS that he was D.
(19)
CD a. Fred read more books than he was supposed to D, and *ERIC also read more books than he was D. b. Fred read more books than he was supposed to D, and *he also read more JOURNALS than he was D.
It should be noted that all of the examples in (16)–(19) involve A0 -movement of an operator (either an empty relative operator or an empty degree operator). In none of these examples, however, is there
A step-by-step guide to ellipsis resolution
217
any contrastively focused material that is c-commanded by the moved operator, which leads to a kind of configuration that is ruled out by the aforementioned generalization by Schuyler. Once this factor is controlled, sloppy interpretations become available or at least easier to detect.5 (20)
a. Why are you so upset with Fred? He bought the books that he was supposed to, right? b. Yeah, but then, (?) he READ the books (Op1) that he WASN’T D. (D = supposed to read t1)
Thus I conclude that an instance of sloppy VP can indeed host a wh-trace. The significance of our finding concerning wh-traces is not simply a matter of quantifying into a pronoun. If it were, we might be able to account for it by using double indexing on the VP pronoun or by changing the semantic denotation to functions from assignment functions to whatever the denotation we ordinarily assign to a linguistic expression. Or at least we find a similar case in nominal anaphora, namely a pronoun of laziness (or a paycheck pronoun). (21)
The man who1 gives his1 paycheck to his wife is wiser than the one who2 gives it to his cat. (it = his1 paycheck)
What our examples of sloppy VP Ellipsis show is more than that, however. There are no known pronouns of any kind which can contain wh-traces within them. The following examples illustrate this point. (22)
a. We know which countries Kim visited. *The question is which countries she didn’t do so. b. Tell me which models you took good pictures of and *which models you took ugly ones.
To sum up the challenges that a pro-form thesis faces, we would need to find a silent pronoun that has the following attributes: (i) it is neither referential nor a bound variable, (ii) it is cross-categorical, corresponding to a VP, an NP (in the DP structure), or IP, and (iii) it can host a wh-trace. As far as I can see, there are no pronominal forms that satisfy all of these requirements. Therefore, we are forced to postulate a new kind of empty anaphora specifically tailored for ellipsis. Such a move comes with great cost, and it is difficult to justify.
218
Satoshi Tomioka
9.3
LF identity ‘step-by-step’
9.3.1
Basics
Section 9.2 has shown that it is not easy to implement a proform approach for sloppy VP Ellipsis. On the other hand, the problem for the LF Structure approach is also obvious. Consider (2) and (4) again. (2)
A. When John had to cook, he didn’t want to. B. When he had to CLEAN, he didn’t, either.
(4)
LF Structure of (2) A. When John had to cook, he didn’t want to cook. B. When he had to CLEAN, he didn’t want to clean, either.
The second VP Ellipsis should not be licensed since there is no VP that has the LF structure of the form ‘‘want to clean.’’ The problem cannot be avoided as long as we are looking at LF representation globally. The solution I propose crucially relies on the idea that the LF identity for ellipsis does not apply in one big step. Or at least it does not have to. Imagine XP dominates YP. YP has an appropriate antecedent for deletion. That licenses the deletion of YP. For the determination of an appropriate ellipsis antecedent of XP, the content of YP doesn’t matter, as long as a potential antecedent of XP also has elided material in the position corresponding to XP. Putting it differently, smaller ellipsis gets the ‘‘seal of approval’’ when it finds an appropriate antecedent for itself, and once this is accomplished, the identity condition for larger ellipsis containing the smaller ellipsis does not care about the content of the small ellipsis. To distill this intuitive idea, it may be useful to examine a case like (23), which involves ‘‘partial’’ VP Ellipsis. (23)
A. When John had to cook, he didn’t want to D. B. When he had to clean, he didn’t want to D, either. (sloppy OK)
There is no problem here. Each instance of VP Ellipsis should be and indeed is licensed. The question is whether the two instances of want to D can be regarded as LF identical. The answer is not obvious. They are identical as far as the surface strings are concerned whereas the semantic contents of the two VPs differ, provided that the missing VPs have full representations at LF. If the surface string argument wins, then we will get (24). (24)
A. When John had to cook, he didn’t want to D. B. When he had to clean, he didn’t want to D, either. (sloppy OK)
A step-by-step guide to ellipsis resolution
219
The outcome in (24) is, of course, exactly what we have in (2). But why should a surface string matter for the LF identity condition? Perhaps, the term ‘‘surface string’’ is misleading. The task here is to determine whether the second instance of want to D can be left unpronounced. I suggest that, in this decision-making process, the content of what has been already decided to be deleted should not play a role. (25) shows a slightly more schematic illustration of how this idea should be cashed out. (25)
Step 1: When he had to clean, he didn’t want to clean Find an antecedent for this VP Step 2: When he had to clean, he didn’t want to clean. A matching VP in the when-clause licenses the deletion Step 3: When he had to clean, he didn’t b want to clean c Find an antecedent for this VP
What we are looking for in Step 3 is a VP of the form ‘‘[VP want to [SILENT ]].’’ Since A’s utterance in (2) furnishes such a VP, the ellipsis is licensed. It is also obvious that the sloppy reading is not available in (5), which constitutes a minimal pair with (2).
VP
(5)
A. When John had to cook, he didn’t want to cook. B. When he had to CLEAN, he didn’t, either.
With this example, Step 3 of (25) fails since there is no VP of the appropriate form in A’s utterance. In a sense, Schwarz and I share the same goal: to prevent the content of smaller ellipsis from coming into play in the resolution of larger ellipsis. In Schwarz’s system, this goal is achieved by making the smaller elided VP a silent bound variable. In the current proposal, the smaller VP is ignored because it has already been decided to be deletable. The phrase ‘‘already been decided to be deletable’’ is very suggestive. It points to the direction that ellipsis resolution is not a one-step process but can involve a combination of smaller steps. 9.3.2
Towards a formalization
In this subsection, I will formalize the ‘from-bottom-to-top’ way of VP Ellipsis resolution. First, I assume, along with Chomsky and Lasnik (1993), Fox (2000), Merchant (2001), and many others, that VP Ellipsis is an instance of PF deletion. The instruction for the PF component to delete
220
Satoshi Tomioka
is featural, and my analysis owes greatly to Merchant’s (2001) e-feature marking idea. (26)
E-feature marking: a. Some heads select XPs marked for an e-feature. b. All e-marked constituents are unpronounced.
This is a slight modification of Merchant’s theory, which puts an e-feature on a head, and what is elided is the complement of an e-marked head. The reason for the change is simply a matter of convenience since I find it a lot easier to assume that e-features are placed on elided constituents. E-feature marking is obligatory in the sense that, if an XP is selected by an e-marking head, then, the XP must bear an e-feature and consequently delete. The optionality of deletion does not come from the optionality of e-marking, but rather from the choice of a head from the lexicon. Hence, e-marking is not constrained by any kind of economy principle.6 Of course, e-features are relevant for the recoverability of elided material as well. I propose the following identity condition for e-marked constituents. (27)
Identity condition for ellipsis: An e-marked constituent a must have an antecedent such that a. The e-marking of is identical to that of , AND b. and are LF identical up to indices and structural content of any e-marked constituents that are properly contained by or .
If an e-marked constituent does not embed any further e-feature, (27) is basically identical to the identity condition that has been proposed in the past (e.g., Rooth 1992a; Heim 1997; Tomioka 1997; Romero 1998). The reference to properly contained e-marking is added so that the content of the smaller ellipsis is spared from the Identity Condition imposed on the larger ellipsis.7 Let us see how this new condition works with the example we have been discussing. (28)
LF of (2) under the sloppy reading: A. When he had to cook, he1 didn’t want PRO1 to [VP E cook] B. When he had to clean, he1 didn’t [VP E want [CP [IP PRO1 to [VP E clean]]]]
For each of the three instances of e-marking, the condition in (27) must be satisfied. (29)
a. p [VP E cook]] because of the presence of ‘‘when he had to [VP cook]’’
A step-by-step guide to ellipsis resolution
221
b. p [VP E clean]] because of the presence of ‘‘when he had to [VP clean]’’ c. p [VP E want [CP [IP PRO1 to [VP E clean]]]] because of the presence of ‘‘he didn’t [VP want PRO1 to [VP E cook]]’’ (29a,b) are straightforward. (29c) is licensed by the new condition. The VP in A’s utterance has the structurally identical e-marking, and everything apart from the material within the embedded e-marking is identical to the elided VP. Thus, the condition is met. We can also correctly predict that (5), the phonologically repeated version of (2), fails to satisfy the condition. (30)
LF of (5) under the sloppy reading: A. When he had to cook, he1 didn’t want PRO1 to [VP cook] B. When he had to clean, he1 didn’t [VP E want [CP [IP PRO1 to [VP E clean]]]]
(31)
a. p [VP E clean]] because of the presence of ‘‘when he had to [VP clean]’’ b. [VP E want [CP [IP PRO1 to [VP E clean]]]] *because there is no VP of the form ‘‘[VP want PRO1 to [VP E]]’’
Since the VP in A’s utterance does not have an e-marked constituent in the structurally parallel position, (31b) fails. Hence, no sloppy reading. One may wonder, however, whether (27) can successfully license the strict reading of (5). If we minimally change (30) in such a way that we have [VP E cook], instead of [VP E clean], then, the condition (27) rules it out for the same reason that we saw in (31b). However, there is another possibility for the strict reading of (5) in which there is no embedded e-marking. (32)
LF of (5) under the strict reading: A. When he had to cook, he1 didn’t want PRO1 to [VP cook] B. When he had to clean, he1 didn’t [VP E want [CP [IP PRO1 to [VP cook]]]]
(33)
VP E want [CP [IP PRO1 to [VP clean]]]] [p because the VP in A’s utterance is LF equivalent
Whether we have an embedded e-feature or not, the outcome is identical. Thus, the LF in (32) should be allowed, and (5) can have the strict reading. It is now the time to ask ourselves the question of why the identity condition should be formulated as in (21). The trend of the recent syntactic theories, most notably the Minimalist framework, is to rely
222
Satoshi Tomioka
heavily on the notion of derivations. The current proposal, though it itself is not derivationally formulated, is in line with this trend. The LF identity condition of an elided constituent ‘‘relives’’ the history of its derivation as it is evaluated in the ‘‘from-bottom-to-top’’ fashion. Once the elided material clears the condition, it is no longer considered as a part of the LF identity of a larger elided material that contains it. This part of the condition is reminiscent of Chomsky’s (2000) idea of derivation by phase, in particular the Phase Impenetrability Condition, which provides that operations in one phase cycle are not visible beyond its phase boundary. Could the Identity Condition that I have proposed be reformulated in a more derivational fashion? It is a legitimate question, and such an attempt is certainly worthwhile. I may add one cautionary note, however, that ellipsis requires global knowledge of the context; in resolving ellipsis, one must go beyond what has been already built up in the derivational process in order to find an antecedent. It is not entirely clear to me whether such a move can be justified. I will leave this issue for future investigation. 9.3.3
An alternative: e-givenness
A possible alternative to the solution proposed in section 9.3.2 can be devised by appealing to the theory proposed in Merchant (1999; 2001). Merchant’s idea of e-givenness, inspired by the focus theory of Schwarzschild (1999), combines the identity condition for ellipsis with the general theory of focus licensing. Merchant’s theory can be summarized as follows. Ellipsis of a constituent is licensed if and only if is selected by a head marked for the ellipsis feature e. The sister of an e-marked head must satisfy the e-givenness condition, a semantic identity condition based on focus structure. (34)
e-givenness: An expression E is e-given iff E has a salient antecedent A and, modulo Existential Type Shifting, i. A entails e-clo(E), and ii. E entails e-clo(A).
(35)
The e-closure of (e-clo()) is the result of replacing all f(ocus)marked elements of with existentially bound variables of the appropriate type (modulo Existential Type Shifting).
The Existential Type Shifting of a constituent existentially closes all the unsaturated arguments of . Here is an example that demonstrates how Merchant’s theory works for VP Ellipsis.
A step-by-step guide to ellipsis resolution
(36)
223
a. Abby left the party, and then, Bob did D. b. LF: Abby [VP left the party], and then Bob didE [VP left the party]
The elided VP is the sister of the e-marked head didE and must satisfy the e-givenness condition. The following steps show how the condition is satisfied. (37)
a. The denotation of the elided VP: lx:x left the party b. E-type-shifting: 9x:x left party c. E-clo (vacuous): 9x:x left party
There is an antecedent (36a) that satisfies the conditions in (32). Both the e-clo and Existential Type Shifting of the VP in the first conjunct left the party would give the proposition ‘‘9x:x left party.’’ As formulated, Merchant’s e-givenness cannot account for the puzzle presented by Hardt and Schwarz. However, Merchant (2005) suggests that a small amendment to the definition of e-closure can be made in order to accommodate the problem. (38)
The e-closure0 of (e-clo0 ()) is the result of replacing all f-marked elements and the sisters of all e-marked heads contained in with existentially bound variables of the appropriate type (modulo Existential Type Shifting).
With this modified definition of e-closure, we existentially close not only f-marked elements but also a constituent selected by a head with an e-feature. Let us see how this modified e-givenness can deal with the puzzle at hand. (2)
A. When John had to cook, he didn’t want toE {cook}. B. When he had to clean, he didn’tE {want toE clean}, either.
The Existential Type Shifting and e-closure of the bigger VP in B’s utterance, want toE clean, are shown in (39). (39)
a. Existential Type Shifting: 9x:x wants to clean b. E-clo: 9x:9P5e;t4 :x wants to do P.
Now we need to find an antecedent whose e-closure entails (39b) and whose Existential Type Shifting is entailed by (39a). The e-closure of the bigger VP in (2a), want toE cook, is precisely that, as shown in (40). (40)
a. Existential Type Shifting: 9x:x wants to cook b. E-clo: 9x:9P5e;t4 :x wants to do P
224
Satoshi Tomioka
(39a) entails (40b), and (40a) entails (39b). So, the sloppy reading in (2) is licensed. However, the situation in (5) is different. (5)
A. When John had to cook, he didn’t want to cook. B. When he had to clean, he didn’tE {want toE clean}, either.
The e-closure and the Existential Type Shifting of the bigger VP in (5A) is (41). (41)
9x:x wants to cook.
It is obvious that (39a,b) and (41) do not have the entailment relations that e-givenness requires: (41) entails (39b) but is not entailed by (39a). So, ellipsis is not licensed. Hence, no sloppy reading in (5). The revised version of Merchant’s theory provides another effective means to make the semantic content of an embedded ellipsis irrelevant. However, there are two aspects of this theory that are crucially different from the proposal above. First, recall that e-marked elements are not the only things that are replaced with variables; f-marked constituents are also replaced with variables of the appropriate type. Therefore, Merchant’s theory predicts that sloppy ellipsis is possible with a non-ellipsis antecedent as long as the antecedent is f-marked. Now, let us look at (5) again. Schwarz notices that, when repeated, the second VP cook can have prosodic prominence. Thus, the following pattern (CAPITAL LETTERS indicate stress) is acceptable.8 (42)
A. When John had to COOK, he didn’t want to COOK. B. When he had to CLEAN, he didn’t, EITHER.
The conventional interpretation of this stress pattern is to mark the stressed constituent with the focus feature f. Then, (5) has the following LF. (43)
A. When John had to [cook]F , he didn’t want to [cook]F. B. When he had to [clean]F, he didn’tE {want toE clean}, either.
Then, the e-closure of the VP [VP want to [cook]F] is indeed (16), making it identical to the elided version of (5). Therefore, Merchant’s theory of e-givenness actually fails to account for the contrast Schwarz noted. The second difference is the treatment of an elided structure that contains a trace (or a bound variable in general). One of the arguments against the proform analysis of sloppy ellipsis was that it can host a wh-trace. The relevant example is repeated below.
A step-by-step guide to ellipsis resolution
(15)
225
a. John has a very indirect way of telling what he thinks. For instance, when he likes someone, he tells you who1 he doesn’t D. (D = like t1) b. Wait a minute. But, when he HATES someone, he doesn’t D. Instead, he tells you exactly who he hates. (D = tell you who2 he doesn’t hate t2)
(20)
a. Why are you so upset with Fred? He bought the books that he was supposed to, right? b. Yeah, but then, (?) he READ the books (Op1) that he WASN’T D. (D = supposed to read t1)
The second instance of ellipsis has the following e-marking pattern. (44)
. . . he doesn’tE [ tell you who2 he doesn’tE [hate t2]]
It is not clear how the e-closure of the smaller VP can be achieved since it contains a variable that remains free within that VP. Imagine, as Tomioka (1997: chapter 3) suggests, that an unbound variable is treated as a missing argument a` la Categorial Grammar.9 However, it would create a different problem. If the variable is closed off at the level of the VP via e-closure, the wh-phrase who2 has no variable to bind. In other words, if an elided constituent has a bound variable and its binder is contained in larger ellipsis, the existential closure of smaller ellipsis cannot be easily incorporated into compositional interpretation. It remains to be seen whether these differences can be eliminated. If they could, then, the proposal here and the e-givenness theory would become notational variants as far as sloppy ellipsis is concerned, and the evaluation of the two hypotheses would have to be done in larger and more general contexts. 9.4
Focus
There is growing consensus among ellipsis researchers that ellipsis should be treated as a part of a larger scheme of linguistically representing redundant information. Particularly influential was Rooth’s (1992a) idea that, in addition to the LF identity condition, a licensing condition of focus is imposed on elliptical sentences, and in this sense, focus structure of sentences with ellipsis must be licensed independently of the LF identity condition. While the main theme of this chapter is to construct a syntactic identity condition motivated by sloppy ellipsis data, it is not a trivial issue
226
Satoshi Tomioka
whether the proposed analysis can be properly embedded in a reasonable theory of focus. Let us once again come back to example (2). (2)
A. When John had to cook, he didn’t want to D. B. When he had to clean, he didn’t D, either.
Since the contrasted VP clean in B’s sentence necessarily receives prosodic prominence, let us assume that it is marked for focus. If nothing else is focused and the missing VP is fully represented, the sentence has the following LF representation. For the sake of simplicity, the presence of either is ignored. (45)
When he had to [CLEAN]F, he didn’t want to clean.
The next step is to replace the f-marked constituent with a variable of the type identical to the denotation of that constituent. After this operation, we get the following proposition. (46)
When he had to do P, he didn’t want to clean.
In Rooth’s theory of focus, the utterance context must furnish one of the two LF representations shown below. (47)
a. An LF the denotation of which is a set of propositions of the form ‘‘When he had to do P, he didn’t want to clean’’ and contains at least one member in which P 6¼ [[clean]]. b. An LF the denotation of which is a proposition of the form ‘‘When he had to do P, he didn’t want to clean’’ in which P 6¼ [[clean]].
It is obvious that such an LF cannot be found. Since B’s sentence is contrasted with what A has said, we would hope that A’s sentence satisfies the condition. Unfortunately it doesn’t, and the problem is not unique to Rooth’s theory of focus. In Schwarzschild’s (1999) theory, for instance, the existentially closed version of (46), shown in (48), must be entailed. (48)
9P: when he had to do P, he didn’t want to clean.
Clearly this proposition is not entailed by A’s sentence. The heart of the problem is that the second instance of clean in (45) is a constant so that its meaning does not co-vary with that of the first instance of clean. The only available option, apart from treating it as a bound variable like Schwarz does, is Kratzer’s (1991) idea of F-indexing. Under this system, each occurrence of focus comes with an index, but unlike an ordinary index, it is invisible to ordinary assignment functions. In computing the focus meaning of a sentence, however, distinguished assignment
A step-by-step guide to ellipsis resolution
227
functions, as well as ordinary ones, are used, and an f-index is sensitive only to distinguished assignments. The following example illustrates how the system works. ‘‘g’’ is an ordinary assignment function whereas ‘‘h’’ is a distinguished assignment function. (49)
a. [FRED]F1 left early. b. Ordinary meaning: [[[FRED]F1 left early]]g = the proposition that Fred left early. c. Basis of focus meaning: [[[FRED]F1 left early]]g,h = the proposition that h(1) left early. d. Focus meaning: {p: 9x 2 De :p = [[[FRED]F1 left early]]g,h1/x} = {p: 9x 2 De :p ¼ x left early}
The addition of f-indexing makes it possible to license the focus marking of sloppy ellipsis. For the LF of (2b), we now have (50), instead of (45). (50)
When he had to [CLEAN]F2, he didn’t {want to [clean]F2}.
Notice that the two instances of the same VP [clean] are F-coindexed, and they co-vary under the distinguished assignment h. Thus, the basis of focus meaning and the focus meaning are (51a) and (51b), respectively. (51)
a. The proposition that, when he had to do h(2), he didn’t want to h(2). b. {p: 9P 2 D5e;t4 :p ¼ when he had to do P, he didn’t want to do P}
Rooth’s condition is now satisfied since (2a) denotes a proposition of the form ‘‘when he had to do P, he didn’t want to do P,’’ and P is not ‘‘clean.’’ (51) can also be used for Schwarzschild’s focus theory. Let (51a) be the input of his e-Closure in the way specified in (52a). Then, the result of applying it to (51b) will be (52b). (52)
a. Let 1, . . ., n to be f-indices within an expression . Then, e-Closure of with respect to g and h ¼ 9 x, . . ., z [[]]g,h 1/x,. . .,n/z b. E-Closure of (51b) ¼ 9P. when he had to do P, he didn’t want to do P.
(52b) is entailed by the meaning of (2a). Thus, Schwarzschild’s condition is met. To sum up this section, the revised LF identity condition appears to make trouble for a focus theory. Embedded ellipsis is not a bound variable, but in order to get the right focus meaning, it must co-vary with its
228
Satoshi Tomioka
antecedent. It is crucial, therefore, that Kratzer’s designated variables are brought into the computation of focus meaning. 9.5
Conclusion
The difficulty that sloppy ellipsis presents is very acute. Although a proform theory seems to be an easy way out, I have pointed out several challenges that such a theory must overcome. Sloppy ellipsis is found with other types of ellipsis, such as Sluicing, and some of the well-known constraints on bound variables do not apply to sloppy ellipsis. With some effort, we can even find some instances in which sloppy ellipsis contains an A0 trace. There are no known pronominal forms that can fend off all these challenges. On the other hand, sloppy ellipsis is equally troublesome for an approach that assumes full-fledged LF structure of elided material. In this chapter, I made an attempt to show that it is still possible to maintain the LF structure approach even in light of sloppy ellipsis data. The idea behind the current proposal is that the LF licensing of VP Ellipsis should allow the LF structure of smaller ellipsis to be ignored in the computation of the LF identity of larger ellipsis that contains it. While the analysis presented in this chapter is in accordance with the recent trend in Minimalist syntax that syntactic constraints apply derivationally, I leave as an open question whether the proposed LF identity condition is properly couched within the notion of phrases. Sloppy ellipsis also has impact on the focus licensing condition that is believed to be essential in ellipsis. Since ellipsis with sloppy meaning is not treated as a bound variable, it is expected to act as a constant in the semantic computation. Although it does not pose a problem for the ordinary semantic meaning, its focus meaning should be represented as a bound variable. Kratzer’s (1991) theory of focus, which makes use of focus indexing and designated variables, provide the machinery that solves this problem.
10
Shared constituents and Linearization Chris Wilder NTNU Trondheim
This chapter investigates basic linear properties of constituent sharing (‘‘directional ellipsis’’) in coordination, proposing an explanation in terms of Kayne’s Linear Correspondence Axiom, according to which the order of words in a phrase marker is determined by c-command within that phrase marker. The explanation depends on a multiple dominance treatment of shared constituents, requiring that the Single Mother Condition on phrase markers be dropped.1
10.1
Outline
Coordinate constituent sharing is realized in both ‘‘forward’’ (Gapping, leftward across-the-board or ‘‘ATB’’ movement) and ‘‘backward’’ (Right Node Raising) dependencies. The directionality of the dependency correlates with a constraint on the placement of ‘‘gaps’’ corresponding to the shared constituent a. (1)
a. If a shared constituent a surfaces in the final conjunct (as in Right Node Raising), then the gaps corresponding to a in all non-final conjuncts must be at the right edge of their respective conjuncts, whereby that position must be a possible surface position for a. b. If a surfaces in or to the left of the initial conjunct (as in Gapping and leftward ATB movement), then gaps corresponding to a in other conjuncts underlie no such edge restriction.
The edge condition (1a) is illustrated by the contrast in (2) (cf. Oehrle 1991). Both (2a) and (2b) involve a shared object that surfaces in the final conjunct. (2a) involves a direct object, which can stand after the PP in the first VP, i.e. it can undergo Heavy NP Shift – cf. (3a). Indirect objects resist heavy shift (3b), so the gap in (2b) cannot be in VP-final position to satisfy (1a). The absence of such a restriction in ‘‘forward’’ dependencies is 229
230
Chris Wilder
illustrated by the possibility for ‘‘medial’’ verb gaps in Gapping sentences like (4). (2)
a. I invited into my house and congratulated the boy who lives next door. b. * I gave a present and congratulated the boy who lives next door.
(3)
a. I invited into my house the boy who lives next door. b. * I gave a present the boy who lives next door.
(4)
John drank beer and [Mary wine]
This chapter shows how (1a) can be derived directly from a modified version of Kayne’s (1994) Linear Correspondence Axiom (LCA). The proposal is based on the following assumptions: (i) syntactic representations are dominance-only trees (ii) constituent-sharing involves multiple dominance (i.e. we need to relax the Single Mother Condition) (iii) precedence is defined for terminals only by the LCA (as in Chomsky, 1995) (iv) coordinations are syntactically asymmetric (the first conjunct c-commands the second, etc.). The role played by the LCA is similar to that in Kayne (1994), viz (5): (5)
The LCA determines (i) whether a tree is permissible, and (ii) if so, it determines the left-right ordering of terminals of that tree.
The LCA includes the hypothesis that word order is related to structure in that precedence among terminals is determined by c-command among categories in the tree – if X asymmetrically c-commands Y, the terminals of X are ordered before the terminals of Y. In addition, the LCA imposes a ‘‘linearizability’’ requirement on phrase markers – if c-command relations are such that they fail to determine a linear ordering of terminals, the tree is ill-formed. In section 10.4, the contrast in (2) is explained in precisely these terms; c-command in the respective phrase-markers allows (2a) to be linearized but renders (2b) unlinearizable. In Kayne (1994), the constraints imposed by the LCA on the shape of possible phrase markers are argued to derive aspects of X0 -theory such as binary branching. Kayne’s LCA also derives the Single Mother Condition, excluding trees involving multiple dominance. Since the account of (1) argued for here depends on the assumption that shared constituents
Shared constituents and Linearization
231
involve multiple dominance, it becomes necessary to modify Kayne’s LCA (see section 10.4). The account is compatible with Chomsky’s (1995) theory of phrase structure (as discussed in section 10.3). In that theory, phrase structure trees are dominance-only trees created by the Merge operation, and the LCA is conceived of as an algorithm of the PF-component which maps dominance-only trees to an ordering of terminals. Merge theory derives X0 -theory in a different way to the LCA, rendering that aspect of Kayne’s proposal largely redundant; the role of the LCA is largely restricted to (5ii). However, in the present proposal, the linearizability requirement takes on a substantial new role in determining possible trees (5i). The main strength of the proposal is that it eliminates the need for either a rightward movement operation of Right Node Raising (RNR) or a directional ellipsis rule of Backward Deletion (with their specific stipulated properties), to account for the RNR construction. In section 10.5, the multiple dominance – LCA approach is shown to be both conceptually and empirically superior to an ellipsis theory in accounting for a range of RNR facts. A possible extension of the proposal to movement dependencies, replacing the copy theory of movement with a treatment in terms of multiple dominance, is sketched in section 10.6. 10.2
Shared constituents
There are three competing approaches to constituent sharing phenomena: (i) Across-The-Board (ATB) movement (Williams 1978); (ii) ellipsis, i.e. PFdeletion of constituents in one conjunct under identity with constituents in another conjunct (e.g. Wilder 1997); (iii) multiple dominance, i.e. having a single constituent belong to two conjuncts at the same time (McCawley 1982; Goodall 1987; Muadz 1991; Moltmann 1992). Examples like (6) have been variously analyzed as involving ATB movement of the shared constituent out of the coordination (Right Node Raising, Postal 1974), as in (7a);2 ellipsis (e.g. the Backward Deletion rule of Wilder 1997) as in (7b); and as literal sharing of a constituent via multiple dominance (7c) (e.g. McCawley 1982, Goodall 1987). Gapping sentences (8), likewise, have been analyzed as involving ATB movement of the verb as in (9a) (Johnson 2003); ellipsis (e.g. the Forward Deletion rule of Wilder 1997) as in (9b); and multiple dominance as in (9c) (Goodall 1987).3 (6)
John bought and Mary read the book.
(7)
a. [John bought tj ] and [Mary read tj ] [the book]j b. [John bought the book ] and [Mary read the book]
232
Chris Wilder
c.
&P
&′
TP1 T ′1
SU1 John
TP2
& and VP1
T1
T ′2
SU2 Mary
V1 bought
T2
VP2
V2 read (8)
John likes beer and Mary, wine.
(9)
a. Johnk likesj [ tk tj beer ] and [ Mary tj wine ]
OB∗ the book
b. [John likes beer ] and [ Mary likes wine ] c.
&P &′
TP1
T ′1
SU1 John T1
TP2
& and
T ′2
SU2 Mary
VP 1
T2
OB 1 beer V* likes
VP2
OB2 wine
Shared constituents and Linearization
233
The three approaches differ in the answers they give to the questions in (10): (10)
a. Is there one constituent or two? b. Is the shared constituent inside the conjuncts or outside?
The ellipsis approach assumes there are two constituents in the syntactic representation (i.e. two copies of the shared constituent), each inside its own conjunct. The multiple dominance (MD) approach assumes that there is only one constituent, situated ‘‘simultaneously’’ inside both conjuncts. The ATB movement approach supposes the shared constituent surfaces outside the coordination, but the notion of sharing crucially relates to a having raised out of the coordination, leaving a trace inside each conjunct: the conjuncts ‘‘share the trace of a.’’ Thus question (10a) arises for the trace of ATB movement: is there one a-trace or two? The answers to (10) proposed here are the following. With respect to (10a), there is only one constituent; therefore the ellipsis approach is wrong. With respect to (10b), the shared constituent is generated inside the conjuncts in all cases – i.e. the multiple dominance approach is correct. However, I suppose that the single constituent surfaces inside the coordination only in the RNR construction. The structure (7c) is the correct surface representation for (6), while (9c) is false for (8). In Gapping constructions, following Johnson (2003), I suppose that the shared constituent (the verb) surfaces outside the coordination, i.e. that the ATB movement analysis (9a) is correct. For ATB movement generally, however, I adopt the view that an ATB-moved constituent is generated as a shared constituent; an ATB trace is a single multiply dominated constituent, shared by both conjuncts. Most cases of forward conjunction reduction can be analyzed as movement out of conjuncts. Thus, shared subject coordinations may be analyzed as instances of Across-The-Board subject raising (see Burton and Grimshaw 1992). Once we give up the Single Mother Condition to allow an MD representation for RNR, then it also makes sense to assume that a trace left by ATB movement – e.g. the copy left by ATB raising of the subject (written twice and marked * in (11)) – also comprises a single multiply dominated constituent, rather than two separate constituents. (11)
Mary [&P [VP1 Mary* bought a book]] and [VP2 Mary* read it]]
However, some cases such as ‘‘verb-gapping’’ in (8) can only be analyzed as ATB raising at the cost of assuming that movement of other constituents violates the Coordinate Structure Constraint. The word order in Gapping sentences, together with the assumption of the CSC, thus provides an argument for the need for a forward ellipsis rule. The shared verb
234
Chris Wilder
appears to be contained within the first conjunct, since it is preceded by a constituent only belonging to the first conjunct (its subject John) (12). (12)
[ John likes beer ] and [ Mary
V
wine ]
The verb-gap in the second conjunct is therefore not c-commanded by its antecedent, so it cannot be a ‘‘trace.’’ Hence, it must be an ‘‘ellipsis site,’’ and some kind of forward ellipsis rule must be assumed.4 In the alternative approach to Gapping (13) proposed by Johnson (2003), the verb in the Gapping sentence has raised across-theboard, and so is outside the coordination. This entails that the subject of the first conjunct, which precedes the shared verb, is also outside the coordination. Movement of the subject represents a violation of the CSC. (13)
*CSC John drinks [&P [ tSU tV beer] and [Mary tV wine]] ATB
Assuming this type of approach can be maintained, it raises the possibility that all ‘‘forward ellipsis’’ is in fact ATB movement. Shared constituents in such cases are always outside the coordination at surface structure, so that ‘‘forward ellipsis’’ gaps are c-commanded by their overt antecedents, i.e. are actually trace gaps. In that case, there is no forward ellipsis; the gaps attributed to a Forward Deletion or Gapping rule actually result from trace deletion.5 If this is correct, which I will assume, then the only case where an overt shared constituent is inside the coordination at surface structure is RNR. The questions raised then concern the properties of RNR: Why do shared constituents that do not raise out of the coordination only surface inside the final conjunct? Why does the edge condition (1a) hold in this case? As shown in section 10.4, this conjunction of properties is the result of constraints on linearizing trees containing shared constituents. The next paragraphs sketch the underlying intuition. Suppose that shared constituents involve multiple dominance – there is only one shared constituent a, and a has more than one mother (one in each conjunct). Then there is only one set of terminals of a to be linearized. Consider (14), where a is a shared constituent, dominated by both conjuncts. Assume a is asymmetrically c-commanded by constituents in both conjuncts (by w in Conj-1 and by y in Conj-2). By the LCA, a must follow both w and y. That is only possible if a surfaces in the string corresponding to the final conjunct. (14)
[&P [Conj-1 w a x ] [&0 & [Conj-2 y a z ] ] ]
Shared constituents and Linearization
235
Assume also that a asymmetrically c-commands x in the first conjunct. According to the LCA, a must precede x. But having a precede x conflicts with having a surface in the second conjunct, following y (x will have to precede y in the second conjunct). Crucially, that conflict does not arise if there is no terminal x in the first conjunct such that c-commands x, as the terminals of a are then ordered following all other terminals belonging to the initial conjunct. The fact that a actually surfaces farther to the right, inside the final conjunct, is unproblematic insofar as a follows all material that c-commands it, and a does not follow any material that it c-commands. This is in essence the explanation for the right edge condition (1a) on gaps in RNR constructions. There is more to be said about the fact that a belongs to the first conjunct in (14), which c-commands, and therefore precedes, the conjunction and the following conjunct. This relates directly to the question of how trees involving multiple dominance can be linearized at all. Central to the proposal is the idea that only those multiple dominance trees which can be linearized by the LCA are well-formed. The modification of the LCA presented in section 10.4 is necessary to allow linearization of any multiple dominance trees. At the same time, it derives the directionality and edge properties of RNR (the first part of (1)), i.e. that if a shared constituent a surfaces inside the coordination, then (i) a surfaces in the final conjunct, and (ii) a-gaps must satisfy the ‘‘right-edge’’ condition. It is then possible to derive the second part of (1), i.e. the fact that if a surfaces outside the coordination (ATB movement), then a-gaps (¼ ATB traces) need not satisfy the ‘‘right-edge’’ condition, by adopting (15):6 (15)
Traces need not be linearized (‘‘trace deletion’’ renders the terminals of ‘‘trace copies’’ invisible to the LCA, cf. Chomsky 1995)
If a shared constituent a happens to c-command some terminal in the first conjunct, then the structure (not linearizable if a is inside the coordination at S-structure) is linearizable if a raises out of the coordination by S-structure. The reason for that is Trace Deletion. Moved a gets linearized in a position outside the coordination; a’s trace inside the coordination is ‘‘invisible’’ at PF, irrelevant to linearization. Hence the conflict described for (14) does not arise. This explains why gaps corresponding to shared constituents appearing to the left of the coordination do not underlie an edge condition like (1a). 10.3
Multiple dominance
How does multiple dominance differ from ellipsis (‘‘copies-and-deletion’’)? In a ‘‘copies’’ theory, one has two things instead of one. Thus, something
236
Chris Wilder
can be done to one copy of a without that thing automatically happening to the other copy of a – deletion, for instance. In a multiple dominance theory, there is only one shared consituent a, which happens to be in two places at once. Nothing can happen to a in one position without that thing automatically happening to a in its other position. In the present proposal about coordination ellipsis, what causes a gap to appear in the string of one conjunct is not deletion of a copy of the antecedent. Rather, what happens is that a single shared constituent simply fails to be positioned at that particular place in the string when the tree is linearized. The multiple dominance theory of constituent sharing ensures that the shared constituent is treated as a single entity for purposes of linearization. In terms of standard phrase structure theory, this involves relaxing the Single Mother Condition, to permit trees like (16) that branch ‘‘upwards’’ as well as ‘‘downwards.’’7 (16)
E C
D α
A
B
The proposal is compatible with derivational (e.g. Merge theory, Chomsky 1995) or representational views of phrase structure, but presupposes that trees encode only dominance (not precedence), as in Chomsky (1995). 10.3.1
Multiple dominance in Merge theory
In Merge theory, multiple dominance structures can be generated by allowing constituents to be merged more than once. Thus, having merged once with A to form C, a may merge a second time with B to form D, giving (17). (17)
a.
C
D α
A b.
α
C A
B
D α
B
Shared constituents and Linearization
237
C and D do not form a single object in (17); they are different sets that share a member, or in phrase-structure terms, two independent trees that share a subconstituent. In multiple dominance theory, given identity between the instances of a in each subtree, (17a) is equivalent to (17b); these are two ways of representing the same thing. With respect to projection (determination of labels), let us assume that a cannot project in only one of two merges, so that a in (17) projects either in both merges, or in neither. Thus a might be a DP that merges with two verbs to form two separate VPs, as in (18). Then DP* is a ‘‘shared object’’ of the two verbs, as in many of the RNR examples discussed here. Alternatively, a might be a verb that merges with two DPs, projecting two different VPs, as in (19). This configuration is involved in Gapping (VP1 and VP2 subsequently merge with different subjects to form clausal conjuncts sharing a verb; those conjuncts are then merged as complement and specifier of & in a singly rooted object &P). (18)
[VP1 V1 DP* ] [VP2 V2 DP* ]
(19)
[VP1 V* DP1 ] [VP2 V* DP2 ]
In (19), VP1 and VP2 are not distinguishable by reference to their labels (the label will be V* in both cases), but only by reference to their internal constitution. It is crucial to the present proposal that C and D in (17) (the output of merging a with distinct objects A and B), count as distinct even if C and D have the same label. If this were not so, (19) would be equivalent to a single ternary branching object (20). (20)
[VP* DP2 V* DP1 ]
If (19) were interpreted in this way, V* and the two DPs (or a, A and B in (17)) would enter a transitive sister relation. This would in turn imply symmetric c-command between DP1 and DP2, so that any tree containing the configuration would not be linearizable by the LCA, as noted by Kayne (1994).8 In the derivational theory, each Merge operation is assumed to create a new ‘‘root’’ node (a new syntactic object). Having two trees ‘‘floating around’’ as in (17) (or (18)–(19)) is permissible at non-final stages of the derivation. However, (17) cannot constitute the final stage of the derivation, since it violates the Single Root Condition. In Merge theory, this takes the form of a convergence condition (21b) (Chomsky 1995): (21)
A derivation converges only if (i) the Numeration is exhausted, and (ii) the output consists of a single syntactic object.
238
Chris Wilder
So a multi-rooted construct like (17) must ultimately be merged to a single object for convergence, minimally by Merging C and D to form E, as in (16) above. The Single Mother Condition does not follow from the Merge operation. Rather, its effects are achieved through the stipulation that Merge only applies ‘‘at the root’’ (i.e. to preexisting syntactic objects).9 This contrasts with the Move operation, which can merge a constituent a already embedded in an object K with the object K. (22)
Merge (unlike Move) may only apply ‘‘at the root.’’
To permit re-merge as in (17), what needs to be given up is the condition (22). 10.3.2
Relations
Two basic relations are established between constituents by the Merge operation: sisterhood and immediate dominance, on the basis of which other relations (dominance and c-command) are defined. A shared constituent enters into two sets of syntactic relations: it has two mothers and two sisters, and enters into two sets of c-command relations. In (16) above, a is ‘‘shared’’ as a daughter by C and D, and as a sister by A and B (but A is not sister of B – sisterhood is not transitive; A was not merged with B). How the treatment of shared constituents as single entities extends to syntactic relations is important, as linearization is determined by syntactic relations. Dominance and c-command are crucial in determining whether a tree containing a shared constituent can be linearized, and if so, where it ends up in the string (and where ‘‘gaps’’ appear in the string). 10.3.2.1 Sharing and full dominance Relaxing the Single Mother Condition makes available a formal distinction between two types of dominance relation: the standard dominance relation, and a stricter relation excluding shared constituents, which I call full dominance. ‘‘Sharing’’ of constituents can be thought of in terms of ordinary dominance as in (23): (23)
a is shared by X and Y if neither of X and Y dominates the other, and both X and Y dominate a.
The intuition is that a category dominates, but does not fully dominate, subconstituents which it shares in this sense with another category. In (16), C and D dominate but do not fully dominate a (while E does fully dominate a). A relation of full dominance (of a by b) can be defined on the basis of the notion of dominance paths leading from a to the root:10
Shared constituents and Linearization
(24)
239
a. A dominance path of a is a sequence of categories < C1, . . ., Cn > such that C1 ¼ the root, Cn ¼ a, and for all j (1 j n) Cj immediately dominates Cj þ 1. b. b fully dominates a iff b is a member of every dominance path of a.
A tree with no shared constituents has exactly one dominance path for each category. But a shared constituent a has more than one dominance path – each mother of a begins a different path from a to the root (multiple dominance paths also exist for any daughter of a, including a’s terminals). In (16), A and B each have only one dominance path: < E,C,A> and < E,D,B > , respectively. a has two dominance paths: < E,C,a> and < E,D,a > . Since E is a member of both, E fully dominates a; and since neither C nor D is a member of both, neither C nor D fully dominates a. Thus (16) displays the following full dominance relations: C fully dominates A; D fully dominates B; neither C nor D fully dominates a; E fully dominates A, B, C, D and a: and each category fully dominates itself. A notion ‘‘shared constituent (of X)’’ can then be defined in terms of full dominance:11 (25)
a is a shared constituent of X iff (i) X dominates a and (ii) X does not fully dominate a.
By (25), a in (16) is a shared constituent both of C and of D but not of E (the relation is irreflexive, so a is not a shared constituent of itself). The notion generalizes to cases like (26) where a is shared by larger categories containing a’s mothers (i.e. where a’s mothers are not sisters of each other). (26)
[G [ E . . . [ C . . . a . . . ] . . . ] [ F . . . [ D . . . a . . . ] . . . ] ]
In (26), a is a shared constituent of C, and of D, its mothers, but also of E and of F. In this sense, we can speak of a multiply dominated constituent a being shared by two conjuncts (corresponding to E and F) in which it is embedded. The full dominance relation figures in two places in the proposal: in the definition of c-command (section 10.3.2.2), and in the modification of the LCA (section 10.4). 10.3.2.2 C-command According to standard definitions, c-command by X and dominance by X are mutually exclusive, cf. clause (ii) of the traditional c-command definition (27). Given (27), C does not c-command a in (16). (27)
X c-commands Y iff (i) X 6¼ Y, (ii) X does not dominate Y, (iii) Y does not dominate X, and (iv) all categories that dominate X dominate Y.
240
Chris Wilder
It is however reasonable to assume that there can be such a c-command relation. Recent proposals (Epstein 1995; Chomsky 2000) have X c-command Y by virtue of Y being contained in X’s sister. Under this view, in (16), C could c-command a by virtue of a being dominated by D (¼ C’s sister). This view is essential to the proposal in section 10.4. We assume therefore that X can c-command a category Y which it dominates, as long as Y is also dominated by a sister of X, i.e. as long as X does not properly dominate Y. In (16), then, C may c-command a and B (but not A), and D may ccommand A and a (but not B). Since C and D in (16) are complex, one is a specifier or adjunct, and the other projects the label of E. Assume that D projects (if C is an adjunct, D in (16) is a segment of a complex category; if C is a specifier, D is an intermediate projection). To ensure that specifiers/adjuncts can be linearized by the LCA, the c-command definition must ensure that D does not c-command C. So we assume that lower segments/intermediate projections do not act as ccommanders. Beyond this, I will not give a concrete c-command definition.12 10.4
The LCA and multiple dominance
Taking precedence to be a relation defined only for the terminals of trees (as in Chomsky’s 1995 reinterpretation of Kayne 1994), we can think of the LCA as an algorithm of PF-component that linearizes the terminals of dominance-only trees, by reading c-command among categories (input), and delivering precedence between terminals (output).13 10.4.1
Kayne (1994)
Kayne’s proposal has two parts, a hypothesis about the relation between structure and order (28a) and a hypothesis about how that relation constrains the form of phrase-markers (28b). (28)
a. Precedence among terminals is determined by asymmetric c-command among categories that contain them: if a category X asymmetrically c-commands a category Y, then the terminals of X precede the terminals of Y b. There is a requirement for a linear ordering of terminals of a tree which determines possible forms of trees via (28a)
Given (28a), any pair of categories ordered by asymmetric c-command maps to precedence between a pair of sets of terminals. The mapping is mediated by the concept of the image of a category (29), such that the set of terminals that is the image of one category, X, precedes the set of terminals in the image of another, Y, which X asymmetrically c-commands.
Shared constituents and Linearization
(29)
241
a. The image of a category X, d(X), is the unordered set of terminals that a dominates. b. The image of an ordered pair of categories < X,Y> is the set of ordered pairs of terminals d(X) d(Y).
Consider (30), assuming K is a projection of D.14 For terminal categories (lower case), the mapping is trivial – the fact that a asymmetrically c-commands b puts a before b, and so on. The pairs boxed in (31), indicating that a, b, and c each precede d, e, and f, are determined by c-command by the complex category A (which c-commands D and all D’s daughters in (30) – a, b, and c are in the image of A, and d, e, and f are in the image of D. In this way, c-command in K determines a linear ordering of its terminals (31). (30)
K A a
D B
b
d C
E e
F
c (31)
a> (that is, it has an explicit argument slot for the domain restriction – and once this is filled it is the identity function over that domain). However, Sauerland (this volume, chapter 8) points out that filling an argument slot with different material does not usually lead to contrastive stress on the function, and so it is unclear why this should happen in this case (see chapter 8, note 1). (Notice, though, that the effect of choosing different arguments is somewhat different than is normally the case: here the effect of choosing a different argument for each pronoun means that the end result is defined for different domains. It may be that this will account for the special behavior with pronouns.) Despite the fact that it is not clear why choosing a different domain argument for each pronoun licenses contrastive stress, I also pointed out in Jacobson (2000b) that this strategy – combined with the variable-free treatment of paycheck pronouns (also known as ‘‘e-type pronouns’’) accounted for some subtle interactions between stress and paycheck pronouns.
266
Notes to page 55 Another way to implement the domain restriction idea is explored in Sauerland (this volume, chapter 8) within the context of a non-variable-free semantics. Thus Sauerland considers the possibility that each pronoun contains within it a free property variable (along the lines suggested originally in Cooper 1979), and that instantiating the variable will supply a presupposition which gives the domain restriction. But again we can ask exactly why it is that this should license contrastive focus on the second pronoun. We might claim that the focus is licensed in view of the fact that the index on each of the property variables is different. But this simply reintroduces the problems discussed above in (1); nothing would stop the two free property variables in (iv), for example, from having different indices. If, on the other hand, we say that the contrastive focus is licensed because the two property variables instead ‘‘pick up’’ different domains, then we have similar problems to the variable-free implementation of the domain restriction idea. More concretely, notice that other cases where domain restrictions are instantiated differently do not license contrastive stress: (vii) In every Linguistics department, every professor teaches four courses a year, but in every PSYCHOLOGY department, every professor/ *PROFESSOR teaches at most three.
(This is good, but only if PROFESSOR is being contrasted with, e.g., teaching assistants; it is not good on the reading where psychology professors are being contrasted with linguistics professors.) This is the classic kind of example where we find a domain restriction; professor in each case is restricted to a different domain, and yet there is no contrast. The fact that stress is impossible here thus provides a challenge both to the ‘‘identity function over different domains’’ proposal in Jacobson (2000b) and the ‘‘free property variable instantiated differently’’ discussed in Sauerland (this volume, chapter 8). Thus I conclude that there are interesting open questions about contrastively stressed bound pronouns, but at the moment it certainly provides no serious argument in favor of the use of variables (or in favor of an ellipsis analysis) over the variable-free treatment. 12. Of course there is a perfectly good precedent for this kind of situation: the treatment of indexicals is exactly a case where what they contribute to the meaning of the whole depends on the context. We have no trouble with these: we simply take the semantic values of things to actually be functions from contexts to other things. Could we import this strategy into the analysis of ellipsis? We could, but it does not seem to me to buy very much. Let us acknowledge that the full semantic value of any expression is a function from contexts to other things. Thus let us represent the value of a VP as a function from contexts to properties. I will use the notation VP0 then to refer to that function. We can now revise the rule in (37) so that the output reads as follows: (37)
.
Notes to pages 58–60
267
In this view, elided material is indexical. While I know of no real problem with this view (other than its obvious complexity), I see little advantage to it either (and so since it is more complex than the alternative view which I will put forth below, the alternative is, at this point, the preferred option). The problem with any rule like (37) or (370 ) – which requires only semantic identity between the elided material and something else – is that it undermines the entire raison d’eˆtre for positing null material in the syntax in the first place. The reason that many authors over the years have argued for actual linguistic material in the position of the ellipsis site is based on the premise that constraints which can only be stated on linguistic form are found to hold at the ellipsis site. But of course we can’t really see what material is there; we infer it in terms of the form of the ‘‘antecedent.’’ But if we are required only to have identity of meaning and not formal identity with the antecedent then nothing actually ensures that the ‘‘elided’’ material does have the right form (as long as there is some other form that can express the same meaning, that other form should do just as well). Thus any approach requiring only semantic identity does not, in fact, succeed in accounting for the supposed constraints on form – and so there is no real reason to posit silent material in that position. My own solution to be discussed below also does not require any kind of identity of form – and also therefore does not account for these apparent formal matching effects. The general program here, however, would say that any time we appear to find constraints on form which need to be satisfied in the ellipsis site, they cannot really be constraints on form. But I think that any semantic identity approach is going to have to say, in the end, exactly that. 13. Notice, incidentally, that (as mentioned in section 3.2.4), cases like (b) and (c) above are just those that were meant to be excluded by Sag’s condition on ‘‘alphabetic variants.’’ But, as also noted above, it is clear that this condition was simply incorrect. These would also pose a challenge for a (hypothetical) modern reconstruction of this condition which used (a) Heim’s condition on No Meaningless Coindexation (Heim 1997), combined with (b) an insistence on complete identity between the antecedent and the elided material, including identity of variable names. (Heim does not require this, and requires this only for the purposes of Rooth’s focus condition [Rooth 1992a]: for identity she allows identity up to variable names. For further discussion of this issue, see Kennedy [this volume, chapter 5].) 14. Many accounts within the Categorial Grammar tradition build the intersection semantics into the meaning of which such that it takes two arguments (the relative clause and the noun) and intersects them. In the first place, though, this provides no simple semantics for cases of Pied Piping; I will suggest in the next section that which is (not surprisingly) an ordinary pronoun, allowing a simple analysis of Pied Piping. Second, this strategy would still need to be supplemented with either a unary rule, silent operator, or new binary combinatory possibility for the case of relative clauses with no overt relative pronoun (e.g. every book Bill read) and so would seem to have no advantage.
268
Notes to pages 61–90
15. There is an interesting remaining question, which is why an extraction ‘‘gap’’ can’t just keep passing up. Put differently, why can’t S/NP occur as a root node, while SNP can? I leave this question unanswered. Note also that in a case like John voted for every girl who thought he would the NP final result could indeed be either S(VP/NP) or SðVP Þ . It makes no difference which of these we get as the end result; they have the same meaning and are both looking for a contextually salient two-place relation. 16. Jacobson (1998) is also concerned with the interaction of this case with Rooth’s focus condition. The discussion there, however, is really neutral between the direct compositional/variable-free account and the standard account and so will not be repeated here. CHAPTER 4: THE VIEW OF QR FROM ELLIPSIS 1. See McCloskey (1991), Lobeck (1995), Lo´pez (1994; 1999), Doron (1999), and Goldberg (2005). 2. See Tancredi (1992), Rooth (1992a; 1992b), Fox (2000), and Merchant (2001). 3. See Coppock (2001), Ackema and Szendro† i (2002), and Lin (2002) for arguments directly targeting my proposal, and Hartmann (1998) for general arguments on behalf of an ellipsis account. 4. See Larson (1985). 5. Indeed, it must be prevented from moving a VP out of another VP, for otherwise it would generate ‘‘V Object V’’ word orders in English. This, plausibly, is how such word orders are generated in certain verb-final Germanic languages; see Zwart (1997), Koopman and Szabolcsi (2000), etc. 6. See Bruening (2001) for this point, and for an account. 7. I find (42) marginal, considerably worse than (43). Because I will suggest that (43) derives from (42), there is something missed by my proposals. 8. That is, there is no way of forming such a coordination that does not involve moving both the eat-VP and kale out of the first conjunct in violation of the Coordinate Structure Constraint. 9. See Lasnik and Saito (1992), Muller and Sternefeld (1993); Muller (1996) and ¨ ¨ Takano (1996). 10. That VP Scrambling of this sort is (at least marginally) possible in German, Muller establishes by way of examples like (i). ¨ (i) . . . dass das Buch2 mal wieder [VP t2 gelesen]1 keiner t1 hat. . . . that the book once again read no one has. ‘‘. . . that once again no one has read the book.’’ (Muller 1996, (47b): 389) ¨ 11. See Collins (1994) and Takano (2000). 12. See Pesetsky (1987). 13. We must understand the prohibition on mixing types of movement that (50) expresses to be global. For it to block remnant VP topicalization in English, it is necessary that it prevent a derivation that has the following steps: (1) an object moves from VP, (2) that VP moves to FP and (3) the VP topicalizes. The steps (1) and (2) are what I’ve suggested underlie Gapping, and so they should
Notes to pages 91–106
269
be permitted. We must understand (50) to block step (3) when step (1) is A movement. 14. See Lebeaux (1988; 1990), Heycock (1995), Fox (1999b), and many others. 15. See Chomsky (1982). 16. This example is adapted from one in Baltin (2003). CHAPTER 5: ARGUMENT CONTAINED ELLIPSIS 1. For now, I will use the term ‘‘identical’’ informally, without giving an explicit definition, and will assume that the basic idea is fairly intuitive: in the case of referring DPs, the subjects of the related VPs should corefer; in the case of quantified DPs, the assignments of values to the subjects should be the same. An explicit statement of this idea, which will play an important role in the development of the analysis, will be developed in section 5.6. 2. The judgments of these sentences reflect the responses of a majority of informants, but it should be noted that some speakers do not find the examples in which subject identity does not hold to be robustly ungrammatical. In particular, for many people, addition of too or instead to the end of the ungrammatical examples tends to make them somewhat more acceptable, although the majority of my informants still consider them unacceptable. (i) a. ??Every man who wants George to buy some salmon should too. b. ??The woman who thought the man from Portland would catch a tuna did instead. There is no comparable improvement in acceptability with the addition of ‘‘too’’: (ii) *A proof that God exists does too. As has been noted frequently, VP Ellipsis is extremely sensitive to contextual factors. It is possible that the improvement in acceptability of examples like (ia) and (ib) results from the fact that words like too and instead, in these constructions, force the hearer to assign an interpretation to what is nevertheless an ill-formed construction. See Hankamer (1978) and Higginbotham (1985) for relevant discussion of this issue. 3. These data are highly problematic for a theory of ellipsis that incorporates some notion of Sag’s 1976 constraints on alphabetic variance. Generally speaking, VPs that contain variables bound by different VP-external operators, which is the case in (30a) and (30b), are not alphabetic variants. It is an interesting property of the analysis of ellipsis that will be developed here that these and similar examples are expected to be grammatical (the Hirschbuhler sentence ¨ may still be problematic). I will not discuss these issues in great detail, but I will return to them briefly in section 5.7. 4. Arguments in favor of this hypothesis, which include the inability of ellipsis to be licensed by context (in the absence of a linguistic antecedent), parallelism requirements between the elided and antecedent VPs, and ‘‘missing antecedents’’ facts, are presented in detail in Hankamer and Sag (1976). See Dalrymple (1991) for counterarguments, however.
270
Notes to pages 106–111
5. Let us assume that a sentence can be construed as an ordered set of structural descriptions hSD1 ; SD2 ; . . . ; SDn i. Under this interpretation, a licensing condition such as the one proposed below states that a null VP is licensed in the phonological representation SDj of some sentence only if it is present in that sentence’s Logical Form SDi and is identical to a VP in the logical form SDi0 of some sentence already in the discourse. In principle, the necessary relation could obtain either through deletion of syntactic material at PF or through reconstruction at LF; to maintain generality, I will not attempt to choose one view over the other in this chapter. 6. (31) is very similar to Sag’s 1976 statement of VP Deletion; the main difference between this approach and Sag’s is that (31) requires strict identity, whereas Sag required alphabetic variance of lambda expressions, a form of what I have referred to as loose identity. 7. I assume the VP internal subject hypothesis (Kitagawa 1986; Koopman and Sportiche 1991; Kuroda 1988), and in particular, the VP structure defended in Huang (1993), in which the subject originates in SpecVP and moves to SpecIP to receive Nominative Case. Moreover, I assume SpecVP to be the position to which V0 assigns its external theta role. Any model of ellipsis that utilizes a syntactic, identity-based theory of VP reconstruction must account for the fact that all VP internal traces must have antecedents (a requirement that can be formalized in a number of ways, e.g. the Empty Category Principle, the definition of chain). In particular, the well-formedness conditions imposed on the subject trace in the specifier of the antecedent VP must be met by the parallel trace in the specifier of the LF representation of the elided VP. 8. There are three logically possible indexing configurations: inner only, outer only, inner and outer. I will represent these configurations in the following way: Xi (inner only), X:i (outer only), Xi:j (inner/outer). 9. Two things should be said about (37a) and (37b). First, by assuming a general constraint against vacuous quantification, (37a) can be understood as including an implicit minimality requirement. That is, in configurations of the form shown in (iii), where i and i0 are like-valued inner indices and i0 is c-commanded by both X:i and Y:i, i0 is bound by Y:i, even though (37a) does not literally prohibit i0 from being bound by X:i. (iii) [X:i . . . X0 i . . . Y:i . . . Y0 i0 . . . ] Second, (37b) is a modified version of Heim’s original translation definition. Heim states that only definite DPs, pronouns, and traces bear inner indices. For the purposes of this chapter, I will assume that indefinites may also bear inner indices, which, when bound by existential closure, introduce a discourse referent (see Heim 1982; Diesing 1992b; and note 12 below). Indices on definites and indefinites are subject to the Novelty – Familiarity Condition, however (Heim 1982). 10. Due to space considerations, I will not give a detailed presentation of how various ambiguities associated with elliptical structures are to be accounted for within a strict identity model (i.e. strict and sloppy interpretations of pronouns); see Kennedy (1994) for a survey of such constructions. The analysis is essentially the same as that in Reinhart (1983a), in which strict readings are correlated with interpretations of VP internal pronouns in which they are
Notes to pages 111–114
271
coreferential with a subject, and sloppy readings with bound variable interpretations of VP internal pronouns. The analysis proposed here will inherit any problems with such an analysis of strict – sloppy ambiguity (see Wescoat 1989 for some particularly problematic examples), but this is not an issue that I will address. The goal of this chapter is not to promote a particular theory of ellipsis, but to propose that elliptical dependence be explicitly represented within a theory of ellipsis. 11. According to Heim, the sister constituent of the moved DP is then prefixed with a lambda operator which bears the outer index of the moved DP. I will assume here that prefixation of the lambda operator to the sister of the moved DP is part of the process of mapping the syntactic representation of an utterance’s Logical Form onto its semantic interpretation. The lambda operator is a logical entity, and presumably not overtly represented in the syntax. The scope of the lambda operator in the semantic interpretation of a given LF is predictable: it is the c-command domain of the prefixed DP. 12. (39b) is not the complete LF of (39a). In particular, I have not shown the indices on the indefinite DPs. The full representation of the LF of (39a) is given in (iv). (iv) Leo2:1 [VPt1 drives a Dart4] and Felix3:1 does [VPt1drive a Dart4] too. (iv) reflects the fact that the strict identity hypothesis as stated in (31) requires the indefinites to bear the same inner indices. The reader may now justifiably point out that according to the claim in section 5.3.2 that all free inner indices are assigned the same referent, (iv) should mean that Leo and Felix drive the same car. Although this is a possible interpretation of (iv), there is also an interpretation in which Leo and Felix drive different cars. On the surface, it appears that the proposed analysis does not allow this reading. There is a way out of this problem, however. The reading in which Leo and Felix drive different cars results when each occurrence of the indefinite DP a Dart in the two VPs introduces a discourse referent. In a theory of the syntax – semantics interface such as Heim (1982) or Diesing (1992b), indefinites that introduce discourse referents are bound by existential closure. We can therefore say that on the relevant reading of (iv), the inner indices of the two indefinites are not free, but bound by existential closure. Since these indices are not free, they do not corefer. 13. This sort of example was first discussed by Jacobson (1977). Jacobson’s example is (v), which is arguably a better illustration of the problem than (45), since the latter is pragmatically a bit odd, since it means that the woman who married her husband loves her husband, which is typically assumed to be true (though of course is not always true). (v) is just as bad as (45) on the intended reading, but doesn’t have this pragmatic interference. (v) His wife killed her childhood sweetheart. 14. In Higginbotham’s system, anaphoric relations are represented with headed arrows. These arrows directly parallel Heim’s inner and outer indices: the tail of an arrow corresponds to an inner index; the head of an arrow corresponds to an outer index.
272
Notes to pages 115–136
15. Alternatively, we could define two types of antecedents: one with respect to linking (coreference) configurations; one with respect to binding configurations, and retain Higginbotham’s definition of dependence. The end result would be the same, but would not reflect as clearly as refining the notion of dependence does the fact that we are really making a distinction between binding and coreference. 16. By allowing outer indices to appear on every maximal projection in the extended projection of N, i.e., on both DP and NP (see Grimshaw 1991), we can assume that both DP and NP are potential binders, modulo the c-command constraint on binding. 17. Note that a structure in which the higher NP is the binder of an element inside YP is ruled out for the same reason that DP cannot be a binder – the bound element in YP would be b-dependent on itself. 18. The basic idea behind Higginbotham’s analysis is that in the translation of a structure like (62) into a logical representation, the relative clause is mapped into the restriction of the determiner every, while the VP is mapped into its nuclear scope (see Heim 1982; Diesing 1992b). Empty categories bound by the DP in the matrix sentence and those bound by the relative operator in the relative clause are interpreted as cobound, i.e. bound by the determiner every. A more intuitive, semantic representation of (62), where the variable x has been substituted for the index 2, is given in (vi). (vi) 8x½manðxÞ ^ cleanðx; x0 s DartÞ½happyðxÞ 19. For clarity, I will only show indexing on the relevant DPs (the ones involved in the argument containment configuration). Although the subjects of the matrix VPs will necessarily bear the same outer indices (because the matrix VP is what is copied), this index does not play a role in the b-dependence relation involved in the ungrammatical sentences. 20. The analysis proposed in this chapter has no account of the ungrammaticality of (76) in terms of ill-formed indexing configurations (and would, in fact, predict it to be grammatical). Given the grammaticality of examples like the ones in (77), I will assume that an explanation of (76) will be found outside of conditions on indexing configurations, and will leave a resolution of this issue for future work. CHAPTER 6: VARIABLE ISLAND REPAIR UNDER ELLIPSIS 1. The Sluicing correlates that will interest us for the moment are all indefinites; see Chung et al. (1995) for discussion. Other correlates are possible (the ‘‘contrast’’ sluices of Merchant 2001, for example), but these show strong locality effects (stronger, in fact, than mere sensitivity to islands: closer to the kinds of locality found in Gapping and multiple Sluicing). The picture is further complicated by the fact that certain Sluicing-like structures in some languages appear to retain island-sensitivity, as reported in Hoji and Fukaya (1999). 2. For Sluicing and selective (‘‘weak’’) islands, see Albert (1993), Sauerland (1996), Romero (1998), Merchant (2000).
Notes to pages 140–158
273
3. This example shows that such contrast-licensed wh-extraction from VP Ellipsis can cross tensed clause boundaries; nevertheless, such examples are highly sensitive to the nature of the intervening material: if the embedded subject is not bound by the matrix subject, the matrix reading for the elided VP becomes extremely difficult to get (cf. identical bound subject restrictions in cross-clausal Gapping and multiple Sluicing as discussed in Nishigauchi 1998; Johnson 2001; and Merchant 2001). CHAPTER 7: ON BINDING SCOPE AND ELLIPSIS SCOPE 1. See Fiengo and May (1994), Larson and May (1990), Sag (1980), among others. 2. The two dichotomies, wide vs. narrow scope, and de dicto vs. de re, are well known to match only partially (wide scope implies de re, but narrow scope is compatible with both de dicto and de re). This inaccuracy does not affect the exposition in the text, though, as the diagnostics to be used do not test for referential transparency. 3. See, e.g., Hankamer (1973b) and Pinkham (1982). In what follows, than will – for expository convenience – be bracketed as a coordinator in a ternary structure. For discussion of coordinate properties of comparatives and the structure of comparative coordination see Lechner (2004) and references therein. 4. Wide scope in (8) cannot be attributed to a non-scoping mechanism (choice functions or non-local binding of world variables for de re readings; see section 7.2.2), as these devices do not evacuate the relative clause from the c-command domain of the antecedent. Moreover, on the standard assumption that VPE may only elide a single node, and not e.g. multiple terminals, reading (8b) leads to regress, as the antecedent still contains the ellipsis site. 5. Similar ambiguities have been discussed in McCawley (1998: 688) and Pinkham (1982: 130). McCawley assumes that the two readings in examples similar to (11) are not derivationally related, but arise from different deep structures. Pinkham considers (i), and concludes that reading (ia) is the product of a construction specific deletion process, while (ib) represents a base-generated PC which does not include any elliptical structure. As will be seen shortly, both views differ from the one advanced here, according to which all PCs derive from a clausal source (potentially a small clause, as in (ib)). (i) John seems taller than Bill D. a. D ¼ seems d-tall b. D ¼ is d-tall 6. In the narrow reading (¼ (11a)), the verb is also underspecified for tense. Note incidentally that atemporal readings are subject to the additional, curious restriction that the correlate has to be structurally higher than the comparative NPs (see Lechner 2004 for a partial analysis): (ii) John will subject more students to this year’s exam than D to last year’s exam. D ¼ John will subject/*subjected d-many students
274
Notes to pages 161–164 (iii) John will subject this year’s students to a harder exam than D last year’s students. D ¼ John will subject/subjected last year’s students to a d-hard exam
7. The than-XP must in some way be allowed to escape the adjunct. This problem is germane to all analyses of adjunct comparatives which scope parts of the DegP (such as e.g. Heim 2000). 8. I borrow here the terminology of Postal (1974). One classic example motivating double indexing in comparatives is (iv), which combines pronominal variable binding with a consistent, de re reading for the than-XP (see Hoeksema 1984; Heim 1985; Kennedy 1995; Rullmann 1995). Variable binding controls for the LF-position of the than-XP, ensuring that it cannot take scope above believe. Moreover, the de re reading requires that the world variable of the than-XP be identified with the matrix world w. These apparently conflicting requirements for the position of the than-XP can be satisfied by anchoring the embedded world variable directly to w0. (iv) We believedW0 that no/every problemj wasW harder [than itj wasw0] 9. Maribel Romero (p.c.) notes that (15a) might also possess a reading in which the world variable is bound locally. If this intuition turns out to be correct, no stipulatory device needs to be added. 10. In addition, (22) lacks the wide ‘‘mixed’’ construal (v), sponsored by TR to the matrix IP node. In (v), the comparative NP functions as the ECM-subject, while the CD-site is understood as the matrix subject: (v) *D ¼ d-many people wanted to write The ban on (v) falls out from a more general restriction, though. As noted first in Hankamer (1973a), PCs lack mixed readings such as (vib), which associate the CD-site and the comparative NP with distinct grammatical functions. This constraint also filters out reading (v). (See Lechner 2001, 2004 for a possible account of Hankamer’s generalization.) (vi) The Irish singers liked more Estonian singersObject [than-XP than the French singers] a. than-XP ¼ than the French singers liked [d-many Estonian singers]Object b. *than-XP ¼ than [d-many Estonian singers]Subject liked the French singers 11. For most speakers, subject remnants in non-predicative PCs cannot surface with accusative case: (vii) *John wanted to write more plays than me. Thus, predicates inside the than-XP of object PCs are always finite. For a (partial) account which does not rely on finiteness see Lechner (2004: chapter 8). 12. This follows from the parallelism requirement on the comparative NP and the CD-site (see note 9), which forces both the comparative NP and the CD-site of (23a) to be accusative ECM-subjects.
Notes to pages 165–174
275
13. The argument rests on the assumption that nominative strictly correlates with finiteness. This view might turn out to be too strict, e.g. in case the system needs to incorporate nominative default case. Lechner (2004: chapter 8) pursues an alternative approach, which is compatible with the latter conception, and which links the contrast between ECM and object comparatives to differences in the respective structural relations between the CD-site and the remnant. 14. There is a second derivation for the wide reading (22b) which involves high comparative coordination, ATB-subject movement of John and long-distance Gapping of the string wanted to write. (viii) Johnk [[IP tk wanted more people to write plays] than [IP tk wanted dmany people to write poems]]. The derivation in (viii) proves unproblematic, though, because the embedded predicate is non-finite, precluding double indexing. Note also the alternative representation for the narrow reading (22a). 15. As for the fine-grained structure of (27a), assume that the than-XP is construed as a small clause embedding the remnant and the CD-site, as in (ix): (ix) . . . than himself [d-tall] 16. But note that although VPE in isolation does not obey Isomorphism, VPE in ACD does so if ellipsis is correlated with principle C. That is, it is not possible to assign (x) a reading that mixes wide QR (bleeding Principle C) and narrow ellipsis. (x) I expected himj to buy everything that Johnj thought I did. (Fox 1999b: 187, (74)) 17. For one, quantifiers inside the than-XP show a strong – but not absolute (see Rullmann 1995) – tendency toward taking scope at least higher than the existential operator binding the degree variable (see Kennedy 1997a; Hoeksema 1984; Heim 1985; Rullmann 1995; Heim 2000; Schwarzschild and Wilkinson 2002). This has the consequence that relative scope of quantificational terms inside the than-XP cannot be taken to directly reflect the position of the than-XP. Moreover, the binding domain of anaphors inside phrasal thanXPs extends in many cases into the matrix clause (see above). Thus, disjoint reference effects offer the most reliable test for the scope of remnants inside the than-XP. 18. (33a)/(34a) also testifies that Gapping differs from VPE in that it fails to license vehicle change of names into pronouns (Fiengo and May, 1994). 19. But also the relative clause in (xi a) arguably requires some mechanism of overt covert movement, as the relative clause must be attached at a node which supports ellipsis resolution only at LF (see (xi b)). (xi)a. ? John believed [everyone you did D] to be a genius. b. [[John [VP believed t to be a genius]] [everyone you did D]] D ¼ [VP believed t to be a genius] 20. It should be noted that Heim compares the direct analysis with ellipsis approaches, but does not endorse one over the other.
276
Notes to pages 176–188
21. Thus, the intermediate status of (47)a – as opposed to (48a) – can be related to the fact that its derivation only fails to obey the ban on Gapping of finite sentence boundaries. 22. More correctly, the empty subject pronoun inside the than-XP is not elided by Gapping, but has been removed by ATB-movement. Nothing bears on this issue, though. See Lechner (2004) for details. 23. Note on the side that (53b) cannot be excluded as a violation of Secondary SCO (Postal 1971; 1993). Secondary SCO prohibits movement of an embedded r-expression over a coindexed term, as in (xii). However, given the plausible assumption that cross-over restrictions are conditions on the formation of operator-variable chains, and given that the remnant does not bind a variable (movement of the remnant is semantically vacuous), Secondary SCO does not apply in (53b). (xii) Whosej sister did hej like tk. 24. In fact, this is more in line with Heim (1985: 52), who generates than-XPs in their surface positions. 25. Example (xiii) indicates that it is necessary to adopt the additional assumption that the than-XP is transparent for c-command, which allows the subject remnant Peter to c-command the pronoun: (xiii) *She introduced himi to more people than Peteri,SUB Given this premise, the two accounts become now empirically indistinguishable as far as their capacity to handle binding facts is concerned. They only differ in that while the ellipsis analysis copies all categories from the matrix clause (except for the correlate) around the remnant, the direct account ‘‘squeezes’’ the remnant into a slot right above the correlate in the matrix clause. 26. See Lechner (2001; 2004) for additional arguments against the direct analysis from word order and surface properties of ellipsis. 27. On the current analysis, TR – which is a prerequisite for Gapping – could apply either subsequent to reconstruction of the object at LF or by ‘‘covert overt movement’’ prior to topicalization. Crucially, lowering can be avoided. CHAPTER 8: THE SILENT CONTENT OF BOUND VARIABLE PRONOUNS 1. A third idea, which Jacobson (2000) develops, is to analyze all pronouns in a combinatorial semantics as identity functions on the domain of individuals, but with an additional open argument position. In this system, all pronouns uniformly denote the same functor of type 55e; t4; 5e; e44. Jacobson’s intuition is that the additional open argument position is an open contextual argument of a sentence containing a pronoun, but that the value it is applied to serves to license focus on the pronoun. However, the intuition is incompatible with standard ideas about focus semantics, and Jacobson does not develop a focus semantics compatible with her proposal. It seems to me that such an account would need to treat pronouns in a special way: Examples like (i) show that the prediction of standard focus semantics that two identical functions
Notes to pages 194–202
277
cannot be contrastively focused if the arguments they apply to differ is in general correct. (i) *Mary is eating cabbage, and then she is EATING SQUASH. 2. As pointed out to me by Mats Rooth (p.c.) and Marga Reis (p.c.), however allows most easily a hat intonation of the type discussed by Krifka (1999) among others as topic-focus intonation. As far as I can see, however, my argument is not affected by the difference between hat intonation and a double focus intonation. Krifka proposes a treatment of association with topic that is analogous to association with focus. Therefore, I do not distinguish between the two intonations in the text. 3. An interesting puzzle is that, when adjoined to a sentence, however requires only one contrast, as shown by (ii) in contrast to (32d). I have at present no idea how to relate these two uses of however. For some speakers of English, (32d) is marginally acceptable – I assume that they can left-adjoin however to VP, rather than having to left-adjoin it to the subject NP. In German examples with hingegen (‘‘however’’), this VP-adjunction analysis can be controlled for because of the verb-second property. The English facts in the text can all be reproduced in German. (ii) John thinks that Carl called Mary. However, Carl called BERta. 4. The need for an e-type representation has only been demonstrated in cases where there is an antecedent with a bound pronoun in the parallel position around. The other alternative to consider is that there is an antecedent like (iii) where the parallel position is occupied by some material other than a bound pronoun – if there is no antecedent parallel up to the focused constituents, destressing of all the other material would not be licensed. (iii) (Only) John brought something Mary likes. It is quite easy to see that when the index-insensitive focus licensing condition (29b) is adopted, the focus on the bound pronoun is required only on the e-type analysis. If the index-sensitive condition is adopted, the issue is more complicated. Under the assumption that then only the whole clause is considered as an FD, it follows that then, too, the focus is licensed only on the e-type analysis of the pronoun. 5. Kratzer (1998, (23)) observes a similar interaction of focus and binding in the example (iv). Her account, however, is very different from mine and does not extend to the examples in the text. (iv) a. Only I answered a question that you did not think I could answer. Nobody else answered a question you did not think I could answer. b. Only I answered a question that you did not think I could answer. Nobody else answered a question you did not think he could answer. 6. It remains to be seen whether all examples behave as predicted. Consider (v), which was provided by an anonymous reviewer. (v) Discourse: Almost every contestant used a battery to power his car. a. One JAPANESE contestant, however, used a match to power HIS car.
278
Notes to pages 205–215
In this case, the contrast might be based on the content this Japanese contestant contrasting with the contestant different from this one Japanese contestant. However, it is difficult to see how this goes together with the semantics of almost. 7. In fact though the situation may be different if one assumes an analysis of ellipsis in terms of semantic licensing. Then the value of Cooper’s relation variable could act to semantically license ellipsis. It seems to me that then Cooper’s analysis would not have any possible empirical advantage. 8. Hans Kamp (p.c.) reports that, in his judgment, the pronouns remain focusable even if hate is replaced with love making the two relative clauses have the same lexical material. While I find (76) odd after this replacement, regardless of focus placement, my judgment concerning (vi) (or rather its German translation) concords with Kamp’s judgment. This may support the conclusion that elided material alone cannot explain the focus licensing since both occurrences of his should contain the same elided material. (vi) A man who loves his mother should talk to ANOTHER man who loves HIS mother.
C H A P T E R 9 : A STEP-BY-STEP GUIDE TO ELLIPSIS RESOLUTION 1. As for NP movement out of a DP, the Split Topicalization in German, which can strand a quantifier, might be considered as movement of that kind. 2. Sluicing is probably the least likely case of ellipsis to be represented as a proform. It would require some serious semantic tools, such as double indexing or addition of assignment functions to basic types. For extensive discussion on the presence of LF structure in Sluicing, see Merchant (2001). 3. The weak crossover effects in preposed adjuncts turn out to be rather complicated. If the quantifier in the matrix clause is in the subject position, variable binding seems possible, as discussed in Reinhart (1976; 1983a) and Chierchia (1995). (i) (?) If his1 favorite toy is missing, [every boy]1 will cry. This example shows that the reconstruction of a fronted adjunct is in principle possible (see Chierchia, 1995, chapter 4 for discussion). Let us assume that the adjunct clause is adjoined back to VP. The following LF structures illustrate the subject–object asymmetry. The numeral next to the moved NP is an index binder in the sense of Heim and Kratzer (1998). (ii) a. Binding by the subject possible [IP [Subj NP] 1 [VP [VP t1 . . . ] [Adjunct CP . . . pronoun1 . . .]]] b. Binding by the object not possible [IP [Subj NP] 1 [VP [VP t1 Object NP ] [Adjunct CP . . . pronoun . . .]]] If the A-movement of the subject has the effect of Predicate Abstraction, then, the pronoun in the CP is successfully bound. However, the pronoun is not bound by the object NP in (ii b). In order to establish the binding by the object, the object must raise. Imagine that this additional movement creates Weak
Notes to pages 215–220
279
Crossover. Now turning back to VP pronouns, sentence (11) has the following LF after the reconstruction of the if-clause. (iii) [IP [VP [VP quit smoking ] [Adjunct CP . . . [VP D]]]] As was the case with the object NP, (iii) requires the movement of the matrix VP so as to yield the bound variable VP interpretation. In this sense, the sentence should pattern with the example (13), rather than (i). The availability of sloppy reading in (11) therefore suggests that the elided VP is not a bound variable. 4. It should be noted that (18b) suggests that Schuyler’s generalization does not completely align with the MaxElide account of it in Jason Merchant’s chapter (6) in this volume (see also Takahashi and Fox 2006). (18b) has undergone the maximal ellipsis (i.e. that of the VP supposed to read t) which nonetheless leaves no focused constituent between the relative operator and its trace, violating Schuyler’s generalization. 5. Unfortunately I could not construct a sensible example with CD. In the ACD example, the negative auxiliary is contrasted with the affirmative one. However, the same strategy does not work for CD cases because a than clause cannot host negation (e.g. *Dan read more books than he didn’t want to). See Rullmann (1995) for relevant discussions. The best that I can think of is (iv). (iv) Fred bought more books than Eric asked him to D, and then he READ more books than BILL did D. It still seems very hard to get the intended sloppy reading for the second ellipsis (i.e. D ¼ ask him to read tdegree 1 many books). It is perhaps due to a much easier resolution being available (i.e. D ¼ read tdegree 1 many books). I do not know how to control for this factor. 6. I am grateful to Nina Zhang and Akira Watanabe for reminding me of this. 7. There are some alternative ways to get the same result. Imagine, for instance, that an e-marked VP can optionally undergo raising at LF. Let us use (2) as an illustration. The LF structure of (2) will be (v). (v) A: . . ., he [does not [VP [VPE cook]1 [VP want to t1]]] B: . . ., he [does not [VP [VPE clean]2 [VPE want to t2]]] With this structure, we no longer need to make reference to any e-marking in the identity condition. All e-marked VPs have appropriate antecedents (we still keep the ‘‘up to indices’’ clause in the condition). Then, the PF outcome is exactly what we have in (ii). (vi) a. . . ., he [does not [VP [VPE cook]1 [VP want to t1]]] b. . . ., he [does not [VP [VPE clean]2 [VPEwant to t2]]] This is certainly a feasible alternative to (27). However, one of my earlier criticisms of Schwarz’s analysis applies here as well: The movement of a smaller VP seems insensitive to islands. For example, (vii) has the sloppy reading. (vii) When John has to cook, he finds someone else who is willing to D, but when he has to clean, he doesn’t D.
280
Notes to pages 224–234
To license this reading with (iv), the VPs cook and clean must raise out of the complex NPs that contain them. In addition, it is necessary to restrict this movement to an e-marked VP only. The LF of (5) would be just like (vi) above except that [VP cook] is not e-marked. That would not affect the deletability of the two VPs in B’s sentence, however. Therefore, the lack of the sloppy interpretation in (5) would not be explained. It is unclear at best whether this kind of movement constraint is well motivated. 8. It is not clear to me whether the F-marking indicated in (42) is the only possibility. The native speakers that I consulted do not agree on this. However, what is important for our current discussion is the fact that prosodic prominence on the second occurrence of cook is possible. 9. For an alternative analysis, see Tomioka (2001). CHAPTER 10: SHARED CONSTITUENTS AND LINEARIZATION 1. This material was presented in various forms in 1998–9 at ZAS Berlin, Tubingen, Konstanz, MIT, CUNY, UConn, Tucson, and UCLA. My thanks ¨ to all audiences for helpful comments, and especially to Hans-Martin Ga¨rtner, Alan Munn, and Michal Starke. 2. For arguments against a rightward ATB movement analysis of RNR constructions, see Wilder (1997) and references cited there. 3. The adoption of multiple motherhood (the ‘‘joining nodes’’ of Muadz 1991) to represent constituents sharing is only part of the story in ‘‘three-dimensional’’ theories of coordination. The other main ingredient is the assumption of ‘‘splitting nodes’’ (Muadz 1991; Moltmann 1992) as a way of representing the conjuncts themselves. The splitting node proposal is inherited from Goodall’s (1987) idea that coordination involves set union of phrase-markers, and is used by Muadz and Moltmann to express the intuition that conjuncts exist ‘‘in parallel’’ (cf. Williams 1978). This idea is not part of the present proposal. 4. The same word order argument for a forward ellipsis rule is provided by shared subject coordinations in German and English of the type (i)–(ii), where the shared subject follows the verb and other material belonging to the first conjunct; and by German examples like (iii), involving a shared object preceded by an adverb belonging to the first conjunct only. (i) [ in den Wald ging ein Ja¨ger] und [ – SU fing einen Hasen] into the wood went a hunter and caught a hare ‘‘A hunter went into the wood and caught a hare’’ (ii) [out of the woods came a bear] and [– SU attacked us] (iii) . . . weil ich [vorgestern ein Buch abholte] und [gestern — OB because I day-before-yesterday a book picked-up and yesterday zuruckbrachte] ¨ brought-back ‘‘because I picked up a book the day before yesterday and returned it yesterday’’
Notes to pages 234–240
281
On (i), see especially Hohle (1990); cases like (ii) are discussed in Heycock & ¨ Kroch (1994). An analysis along the lines of (13) is conceivable also for (i)–(iii), whereby the shared XP (the subject in (i–ii), the object in (iii)) undergoes ATBmovement out of coordination, while other constituents preceding the shared XP are raised out of the first conjunct only. 5. For arguments that Gapping dependencies are indeed movement dependencies, see Johnson (2003). If the ATB-movement approach to Gapping should turn out not to be correct, then either (i) the shared constituents involved do not involve multiple dominance, but distinct copies, and we need a forward ellipsis rule after all, or (ii) these shared constituents do involve multiple dominance, and (some part of) the Linearization story told here is incorrect. 6. This formulation presupposes the copy-and-deletion theory of movement. As discussed in section 10.6, once the Single Mother Condition is given up, there are good reasons for treating movement dependencies as multiple dominance, i.e. there are no ‘‘copies’’ to be deleted. A movement chain consists of a single multiply dominated category, linearized in a way that replicates the effects of (15). 7. The relations of dominance, immediate dominance and sisterhood obtain as usual. Sisterhood is not transitive: taking sisterhood to be defined as (iv), A and a in (16) are sisters, B and a are sisters, but A and B are not sisters. (iv) X is sister of Y iff there is a Z s.t. Z immediately dominates both X and Y. 8. The assumption that VP1 and VP2 in (19) are distinct objects despite having the same label is reasonable. It is made also in the case where DP2 in (19) is not merged with V*, but with VP1, yielding a second distinct object with the same label, containing the first as a term, as in (v) (in which VP1 ‘‘no longer’’ counts as maximal): (v) [VP2 DP2 [VP1 V* DP1]] 9. Chomsky (1995) distinguishes between syntactic objects, whole trees that have not (yet) undergone merger, and terms: whole trees (current objects) and their subconstituents (one-time objects that have undergone merger). 10. In this formulation, full dominance is a reflexive relation. 11. A difference between (23) and (25) is addressed in section 10.6. 12. Among the issues left open is that of c-command by adjoined heads. According to the ‘‘sister containment’’ view of c-command, X in (vi) does not c-command its trace. (vi) [[Y X Y ] [ . . . tX . . . ]] For concreteness, in this chapter raised heads are treated as if their c-command domain is that of the head to which they adjoin (as in Kayne, 1994). 13. Some theories of phrase structure adopt a dominance-precedence axiom forbidding trees with ‘‘crossing branches.’’ The multiple dominance structures proposed here are depicted in tree diagrams in which branches appear to cross, but the notion is meaningless in a theory in which trees do not encode precedence; what matters is whether the terminals are linearizable or not.
282
Notes to pages 241–251
14. There are differences between Kayne’s proposal and Chomsky’s Merge theory that are not important here. For concreteness, I assume that terminal categories are lexical items (cf. Chomsky 1995). Merge theory does not permit vacuous projection, so a substructure like (vii) is not licit. Chomsky proposes that asymmetric c-command (hence precedence) is established between terminals that are sisters by having one raise to a position where it asymmetrically c-commands the other (crucially, traces do not need to be linearized). Thus, the ‘‘vacuous projection’’ notation (vii) is understood as shorthand for (viii). (vii) [B b [C c ]] (viii) [B b [C c t ]] 15. In most cases, the standard LCA also delivers symmetry (and transitivity) violations. If a in (35) contains more than one terminal, say a and b, we get a < b and b < a. For any terminal c that follows a in A or precedes a in B, we get a < c and c < a. But if a is itself a terminal, rightmost in A and leftmost in B, only the irreflexivity requirement is violated (a < a). 16. TP1 shares OB* with &0 , TP2, T0 2 and VP2 in (41). TP1 is the only category which both shares OB* and asymmetrically c-commands anything. 17. In (46), the terminals the and book are not included in d(VP1) since VP1 shares OB* (with &0 , VP2 and VP3); but they are included in the image of &P and categories containing &P, since OB* is fully dominated by those categories. 18. (52a) is taken from Hohle (1991: 146); (52b,c) and (53) from Wiese (1993). ¨ 19. Two further considerations support the view that status as a syntactic atom rather than a phonological word is decisive in determining ‘‘sub-word deletion’’ possibilities. T. A. Hall (p.c.) points out that the German diminutive suffix -chen, shared in e.g. Bruderund Schwester-chen (‘‘brother- and sister-DIM ’’), ¨ begins with a palatal fricative, and is a schwa monosyllable, both properties conflicting with requirements for being a phonological word in German. The diminutive is however a plausible candidate for a syntactic suffix. Secondly, not any subpart of a lexical item that corresponds to a phonological word is a legitimate target; part of a single stem cannot be shared. German Ameise (‘‘ant’’) is parsed as two phonological words (a) þ (meise). The second is homophonous with Meise (‘‘titmouse’’), which also figures in the compound Blaumeise (‘‘blue tit’’). Sharing of meise from Ameise is not possible: *Blau- und Ameise / * A- und Blaumeise. A similar case concerns the chester and caster of English placenames, which cannot be shared, cf. *Man- or Winchester, *Lanor Doncaster vs. court- or town jester, rope- or shoe-maker, etc. 20. The coordination of a clause involving topicalization of an adjunct with a clause involving topicalization of a complement is not itself the cause of the problem in (60). This is possible in the absence of sharing, cf. (ix): (ix) In London, Peter spoke, and about soccer, Mary spoke. 21. The dominance trees for (59) involve shared heads, as in forward Gapping, which in these cases are linearizable inside the conjuncts (as in the German example (45) above involving multiple shared verbs). Thus in (59a), the shared verb speak (transitive) forms two VPs, one with each PP; and the shared to
Notes to pages 253–256
283
forms two infinitives, one with each VP. Each infinitive is contained in a different finite conjunct, and within each conjunct, the PP fronts, giving (x). (x) [&P [S1 [PP1 about tennis], we persuaded Peter [ to [VP1 speak tPP1 ] ] ] [&0 and [S2 [PP2 about soccer], we persuaded Mary [ to [VP2 speak tPP2] ] ] ] ] The relative ordering of the shared terminals to and speak is the same in both conjuncts, and neither c-commands any non-shared overt terminal of the initial conjunct. Hence Linearization is possible, in the RNR pattern. 22. Examples like (71) and (72) are noted by Abbott (1976) and (73) by Perlmutter and Ross (1970); see also Moltmann (1992). The Spanish example (74) is cited from Camacho (1996). 23. Recall from (24) (section 10.3.2.1) that ‘‘b fully dominates a iff b is a member of every dominance path of a.’’ In (78), D is a member of both of a’s dominance paths, and ), but B is not; hence a is fully dominated by D but not by B. 24. Calling the constituents identified by (25) ‘‘shared’’ may seem counterintuitive with respect to this case, where a counts as a ‘‘shared constituent’’ of only one of its mothers. However, the notion of distinct mothers (and distinct sisters) ‘‘sharing’’ a applies to (78) just as to the ordinary sharing configuration (16); the key difference concerns the dominance relation between a’s sisters/mothers. The definition (25) correctly distinguishes (16) and (78) from ‘‘non-sharing’’ configurations like (xi), where a has only one mother but is dominated by more than one category. Since both X and Y fully dominate a, a is not a shared constituent in (xi). (xi) [X . . . [Y . . . a . . . ] . . . ] 25. The multiple dominance approach to movement is favored, independently of the Linearization issue, in that it permits the copying operation to be eliminated. Once the Single Mother Condition is dropped (e.g. to permit multiple dominance for constituent sharing in coordination), a treatment of movement as multiple dominance becomes desirable for another reason. To maintain the copy-theory alongside multiple dominance, an extra indexing is necessary, specifically in order to distinguish ‘‘copies’’ of a (two distinct a’s, each in one position) from sharing of a (one a in two positions). This indexing would be needed in addition to Chomsky’s (1995, 2000) ‘‘numeration index,’’ which serves to distinguish ‘‘movement copies’’ (sharing a numeration index), from chance lexical repetitions (having distinct numeration indices). 26. Strictly speaking, there are three configurations to consider, (xii)–(xiv), where A and B are sisters immediately dominated by C, and A c-commands (into) B. In (xii), a is shared by both A and B. In (xi), a is shared by A, and is at the same time A’s sister. In (xiv), a is shared by B, and is at the same time B’s sister. (xii) [C [A . . . a . . . ] [B . . . a . . . ] ] RNR (xiii) [C [A . . . a . . . ] a ] B ¼ a (string-vacuous RNR) (xiv) [C a [B . . . a . . . ] ] A ¼ a (Movement)
284
Notes to page 257
(xii) is the usual RNR-configuration; A c-commands a as well as dominating a. In (xiii) also, A both dominates and c-commands a; if the configuration exists, it will behave like RNR (string-vacuous in C, since a must be final in A or the structure does not linearize). In (xiv), the case of interest here, a is both dominated by and c-commands into B. 27. Recall that full dominance is (xii) reflexive and (xiii) excludes c-command. DP in (79) fully dominates itself and its daughter D, so DP c-commands neither itself nor D, even though both are dominated by T0 , a sister to DP. Thus the set A includes neither nor . In general, the modified LCA does not yield reflexivity or symmetry violations involving pairs of terminals belonging to a moved category.
References
Abbott, Barbara. 1976. Right node raising as a test for constituency. Linguistic Inquiry 7: 639–642. Abney, Steven. 1987. The English noun phrase in its sentential aspect. Doctoral Dissertation, Massachusetts Institute of Technology. Ackema, Peter, and Kriszta Szendro†i. 2002. Determiner sharing as an instance of dependent ellipsis. The Journal of Comparative Germanic Linguistics 5: 3–34. Albert, Chris. 1993. Sluicing and weak islands. Unpublished manuscript, UC, Santa Cruz. Asher, Nicholas. 1993. Reference to abstract objects in English. Kluwer Academic Press. Asher, Nicholas, Daniel Hardt, and Joan Busquets. 1997. Discourse parallelism, scope, and ellipsis. In Proceedings of the Seventh Conference on Semantics and Linguistic Theory, ed. Aaron Lawson, 19–36. The Chicago Linguistics Society, University of Chicago, IL. 2001. Discourse parallelism, ellipsis, and ambiguity. Journal of Semantics 18, 1–25. Bach, Emmon. 1979. Control in Montague Grammar. Linguistic Inquiry 10: 515–531. 1980. In defense of passive. Linguistics and Philosophy 3: 297–341. Bach, Emmon, and Robin Cooper. 1978. The NP-S analysis of relative clauses and compositional semantics. Linguistics and Philosophy 2: 145–149. Bach, Emmon, and Barbara H. Partee. 1980. Anaphora and semantic structure. In Papers from the Parasession on Pronouns and Anaphora, eds. Jody Kreiman and Almerindo E. Ojeda, 1–28. Chicago, IL: University of Chicago. Baker, C. L., and Michael Brame. 1972. ‘‘Global rules’’: a rejoinder. Language 48: 51–75. Baltin, Mark. 1987. Do antecedent-contained deletions exist? Linguistic Inquiry 18: 579–596. 2003. The interaction of ellipsis and binding: implications for the sequencing of Principle A. Natural Language and Linguistic Theory 21: 215–246. den Besten, Hans, and Gert Webelhuth. 1990. Stranding. In Scrambling and barriers, eds. Gunther Grewendorf and Wolfgang Sternefeld, 77–92. ¨ Amsterdam: John Benjamins. Bhatt, Rajesh, and Roumyana Pancheva. 2004. Late merger of degree clauses. Linguistic Inquiry 35: 1–46. Booij, Geert. 1985. Coordination reduction in complex words: a case for prosodic phonology. In Advances in nonlinear phonology, eds. Harry van der Hulst and N. Smith, 143–160. Dordrecht, The Netherlands: Foris. 285
286
References
Bouton, L. F. 1970. Antecedent contained pro-forms. In Chicago Linguistics Society, 154–167. The Chicago Linguistics Society: University of Chicago, IL. Brame, Michael. 1983. Ungrammatical notes 4: smarter than me. Linguistic Analysis 12: 323–328. Bresnan, Joan. 1973. Syntax of the comparative clause construction in English. Linguistic Inquiry 4: 275–343. 1975. Comparative deletion and constraints on transformations. Linguistic Analysis 1: 25–74. Bruening, Benjamin. 2001. QR obeys superiority: frozen scope and ACD. Linguistic Inquiry 32: 233–273. Buring, Daniel. 2005. The syntax and semantics of binding theory. Cambridge: ¨ Cambridge University Press. Burton, Strang, and Jane Grimshaw. 1992. Coordination and VP-internal subjects. Linguistic Inquiry 23: 305–312. Camacho, Jose.´ 1996. The structure of NP coordination. Doctoral Dissertation, University of Southern California. Carlson, Gregory N. 1977. Reference to kinds in English. Doctoral Dissertation, University of Massachusetts, Amherst. Chao, Wynn. 1987. On ellipsis. Doctoral Dissertation, University of Massachusetts, Amherst. Chierchia, Gennaro. 1995. Dynamics of meaning. Chicago: University of Chicago Press. Chomsky, Noam. 1972a. Some empirical issues in the theory of transformational grammar. In The goals of linguistic theory, ed. Stanley Peters, 63–130. Englewood Cliffs, NJ: Prentice-Hall. 1972b. Spheres of influence in the age of imperialism; papers submitted to the Bertrand Russell centenary symposium, Linz, Austria, September 11th to 15th, 1972. Nottingham: Bertrand Russell Peace Foundation for The Spokesman. 1981. Lectures on government and binding. Dordrecht, The Netherlands: Foris. 1982. Some concepts and consequences of the theory of government and binding. Cambridge, MA: MIT Press. 1986a. Barriers. Cambridge, MA: MIT Press. 1986b. Knowledge of language. New York: Praeger. 1993. A minimalist program for linguistic theory. In The view from building, volume XX, eds. Ken Hale and Jay Keyser, 1–52. Cambridge, MA: MIT Press. 1995. The minimalist program. Cambridge, MA: MIT Press. 2000. Minimalist inquiries: the framework. In Step by step: essays on minimalist syntax in honor of Howard Lasnik, eds. Roger Martin, David Michaels, and Juan Uriagereka, 89–156. Cambridge, MA: MIT Press. Chomsky, Noam, and Howard Lasnik. 1993. Principles and parameter theory. In Syntax: an international handbook of contemporary research, eds. Arnim von Stechow, W. Sternefeld, and T. Vennemann, 506–569. Berlin: de Gruyter. Chung, Sandra, William A. Ladusaw, and James McCloskey. 1995. Sluicing and Logical Form. Natural Language Semantics 3: 239–282. Collins, Chris. 1994. Economy of derivation and the generalized proper binding condition. Linguistic Inquiry 25: 45–62.
References
287
Cooper, Robin. 1979. The interpretation of pronouns. In Syntax and Semantics, volume X, Selections from the Third Groningen Round Table, eds. Frank Heny and H. Schnelle, 61–92. New York: Academic Press. Coppock, Elizabeth. 2001. Gapping: in defense of deletion. In Chicago Linguistics Society, eds. Mary Andronis, Christopher Ball, Heidi Elston, and Sylvain Neuvel, volume XXXVII, 133–148. University of Chicago. Cormack, A. 1984. VP anaphora: variables and scope. In Varieties of formal semantics, eds. F. Landman and F. Veltman, 81–102. Dordrecht: Foris. Corver, Norbert. 1990. The syntax of left branch extractions. Doctoral Dissertation, Katholieke Universiteit Brabant. Culicover, Peter W., and Michael S. Rochemont. 1990. Extraposition and the complement principle. Linguistic Inquiry 21: 23–47. ¨ Dahl, Osten. 1974. How to open a sentence: abstraction in natural language. In Logical grammar reports, volume XII. Dalrymple, Mary. 1991. Against reconstruction in ellipsis. Unpublished manuscript, Xerox-PARC and CSLI. Dalrymple, Mary, Stuart M. Sheiber, and Fernando C. N. Pereira. 1991. Ellipsis and higher-order unification. Linguistics and Philosophy 14: 399–452. Diesing, Molly. 1992a. Bare plural subjects and the derivation of logical representations. Linguistic Inquiry 23: 353–380. 1992b. Indefinites. Cambridge, MA: MIT Press. Dimitriadis, Alexis. 2001. Function domains in variable free semantics. In Proceedings of SALT 11, eds. Rachel Hastings, Brendan Jackson, and Zsofia Zvolenszky, 134–151. Cornell University, Ithaca, NY: CLC Publications. Doron, Edit. 1999. V-movement and VP ellipsis. In Studies in ellipsis and gapping, eds. Shalom Lappin and Elabbas Benmamoun, 124–140. New York: Oxford University Press. Dougherty, Ray C. 1970. A grammar of coordinate conjoined structures, I. Language 46: 850–898. Dowty, David. 1982. Grammatical relations and Montague grammar. In The nature of syntactic representation, eds. Pauline Jacobson and Geoffrey K. Pullum, 79–130. D. Reidel Publishing Company. 1988. ‘‘Type raising, functional composition, and non-constituent conjunction,’’ in Categorial Grammars and natural language structures, eds. R. Oerhle, E. Bach, and D. Wheeler. Dordrecht: Reidel. Drubig, Hans Bernhard. 1994. Island constraints and the syntactic nature of focus and association with focus. Arbeitspapiere des Sonderforschungsbereichs 340 (working paper). University of Tubingen. ¨ Eisenberg, Peter. 1973. A note on ‘‘Identity of constituents.’’ Linguistic Inquiry 4: 417–420. Elbourne, Paul. 2001. E-type anaphora as NP deletion. Natural Language Semantics 9: 241–288. Engdahl, Elisabet. 1986. Constituent questions. Dordrecht, The Netherlands: D. Reidel Publishing Company. Epstein, Samuel. 1995. A derivational approach to syntactic relations. Unpublished manuscript, Harvard University.
288
References
Evans, Frederic. 1988. Binding into anaphoric verb phrases. In Proceedings of the 5th Annual Eastern States Conference on Linguistics, 122–129. ESCOL Publications Committee: Linguistics Department, The Ohio State University. Evans, Gareth. 1977. Pronouns, quantifiers, and relative clauses. Canadian Journal of Philosophy 7: 467–536. 1980. Pronouns. Linguistic Inquiry 11: 337–362. Farkas, Donka. 1981. Quantifier scope and syntactic islands. In Papers from the 17th regional meeting of the Chicago Linguistic Society, eds. Randall Hendrick, Carrie Masek, and Mary Frances Miller, 59–66. Chicago Linguistic Society, University of Chicago, IL. Farkas, Donka, and Anastasia Giannakidou. 1995. How clause-bounded is the scope of universals? In Proceedings of SALT 6, ed. T. Gallway, 35–52. Ithaca, NY: CLC Publications. Fiengo, Robert. 1980. Surface structure. Cambridge, MA: Harvard University Press. Fiengo, Robert, and Robert May. 1994. Indices and identity. Cambridge, MA: MIT Press. Fintel, Kai von, and Irene Heim. 2002. Notes on intensional semantics. URL www.phil-fak.uni-duesseldorf.de/summerschool2002/fintel.pdf, unpublished manuscript, MIT, Cambridge. Fox, Danny. 1998. Locality in variable binding. In Is the best good enough?, eds. Pilar Barbosa, Danny Fox, Paul Hagstrom, Martha McGinnis, and David Pesetsky, 129–156. Cambridge, MA: MIT Press. 1999a. Focus, parallelism, and accommodation. In Proceedings of the Ninth Conference on Semantics and Linguistic Theory, eds. T. Matthews and Devon Strolovitch, 70–90. Cornell University, Ithaca, NY: CLC Publications. 1999b. Reconstruction, binding theory, and the interpretation of chains. Linguistic Inquiry 30: 157–196. 2000. Economy and semantic interpretation. Cambridge, MA: MIT Press. 2002. Antecedent-contained deletion and the copy theory of movement. Linguistic Inquiry 33: 63–96. Fox, Danny, and Howard Lasnik. 2003. Successive-cyclic movement and island repair: the difference between Sluicing and VP-ellipsis. Linguistic Inquiry 34: 143–154. Fox, Danny, and Jon Nissenbaum. 1999. Extraposition and scope: a case for overt QR. In Proceedings of the West Coast Conference on Formal Linguistics 18, eds. Sonya Bird, Andrew Carnie, Jason D. Haugen, and Peter Norquest, 132–144. Somerville, MA: Cascadilla Press. Frampton, John, and Sam Gutmann. 1999. Cyclic computation: A computationally efficient Minimalist syntax. Syntax 2: 1–27. Frege, Gottlob. 1879. Begriffsschrift. eine der arithmetischen nachgebildete Formelsprache des reinen Denkens. Halle, Germany: Neubert. Ga¨rtner, Hans-Martin. 1997. Generalized transformations and beyond. Doctoral Dissertation, University of Frankfurt/Main. Gawron, Jean Mark, and Stanley Peters. 1990. Anaphora and quantification in Situation Semantics. Stanford University: Center for the Study of Language and Information. Geach, P. 1972. A program for syntax. In Semantics of natural language, eds. Donald Davidson and Gilbert H. Harman, 483–497. Reidel Publishing Company.
References
289
Giannakidou, Anastasia. 1998. Polarity sensitivity as (non)veridicality. Amsterdam: John Benjamins. 2001. The meaning of free choice. Linguistics and Philosophy 24: 659–735. Goldberg, Lotus Madelyn. 2005. Verb-stranding VP ellipsis: a cross-linguistic study. Doctoral Dissertation, McGill University, Montreal. Goodall, Grant, 1987. Parallel structures in syntax. Cambridge: Cambridge University Press. Grimshaw, Jane. 1991. Extended projection. Unpublished manuscript, Brandeis University. Grinder, John, and Paul M. Postal. 1971. Missing antecedents. Linguistic Inquiry 2: 269–312. Grodzinsky, Yosef, and Tanya Reinhart. 1993. The innateness of binding and coreference. Linguistic Inquiry 24: 69–102. Groenendijk, J., and M. Stokhof. 1983. Interrogative quantifiers and skolem functions. In Connectedness in sentence, discourse and text, eds. K. Ehlich and Henk van Riemsdijk, 71–110. Tilburg University. Hackl, Martin. 2000. Comparative quantifiers. Doctoral Dissertation, Massachusetts Institute of Technology. Ha ¨ık, Isabelle. 1984. Indirect binding. Linguistic Inquiry 15: 185–224. 1985. The syntax of operators. Doctoral Dissertation, MIT. 1987. Bound VPs that need to be. Linguistics and Philosophy 10: 503–530. Hankamer, Jorge. 1973a. Unacceptable ambiguity. Linguistic Inquiry 4: 17–68. 1973b. Why there are two than’s in English. In Proceedings of the Chicago Linguistic Society 9, 179–191. The Chicago Linguistics Society, University of Chicago, IL. 1978. On the nontransformational derivation of some null VP anaphors. Linguistic Inquiry 9: 66–74. 1979. Deletion in coordinate structures. New York: Garland Publishing, Inc. Hankamer, Jorge, and Ivan Sag. 1976. Deep and surface anaphora. Linguistic Inquiry 7: 391–428. 1977. Syntactically vs. pragmatically controlled anaphora. In Studies in language variation, eds. R. Fasold and R. Shuy, 121–135. Washington, DC: Georgetown University Press. Hardt, Daniel. 1992a. VP ellipsis and semantic identity. In Proceedings of SALT 2, 145–162. Cornell University, Ithaca, NY: CLC Publications. 1992b. VP ellipsis and semantic identity. In Proceedings of the Stuttgart Ellipsis Workshop, eds. Steve Berman and Arild Hestvik. Stuttgart. 1993. Verb phrase ellipsis: Form, meaning and processing. Doctoral Dissertation, University of Pennsylvania. 1999. Dynamic interpretation of verb phrase ellipsis. Linguistics and Philosophy 22: 185–219. Harris, Zellig. 1957. Co-occurrence and transformation in linguistic structure. Language 33: 283–340. Hartmann, Katharina. 1998. Right node raising and gapping. Doctoral Dissertation, Goethe-Universita¨t zu Frankfurt am Main, Frankfurt. Heim, Irene. 1982. The semantics of definite and indefinite Noun Phrases. Doctoral Dissertation, University of Massachusetts, Amherst.
290
References
1983. On the projection problem for presuppositions. In Proceedings of WCCFL 2, ed. Daniel P. Flickinger, 114–125. Stanford University, Stanford, CA: CSLI Publications. 1985. Notes on comparatives and related matters. Unpublished manuscript, University of Texas-Austin. 1990. E-type pronouns and donkey anaphora. Linguistics and Philosophy 13: 137–178. 1993. Anaphora and semantic interpretation: a reinterpretation of Reinhart’s approach. SfS-Report 7–93, Universita¨t Tubingen, Germany. ¨ 1994. Puzzling reflexive pronouns in de se reports. Unpublished Handout, March, MIT, Cambridge, MA. 1997. Predicates or formulas? evidence from ellipsis. In Proceedings of SALT VII, eds. Aaron Lawson and Enn Cho, 197–221. Cornell University: CLC Publications. 2000. Degree operators and scope. In Proceedings of SALT X, eds. Brendan Jackson and Tanya Matthews, 40–64. Cornell University: CLC Publications. 2001. Degree operators and scope. In Audiatur vox sapientiae. A festschrift for Arnim von Stechow, eds. Caroline Fery and Wolfgang Sternefeld, 214–239. Berlin: Akademie Verlag. Heim, Irene, and Angelika Kratzer. 1998. Semantics in generative grammar. Malden, MA: Blackwell. Hellan, Lars. 1981. Towards an integrated theory of comparatives. Tubingen: ¨ Gunter Narr. Heycock, Caroline. 1995. Asymmetries in reconstruction. Linguistic Inquiry 26: 547–570. Heycock, Caroline, and Anthony Kroch. 1994. Verb movement and coordination in a dynamic theory of licensing. The Linguistic Review 11: 257–284. Higginbotham, James. 1983. Logical form, binding, and nominals. Linguistic Inquiry 14: 395–420. 1985. On semantics. Linguistic Inquiry 16: 547–594. Hirschberg, Julia, and Gregory Ward. 1991. Accent and bound anaphora. Cognitive Linguistics 2: 101–121. Hirschbuhler, Paul. 1978. The syntax and semantics of wh-constructions. Doctoral ¨ Dissertation, University of Massachusetts at Amherst. 1982. VP deletion and across-the-board quantifier scope. In NELS, eds. James Pustejovsky and Peter Sells, 132–139. University of Massachusetts, Amherst: GLSA. Hobbs, Jerry, and Andrew Kehler. 1997. A theory of parallelism and the case of VP Ellipsis. In Proceedings of the ACL 35, 394–401. Madrid, Spain. Hoeksema, Jack. 1984. To be continued: the story of the comparative. Journal of Semantics 3: 93–107. Hohle, Tilman N. 1990. Assumption about asymmetric coordination in German. ¨ In Grammar in progress: GLOW essays for Henk van Riemsdijk, eds. Joan Mascaro´ and Marina Nespor, 221–235. Dordrecht, The Netherlands: Foris. 1991. On reconstruction and coordination. In Representation and derivation in the theory of grammar, eds. Hubert Haider and Klaus Netter,
References
291
volume XXII, 139–198. Dordrecht, The Netherlands: Kluwer Academic Publishers. Hoji, Hajime, and Teruhiko Fukaya. 1999. Stripping and sluicing in Japanese and some implications. In Proceedings of the 18th West Coast Conference on Formal Linguistics, eds. S. Bird, A. Carnie, J. Haugen, and P. Norquest, 145–158. Somerville, MA: Cascadilla Press. Hornstein, Norbert. 1995. Logical Form: from GB to Minimalism. Cambridge, MA: Blackwell. Huang, C.-T. James. 1993. Reconstruction and the structure of VP: some theoretical consequences. Linguistic Inquiry 24: 103–138. Hudson, Richard A. 1976. Conjunction reduction, gapping and right-node raising. Language 52: 535–562. Jacobson, Pauline. 1977. The syntax of crossing coreference sentences. Doctoral Dissertation, University of California, Berkeley. 1982. On the syntax and semantics of multiple relatives in English. Bloomington, IN: Indiana University Linguistics Club. 1987. Phrase structure, grammatical relations, and discontinuous constituents. In Syntax and semantics, volume XX: Discontinuous constituency, eds. Geoffrey Huck and Almerindo E. Ojeda, 27–69. New York Academic Press, Inc. 1992a. Antecedent contained deletion in a variable-free semantics. In SALT 2, eds. Chris Barker and David Dowty, 193–213. Linguistics Department, Ohio State University. 1992b. Flexible categorical grammars: questions and prospects. In Formal grammar: Theory and implementation, ed. Robert D. Levine, 129–167. Oxford: Oxford University Press. 1994. Binding Connectivity in Copular Sentences. In Proceedings of the Fourth Conference on Semantics and Linguistic Theory, eds. M. Harvey and L. Santelmann. Cornell University, Ithaca, NY: CLC Publications. 1998. ACE and pied-piping: evidence for a variable-free semantics. In Proceedings of SALT 8, eds. Devon Strolovitch and Aaron Lawson, 74–91. Cornell University, Ithaca, NY: CLC Publications. 1999. Towards a variable-free semantics. Linguistics and Philosophy 22: 117–184. 2000a. Paycheck pronouns, Bach–Peters sentences, and variable-free semantics. Natural Language Semantics 8: 77–155. 2000b. Paychecks, stress, and variable free semantics. In Proceedings of SALT X, eds. T. Matthews and B. Jackson, 65–82. Cornell University, Ithaca, NY: CLC Publications. 2002. The (dis)organization of the grammar: 25 years. Linguistics and Philosophy 25: 601–626. 2003. Binding without pronouns (and pronouns without binding). In Binding and resource sensitivity, eds. G.-J. Kruiff and Richard Oehrle, 57–96. Dordrecht: Kluwer Academic Publishers. 2004. Kennedy’s puzzle: What I’m named or who I am? In Proceedings of SALT 14, ed. Robert B. Young, 145–162. Ithaca, NY: CLC Publications, Cornell University. Jayaseelan, K. A. 1990. Incomplete VP deletion and Gapping. Linguistic Analysis 20: 64–81.
292
References
Johnson, Kyle. 2001. What VP ellipsis can do, and what it can’t, but not why. In The handbook of contemporary syntactic theory, eds. Mark Baltin and Chris Collins, 439–479. Oxford: Blackwell. 2003. In search of the English middle field. Unpublished manuscript, University of Massachusetts, Amherst. 2004. How to be quiet. In Proceedings of the Chicago Linguistic Society 39. Kadmon, Nirit. 1987. On unique and non-unique reference and asymmetric quantification. Doctoral Dissertation, University of Massachusetts, Amherst. Kayne, Richard S. 1994. The antisymmetry of syntax. Cambridge, MA: MIT Press. 1998. Overt vs. covert movement. Syntax 1: 128–191. Kehler, Andrew, and Stuart Shieber. 1997. Anaphoric dependencies in ellipsis. Computational Linguistics 23: 457–466. Kennedy, Christopher. 1994. Argument contained ellipsis. Linguistics Research Center Report LRC-94-03, University of California, Santa Cruz. 1995. An indexical account of certain ambiguities. In Proceedings of ESCOL 95, 103–115. Ohio State University: ESCOL Publications Committee, Linguistics Department. 1997a. Antecedent contained deletion and the syntax of quantification. Linguistic Inquiry 28: 662–688. 1997b. Projecting the adjective: the syntax and semantics of gradability and comparison. Doctoral Dissertation, University of California, Santa Cruz. Kennedy, Chris, and Jason Merchant. 2000. Attributive comparative deletion. Natural Language and Linguistic Theory 18: 89–146. Kennedy, Christopher, and Jeffrey Lidz. 2001. A (covert) long distance anaphor in English. In Proceedings of the 20th West Coast Conference in Formal Linguistics, eds. Karine Megerdoomian and Leora Anne Bar-el, 318–331. Somerville, MA: Cascadilla Press. Kitagawa, Yoshihisa. 1986. Subjects in Japanese and English. Doctoral Dissertation, University of Massachusetts, Amherst. 1991. Copying identity. Natural Language and Linguistic Theory 9: 497–536. Kitahara, Hisatsugu. 1999. Eliminating * as a feature (of traces). In Working minimalism, eds. Samuel Epstein and Norbert Hornstein, 77–93. Cambridge, MA: MIT Press. Klein, E. 1986. VP ellipsis in DR theory. In Studies in DRT and the theory of generalized quantifiers, eds. J. Groenendijk, de Jongh, and Stokhof, 161–187. Dordrecht: Foris. Koopman, Hilda Judith, and Dominique Sportiche. 1991. The position of subjects. Lingua 85: 211–258. Koopman, Hilda Judith, and Anna Szabolcsi. 2000. Verbal complexes. Cambridge, MA: MIT Press. Koster, Jan. 1978. Locality principles in syntax. Dordrecht, The Netherlands: Foris. Kratzer, Angelika. 1991. The representation of focus. In Semantics: an international handbook of contemporary research, eds. Arnim von Stechow and Dieter Wunderlich, 825–834. Berlin: Walter de Gruyter. 1998. More structural analogies between pronouns and tense. In Proceedings of SALT VIII, eds. Devon Strolovitch and Aaron Lawson, 92–110. Cornell University, Ithaca, NY: CLC Publications.
References
293
Krifka, Manfred. 1999. Additive particles under stress. In Proceedings of SALT 8, eds. Devon Strolovitch and Aaron Lawson, 111–128. Cornell University, Ithaca, NY: CLC Publications. Kuroda, S.-Y. 1988. Whether we agree or not. Linguisticae Investigationnes 12: 1–47. Lakoff, George. 1971. On generative semantics. In Semantics, eds. D. Steinberg and L. Jakobovits, 232–296. Cambridge: Cambridge University Press. 1972. The arbitrary basis of transformational grammar. Language 48: 76–87. Lappin, Shalom. 1996. The interpretation of ellipsis. In The handbook of contemporary semantic theory, ed. Shalom Lappin, 145–175. Oxford: Blackwell. Larson, Richard K. 1985. On the syntax of disjunction scope. Natural Language and Linguistic Theory 3: 217–264. Larson, Richard K., and Robert May. 1990. Antecedent containment or vacuous movement: reply to Baltin. Linguistic Inquiry 21: 103–122. Lasnik, Howard. 1976. Remarks on coreference. Linguistic Analysis 2: 1–22. 1995. Verbal morphology: Syntactic structures meets the minimalist program. In Evolution and revolution in linguistic theory, eds. Hector Campos and Paula Kempchinsky, 251–275. Georgetown: Georgetown University Press. 1999a. A note on pseudogapping. In Minimalist analysis, ed. Howard Lasnik, 151–174. Oxford: Blackwell. 1999b. On feature strength: three minimalist approaches to overt movement. Linguistic Inquiry 30: 197–217. 1999c. Pseudogapping puzzles. In Fragments: Studies in ellipsis and gapping, eds. Shlom Lappin and Elabbas Benmamoun, 141–174. Oxford: Oxford University Press. 2001. When can you save a structure by destroying it? In Proceedings of the North East Linguistic Society 31, eds. Minjoo Kim and Uri Strauss, 301–320. Georgetown University: GLSA. Lasnik, Howard, and Mamoru Saito. 1984. On the nature of proper government. Linguistic Inquiry 15: 235–290. 1992. Move . Cambridge, MA: MIT Press. Lebeaux, David. 1988. Language acquisition and the form of the grammar. Doctoral Dissertation, University of Massachusetts, Amherst. 1990. Relative clauses, licensing, and the nature of the derivation. In Proceedings of the North East Linguistic Society 20, eds. Juli Carter, Dechaine Rose-Marie, Bill Philip, and Tim Sherer, 318–332. University of Massachusetts at Amherst: Graduate Linguistic Student Association. Lechner, Winfried. 1998. Comparative deletion. Doctoral Dissertation, University of Massachusetts, Amherst. 1999. Comparatives and DP-Structure. Doctoral Dissertation, University of Massachusetts, Amherst. 2000. Conjunction reduction in subordinate structures. In Proceedings of the North East Linguistic Society 30, eds. Masako Hirotani, Andries Coetzee, Nancy Hall, and Ji-yung Kim, 455–468. Rutgers University: Graduate Linguistic Student Association. 2001. Reduced and phrasal comparatives. Natural Language and Linguistic Theory 19: 683–735. 2004. Ellipsis in Comparatives. Berlin/New York: Mouton de Gruyter.
294
References
Levin, Beth, and Malka Rappaport. 1986. The formation of adjectival passives. Linguistic Inquiry 17: 623–661. Levin, Lori. 1982. Sluicing: A lexical interpretation procedure. In The mental representation of grammatical relations, ed. Joan Bresnan, 590–654. Cambridge, MA: MIT Press. Levin, Nancy. 1986. Main-verb ellipsis in spoken English. New York: Garland Publishing, Inc. Levine, Robert D. 1985. Right Node (non)-Raising. Linguistic Inquiry 16: 492–497. Lin, Jo-Wang. 1996. Polarity licensing and wh-phrase quantification in Chinese. Doctoral Dissertation, University of Massachusetts, Amherst. Lin, Vivian. 2002. Coordination and sharing at the interfaces. Doctoral Dissertation, Linguistics Department, MIT. Lobeck, Anne. 1991. The phrase structure of ellipsis. In Perspectives on phrase structure, ed. Susan Rothstein, 81–103. San Diego: Academic Press. 1995. Ellipsis: Functional heads, licensing and identification. New York: Oxford University Press. Lo´pez, Luis. 1994. The syntactic licensing of VP-ellipsis: a comparative study of Spanish and English. In Issues and theory in Romance linguistics: selected papers from the linguistic symposium on Romance languages XXIII, ed. Michael L. Mazzola, 333–354. Washington, DC: Georgetown University Press. 1999. VP-ellipsis in Spanish and English and the features of Aux. Probus 11: 263–297. Lo´pez, Luis, and Susanne Winkler. 2003. Variation at the syntax-semantics interface: evidence from gapping. In The interfaces: Deriving and interpreting omitted structures, eds. Kerstin Schwabe and Susanne Winkler, 227–248. John Benjamins. May, Robert. 1977. The grammar of quantification. Doctoral Dissertation, Massachusetts Institute of Technology. 1985. Logical form: its structure and derivation. Cambridge, MA: MIT Press. McCawley, James D. 1970. Where do noun phrases come from? In Readings in English transformational grammar, eds. R. A. Jacobs and Peter S. Rosenbaum, 166–183. Waltham, MA: Ginn and Company. 1982. Parentheticals and discontinuous constituent structure. Linguistic Inquiry 13: 91–106. 1998. The syntactic phenomena of English. Chicago: University of Chicago. McCloskey, James. 1991. Clause structure, ellipsis and proper government in Irish. Lingua 85: 259–302. Merchant, Jason. 1999. The syntax of silence: sluicing, islands, and identity in ellipsis. Doctoral Dissertation, University of California, Santa Cruz. 2000. Islands and LF-movement in Greek sluicing. Journal of Greek Linguistics 1: 39–62. 2001. The syntax of silence: sluicing, islands, and the theory of ellipsis. Oxford: Oxford University Press. 2002. Swiping in German. In Studies in comparative Germanic syntax, eds. C. Jan-Wouter Zwart and Werner Abraham, 295–321. Amsterdam: John Benjamins.
References
295
2005. On the role of unpronounced syntactic structures. Talk given at the University of Illinois, Urbana-Champaign. Moltmann, Friederike. 1992. Coordination and comparatives. Doctoral Dissertation, MIT. Montague, Richard. 1974. The proper treatment of quantification in ordinary English. In Formal philosophy: Selected papers of Richard Montague, ed. Richmond H. Thomason, 247–270. New Haven: Yale University Press. Muadz, H. 1991. Coordinate structures: a planar representation. Doctoral Dissertation, University of Arizona, Tucson. Muller, Gereon. 1996. A constraint on remnant movement. Natural Language and ¨ Linguistic Theory 14: 355–407. Muller, Gereon, and Wolfgang Sternefeld. 1993. Improper movement and unam¨ biguous binding. Linguistic Inquiry 24: 461–508. Muskens, Reinhard. 1996. Combining Montague semantics and discourse representation. Linguistics and Philosophy 19: 143–186. Napoli, Donna Jo. 1983. Comparative ellipsis: a phrase structure analysis. Linguistic Inquiry 14: 675–694. Nishigauchi, Taisuke. 1998. ‘‘Multiple sluicing’’ in Japanese and the functional nature of wh-phrases. Journal of East Asian Linguistics 7: 121–152. Nunes, Jairo. 1995. The copy theory of movement and linearization of chains in the Minimalist Program. Doctoral Dissertation, University of Maryland. Oehrle, Richard T. 1991. Categorial frameworks, coordination, and extraction. In The Proceedings of the Ninth West Coast Conference on Formal Linguistics, ed. Aaron L. Halpern, 411–426. Stanford, CA: Center for the Study of Language and Information. Partee, Barbara Hall. 1975. Montague grammar and transformational grammar. Linguistic Inquiry 6: 203–300. 1976. Some transformational extensions of Montague grammar. In Montague grammar, ed. Barbara Hall Partee, 51–76. Academic Press, Inc. Partee, Barbara Hall, and Mats Rooth. 1983. Generalized conjunction and type ambiguity. In Meaning, use and interpretation of language, eds. R. Ba¨uerle, C. Schwarze, and Arnim von Stechow, 362–383. De Gruyter. Percus, Orin. 2000. Constraints on some other variables in syntax. Natural Language Semantics 8: 173–229. Perlmutter, David. 1971. Deep and surface structure constraints in syntax. New York: Holt, Rinehart and Winston. Perlmutter, David, and John Robert Ross. 1970. Relative clauses with split antecedent. Linguistic Inquiry 1: 350. Pesetsky, David. 1987. Wh-in-situ: Movement and unselective binding. In The representation of (in)definiteness, eds. Eric J. Reuland and Alice G. B. ter Meulen, 98–129. Cambridge, MA: MIT Press. 2000. Phrasal movement and its kin. Cambridge: MIT Press. Phillips, Colin. 1996. Order and structure. Doctoral Dissertation, Massachusetts Institute of Technology, Cambridge, MA. Pinkham, Jessie E. 1982. The formation of comparative clauses in English and French. Indiana University Linguistics Club.
296
References
Pollard, Carl. 1984. Generalized phrase structure grammar. Doctoral Dissertation, Stanford University. Pollard, Carl Jesse, and Ivan A. Sag. 1994. Head-driven phrase structure grammar. Stanford: Center for the Study of Language and Information. Postal, Paul M. 1969a. Anaphoric islands. In Papers from the Fifth Regional Meeting Chicago Linguistics Society, eds. Robert I. Binnick, Alice Davison, Georgia M. Green, and Jerry L. Morgan, 205–239. Chicago, IL: University of Chicago. 1969b. On so-called ‘‘pronouns’’ in English. In Modern studies in English, eds. D. Reibel and Sandford Schane, 201–224. Englewood Cliffs, New Jersey: Prentice-Hall. 1971. Cross-over phenomena. New York: Holt, Rinehart and Winston. 1974. On raising. Cambridge, MA: MIT Press. 1993. Some defective paradigms. Linguistic Inquiry 24: 347–364. 1998. Three investigations of extraction. Cambridge: MIT Press. Potts, Christopher. 1999. Vehicle change and anti-pronominal contexts. University of California, Santa Cruz. Prust, ¨ Hub, Remko Scha, and Martin van den Berg. 1994. Discourse grammar and verb phrase anaphora. Linguistics and Philosophy 17: 261–327. Pullum, Geoffrey, and Arnold Zwicky. 1986. Phonological resolution of syntactic feature conflict. Language 62: 751–773. Reinhart, Tanya. 1976. The syntactic domain of anaphora. Doctoral Dissertation, Massachusetts Institute of Technology. 1983a. Anaphora and semantic interpretation. Chicago, IL: University of Chicago Press. 1983b. Coreference and bound anaphora: a restatement of the anaphora questions. Linguistics and Philosophy 6: 47–88. Reinhart, Tanya, and Eric Reuland. 1993. Reflexivity. Linguistic Inquiry 24: 657–720. Richards, Norvin. 1998. In full pursuit of the unspeakable. In Proceedings of the North East Linguistic Society, eds. Pius N. Tamanji and Kiyomi Kusumoto, 153–168. University of Toronto: Graduate Linguistic Student Association. Riemsdijk, Henk van. 1982. A case study in syntactic markedness. Dordrecht, The Netherlands: Foris. Rizzi, Luigi. 1990. Relativized minimality. Cambridge, MA: MIT Press. Roberts, Craige. 1991. Distributivity and reciprocal distributivity. In Proceedings of the First Semantics and Linguistic Theory Conference, eds. Steven Moore and Adam Zachary Wyner, 209–230. Cornell University, Ithaca, NY: Cornell University Working Papers in Linguistics. Romero, Maribel. 1998. Focus and reconstruction effects in wh-phrases. Doctoral Dissertation, University of Massachusetts, Amherst. 2000. Antecedentless sluiced wh phrases and islands. In Ellipsis in conjunction, eds. Kerstin Schwabe and Ning Zhang, 195–220. Tubingen: Niemeyer. ¨ Rooth, Mats E. 1981. A comparison of three theories of verb phrase ellipsis. In University of Massachusetts Occasional Papers in Linguistics, eds. Wynn Chao and Dierdre Wheeler, 212–244. GLSA. 1985. Association with focus. Doctoral Dissertation, University of Massachusetts, Amherst.
References
297
1992a. Ellipsis redundancy and reduction redundancy. In Proceedings of the Stuttgart Ellipsis Workshop, eds. Steve Berman and Arild Hestvik. Stuttgart. 1992b. A theory of focus interpretation. Natural Language Semantics 1: 117–121. 1996. Focus. In The handbook of contemporary semantic theory, ed. Shalom Lappin, 271–297. Oxford: Blackwell. Ross, John Robert. 1967. Constraints on variables in syntax. Doctoral Dissertation, Massachusetts Institute of Technology. 1969. Guess who? In Chicago Linguistics Society, eds. Robert I. Binnick, Alice Davison, Georgia M. Green, and Jerry L. Morgan, 252–286. Chicago, IL: The Chicago Linguistics Society, University of Chicago. 1970. Gapping and the order of constituents. In Progress in linguistics, eds. Manfred Bierwisch and Karl E. Heidolph, 249–259. The Hague: Mouton. Rullmann, Hotze. 1995. Maximality in the semantics of wh-constructions. Doctoral Dissertation, University of Massachusetts, Amherst. Russell, Bertrand. 1905. On denoting. Mind 14: 479–493. Safir, Ken. 1986. Relative clauses in a theory of binding and levels. Linguistic Inquiry 17: 663–689. Sag, Ivan. 1976. Deletion and logical form. Doctoral Dissertation, Massachusetts Institute of Technology, Cambridge, Massachusetts. 1980. Deletion and logical form. New York: Garland Publishing. Sag, Ivan, and Jorge Hankamer. 1984. Toward a theory of anaphoric processing. Linguistics and Philosophy 7: 325–345. Sauerland, Uli. 1996. Guess how? In Proceedings of ConSoLE, 297–311. Leiden, The Netherlands: SOLE, University of Leiden/Luch. 1998. The meaning of chains. Doctoral Dissertation, Massachusetts Institute of Technology, Cambridge, MA. 1999. Why variables? In North East Linguistics Society, eds. Tamanji Pius, Masako Hirotani, and Nancy Hall, 323–338. University of Delaware: GLSA. 2000. The content of pronouns: Evidence from focus. In Proceedings of SALT X, eds. Tanya Matthews and Brendan Jackson, 167–184. Cornell University, Ithaca, NY: CLC Publications. 2003. A new semantics for number. In Proceedings of SALT 13, 258–275. Cornell University, Ithaca, NY: CLC Publications. 2004. The interpretation of traces. Natural Language Semantics 12: 63–127. Schachter, Paul. 1977. Constraints on coordination. Language 53: 86–103. Schuyler, Tamara. 2001. Wh-movement out of the site of VP ellipsis. In Syntax at ´ Santa Cruz, ed. Seamas Mac Bhloscaidh, 1–20. UC, Santa Cruz: Linguistics Department. Schwarz, Bernhard. 1999. On the syntax of either . . . or. Natural Language and Linguistic Theory 17: 339–370. 2000. Topics in ellipsis. Doctoral Dissertation, University of Massachusetts, Amherst. Schwarzschild, Roger. 1999. Givenness, avoid F and other constraints on the placement of accent. Natural Language Semantics 7: 141–177. Schwarzschild, Roger, and Karina Wilkinson. 2002. Quantifiers in comparatives: a semantics of degree based on intervals. Natural Language Semantics 10: 1–41.
298
References
Sharvit, Yael. 1998. Possessive wh-expressions and reconstruction. In Proceedings of the North East Linguistic Society, eds. Pius N. Tamanji and Kiyomi Kusumoto, 409–424. University of Toronto: Graduate Linguistic Student Association. 1999. Functional relative clauses. Linguistics and Philosophy 22: 447–478. Simpson, Andrew. 2000. Wh-movement and the theory of feature-checking. Amsterdam: John Benjamins. Soames, Scott. 1989. Presupposition. In Handbook of philosophical logic, eds. Dov Gabbay and Franz Guenther, volume IV, 553–616. Dordrecht: Reidel. Sproat, Richard. 1988. On anaphoric islandhood. In Theoretical morphology, eds. Michael Hammond and Michael Noonan, 291–304. San Diego, CA: Academic Press, Inc. von Stechow, Arnim. 1984. Comparing semantic theories of comparison. Journal of Semantics 3: 1–77. Steedman, Mark. 1987. Combinatory grammars and parasitic gaps. Natural Language and Linguistic Theory 5: 403–439. 1996. Surface structure and interpretation. Cambridge, MA: MIT Press. Stockwell, Robert P., Paul Schachter, and Barbara Hall Partee. 1968. Integration of transformational theories of English syntax. Los Angeles: UCLA. Takahashi, Shoichi. 2004. Pseudogapping and cyclic linearization. In Proceedings of the thirty-fourth meeting of the North East Linguistic Society, eds. Keir Moulton and Matthew Wolf, volume XXXIV, 571–585. University of Massachusetts, Amherst: Graduate Linguistic Student Association. 2006a. Decompositionality and identity. Doctoral Dissertation, Massachusetts Institute of Technology. 2006b. More than two quantifiers. Natural Language Semantics 14: 57–101. Takahashi, Shoichi, and Danny Fox. 2006. MaxElide and the re-binding problem. In Proceedings of Semantics and Linguistic Theory, eds. Effi Georgala and Jonathan Howell, volume XV, 223–240. Ithaca, NY: CLC Publications. Takano, Yuji. 1996. Movement and parametric variation. Doctoral Dissertation, University of California, Irvine. 2000. Illicit remnant movement: an argument for feature-driven movement. Linguistic Inquiry 31: 141–156. Tancredi, Christopher. 1992. Deletion, deaccenting and presupposition. Doctoral Dissertation, Massachusetts Institute of Technology, Cambridge. Taraldsen, Knut Tarald. 1979. The theoretical interpretation of a class of marked extractions. In Theory of markedness in generative grammar, eds. A. Belletti, L. Brandi, and R. Rizzi, 475–516. Pisa: Scuola Normale Superiore. Tiedeman, Robyne. 1995. Some remarks on antecedent contained deletion. In Minimalism and Linguistic Theory, eds. Shosuke Haraguchi and Michio Funaki, 67–103. Tokyo: Hituzi Syobo. Tomioka, Satoshi. 1997. Focussing effects in VP ellipsis and NP interpretation. Doctoral Dissertation, University of Massachusetts, Amherst. 1999. A sloppy identity puzzle. Natural Language Semantics 7: 217–241. 2001. A certain scope asymmetry in VP ellipsis contexts. In Linguistic form and its computation, eds. Christian Roher, Antje Rossdentscher, and Hans Kamp, 183–203. Stanford University: CSLI Publications.
References
299
2004. Another sloppy identity puzzle. In Generative Grammar in a Broader Perspective: Proceedings of the Fourth GLOW in Asia, ed. Han-Jin Yoon, 383–404. Utrecht, The Netherlands: Generative Linguistics of the Old World. Truckenbrodt, Hubert. 1995. Phonological phrases: their relation to syntax, focus, and prominence. Doctoral Dissertation, MIT. Wasow, Thomas. 1972. Anaphoric relations in English. Doctoral Dissertation, MIT, Cambridge, MA. Webber, Bonnie. 1978. A formal approach to discourse anaphora. Doctoral Dissertation, Harvard University. Wesche, B. 1995. Symmetric coordination. Tubingen: Niemeyer. ¨ Wescoat, Michael. 1989. Sloppy readings with embedded antecedents. Manuscript, Stanford University. Wiese, R. 1993. Tilgung prosodischer konstituenten. Paper presented at 15th annual meeting of the Deutsche Gesellschaft fur ¨ Sprachwissenschaft. Mainz, Germany: University of Mainz. Wilder, Chris. 1995. Antecedent containment and ellipsis. In FAS Papers in Linguistics, volume IV, 132–165. 1997. Some properties of ellipsis in coordination. In Studies on universal grammar and typological variation, eds. Artemis Alexiadou and T. Hall, 59–107. Amsterdam: John Benjamins. Williams, Edwin. 1974. Rule ordering in syntax. Doctoral Dissertation, Massachusetts Institute of Technology. 1977. Discourse and logical form. Linguistic Inquiry 8: 101–139. 1978. Across-the-board rule application. Linguistic Inquiry 9: 31–43. 1986. A reassignment of the functions of LF. Linguistic Inquiry 17: 264–300. 1990. The ATB-theory of parasitic gaps. The Linguistic Review 6: 265–279. Wiltschko, Martina. 1997. Superiority in German. In West Coast Conference on Formal Linguistics, eds. Emily Curtis, James Lyle, and Gabriel Webster, 431–445. University of Washington: Stanford Linguistics Association. 1998. On the syntax and semantics of (relative) pronouns and determiners. The Journal of Comparative Germanic Linguistics 2: 143–181. Winkler, Susanne. 1996. Focus and secondary predication. Berlin/New York: Mouton de Gruyter. Wurmbrand, Susi. 2001. Infinitives: Restructuring and clause structure. Berlin/ New York: Mouton de Gruyter. Zwart, C. Jan-Wouter. 1997. Morphosyntax of verb movement: A minimalist approach to the syntax of Dutch. Dordrecht: Kluwer Academic Publishers.
Index
Across-the-Board Movement 13, 51, 76, 78, 79, 156, 229, 231, 233, 234, 252, 258, 275, 276, 280 alphabetic variants 41, 104, 128–130, 194, 195, 267, 269 anaphoric islands 5 Antecedent Contained Deletion (ACD) 2, 9–12, 30, 34–68, 96, 101, 102, 126, 154, 167, 174, 181–182, 208, 216, 263, 279 Binders Out approach 37, 39–40, 41, 42, 44, 46, 49, 263 Transitive VP approach 40, 41, 44, 52, 58 Argument Contained Ellipsis (ACE) 96, 104, 113, 122–128 auxiliaries, semantics of 34–35, 37, 57, 59–60 Background Matching 15–16, 17–19, 22–24, 25, 26–27 Background Construction Algorithm 17, 18–19 Backwards anaphora 77 c-command 239–240, 249, 256, 281 colinked 109 comparatives direct analysis 174–175, 179–181 ellipsis in 3, 11–12, 156, 216, 279 phrasal 155 completeness requirement 19, 21, 22, 25, 260 conjecture on double indexing 164, 165, 166, 167, 274 Coordinate Structure Constraint (CSC) 81, 136, 139, 144, 233–234 Crossover Strong Crossover 276 Weak Crossover 214–215, 278–279 deaccentuation 5 de dicto 154, 165, 166, 167, 273 dependence relation 96, 105
300
elliptical 97, 105–113, 123, 128–130 referential 97, 105, 113–120, 130 b-dependence 116, 117, 118, 119, 123–125, 126, 127, 269 r-dependence 115, 116, 269 de re 162, 166, 167, 273, 274 Derived VP Rule 45–46, 49 Direct Compositionality 30–31, 32–33, 45, 261 Strong 31, 54, 261 Weak 31, 261 dominance full dominance 238–239, 242, 281, 282, 283 multidominance 14, 230, 231, 233, 235–239, 242–244, 247–258, 281, 283 double object construction 73, 77 ECM infinitives 163–165, 167 Edge Restriction 229, 234, 235, 243, 258 feature agreement 188–189 focus 8, 17–18, 52, 66, 148, 184–190, 200–201, 202, 206–208, 224, 225–228 Alternative Semantics 17 contrastive 186, 193, 194, 198, 204, 207, 208, 215, 217 focus antecedent 185, 186, 201–202 focus domain 185, 186, 201–202, 277 focus indexing 226–227, 228 focus movement 149–150 focus sensitive particles 194–195, 196, 277 givenness 186 licensing 186, 192–193, 198, 203 formal variable 120–121, 123 functional gaps 62, 63 Gapping 11–13, 74–86, 156–161, 176, 181, 229, 230, 231, 233–234, 237, 245–246, 275, 276, 281 Isomorphism requirement 157, 158, 161, 167, 275 Geach rule 35, 39, 47, 48–49, 50, 59, 60, 63
Index Identity Condition on ellipsis 41, 55–56, 64, 65–66, 97, 98, 104, 105–113, 128–130, 134–135, 213, 218, 220, 262, 266–267, 269, 270–271 e-feature marking 212, 220, 221, 279–280 e-givenness 54–55, 134–135, 148, 150, 222–225 strict identity hypothesis 106 on Right Node Raising (RNR) 249–251 implicational bridging 203 islands 2, 3, 13, 71–72, 74, 132, 135–140, 149–150, 272 repair 132, 137, 143–146, 148 lambda abstraction 38 licensing of ellipsis 133–134, 278 Linear Correspondence Axiom 14, 229, 230, 234, 237, 240–244, 282 Linearization 14, 234–235, 240–244, 245, 256 Linked 109, 110 Max Elide 141–143, 152, 279 Merge 236–238, 249, 252, 282 NP Ellipsis 205–206, 207, 208, 209, 214 Parallel Dependencies 15–16, 19–20, 21–22, 27–28, 150 Parasitic Gaps 92 Phase Impenetrability Condition 222 Phonological Word 248, 282 Pied Piping 31, 32, 50, 61–63 Principle B 172–173 Principle C 52, 91–92, 154, 169, 171, 177, 179, 180, 275 proform account of ellipsis 4–5, 6, 55, 57, 210, 212, 215, 217, 228, 278 pronouns bound variable 183–184, 186, 187, 190, 192, 195–198, 202, 208–209, 212 Donkey anaphora 204 e-type 8–9, 197, 199, 200, 202, 203–204, 208, 209, 277 paycheck 217 Proper Binding Condition 73, 86, 87
301 Pseudogapping 3, 13, 69–70, 76, 77, 82, 92, 93, 259, 263 quantificational NPs 42, 151, 199–200, 202–203 scope of 71–73 Quantifier Raising (QR) 71, 73, 82, 111, 126, 127, 152, 199, 213 reciprocals 191 reconstruction 91–92, 276, 278 referential circularity 114–115, 116, 118–119, 128 reflexives 166, 190 relative clauses, semantics of 35, 36, 38–40, 60–61, 118–119, 120–122, 262 re-merge 255 remnant movement 87, 268 Right Node Raising (RNR) 3, 13, 156, 229, 231, 233, 234–235, 244–254, 256, 258 Scrambling 72, 87, 88–89, 90–91 silent constituent account of Ellipsis 5–6, 54, 76, 210, 219 sloppy anaphora 4, 7–9, 18, 19–20, 28, 191–192, 211–212, 214, 215, 216, 217, 221, 224, 228 Two Pronouns Puzzle 16 Sluicing 1–2, 3, 132–135, 143, 214, 278 small clause 77, 84, 101 superiority 88, 89–90 symmetry effect 24–25 type shift 32, 38–39, 47, 48, 49, 60 shift rule ‘‘z’’ 49–50, 53, 60, 63, 64 variable-free semantics 31, 39–40, 41, 44, 52, 55 vehicle change 260 Verb Phrase Ellipsis (VPE) 3, 15–68, 205, 206–207, 210, 222 Wrap 33, 43–44, 48 X-bar Theory 230