Вы находитесь на странице: 1из 229

cover next page >

title: Three Investigations of Extraction Current Studies in


Linguistics Series ; 29
author: Postal, Paul Martin.
publisher: MIT Press
isbn10 | asin: 0262161796
print isbn13: 9780262161794
ebook isbn13: 9780585078533
language: English
subject Extraction (Linguistics)
publication date: 1998
lcc: P158.17.P67 1998eb
ddc: 415
subject: Extraction (Linguistics)

cover next page >


If you like this book, buy it!
< previous page page_i next page >
Page i

Three Investigations of Extraction

< previous page page_i next page >


If you like this book, buy it!
< previous page page_ii next page >
Page ii
Current Studies in Linguistics
Samuel Jay Keyser, general editor
1. A Reader on the Sanskrit Grammarians, J. F. Staal, editor
2. Semantic Interpretation in Genera tire Grammar, Ray Jackendoff
3. The Structure of the Japanese Language, Susumu Kuno
4. Speech Sounds and Features, Gunnar Fant
5. On Raising: One Rule of English Grammar and Its Theoretical Implications, Paul M. Postal
6. French Syntax: The Transformational Cycle, Richard S. Kayne
7. Panini * as a Variationist, Paul Kiparsky, S. D. Joshi, editor
8. Semantics and Cognition, Ray Jackendoff
9. Modularity in Syntax: A Study of Japanese and English, Ann Kathleen Farmer
10. Phonology and Syntax: The Relation between Sound and Structure, Elisabeth O. Selkirk
11. The Grammatical Basis of Linguistic Performance: Language Use and Acquisition, Robert C. Berwick and Amy S.
Weinberg
12. Introduction to the Theory of Grammar, Henk van Riemsdijk and Edwin Williams
13. Word and Sentence Prosody in Serbocroatian, Ilse Lehiste and Pavle Ivic*
14. The Representation of (In)-definiteness, Eric J. Reuland and Alice G. B. ter Meulen, editors
15. An Essay on Stress, Morris Halle and Jean-Roger Vergnaud
16. Language and Problems of Knowledge: The Managua Lectures, Noam Chomsky
17. A Course in GB Syntax: Lectures on Binding and Empty Categories, Howard Lasnik and Juan Uriagereka
18. Semantic Structures, Ray Jackendoff
19. Events in the Semantics of English: A Study in Subatomic Semantics, Terence Parsons
20. Principles and Parameters in Comparative Grammar, Robert Freidin, editor
21. Foundations of Generative Syntax, Robert Freidin
22. Move a: Conditions on Its Application and Output, Howard Lasnik and Mamoru Saito
23. Plurals and Events, Barry Schein
24. The View from Building 20: Essays in Linguistics in Honor of Sylvain Bromberger, Kenneth Hale and Samuel Jay
Keyser, editors
25. Grounded Phonology, Diana Archangeli and Douglas Pulleyblank
26. The Magic of a Common Language: Jakobson, Mathesius, Trubetzkoy, and the Prague Linguistic Circle, Jindrich*
Toman
27. Zero Syntax: Experiencers and Cascades, David Pesetsky
28. The Minimalist Program, Noam Chomsky
29. Three Investigations of Extraction, Paul M. Postal

< previous page page_ii next page >


If you like this book, buy it!
< previous page page_iii next page >
Page iii

Three Investigations of Extraction

Paul M. Postal

The M IT Press
Cambridge, Massachusetts
London, England

< previous page page_iii next page >


If you like this book, buy it!
< previous page page_iv next page >
Page iv
1998 Massachusetts Institute of Technology
All rights reserved. No part of this book may be reproduced in any form by any electronic or mechanical means
(including photocopying, recording, or information storage and retrieval) without permission in writing from the
publisher.
This book was set in Times New Roman on the Monotype "Prism Plus" Post-Script Imagesetter by Asco Trade
Typesetting Ltd., Hong Kong. Printed and bound in the United States of America.
Library of Congress Cataloging-in-Publication Data
Postal, Paul Martin, 1936-
Three investigations of extraction / Paul M. Postal.
p. cm. (Current studies in linguistics; 29)
Includes bibliographical references and index.
ISBN 0-262-16179-6 (alk. paper)
1. Extraction (Linguistics) I. Title. II. Series: Current
studies in linguistics series; 29.
P158.17.P67 1998
415dc21 98-39766
CIP

< previous page page_iv next page >


If you like this book, buy it!
< previous page page_v next page >
Page v
Preface ix
Chapter 1 Introduction 1
1
1.1 Background
4
1.2 Remarks on Chapter 2
22
1.3 Remarks on Chapter 3
24
1.4 Remarks on Chapter 4
Chapter 2 Contrasting Extraction Types 25
25
2.1 Two Types of Left Extraction
26
2.2 A-Extraction/B-Extraction Differences
32
2.3 Antipronominal Contexts
35
2.4 Why Are B-extractions Incompatible with Antipronominal
Contexts?
42
2.5 Two Subtypes of A-Extraction

< previous page page_v next page >


If you like this book, buy it!
< previous page page_vi next page >
Page vi
Chapter 3 The Status of the Coordinate Structure Constraint 51
51
3.1 Background
56
3.2 A-Scenarios
77
3.3 B-Scenarios
90
3.4 C-Scenarios
92
3.5 Conclusion
Chapter 4 Right Node Raising and Extraction 97
97
4.1 Background
98
4.2 The Unity of Right Node Raising and Left Extractions
138
4.3 Right Node Raising and Slash Category Approaches to
Extraction
Appendix A Mistaking Selective Islands for Nonislands 165
165
A.1 Nonuniqueness of the Key Error
165
A.2 A Putative Case of "Reanalysis"
166
A.3 Preserving the Right Selective Island Constraint
167
A.4 A Denial That Certain Complex NPs Are Islands
167
A.5 The Island Status of Irrealis if Complements
168
A.6 Deep Extraction
169
A.7 Apparent Support for a Novel Treatment of Finite Subjects

< previous page page_vi next page >


If you like this book, buy it!
< previous page page_vii next page >
Page vii
171
A.8 Apparent Support for a "Vacuous Movement Hypothesis"
172
A.9 Implications
Appendix B Additional Arguments That Right Node Raising Is an
Extraction 173
173
B.1 Remarks
173
B.2 Strange Plural Right-Node-Raising Pivots
173
B.3 Inverse Copula Constructions
174
B.4 The Unextractability of Right-Dislocated Phrases
Appendix C Reaction to Referee Comments 175
Notes 181
References 201
Index 203

< previous page page_vii next page >


If you like this book, buy it!
< previous page page_i next page >
Page i

Three Investigations of Extraction

< previous page page_i next page >


If you like this book, buy it!
< previous page page_ix next page >
Page ix

Preface
This volume brings together three theoretical studies of extraction that were written independently in 1991-1993. These
appear as chapters 2, 3, and 4.
Chapter I, the introduction, is newly written. Section 1.1 extracts common introductory material from the original
versions of chapters 2-4, which are concomitantly simplified in nonsignificant ways. Sections 1.2-1.4 discuss in part
some of the theoretical underpinnings of the following chapters and address certain theoretical issues these chapters
raise but do not treat. In particular, certain referee criticisms are partially answered in chapter 1.
Chapter 2 began as "A Novel Extraction Typology," a paper read at the annual meeting of the Linguistics Association
of Great Britain, at the University of Birmingham in March 1993. A revised version was published under the title
"Contrasting Extraction Types" in Journal of Linguistics 30, 159-186. Chapters 3 and 4 have not previously been
published.
I have made only modest attempts to update the papers that appear as chapters 2-4, mostly with the goal of eliminating
redundancies and adding cross-references between them. Here and there I have added references to works postdating
the original completion of the papers, but this process has not been systematic. In general, it has not been possible to
take serious account of work seen after the original papers were completed. I have also modified the originals of
chapters 2-4 in ways pursuant to producing a joint bibliography and index.
The term extraction, due I believe to Jorge Hankamer, covers in a theoretically neutral way phenomena like the left
front position of English wh forms, topicalized phrases, and the like. Overall, the present papers, which focus on the
facts of English, treat aspects of extraction including

< previous page page_ix next page >


If you like this book, buy it!
< previous page page_x next page >
Page x
evidence for a new typology of extractions, island constraints and their proper characterization, the role of resumptive
pronouns in freeing extractions from island constraints, the relations between right extractions (e.g., so-called right node
raising) and left extractions, the feasibility of slash-category accounts of right extractions, and, briefly, relations
between ex-traction and plurals, coordination, and respectively-type constructions.
Preparation of this book has been greatly facilitated by National Science Foundation grant SBR-9510984 (to P.M.
Postal and M. R. Baltin), "Extraction from Selective Islands." I would like to thank Ann Delilkan, whose work as
research assistant has improved this project as well.
A number of linguists have commented on earlier versions of one or another part of this book in ways that have led to
significant improvements. I would like to thank Edwin Battistella, Robert Borsley, Joseph Emonds, David E. Johnson,
Warren Plath, Geoffrey K. Pullum, and John Robert Ross for their help in this regard. Special appreciation is due to
James D. McCawley for detailed substantive criticisms and suggestions with respect to what appears here as chapter 4. I
am also indebted to two anonymous MIT Press referees for their commentary on and criticisms of chapters 3 and 4. No
one but the author is responsible for whatever deficiencies nonetheless remain.

< previous page page_x next page >


If you like this book, buy it!
< previous page page_1 next page >
Page 1

Chapter 1
Introduction
1.1 Background
I use the term extraction in this book to refer descriptively to relations like those in (1) between the sometimes null
italicized constituents and the cosubscripted gaps (equivalently, the positions of those gaps) to their right. 1 For clarity,
when the constituent cosubscripted with a gap is complex, I place it in square brackets.
(1) a. Whol did they nominate t1 to be director?
b. the gun (which1) they claimed was used in the crime.
c. What1 Ellen wants t1 is a Mercedes-Benz.
d. [No such gorilla]1 did I ever see t1.
e. Stella tickled more chimps than (what1) I said that Dwight tickled t1.
f. [What a lovely woman]1 I found out that he married t1!
g. Frank, who1 they adore t1, is dishonest.
h. Frank1, I would never hire t1.
i. It was Frank who1 they hired t1.
In early work on transformational grammar, extractions tended to be viewed as independent entities, each defined by
some particular movement rule: the question extraction in (1a) was different from the restrictive relative extraction in
(1b), which in turn was different from topicalizations like (1h), and so on (see, e.g., Ross 1967; Postal 1971). But since
at least the publication of Chomsky 1977b, it has been widely assumed that all extractions to the left (hereafter L(eft)-
extractions) represent fundamentally the same phenomenon. In the Government-Binding (GB) terms, advocated by
Chomsky (1981, 1982, 1986a,b), this commonality is represented by analyses involving trace-leaving movements to
nonargument positions.

< previous page page_1 next page >


If you like this book, buy it!
< previous page page_2 next page >
Page 2
However, other frameworks not appealing to such notions also recognize the commonality of L-extractions. Hukari and
Levine mention several of these frameworks.
(2) ''The publication in 1977 of Chomsky's paper "On wh movement" marks the beginning of an era of remarkable
consensus in linguistic theory, not only among transformational grammarians but throughout the emerging community
of theorists committed to generative models of grammar in which transformations or equivalent processes play no role.
Chomsky's claim that all unbounded dependency constructions (UDCs) behave uniformly with respect to certain
logically independent syntactic criteria and therefore, by Ockham's razor, should be treated as a single species was
central to the emergence of the Binding Theory of current transformational grammar; but it was incorporated without
any apparent difficulties into Lexical Functional Grammar (e.g., Kaplan and Bresnan 1982) and various avatars of
Generalized Phrase Structure Grammar (e.g., Gazdar 1982, Gazdar et al. 1985 (hereafter GKPS), Pollard and Sag 1987).
In Kaplan and Bresnan (1982) the use of bounded domination metavariables ( and ) to express the linkage between
gaps and their fillers extended to indirect wh-questions, topicalizations, and tough constructions, while in GKPS a
mechanism of SLASH propagation accounts in addition for relative clauses, constituent questions and parasitic gaps.
From the end of the 1970s on, the unified treatment of UDCs emerges as a leitmotif of the increasingly fractionated
discipline of syntactic theory." (Hukari and Levine 1991, 97-98)
Other work also accepts the unity of L-extractions:
(3) a. The base-generated syntax of Brame (1978). Brame treats all L-extractions as involving parallel "binding" by
"operators."
b. The Head-Driven Phrase Structure Grammar of Pollard and Sag (1987, 1994)
c. The Categorial Grammar of Steedman (1985, 1988, 1989, 1996)
d. The Lexical-Functional Grammar described by Kaplan and Zaenen (1989)
e. Tree-Adjoining Grammar, as characterized by Kroch (1989)
Hukari and Levine (1991) go on to propose a partially new typology of extractions, which separates (among others)
those listed in (1) from those involving object raising, parasitic gaps (P-gaps), and so on.

< previous page page_2 next page >


If you like this book, buy it!
< previous page page_3 next page >
Page 3
Of course, unchallenged facts initially motivate viewing L-extractions as a single phenomenon. Restricting attention to
NP cases, (4a-i) are among the features shared by the extractions in (1).
(4) Properties of English NP extractions
a. They contain a visible gap.
b. The structural "distance" between the gap and its binder is unbounded.
c. They are island-sensitive (see Chomsky 1977b, 86).
d. They "license" P-gaps. 2
e. They induce strong crossover violations.
f. They determine crossing dependencies in the same way.
g. They have across-the-board instances.
h. They can strand prepositions in the same contexts.
i. They are subject to the same specific "pure extraction" constraints.
Space considerations preclude documenting these properties for all the construction types. However, (5) shows that
(4d,e) hold for what might be considered the most exotic case, namely, comparative extraction.
(5) a. Jerome followed more suspects than (what1) Arthur interrogated t1 without arresting pg1.
b. Jerome followed more suspects than (what1) Arthur convinced me/*them1 that you would help t1. (see Bresnan
1975)
Evidence for (4i) is that indirect objects are unextractable (in my English), regardless of the extraction construction.
(6) a.*Who1 did you give t1 perfume?
b.*the woman who1 I gave t1 perfume.
c.*I dated more women than (what1) he gave t1 perfume.
d.*Joan1, I gave t1 perfume.
e.*It was Joan who1 I gave t1 perfume.
Similarly, Langendoen and Pullurn (1977) note that with certain verbs it is impossible for some speakers to extract an
NP from the context to S. This also holds for all the types, as (7) indicates.
(7) a. I mentioned to Malcolm that it was snowing.
b.*Who1 did you mention to t1 that it was snowing?
c.*the person who1 I mentioned to t1 that it was snowing
d.*Alvin tickled more kids than (what1) I mentioned to t1 that it was snowing.
e.*Jerome1, I mentioned to t1 that it was snowing.

< previous page page_3 next page >


If you like this book, buy it!
< previous page page_4 next page >
Page 4
1.2 Remarks on Chapter 2
Starting from the assumption that L-extraction is a unitary phenomenon, chapter 2 investigates previously apparently
largely unnoticed systematic distributional distinctions among English L-extractions. This leads first to a major
subdivision of L-extractions into types referred to arbitrarily as A-extractions and B-extractions. The latter category
contains NP topicalization, NP clefting, and nonrestrictive relative clause extraction. The former contains all the rest,
including most types of question extraction and restrictive relative extraction as well as negative extraction, pseudo-
clefting, and others.
I argue first that B-extractions are notably more restricted than A-extractions with respect to their possible extraction
sites. Next I show that this extra restrictiveness correlates with constraints on the distribution of weak definite
pronouns. Essentially, unlike the site linked to an A-extraction, a B-extraction site cannot be what is called an
antipronominal context (AC), a position that is incompatible with weak definite pronouns. An example of an AC not
discussed in chapter 2 is provided by the object position of the verb tell (in its 'determine' sense). That such positions
ban B-extractions but not A-extractions is shown by the contrasts in (8), which differentiate the object position of this
verb from that of (e.g.) determine when B-extraction is involved but not otherwise.
(8) a. We could easily determine/*tell it.
b. What1 we could easily determine/tell t1 was that Mike was a spy.
c. The first thing that1 we could determine/tell t1 was that Mike was a spy.
d. [That Mike was a spy]1, we could easily determine/*tell t1.
e. That1, which1 I wish we had been able to determine/*tell t1 sooner, is surprising.
f. It was that which1 we could immediately determine/*tell t1.
A theoretical link between the two phenomena at issue, B-extractions and sensitivity to ACs, is established by positing
that a B-extraction inherently requires an (invisible) resumptive pronoun (RP) in its extraction site. By "inherently" I
mean that this is necessary independently of questions of extraction from islands, which, I argue below, demands the
presence of RPs for all types of L-extractions.

< previous page page_4 next page >


If you like this book, buy it!
< previous page page_5 next page >
Page 5
The association of B-extractions with RPs apparently raises a paradox for a theory of how extractions relate to islands,
since if one accepts Ross's (1967) claim (as I do), an extraction that leaves an RP in the ex-traction site is not island-
sensitive. The RPs that B-extractions determine would then apparently render them insensitive to island boundaries,
contrary to (well-known) fact. A sketch of the beginnings of a resolution for this paradox is advanced in section 2.4.3.
It involves the claim that the (invisible) RPs associated with English B-extractions must themselves L-extract,
essentially to a position rendering them sisters of the extractees. I suggest that this extraction is required in order for
these RPs to be con-trolled, control being the mechanism underlying their invisibility. The unchallenged island
sensitivity of B-extractions is then attributed not to the RP-inked extraction of the (usually) visible explicit extractee,
but to the extraction of the invisible RP. In chapter 2 I claim that the latter is an A-extraction. This is supposed to make
the overall B-extraction constructions island-sensitive for essentially the reasons Ross originally gave: there is an
extraction not linked to an RP.
Although I believe the above account contains a core of valid assumptions, many problems need to be dealt with and
the proposal remains too sketchy and vague to be anything more than programmatic. I return to key issues below.
Having established a distinction between A-extractions and B-extractions, I argue that A-extractions themselves divide
into at least two subtypes. This division is motivated by observing contrasts like those in (9) and (10).
(9) a. the guy who1 they asked Jane whether the police questioned t1
b. Joe saw more students than (what1) they (*asked Jane whether the police) questioned t1.
(10) a. [Which spy]1 did the police arrest anyone who had contacted t1?
b. The more people they praised, the more people (that,) the police arrested (*anyone who had contacted) t1.
These examples involve L-extraction out of what are called selective islands (see section 3.2.2). Roughly, a selective
island permits the extraction of only a very limited subset of all constituent types, mostly NPs. The generalization
reached is that the extraction site of all B-extractions but only some A-extractions can be inside a selective island when
the extractee is outside that selective island. A-extractions that cannot extract from selective islands are called A2-
extractions and include, as in (9b) and

< previous page page_5 next page >


If you like this book, buy it!
< previous page page_6 next page >
Page 6
(10b), comparative extraction and extraction associated with the more ... the more. A-extractions capable of spanning
selective island boundaries, as in (9a) and (10a), are called A1-extractions.
I might add, though it is not possible to explore these matters in any depth here, that work carried out after chapter 2
was written permits both a more extensive specification of A2-extractions than is given in chapter 2 and a richer
characterization of their properties. 3 This work leads me to believe that the extractions in (11), not mentioned in
chapter 2, also fall under the rubric of A2-extraction.
(11) a. The relative-like extraction associated with such + noun + as (see Bresnan 1977)
b. The relative-like extraction associated with the same + noun + that
c. The relative-like extraction associated with a different + noun + than
The failure of these constructions to permit extraction from the selective island in (9b) is illustrated in (12).
(12) a.*such men as1 they asked Jane whether the police questioned h
b.*the same men that1 they asked Jane whether the police questioned h
c.*a different guy than1 they asked Jane whether the police questioned t1
Furthermore, work carried out since chapter 2 was written has revealed that beyond their inability to extract from
selective islands, A2-extractions share other properties. First, they seem unable to extract a "backward" controller of the
sort found in (13a), as shown in (13b-e). (p marks the site of the invisible controlled pronoun.)
(13) a. p1 meeting the director thrilled everyone1/no one1/Mary1.
b.*[What a fool]1 p1 meeting the director thrilled t1!
c.*Jacqueline interviewed more employees than p1 meeting the director thrilled t1.
d.*The more people she talks to, the more people p1 meeting the director seems to thrill t1.
e.*Such employees as Pt meeting the director thrilled h should be promoted.
A1-extraction or B-extraction from the same positions is fine.4

< previous page page_6 next page >


If you like this book, buy it!
< previous page page_7 next page >
Page 7
(14) a. [Which employees]1 did p1 meeting the director seem to thrill t1?
b. [No other employees]1 did p1 meeting the director seem to thrill t1.
c. It was those employees who1 p1 meeting the director thrilled t1.
Second, A1-extractions seem incapable of yielding the "real" gap for a P-gap to the left of the "real" gap. 5,6
(15) a.*[What a fool]1 your interviewing pg1 thrilled t1!
b.*Jacqueline met more employees than your interviewing pg1 thrilled t1.
c.*I saw the same guy that your interviewing pg1 thrilled t1.
Again, neither A1-extractions nor B-extractions suffer from this limitation.
(16) a. The woman who1 your interviewing pg1 thrilled t1 is outside.
b. Jane1, I am sure your interviewing pg1 thrilled t1.
Third, as noted in Postal 1993b, in certain cases weak crossover effects do not appear in the presence of forms like
only, even, or own in contexts otherwise manifesting them, yielding contrasts like the one in (17).
(17) a.*Who1 did his1 mother criticize t1?
b. Who1 did only/even his1 mother criticize t1?
c. Who1 did his1 own mother criticize t1?
Notably, though, the sort of amelioration seen in (17b,c) is not found with A1-extractions.
(18) a.*[What a fool]1 (even/only) his1 mother criticized t1!
b.*Jane invited the same guy that1 his1 own mother criticized t1.
c.*Clara investigated more soldiers than their1 own parents praised t1.
Although constructing a theory predicting that A2-extractions manifest the particular array of features that they do is
now no doubt significantly more difficult than assumed in chapter 2, the correlation of properties supports the claim
made there that A2-extractions form a significant category in English syntax.
A key claim of chapter 2 is that each subtype of L-extractionA1, A2, and Bis properly characterized by its relation to
RPs. Essentially, as explored in the discussion around (74) of chapter 2:
(19) a. A B-extraction requires an RP in its extraction site.
b. An A1-extraction allows an RP in its extraction site.
c. An A2-extraction forbids an RP in its extraction site.

< previous page page_7 next page >


If you like this book, buy it!
< previous page page_8 next page >
Page 8
Given this division and Ross's (1967) principle linking extraction from islands to RPs in extraction sites, it follows that
B-extractions and A1-extractions but not A2-extractions should in general be able to extract from selective islands. In
contrast, A1- and A2-extractions but not B-extractions should be able to have extraction sites in ACs. But no L-
extraction should then allow an extractee external to an island I, and an extraction site internal to I, if that site is an AC.
In chapter 2 I seek to support the correctness of all of these implications.
The distinctions in (19) relate to the issue of the island sensitivity of B-extractions touched on above. I suggest in
chapter 2 that this sensitivity can be taken to follow from Ross's (1967) claim that extraction from islands is possible
only if the extraction links to an RP. To get an over-view of the issues, consider a typical example of the sensitivity in
question.
(20) a.*The woman who1 they notified Jack, who lived with h, was missing.
b.*Linda1, they notified Jack, who lived with t1.
The topicalization in (20b) is no less ill formed than the restrictive relative extraction in (20a), although the latter is an
A-extraction and the former a B-extraction. One of course wants to say that both (20a) and (20b) are ill formed for the
same reason, namely, because the extraction site is inside the nonrestrictive relative clause, an island, whereas the
extractee, who/Linda, is external to it.
However, such a description does not differentiate the B-extraction in (20b) from an A-extraction; moreover, it fails to
indicate the role of the RP that I argue to be associated with each B-extraction. The assumptions in that account are
roughly as shown in (21).
(21) a. Adoption of Ross's (1967) fundamental claim that a constituent C can licitly extract even across an island
boundary if C links to an RP in C's extraction site (see (25) of chapter 3 and following discussion for a more articulated
statement of Ross's original proposal)
b. Recognition that many of the RPs relevant to (21a) are invisible
c. A general requirement that the RP associated with any B-extraction must itself extract, roughly to be a sister of the
B-extractee's final/surface position
Such assumptions suggest analyzing (20b) as shown schematically in (22).

< previous page page_8 next page >


If you like this book, buy it!
< previous page page_9 next page >
Page 9
In this and following diagrams, I use squared-off lines connecting phrases to represent the relation between an extractee
and the RP it determines. 7 The idea advanced in chapter 2 is that, like any A-extraction from a nonrestrictive relative
clause, (20b)/(22) is ruled out because, although the extraction of Linda links to an RP, permitting its extraction from an
island like the nonrestrictive relative clause in (20b) without violating Ross's (21a), the requirement in (21c) that the RP
in question itself extract to a position sister to that of Linda and hence external to the island leads to a violation of (2la)
with respect to the invisible RP.
However, that earlier account raises as yet unanswered questions. What, for instance, prevents the RP extraction in (22)
from linking to a second RP in the original extraction site, potentially (wrongly) freeing the second extraction from
island constraints under Ross's principle? This could be precluded if something like (23b) either followed from some
more general principlesay, (23a)or was simply a grammatical axiom.
(23) a. Any extraction of an RP is an A2-extraction.
b. The extraction of the RP linked to a B-extraction extractee is an A2-extraction.
Principle (23) correctly precludes a second RP in the extraction site of (22) and, given previous assumptions, would
serve to correctly block (20b), which then involves extraction of an element (RP1) from an island without a linked RP
in the extraction site.
The key issue, though, is how a framework incorporating the con-junction of (21) and (23) can be kept compatible with
a basic idea that underlies the last part of chapter 2, much of chapter 3, Postal and Baltin 1994, and Postal, in
preparation a,b. This idea is that particular L-extractions from some islandsnamely, selective islandsare possible
because invisible RPs appear in their extraction sites, rendering such extractions consistent with Ross's (21a).
So, in the work just cited, examples like (24a,b) involve extraction from the marked sites of the extractees (in small
capitals), even though the final loci of the latter are separated from those sites by the relative clause boundaries. (Here
and below, where useful, island boundaries are indicated by angled brackets.)
(24) a. WHAT1 the secret police arrested everyone who saw t1 was a video.
b. [THAT VIDEO]1, the secret police arrested everyone who saw t1.

< previous page page_9 next page >


If you like this book, buy it!
< previous page page_10 next page >
Page 10
Moreover, in the work in question, examples like (24a,b) are taken to represent extraction from (selective) islands. In
general in this chapter, I will not argue for the islandhood of various constituents asserted to be selective islands. The
reader can verify, however, that all such constituents satisfy the various conditions on islands specified in chapters 2
and 3. In particular, none of them permit extractions of non-NP constituents of the types how long, why, and so on.
Moreover, they do not permit non-extractive cross-constituent relations of the sort characteristically blocked by islands,
several of which are discussed in chapter 3. Notably, they also bar NP extractions from ACs. Recalling the non-AC/AC
contrast in object positions in (8), observe the difference between the two cases in (25).
(25) WHAT1 the secret police arrested everyone (who might be able to determine/*tell t1) was that the minister of
security was an extraterrestrial.
Such differences would, I think, be essentially incomprehensible if examples like (24) were taken simply to show that
the restrictive relative clauses in question were nonislands.
Consider (24b). Given claim (21c), the RP associated with the B-extraction topicalization must end up a sister of the
topic, meaning that it as well as the topic itself needs to extract from the relative clause. Principle (23) would preclude a
second RP. Nonetheless, (24b) is essentially well formed for many speakers (including me), contrasting sharply with
(20a,b), which are truly impossible (probably for everyone). 8 The problem is this: if (20a,b) are ungrammatical
because they violate (21a), how can (e.g.) (24b) both involve RP extraction from an island, to satisfy (2lc), and satisfy
both (21a) and (23)?
Various approaches to these problems are possible, but I have not been able to find one that is truly satisfying. I will
nonetheless sketch an ac-count that is consistent with the basic facts and with most of the key assumptions of chapters 2
and 3. At the worst, it can be taken as clarifying the issues that need to be faced.
The analysis I will describe rejects (23) and assumes among other things that cases like (24) involve two (obviously)
invisible RPs. Given this, plus a regimentation of island types to allow for the distinction between islands like the one
in (20) and the one in (24), it is possible to show how a framework incorporating Ross's (21a) blocks extraction from
the former island type but not from the latter.

< previous page page_10 next page >


If you like this book, buy it!
< previous page page_11 next page >
Page 11
The initial motivations for recognizing two RPs in cases like (24) can be summarized as follows. (a) Data independent
of such cases discussed at length in chapter 2 argue that B-extractions link to invisible RPs independently of island
facts. (b) The contrast between B-extractions and A-extractions shows that the latter are not so linked. (c) However,
neither in cases like (20) nor in those like (24) is there any contrast between A-extractions and B-extractions. So what
distinguishes type (20) islands from type (24) islands is arguably independent of the invisible RP systematically linked
to B-extractions. (d) Other evidence, some discussed in chapters 2 and 3, shows that extraction from islands like (24)
(i.e., selective islands) systematically involves invisible RPs, even in the case of A-extractions. Combining these
considerations with my adoption of Ross's basic claim in (21a) suggests that a B-extraction from a selective island
involves two RPs, and I believe the simplest assumption is that in that case A-extractions do as well. This means that
both cases of (24) are treated as involving pairs of invisible RPs.
To develop this approach further, one should first distinguish two types of RPs associated with L-extractions, call them
controlled RPs and non-controlled RPs. 9 Controlled RPs are the ones relevant for cases like (24). They are always
invisible. Noncontrolled RPs are either visible surface pronouns or, if not, invisible only in contexts where the language
in question permits/requires non-RP pronouns to be invisible (see the French case in note 9). In English, noncontrolled
RPs are found only rather marginally, in the left dislocation construction of (26), in right dislocations, and in certain
extractions from relative clauses in nonstandard English; see (61) of chapter 2 and (30)-(31) below.
(26) Marilyn1, I suspect Bob is going to fire her1.
As is well known, other languages (e.g., Hebrew, Irish) have non-controlled RPs in more central constructions (see,
e.g., Borer 1984; McCloskey 1979). As Ross (1967) observed, noncontrolled RP con-structions are essentially
insensitive to islands.
I assume that all controlled RPs must themselves extract, but not (as claimed earlier for B-extractions) always to a
sister position of the extractee. Given that requirement and (21a), extraction from islands via invisible (controlled) RPs
as in, for example, (24) is possible because of the assumptions in (27).

< previous page page_11 next page >


If you like this book, buy it!
< previous page page_12 next page >
Page 12
(27) a. The (controlled) RP left by an extractee like WHAT in (24a)call it the primary (controlled) RPextracts via a non-
A2-extraction (i.e., itself links to a secondary (controlled) RP); a primary RP is, of course, a controlled RP linked to an
extractee that is not an RP.
b. Primary controlled RPs extract to sister positions of the extractees that determine them; this principle subsumes (21c)
as a special case.
c. A key issue is what should be said about secondary RP extraction. Tentatively, I suggest that a secondary controlled
RP, RPx,, extracts no further than to the left boundary of the lowest island from which the primary RP that RPx links to
extracts. This means that secondary RPs never extract from islands.
d. Tertiary RPs (i.e., RPs linked to extracted secondary RPs) are excluded entirely. In other terms, secondary RP
extraction is always an A-extraction. Given (2la), this is why secondary RPs cannot extract from islands.
e. Islands like the one in (20) and the one in (24) can now be distinguished as described in (28).
(28) The class of islands partitions into the disjoint sets locked and unlocked such that an island is unlocked if and only
if it permits extraction of an RP.
In the terms just sketched, it would have to be specified that English nonrestrictive relative clauses are locked islands,
whereas certain restrictive relative clauses including those of the type in (24) are unlocked. It is obviously an important
task to seek principles predicting how islands are subclassified as locked or unlocked and to determine their degree of
universality. I cannot consider these matters here. The key point is that what are called selective islands throughout this
book will all be unlocked islands. The double terminology is not really redundant because the notion ''unlocked"
provides the beginning of a theoretical account of what deter-mines that an island falls into the purely descriptive
category "selective."
In these terms, a grammatical example like (24a) would have the schematic structure in (29), where {...} represents the
secret police arrested everyone.

< previous page page_12 next page >


If you like this book, buy it!
< previous page page_13 next page >
Page 13
In (29) the path between the extractee WHAT1 and its extraction site can span the relative clause island boundary
because the former links to a (primary) RP, in accord with (21a). That RP must itself extract because it is a controlled
RP and must extract to the point of the extractee to which it is linked and hence out of the same island because it is
primary. More-over, the primary RP can extract from that island because the latter is unlocked. But to take advantage
of that property, the primary RP extractee must, in accord with Ross's principle, itself link to a secondary RP. That
controlled RP must also extract and does, but nothing forces it to extract further than to the boundary of the lowest
island the primary RP extracts from, here the relative clause. There is and can be no extraction from an island of the
secondary RP because, given my interpretation of Ross's principle as (21 a), that would require a tertiary RP, banned by
(27d).
Note that recognition of a primary RP in (29) subsumes this A-extraction case under the notion of unlocked island as
given in (28). In the same terms, the ungrammatical example (20a) would have the structure shown in (30).

Here ill-formedness results because the primary RP has extracted from a locked island, and locked islands, by
definition (28), do not permit this.
The RP extraction heavily appealed to here is evidently subject to skepticism not least because both the primary and
secondary RPs at issue are invisible and hence their existence and claims that they extract can only be based on abstract
theoretical considerations. Two relevant considerations are these. First, positing that RPs extract is independently
motivated since the clear, visible RPs known in other languages (e.g., German, Hebrew, Irish) sometimes extract (see
Koster 1987, 63; Borer 1984, 220-222; McCloskey 1979, 94-97; Sells 1984, 91-94). Second, it is arguable that even
certain visible English RPs are subject to L-extraction. In chapter 2 I assume that both the English left dislocation and
right dislocation constructions, respectively illustrated in (3lb) and (3lc), involve RPs (italicized here).
(31) a. They realize I will never support Alfred again.
b. Alfred1, they realize I will never support him1 again.
c. They realize that I will never support him1 again, Alfred1.

< previous page page_13 next page >


If you like this book, buy it!
< previous page page_14 next page >
Page 14
And these RPs can apparently be the extractees of certain L-extractions.
(32) a. Alfred1, they realize that him1, I will never support t1 again.
b. They realize that him1, I will never support t1 again, Alfred1.
I would interpret (32a,b) as involving topicalization of the RPs linked to left dislocation and right dislocation. And they
do not seem at all bad. Thus, it is fairly clear that grammatical theory must allow for RP extraction independently of the
cases involving selective islands. 10
Further, one can briefly defend the claim that, for example, (24a,b) manifest double RP extraction. Although one might
initially view this idea as unnecessarily complex, it permits an account of the following fact. Not only does B-extraction
from an unlocked island as in (24b) obey the same conditions on unlocked island extraction as the A-extraction in (24a)
but in fact these conditions overlap with those on object raising, object deletion, and P-gap constructions, as touched on
in Cinque 1990, sec. 3.2.3, Postal 1993a, 750-753. Moreover, these conditions contrast with the conditions holding for
B-extractions that do not involve the spanning of island boundaries. I cannot fully document this here, but I will cite
two restrictions.
First, although it is licit for a finite-clause subject to B-extract (e.g., topicalize), it is impermissible in a wide range of
cases for such a subject to be an object-raising gap (see (33a,b)), an object deletion gap (see (33c,d)), or a P-gap (see
(33e,f)). In the same way, it cannot be a gap of extraction from an unlocked island, regardless of whether the relevant
extraction is of type A or B (see (33g,h)). Compare the illicit cases of (33) with the licit B-extraction of (34).
(33) a. Mike is difficult to imagine Mary loves.
b.*Mike is difficult to imagine loves Mary.
c. Mike is too young to imagine Mary would marry.
d.*Mike is too young to imagine would marry Mary.
e. Who1 did Mike call t1 after concluding Mary would marry pg1?
f.*Who1 did Mike call t1 after concluding Pg1 would marry Mary?
g. Mike1, the police regret (that) you notified t1.
h.*Mike1, the police regret (that) t1 notified you.
(34) Mike1, I cannot imagine t1 would ever marry Mary.
However, when the same position is embedded in an unlocked island, the restrictions illustrated in (33) appear.

< previous page page_14 next page >


If you like this book, buy it!
< previous page page_15 next page >
Page 15
(35)*Mike1, they regret t1 married Mary).
Second, for reasons that need not concern us here, the constructions of (33) are incompatible with the object of the
preposition in (36) when the verb is matter but not when it is speak; however, independently of islands, B-extractions
are not sensitive to this difference.
(36) a. Mike matters/speaks to Mary.
b. Mary is impossible for Mike to*matter/speak to.
c. Mary is too silly for Mike to*matter/speak to.
d. Who1 did the police interrogate h because Mike*mattered/spoke to pg1?
So simple B-extractions from this position are fine with either verb.
(37) a. Mary,, Mike does matter/speak to t1.
b. Mary1, who1 Mike does matter/speak to t1, is outside.
However, if the position is embedded in an unlocked island, B-extraction fails (only) with matter.
(38) a. Mary1, they asked me whether Mike*mattered/spoke to t1.
b. Mary1, they arrested everyone who*mattered/spoke to t1.
c. Mary1, they would prefer it if no one*mattered/spoke to t1.
One can begin to make sense of these facts under the double RP ex-traction view. This is so because simple B-
extractions (more precisely, the extraction of the primary RP) can then be taken to manifest one kind of control (call it
extraction control A (EXCA)) whereas the other cases, involving less distant extraction of pronouns, including
extraction of the controlled secondary RP, can be taken to manifest a second kind of control (call it extraction control B
(EXCB)). It can further be assumed that EXCA requires the presence of the controlled RP at the surface locus of the
controlling extractee (guaranteed by (21c)) and that EXCB does not. The tighter restrictions illustrated in (33)-(38) are
then conditions on EXCB. And ill-formed B-extractions from unlocked islands like those in (35) can be ruled out
because, under the double RP view of such cases, they manifest an instance of EXCB (involving the secondary RP) as
well as an instance of EXCA.
The scheme just sketched remains entirely informal; it is moderately complex; and certainly not all of its components
have been extensively justified. The task of developing these ideas in a truly serious way is beyond the scope of this
book. But it is worth stressing several virtues. First,

< previous page page_15 next page >


If you like this book, buy it!
< previous page page_16 next page >
Page 16
the above remarks provide the mechanisms for allowing RPs to serve as catalysts for extraction from islands along the
lines basically uncovered by Ross (1967). Second, it accounts for the fact that B-extractions, shown in chapter 2 to link
to (primary) invisible RPs quite independently of all island facts, are nonetheless island-sensitive.
Third, the framework sketched above accounts for the fact that ex-traction from a selective island is blocked if a locked
island intervenes along the path between the extraction site and the extractee position. This is illustrated in (39b),
whose partial structure would be schematically that shown in (39c).
(39) a. WHAT1 Joan would prefer it if you read t1 is Das Kapital.
b.*WHAT1 Joan {would prefer it if you called} Mike, who read t1, is Das Kapital.

Here, WHAT1 can successfully extract from the locked island represented by the nonrestrictive relative clause
embedded inside the unlocked island represented by the complement of prefer. This follows since that extraction
determines a (primary controlled) RP. However, for that RP to ex-tract to a sister position of the non-RP extractee, as
required by (27b), it also must extract from the locked island (requiring a further RP). But RP extraction is just what
locked islands (by definition) do not permit.
Fourth, the ideas sketched above address, I suggest, an MIT Press referees criticism with respect to chapter 3:
"Apparently, the possibility of null resumptive pronouns in English must be severely restricted; other-wise, Postal
would end up making a prediction that even extraction from what he calls strong islands should not exhibit any island
(i.e. Subjacency and/or ECP) effects, a seriously incorrect prediction." The referee's claim that the distribution of
controlled (hence null) RPs in English must be severely restricted is, evidently, essentially correct. And by
distinguishing locked from unlocked islands and by banning tertiary RPs, I believe the excess freedom whose existence
the referee foresees is excluded. That is, I have shown, for example, why the ungrammatical (20), involving extraction
from a nonrestrictive relative, cannot be salvaged via the mechanism of invisible controlled RPs that do save the island
extractions in, for ex-ample, (24). The difference of course is that the former involves a locked

< previous page page_16 next page >


If you like this book, buy it!
< previous page page_17 next page >
Page 17
island, the latter an unlocked one. Another way to put the point is that the referee's remarks seem to assume that the
invisible RPs hypothesized and argued for in chapter 3 differ from the visible in-situ RPs found in languages like
Hebrew and Irish only in being phonetically null (see also note 9). But the present chapter has sketched a much more
articulated account of the controlled RPs in question.
Certain other factors may suggest that the ideas outlined here are in-adequate, although the problems do not yield
counterexamples. First, the proposals do not account for the property that all of the secondary con-trolled RPs I have
posited are controlled by primary controlled (hence invisible) RPs. Logically, it would be possible for a controlled RP
to be controlled by a visible RP. If such cases cannot be attested, some as yet unformulated principle would have to
guarantee their nonexistence.
Second, although I have provided an account of the locked/unlocked island contrast that appeals to the fact that locked
islands do not allow RP extraction, there are clearly also distinctions to be made among unlocked islands. Cinque
(1990, 109) observes that nonfinite adjuncts are what in my terms would have to be analyzed as unlocked islands and
further that in effect compounds of these adjuncts do not permit extraction. That is, extraction from one such adjunct
inside another is not possible.
(40) a. the article that1 we went to England without reading t1
b.*the book that1 we left Russia without being arrested after distributing t1
However, it would be a mistake to conclude that extraction from compounds of unlocked islands is uniformly
impossible. Unlocked islands seem to subdivide in this respect. For instance, compounds of the unlocked island
represented by emotive factive complements like those of sorry seem to permit extraction relatively freely.
(41) the book which1 Greta is sorry (that Mike is sorry (that Sally is sorry)) that you read t1
I do not find any decay associated with the extra degrees of embedding in such cases beyond what is expected from
sheer complexity. If that is the case, then a full theory of islands must account for the difference between allowed and
disallowed extraction from compounds of unlocked islands. Although this complex matter has not been studied in
depth, there unfortunately seems to be no natural way to express this distinction consistently with all previous
assumptions.

< previous page page_17 next page >


If you like this book, buy it!
< previous page page_18 next page >
Page 18
But there is a rather straightforward account that is not consistent with previous assumptions. By definition, every
unlocked island permits (primary) RP extraction. This is required in order to allow extraction of primary controlled RPs
to the positions of the extractees to which they link. The contrast between selective islands whose compounds permit
extraction and those whose compounds do not indicates that there are two types of unlocked islands. It might be
hypothesized that this subdivision has to do with the behavior of secondary RPs.
If secondary RPs were required to end up in the same position as primary RPs, they would have to extract from some
islands (e.g., in (40a)), requiring either that one posit tertiary RPs or that secondary RPs be excluded from the scope of
Ross's claim (21a). Instead, I have suggested (27c). But I have found no way to develop an account of the distinctions
among unlocked islands in terms of (27).
Suppose one claimed:
(42) The class of unlocked islands partitions into the categories rigid and flexible such that an unlocked island I is rigid
if and only if any constituent extracted from I must link to an RP at the left boundary of l.
The assumption underlying (42) is that, for example, nonfinite adjuncts are rigid unlocked islands. Therefore, in cases
where present terms require primary RP extraction from such islands, this will only be possible if the secondary RP
linked to a primary RP can reach that island.
In cases involving a single rigid island, like (40a), (42) has no relevant consequences. But for compounded rigid
islands, (42) will require the secondary RP to extract from the lowest rigid island. Given earlier assumptions, this would
require tertiary RPs, which I have excluded. To see this, assuming that English nonfinite adjuncts are rigid unlocked
islands, consider the ill-formed (40b), whose partial schematic structure would be (43).

The non-RP extractee, taken here inessentially to be that, can extract from the lowest unlocked island because of the
primary RP. The latter can in turn extract without violation from the lowest island and the con-

< previous page page_18 next page >


If you like this book, buy it!
< previous page page_19 next page >
Page 19
taining adjunct island because of the secondary RP, given that these islands are unlocked. But both of these unlocked
islands are rigid. If the containing one did not exist, the secondary RP could satisfy the definition of rigid in (42)
without running afoul of the principle (27d) banning tertiary RPs by extracting only to the front of the after phrase.
That is why extraction from one nonfinite adjunct can yield grammatical (rigid) selective island extraction in English,
like Cinque's (40a). But in (40b) a second (outer) island exists and is also rigid. Since the primary RP has extracted
from that island as well, it follows that the secondary RP must appear at the left boundary of the without constituent, as
in (43). For that to be the case, the secondary RP must extract from the smaller rigid unlocked island, requiring, under
my interpretation (27a) of Ross's (1967) claim, that it also link to an RP. But that would be a tertiary RP, in violation of
principle (27d).
Another instance of rigid unlocked islands is provided by subject complements, which I take to be a highly restricted
example of this category. That it is, contrary to widespread belief, possible to extract from some subjects
(independently of P-gaps) under limited conditions is shown by cases like (44a), due to Kuno and Takami (1993, 49).
11
(44) a. That is something which1for you to try to understand t1 would be futile.
b. That is something which1for him to be able to determine/*tell t1 would be quite surprising.
But that such examples illustrate selective island extraction rather than providing evidence for the nonislandhood of
subjects is shown among other things by the fact that the relevant type of extraction is impossible from an AC.
Recalling the distinction in this regard between determine and tell, note that the contrast in (44b) correlates with the AC
contrast in (8a). The same point is made by the fact that parallels to (44a) involving non-NPs are quite bad.
(45)*That is something [under which]1for you to try to hide t1 would be futile.
Given then that cases like (44) represent RP-facilitated selective island extraction, (42) determines correctly that such
extraction is not possible from compounded subject island complements.
(46)*That is something which1Bill's trying to understand t1 being sneered at by the teachers) was criticized by the
principal.

< previous page page_19 next page >


If you like this book, buy it!
< previous page page_20 next page >
Page 20
(42) also predicts that it is impossible to extract from a nonfinite adjunct inside a subject complement or from a subject
inside a nonfinite adjunct, since both represent rigid islands. These predictions seem true.
(47) a.*That is something which1 for you to get drunk before understanding t1 would be silly.
b.*That is something which1 Henry got drunk without Bob's being able to understand t1 having been discussed.
Hence, (42) would seem to offer a basis for the ban on extraction from compounded rigid islands despite the possibility
of extraction from a single such island. However, this desirable result, consistent with previous assumptions, runs up
against the unhappy fact that (42) combines with earlier ideas to make very strong but (I think) untenable claims.
Namely, it should be impossible to extract from compounded islands when the outermost is a rigid unlocked island and
all the others are flexible. This is so because extracting the primary RP from the rigid island requires the secondary RP
that links to it to appear at the level of that constituent. But this it can do only by itself extracting from all of the
embedded flexible islands, requiring under principle (21a) previously banned tertiary RPs. The situation is illustrated in
(48).
(48) a. Mike went home without criticizing anyone who defended Lucille.
b. Mike criticized (everyone who went home without defending Lucille.
c. It was Lucille that1 Mike went home without criticizing anyone who defended t1.
d. It was Lucille that1 Mike criticized everyone who went home without defending t1.
In (48) I again take the nonfinite adjuncts to be rigid unlocked islands but the quantifier-headed restrictive relatives to
be flexible unlocked islands. (42) predicts that extractions from combinations of these should differ depending on
which is embedded in the other. (48c) should be ungrammatical because the rigidity of the adjunct island allows
extraction from it only in the presence of the secondary RP at that level, which would illicitly force the secondary RP to
extract from the flexible unlocked island defined by the relative. (48d) is not blocked in a parallel way since the
primary RP can extract from the exterior flexible island without the secondary RP having to exit the inner (in this case,
rigid) island. But neither

< previous page page_20 next page >


If you like this book, buy it!
< previous page page_21 next page >
Page 21
(48c) nor (48d) seems to be ungrammatical, and I do not see that they differ in status. That is, both seem to contrast
rather clearly with cases like Cinque's (40b).
A parallel pattern results when emotive factive complements of predicates like shock are chosen to represent a flexible
unlocked island.
(49) a. Laura smiled despite being shocked that they fired Ed.
b. Mike was shocked that Laura smiled despite having to fire Ed
c. the guy who1 Laura smiled despite being shocked that they fired
d. the guy who1 Mike was shocked that Laura smiled despite having to fire
Given (42), (49c) should be ungrammatical and (49d) should be grammatical. But again I see no difference. Once more,
extractions from island combinations seem grammatical even though, to satisfy (42), certain sec-ondary RPs would
have to extract from islands in violation of earlier assumptions.
In short, (42) predicts a stronger asymmetry between extractions from island compounds than is warrantedfor instance,
apparently wrongly predicting cases like (48c) and (49c) to be impossible.
One could maintain (42) by weakening earlier assumptions. For in-stance, one could posit, contrary to (27), that
secondary RPs can extract from (only some) islands, namely, flexible ones. This would mean either keeping (21a) and
allowing tertiary RPs or continuing to bar tertiary RPs but weakening (21a) to allow secondary RPs to extract from
islands without linking to additional RPs. Neither seems like a happy conclusion.
Although (42) thus has certain attractions, the overall facts about island embedding do not permit any obviously neat
way of combining it with other assumptions. If (42) is simply rejected, as is certainly possible, then I currently can
offer no account of the genuine contrasts between rigid and flexible unlocked islands. That is, it seems that minimally a
proper ac-count of extraction from islands should yield the following (no doubt partial) typology of islands or its
equivalent:
(50) a. A locked island allows no RP extraction.
b. An unlocked island allows some (namely, primary) RP extraction.
i. A rigid unlocked island allows extraction only if an RP linked to the extractee can reach the island left periphery.
Extraction from compounds of rigid islands is banned.

< previous page page_21 next page >


If you like this book, buy it!
< previous page page_22 next page >
Page 22
ii. A flexible unlocked island allows extraction without meeting (50bi) so that extraction from compounds of flexible
islands is allowed, as is extraction from a complex of flexible islands embedded in a single rigid island.
The ideas proposed in this section do not achieve the goal of providing such an account, and it cannot be advanced
further here.
1.3 Remarks on Chapter 3
Chapter 3 takes as its point of departure the purported demonstration by Lakoff (1986) that the Coordinate Structure
Constraint (CSC), first pro-posed by Ross (1967), is falsified by a range of data. Lakoff's claim is remarkable because
(a) the CSC is often taken to be the most robust island constraint known, and (b) the data supposedly falsifying it are
not from some poorly studied language, but from English.
Although the facts Lakoff cites for his radical negative conclusion are in essence all correct, I argue that the conclusion
itself is entirely unfounded. This curious conjunction is defensible because it can be argued that Lakoff's argument in
part misconstrues what Ross's formulation of the CSC claims. This misstep takes the form of ignoring the codicil added
in chapter 6 of Ross 1967, limiting the effects of island boundaries in cases where extractions are linked to RPs.
Basically, although syntacticians often tend, informally, to think of island boundaries as absolute barriers to extraction
relations across them, Ross's account allows such extraction if the extraction sites contain RPs. I of course developed a
version of Ross's ideas in my terms in section 1.2.
This feature of the CSC can be used as a basis for rejecting Lakoff's conclusions although none of the cases he cites as
counterexamples involve visible RPs. For it is not only possible to posit invisible RPs in the relevant cases, the
theoretical basis for which I sketched in section 1.2; in fact, it is possible to establish a methodology that supports their
existence. Since no arguments preclude recognizing nonphonetic RPs and since chapter 2 in particular provides strong
support for their existence in English, appealing to such RPs as part of an attempt to refute Lakoff's claims is in no way
illegitimate. Given the existence of covert RPs, it can be seen that rather than instantiating illicit extraction from islands
of the sort the CSC was intended to ban, the cases Lakoff advances mostly involve a subspecies of the type of
extraction from islands that Ross's (1967)

< previous page page_22 next page >


If you like this book, buy it!
< previous page page_23 next page >
Page 23
conception of islands permits. In the terminology of section 1.2, the majority of cases that Lakoff takes to show that
coordinate structures are not islands can be analyzed as extractions from selective islands. That is, they can be treated as
unlocked islands, extraction from which is permitted via the mechanisms of controlled (hence invisible) RPs.
In addition, I present other evidence arguing that the cases Lakoff takes to involve illicit (under the CSC) extraction
from coordinate constituents actually involve licit extraction from islands, licit because of the presence of RPs.
In defense of Lakoff's discussion, I should note that the argument flaw that I believe undermines his conclusion is easy
to fall into given that Ross exclusively considered visible RPs, whereas those relevant to Lakoff's discussion and the
CSC are nonphonetic. That is, it is hardly surprising that Lakoff (and many other linguists) has misconstrued certain
RP-facilitated L-extraction from certain selective islands as L-extraction from nonislands, because of (among other
things) the invisibility of the RPs involved.
I also show in chapter 3 that certain contexts Lakoff took to invalidate the CSC are not in fact coordinate and thus are
logically incapable of providing counterexamples to the CSC regardless of their relation to extraction.
Having rejected Lakoff's conclusions and hence having defended the CSC against what appear initially to be severe
counterexamples, I also show (in appendix A) that Lakoff's mistaken conclusion is in effect a special ease of a rather
common errornamely, confusing extractions from islands legitimized by controlled RPs (i.e., extractions from selective
islands) with extractions from nonislands. This error has considerable implications for any general theory of extraction
and for particular principles that have been proposed and defended in the literature. I argue that several principles and
their supporting arguments are vitiated by the same sort of mistake that underlies Lakoff's conclusion about the CSC.
Specifically, I discuss conclusions by Chomsky (1977b, 1986a), Kayne (1985, 1994), Pollard and Sag (1994), Pullum
(1987), Deane (1988, 1991), and Chung and McCloskey (1983), arguing that none are tenable since all overlook the
fact that the extractions they deal with are of the special sort involving unlocked islands and invisible RPs, hence are
not barred by island boundaries in the same way that non-RP-linked extractions are.
Beyond its defense of the CSC, the key import of chapter 3 thus involves its focus on the existence of ignored invisible
(because controlled)

< previous page page_23 next page >


If you like this book, buy it!
< previous page page_24 next page >
Page 24
RPs (in English) and their strong relevance to a wide range of conclusions about extractions, islands, and the conditions
governing them.
1.4 Remarks on Chapter 4
In chapter 4 I critique claimsprincipally by McCawley (1982, 1987, 1988), but also by Levine (1985)that the
construction type called right node raising (RNR) is not an extraction in the same sense as L-extractions like those
involving wh forms, topicalization, and so on. I develop a range of arguments to counter this view and to support the
conclusion that RNR is properly regarded as a subspecies of the same sort of phenomenon as L-extraction, albeit one
with special properties. This conclusion has various implications, for despite the commonalities between RNR and L-
extractions, there are also clear contrasts, which interfere with certain theoretical accounts of extraction based almost
exclusively on L-extractions.
In chapter 4 I also investigate the relation between RNR and Slash category approaches to extraction as developed in
work on Generalized Phrase Structure Grammar and Head-Driven Phrase Structure Grammar. Although RNR has at
least once been claimed to support Slash category conceptions (see Gazdar 1981,179-180), I argue in section 4.3 that
this is not the case. Although I defend Slash ideas against certain criticisms arising in transformational grammar work, I
argue that English RNR structures are actually incompatible with current formulations of Slash approaches. A key basis
for this claim is a phenomenon noted by Piera (1985), Levine (1985, 494-495), and Hukari and Levine (1989, 1991),
which I refer to as quasi exfiltration (Q-exfiltration). At issue is a kind of ''recursive" compounding of RNR represented
by examples like (51).
(51) Carlotta may believe t1 and Sandra certainly does believe t1 [that Mike t2]1 and Barbara can prove that Ted t2
[betrayed your confidence]2.
I argue that in general the range of such cases cannot be consistently described under the Slash regimes hitherto
proposed.

< previous page page_24 next page >


If you like this book, buy it!
< previous page page_25 next page >
Page 25

Chapter 2
Contrasting Extraction Types
2.1 Two Types of Left Extraction
Despite the unchallenged similarities between L-extractions discussed in chapter 1, in this chapter I seek to ground a
nontraditional division of English (NP) L-extractions. My aim is to argue that even the L-extractions illustrated in (1)
of chapter I do not form a homogeneous class. Two distinguishable types are represented in that list. Referred to
arbitrarily as types A and B, they are extensionally characterized roughly as in (1).
(1) Basic NP L-extraction types (subject to refinement)
a. A-extractions
b. B-extractions
question extraction
topicalization
restrictive relative extraction
nonrestrictive relative extraction
pseudoclefting
clefting
negative-NP extraction
comparative extraction
exclamatory extraction
So (1a-f) of chapter I illustrate A-extractions and (1g-i) B-extractions. After documenting multiple contrasts between
A- and B-extractions of NPs, I propose an account of this nontraditional distinction that appeals to partly traditional
elements. Then I argue that the A/B-extraction distinction is insufficiently delicate. A-extractions themselves divide
into two contrasting subtypes.

< previous page page_25 next page >


If you like this book, buy it!
< previous page page_26 next page >
Page 26
The basic theoretical idea to be developed is that whereas the gap positions in A-extractions do not in general manifest
properties distinct from those of the positions in which the extractees occur, the gap positions in B-extractions do. More
precisely, the gap positions of B-extractions manifest properties of contexts involving (weak) definite pronouns. This
suggests that B-extractions obligatorily involve (invisible) resumptive pronouns (RPs) in their extraction sites, whereas
A-extractions do not. If viable, this claim about B-extractions means that they are properly characterized in part in the
way that Perlmutter (1972) claimed that all L-extractions should be.
Since full documentation of the contrasts for all the extractions would be highly repetitive, I take negative-NP fronting
and/or question formation as the default instantiation of A-extractions and topicalization as the default for B-
extractions.
2.2 A-Extraction/B-Extraction Differences
2.2.1 Remarks
That A- and B-extractions differ in some fundamental way can be argued by showing that A-extractions are possible
from numerous environments that preclude B-extractions. I am aware of at least seventeen such con-texts, although
space considerations prevent full documentation. These contexts have a common property, which can provide a basis
for their exclusion of B-extractions if the A/B difference is analyzed in a certain way.
2.2.2 Existential there Constructions
First, A-extractions are compatible with the "focus" position of existential there constructions, whereas B-extractions
are not.
(2) a. He knew that there were (no) such chemicals in the bottle.
b. [No such chemicals]1 did he know that there were t1 in the bottle.
c.*[Such chemicals]1, he knew that there were t1 in the bottle.
The contrast between (2b) and (2c) is striking since negative-NP fronting is a formal, even stilted construction, whereas
topicalization is more colloquial. A priori, one would expect that the negative member of such pairs would be worse.
Other cases cited below show differences in the same direction. One cannot account for (2c) on the grounds that the
extracted NP violates general restrictions defining topicalizable NPs.

< previous page page_26 next page >


If you like this book, buy it!
< previous page page_27 next page >
Page 27
Generic indefinite NPs like that in (2c) are not inherently incompatible with topicalization, as (3) indicates.
(3) a. [Such chemicals]1, he said t1 had never been in the bottle.
b. [Such chemicals]1, he would never have placed h in the bottle.
Parallel remarks hold for all the cases considered.
2.2.3 Change-of-Color Contexts
A-extractions are compatible with positions of NPs designating changes of color, whereas B-extractions are not.
(4) a. He painted the car green/that color.
b. [What color]1 did he paint the car t1?
c. [No such color]1 would I ever paint my car t1.
d.*[Green/That color]1, he never painted the car t1.
Again, the NPs that cannot be topicalized in (4d) can be topicalized in general.
(5) [Green/That color]1, he never discussed t1 with me.
2.2.4 Name Positions
In a variety of positions, NPs refer to names and not to things named by those names. Whereas A-extractees can be
linked to such positions, B-extractees cannot.
(6) a. They named him Raphael/that/something ridiculous.
b. They called him fathead/something obscene.
c. [Nothing of the sort]1 did I ever name him/refer to him as t1.
d. What1 did they name him t1/refer to him as t1?
e.*Raphael1, I wouldn't name anybody t1.
f.*That1/Fathead1, you shouldn't call a person t1 in public.
As in the change-of-color case, the NPs that cannot be topicalized in this context can be in others.
(7) a. Raphael1, we never discussed h as a possible name for him.
b. Fathead1, no one seems to know who first used t1 as a pejorative.
2.2.5 Inalienable Possession Contexts
English has different ways of representing inalienable possession, shown in (8).

< previous page page_27 next page >


If you like this book, buy it!
< previous page page_28 next page >
Page 28
(8) a. They touched his ear.
b. They touched him on the/his ear.
Whereas A-extractions of body part NPs are possible for both varieties, (8a) and (8b), B-extractions are possible only
for the former.
(9) a. [What part of the/his body]1 did they touch (him on)
b. [His ear]1, they never touched (*him on) t1
2.2.6 Predicate Nominals
A fifth contrast involves predicate nominals (PNs). 1
(10) a. Frank is a bodyguard.
b. I regarded Frank as a bodyguard.
c. Frank became a bodyguard.
d. Frank turned into a bodyguard.
e. They made a bodyguard out of Frank.
f. Italians make good cannibal snacks. (= OK 'are' or OK 'manufacture')
The A/B-extraction contrast holds systematically. Examples (11a-f) show that all of (10a-f) permit A-extraction.
(11) a. [What kind of dancer]1, do you want to be t1?
b. [What kind of idiot]1 did they regard him as t1?
c. What1 are you going to become t1?
d. [what sort of thing]1 did he turn into t1?
e. [what sort of brain surgeon]1 can we make t1 out of Percy?
f. [What kind of cannibal snacks]1 do Italians make t1? (= OK 'are' or OK 'manufacture')
However, (12) illustrates that PNs are incompatible with topicalization.
(12) a.*[A good bodyguard]1, Frank is t1.
b.*[The best bodyguard in the world]1, I never referred to Frank as t1.
c.*[That kind of surgeon]1, Frank never became t1.
d.*[A vicious werewolf]1, Frank immediately turned into t1.
e.*[The country's worst drug dealer]1, they made t1 out of Frank.
f.*[Good cannibal snacks]1, Italians never make t1. (*unless = 'manufacture')
(13) is an apparent counterexample to this claimed incompatibility.
(13) [A good doctor]1, she isn't t1

< previous page page_28 next page >


If you like this book, buy it!
< previous page page_29 next page >
Page 29
If English NP topicalization is incompatible with PNs, (13) cannot be a topicalization of a PN. Notably, there is an
analysis for (13) distinct from PN topicalization, namely, the non-NP extraction in (14).
(14) [Fond of Mike]1, she isn't t1.
If (13) is like (14), it represents extraction not of a PN but of some larger predicational constituent containing a PN, just
as (14) involves extraction of a constituent containing (of) Mike.
Significantly, constructions (13) and (14) share restrictions. Both are blocked in simple positive cases such as (15).
(15) a.*[A good doctor]1, she is t1.
b.*[Fond of Mike]1, she is t1.
But both are well formed in certain contrastive positive contexts.
(16) a. They said she was a good doctor, and [a good doctor]1 she may be t1.
b. They said she was fond of Mike, and [fond of him]1 she may be t1.
Moreover, the two types manifest parallel restrictions on embedding.
(17) a.*I learned that [a good doctor]1, Joan wasn't t1.
b.*I learned that [fond of Mike]1, Joan wasn't1.
Most importantly, both constituent types can be the sort of extractees associated with though, whose associated L-
extraction is in general in-compatible with NPs.
(18) a. [Fond of Mike]1 though she was t1, ...
b. [A good doctor]1 though she was t1, ...
c.*[A good doctor]1 though she consulted t1, ...
d.*[A good doctor]1 though she planned to have Ted hire t1, ...
Given these similarities, it seems reasonable to conclude that (13) fails to counterexemplify the claim that PNs cannot
be topicalized, for the extracted constituent can be taken to be a non-NP containing a PN.
2.2.7 Adverbial NPs
A-extractions are compatible with a variety of "adverbial" NPs that B-extractions are not compatible with.
(19) a. Harry often talks that way.
b. [What way]1 does Harry talk t1?

< previous page page_29 next page >


If you like this book, buy it!
< previous page page_30 next page >
Page 30
c. the way that1 Harry talks t1
d.*[That way]1, Harry often talks t1.
e.*That way, which1 Harry talks t1, ...
(20) a. Harry resigned for that reason.
b. [What reason]1 did he resign for t1?
c. [No such reason]1 did he ever resign for t1.
d.*[That reason]1, he resigned for t1.
2.2.8 Extraposed Prepositional Phrases
A significant distinction between A- and B-extractions linked to prepositional phrase extraposition is illustrated in (21).
(21) a. They published a scurrilous review of his book last year.
b. They published a scurrilous review t1 last year [of his book]1.
c. [No such scurrilous review of his book]1 did they publish t1
d. [Such a scurrilous review of his book]1, they published t1 last year.
e. [No such scurrilous review t1]1 did they publish t1[ of his book]1.
f.*[Such a scurrilous review t1]2 they published h last year [of his book]1.
Contrasts like that between (21e) and (21f), based on parallel NP types, seem quite remarkable.
2.2.9 Infinitival Extraposition
A further contrast between A- and B-extractions, related to infinitival extraposition, is found in (22).
(22) a. I did not perceive a definite wish to retire in Sylvia.
b. I did not perceive a wish t1 in Sylvia [to retire]1.
c. [No wish to retire]1 did I perceive t1 in Sylvia.
d. [No wish t1]2 did I perceive t2 in Sylvia [to retire]1.
e. [A definite wish to retire]1, I did not perceive t1 in Sylvia.
f.*[A definite wish t1]2, I did not perceive t2 in Sylvia [to retire]1.
2.2.10 Exceptive Shifting
Another contrast between A- and B-extractions is that only the former are compatible with shifted expressions such as
except + NP and other than + NP.

< previous page page_30 next page >


If you like this book, buy it!
< previous page page_31 next page >
Page 31
(23) a.He handed something other than the gun to Rim.
b. He handed something t1 to Rita [other than the gun]1.
c.[What other than/except the gun]1 did he hand t1 to Rim?
d. [What t1]2 did he hand t2 to Rim [except/other than the gun]1 ?
e. [Nothing other than the gun]1 would he have handed t1 to Rita.
f. [Nothing t1]2, would he have handed t2 to Rita [other than the gun]1.
g. [Something other than the gun]1, he might have handed t1 to Rita.
h.*[Something (dangerous) t1]2, he might have handed t2 to Rita [other than the gun]1.
2.2.11 Temporal NPs
A sharp A/B-extraction difference is linked to postverbal temporal NPs with the verbs stay and spend. With the former,
the temporal NP can be extracted only by an A-extraction.
(24) a. Frank spent/stayed that much time in Ireland.
b. [How much time]1 did Frank spend/stay t1 in Ireland?
c. [That much time]1, Frank could never spend/*stay t1 in Ireland.
2.2.12 Idiomatic Verb + NP Structures
Several idiomatic verb + object structures are such that the NP is subject to A- but not B-extractions. One is the well-
known idiom make headway (see Emonds 1979, 233, citing Vergnaud 1974).
(25) a. They made a lot of/that much headway on the job.
b. [How much headway]1 did they make t1 on the job?
c. [Not much headway]1 did they make t1 on the job.
d. the headway that1 they made t1 on the job
e.*[That much headway]1, I am sure they made t1 on the job.
f.*[That much headway], which1 they made t1 on the job,...
2.2.13 A Locative Case
With some verbs, including born, an A/B-extraction contrast is linked to the type of NP heading a locative phrase. For
instance, when the NP designates a building, there is no contrast.
(26) a. Ed was born in that house.
b. [Which house]1 was Ed born in t1?

< previous page page_31 next page >


If you like this book, buy it!
< previous page page_32 next page >
Page 32
c. [No such house]1 was he born in t1.
d. [That house]1, our president is said to have been born in t1.
However, when the locative NP is a proper noun designating a country, a contrast appears.
(27) a. Ed was born in Argentina.
b. [What country]1 was Ed born in t1?
c. [No country like that]1 could he have been born in t1.
d.*[Argentina]1, our president is said to have been born in t1.
That is, in the country case, a locative prepositional phrase head can be extracted via an A- but not a B-extraction.
2.2.14 Summary
I have documented twelve English contexts that permit A- but not B-extractions, showing that some hitherto ignored
factors distinguish these two collections of structures. As far as I know, no contexts permit B- but not A-extractions.
The B-extraction environments therefore appear to be a proper subset of the A-extraction environments. However, this
claim must be formulated with some care since there are environments that permit B-extractions but not certain A-
extractions.
2.3 Antipronominal Contexts
The documentation of pervasive if hitherto overlooked A/B-extraction contrasts raises various questions. Most
fundamentally, what if any common factor accounts for the various B-extraction blockages and why is every A-
extraction environment a B-extraction environment but not conversely?
A commonality linking the contexts that prohibit B-extractions is revealed by reconsidering existential there structures.
Notably, the position precluding B-extractions is also incompatible with definite pronouns, yielding correlations like
(28).
(28) a. There are such apples on the table.
b.*There are them on the table.
c.*[Such apples]1, there are h on the table.
To facilitate discussion, I refer to a particular environment accepting NPs as an antipronominal context if and only if it
precludes the occurrence of anaphoric (i.e., linked to antecedents) definite pronominal NPs (e.g., he,

< previous page page_32 next page >


If you like this book, buy it!
< previous page page_33 next page >
Page 33
she, it, them). I will refine this notion as the discussion proceeds. The data in (28) show that the ban on B-extraction
with existential there can be characterized as in (29).
(29) The NP position that can be a gap for A- but not B-extractions is antipronominal.
Crucially, property (29) is not specific to the context in (28) but holds for every case previously documented where B-
extractions are banned.
Property (29) holds of change-of-color environments, where, for example, topicalization is blocked.
(30)*They painted their porch green1 but I refused to paint mine it1.
Recall that color phrases themselves can be topicalized from other environments such as the object position of discuss.
Significantly, (31) illustrates that that environment is not an antipronominal context.
(31) I wanted to talk about [(the color) green]1 but he refused to discuss it1.
Name positions are also antipronominal contexts.
(32) a.*He named his daughter Lucille1 but I didn't name mine it1.
b.*One shouldn't call people [fathead or idiot]1 but he calls students them1.
But contexts like (7), which permit B-extraction of name NPs, are not.
(33) a. We discussed it as a possible name for him.
b. As for "fathead"1, no one seems to know who first used it1 as a pejorative.
In neither the color nor the name case does the antipronominal condition reduce to a general ban against the structure
NP + pronoun, for the relevant restrictions are maintained in corresponding passives, as (34) indi-cates. Compare the
different restriction in (35a), which vanishes in the corresponding passive, as in (35b).
(34) a.*His car was painted green1 but mine will never be painted it1.
b.*His daughter was named Marsha1 but mine was not named it1.
(35) a.*They gave Ted it.
b. Ted was given it.
The inalienable position that resists B-extraction is antipronominal but the one that permits it is not. Therefore, the
contrast between the short

< previous page page_33 next page >


If you like this book, buy it!
< previous page page_34 next page >
Page 34
and long versions of (9b) correlates directly with that between (36a) and (36b), respectively.
(36) a. They wanted to touch [his arm]1 and they did touch it1.
b.*They wanted to touch him on [the/his arm]1 and they did touch him on it1.
PN positions are in general antipronominal.
(37) a.*Frank is it.
b.*I referred to Frank as it.
c.*Frank became it.
d.*Frank turned into it.
e.*They made it out of Frank.
f.*Italians make them. (* unless = 'manufacture')
These cases correlate precisely with the ill-formed B-extractions of (12).
The adverbial contexts are antipronominal, correlating with the impossibility of B-extractions in (19d) and (20d).
(38) a.*I talk [that way]1 but Harry rarely talks it1.
b.*I resigned for [those reasons]1 but Harry did not resign for them. t1
The environments associated with extraposed prepositional phrases, infinitives, and exceptive phrases, which are
incompatible with B-extractions, are all antipronominal.
(39) a.*They published it last year [of his best book].
b.*I did not perceive it in Sylvia [to retire].
c.*He might have handed it to Rita [other than the gun].
The parallel environments without the extrapositions, which permit B-extractions, are not antipronominal; the examples
formed from (39a-c) by suppressing each one's final constituent are grammatical.
The temporal-NP A/B-extraction contrasts distinguishing stay from spend correlate with the fact that the temporal
position of the former is antipronominal whereas that of the latter is not.
(40) a.*He is going to stay [(the next) two weeks]1 in Greece but I am going to stay them1 in Spain.
b. He is going to spend [(the next) two weeks]1 in Greece but I am going to spend them1 in Spain.
Idiomatic contexts that resist B-extraction are also antipronominal.

< previous page page_34 next page >


If you like this book, buy it!
< previous page page_35 next page >
Page 35
(41) a.*Herbert claimed to have made [(that) (much) headway]1 on the project but he never made it1.
b.*Herbert said he gave [a great deal of/some thought]1 to those problems but I know that he didn't give it1 to them.
Earlier I documented that from the environment [born in], A-extraction of an NP referring to a building is possible but
A-extraction of one referring to a country is not. Correspondingly, that environment is antipronominal with respect to
antecedents designating countries but not with respect to those designating buildings.
(42) They said Ed was born in [that house]1/Argentina2 but he wasn't born in it1/*it2.
Examples like (42) reveal that the notion ''antipronominal context" is less simple than so far implicitly suggested. Given
that pronouns enter into relations with antecedents, an adequate notion of antipronominal context ultimately has to be
relational. It must specify that a context is antipronominal with respect to certain antecedents of certain types. Later I
will point out further complications.
2.4 Why Are B-Extractions Incompatible with Antipronominal Contexts?
2.4.1 Remarks
I have so far tried to support the conclusions in (43).
(43) a. A large collection of English environments E can contain A- but not B-extraction sites.
b. Each member of E is an antipronominal context.
c. No attested environments allow B- but not A-extraction gaps.
Assuming that this state of affairs is correct, how can one account for it?
2.4.2 The Resumptive Pronoun Hypothesis
Conclusions (43a,b) suggest that B-extractions, unlike A-extractions, have pronominal properties. For nonrestrictive
relative extraction, this is unsurprising. Given semantic parallelisms like those in (44), the extracted constituent itself is
arguably a definite pronoun. 2
(44) a. Marianne1, who1 is very lovely, wants to ...
b. Marianne1, and she1 is very lovely, wants to ...

< previous page page_35 next page >


If you like this book, buy it!
< previous page page_36 next page >
Page 36
So special assumptions linking nonrestrictive relative extraction to pro-nominal features might be unnecessary. If so,
nonrestrictives would not bear much on current concerns. 3 However, an appeal to the pronominal character of the
extractee certainly has no general application to topicalization and probably none to clefting.
The pronominal properties of the latter B-extractions can be captured naturally by taking them to be obligatorily
associated with RPs in the extraction site. Evidently, the pronouns thereby posited are invisible, a feature I return to
below. In these terms, roughly, B-extraction gaps are banned from antipronominal contexts because those gaps actually
represent invisible pronouns. This proposal also offers a straightforward account of (43c), which has, however, only
limited importance. Namely, claim (43c) then follows from the fact that, other than the NP environments that take
expletives, resumptive reflexives, or bound pronouns such as those italicized in (45), there are no attested NP
environments that preclude nonpronominal NPs.
(45) a. He availed himself of he opportunity.
b. She blinked her eyes.
But if there were an environment Z that allowed anaphoric pronouns but no nonanaphoric NPs, Z could well permit B-
but not A-extraction gaps.
Accounting for (43a,b) by appealing to invisible pronouns is not original. The proposal's historical antecedents are too
complex to be discussed adequately here. Recall though that in early transformational studies, extractions were almost
uniformly analyzed as displacements of categories by movement transformations. Ross (1967) recognized two types of
displacement, called choppings and copying transformations. The former reordered constituents from gap positions
without leaving anything (other than the null element) in those positions; the latter left RPs, which seem to have always
been visible, surface pronouns. Nothing in Ross's framework corresponded exactly to the A/B-extraction distinction.
Modifying Ross's then standard transformational position, Perimutter (1972) proposed that all extractions left RPs
(called shadow pronouns) in extraction sites, an effect that Ross's assumptions had only allowed as a possibility.
Perlmutter's approach determined pronominal invisibility via a rule applicable only to some pronouns. This proposal is
arguably a key root of what in the work of Chomsky and those he influenced came to be called the trace theory of
movement rules. The A/B-extraction distinction falsities Perlmutter's overall claim, which was excessively general; it

< previous page page_36 next page >


If you like this book, buy it!
< previous page page_37 next page >
Page 37
implied in effect that all extractions are B-extractions. Nonetheless, Perimutter deserves much credit for apparently
being the first to suggest that invisible RPs play a role in extraction constructions. The A/B-extraction distinction
likewise shows that standard versions of Chomsky's trace theory are wrong in claiming in effect that all extractions are
A-extractions.
Beyond making substantive claims about language that amount really to the first recognition of B-extractions,
Perimutter (1972) applied a novel methodology to the domain of extractions. This was to seek specific properties l. of
definite pronouns in nonextraction structures and then to show that l manifest themselves in extraction structures.
Earlier I used exactly this strategy to show that B-extractions are blocked from anti-pronominal contexts.
A similar strategy was in effect adopted by Hans-Georg Obenauer in the early 1980s and was also used by Guglielmo
Cinque, neither of whom explicitly referred to Perlmutter's earlier proposals. Work positing invisible RPs in extraction
structures includes Obenauer 1984, 1985, 1986, 1992, Cinque 1990, and Koster 1987, 153-171. Cinque's now available
work (1990) proposes essentially another extraction typology, developed over many years of work in transformational
grammar in terms of Chomsky's GB assumptions.
Although adopting the overall GB position, Cinque makes fairly radical proposals internal to that framework.
Specifically relevant to present interests, he claims that a range of constructions treated in GB terms as involving
extractions do not represent what standard views would analyze in terms of movement and associated traces. Rather,
Cinque takes them to manifest what he calls invisible RPs, although in his termsunlike those of, say, Ross (1967) or
Perlmuttter (1972)such pronouns do not arise through movement. Cinque's ideas are also an important forebear of my
postulation of B-extractions. But not only did Cinque not recognize the pronominal character of what I have called B-
extractions, he explicitly denied it.
(46) "This requires a different treatment for the absence of weak crossover effects in topicalization and appositive
relatives, which I regard as true movement constructions leaving a real variable...." (Cinque 1990, 199, n. 59)
Cinque recognizes pronominal characteristics only in the empty operator constructions of the GB literature, those
involved in object raising (47a), object deletion (47b), P-gaps (47c), and so on, plus extractions from certain islands
discussed later.

< previous page page_37 next page >


If you like this book, buy it!
< previous page page_38 next page >
Page 38
(47) a. Marian1 is easy to please t1.
b. Marian1 is too busy for us to invite t1.
c. Marian1, he hired t1 without first investigating pg1.
Cinque is surely right to recognize pronominal aspects to these structures. Every environment supporting the claim that
B-extractions are excluded from antipronominal contexts supports a parallel claim for the constructions of (47). (48)
illustrates only with the inalienable possession cases.
(48) a. His arm1 was difficult to touch (*him on) t1.
b. His arm1 was too sore to touch (*him on) t1.
c. [Which part of his body]1 did they photograph t1 after touching (*him on) pg1?
2.4.3 Island Sensitivity
Postulating invisible RPs in (certain) B-extractions raises an obvious issue with respect to island constraints. A
fundamental principle stated by Ross (1967) is the Island Law (see (25) of chapter 3).
(49) The Island Law
Only choppings (extractions not involving RPs) are subject to island constraints.
Therefore, positing invisible RPs in B-extraction structures might appear to entail that these are insensitive to island
boundaries, contrary to well-known fact (see (4c) of chapter 1). Although this complex issue demands far more space
than is available here, recognizing invisible RPs in B-extractions need not be incompatible with the conjunction of the
Island Law with (4c) of chapter 1, for one can associate the fact that the RPs posited in B-extractions are invisible with
a relation of control between the binder and its RP. I view control minimally as a relation between an antecedent and a
pronominal that determines that the latter is not visible. It can be assumed that general principles of control require
controller and controllee to be in a local relation. This will force the RP itself to extract.
A B-extraction like English topicalization is therefore actually an amalgam of at least four elements.
(50) A B-extraction involves
a. extraction of the binder (extractee);
b. presence of an RP in the same role as the binder;
c. extraction of the RP, possibly to the same point as the binder; 4
d. control of the extracted RP by the binder.

< previous page page_38 next page >


If you like this book, buy it!
< previous page page_39 next page >
Page 39
In such terms, the island sensitivity of, say, topicalization is due not to the binder extraction, which is associated with
an RP in the extraction site, but to the RP extraction, which is not (i.e., which is an A-extraction). Expressed in Ross's
(1967) terms, English topicalization is island-sensitive because it involves chopping of an RP. However, French
topicalization (of direct objects) is not island-sensitive because the associated RP is not extracted.
(51) a. Marcel1, they arrested (*the only woman who ever loved) t1.
b. Marcel1, ils ont arrt (la seule femme qui /1 'a jamais aim) t1.
The presence of the unextracted RP represented by the italicized accusative clitic in (5lb) preserves the latter from the
type of island violation seen in the long form of (51a).
2.4.4 Wide versus Narrow Pronominalization Restrictions
So far I have made (among others) claim (52).
(52) An English B-extraction gap cannot appear in an antipronominal context.
This principle predicts the documented incompatibilities between B-extractions and the various antipronominal contexts
attested earlier. However, pronominal restrictions interact with extractions in a more complex way than is specified in
(52). To see this, observe (53).
(53) a.*Katie attends Yale1 but Amanda does not attend it1.
b.*Katie attends Yale1 but Amanda wouldn't even apply to it1.
Apparently, the objects of attend and apply to are antipronominal with respect to NPs designating educational
institutions. However, as (54) illustrates, B-extractions of object NPs with such designations are in general not blocked.
(54) a. Yale1, Katie would never apply to t1/attend t1.
b. It was Yale that1 Katie refused to apply to t1/attend t1.
One reaction to the failure of (54a,b) to manifest tile ill-formedness of (53a,b) would deny the existence of a systematic
connection between antipronominal contexts and extractions. Under that assumption, the data about attend and apply to
pose no problem. The drawback is that the correlations documented earlier reduce to nothing but accidents.
An alternative approach would maintain a systematic connection between antipronominal contexts and blockage of
(among other things)

< previous page page_39 next page >


If you like this book, buy it!
< previous page page_40 next page >
Page 40
B-extractions, by differentiating types of antipronominal contexts on the basis of the sorts of constraints that determine
them. The minimal viable categorization would differentiate wide pronominal bans from narrow pronominal bans. A
narrow ban precludes occurrence of a pronoun in a position R only if R is a surface context. That is, a narrow ban
blocks only visible pronouns in surface positions. Since the RPs I have posited in, say, English NP topicalizations are
invisible, (53) can be made compatible with (54) by taking the constraint relevant to the former to be a narrow ban.
Wide bans, on the contrary, block pronouns whether or not they are surface forms and thus apply to both visible and
invisible pronouns. All the antipronominal contexts defining environments in which B-extractions are blocked would
then be defined by wide bans.
Taking wide and narrow bans to define wide and narrow anti-pronominal contexts, respectively, one would replace (52)
by the slightly weaker (55).
(55) An English B-extraction gap cannot appear in a wide antipronominal context.
Although consistent with all the data, the second approach raises the issue of vacuity. Does it amount to anything more
than calling an anti-pronominal context that correlates with a B-extraction blockage a wide antipronominal context and
one that does not, a narrow one?
Although at the moment the answer is negative, this need not lead to (55)'s being vacuous. First, the majority of
antipronominal contexts with which I am familiar seem to be wide. Second, that the narrow/wide anti-pronominal
context distinction is not empty game-playing is shown by correlations. As suggested by (48), the antipronominal
contexts cited earlier are incompatible not only with B-extractions but also with object raising, object deletion, parasitic
gaps, and several other constructions. Notably, the narrow antipronominal context in (53) is, on the contrary,
compatible with all of these.
(56) a. Yale1 would be hard for me to attend t1.
b. Yale1 is too expensive for me to apply to t1.
c. [Which college]1 did she want to attend t1 without applying to pg1?
Thus, the factual content of calling an antipronominal context narrow is that every posited invisible pronoun
construction in the language will be insensitive to it. And specifying an antipronominal context as wide entails

< previous page page_40 next page >


If you like this book, buy it!
< previous page page_41 next page >
Page 41
that every invisible pronoun construction is incompatible with it. These considerations suggest that the mild weakening
involved in the narrow/ wide antipronominal context distinction by no means deprives of content claims that
antipronominal contexts and extraction blockages are correlated. If there really were no systematic connection, one
would expect a relatively random distribution of extraction acceptability with respect to antipronominal contexts. But
even in the presence of data like (53) and (54), this is not the observed state of affairs.
2.4.5 Two Types of Definite Pronoun
An apparent exception to my earlier claim that PNs form antipronominal contexts is the italicized environment of (57).
(57) It was her that they hired.
That is, despite (57), the post-be position behaves like a wide anti-pronominal context.
(58) a. Who1 was it t1 that they hired?
b. No one else1 could it have been t1 that they hired.
c.*Karen1, it was t1 that they hired.
d.*I like Karen, who1 it was t1 that they hired t1.
The data in (58) are not counterexamples to any earlier claims. Nonetheless, they seem anomalous in manifesting the
same pattern attributed to wide antipronominal contexts even though they apparently lack such contexts. Thus, they
represent currently unexplained data.
Although this conclusion is quite tentative, (57) suggests the need to further refine the concept "wide antipronominal
context." Suppose that an NP position qualifies as a wide antipronominal context if it excludes what I will refer to as
weak definite pronouns; and suppose the pronoun in (57) is a strong definite pronoun. Although this contrast remains
obscure, its existence is suggested by the properties of it, which I take to be weak, in contrast to those of that, which
(on one analysis anyway) is strong. Example (59) shows that the weak form it is impossible in the context of (57).
(59) It was*it/that that they believed/discussed/denied.
Assuming then that wide antipronominal contexts are defined by the exclusion of weak definite pronouns, the PN
position in (57)/(59) would still be a wide antipronominal context, and (58) would become perfectly regular in terms of
earlier assumptions.

< previous page page_41 next page >


If you like this book, buy it!
< previous page page_42 next page >
Page 42
The view just sketched can be interpreted as claiming something like (60).
(60) All RPs are weak definite pronouns.
That (60) is true at least for English can be supported by finding plausible candidates for visible English RPs in certain
contexts and showing that in those contexts the weak pronoun it does not alternate with the strong pronoun that. The
examples in (61) represent what I consider to be such cases.
(61) a. Right dislocation
I like it1/*that1 very much, [your idea]1.
b. Left dislocation
[Your proposal]1, the committee is considering it1/*that1 very seriously.
c. Copy raising
[That idea]1 seems/sounds like it1/*that1 will be very popular.
d. Substandard extraction from islands
He expressed the sort of idea which1 the secret police will arrest you if they hear about it1/*that1.
In each case the environment that excludes that as a putative RP accepts it as an ordinary NP.
Given the facts motivating a weak/strong distinction, it seems plausible that (60) is correct, that only exclusion of weak
definite pronouns defines wide antipronominal contexts, that (57) thus involves a wide anti-pronominal context, and
that the ungrammaticality of (58c,d) therefore follows from principle (55), which depends on the proposed RP analysis
of B-extractions.
2.5 Two Subtypes of A-Extraction
2.5.1 Basics
Recent years have seen extensive discussion of what I refer to as extraction from selective islands. For the phenomena
analyzed in Ross 1967, constituents seem to be either nonislands or absolute islands; they allow all types of constituents
in all functions to be extracted, or none. But since the late 1970s, certain constituents have been observed to be selective
with respect to permissible extractions. Earlier work reprinted in Kayne 1984 revealed what have come to be called
subject-object asymmetries, and

< previous page page_42 next page >


If you like this book, buy it!
< previous page page_43 next page >
Page 43
Koster (1978a) and Huang (1982) noticed that in the same contexts, adjunct extraction is blocked; see (62) (from Kayne
1984, 3) and (63).
(62) a. The only person who1 it's not essential she talk to t1 is Bill.
b.*The only person who1 it's not essential t1 talk to her is Bill.
(63) a.*How1 is it essential that we fix the car t1?
b.*the reason why1 it is essential that I talk to her t1
c.*[For whose sake]1 is it essential that Anthony resign t1?
The most commonly discussed selective islands are embedded interrogative clauses. However, there are many others.
Clear examples are the irrealis if clauses studied by Pullum (1987).
(64) a. [Which car]1 would you prefer it if I fixed t1?
b.*[How rapidly]1 would you prefer it if I fixed the car t1?
There is now a large, factually rich, and growing literature on selective islands including Kayne 1984, Obenauer 1984,
1985, 1986, Koster 1987, Cinque 1990, Rizzi 1990, Frampton 1991, Manzini 1992, and Lasnik and Saito 1992. 5 A
contribution to this literature of special importance from my point of view is made by Cinque (1990), who notices, at
least sporadically, that extractions from selective islands are banned, even when general principles would otherwise
allow them, if the extraction site is antipronominal. (Previously mentioned contrasts like that between stayand spend
derive from Cinque's observations with respect to their Italian equivalents.) Among other considerations, the general
observation leads Cinque to suggest that extractions from selective islands involve in-visible RPs, and no
movemententirely different mechanisms than are posited for standard extractions (like (1a-i) of chapter 1) within the
GB framework.
Developing Obenauer's and Cinque's observations, I suggest (among other things) the generalization in (65).
(65) Extraction from a selective island Pi is impossible if the extraction site in Pi is a wide antipronominal context.
Beyond Cinque's (1990) mostly Italian data, other facts in the literature support (65). Frampton (1991) notes (without
linking the observation to the antipronominal character of the context) that the NP "focused" in an existential there
construction cannot be extracted from an embedded question; see (66) (from Frampton 1991, 41).

< previous page page_43 next page >


If you like this book, buy it!
< previous page page_44 next page >
Page 44
(66) a. [How many bags]1 do you wonder whether I think t1 are on the table?
b.*[How many bags]1 do you wonder whether I think there are t1 on the table?
Of course, the position marked t1 in (66b) was shown to be a wide antipronominal context. The irrealis type of
selective island is illustrated in (67).
(67)*[How many books]1 would you prefer it if (he believed) there were t1 on the table?
Obenauer notes that change-of-color phrases cannot be extracted from interrogative selective islands, as illustrated in
(68) (from Obenauer 1985, 199, n. 42).
(68) a. ?[This boat]1, I was wondering whether to paint t1 pea green.
b.*[Pea green]1, I was wondering whether to paint this boat t1.
Rizzi observes that idiomatic NPs cannot be extracted from selective islands (without linking this limitation to
antipronominal contexts); see (69) (from Rizzi 1990, 79).
(69) a. [What headway]1 do you think you can make t1 on this project?
b.*[What headway]1 do you wonder how to make t1 on this project?
(70) and (71) show the point for various other cited antipronominal contexts.
(70) a.*What1 did he ask you whether I nicknamed my cat t1?
b.*What1 did he ask you whether I referred to him as t1?
c. [What part of the body]1 did they ask you whether he touched (*her on) t1?
(71)*[How strong a wish t1]2, did they ask her whether she perceived t2 in Sylvia [to retire quite soon]1?
Note that a sentence identical to (71) without the infinitival extraposition is grammatical.
Principle (65) can provide a further argument justifying the distinction drawn earlier between weak and strong definite
pronouns. On the basis of that distinction, but not without it, principle (65) correctly predicts that pronoun-accepting
contexts like that of (57) are not suitable contexts for extraction from selective islands. This explains contrasts like
those in (72).

< previous page page_44 next page >


If you like this book, buy it!
< previous page page_45 next page >
Page 45
(72) a. I don't know who1 (*they asked him whether) it was t1 that he saw.
b. What1 (*they would have preferred it if) it was t1 that he bought was a Toyota.
2.5.2 Why A1-Extractions Resemble B-Extractions with Respect to Selective Islands
Supposing now that (65) is true, what is the explanation? I would propose, contrary to Cinque (1990), that the same
mechanisms involved in ordinary extractions like those in (1) of chapter I function for selective islands. This view is
based on the initially puzzling principle in (73).
(73) Selective islands are absolute islands, that is, just islands.
Given (73), the Island Law determines that a constituent can extract from a selective island only if an RP is present. As
with B-extractions, the in-visibility of the RP and the sensitivity of the construction to (certain) islands internal to the
selective island can be accounted for by claiming that the RPs are extracted and controlled by their binders. This view
accounts directly for principle (65).
This approach to selective islands has a specific consequence for A-extractions. Earlier sections might have suggested
that, whereas a B-extraction obligatorily determines an RP in its extraction site, an A-extraction obligatorily fails to do
so. But such a view is now untenable: a typical successful extraction from a selective island (e.g., (66a)) involves an A-
extraction and, given (73) and the Island Law, an RP in the extrac-tion site.
One might then conclude that A-extractions differ from B-extractions in not requiring but allowing a RP. However, the
real situation is at least as differentiated as shown in (74).
(74) L-extraction types
a. Those that require RPs in their extraction sites; these are the B-extractions discussed earlier
b. All others; these are the A-extractions discussed to this point
c. A-extraction subtypes
i. A1-extractions, which allow RPs in their extraction sites
ii. A2-extractions, which forbid RPs in their extraction sites
The argument for the A1/A2 distinction is this. If (65) is a correct principle about selective islands, then without the
A1/A2 distinction, all

< previous page page_45 next page >


If you like this book, buy it!
< previous page page_46 next page >
Page 46
A-extractions (at least of NPs) from the permissible extraction sites of selective islands would be allowed. Past work
seems to have assumed that this is the case. But it is not. The NP gaps of certain A-extractions cannot appear internal to
selective islands. So, contrasting with cases analyzed as A1-extractions because they do extract from selective islands,
there is comparative extraction, which cannot.
(75) a. What1 did they ask you whether you had seen t1?
b. the things which1 they asked me whether I had seen t1
c. What: they asked me whether I had seen t1 was a blue Mercedes-Benz.
d. [No such car]1 did they ever ask whether I had seen t1.
e.*They are looking for more cars than (what1) they asked me whether I had seen t1.
Consider also (76).
(76) a .What1 would they prefer it if you had bought t1?
b.the things which1 they would have preferred it if I had bought t1
c.What1 they would have preferred it if I had bought t1 was a blue Mercedes-Benz.
d. [No such car]1 would they have preferred it if I had bought t1.
e.*They agreed to look for more cars than (what1) I would prefer it if you bought t1.
Another A2-extraction is free relative extraction, whose gaps also seem incompatible with selective islands.
(77) a. the pilots Who1 we asked them whether you had contacted t1
b. [Which pilots]1 did we ask them whether you had contacted t1?
c. [whatever pilots]1 we (*asked them whether you had) contacted t1
(78) a. the pilots Who1 we would prefer it if she contacted t1
b. [Which pilots]1 would you prefer it if she contacted t1?
c.*[whatever pilots]1 we would prefer it if she contacted t1
The ''more the merrier" construction discussed by Ross (1967) is also an A2-extraction.
(79) a. The more stars he dates the more stars (that1) he claims I should date t1.
b.*The more stars he dates the more stars (that1) he would prefer it if I dated t1.

< previous page page_46 next page >


If you like this book, buy it!
< previous page page_47 next page >
Page 47
c.*The more stars he dates the more stars (that1) he asks me whether you have dated t1.
Yet another is, I believe, the extraction linked to no matter + wh.
(80) a. [No matter]1 what they believe that she did t1,...
b.*[No matter what]1 they asked you whether she did1,...
c.*[No matter what]1 they would prefer it if she did t1,...
Thus, the following a priori curious situation exists. Neither A1- nor A2-extractions are in general sensitive to the
distribution of weak pro-nouns, and their gaps therefore can appear in wide antipronominal con-texts. B-extraction sites
cannot appear in wide antipronominal contexts. But selective islands group B-extractions with A1-extractions. In
contrast to A2-extractions, both can extract from selective islands.
(81) a. [That car]1, they never asked me whether I had seen t1.
b. [That car]1, they would have preferred it if I had bought t1
The conceptualization suggested here for the A1/A2-extraction contrast accounts for this otherwise anomalous
situation. Given characterization (74), A1- and B-extractions are alike but distinct from A2-extractions in that both are
compatible with RPs in their extraction sites. This explanatory success argues that the basic idea of these remarksthat
(only certain) L-extractions are linked with invisible RPsis correct.
2.5.3 Further Diversity
Even an analysis of extractions along the lines of (74) is insufficiently ramified. It cannot be presumed possible to
specify for each overall ex-traction construction (e.g., questioning) that it behaves uniformly with respect to RPs in all
environments. So far, for example, I have taken question extraction to fall into category A1. But consider (82a-b) (the
latter from Heim 1988, 32).
(82) a. What1 is there t1 on the table?
b.*[Whose drink]1 is there t1 on the table?
Heim notes that constituents with preposed genitives cannot be question extractees from the "focus" position of an
existential there construction, a wide antipronominal context. The blockage in (82b), which contrasts with simple
question extraction from the relevant context as in (82a), would follow from already specified principles if the
informally stated principle in (83) held.

< previous page page_47 next page >


If you like this book, buy it!
< previous page page_48 next page >
Page 48
(83) Extractees with a preposed genitive wh form obligatorily link to RPs in their extraction sites.
This states in other terms that in the genitive case, question extraction is a B-extraction, not an A1-extraction.
Moreover, principle (83) generalizes from question structures and the particular there context. Thus, it entails as well
the contrast in (84).
(84) a. the (only) people who1 there were t1 at the party
b.*the (only) people [whose parents]1 there were t1 at the party
It also predicts the contrast in (85) for the change-of-color antipronominal context.
(85) a. the color which1 we painted the cabin t1
b.*the doctor [whose favorite color]1 we painted the cabin t1
Initially, though, the predictions of (83) seem wrong for PN contexts like (86).
(86) a. [Whose bodyguard]1 did he become t1?
b. [Whose favorite comedian]1 did she turn into t1?
However, like the initially troublesome topicalization of (13), such cases can possibly be analyzed as involving
extraction of a larger constituent containing a PN rather than the PN itself. Thus, they would parallel (87), for example.
(87) [How fond of Mike]1 was Barbara t1?
Such a proposal is possible only because predicational phrases permit question extraction. They do not, however, permit
most forms of relativization. Therefore, the proposal predicts that relative clause analogs of (86) should be
unacceptable; (88) shows that they are.
(88) a.*the rock star [whose bodyguard]1 Ted became t1
b.*the doctor [whose favorite comedian]1 she turned into t1
If anything like (83) is correct, the conditions linking extraction types to the presence of RPs must be more complicated
than specifications of the form "Question extraction allows an RP." This conclusion is strengthened by the existence of
question cases that are neither standard A1-extractions nor B-extractions like those involving preposed genitives.
Rather, certain question structures seem to behave like A2-extractions. Examples (89a-c) present a now expected
paradigm.

< previous page page_48 next page >


If you like this book, buy it!
< previous page page_49 next page >
Page 49
(89) a. [Which steak]1 would you prefer it if I ate t1?
b. [That steak]1, I would prefer it if you ate t1 last.
c.*I ate more steaks than (what1) they would prefer it if you ate t1.
That is, (89a) shows typical A1-extraction behavior from a selective island, (89b) typical B-extraction behavior, and
(89c) typical A2-extraction behavior. Next consider the contrasts in (90)-(93), where each involves extraction from
some selective island.
(90) a. [What size steak]1 would you prefer it if we ordered t1?
b.*[How big a steak]1 would you prefer it if we ordered t1?
(91) a. [What size steak]1 will they arrest anyone who eats t1?
b.*[How big a steak]1 will they arrest anyone who eats t1?
(92) a. [What size steak]1 did he reject the idea of eating t1?
b.*[How big a steak]1 did he reject the idea of eating t1?
(93) a. [What size steak]1 did it shock people that he ate t1?
b.*[How big a steak]1 did it shock people that he ate t1?
NP question extraction involving how + adjective seems to be subject to much the same restrictions as a corresponding
how + non-NP extraction: neither permits the gap to be separated from the extractee by a selective island boundary.
(94) a.*[How big]1 he would prefer it if the steak was t1 is unknown.
b.*[How big]1 they will arrest everyone who says the steak was t1 is unknown.
c.*[How big]1 he rejected the idea that the steak was t1 is unknown.
This will follow in current terms if one assumes that, like adjective phrases, NPs with initial how + adjective cannot
antecede RPs. But that amounts to saying that NP question extraction, normally of type A1, probably of type B with
preposed genitives, is of type A2 in cases like the (b) questions of (90)-(93). If so, then the assignment of question
extractions to types depends on the internal structure of the questioned NP, all three assignments allowed for in (74)
being required for some English question extractees.

< previous page page_49 next page >


If you like this book, buy it!
< previous page page_i next page >
Page i

Three Investigations of Extraction

< previous page page_i next page >


If you like this book, buy it!
< previous page page_51 next page >
Page 51

Chapter 3
The Status of the Coordinate Structure Constraint
3.1 Background
The term Coordinate Structure Constraint (CSC), introduced into grammatical theorizing by Ross (1967), designates
the proposed natural language universal formulated as (1).
(1) "In a coordinate structure, no conjunct may be moved, nor may any element contained in a conjunct be moved out
of that conjunct." (Ross 1967, 98-99)
As formulated in (1), the CSC is an informal restriction on the application of transformational rules, specifically,
movement rules. 1 Logically, this, formulation has two distinct parts (Grosu 1973); see section 3.3. Hereafter, I regard
the part barring the extraction of conjuncts themselves as a separate constraint, the Conjunct Constraint.
From here on, then, CSC designates only the constraint against extracting proper subconstituents of conjuncts. In
another terminology also introduced by Ross, the CSC states that coordinate conjuncts are islands, as illustrated in (2).
(2) a. [Which surgeon]1 did Sally date friends of t1 (*and a lawyer)?
b. [Which surgeon]1 did Sally date friends of t1 (*and hope to date Bob)?
c. [Which surgeon]1 did they say that Sally dated friends of t1 (*and Claude believed that Gwen was jealous)?
Notably, the CSC generalizes across category types and thus predicts the ill-formedness of (e.g.) NP conjuncts, as in
(2a), VP conjuncts, as in (2b), and S conjuncts, as in (2c).

< previous page page_51 next page >


If you like this book, buy it!
< previous page page_52 next page >
Page 52
Ross (1967) went on to observe that the CSC does not bar extraction from coordinates in which one extractee
corresponds to a gap in each conjunct, so-called across-the-board (ATB) extraction.
(3) "There is an important class of rules to which [(l)] does not apply. These are rule schemata which move a
constituent out of all the conjuncts of a coordinate structure." (Ross 1967, 107)
So Ross noted that, despite (1), examples like (4) are grammatical.
(4) [Which car]1 did Sally buy t1, Marilyn borrow t1, and Lucille wreck t1?
Restricting (1) via (3) can be interpreted as claiming that a coordinate conjunct is an island unless it and each of its co-
conjuncts contain extraction sites, all linked to the same extractee. Given that, genuine coun-terexamples to the CSC
must have at least the form of single extractions from fewer than all sister conjuncts. 2
The CSC can be taken to refer to certain sentential properties rather than, as in (l), properties of operations of
transformations. Such a view is necessary for those who believe, as I do, that the CSC expresses an important universal
truth about natural languages while denying the relevance of transformations (or even all generative devices; see
Langendoen and Postal 1984) to correct natural language grammars. Modulo that, however, my impression is that the
CSC is widely regarded as the most problem-free syntactic constraint ever discovered.3 As Gazdar (1982, 175) states,
"Numerous island constraints other than the CSC have been proposed in recent years. Unfortunately, few if any of them
are as resilient to counterexamples as the CSC is." Besides the fact that in natural language after natural language, the
CSC seems to have essentially the status Ross (1967) proposed, there are other reasons for attributing to it an unusual
degree of confirmation. For example, Georgopoulos (1983, 1984a,b, 1985a,b) notes that no other standard island
constraints seem to constrain extractions in the Western Austronesian language Belauan (spoken in the Caroline
Islands);4 but the CSC essentially limits relevant Belauan structures just as Ross's principle specifies (see Georgopoulos
1983, 137; 1984a, 80; 1985a, 87-88).
Against this background, it would be remarkable to find the CSC refuted, especially on the basis of facts no farther
afield than English. But just such a putative refutation is advanced by Lakoff (1986), who cites

< previous page page_52 next page >


If you like this book, buy it!
< previous page page_53 next page >
Page 53
three classes of English data as counterexemplifying the CSC. He refers to them as the A, B, and C scenarios (hereafter,
I abbreviate scenario by S). They are respectively illustrated in (5).
(5) a. the stuff which1 Arthur sneaked in and stole t1
b. [How many dogs]1 can a person have t1 and still stay sane?
c. That is the drug which1 athletes take t1 and become quite strong.
As Lakoff notes, the first involves examples of a type already cited by Ross (1967, 103-104); the second involves data
discovered by Goldsmith (1985) and also taken by him to involve incompatibility with the CSC; 5 and the third is
represented by observations attributed to Peter Farley. Although for simplicity I use Lakoff's terminology for the
structures of (5), his categories might seem to treat the relevant VP structures as indivisible wholes. But it would
probably be more accurate to pick out particular conjuncts as having crucial properties. This permits seeing key aspects
of the differences in (5) and those between any of (5) and ordinary symmetrical logical coordination as relational. All
the cases in (5) manifest a semantic relation between their conjuncts that, among other things, differentiates them from
the most typical structures subject to the CSC, logical coordination. To avoid question-begging, I informally refer to all
such constructions as linear.
In logical coordination, conjunct order is semantically irrelevant. But in other types of linear structures, order does play
a role, in that distinct conjuncts are linked by asymmetrical relations. The Ss in (5) are asymmetrical linear structures,
but not the only ones. Consider (6).
(6) a. Frank criticized de Gaulle and hence criticized a Frenchman.
b. Oedipus hugged Jocasta and consequently hugged his own mother.
These examples instantiate a linear structure, call it a D-S, in which one conjunct stands in some kind of logical
consequence relation to preceding conjuncts.6 Lakoff (1986) does not cite D-Ss, which do not permit analogs of the
sort of extractions he claims refute the CSC.
(7) a.*It was de Gaulle who1 Frank criticized t1 and hence criticized a Frenchman.
b.*It was a Frenchman who1 Frank criticized de Gaulle and hence criticized t1.
But D-Ss illustrate the sense in which a characterization of linear structures as Ss hinges on the relational features of
particular conjuncts.

< previous page page_53 next page >


If you like this book, buy it!
< previous page page_54 next page >
Page 54
A-Ss manifest a relation between a conjunct (e.g., the first in (5a)) and the one that follows it, in which, minimally, the
first must characterize an event/state of affairs temporally preceding that characterized by the second. Of course, this
hardly begins to capture what is essential to these structures, which is somewhat obscure (for relevant discussion, see
Deane 1991; Na and Huck 1992; Lakoff 1986). What appears to define a B-S is that one conjunct represents a state of
affairs that might be regarded as "unexpected" given the state of affairs represented by the preceding conjunct. As
Goldsmith (1985, 133) puts it, "the and here ... could be paraphrased as 'and nonetheless'." Finally, in a C-S the state of
affairs represented by one conjunct causes the state of affairs represented by the following conjunct.
Consistent with their relational character, the conjuncts of a single S need not be homogenous with respect to S types.
Lakoff (1986, 153) cites (8) as a mixed A/B-S.
(8) [How many courses]1 can you take t1 for credit, still remain sane, and get all As in t1?
The relation between the second and first VP conjuncts is of the B-S type; that between the second and third is of the
A-S type. It is also possible to mix the A-S and C-S types (see (9a)), perhaps the B-S and C-S types (see (9b)), and
perhaps all three types (see (9c)).
(9) a. the antidote which1 Frank went to the store, bought t1, drank t1, and survived the poisoning
b. ?the antidote which1 Frank drank t1, still died, and interrupted the wedding
c. ?the antidote which1 Frank went to the store, bought t1, drank t1, still died, and interrupted the wedding
Since Ss seem to be characterized by specific relations between conjuncts of special types, it is useful to have names
for the defining conjunct types. I will speak of A-conjuncts, B-conjuncts, C-conjuncts, and D-conjuncts. In (5a) the first
conjunct is an A-conjunct; in (5b) the second conjunct is a B-conjunct; in (5c) the second conjunct is a C-conjunct; and
in (6a) the second conjunct is a D-conjunct. In (9a) the first conjunct is an A-conjunct and the fourth a B-conjunct. The
status of the second and third conjuncts is less clear. Later I will show that such constituents differ from the clear initial
A-conjunct in syntactic ways. In (9b) the second

< previous page page_54 next page >


If you like this book, buy it!
< previous page page_55 next page >
Page 55
conjunct is a B-conjunct and the third a C-conjunct. In (9c) the first is an A-conjunct, the fourth a B-conjunct, and the
fifth a C-conjunct. 7 The status of the second and third conjuncts is as in (9a).
From the data regarding A/B/C-Ss, Lakoff reaches the following conclusion:
(10) It should be clear that the coordinate structure constraint simply does not exist as a purely syntactic phenomenon.
It is just not true that if you extract an element from one conjunct, you must extract it from all of them. So far as VP
conjunctions are concerned, extraction is permitted from any or all conjuncts, so far as syntax alone is concerned."
(Lakoff 1986, 156)
Certain later discussions apparently accept this view (see Deane 1991, 13; Na and Huck 1992). However, my goal here
is to show that conclusion (10) is erroneous. Despite the apparently severe problems they raise for the CSC, I argue that
none of the data Lakoff cites refute it; nor do any other English data that I known of.8
Lakoff's (1986) A/B/C-Ss do not all fail to counterexemplify the CSC for the same reasons. I argue that B/C-Ss are
consistent with the CSC because they are not properly analyzed as coordinate structures in the sense referenced by the
CSC. The reason why A-Ss are consistent with the CSC is more involved.
The conclusion that A-Ss counterexemplify the CSC is, I suggest, based on two types of misstep. First, on a theoretical
level, Lakoff does not properly determine what the CSC claims. In effect, he takes the CSC to be specified entirely by
chapter 4 of Ross 1967 (i.e., by my (1) + (3)), ignoring that in chapter 6 Ross imposes a further general subrestriction on
all island constraints. Second, on a factual level, Lakoff overlooks a range of properties and restrictions characterizing
the A-S extraction structures he cites as counterexamples to the CSC. These oversights are of more general
significance, as they parallel oversights regarding different structures characteristic of much other work (see appendix
A). Once these factors are uncovered, it can be argued that none of the A-S data cited by Lakoff are in fact
counterexamples to the CSC as properly understood because all A-S conjuncts are islands. An important aspect of my
claims about A-Ss is that they are consistent with the CSC whether they are analyzed as truly coordinate or not. This
seemingly paradoxical claim is clarified in section 3.2.

< previous page page_55 next page >


If you like this book, buy it!
< previous page page_56 next page >
Page 56
3.2 A-Scenarios
3.2.1 The Logic of the Case
Apparent A-S counterexamples to the CSC are typified by (11).
(11) the cheese which1 Frank drove to the store, bought t1, went home, and gave t1 to Greta
Lakoff appears to reason from such data as follows: 9
(12) Logic of Lakoff's argument from A-Ss against the CSC
a. Various pieces of evidence indicate that the four VPs in (11) are coordinated.
b. Nonetheless, these constituents do not instantiate full ATB extraction since the first and third conjuncts contain no
extraction sites. Hence, Ross's exclusionary stipulation (3) is not relevant.
c. Therefore, via the CSC, all the conjuncts in (11) are islands.
d. The second and fourth (VP) conjuncts of (11) contain extraction gaps bound by an extractee external to those
conjuncts.
e. But the basic property of an island I is that it cannot contain an extraction gap linked to an extractee external to I.
f. Since the conjunction of steps (c) and (d) contradicts step (e), the CSC is false.
It would be natural to attempt to avoid conclusion (12f) by rejecting the view that the apparently coordinate structures
of A-Ss are really coordinate. That is, one could quarrel with assumption (12a). This was in effect the position that
Lakoff took in 1967 in the face of then sporadic observations of A-Ss and that Ross (1967, 103-104) adopted. It is also
essentially the view of Pauline Jacobson, whose informal suggestion Lakoff (1986) criticizes in an appendix (see also
Na and Huck 1992). If premise (12a) could properly be rejected, A-Ss would trivially satisfy the CSC in just the way
every other noncoordinate structure does.
Is there a basis independent of apparent CSC violations for denying that the elements of A-Ss are really coordinate?
Consider what evidence suggests to the contrary that these constructions manifest genuinely coordinate constituents.
First is the presence of and, which functions as a coordinating conjunction in uncontroversial coordinate cases.
Moreover, A-Ss can contain but, which is a coordinating conjunction in other cases.
(13) The cheese which1 Frank went to the store but didn't buy t1 later spoiled.

< previous page page_56 next page >


If you like this book, buy it!
< previous page page_57 next page >
Page 57
Second, the ''recursive" possibilities of cases like (11) seem to be the same as those of unquestioned coordinations: the
number of conjuncts is expandable without limit, yielding apparently "fiat" bracketings.
(14) the cheese which1 Harry went to the store, took out his wallet, grabbed a five dollar bill, bought t1, went home,
took a shower, and then ate t1
For Lakoff (1986, 152), the discovery of the unlimited iterability of A-Ss seems to have been decisive for his
conclusion that they are coordinate and hence that the attested extractions from (certain) A-S conjuncts
counterexemplify the CSC.
Third, the and in A-Ss is subject to the same principles governing its nonappearance on some conjuncts in
unquestioned coordinate forms. For instance, as in (11), and can but need not be absent from all but the last conjunct;
but if absent from the kth, it must also be absent from the k - 1th.
(15) a. the book which1 she went to the library, (and) defaced t1, (and) tried to steal t1, and had to pay for t1
b.*the book which1 she went to the library, and defaced t1, tried to steal t1, and had to pay for t1
Fourth, Lakoff (1986, 153) states that "[t]he very existence of across-the-board extraction in such cases shows that true
conjunction is required, ..." But the logic of this aspect of Lakoff's argument, a claim that the mere presence of (partial)
ATB extraction in cases like (11) indicates coordinate status, is neither straightforward nor well spelled out. 10
Nonetheless, there do appear to be some initial grounds for taking A-Ss to represent true coordination, partially
summed up by Deane (1991, 13) with respect to cases like (14) as follows: "The crucial evidence lies in sentences like
..., which involve multiple conjuncts, across-the-board extraction from several of the conjuncts, and comma intonation
before the conjunction each of which is, as Lakoff notes, a direct indication of coordinate structure."
However, there are also arguments to the contrary. First, whereas the standard coordinate conjunction and can generally
alternate with the disjunctive form or, that does not hold for A-Ss.
(16)*the cheese which1 Frank went to the store, bought t1, went home, or gave t1 to Greta

< previous page page_57 next page >


If you like this book, buy it!
< previous page page_58 next page >
Page 58
Second, Ross (1967, 103-104) cites several properties of A-Ss by which they differ from unchallenged coordinate
cases, properties whose recognition he attributes to Lakoff.
(17) a. The main verb of the second conjunct must be nonstative.
b. The second conjunct cannot be negative.
c. There are tense restrictions.
As evidence for (17b), Ross offers (18).
(18)*The shirts which I went to the movies and didn't pick up will cost us a lot of money.
But (18) yields a good sentence, I believe, if still is inserted before the auxiliary, and it certainly yields one if and is
replaced by but.
Third, despite the cross-categorial generality of the CSC illustrated in (2), as Lakoff (1986) observes, the A-S
extractions putatively refuting the CSC are possible only with VP conjunction, and not with (e.g.) S conjunction. For
example, compare (11) and (19).
(19)*the cheese which1 Frank went to the store, his wife bought t1, they went home, and we gave t1 to Greta
Moreover, a limitation to VP conjunction is not nearly delicate enough. A further restriction partially paralleling that
observed by Goldsmith (1985) for B-Ss (see the discussion of (104)) seems to require that the VPs involved be
minimal. It precludes VPs with modals or other nonfinite complement-taking verbs.
(20) a. the book which1 Gail will drive there and (*will) borrow t1
b. [What model Cadillac]1 do you think that Ernest might come down and (*might/*should) drive t1?
c. [What model Cadillac]1 could Ernest want to come down and (*want/expect to) drive t1?
Fourth, the quantifier both, possible with true coordinate (VP) structures, is impossible in A-Ss. 11
(21) a. Jacqueline (both) went to Sears and bought a new dress.
b. the dress which1 Jacqueline (*both) went to Sears and bought t1
Although (21a) is grammatical with both, that version is not interpretable as an A-S, correlating with the impossibility
of extraction in (21b) in the presence of the quantifier.

< previous page page_58 next page >


If you like this book, buy it!
< previous page page_59 next page >
Page 59
Fifth, A-S conjuncts evidently have semantic properties distinct from those of conjuncts in ordinary conjunction cases
(in addition to Lakoff 1986, see Schmerling 1975; Na and Huck 1992; Deane 1991). 12 Specifically, if conjunct C1
precedes conjunct C2, then the event depicted by C1 predates that depicted by C2. This need not hold of true
coordinations.
(22) Chris (either) will go to the mall tomorrow or bought the cassette yesterday.
Sixth, unlike true coordinations, A-Ss do not have interwoven dependency correspondents; see note 2.
(23) the wine and beer which1 Jack and Bob will go to the store and buy t1 (* respectively)
Without respectively, (23) can be interpreted as an A-S. But with that form, it must be interpreted as a true
coordination, which bars extraction.
Given all these factors, it is unclear whether A-Ss involve coordination in the sense referred to by the CSC. But for my
limited purposes here, it doesn't matter at all. My defense of the CSC does not appeal to rejecting the coordinate
character of A-Ss. I argue instead that even if A-S constituents are coordinate in the relevant sense, they fail to yield
genuine counterexamples to the CSC. That is so because argument (12) contains a theoretical flaw independent of step
(12a). This flaw is linked to step (12e), which has evidently been interpreted to mean (24).
(24) Extraction from any island (hence from a coordinate island) is absolutely banned.
Although (24) may look innocuous, it is both inconsistent with the system developed in Ross 1967 and arguably false.
What Ross actually said was:
(25) "Chopping rules are subject to the constraints ... copying rules are not." (Ross 1967, 257)
For Ross, a chopping rule was an extraction associated either with nothing or with the null element in the extraction
site, whereas a copying rule was one associated with an RP in the extraction site. If the core of view (25) is correct, it
cannot be concluded that nothing can extract from an island.13 a Rather, the strongest inference possible is that if a
constituent to extracts from an island, then an RP occurs in js extraction site. Given that examples like (11) do not
involve visible RPs, this weaker conclusion

< previous page page_59 next page >


If you like this book, buy it!
< previous page page_60 next page >
Page 60
would of course reduce in the A-S case to the stronger one implicitly assumed by Lakoff, if (26) were a true principle
of natural language.
(26) All (resumptive) pronouns are visible (i.e., surface pronouns).
But there is overwhelming evidence against (26). 14
Lakoff's claimed refutation of the CSC from A-Ss amounts to instantiating coordinated constituents that are not islands.
The logic is thus that in, for example, (11), if the constituent bought t1 is taken to be coordinate, the CSC claims that it
is an island. Were the CSC true, such a constituent could not then exist. Therefore, the putative refutation concludes,
the CSC is false. The flaw is that no demonstration has been provided that the relevant (VP) constituents are not
islands. Lakoff's discussion implicitly takes it as self-evident that a constituent permitting any grammatical extraction is
a nonisland, in effect ignoring Ross's proviso (25), which claims that an extraction from an island is ill formed only if it
fails to link to an RP in the extraction site.
In the absence of a principle requiring that RPs be visible surface forms, a genuine refutation of the CSC from A-Ss
would have to argue that an appeal to invisible RPs in the relevant cases fails. Lakoff (1986) attempted no such
demonstration; nor, to my knowledge, has any other author. Thus, at best, his (1986) argument against the CSC based
on A-Ss is unfinished, having a hidden, unjustified premise. Moreover, good evidence can be found that extractions
from A-S conjuncts do depend on the presence of nonphonetic RPs. The argument from A-Ss is then not completable,
be cause one of its key premises, (12e), is untenable. A-Ss remain consistent with the CSC, since, despite the well-
formedness of examples like (11), their conjuncts can viably be considered islands.
3.2.2 Selective Islands
There is now a rich literature on what Culicover (1990) called strange extractions, what I elsewhere (Postal
1990c)called weird extractions, and what many in the GB tradition have called extraction from weak islands or long
movement. In chapter 2 and in Postal and Baltin 1994, it is referred to as extraction from selective islands. Involved are
constituents that seem to behave with respect to extractions partly like nonislands (because they permit some
extractions) and partly like islands (because many constituent types extractable from "normal" nonislands are not
extractable from these). Space considerations preclude giving this topic the attention

< previous page page_60 next page >


If you like this book, buy it!
< previous page page_61 next page >
Page 61
it merits (for recent discussions in the GB framework, see Koster 1987, esp. 153-159; Cinque 1990; Rizzi 1990;
Frampton 1991; Lasnik and Saito 1992; Manzini 1992). A radically different perspective was of course presented in
chapter 2.
Suffice it to say that various constructions permit what appear to be ordinary extractions, but only subject to conditions
not otherwise generally limiting extractions. These conditions properly include the following:
(27) Conditions on extraction from selective islands
If a constituent C external to a selective island V is extracted from site S internal to V, then
a. Setting aside certain cases of prepositional phrase (PP) extraction, 15 C is an NP.16
b. S is not a finite subject position.17
c. S is not an antipronominal context (AC).
d. C is not reflexive.18
e. C is not inherently unpassivizable.19
As explicated in chapter 2, an AC is, roughly, a position that accepts (some) NPs but not weak definite pronouns.
The great variety of English selective islands includes the types in (28).
(28) a. Interrogative clauses
b. Irrealis if complements (see Pullum 1987)
c. Complements of factive predicates like regret
d. Rationale clauses
e. Clausal complements of (certain) nouns
In general, NP objects extract from these constituents but (subject to the remarks about PPs in note 15) nothing else
canhence, not finite subjects, adverbials, or predicafional phrases. The constraints in (27a,b) lead to contrasts such as
those illustrated for types (28a-e) in (29)-(33), respectively.
(29) Interrogative clauses
a. the person who1 I asked them whether (they believed) Carla tickled t1
b.*the person who1 I asked them whether (they believed) t1 tickled Melissa
c. the person who1 I asked them whether Carla played with t1
d.*the person [with whom]1 I asked them whether Carla played t1

< previous page page_61 next page >


If you like this book, buy it!
< previous page page_62 next page >
Page 62
(30) Irrealis if complements
a. the person who1 I would prefer it if you hired t1
b.*the nurse who1 I would prefer it if t1 hired you
c. the way [in which]1 (*I would prefer it if) you behaved t1
d.*Where1 would you prefer it if we lived t1?
(31) Complements of factives
a. the person who1 I regret (that) Carla tickled t1
b.*the person who1 I regret t1 tickled Melissa
c. the person who1 I regret that Carla played with t1
d. the person [with whom]1 (*I regret that) Carla played t1
e.*the reason why1 I regret that he resigned t1
(32) Rationale clauses
a. [Which scientist]1 did Ed go to England (in order) to consult with t1?
b.*[With which scientist]1 did Ed go to England (in order) to consult t1?
c.*How1 did Ed go to England (in order) to treat Louise t1?
(33) Clausal complements of nouns
a. [Which doctor]1 did he formulate a plan to prove Joan had betrayed t1?
b.*[Which doctor]1 did he formulate a plan to prove t1 had betrayed Joan?
c.*How1 did he formulate a plan to prove Sarah treated the baby t1?
Condition (27c) takes on substance given the English ACs discussed in chapter 2, some of which are illustrated in (34),
as well as the ones listed in (35), only the first of which was previously discussed (in chapter 1).
(34) a. There are gorillas/* them in the meadow.
b. He dyed his beard green/that color/*it.
c. Mary was born in Bosnia/*in it. (where it designates a country)
d. He gave those proposals/them that much thought/*it.
e. She became a lawyer/*it/*her last year.
f. They canceled many flights/*them last year to Cuba.
(35) a. The object of the verb tell (='determine')
b. The object of manner-of-speaking verbs like grunt and whine
c. The language-designating object of the preposition in after speak

< previous page page_62 next page >


If you like this book, buy it!
< previous page page_63 next page >
Page 63
(35a-c) are respectively illustrated by the contrasts in (36a-c).
(36) a. They said they could determine/tell how he did it, and they could determine/*tell it.
b. If you know the name, then say/*grunt it.
c. He said the delegates were speaking in Thai and they were speaking in that language/*it.
If (27c) is correct, none of these ACs should be possible sites for extraction from selective islands. The general point
was already partly documented in chapter 2 and is documented further here, for the ACs of both chapter 2 and (35), by
utilizing in (37) and (38) the selective islands illustrated respectively in (29) and (31).
(37) a.*[What sort of gorilla]1 did he ask you whether there was t1 in the meadow?
b.*[What color]1 did he ask you whether Melissa dyed her hair t1?
c.*[What country]1 did he ask you whether she was born in t1?
d.*[How much thought]1 did he ask you whether you gave those proposals t1 ?
e.*What1 did he ask you whether she became t1?
f.*[How many flights]1 did he ask you whether they canceled t1 last year to Cuba?
g.*What1 they asked me whether I could tell t1 was whether he was a vampire.
h.*That's what1 they asked me whether he grunted t1.
i.*He knows the language that1 they asked me whether we were speaking in t1.
(38) a.*[What sort of gorilla]1 did he regret (that) there was t1 in the meadow?
b.*[What color]1 did he regret that Melissa dyed her hair t1?
c.*[What country]1 did he regret that she was born in t1 ?
d.*[How much thought]1 did he regret that you gave those proposals t1 ?
e.*What1 do you most regret that she became t1?
f.*[How many flights]1 did he regret that they canceled t1 to Cuba?
g.*What1 do you most regret that you couldn't tell t1?
h.*What1 they really regret that he grunted t1 was "To hell with the president."
i.*That is the language which1 they probably regret they were speaking in t1.

< previous page page_63 next page >


If you like this book, buy it!
< previous page page_64 next page >
Page 64
Continuing to focus on the relevance of the constraints in (27) to extraction from selective islands, let us turn to
condition (27d). This accounts for facts like those in (39).
(39) a. Himself1, (*they asked me when) I talked to Harry1 about t1.
b. It is herself1 (*who1 I would prefer it if) you talked to Jane1 about t1.
c. Themselves1, (*I regret that) you talked to the applicants1 about t1.
d. Herself1, I never (*went there (in order) to) talk to Jane1 about t1.
e. Herself1, they (*formulated a plan to) talk to Louise1 about t1.
Condition (27e), linking extraction from selective islands to the passivizability of the extracted NP, is designed to
account for the fact that restrictions like those in (40) and (41) manifest themselves (among other things) as constraints
on selective island extraction.
(40) a. Abigail felt the rocks move.
b.*The rocks were felt move by Abigail.
(41) a. The victims matter to the doctor.
b.*The victims were mattered to by the doctor.
In each case the unpassivizable NPs are extractable in ordinary environments.
(42) a. What1 did they say that she felt t1 move?
b. Who1 did she say the victims mattered to t1?
But, as (43) and (44) illustrate for each of the five selective island types in (28), the passivization constraints show up
in selective island extraction.
(43) a.*the rocks which1 they asked me whether Abigail felt t1 move
b.*the rocks which1 I would prefer it if you felt t1 move
c.*the rocks which1 I regret that she felt t1 move
d.*the rocks which1 she went to Bhutan in order to feel t1 move
e.*the rocks which1 she formulated a plan to feel t1 move
(44) a.*the doctor who1 they asked me whether Abigail mattered to t1
b.*the doctor who1 they would prefer it if Abigail mattered to t1
c.*the doctor who1 I regret that Abigail matters to t1
d.*the doctor who1 Abigail had a nose job in order to matter to t1
e.*the doctor who1 they formulated a plan to make Abigail matter to t1

< previous page page_64 next page >


If you like this book, buy it!
< previous page page_65 next page >
Page 65
Facts like those in (37) and (38) raise a question:
(45) Why is extraction from selective islands sensitive to ACs?
Following the basics of the analysis in chapter 2, the essential answer is as follows:
(46) a. Selective islands are islands.
b. Therefore, given Ross's principles, extraction from a selective island is possible only when an RP is present in the
extraction site. 20
c. Since RPs are (weak) pronouns, it follows that, whether visible or not, they should in general obey constraints on
visible (weak) pronouns and hence be banned from ACs.21
In these terms, extraction contrasts between strict and selective islands depend not on differences in islandhood per se
but, I suggest, on the principles determining that the RPs posited in (46b) are invisible. I propose that these principles
are subsumed under the more general control phenomenon. Thus, an island (e.g., an English parenthetical phrase) that
is not a selective island fails to be such because it precludes the kind of control (of RPs) characteristic of selective
island extraction, as touched on in chapter 1.22 Of course, some English-particular condition must guarantee that the
posited RPs can only be invisible, a condition that can be formulated as requiring that they be controlled; see note 8 of
chapter 1. This need not concern us here.
Although space considerations prevent discussion, a control view of the invisibility of the posited RPs in selective
island extraction offers a reason why such extraction is subject to conditions (27a,b,d,e) as well. Namely, these can be
taken as conditions on a certain type of (nonsubject) control characteristic not only of selective islands but also of
(among other things) object raising, object control, and P-gap structures. This is the type of control referred to as
EXCB in chapter 1.
3.2.3 A-Scenario Conjuncts as Selective Islands
Let us return to the central point, extractions from A-Ss. The portion of Lakoff's (1986) critique of the CSC based on
A-Ss fails because it ignores the key fact in (47).
(47) Those conjuncts of A-Ss [that permit extractions] are selective islands (hence, as claimed in (46a), islands).23

< previous page page_65 next page >


If you like this book, buy it!
< previous page page_66 next page >
Page 66
The bracketed clause in (47) is motivated by the fact implicit in Lakoff's (1986, 152) remark that in a binary A-S,
extraction is possible from the second conjuncthence, by implication, not from the first 24 (see also Na and Huck 1992,
260).25 For example, compare (48a) and (48b) (from Deane 1991, 24).
(48) a. ?*[Which store]1 did he go to t1 and buy groceries?
b.*What1 did he pick up t1 and call me?
More generally, extraction is possible only from non-A-conjuncts.
(49) a. Harry went to the store, bought something, went home, and ate it.
b. the stuff which1 Harry went to the store, bought t1, went home, and ate t1
c.*the store which1 Harry went to t1, bought stuff, went home, ate it, and returned to t1 for more
d.*the store which1 Harry bought stuff, ate it, and returned to t1 for more
However, although certain A-S conjuncts permit extractions, those are subject to conditions (27a-e) and hence are
selective island extractions. Specifically, only NPs and PPs (subject to the remarks of note 15) can be extracted from
A-S conjuncts; thus, it is not surprising that the extractee in every A-S extraction cited by Lakoff (1986) is an NP.
Differences between NP extractions from A-Ss (satisfying principles (27b,c,d,e)) and others are revealed by the NP and
non-NP extractions in (50)-(52).26
(50) a. [Which student]1 did Nora (go to the drugstore, come home and), talk to t1 for an hour?
b. [To which student]1 did Nora (*go to the drugstore, come home, and) talk t1 for an hour?
c. [How long]1 did Nora (*go there, come home, and) talk to that student t1?
(51) [Very fat]1 though Nora (*went to Italy, ate a lot, and) became t1,...
(52) [Interviewed by Myra]1, Edgar (*went home, got drunk, and) refused to (allow himself to) be t1.
Condition (27b) also governs extractions from A-Ss. (I owe (53b) to Peter Culicover (personal communication, 4
December 1990).)
(53) a. Who1 did you (go right up to the stage and) prove Mary had betrayed t1?

< previous page page_66 next page >


If you like this book, buy it!
< previous page page_67 next page >
Page 67
b. Who1 did you (*go right up to the stage and) prove t1 had betrayed Bill?
(54) a. the immigrants who1 the district attorney (jumped up,) waved his arms, and) claimed she defrauded t1
b.*the immigrants who1 the district attorney (jumped up, waved his arms, and) claimed t1 defrauded her
Most strikingly, A-S extraction is subject to condition (27c). Extraction sites in A-Ss cannot be any of the ACs
illustrated in (34) or (36); see (ic) of note 23.
(55) a. [How many gorillas]1 did he (*run in and) claim there were t1 in the meadow?
b. [What color]1 did she (*fly to Vancouver and) dye her hair t1?
c. [What country]1 did Mike (*go bar hopping, and) claim he was born in h1?
d. [How much thought]1 did they (*get drunk, drive home, and) give those proposals t1?
e. [What type of vampire]1 did Chris (*move to Transylvania, buy a castle, and) turn into t1?
f. [How many flights]1 did they (*get drunk, drive home, and) cancel t1 to Cuba?
g. What1 Mike could (*use a microscope and) easily tell t1 was that the specimen was dead.
h. [What sort of thing]1 did the lunatic (*rush into the lounge and) grunt t1?
i. [What language]1 did the delegates (*dash outside and) speak in t1?
Condition (27d) also limits A-S extraction.
(56) a. Himself1, Jane (*went home and) talked to Ed1 about t1.
b. Himself1, Jane (*rushed in and) claimed (that) Mary had talked to Harry1 about t1.
c. Herself1, I never (*went to her bedside and) described Ida1 to t1.
Finally, the passivizability restrictions that limit extraction from other selective islands (27e) also constrain extraction
from A-Ss.
(57) a. What1 did she (*climb up there and) feel t1 move?
b. Who1 did Greta (*rush home and) claim to matter very much to t1?

< previous page page_67 next page >


If you like this book, buy it!
< previous page page_68 next page >
Page 68
Further evidence that the conjuncts of A-Ss are selective islands, independent of conditions (27a-e), derives from
observations developed in chapter 2. There I claimed that English NP extractions divide into the types in (58) with
respect to their compatibility with RPs.
(58) Categories of English NP extractions
a. A1-extractions permit RPs in their extraction sites.
b. A2-extractions preclude RPs in their extraction sites.
c. B-extractions require RPs in their extraction sites.
Assignments of extractions to the categories in (58) include the following:
(59) a. A1-extractions: question extraction, restrictive relative extraction
b. A2-extractions: comparative extraction, free relative extraction
c. B-extractions: NP topicalization, NP clefting
An implication of categorization (58) is that A1- and A2-extractions are both compatible with ACs, whereas B-
extractions are not. A further consequence, under my assumption that extraction from selective islands depends on RPs,
is that such extractions can be of type A1 or type B but never of type A2.
In chapter 2 I argued that these consequences are in general correct. Here I extend documentation to the selective island
types of (30) and (32), showing that questioning (A1-extraction) and topicalization (B-extraction) are possible from
such islands, but not comparative formation or free relative (A2-extraction).
(60) a. Who1 would you prefer it if they hired t1?
b. Melvin1, I would prefer it if they didn't hire t1.
c.*He could end up hiring more waiters than (what1) I would prefer it if he hired t1.
d.*He could interview [whatever waiters]1 you would prefer it if he hired t1.
(61) a. Who1 did they go to Ottawa in order to hire t1?
b. Melvin1, I would go to Ottawa in order to hire t1.
c.*He could end up contacting more waiters than (what1) I could go to Ottawa in order to hire t1.
d.*He could interview [whatever waiters]1 you would go to Ottawa in order to hire t1.
By the criteria just enunciated, those conjuncts of A-Ss that permit extraction again qualify as selective islands, for they
allow A1- and B-extractions but not A2-extractions.

< previous page page_68 next page >


If you like this book, buy it!
< previous page page_69 next page >
Page 69
(62) a. [Which neighbors]1 did Frank go to the store, buy a water pistol, go home, and squirt t1?
b. It was those people who1 Frank went to the store, bought a water pistol, returned home, and squirted t1.
c. Ed interviewed more people than (what1) Frank (*went to the store, bought a water pistol, returned home, and)
squirted t1.
d. Ed counseled [whatever people]1 Frank (*went to the store, bought a water pistol, returned home, and) squirted t1.
(63) a. [Which suspects]1 did the police break in and arrest t1?
b. [Those suspects]1, the police broke in and arrested t1.
c. Miles interrogated [whatever suspects]1 the police (*broke in and) arrested t1.
d. Miles interrogated fewer suspects than (what1) the police (*broke in and) arrested t1.
A different type of evidence indicates that ASs are selective islands rather than nonislands. This depends on the fact
that other selective islands behave like islands in blocking the object-raising, object deletion, and P-gap constructions. I
document this point with the selective island types illustrated in (31) and (32).
(64) a. Jane1 is hard to believe people think/*regret that you have dated t1.
b.*Jane1 is hard to get people to go to England in order to consult t1.
(65) a. Jane1 is too rich for people to believe/*regret that you dated t1.
b.*Jane1 is too reclusive to get people to go to England in order to consult t1.
(66) a. It was Jane who1 he hired h despite believing/*regretting that you had dated pg1.
b.*It was Jane who1 he fired t1 after calling Louise in order to locate Pg1.
And A-Ss also block these constructions.
(67) a. Jane1 will be hard to (*go to England and) contact t1.
b. Jane1 is too reclusive for us to (*jump in the car, drive 300 miles, and) visit t1.
c. It was Jane who1 he hired t1 after (*jumping in the car, driving 300 miles, and) visiting pg1.

< previous page page_69 next page >


If you like this book, buy it!
< previous page page_70 next page >
Page 70
In blocking the object-raising, object deletion, and P-gap constructions, A-Ss behave exactly like other selective
islands.
3.2.4 Nonextraction Evidence
Several types of evidence independent of extraction phenomena support the conclusion that those conjuncts of Lakoff's
A-Ss that permit extraction are selective islands and hence, via principle (46a), islands. First, as initially noted by Ross
(1967, 272), the distribution of certain negative- ''licensed" "affective" forms including any and ever is partially
controlled by islands (see (68)). Certain indefinites like say a word and eat a bite are similarly restricted (see
Schmerling 1970 and (69)).
(68) a. Nobody bought a camera because he planned to spy on someone/*anyone.
b. Ted did not buy the camera so that you could spy on someone/*anyone.
c. Ted did not (*enter the restaurant because he planned to) eat a bite.
(69) a. Nobody believed (*the claim) that Isabelle had ever studied Turkish.
b. Nobody believed that Isabelle was married to the guy who had (*ever) studied Turkish.
c. Nobody believed (*the claim) that Isabelle would say a word.
In the ungrammatical cases, the negative "licenser" of any, ever, or the special indefinites is separated from them by an
island boundary. By this criterion, the non-A-conjuncts of A-Ss are islands.
(70) a. Nobody believed that Frank (went home and) drank some beer.
b. Nobody believed that Frank (*went home and) drank any beer.
c. Nobody believed that Frank ran out, bought some flowers, and (*ever) gave them to Lois.
d. Nobody believed that Frank would (*go back and) say a word.
A second type of independent support for the island character of A-Ss also derives from work by Ross (see Ross 1971
as well as Hooper and Thompson 1973 and Authier 1992). Ross observes that topicalization is impossible in many
embedded contexts in English (see (71) (from Ross 1971) as well as Ross 1967, 255).
(71) a. Nostradamus predicted that [these golfers]1, Spiro would only wound t1.

< previous page page_70 next page >


If you like this book, buy it!
< previous page page_71 next page >
Page 71
b.*That [these golfers]1, Spiro would only wound t1 was predicted by Nostradamus.
c. It was predicted by Nostradamus that [these golfers]1, Spiro would only wound t1.
Ungrammaticality results in (71) in the single case where the complement is an island, namely, when it is preverbal.
Other examples of islands precluding topicalization include (72a-c). 27
(72) a. I am sure (*of the fact) that snails1, he won't eat t1.
b.*That snails1, he won't eat t1 was determined by Linda.
c.*Linda is annoyed because snails1, he won't eat t1.
Suppose then that (73) is true.
(73) In English, topicalization cannot occur inside an island28
By this criterion, selective islands are islands, as (46a) claims. This is illustrated for the type in (28b) by (74), for the
type in (28c) by (75) (from Authier 1992, 334), and for the type in (28f) by (76).
(74) a. I would prefer it if you contacted Clarence directly.
b.*I would prefer it if Clarence1, you contacted t1 directly.
(75) a. John regretted that we went to see Gone with the Wind.
b.*John regretted that [Gone with the Wind]1, we went to see t1.
(76) a. Jean made the proposal that we try to fire Glen.
b.*Jean made the proposal that Glen1, we try to fire t1.
Principle (73) then supports the claim that A-Ss are islands. If they are, cases like (77)-(78) follow without further
stipulation.
(77) a. Harry sat down, took the oath, and swore that he never slandered Linda.
b. Harry (*sat down, took the oath, and) swore that Linda1, he had never slandered t1.
(78) a. Harry got drunk and admitted that he had repeatedly cheated your mother-in-law.
b. Harry (*got drunk and) admitted that [your mother-in-law]1, he had repeatedly cheated t1.
A third type of support independent of selective island extraction for the claim that A-Ss are islands depends on
negative polarity items such as

< previous page page_71 next page >


If you like this book, buy it!
< previous page page_72 next page >
Page 72
lift a finger (see Baker 1970). Such elements must cooccur with a negative form, subject to the following condition:
(79) The negative element "licensing" a negative polarity item r cannot be separated from r by an island boundary.
Support for principle (79) is found in (80), where the negative "licensers" are italicized.
(80) a. Nobody believed (*the claim) that she would lift a finger.
b.*Nobody hates people who lift a finger.
c. It would not be legal to lift a finger.
d.*To lift a finger would not be legal.
e.*Jose was not arrested because he lifted a finger.
f.*It does not amuse me, that you lifted a finger.
On this basis also, A-Ss qualify as islands since they cannot contain negative polarity items linked to external
"triggering" negatives.
(81) a. Nobody believed that Frank would (*go to Idaho, get a job, and) lift a finger.
b. Estelle did not say that Frank would (*climb onto the stage and) lift a finger.
c. I guess that he can't (*come back and) lift a finger.
In this respect A-Ss do not differ from uncontroversial coordinate structures.
(82) Nobody believed that Frank would support Joan (*and lift a finger).
A fourth nonextraction argument for the island status of A-Ss depends on generalization (83).
(83) In multiple interrogatives, an in-situ wh form cannot be separated from a co-questioned wh form by an island
boundary.
This claim would be supported by the ungrammaticality of the starred forms of (84), where the intervening island
boundaries are marked by . ((84e) is taken from Aoun, Hornstein, and Sportiche 1980, 78.)
(84) a. Who believes (* the claim) that Bill saw who(m)?
b.*Who likes books that criticize who(m)?
c.*Who was arrested because Jane described who(m)?
d.*Who succeeded in that business despite who(m)?
e.*Who knows children who study what?

< previous page page_72 next page >


If you like this book, buy it!
< previous page page_73 next page >
Page 73
The stars in (84) represent my own judgments. Such data can hardly be used as the basis of a general argument,
however, since the literature contains many conflicting reports. Principle (83) is incompatible with Chomsky's (1986b,
153) statement that, for example, the long variant of (84a) is grammatical and with Koster's (1987, 212) claim that (85)
is well formed.
(85) Who left despite whom?
It is likewise incompatible with the judgments of Lasnik and Saito (1992, 12-13) in (86).
(86) a. Who wonders whether John saw what?
b. Who read a report that John bought what?
c. Who went to class after he read which book?
d. Who said that friends of who hit Bill?
I would tend to reject all of these. Similarly, an unvarnished version of (83) is incompatible with the claims of Fiengo
et al. (1988, 81) that (84b) and (87) are well formed.
(87) a. Who got jealous because I spoke to who?
b. Who met students with what color hair?
Again I would reject these. Fiengo et al. (1988, 81) further cite (88) with only a question mark.
(88) ?Who saw John and who?
I reject this totally, in accord with parallel judgments by Bresnan ((89a), from 1975, (37)) and Pesetsky ((89b), from
1982, 618). 29
(89) a.*Who played checkers and what?
b.*Which article proves your theorem and defends which theory?
Why these judgment conflicts arise is rather mysterious, and it is interesting to explore the principles at work. I believe
Koster (1987, sec. 4.6) makes very important observations. The key point is that cases like those just cited, even when
taken to be grammatical, span island boundaries in ways that rather strongly resemble the ways that extractions from
selective islands do. Particularly, as Koster observes, all of the examples of this sort involve in-situ wh forms that are
NPs. Moreover, I believe there is an important correlation between the cases where multiple wh forms span island
boundaries and the class of selective island boundaries. It is an

< previous page page_73 next page >


If you like this book, buy it!
< previous page page_74 next page >
Page 74
important task to develop the connection between selective island extraction and multiple-wh cases that appear to
violate island conditions. That is beyond the scope of this discussion. However, the troubling data just presented can be
rendered irrelevant to a broader argument in defense of the CSC by restricting principle (83) to cases where the in-situ
forms are not NPs. This yields (90).
(90) In multiple interrogatives, an in-situ non-NP wh form cannot be separated from a co-questioned wh form by an
island boundary.
Supporting (90) is unfortunately difficult because in-situ non-NPs are often not very acceptable even in simple
structures. However, I agree with Koster's (1987, 227) informants that the following questions are grammatical:
(91) a. Who worked how long?
b. Who complained (about it) how often?
Given that, principle (90) is supported by the fact that separation of such wh adverbials from co-questioned wh forms
leads to sharp unacceptability.
(92) a. Who believes (* the claim) that Bill played how long?
b.*Who likes concerts that last how long?
c.*Who was arrested because Jane complained how often?
d.*Who succeeded in that business despite vacationing how often?
e.*Who wonders whether John swam how long?
f.*Who read a report that John lied how often?
g.*Who went to class after he rested how long?
h.*Who said that running how long exhausted Bill?
I take these data to strongly support principle (90), making it available as a diagnostic for the analysis of A-Ss.
Notably, then, by the standard of (90) A-Ss qualify as islands, because multiple interrogatives with one question form
outside an A-S and an in-situ non-NP inside (a non-A-conjunct) are clearly blocked.
(93) a. Who claimed that Jane (*rushed home and) rested how long?
b. Who said Mike used to (*dash out, find a taxi, and) rush to the track how often?
c. Who (*drove to the pool, dove in, and) stayed under how long?
With respect to multiple interrogatives, A-Ss resemble uncontroversial coordinate structures.

< previous page page_74 next page >


If you like this book, buy it!
< previous page page_75 next page >
Page 75
(94) a. Who claimed that Jane (*called Greg and) stayed on the phone how long?
b. Who learned that Anthony stayed under how long (*and almost drowned)?
A final consideration supports the view that English A-Ss are selective islands, hence islands, and that the grammatical
extractions from them depend on RPs. This indirect evidence derives from French, which has sentences analogous to
A-Ss, illustrated in (95).
(95) a. Jacques a couru au march, a achet du pain1, a fonc chez lui, et l1 'a mang.
'Jacques ran to the market, bought some breads, rushed home, and ate it1.'
b. Jacqueline compte flaner en ville, trouver une bague1, 1a1 voler, fuir, et la1 vendre Marcel.
'Jacqueline intends to stroll around in town, find a ring1, steal it1, flee, and sell it1 to Marcel.'
Unlike the situation in English, extraction from any French A-S conjunct is entirely impossible.
(96) a.*le pain que1 Jacques a couru au march, (a) achet t1, (a) fonc chez lui, et (a) mang t1
'the bread which1 Jacques ran to the market, bought t1, rushed home, and ate t1'
b.*Quelle bague1 est-ce que Jacqueline compte flaner en ville, trouver t1, voler t1, fuir, et vendre t1 Marcel?
'Which ring1 does Jacqueline intend to stroll around in town, find t1, steal t1, flee, and sell t1 to Marcel?'
One can reason from these facts as follows. French A-Ss do not differ from their English correspondents with respect to
islandhoodthey are islands. They differ in not being selective islands, which is to say that French A-Ss block the kind
of control required for selective island extraction. So, where English countenances extraction from A-Ss of only those
elements (very roughly, object NPs) that can in general extract from selective islands, French allows no extractions at
all.
Accounting for English/French extraction differences from A-Ss in terms of differences in the distribution of RP
control is quite plausible. Independently of A-Ss, the two languages arguably differ systematically along this
dimension, with English allowing such control in a broader

< previous page page_75 next page >


If you like this book, buy it!
< previous page page_76 next page >
Page 76
range of contexts. Consider for example rationale adjuncts and without adjuncts, which both permit (selective island)
extraction. 30
(97) a. She went to England without reading the book.
b. the book which1 she went to England without reading t1
c. She flew there in order to confront the director.
d. the director who1 she flew there in order to confront t1
Although French has fairly close analogs of both (97a) and (97c), extractions paralleling (97b) and (97d) are sharply
ungrammatical.
(98) a. Elle est alle en Angleterre sans lire le livre.
b.*le livre qu'1 elle est alle en Angleterre sans lire t1
c. Elle y est alle en avion pour confronter le directeur.
d.*le directeur qu'1 elle y est alle en avion pour confronter t1
However, (97b,d) involve selective island extractions; see (32), (39d), (43d), (44d), and the following examples, which
indicate that for both contexts, only NP extraction is permitted and even then not from ACs.
(99) a.*[How well]1 did she move to Greece without learning Greek t1?
b.*[What color]1 did she move to Greece after dyeing her hair t1?
Thus, if one says (as is optimal) that for the constructions at issue, islandhood is assigned identically in English and
French, the contrastive extraction facts for A-Ss follow from an independently existing difference between these
languages. French has a much narrower range of control contexts permitting invisible RPs and hence the sort of
"deceptive" extraction characteristic of selective islands than English does. Lakoff's approach, which takes (NP)
extraction from English A-Ss to indicate that these are not islands (and hence that the CSC is false), leaves the contrasts
with French A-Ss unrelated to the independently occurring contrasts between extractions from adjuncts.
3.2.5 A-Scenarios: Summary
In this section I have argued that the structures Lakoff (1986) called A-Ss are consistent with the CSC, regardless of
whether they are properly analyzed as manifesting true coordination. This approach is feasible because even those A-S
conjuncts that permit some extractions are selective islands; thus, extraction from them depends on RPs, in accord with
Ross's (1967) original insight. As a consequence, A-Ss manifest the sensitivity to the possible occurrence of weak
pronouns characteristic of selective islands in general, that is, an extraction site in an A-S cannot be an AC.

< previous page page_76 next page >


If you like this book, buy it!
< previous page page_77 next page >
Page 77
Further, extraction from A-Ss is subject to other conditions on selective island extraction, which bar the extraction of
finite subjects, most non-NPs, reflexive forms, and inherently unpassivizable NPs. Also, as with other selective islands,
A2-extractions (see (58) and (59)) are impossible. Additional evidence from "affective" elements, topicalization,
negative polarity items, and multiple interrogations confirms the island status of A-Ss. Finally, the impossibility of
extracting from A-Ss in French combines with the narrower possibilities of extracting from selective islands in that
language to further support the islandhood of English A-Ss.
Thus, Lakoff's conclusion that A-Ss counterexemplify the CSC blends theoretical and factual mistakes: not recognizing
the role of RPs in permitting extraction from islands, and not recognizing the selective island character of the apparent
nonislands at issue, A-Ss.
The unsoundness of the anti-CSC argument based on A-Ss by itself enormously weakens the overall conclusion reach
in Lakoff 1986, for the bulk of the evidence Lakoff presents against the CSC involves A-Ss in one way or another.
Given that Goldsmith (1985) had already discussed B-Ss and that C-Ss play only a marginal role in his discussion, the
prime stimulus for the writing of Lakoff (1986) would appear to have been a recognition that unboundedly iterative
A-Ss are possible. Documentation that extractions from A-Ss do not threaten the CSC therefore goes a very long way
toward undermining the claim that this principle is incompatible with attested English data.
3.3 B-Scenarios
3.3.1 Basics
What Lakoff (1986) calls B-Ss seem to have been discovered by Goldsmith (1985) and are illustrated in (100a) (from
Goldsmith 1985, 133) and (100b) (from Lakoff 1986, 152).
(100) a. [How many courses]1 can we expect our graduate students to teach t1 and (still) finish a dissertation on time?
b. [How much]1 can you drink t1 and still stay sober?
Lakoff concludes that B-Ss also counterexemplify the CSC because they too instantiate true coordination.
He argues this by indicating that the number of possible conjuncts is unlimited, as in unquestioned coordinations (and
A-Ss), and that (partial) ATB extraction is possible. ((101a) is from Lakoff 1986, 153.)

< previous page page_77 next page >


If you like this book, buy it!
< previous page page_78 next page >
Page 78
(101) a. [How many courses]1 can you take t1 for credit, still remain sane, and get all As in t1?
b. the poison which1 he planned to grab a glass, drink t1, and still not die
These properties do seem to support a coordinate analysis, a conclusion strengthened by the fact that the conditions on
the nonoccurrence of and in successive conjuncts holding for unquestioned coordination cases (discussed in section
3.2.1) hold here as well. Hence, the final conjuncts of (100a,b) have the same status as the second conjunct of (101a);
that is, they are B-conjuncts. But the former have initial and, which the latter lacks.
Nonetheless, certain considerations, independent of the extractions in question, weigh against a coordinate analysis.
First, as in A-Ss, or conjuncts are impossible in B-Ss.
(102)*[How many courses]1 can we expect our graduate students to teach t1 or (still) finish a dissertation on time?
Second, unlike clear coordinations, B-Ss are incompatible with both.
(103) a. the courses which1 Bob can both take t1 for credit and get all As in t1
b. the courses which1 Bob can (*both) take t1 for credit and (still) stay sane
Third, there is the obvious semantic specificity of cases like (101), not characteristic of ordinary coordination with and.
Fourth and fifth, as noted by Goldsmith (1985), when extraction takes place, the apparent conjuncts must be (a) VPs (as
in A-Ss) and (b) "bare" VPs. Goldsmith illustrates property (b) with contrasts such as these:
(104) [How many courses]1 can we expect our students to teach t1 and still/*to still/*still to lead a normal life?
Presumably, it is the "bareness" condition that blocks extraction from B-Ss when the VPs are finite or both contain
modals.
(105) a.*the number of courses which1 his students teach t1 and still are happy
b.*the kind of handicap which1 he had t1 and still outperformed the opposition
c. [How many pills]1 do you think he can swallow t1 and (*can) still survive?

< previous page page_78 next page >


If you like this book, buy it!
< previous page page_79 next page >
Page 79
Sixth, as with A-Ss, interwoven dependency correspondents of B-Ss are ill formed.
(106) [How many courses]1 did Frank and Jim (*respectively) take t1 and still say sane (*respectively)?
Seventh, although apparent multiple-conjunct cases like (101a) exist, it is unclear that they involve multiple B-
conjuncts. Lakoff (1986, 154) characterizes (101a) as a mixture of an A-S and a B-S. Although the second conjunct has
the distinctive semantics of a B-conjunct (i.e., represents a state of affairs that is "surprising," given the previous
conjunct), the third conjunct does not. One should then consider linear cases where several conjuncts in a row have the
"surprise" semantics. But I delay discussion of the matter until section 3.3.3.
Summing up, although certain considerations support a coordinate analysis for B-Ss, as with A-Ss there are also
grounds for denying their coordinate character. In this case, in contrast to my analysis of A-Ss, I suggest that the latter
is the correct course. Such a step is more or less mandated to defend the CSC from B-S extractions; the approach
invoked for A-Ss is quite clearly not (fully) applicable here.
3.3.2 B-Scenario Extraction Does Not Fully Reduce to Selective Extraction
Although B-Ss partially resemble A-Ss in being syntactically asymmetrical, that is, in permitting extractions only from
certain conjuncts, as illustrated in (107), the B-S and A-S extraction patterns differ.
(107) a. [How long]1 did they avoid sweets t1 and still remain obese?
b.*[How slim]1 did Mike eat only ice cream and still remain t1?
c. [What color]1 did Mike paint his ears t1 and still make it seem he was normal?
d.*[What color]1 did Mike go color-blind and still paint his ears t1? 31
e. Himself1, Francine hopes to talk to Edward1 about t1 constantly and yet still stay uninvolved.
f.*Himself1, Francine hoped to be professional and yet still talk to Edward1 about t1 constantly.
g. That1, Francine could hardly feel shake t1 and still stay calm.
h.*That1, Francine might stay home and still feel shake t1.
i. Mike dated more nurses than (what1) I could date t1 and still stay sane.

< previous page page_79 next page >


If you like this book, buy it!
< previous page page_80 next page >
Page 80
j.*Mike could date more nurses than (what1) I could live here and still date t1.
As Lakoff (1986, 153-154) observes, A-Ss involving an extraction always require an extraction site in their final
conjunct but B-Ss do not. More significantly, unlike extractions from A-Ss, extractions from B-Ss need not obey the
conditions on selective island extraction and hence can violate the restrictions in (27): they are not categorically limited
(see (107a)), are not blocked by ACs (see (107c)), can involve reflexives (see (107e)), and can involve inherently
unpassivizable NPs (see (107g)). Further, extractions from the ''initial" conjuncts of B-Ss contrast with A-S extractions
in that they can be of type A2 (see (58b) and (107i)).
One might conclude from the starred examples of (107) that extraction from B-conjuncts (i.e., those representing
"surprising" states of affairs) is impossible. But that inference is unjustified, since (107b,d,f,h,j) all violate conditions on
extraction from selective islands. When these conditions are respected, extractions seem possible.
(108) a. Who1 did Mike remain celibate and yet still date t1?
b. [Which room]1 did Mike go color-blind and still want to paint t1 red?
c. the nurse who1 Mike could be in pain and still not call t1
None of (108a-c) seem ungrammatical to me. 32 So, whereas all extractions from A-Ss are selective island extractions,
with B-Ss this is likely true only of extractions from B-conjuncts, those representing what Goldsmith (1985) calls the
'nonetheless' reading.
Overall, though, since extraction from non-B-conjuncts of B-Ss is not selective island extraction, B-Ss cannot be kept
fully consistent with the CSC in the same way that A-Ss can. One cannot simply claim that those B-S conjuncts that
permit extraction are islands and that constituents extract from them only via the device of RPs. Defending the CSC
against B-Ss requires arguing, rather, that at least their "initial" constituents are not islands. But if their "initial"
constituents are not islands, B-Ss cannot be true coordinate structures in the sense referred to by the CSC. If they were,
the CSC would of course characterize all of their conjuncts as islands.
If the "initial" conjuncts of B-Ss are not islands, then, in contrast to the situation documented earlier with A-Ss, these
conjuncts should not behave like islands with respect to nonextraction phenomena that are

< previous page page_80 next page >


If you like this book, buy it!
< previous page page_81 next page >
Page 81
sensitive to island boundariesfor example, "affective" forms, negative polarity items, and multiple interrogations.
Although constructing relevant examples is not easy, the evidence seems to support the view that the "initial" conjuncts
of B-Ss are nonislands.
The pattern for "affective" elements in true (VP) coordinations is, roughly, that an "affective" cannot occur in only one
conjunct. 33
(109) a. Nobody said that Sally met a doctor somewhere/*anywhere and charmed a lawyer there.
b. Nobody said that Sally (*met a doctor here and) charmed a lawyer anywhere.
But the B-S situation contrasts with that of (109).
(110) Nobody said that Sally would (ever) meet a doctor (anywhere) and still visit a lawyer.
The acceptable "affectives" in (110) are in the "initial" conjunct of the B-S, the conjunct that permits nonselective
extraction. An "affective" in the B-conjunct is not acceptable, correlating with the (selective) island character of that
conjunct.
(111) Nobody said that Sally would meet a doctor and (still) (*ever) visit a lawyer (*anywhere).
A similar asymmetrical "affective" pattern is seen in (112).
(112) a. At UCLA, I don't think anyone (ever) managed to fail nine courses and (still) graduate in four years.
b. At UCLA, I don't think anyone managed to fail 9 courses and (still) (*ever) graduate in four years.
Next consider negative polarity items. As (113) shows, these cannot occur in only one conjunct of a true coordinate
VP.34
(113) a. Nobody thought he would lift a finger (*and worry about Bob).
b.*Nobody thought he would worry about Bob and lift a finger.
But negative polarity items in B-Ss manifest an asymmetrical pattern, correlating with the extraction facts.
(114) a. Nobody thought that Frank would lift a finger and still get criticized.
b.*Nobody thought that Frank would get criticized and still lift a finger.

< previous page page_81 next page >


If you like this book, buy it!
< previous page page_82 next page >
Page 82
Now consider multiple interrogations. It is impossible, at least in my idiolect, for one of several co-questioned
interrogative forms to appear in a VP conjunct of a standard coordination; see (90).
(115) a. Who claimed Frank studied how long (*and failed math)?
b. Who claimed Frank (*failed math and) studied how long?
B-Ss once more show an asymmetrical distribution.
(116) a. Who claimed that Frank could study how long and still visit Lydia often?
b. Who claimed that Frank could (*visit Lydia often and still) study how long?
Thus, various nonextraction facts support the conclusion drawn from extraction cases that B-Ss are not true coordinate
structures. Overall, whereas the non-"initial" conjuncts (B-conjuncts) of B-Ss behave like (selective) islands, the
"initial" conjuncts behave like nonislands.
An additional argument that B-Ss are not coordinate structures can be based on principle (117).
(117) If individual constituents of type C permit extraction of phrases of category j, then coordinated constituents of
type C permit ATB extraction of phrases of category j.
This principle, which seems true for English independently of the structures at issue in this discussion, claims that as
such, true coordination imposes no special categorical constraints on extraction. Given (117), if B-Ss represent (VP)
coordination, then, since independently English VPs permit extraction of arbitrary categories, B-Ss should permit ATB
extraction of arbitrary categories.
But apparent ATB extraction from B-Ss is in general limited to NPs.
(118) a. Who1 did Bob work for t~ and still not have respect for t1?
b.*[For whom]1 did Bob work t1 and still not have respect t1?
(119) a.*[In what way]1 did Lydia unsuccessfully repair the VCR t1 and still try to repair the TV t1?
b.*[How long]1 does Bob have to work t1 and still stay alert t1?
That is, apparent ATB extraction from an "initial" conjunct and a B-conjunct of a B-S is subject to the same constraint
as individual extraction from a B-conjunct.

< previous page page_82 next page >


If you like this book, buy it!
< previous page page_83 next page >
Page 83
These facts strongly count against a coordinate analysis of B-Ss. If such an analysis is rejected, then apparent ATB NP
structures like (119a) can be assimilated to the P-gap phenomenon, with the rightmost gap taken to be a P-gap; thus,
(119a,b) turn out parallel to the uncontroversially non-coordinate (120a, b).
(120) a. Who1 did Bob work for t1 without having any respect for t1?
b.*[For whom]1 did Bob work t1 without having any respect t1?
That is, under a noncoordinate analysis of B-Ss, (119b) and (120b) reflect the restriction of English P-gaps to NPs (see
Postal 1993a, 1994). Moreover, NP cases like (119a) further resemble P-gap forms in being in effect subject to the
other conditions on selective island extraction in (27) (see chapter 2 and Postal, in preparation a). Compare the cases in
(121), for example.
(121) a. [What color]1 did Valerie paint her house h (*without painting/*and still not paint her yacht t1)?
b. [Those rocks]1 Sandra will touch h (*without feeling move/*and still not feel t1 move).
c. Herself1, Edna1 can talk to h*without boring/*and still not bore t1).
Thus, ATB issues provide two reasons for denying coordinate status to B-Ss: non-NP extractions resembling ATB
cases are not permitted, and what look like ATB NP extractions from B-Ss satisfy independently known conditions for
an analysis in which the rightmost gap is a P-gap.
3.3.3 B-Scenarios and the Conjunct Constraint
An unusual argument reinforces the conclusion that B-Ss are not true coordinate structures. As remarked earlier, it
seems correct to divide Ross's original formulation of the CSC into separate principles. The one I called the Conjunct
Constraint in section 3.1 forbids the extraction of coordinate conjuncts themselves. The other, the CSC, bans (non-
ATB) extraction from true conjuncts. The Conjunct Constraint is almost never questioned; 35 nothing in Lakoff 1986 is
intended to challenge it. The argument against coordinate status for B-Ss based on the Conjunct Constraint is that to
suppose that B-Ss instantiate true coordination requires rejecting the Conjunct Constraint.
This conclusion depends on observations made by Lawler (1974) about an English negative construction, call it double
neg (DN), associated with

< previous page page_83 next page >


If you like this book, buy it!
< previous page page_84 next page >
Page 84
the L-extraction of certain constituents; see also Goldsmith 1985. DN seems possible only as a response to a previous
sentence; most commonly and naturally, DN sentences answer preceding questions. But it is not limited to the latter use
(see Goldsmith 1985, 140). DN examples include 122a-d), where the fronted constituents are italicized.
(122) a. Did you visit any churches? Not in Spain I didn't.
b. Can we get a decent meal? Not around here you can't.
c. Can this stain be removed? Not with that solvent it can't.
d. I am going to the movies. Not without doing the dishes you're not.
As in all of (122), DN seems most natural when it instantiates VP anaphora linked to a VP in the antecedent sentence.
But this is probably not required; (123), for instance, seems to be a viable DN example.
(123) Can we send astronauts to Mars? Not without bankrupting the country we can't send any there.
Lawler's key observation is that B-conjuncts can be DN L-extractees. ((124) is from Lawler 1974, 370.)
(124) Can linguists study negation? Not and stay sane they can't.
Lawler (1974, 370) concludes in particular that "the Coordinate Structure Constraint seems to be fractured beyond
repair ..." Of course, Coordinate Structure Constraint can refer here only to the Conjunct Constraint.
No doubt, if B-conjuncts were true coordinate conjuncts, DN cases like (124) would counterexemplify the Conjunct
Constraint, along the lines of (125).
(125) The logic of the claim that if they are coordinate, B-S structures like (124) counterexemplify the Conjunct
Constraint is as follows:
a. DN cases like (122)/(124)
i. involve extractions,
ii. manifest invisible VPs anaphorically linked to VPs in the question antecedent.
b. If B-Ss involve coordinate VPs in the sense referred to by the Conjunct Constraint, then, for example,
(124)involves extraction of a coordinate conjunct.
c. Therefore, the Conjunct Constraint is false.
Moreover, a variant of (125) can be provided for cases where DN does not involve VP anaphora.

< previous page page_84 next page >


If you like this book, buy it!
< previous page page_85 next page >
Page 85
Evidently, though, abandoning the Conjunct Constraint would solve a quite restricted problem (the analysis of B-Ss) at
the cost of creating many much graver ones. If there is no such condition, it is unexplained why no other (English)
constructions manifest analogs of the kind of extraction in (124). Analogs of (124) are not even possible for the non-B-
conjuncts of B-Ss or mixed A/B-Ss.
(126) a. Jerome can swallow the poison and still avoid death.
b. Can Jerome avoid death?*Not (and) swallow the poison he can't (and).
c. Jerome can go to the bar, drink 11 daiquiris, still stay sober, and (then) eat 4 burgers.
d. Can Jerome go to the bar, drink 11 daiquiris, and still stay sober?*Not and (then) eat 4 burgers he can't.
Similarly, analogs of (124) are not even possible for simple A-Ss.
(127) a. Can I go to the store?*Not and buy beer you can't.
b. Can he use the phone?*Not and make a date with Laura he can't (use it).
And, anticipating section 3.4 a bit, they are also impossible for the C-conjuncts of Lakoff's C-Ss.
(128) Do guys in the Caucasus eat snails?*Not and live to be 100 they don't.
DN structures are likewise barred for what I called D-Ss in section 3.1.
(129) Can he criticize de Gaulle?*Not and therefore criticize a Frenchman he can't.
And, evidently, they are impossible for true coordinate structures.
(130) Can tigers fly?*Not and eat Wheaties they can't.
So, abandoning the Conjunct Constraint in order to allow (only the B-conjuncts of) B-Ss to both be coordinate and
participate in the DN construction misfires in predicting falsely that, for example, (126b,d), (127a,b), (128), (129), and
(130) are well formed.
All the cases mentioned so far with respect to the possibility of abandoning the Conjunct Constraint involve VPs. But
without that constraint, nothing known would preclude the extraction of conjuncts of other extractable constituent types
as well (e.g., NPs, PPs, Ss). Since such cases

< previous page page_85 next page >


If you like this book, buy it!
< previous page page_86 next page >
Page 86
are unattested, support for a proposal to abandon the Conjunct Constraint merely to facilitate the description of English
B-Ss must be regarded as unfounded, for reasons parallel to those cited in note 8.
A fact about the interaction of B-Ss and DNs not noted by Lawler (1974) or Goldsmith (1985) is relevant to the
question whether B-Ss represent true coordinate structures or adjuncts. It also relates to the issue touched on earlier of
the extent to which B-conjuncts iterate. Cases like (131) might be true, iterative B-Ss; that is, they might have
successive B-conjuncts.
(131) Carol can take seven courses for credit, (still) stay sane, and (still) make the varsity hockey team.
Strikingly, though, such examples do not have DN analogs.
(132) Can Carol take seven courses for credit?*Not (still) stay sane and (still) make the varsity hockey team she can't.
I conclude from (132) that despite their semantics, iterated structures like (131) really lack the syntactic structure of
those B-Ss that permit extraction. If so, one might expect that such iterative structures would also preclude extraction
from non-B-conjuncts, which seems correct.
(133) a.*[How many pecan pies]1 can one eat t1, (still) not get sick, and (still) stay slim?
b.*the number of pecan pies which1 one can eat t1, (still) not get sick, and (still) stay slim
c.*He ate more pecan pies than (what1) I can eat t1, (still) not get sick, and (still) stay slim.
d.*[How many courses]1 can you take t1 for credit, (still) stay sane, and (still) not collapse from exhaustion?
e.*[How long]1 can one eat nachos t1, (still) not get sick, and (still) want a pizza?
f.*[What drugs]1 can you take t1, (still) feel OK, and (still) drive?
In this respect, those conjuncts that are distinctive to B-Ss contrast with true coordinate conjuncts in their "recursive"
possibilities.
In contrast, (133a-f) all map into fine sentences if and is added at the beginning of the second conjunct. What follows
the first conjunct is then presumably not an iteration of B-conjuncts but a single coordinate B-conjunct, parallel to a
coordinate adjunct like that in (134).
(134) He took those drugs [after [getting sick and feeling faint]].

< previous page page_86 next page >


If you like this book, buy it!
< previous page page_87 next page >
Page 87
This view predicts that a correspondent of (132) with an added and should also be good, for it would contain a single
fronted coordinate adjunct. This is correct.
(135) Can Carol take seven courses for credit? Not and (still) stay sane and (still) make the varsity hockey team she
can't.
What emerges from this discussion is the likelihood that manifestating the 'and nonetheless' semantics typical of B-Ss
by no means guarantees the presence of the type of B-S syntactic structure that permits extraction. An additional
consideration supports this conclusion. French has apparent coordinate VP structures with the 'and nonetheless'
semantics; but these rigorously resist any extraction from any conjunct.
(136) a. Arnaud peut boire 11 calvados et rester lucide.
'Arnaud can drink 11 calvados and stay clearheaded.'
b.*[Combien de calvados]1 Arnaud peut-il boire t1 et rester lucide?
'How many calvados can Arnaud drink and stay clearheaded?'
c. Arnaud peut boire 11 calvados et rester capable de conduire certains vhicules.
'Arnaud can drink 11 calvados and remain capable of driving certain vehicles.'
d.*[Quels vhicules]1 Arnaud peut-il boire 11 calvados et rester capable de conduire t1?
'Which vehicles can Arnaud drink 11 calvados and remain capable of driving?'
Since (136a,c) seem to have the same general properties as corresponding English examples, if English B-Ss are
correctly analyzed as coordinate, it is hardly plausible that French B-Ss are noncoordinate. So, in the absence of ad hoc
constraints, merely taking English B-Ss to be coordinate and abandoning the CSC to permit extraction from them
would have the false consequence that French B-Ss permit extraction (from non-B-conjuncts). But if one rejects
Lakoff's view of B-Ss, one can say that whereas English and French share the coordinate structure that blocks all
extraction via the CSC, English assigns B-Ss in addition a noncoordinate structure (of some kind); only the latter then
allows (nonselective) extraction.
3.3.4 B-Conjuncts as Adjuncts
If extraction-permitting B-Ss do not represent true coordination, a reasonable further assumption would be that B-
conjuncts represent an unusual type of adjunct; the "initial" conjuncts would then represent

< previous page page_87 next page >


If you like this book, buy it!
< previous page page_88 next page >
Page 88
main-clause VPs. Thus, the relation between a truly coordinate B-S like that found in French and the type that permits
some extractions is that in the latter, one VP conjunct, a B-conjunct, is treated as an adjunct. That would explain why
the "initial" conjuncts are not islands and why B-conjuncts allow selective extractions. It would also explain why B-
conjuncts in DN behave like uncontroversial adjuncts. 36
(137) Can linguists VP study B-Ss1?
a. Not [without going insane]2 they can't VP1 t2.
b. Not [and stay sane]3 they can't VP</SM 1 t3.
My suggestion here about B-Ss significantly parallels that of Goldsmith (1985, 141-142), who observes that an adjunct
analysis of B-conjuncts accounts both for the extraction possibilities and for Lawler's cases like (138).
(138) a. Can I go outside without any clothes on?
b. Not and stay healthy, you can't.
Goldsmith proposes informally that B-Ss involve a special usage of and as a subordinator rather than a coordinator.
An adjunct view of B-conjuncts is strengthened by other restrictions. In various VP-fronting constructions, where
adjuncts can either be "carried along" or "left behind," so can B-conjuncts.
(139) a. They said Val can eat 13 burgers without vomiting and
i. [eat 13 burgers without vomiting]1 he can t1.
ii. [eat 13 burgers]1 he can t1 without vomiting.
b. They said Val can eat 13 burgers and still not vomit and
i. [eat 13 burgers and still not vomit]1 he can t1.
ii. [eat 13 burgers]1 he can t1 and still not vomit.
A different but related class of cases makes the same point.
(140) a. [Eat 13 burgers without vomiting]1 though he can t1, ...
b. [Eat 13 burgers]1 though he can t1 without vomiting, ...
c. [Eat 13 burgers and still not vomit]1 though he can1, ...
d. [Eat 13 burgers]1 though he can t1 and still not vomit, ...
The properties just documented for B-S conjuncts do not hold of conjuncts whose coordinate status is not in question.
Compare:
(141) They said Val can eat 13 burgers and drink 14 beers and
a. [eat 13 burgers and drink 14 beers]1 he can t1.
b.*[eat 13 burgers]1 he can t1 and drink 14 beers.

< previous page page_88 next page >


If you like this book, buy it!
< previous page page_89 next page >
Page 89
(142) a. [Eat 13 burgers and drink 14 beers]1 though he can t1,...
b.*[Eat 13 burgers]1 though he can t1 and drink 14 beers,...
Furthermore, B-conjuncts behave like adjuncts with respect to forms of VP anaphora.
(143) a. Helen said she could VP eat 13 burgers1 without vomiting and she can VP1 without vomiting.
b. Helen said she could VP eat 13 burgers1 and still feel hungry and she can VP1 and still feel hungry.
Again, the adjunct-like behavior of B-conjuncts with respect to this form of anaphora is highlighted by comparing
(143b) with a partially parallel structure differing in that it is a true coordination rather than a B-S.
(144)*Helen said she could VP eat 13 burgers1 and drink 9 beers and she can VP1 and drink 9 beers.
One notable difficulty already touched on argues against taking B-conjuncts as adjuncts. This is the fact that and can
fail to appear on B-conjuncts in n-ary mixed Ss under the same conditions as it does in true coordinates. Compare:
(145) a. [What courses]1 can you register for t1, fail to attend t1, and get As in t1?
b. [What courses]1 can you register for t1, still have a social life, and get As in t1?
In each case, and can fail to appear on the bracketed conjunct because of its nonfinal position in a linear structure. But
(145a) is a true coordinate structure whereas (145b) is a mixed Scontaining a B-conjunct. The extraction in the latter
shows that the absence of and cooccurs with extraction and thus must, in my terms, be allowed even in the
noncoordinate structure that I have suggested underlies nonselective extraction from B-Ss. Although genuine, nontrivial
parallelisms thus exist between B-Ss and coordinates, arguing against their adjunct status, I do not believe such facts
overwhelm the evidence that those B-conjuncts cooccurring with extractions are adjuncts. However, until principles
capturing the generalization covering (145a,b) are found, this conclusion remains partially vulnerable.
3.3.5 B-Scenarios: Summary
Unlike the situation with A-Ss, it is not possible to analyze all extractions from B-Ss as selective island extraction.
Extraction from non-B-conjuncts

< previous page page_89 next page >


If you like this book, buy it!
< previous page page_90 next page >
Page 90
does not respect the appropriate conditions. Attested extractions from BSs are nonetheless arguably consistent with the
CSC because as considerable evidence indicates, the B-conjuncts occurring in B-Ss that permit extraction are not true
coordinate conjuncts and there is a substantial basis for considering them adjuncts. This permits analyzing the other
conjuncts of such B-Ss, those that permit nonselective extraction, as main-clause VPs. Under these conditions, the
compatibility of English B-Ss with various extractions raises no problems for the CSC, which refers to true coordinate
constituents and not adjuncts or main-clause VPs.
3.4 C-Scenarios
3.4.1 Basics
What Lakoff (1986) refers to as C-Ss are illustrated in (146).
(146) a. The guys in the Caucasus drink that stuff and live to be 100.
b. the stuff which1 the guys in the Caucasus drink t1 and live to be 100
Their semantics differs from that of B-Ss in that the situation represented by the first conjunct is the cause of that
represented by the last, the C-conjunct. As with other Ss, the extraction possibilities for the different types of conjuncts
contrast. From the first conjunct, it is possible to extract not only NPs, as in (146b), but also non-NPs.
(147) a. the conditions [under which]1 the guys in the Caucasus can eat that stuff t1 and live to be 100
b. [With what kinds of sauces]1 do the guys in the Caucasus eat that stuff h and live to be 100?
c. [For how long]1 have guys in the Caucasus eaten that stuff t1 and lived to be 100?
Thus, these extractions are not reducible to a selective island phenomenon. However, extraction from C-conjuncts does
seem to be selective.
(148) a. a disease which1 the guys in the Caucasus eat snails and don't suffer from t1
b. a disease [from which]1 the guys in the Caucasus (*eat snails and) don't suffer t1
c.*[How rapidly]1 do the guys in the Caucasus eat snails and recover from that disease t1?

< previous page page_90 next page >


If you like this book, buy it!
< previous page page_91 next page >
Page 91
Overall, then, C-Ss resemble B-Ss and not A-Ss in that maintenance of the CSC cannot depend completely on the view
that the C-S conjuncts are selective islands. Rather, it must be denied that C-Ss are coordinate structures.
3.4.2 Noncoordinate Properties of C-Scenarios
Significantly, alongside properties that suggest a coordinate analysis of C-Ss are others that oppose such a treatment.
Disjunction and the coordinate quantifier both are impossible.
(149) a.*the stuff which1 the guys in the Caucasus drink t1 or live to be 100
b.*The guys in the Caucasus (all) both drink that stuff and live to be 100.
c.*the stuff which1 the guys in the Caucasus (all) both drink t1 and live to be 100
Of course, the short variant of (149b) is fine if interpreted as an ordinary logical coordination. But the extraction in
(149c) precludes an ordinary coordinate analysis. Similarly, C-Ss of the interwoven dependency type are not possible.
(150)*the stuff which1 the guys in the Caucasus and the guys in the Himalayas drink t1 and live to be 100, respectively
Further, it is difficult to imagine a non-binary C-S. Structures with three or more conjuncts always seem to be non-
C-Ss, specifically, logical coordinations.
(151) a. The guys in the Caucasus drink that stuff, eat snails, and live to be 100.
b.*the stuff which1 the guys in the Caucasus drink t1, eat snails, and live to be 100
c.*the stuff which1 the guys in the Caucasus eat snails, drink t1, and live to be 100
That (151a) is not a C-S but a logical coordination is supported by (151b,c), whose ungrammaticality then follows from
the usual interaction of the CSC with true coordinations. So those C-Ss permitting extraction resemble B-Ss and
contrast with A-Ss in not really being ''recursive."
Further evidence for rejecting the coordinate status of C-Ss involves VP anaphora. This is impossible when the
antecedent is outside true coordinate structures and the anaphoric VP is inside, as already in effect illustrated in (144);
see also (152).

< previous page page_91 next page >


If you like this book, buy it!
< previous page page_92 next page >
Page 92
(152) a.*My father won't VP fire a gun1 because Joan is careless and once did VP1.
b.*My father won't VP fire a gun1 even though he has one and wants to VP1.
McCawley (1988, 502) cites the well-formed (153).
(153) Sam refuses to VP fire a gun1, because his father once did VP1 and killed someone.
But unlike the constituents in (152), the constituent once did and killed someone is a C-S 37 Therefore, (153)
parallels adjunct cases like (154).
(154) Sam refuses to VP fire a gun1, because his father once did VP1 in order to kill someone.
Yet more evidence distinguishes C-Ss from true coordinate structures. Multiple interrogations involving non-NPs,
shown earlier to be impossible across true coordinate boundaries (see (115)), are possible into (the first conjuncts of)
C-Ss.
(155) Who said the guys in the Caucasus eat what and live to be 100?
Similarly, in contrast to true coordinate structures, C-Ss permit negatively "licensed" "affective" elements in their
initial conjuncts.
(156) Nobody claimed that the guys in the Caucasus eat anything like that and live to be 100/*drink rum.
Further, in contrast to the true coordinate facts illustrated in note 28, topicalization is possible in the first conjunct.
(157) Henry claimed that [that stuff]1, the guys in the Caucasus can eat t1 with yogurt and live to be 100/*and drink
wine.
3.4.3 C-Scenarios: Summary
If they are true coordinates, C-Ss counterexemplify the CSC. Since there are good grounds for doubting their
coordinate status, though, at this stage there is little basis for thinking they genuinely threaten that principle.
3.5 Conclusion
The preceding sections of this chapter had an extremely limited goal. I have tried to show only that data like those taken
by Lakoff (1986) to refute the CSC fail to do so, for partially diverse reasons. I have claimed

< previous page page_92 next page >


If you like this book, buy it!
< previous page page_93 next page >
Page 93
that A-Ss are consistent with the CSC since the extractions permitted from their conjuncts turn out to be systematically
restricted to the sorts of extractions possible from (other) selective islands. Therefore, it can be assumed that all the
extraction-permitting conjuncts of A-Ss are selective islands, hence, via (46a), islands. Therefore, arguably, extraction
from them is possible only because, as a full version of the CSC in Ross's (1967) formulation claims, extraction is
permitted even from islands if RPs occur in the extraction sites. To support the island character of A-S conjuncts, I
advanced evidence from "affective" elements, negative polarity items, embedded topicalization, and multiple
interrogations. The impossibility of extraction from what appear to be French equivalents of A-Ss also suggests that
extractions from English A-Ss are not a consequence of a nonisland status.
For B-Ss, the situation is rather different. Although extraction from B-conjuncts (i.e., the non-"initial" conjuncts of
B-Ss) is arguably selective island extraction, extraction from "initial" conjuncts does not respect the constraints on
extraction from selective islands. Therefore, an approach to keeping B-Ss consistent with the CSC that parallels the
approach invoked for A-Ss will not work. The alternative is to deny that those B-Ss that involve extraction represent
true coordinate structures, a position for which I presented diverse evidence. Among other things, I noted that taking all
B-Ss to be true coordinate structures would require rejecting the Conjunct Constraint as well as the CSC. Additionally,
since only a minor group of conjoined VP types permit B-S type extraction, merely rejecting the CSC would vastly
"overgenerate" even within the VP realm. Further, I presented evidence that those B-Ss that manifest extractions do not
allow B-conjuncts to iterate, thereby contrasting with truly coordinate structures. Treating English B-Ss as coordinate
would also fail to account for solid parallels between them and clear instances of adjuncts. Finally, denying at least the
possibility of a noncoordinate analysis for English B-Ss leaves no way to account for the extraction contrasts between
semantically parallel English and French cases.
For C-Ss also, I suggested that the proper analysis is to deny that they are true coordinate structures. However, whereas
B-Ss may have a dual analysis, some being truly coordinate and some involving adjuncts, no reasons were found to
assign dual structures to C-Ss. For both B-Ss and C-Ss, denial of coordinate status is supported by the fact that, in the
presence of extractions, there are no clear cases of other than binary structures. Instances in which extraction-permitting
B-Ss might seem to

< previous page page_93 next page >


If you like this book, buy it!
< previous page page_94 next page >
Page 94
be truly "recursive" actually involve mixtures of A-Ss and B-Ss. Therefore, in the apparent cases of B-S "recursion"
reported by Lakoff (1986), the true "recursion" is associated with the A-S structure, as in, for example, (8). Thus,
among the Ss claimed to counterexemplify the CSC, the property of "recursion'' characteristic of true coordinate
structures of all kinds is found only with A-S type conjuncts, those that were shown to be (selective) islands.
I stress the limitations of current goals in order to avoid possible criticisms. It might be claimed that since I have not
offered analyses of any of the S types, some conclusion follows. But my goal was not to analyze diverse structures
related to coordinate structures in one or more ways; rather, it was only the much weaker task of showing that no real
basis for rejecting the CSC has been developed. Although providing valid analyses consistent with earlier conclusions
would strengthen my defense of the CSC, it is not necessary for that defense.
Since Lakoff (1986) does not ultimately show that there are any genuine counterexamples to the CSC, claims like (158)
are unfounded:
(158) "Moreover, any theory of syntax that requires that the coordinate structure constraint exist in the syntax is simply
incorrect." (Lakoff 1986, 152)
I interpret Lakoff's perhaps obscure phrase require ... exist ... as embodying a claim that his work falsities any
framework from which the CSC (or its consequences) follows as a theorem.
This might be an implicit suggestion that his S observations thereby falsify (e.g.) the GPSG framework, given (158).
(159) "Section 3 ... demonstrates that Ross's Coordinate Structure Constraint (CSC) and the "across-the-board" (ATB)
violations of it follow as theorems from the grammar fragments ... in ... previous ... sections." (Gazdar 1981, 155-156;
emphasis mine)
GPSG writers have rightly stressed that various other purported "theorems" that have been influential in recent
linguistics turn out not to be genuine theorems (see, e.g., Pullum and Gazdar 1982); but in fact the claimed
theoremhood of the CSC is one of these. No proof of the result claimed in (159) is offered, and no work containing
such proof is cited. Purported theorems disassociated from proofs always deserve deep skepticism.

< previous page page_94 next page >


If you like this book, buy it!
< previous page page_95 next page >
Page 95
Moreover, taken literally, the supposed "theorem" is incoherent. First, Ross's CSC was intended as a linguistic
universal, whereas (159) claims that the CSC follows from fragments of English grammar. Since no universal can
follow from a grammar of English, the very statement of (159) indicates that no proof had been constructed.
Second, the CSC (see (1)) literally constrains the operation of transformations, whose existence in correct natural
language grammars GPSG denies. But a theorem about limitations on the operation of transformations cannot be
derived from a framework that claims there are none. Minimally, then, (159) must be be taken to mean only what (160)
says.
(160) "Gazdar (1981) shows in detail how all the phenomena covered by the CSC ... follow from the analysis of
unbounded dependencies developed ... above, and the rule schemata for coordination ..." (Gazdar 1982, 175)
This states only that the facts underlying Ross's postulation of the CSC follow from the GPSG analysis. Setting aside
the universality aspect, that informal assertion is coherent but associated with no more of a proof than (159).
Significantly, claims that the CSC is a theorem are not repeated in later, more careful formulations of GPSG ideas, such
as Gazdar et al. 1985. Independently of whether genuine counterexamples to the CSC exist, then, an assertion that
GPSG has been falsified because it has the CSC as a theorem has no known basis.
To conclude, the current status of the CSC would appear to be this: it is supported by massive evidence from many
languages, and, if careful account is taken of what it claims in combination with Ross's insight about the role of RPs in
permitting extraction from islands, it has no known clear counterexamples in English.

< previous page page_95 next page >


If you like this book, buy it!
< previous page page_i next page >
Page i

Three Investigations of Extraction

< previous page page_i next page >


If you like this book, buy it!
< previous page page_97 next page >
Page 97

Chapter 4
Right Node Raising and Extraction
4.1 Background
The term right node raising (RNR) is an atheoretical designation for the phenomenon (not for any type of rule or
characterization of it) illustrated in (1).
(1) a. Ernest suspected t1, Louise believed t1, and Michael proved t1 [that she was guilty]1.
b. She may have t1 and should have t1 [defrosted the roast]1.
c. They know when t1 but they don't know where t1 [he abused the dog]1.
d. Eloise peeled t1 and Frank ate t1 raw [the large Spanish onion]1.
I refer to the right-hand constituent that seems in such constructions to correspond to n (n > 1) gaps in the various
conjuncts to the left as the RNR pivot. Evidently, in many respects RNR pivots relate to gaps in the way that the
extractees of standard L-extractions do.
In this chapter I consider RNR, without, however, providing an explicit overall account of its nature. Rather, I deal with
two major points that, I suggest, directly or indirectly reveal features that a viable conception of RNR must have. First, I
argue that RNR falls into the same general class of phenomena as the L-extractions cited in chapters 1-3. Hence,
whatever descriptive mechanism is appropriate for these phenomena is correct for RNR. Put in other terms, I argue that
RNR is an extraction phenomenon. More precisely, coordinate RNR cases like (1a-d) involve the same kind of
interaction of an extraction phenomenon with coordination that across-the-board (ATB) L-extractions like that in (2)
do.
(2) [What kind of large onion]1 did Eloise peel t1, Marsha cook t1, and Frank eat t1?

< previous page page_97 next page >


If you like this book, buy it!
< previous page page_98 next page >
Page 98
Although the conclusion that RNR is an extraction phenomenon might seem banal, it directly contradicts proposals by
McCawley (1982, 1987, 1988), Ojeda (1987), Levine (1985), McCloskey (1986), and Kayne (1994), among others, to
the effect that RNR is radically different from L-extractions.
Second, I argue that despite their having been taken to support a Slash category view of extraction (see Gazdar 1981,
179-180), RNR constructions are incompatible with current formulations of Slash approaches. 1
4.2 The Unity of Right Node Raising and Left Extractions
4.2.1 The Context
4.2.1.1 A Novel Proposal about Right Node Raising
McCawley (1982, 98-101; 1987, 186-191; 1988, 528-533) has argued for a conception of RNR that among other things
radically distinguishes it from L-extractions.
(3) a. "... my treatment of RNR, which I argued to involve fusion of identical constituents without change of constituent
structure, and consequently a surface structure involving not only discontinuity but also a node with multiple mothers
..." (McCawley 1987, 186; emphasis mine)
b. "... (RNR) ... alters word order without altering constituent structure, ... "(McCawley 1982, 98)
McCawley (1987, 186) gives (4b) as the surface structure that his approach associates with (4a).
(4) a. John loves, and Mary hates oysters.

Although certain aspects of this account are less than clear, the key feature for present purposes is that under its
assumptions, RNR and L-extractions are radically different. On McCawley's view as on almost

< previous page page_98 next page >


If you like this book, buy it!
< previous page page_99 next page >
Page 99
all others, L-extractions would, for example, systematically yield constituent structure relations distinct from those
found in corresponding non-extraction cases. 2
A position similar to, though even less explicit than, McCawley's is stated in (5).
(5) a. "However, one might follow McCawley's approach ... and argue that in some respect or other the "raised" node is
simultaneously present in all conjuncts in RNR. There are very plausible grounds for this proposal." (Levine 1984, 16)
b. "... (RNR) ... exhibits a number of peculiarities that make it extremely unlikely to be ... a displacement phenomenon."
(Levine 1985, 492)
c. "Such evidence further reinforces the conclusion that the so-called Raised element in RNR is actually present in all
conjuncts in a way that makes it sensitive to island constraints, pronominalization constraints, and so onthat, in fact, it is
present in some sense in the phrase structure sites from which it appears to have been displaced." (Levine 1985, 496)
To justify the view that RNR is fundamentally distinct from other phenomena, one must do two things. Positively, one
must argue that RNR's properties sharply differentiate it from the phenomena from which it is theoretically
distinguished. Negatively, one must show that RNR does not share characteristic properties of other phenomena,
specifically, properties of L-extractions. Any conclusion must balance attested differences and similarities. But
McCawley's works on RNR do not deal with the second issue; nor do Levine's. Both authors merely try to attest ways
in which RNR constructions are unique. I argue for the extraction character of RNR in the context of a critique of
McCawley's and Levine's proposals, concentrating on McCawley's since they largely subsume Levine's discussion.
4.2.1.2 Objections to the Proposal
4.2.1.2.1 Outline
At first glance, the evidence McCawley gives for sharply distinguishing RNR from L-extractions is not negligible. It
consists of what can, I believe, be analyzed into at least seven arguments. Each can be viewed as putative support for
the multiple-mother, constituent-structure-preserving features of (4b).3

< previous page page_99 next page >


If you like this book, buy it!
< previous page page_100 next page >
Page 100
4.2.1.2.2.Antecedent-Anaphor Linkages
One argument, due to Levine (1984, 1985), hinges on the assertion that RNR preserves all the antecedent-anaphoric
pronoun linkages in corresponding non-RNR structures. This would putatively support structures like (4b) since the
linkages are presumed to be stated in terms of configurational properties, which RNR surface structures like (4b) would
fully preserve. However, the underlying factual claim, although seemingly valid for the examples Levine and
McCawley discuss, seems not to be true. It is counter-exemplified in my idiolect by, for example, (6).
(6) a. I described [some blind victim's]1 sister to him1.
b.*I should have described t1 to him2 and would have described t1 to him2 [some blind victim's2 sister]1.
c.*I should have described himself1 (and only himself1) to Bob1.
d. I should describe t1 to Bob1 and probably will describe to Bob1/him1 himself1 (and only himself1).
Consider also (7a,b), involving nonpronominal anaphora.
(7) a. I could have taken pictures of [some woman]1 for [that woman]1.
b.*I could have taken pictures t1 for [that woman]2 and should have taken pictures t1 for [that woman]2 [of [some
woman]2]1.
The very assumptions underlying the pronominal anaphora argument combine with the contrasts in (6) and (7) to
support the conclusion that in relevant ways RNR structures involve structural relations distinct from those of their
simpler counterparts, not just word order differences.
4.2.1.2.3 Relative Clause Right-Node-Raising Pivots
McCawley's second basis for in effect distinguishing RNR from L-extractions is that a relative clause that is an RNR
pivot is as much an island as one that is in a standard position (see McCawley 1982, 100; also Ojeda 1987, 262). For
example:
(8) a. Tom bought a can opener t1 and Alice bought a dictionary t1 [that were once owned by Leonard Bloomfield]1.
b.*[Which linguist]1 did Tom buy a can opener t2 and Alice buy a dictionary t2 [that were once owned by t1]2?
The argument from (8) is quite undeveloped, however.
First, the relative clauses and NPs at issue are the sort that permit relative clause extraposition. 4 But, as Ross (1966)
first noted, extraposed relatives are islands even in non-RNR contexts.

< previous page page_100 next page >


If you like this book, buy it!
< previous page page_101 next page >
Page 101
(9) a. Ted bought a can opener t1 last week [which belonged to Bloomfield]1.
b.*[Which linguist]1 did Ted buy a can opener t2 last week [which belonged to t1]2?
So there is no reason to assume that it is a property special to RNR constructions that is relevant to (8b). Moreover,
anaphora facts clash with the claim that the same structural relations hold of both NP-embedded relatives and
extraposed ones.
(10) a. A woman who hated Gladys1 ran up to her1 in the street.
b.*A woman t1 ran up to her: in the street [who hated Gladys2]1.
c. A boy who adored [some movie star]1 ran up to [that movie star]1 on the stage.
d.*A boy t1 ran up to [that movie star]2 on the stage [who adored [some movie star]2]1.
Therefore, an analysis assigning extraposed relatives the same structural relations as NP-embedded ones just seems
incorrect.
Moreover, the logic of the argument from the contrast in (8) is obscure. The assumption seems to be that an RNR pivot
relative clause could be an island only if it were, in some sense, at every level a part of its "pre-RNR" island. But if
relative clauses are identifiable independently of their contextas in, for example, Gazdar's (1981, 1982) proposals,
where they are defined by unique node labelingMcCawley's argument would not begin to go through. To derive support
from the contrast in (8) for his view of RNR, then, it would be necessary to argue that there is no way to identify an
RNR pivot relative clause as a relative clause per se without referring to its presumed status as a subconstituent of a
containing NP. This never-undertaken task seems at the least enormously arduous.
4.2.1.2.4 VP Anaphora
McCawley's (1982, 100) third argument depends on cases like (11).
(11) Tom admires t1 and is sure that everyone else admires t1 [Adolf Hitler]1, but of course you and I don't (= admire
Adolf Hitler).
Here VP anaphora seems to take as its antecedent a phrase that would not exist (in surface structure) under ordinary
views of RNR, but does exist there under McCawley's view. This might constitute an objection to certain views. But it
is unclear why it argues against, for example, a position according to which the relevant verbal phrases exist in abstract
structures

< previous page page_101 next page >


If you like this book, buy it!
< previous page page_102 next page >
Page 102
whereas the surface forms are of the conventional (single-mother) sort. That is, it is obscure how (11) could disconfirm
a view that takes it to have a conventional surface structure as well as a more abstract form something like (12).
(12) Tom admires Adolf Hitler and is sure that everyone else admires Adolf Hitler, but of course you and I don't (=
admire Adolf Hitler).
McCawley's modified (via multiple mother nodes) transformational view of RNR also assumes underlying structures
like (12) but claims that the surface structure preserves the VP constituency of the RNR pivot. This seems to assume
among other things that VP anaphora in general is defined exclusively on surface relations. However, reference to
nonsurface aspects of structure is motivated since the antecedent phrase for VP anaphora cannot be required to be a
surface constituent, even given McCawley's richer notion of the latter. For instance, in (13) the antecedents clearly
cannot be surface phrases. ((13c) is from Fiengo and May 1994, 220, attributed to William Ladusaw.)
(13) a. Gordon may have discovered no counterexamples but Elissa did ( discover no counterexamples).
b. The doctor didn't do anything but the nurse did ( do anything).
c. Max didn't talk to anyone but Oscar did ( Oscar talked to anyone).
4.2.1.2.5 Contrasting Island Behavior
The fourth and fifth arguments that McCawley gives (1982, 100-101; 1988, 530-532) derive from observations and
conclusions by Wexler and Culicover (1980, 299-303). The arguments are based on contrasts like the one in (14).
(14) a. Mary buys t1 and Bill knows a man who sells t1 [pictures of Elvis Presley]1.
b.*Who1 does Mary buy t2 and Bill know a man who sells t2 [pictures of t1 ]2?
With respect to such cases, Wexler and Culicover induce generalization (15).
(15) "[A] raised [by RNR] node always behaves, vis--vis all constraints on analyzability, just as it would if it were in
its original underlying position. Hence, whereas it is apparently possible to apply RNR to a constituent of a relative
clause, if we then try to analyze this raised node, we find that it acts as though it were still within the relative clause." 5
(Wexler and Culicover 1980, 301)

< previous page page_102 next page >


If you like this book, buy it!
< previous page page_103 next page >
Page 103
Notably, under McCawley's nonextraction view of RNR, at every stage the RNR pivot is part of the island-defining
relative clause.
One might then define the fourth argument for McCawley's view of RNR as saying this. At least certain island
constraints (roughly, so-called Subjacency) govern L-extractions, but not RNR, as seen in (14). Hence, RNR is not an
extraction. This is, I assume, the argument implicit in Levine's (1985, 492) remark that "elements [i.e., RNR pivots:
PMP] may be an unlimited number of bounding nodes removed from their associated gaps, representing massive
violations of island constraints...." Such an argument is relatively weak, however. An alternative would say that
although all extractions, including RNR, share many properties, the island types in question constrain only L-
extractions. This would be a way of capturing similarities between RNR and other extractions without denying certain
differences, a natural and common way of proceeding. It cannot be insisted that all extractions share all properties. For
example, as indicated in chapter 2, among English L-extractions, the one associated with restrictive relatives can extract
a wide variety of constituents from ACs but the ones associated with nonrestrictive relatives and topicalization cannot.
This is because the restrictive relative extraction is an A-extraction whereas the extraction associated with
nonrestrictives and topicalization are B-extractions and hence incompatible with ACs. One can hardly conclude from
this that, for example, one of the two types of L-extraction is not an L-extraction (see also note 9).
Moreover, a case can be made that some of the restrictions on L-extractions often subsumed under Subjacency in fact
do also hold for RNR. This is true, for instance, of extraction from subjects.
(16) a. the guy [of whom]1 Mary bought pictures t1 and Sally bought carvings t1
b. Mary bought pictures t1 and Sally bought carvings t1 [of the famous blind poet]1.
c. Pictures of the poet were bought by Mary and carvings of the poet were bought by Sally.
d.*the poet [of whom]1 pictures t1 were bought by Mary and carvings t1 were bought by Sally
e.*Pictures t1 were bought by Mary and carvings t1 were bought by Sally [of the famous blind poet]1.
An MIT Press referee raises a number of objections to the original form of this section. I examine these remarks in
appendix C.

< previous page page_103 next page >


If you like this book, buy it!
< previous page page_104 next page >
Page 104
4.2.1.2.6 Islandhood of Right-Node-Raising Pivots
One might define the other argument linked to (14), the fifth argument putatively supporting McCawley's nonextraction
view, as claiming that RNR could not be an extraction because an analysis that assumes it is fails to capture the
following fact. Although RNR operates out of (some) islands, subsequent L-extractions are still constrained by those
apparently "destroyed" embeddings, as shown in (14b). The same argument is cited by Levine (1985, 492-494) and
Oehrle (1990, 411-412); see (151).
However, the issues surrounding this argument are complex and little explored. In the case of any potential L-
extraction-RNR interaction like (14b), at least two distinct conceptualizations are a priori possible.
(17) Schematically,
a. the L-extraction could be defined on a predefined RNR structure; or
b. the RNR structure could be defined on a predefined L-extraction structure.
In the latter case, RNR would in effect be taking as its targets what Mller (1996) calls remnants. It seems that previous
work (e.g., Wexler and Culicover 1980) may have assumed without real justification that only (17a) was possible. 6 But
without supporting argument, this need not be accepted. If there are two possibilities, the ill-formedness of cases like
(14b) noted by Culicover and Wexler (1980) would have to involve more than one constraint, one to block each
possible description.
Consider the following as a possible partial theory for English RNR:
(18) a. RNR pivots block L-extraction. For example, they are islands moreover, in terms of chapter 1, locked islands.
b. The targets for RNR can be L-extraction remnants.
(18a) blocks description (17a) for cases like (14b), leaving only (17b). But that description is surely independently ill
formed in (14b), since it involves L-extraction out of restrictive relative clause islands known to block it. Examples like
(19), lacking non-RNR-dependent islands, are predictably acceptable.
(19) Who1 does Mary buy t2 and Bill sell t2 [pictures of t1]2?
This follows since the ban on describing (19) as in (17a) says nothing about an analysis in terms of (17b). In (19) L-
extraction on pre-RNR structures yields no violation. It follows that (19) would be blocked only if

< previous page page_104 next page >


If you like this book, buy it!
< previous page page_105 next page >
Page 105
there were some principle precluding RNR from taking L-extraction remnants as its targets. But at this stage it is
unclear that there is any such principle.
In fact, it is not even clear that allowing (18b) yields a contrast with the behavior of L-extractions, because English
allows little interaction between multiple L-extractions from the same constituent. The closest I have come to testing
the question involves a contrast between pairs like (20c) and (20d) (see Lasnik and Saito 1992, 101, (140c)).
(20) a. Marilyn believes I talked to John about pictures of Mary.
b. Marilyn believes that John1, I talked to t1 about pictures of Mary.
c. ??[Pictures of Mary]2, Marilyn believes that John1, I talked to t1 about t2.
d.*[Pictures of t1]2, Marilyn believes that John:, I talked to t1 about t2.
Although (20d), which would be an L-extraction remnant extraction, is indeed bad, the closest nonremnant extraction,
(20c), is hardly perfect either. One would obviously be hesitant to draw major theoretical conclusions distinguishing L-
extractions from RNR in fundamental ways because of what pairs like (20c) and (20d) indicate about a possible ban on
L-extracting L-extraction remnants. Moreover, even if there were such a contrast, it would support a relatively weak
argument against an extraction view of RNR, one showing only that RNR is different in a particular way.
Part of my discussion here depends on (18a), which has not been justified. I would like to raise the possibility that (18a)
is just a special case of the principle referred to as the Right Selective Island Constraint (RSIC) in appendix A. That
principle represents the now well-known observation that complex NP shift (CXS) cases do not permit extraction; see
section A.3. What I am suggesting is that this is true in general for right extractees including RNR pivots despite cases
like the grammatical (19). The latter might seem to show that the RSIC does not hold for RNR pivots as opposed to
CXS pivots because of contrasts with cases of CXS like (21).
(21)*Who1 did Mary buy t1 from Lois pictures of t1?
However, the well-formedness of (19) does not necessarily indicate a contrast between RNR and CXS with respect to
(18a). This is because one can allow for the contrast between (19) and (21) even assuming the

< previous page page_105 next page >


If you like this book, buy it!
< previous page page_106 next page >
Page 106
RSIC governs RNR pivots by localizing the difference between the two constructions in the domain of remnant
extraction. That is, one can say that an RNR pivot can be an L-extraction remnant but a CXS pivot cannot be. Although
this difference remains unexplained, these constructions are known to differ in other mysterious ways; for example, the
former can strand prepositions, the latter cannot.
What I hope to have shown is that given the possibilities in (17), which to my knowledge have never been seriously
challenged, facts like (14b) are far from yielding a strong argument for the nonextraction character of RNR
constructions. 7
4.2.1.2.7 Target Constituents
A sixth argument differentiating RNR from L-extractions might be based on facts McCawley (1988, 529) cites about
the scope of RNR drawn from observations by Bresnan (1974). These data show that RNR pivots can be types of
constituents that never function as binders for English L-extractions, for example, the types of constituents in (lb,c) and
(22).
(22) a. I know when t1 but I don't know where t1 [Amanda met Steve]1.
b. He suspects that the captain t1 but knows that the major t1 [detests goat cheese]1.
c. He wants her to t1 but is afraid to ask her to t1 [go to the prom]1.
Although McCawley does not formulate an explicit argument on this basis, the greater generality of RNR with respect
to target constituents might be taken to belie its unity with L-extractions.8 And although McCawley formulates no such
argument, a referee does. This is considered in appendix C.
But the difference in question shows only that RNR is distinct, not that it is fundamentally distinct in nature. Since it is
logically possible that some extractions are relevant to a wider range of constituents than others, the facts cited cannot
strongly indicate that RNR is not an extraction.9 Moreover, along the very parameter at issue, target scope, RNR and L-
extractions nonetheless share many restrictionsfor example, those in (23). ((23g) is from McCawley 1988, 529.)
(23) a. Tony may be t1 and Glen certainly is t1 [upset]1.
b. [Upset]1, Tony may be t1.

< previous page page_106 next page >


If you like this book, buy it!
< previous page page_107 next page >
Page 107
c.*Tony is slightly t1 and Fred is greatly t1 [upset]1.
d.*[Upset]1, Tony is slightly t1.
e.*Ellen objects to your t1 and Betty objects to my t1 [calling Isabelle]1.
f.*I said he objected to your calling her and [calling her]1, he did object to your t1.
g.*Ted has always wanted a t1 (and) so I've given him my t1 [coffee grinder]1.
h.*It was [coffee grinder]1 that Ted had always wanted a t1 and I decided to give him my t1.
4.2.1.2.8 Constraints on that Clauses
A seventh argument McCawley (1987, 188, n. 1) considers for distinguishing RNR from L-extractions via appeal to
multiple mothers is based on an observation by Grosu and Thompson (1977). These authors note that Ross's (1967)
constraint banning ''internal Ss," discussed and revised by Kuno (1973), is maintained in RNR structures.
(24) a.*Mary considers that three and three are seven to be definitely false.
b.*John considers that two and two are five t1 and Mary considers that three and three are seven t1 [to be definitely
false]1.
Such an argument depends on complicated and controversial assumptions about the nature of the violation in (24a). In
my own relational view of this constraint, the ungrammaticality of (24a) arises from the fact that, for example, that
clauses are only allowed to head a restricted class of final arcs; see note 10. One entailment of the appropriate condition
is that that clauses cannot head final 2-arcs. But, under a relational raising analysis of structures like (24a) (see Postal
1974; Johnson and Postal 1980; Postal and Pullum 1988), this condition is violated. Moreover, as RNR has no effect on
the final status of the that clauses in (24b), the relational condition applies as well to the RNR case. Therefore, if
anything like such a relational proposal is correct, facts like (24b) provide no real information about RNR and fail to
support McCawley's specific view about it. The relational condition is discussed further in section 4.2.1.3.2.
4.2.1.2.9. An Agreement Phenomenon
McCawley (1987, 187) offers what would be an eighth argument favoring his view, based on the agreement facts in
(25).

< previous page page_107 next page >


If you like this book, buy it!
< previous page page_108 next page >
Page 108
(25) a. Historical and scientific knowledge are/*is different in nature.
b. Thai and Burmese food are/*is quite similar.
He says that such examples "are evidently instances of RNR applying to a conjoined NP." But these cases are distinct
from those underlying the seven previous arguments, which all deal with instances of RNR involving full clauses.
Moreover, they lack the intonational characteristics of clausal RNR cases; there is no hiatus after the conjoined
adjectives. McCawley gives no argument that (25a,b) are instances of RNR. I suspect that instead they instantiate the
more general phenomenon once called conjunction reduction.
McCawley rightly focuses on the key feature of (25a,b): in the presence of specific conjoined modifiers, certain mass
nouns behave like plurals for agreement. But, contrary to his suggestion, this striking property cannot be attributed to
RNR surface structures manifesting multiple mother nodes, even if (25a,b) do instantiate RNR. The same agreement
property appears in, for example, (26a,b), which are entirely refractory to such an analysis.
(26) a. Logical and empirical truth are/*is necessary and contingent, respectively.
b. Thai and Burmese food are/*is light and greasy, respectively.
Such cases, called interwoven dependencies below, show that something more fundamental is going on than
McCawley's treatment allows for. See section 4.2.2.14.
4.2.1.3 Potential Arguments
4.2.1.3.1 Comments
This completes my survey of the grounds McCawley and others have recently given for a view of RNR that, among
other things, fundamentally differentiates it from L-extractions. However, three potential arguments for his position not
previously offered deserve consideration.
4.2.1.3.2 express-Class Verbs
The first argument derives from work by Grimshaw (1982), who attests English verbs that do not take postverbal that
clause complements but permit them as passive subjects. Jacobson (1992) points out that such clauses can also be
object-raising subjects with complements based on these verbs and cites Larry Horn's observation

< previous page page_108 next page >


If you like this book, buy it!
< previous page page_109 next page >
Page 109
that such clauses topicalize (see also Dowty and Jacobson 1988, 103; Hukari and Levine 1991, 116-117).
(27) a. This theory captures/expresses/reflects*(the fact) [that verbs exist].
b. [That verbs exist] is captured/expressed/reflected by this theory.
c. [That verbs exist]1 is impossible for such a theory to capture/express/reflect t1.
d. [That verbs exist]1, his theory fails to capture/express/reflect t1.
Observe, then, that RNR structures with such verbs cannot be based on that clauses.
(28) A good theory should capture/express/reflect t1 and would capture/ express/reflect t1
a.*[that verbs exist]1.
b. [the fact that verbs exist]1.
Compare (28) with (29), formed with verbs for which the analog of the starred form of (27a) is grammatical.
(29) A good theory should entail/explain t1 and would entail/explain t1
a. [that verbs exist]1.
b. [the fact that verbs exist]1.
The RNR restriction in (28a) would seem to support McCawley's multiple-mother-node view of RNR surface structures
and hence the claim that RNR is radically distinct from L-extractions. Under that view, (28a) would directly reflect the
restriction in the starred form of (27a). So (28a) would seem to justify McCawley's and Levine's view that RNR pivots
occur as surface constituents in the positions of (only) apparent RNR gaps.
However, the tempting argument just sketched has less force than it appears to. I suggest that the ill-formedness of
(28a) and the starred form of (27a) derives from a relational constraint on that clauses cited in section 4.2.1.2.8 in
connection with the seventh argument putatively sup-porting McCawley's position. By relational I refer to a condition
statable in the framework of Johnson and Postal 1980 and derivative work (e.g., Postal 1996). The correct formulation
of the condition relevant for that clauses remains problematic, largely because of issues related to subjects. I will
propose an elegant formulation, which, however, involves certain

< previous page page_109 next page >


If you like this book, buy it!
< previous page page_110 next page >
Page 110
no doubt controversial assumptions. To implement it, I assume that the class of Central-arcs includes those with the
Relational-signs (R-signs) in (30).
(30) Central-arc R-signs
subject {1} direct object {2} indirect object {3} subobject {4} semiobject {5} quasi object {6} chomeur {8}
extraposition {9}
Necessary to my formulation is a refinement of past relational approaches to extraposition, which all take extraposed
clauses to head 8-arcs. The refinement amounts to dividing the former collection of 8-arcs into two distinct sets, by
recognizing an additional relation called extraposition (R-sign = 9), whose extension is probably limited to extraposed
clausal constituents. This relation's extension covers at least the union of the relations COMP and XCOMP posited in
Bresnan 1982. In particular, then, extraposed that clauses would head 9-arc local successors of other Central-arcs,
specifically, l-arcs or 2-arcs. The that clause condition at issue then precludes that clauses from heading most types of
final Central-arcs and can be formulated roughly as in (31).
(31) The Complement Clause Condition
A final Central-arc headed by a that clause is a 9-arc. 10
Rule (31) claims among other things that that clause constituents cannot head final l-arcs, 2-arcs, 3-arcs, 5-arcs, 6-arcs,
or 8-arcs. However, consider (32), in which the complement might be assumed in relational terms to head both an
initial and a final 2-arc.
(32) He believes [that verbs exist].
Given the posit of 9-arcs, such examples can be grammatical despite (31) because they can be taken to instantiate the
extraposition option, with the associated extraposition expletive (it) null (see (37)). This allows the that clause in (32)
heading an initial 2-arc to head a final 9-arc (see Postal 1986a, sec. 3.1 and chap. 6).
Further, formulation (31) is intended to capture the distinction noted by Higgins (1973) and Kuno (1973) between such
cases as (33a) and (33b,c).
(33) a. That Tony is a spy is quite obvious.
b.*How obvious is that Tony is a spy?
c. Clear to everyone was*(the fact) that he was dying.

< previous page page_110 next page >


If you like this book, buy it!
< previous page page_111 next page >
Page 111
That is, a that clause can appear in so-called subject position only in structures not involving subject-auxiliary
inversion. My view of this is as follows. In the latter case, the "inverted" constituent heads a final and surface l-arc and
thus, if it is a that clause, violates (31). The problematic case is (33a), which logic determines must have an analysis in
which the complement does not head a final l-arc. This combines with the traditional relational idea that full clauses
require a final l-arc to indicate that cases like (33) involve an invisible final 1. I take this to be an RP, linked to the fact
that the clause has raised to a nontraditional relation called Prime, mentioned in Postal 1989, 124, n. 1. It must be
specified that, at least for clausal 1s, raising to this relation requires an invisible RP, which is then the final 1 of the
tensed clause, yielding consistency with (31). 11
In these terms, then, the difference between structures with and without subject-auxiliary inversion is among other
things the difference between structures that do not manifest raising to Prime and those that do. Further, the binary
"subject + VP" surface structure of tensed clauses is a function of raising to Prime, even though the initial structures
impose no such bracketing. Evidently, then, the standard order is "verb + subject." I should note that such an analysis
does not require that an RP obligatorily be linked to raising to Prime. That can be an option in generalthe option being
forced by principle (31) in the case where the I is a that clause, since without the RP the raised that clause would head
a final l-arc.
Additionally, (31) is intended to block all cases where a that clause would represent the head constituent of a PP. This
consequence of (31) is dealt with below in connection with specific examples; see the discussion of (41) and (46).
Turning to the starred form of (27a), one sees that an analysis in which the initial 2-arc headed by the that clause is a
final 2-arc violates condition (31). Why, though, is the extraposition analysis posited for (32) unavailable for the verb
class noted by Grimshaw (1982)? An informal answer is stated in (34).
(34) A lexically linked clausal constraint
Verbs in the class {capture, express,...} are incompatible with clausal extraposition structures.
In the terms of Postal 1986a, revised so that extraposition involves demotion to 9 rather than to 8, a more precise
version of (34) would say that a Predicate-arc headed by one of these verbs cannot have the same tail node as a 9-arc
(local successor of a 1-arc or 2-arc) headed by a that

< previous page page_111 next page >


If you like this book, buy it!
< previous page page_112 next page >
Page 112
clause. 12 Given (34), there is in effect no viable analysis of the starred form of (27a). If the that clause heads a final 2-
are, it violates general rule (31); if it heads a final 9-arc under extraposition, it violates the lexically linked (34).
The latter condition might, of course, be criticized as ad hoc. However, despite the acceptable passives in (27b), there
exists the possibly previously unnoticed contrast in (35).
(35) a. It is believed (by most experts) [that verbs exist].
b.*It was captured/expressed/reflected (by that theory) [that verbs exist].
Ill-formed expressions like (35b) motivate condition (34) independently of cases like (27). The RNR case in (28a) is
then ill formed even under the view that RNR is an extraction phenomenon. This follows because the complement
cannot licitly head either type of final arc generally available for a that-clause-initial 2. It cannot head a 2-arc because
of the general rule (31); it cannot head a 9-arc (associated with extraposition) because of the lexically restricted (34).
To complete the argument that (28a) fails to support McCawley's view of RNR or a claim that RNR is not an extraction
phenomenon, it remains to account for the grammaticality of L-extractions like the topicalization in (27d). Given (31),
logic requires that my description of this extraction not posit that the topicalized that clause head a final 2-arc. It also
cannot be taken to head a final 9-arc because of Higgins's (1973) generalization that extraposed clauses are not subject
to topicalization (see Sag and Klein 1982; Emonds 1985, 314-315; Postal 1986a, 99). This is motivated by
unambiguous extraposition structures like (36).
(36) a. It seemed to me/was believed by everyone [that verbs exist].
b.*[That verbs exist]1, it seemed to me/was believed by everyone t1.
But an account of why topicalization is nonetheless possible in (27d) can appeal to the view that English NP
topicalization is always linked to an (invisible) RP in the extraction site, a view already given substantial support in
chapter 2, where topicalization was shown to be a B-extraction13 (see also section 4.2.1.3.3). More precisely, if, in a
constituent C, the last Central-arc headed by a topicalized phrase is an w-arc, then an invisible RP heads a final w-arc
in C. In (27d), then, the topicalized clause that heads an initial 2-arc but not a final 9-arc linked to extraposition
nonetheless does not, as might otherwise be expected, head a final 2-arc. This is

< previous page page_112 next page >


If you like this book, buy it!
< previous page page_113 next page >
Page 113
because topicalization determines the existence of an invisible RP, which heads the final 2-arc in the basic clause. In
these terms, the structure of (27d) would include the elements shown in (37). In such diagrams, broken arrows
represent the Sponsor relation, double arrows the Erase relation. An arc at the point of an Erase arrow is erased,
entailing, among other things, that it is not part of the surface representation.
(37) Partial structure of (27d)

Here B, although an initial 2-arc of node 100, is not final; rather, C (headed by an RP) is. It should be stressed that the
contrast between ungrammatical RNR in (28a) and grammatical topicalization in (27d) depends not only on the claim
that the latter involves invisible RPs, but also on the view that the former does not.
The idea that cases like (32) are grammatical because they involve a masked extraposition structure receives some
support from facts related

< previous page page_113 next page >


If you like this book, buy it!
< previous page page_114 next page >
Page 114
to those cited in section 4.2.1.2.8. There I considered an argument linked to examples like (24) mentioned by
McCawley (1987). It involved an interaction between RNR and the "internal S" constraint. Consider then:
(38) a.*Ted considers that three and three are seven to be unimportant.
b.*Ted considers to be unimportant t1 but Frank considers to be crucial t1 [that three and three are seven]1.
Unlike (24b), (38b) involves an RNR pivot that is itself a that clause. Expression (38a) can be taken to violate principle
(31) under a (relational) raising-to-object analysis of such infinitival clauses, for under this analysis the that clause
heads a final 2-arc. The same assumptions would block (38b) as well, if an extraposition analysis of that construction
were unavailable. But a combination of raising to object plus extraposition in English in general yields an output with a
visible expletive itand when that is present, the analogs of both (38a) and (38b) are grammatical.
(39) a. Ted considers it to be unimportant that three and three are seven.
b. Ted considers it to be unimportant t1 but Frank considers it to be crucial t1 [that three and three are seven]1.
Contrastively, if the that clause in a case like (38b) is topicalized, the result is well formed.
(40) [That three and three are seven]1, Ted considers to be unimportant t1 (but Frank considers to be crucial t1).
This is so under my present assumptions because the that clause then does not head a final 2-arc, given that
topicalization, unlike RNR, is linked to RPs, as shown in (37). The final 2 in each clause would be the invisible RP.
4.2.1.3.3 that Clause Extractions
A second observation not previously discussed in the literature on the character of RNR might also seem to support
McCawley's position in a way parallel to the argument just analyzed. L-extractions of that clauses in contexts like those
of (41) must strand prepositions (see Kaplan and Bresnan 1982, 242; Postal 1994). But RNR in such cases cannot do so.
(41) a. I am no longer certain /* of that Nancy is an extraterrestrial.
b. [That Nancy is an extraterrestrial]1, I am no longer certain*/of t1.
c. Frank may be certain /* of t1 and should be certain /* of t1 [that Nancy is an extraterrestrial]1.

< previous page page_114 next page >


If you like this book, buy it!
< previous page page_115 next page >
Page 115
McCawley's view of RNR would correctly block the starred version of (41c) on the basis of the known constraint
operative in the starred variant of (41a). Moreover, it would correctly predict that the pattern of
grammaticality/ungrammaticality in (41c) reverses when the fact is inserted before the that clause, on the basis of the
contrast between the starred version of (41a) and (42).
(42) I am no longer certain of the fact that Nancy is an extraterrestrial.
However, as with the previous potential argument, the contrast between (41b) and (41c) arguably depends not on the
fact that topicalization is an extraction and RNR is not, but partly on the fact that English topicalization is inherently
linked to (invisible) RPs, whereas RNR is not. The first of these claims, previously appealed to in my treatment of
constructions with the verbs discussed by Grimshaw (1982), was supported in chapter 2, since at issue is, of course, the
view that topicalization is a B-extraction.
To understand the topicalization/RNR contrast, consider, for example, the "coat of paint" position in contexts like
(43a). This is an AC; that is, it cannot be filled by a weak definite pronoun. Therefore, although the position is not
incompatible with L-extractions, it is incompatible with topicalization of otherwise topicalizable NPs, because
topicalization is a B-extraction.
(43) a. Francine gave the wall the last coat of paint.
b. What1 Francine gave the wall t1 was its final coat of paint.
c.*Francine gave the wall it.
d. [Some other coat of paint]1, I would have applied t1 more rapidly.
e.*[Some other coat of paint]1, I alone would have given the wall t1.
A similar pattern occurs with color-designating phrases in contexts like (44a-c).
(44) a. Frank painted his house(s) green/that color.
b. Frank painted it/them green/that color.
c.*Frank painted his houses/them it.
d. [Some other color]1, I would have agreed to t1.
e.*[Some other color]1, Frank painted them t1 yesterday.
Examples (43e) and (44e) are typical of those justifying the view that

< previous page page_115 next page >


If you like this book, buy it!
< previous page page_116 next page >
Page 116
topicalization is linked to RPs, because that assumption permits reduction of the ill-formed extraction cases to the
independently needed restrictions applying in (43c) and (44c). Without the RP assumption, the ungrammaticality of
(43e) and (44e) is an unexplained anomaly.
Next consider RNR cases based on parallel extraction sites.
(45) a. Francine may have given the south wall t1 and she certainly gave the north wall t1 [its last coat of paint]1.
b. Allan may have painted his barn t1 and certainly did paint his house t1 [a horrible shade of bright pinkie.
By the logic applied to (43e) and (44e), the grammaticality of (45a,b) supports the view that apparent RNR gaps do not
(obligatorily) represent invisible RPs. If they did, the restriction applying in (43c) and (44c) would block (45a,b). Thus,
both elements of the earlier account of the contrast between (27d) and (28a) can be motivated.
However, the joint claims that topicalization involves RPs but RNR does not yield only part of a description of more
complex paradigms like (41). A full treatment must deal with alternations like the ones in (46) and account for the
possibility of stranded prepositions in, for example, (41b) and their impossibility in (41c).
(46) a. I am certain*/ of that.
b. I am certain /* of that Nancy is an extraterrestrial.
My assumptions extend those made in Postal 1986a. First, I take the possibility of the good form of (46b) to be a
function of extraposition, that is, in presently revised terms, 9-arc local successors. This is in effect the only analysis
consistent with rule (31). In the terms used in Postal 1986a, extraposition always involves an expletive nominal, but in
these constructions, as in (32), it is required to be null. Grammatical RNR cases like (41c) then involve extraposed
clauses. If the grammatical variant of (46b) is good because the clause is extraposed and heads a final 9-arc in accord
with principle (31), the ungrammatical variant is bad because no extraposition exists and the clause heads the same kind
of final arc as nonclausal arguments of the predicate. Let us assume nonessentially that the relevant final relation is the
Central relation quasi object (R-sign = 6) described in Postal 1990a. Then the ill-formed version of (46b) would be
blocked by rule (31) on the assumption that the relational structure of the prepositional phrase (PP) is like (47).

< previous page page_116 next page >


If you like this book, buy it!
< previous page page_117 next page >
Page 117
(47) Partial structure of (46b)

This violates (31) since the that clause heads a final Central-arc C, a non-9-arc. Structure (47) differs from relational
descriptions of PPs like that in Johnson and Postal 1980 in eliminating the relation Marquee and in marking those arcs
previously taken to be Marquee-arcs with the same Central R-signs as their predecessors. If previous representations
were maintained, (31) would fail to block the starred version of (46b)more generally, would fail to impose the
exceptionless incompatibility of that clauses with PP head status.
Next consider (41c). Under the assumption that RNR is not linked to RPs, it must be explained why the preposition-
stranding version of (41c) is ill formed, in contrast to that of (41b). An analysis parallel to that given for the contrast
between (28a) and (27d) is clearly desirable and is available under certain (very) nonstandard general assumptions
about extraction-linked preposition stranding. These assumptions deserve extensive discussion, but here I will only
sketch the relevant view (see Postal 1991). The basic idea is that preposition stranding involves invisible RPs, which,
however, are the result not of extraction but of demotion. 14 I am uncertain what to call the relation to which phrases
demote in this case,

< previous page page_117 next page >


If you like this book, buy it!
< previous page page_118 next page >
Page 118
but I take the relevant R-sign to be 10. Crucially, like 2-arcs and 6-arcs, 10-arcs are Central-arcs.
I take it as a brute fact about English that where a phrase advances to 10 from some relation to, an RP must head an w-
arc in the same clause. Since w is, like the relation 6 of (47), always the sort of relation that determines a PP structure
(whereas 10 is like 1, 2, 9, etc., in not deter-mining such a structure), the RP would, other things being equal, then be
expected to show up as the (reflexive) head of a PP. But English requires in addition that this reflexive RP be null.
Moreover, it is true, at least for English, that the head of a 10-arc A must always be extracted; technically, A must have
a foreign successor. 15 That is why (48), for example, is ungrammatical.
(48)*They talked the presidential race about.
Hence, a stranded-preposition extraction structure will have the general form (49).
(49) Stranded-preposition extraction structure

< previous page page_118 next page >


If you like this book, buy it!
< previous page page_119 next page >
Page 119
Under this analysis, what appears to be extraction of the head of a PP is actually extraction of an element advanced to a
fixed relation 10 that antecedes the null reflexive RP head of a cooccurring pp. 16 Since 10-arcs are Central-arcs but
not 9-arcs, rule (31) determines that a that clause cannot head a final 10-arc. This yields an account of why the
preposition-stranding versions of (41c) are ill formed. The licit versions involve the extraposition type of demotion,
yielding that clauses heading final 9-arcs in accord with (31). But the stranding versions could only involve
advancement of the that clauses to 10, as in (49), leaving them heading final 10-arcs in violation of rule (31). In
contrast, topicalizations like (41b) are grammatical, because of the RP-linked feature of topicalization. This determines
that although the that clause in (41b) heads a 10-arc, that arc is not final, as in (37). Rather, the invisible RP heads the
final 10-arc, so that, unlike the stranding version of (41c), (41b) satisfies constraint (31).
4.2.1.3.4 Two Types of infinitive
A third potential basis of support for a position like McCawley's merits brief consideration. Elsewhere, following an
observation by Howard Lasnik, I noted that, in contrast to NPs in contexts like (50ai), those in (50aii) are not subject to
CXS (see Postal 1974, 92-93, 406-411). This correlates with the fact that the NP in (50ai) is passivizable and the one in
(50aii) is not.
(50) a. Steve i. believes/ii. wants the visitor to be alert.
b. Steve i. believes/ii.*wants t1 to be alert [the attractive visitor from the dark galaxy]1.
c. The visitor was i. believed/ii.*wanted to be alert by everyone.
However, either structure permits L-extraction.
(51) [Which visitor]1 did Steve i. believe/ii. want t1 to be alert?
The latter fact might suggest a further potential argument for a position about RNR like McCawley's since, as perhaps
not previously noted, a constraint parallel to the one that holds in (50) holds for RNR.
(52) Steve may have i. believed/ii.*wanted t1 to be alert and probably did believe/want t1 to be alert [the attractive
visitor from the dark galaxy]1.
The contrast between (51ii) and (52ii) might be invoked to argue that RNR is not an extraction phenomenon.
However, such a conclusion is very weak. Even under the view that RNR is an extraction, the constraint can be stated
as blocking

< previous page page_119 next page >


If you like this book, buy it!
< previous page page_120 next page >
Page 120
R-extractions. Given the passive constraint operative in (50cii), one possible approach would link the structure of
(50aii) to that of cases like (53), for the post-for NP here is also not subject to passivization, CXS, or RNR.
(53) I would prefer for the visitor to be alert.
A natural suggestion would then be that, contrary to the view I took in Postal 1974, unlike (50ai), (50aii) does not
involve main-clause constituency (raising) for the postverbal NP, which is then regularly not subject to passivization.
RNR and CXS would be blocked because these are in general impossible for (unraised) final subjects. A quite different
approach, which I currently believe to be far more plausible, would recognize raising in both structures of (50a), as in
Postal 1974. But the resulting main-clause constituents would be distinguished because in the want structure, but not the
believe structure, the raising-induced object is associated with a kind of abstract quirky case marking, probably
nominative. To make such an analysis work, it is then necessary to view passivization, CXS, and RNR (but not
standard L-extractions) as incompatible with the quirky marking in question. This is not the place to seriously develop
these ideas. In any event, the multiple-mother-node analysis provides no account for (52ii) since it claims in effect that
the surface structure of that example is essentially like that of the perfectly grammatical (50aii). So far, then, the
contrast between (51) and (52ii) appears to show very little about the issues with which this study is concerned.
I conclude that none of the three potential arguments just discussed really offers any support for a view of RNR like
McCawley's. Although the cases do reveal contrasts between RNR and topicalization, the first two can ultimately be
argued to follow from the fact that the latter is linked to RPs and the former is not. This independently justifiable
distinction interacts with (a) the relational characterization of the limitation on verbs like capture in (34), (b) a
relational view of the basic clausal constraints on that clauses first noticed by Ross (1967) and stated in (31), and (c) a
relational view of preposition stranding, to yield the contrasting facts in fairly general ways. The third type of contrast
could well be due to the way a contrast between the presence and absence of subject raising with different types of
infinitival complement interacts with an independent ban on RNR of final subjects. Alternatively, it could be due to a
contrast in abstract case marking. Beyond suggesting that these potential arguments are not real arguments for the sort
of position McCawley has

< previous page page_120 next page >


If you like this book, buy it!
< previous page page_121 next page >
Page 121
advocated, this discussion underlines the general point that even quite sharp real contrasts between RNR and L-
extractions can fail to argue that RNR is not an extraction phenomenon.
4.2.1.4 Summary
I have shown that the positive evidence that McCawley and others have provided for the position that RNR is
fundamentally distinct from extraction phenomena is at best much less than has been assumed and that certain apparent
evidence for that assumption is in fact not evidence for it. Moreover, McCawley's accounts of RNR, like that of Levine
(1985), fail to consider whether there is evidence that shows, to the contrary, that RNR and L-extractions are quite
similar in various respects. I take up this issue in the following section.
4.2.2 Common Properties of Right Node Raising and L-Extractions
4.2.2.1 Comments
Support for a position that recognizes the essential similarity of L-extractions and RNR is readily available.
Considerable evidence favors the view that RNR is an extraction phenomenon. This evidence consists of myriad
restrictions that have been, or can be, shown to hold for L-extractions and that also constrain the formation of RNR
constructions.
4.2.2.2 The Coordinate Structure Constraint
Implications that RNR and L-extractions are clearly similar have been neglected even when evidence for such
similarity has been given by advocates of a distinct view of RNR. For instance, McCawley (1982, 101, n. 11) observes
that the formation of RNR constructions is sensitive to Ross's (1967) CSC, as previously observed in effect by Wexler
and Culicover (1980, 302). But the CSC is, of course, a fundamental condition on L-extractions, as shown in chapter 3.
McCawley states:
(54) ''The CSC does, however, rule out applying RNR to material that is in coordinate constituents of the conjuncts of a
coordinate structure." (McCawley 1982, 101, n. 11)
And he gives an example that illustrates this nicely.
(55) Tom is writing an article on Aristotle and Freud, and Elaine has just published a monograph on Mesmer and Freud
*Tom is writing an article on Aristotle t1 and Elaine has just published a monograph on Mesmer t1 [and Freud]1.

< previous page page_121 next page >


If you like this book, buy it!
< previous page page_122 next page >
Page 122
*Tom is writing an article on Aristotle and t1 and Elaine has just published a monograph on Mesmer and t1 [Freud]1.
Here RNR seems to obey the aspect of the CSC that has been called the Conjunct Constraint (see Grosu 1972; Pollard
and Sag 1994; and chapter 3). Namely, an RNR pivot cannot be linked to a conjunct, regardless of whether the
associated conjunction occurs on the pivot.
Example (56) illustrates that an RNR pivot also cannot be linked to a piece of a conjunct except in the same ATB
fashion (relating to one element from each conjunct) that L-extractions can.
(56) Tom may have bought sketches of Gail and photos of Louise and Bob saw Louise RNR
*Tom may have bought sketches of Gail and photos of t1, and Bob saw t1 Louise1.
Hence, RNR cases also obey what has been called the Element Constraint and thus are governed by both aspects of the
original CSC.
The CSC facts could be taken to yield a kind of standoff with the argument of section 4.2.1.2.5, which sought to
differentiate RNR from L-extractions on the grounds that only the former fails to obey the island condition defined by
relative clauses. The argument is at best neutralized by the observation that RNR is controlled by the CSC; for under
McCawley's view, both aspects of the CSC must control radically different grammatical phenomenaextractions and the
nonextraction phenomenon he takes RNR to be.
4.2.2.3 The Indirect Object Constraint
A constraint noted in the 1960s and valid for many types of English involves traditional indirect objects (IOs) 17 (see
Fillmore 1965; Kuroda 1968; Hankamer 1973; Langendoen, Kalish-Landon, and Dore 1974; Oehrle 1975, 236-237;
Culicover 1982, 337). It is illustrated in (57).
(57) a. Ernest sold drugs to Lydia.
b. Ernest sold Lydia drugs.
c. [Which hostess]1 did Ernest sell drugs to t1?
d.*[Which hostess]1 did Ernest sell t1 drugs?
e. [Which drugs]1 did Ernest sell Lydia t1?
(58) a. The director handed the report to the consultant.
b. The director handed the consultant the report.
c. It was that consultant who1 the director handed the report to t1

< previous page page_122 next page >


If you like this book, buy it!
< previous page page_123 next page >
Page 123
d.*It was that consultant who1 the director handed t1 the report.
e. It was that report which1 the director handed the consultant t1.
One might formulate this constraint informally as in (59).
(59) The Indirect Object Constraint (first version)
An 10 cannot be L-extracted.
This covers cases like (57) and (58) properly. To be viable, object-raising, object deletion, and P-gap structures must
then involve L-extractions, as these also manifest the effect.
(60) a. Martin1 is hard for us to sell things to t1.
b.*Martin1 is hard for us to sell t1 things.
c. Martin1 is too poor for us to sell things to t1.
d.*Martin1 is too poor for us to sell t1 things.
e. It was Martin who1 I hired t1 after selling things to t1.
f.*It was Martin who1 I hired t1 after selling t1 things.
However, such a conclusion is already fairly well established in at least a good part of the literature (see, e.g., Chomsky
1977b, 1981, 1982, 1986a; Browning 1987a,b). Under the presumbably uncontroversial view that passivization does not
involve extraction, (59) is correctly not incompatible with the passivization of some IOs.
(61) Joanne was handed the tragic telegram by the secretary.
Moreover, (59) does correctly block extraction of the IO from a passive for those who, as I do, allow passives like
(62a).
(62) a. The telegram was handed Joanne by the secretary.
b.*[Which employee]1 was the telegram handed t1 by the secretary?
However, (59) fails to predict that the Indirect Object Constraint governs RNR.
(63) a. I first offered apples to t1 and then sold peaches to t1 [the immigrant from Paraguay]1.
b. I first offered apples t1 and then sold peaches t1 [to the immigrant from Paraguay]1.
c.*I first offered t1 apples and then sold t1 peaches [the immigrant from Paraguay]1.
In (63c) the RNR pivot corresponds to an IO in each conjunct. 18 Ungrammaticality also results when the pivot
corresponds to an IO in (either) one and a non-IO in others.

< previous page page_123 next page >


If you like this book, buy it!
< previous page page_124 next page >
Page 124
(64) a.*I first offered apples to t1 and then sold t1 peaches [the immigrant from Paraguay]1.
b.*I first offered t1 apples and then sold peaches to t1 [the immigrant from Paraguay]1.
A unified formulation of the Indirect Object Constraint that predicts its application in the full range of structures where
it actually holds then be at least as general as (65).
(65) The Indirect Object Constraint (second version)
An IO cannot be extracted.
For (65) to correctly predict the ungrammaticality of (63), (64b) and the like, RNR must be regarded (at least in part) as
an extraction phenomenon. 19
4.2.2.4 The Genitive Constraint
There is a constraint precluding the L-extraction of genitive-marked NPs or of the PPs that contain them. To avoid
questions of more general limitations on extraction of elements on left branches, I restrict the discussion exclusively to
post-of genitives. Consider:
(66) a. She saw several children of Ted's.
b. She saw several pictures of Ted.
c.*It was Ted's1 that she saw several children of t1.
d. It was Ted1 that she saw several pictures of t1.
e.*It was [of Ted's]1 that she saw several children t1.
f. ?It was [of Ted]1 that she saw several pictures t1.
Call the restriction that blocks (66c,e) the Genitive Constraint. Like the Indirect Object Constraint, this restriction also
manifests itself in object-raising, object deletion, and P-gap cases, with the same implication.
(67) a.*Ted('s)1 is hard to find children of t1.
b. Ted1 is hard to find pictures of t1.
c.*Ted('s)1 is too remote to find children of t1.
d. Ted1 is too remote to find pictures of t1.
e.*It was Ted('s)1 that they arrested t1 after finding children of t1.
f. It was Ted1 that they arrested t1 after finding pictures of t1.
These facts could lead to the following informal formulation:
(68) The Genitive Constraint (first version)
A constituent of the form [(of) NP's] cannot be L-extracted.

< previous page page_124 next page >


If you like this book, buy it!
< previous page page_125 next page >
Page 125
Notably, parallel genitive restrictions manifest themselves in RNR structures.
(69) a. Glen was looking for photos of t1, but only found sketches of t1 [Ted and Alice]1.
b. Glen was looking for photos t1 but only found sketches t1 [of Ted and Alice]1.
c.*Glen was looking for nieces of t1 but only found cousins of t1 [Ted and Alice's]1.
d.*Glen was looking for nieces t1 but only found cousins t1 [of Ted and Alice's]1.
So (68) must be generalized.
(70) The Genitive Constraint (second version)
A constituent of the form [(of) NP's] cannot be extracted.
As before, such a generalization makes sense only if both L-extractions and RNR instantiate some common linguistic
phenomenon called "extraction."
4.2.2.5 The First Reflexive Constraint
Several authors observe a limitation on PP antecedents of reflexives (see Postal 1974, 275-276, n. 5; Chomsky and
Lasnik 1977, 485, n. 106; Van Riemsdijk and Williams 1986, 203; Baltin and Postal 1996).
(71) a. I spoke to Lydia about herself.
b. [Which woman]1 did you speak to t1 about herself?
c.*[To which woman]1 did you speak t1 about herself?
d. the woman who1 I spoke to t1 about herself
e.*the woman [to whom]1 I spoke t1 about herself
One might formulate this limitation as follows:
(72) The First Reflexive Constraint
If D is a PP whose head NP is the antecedent of a reflexive NP head of a PP, then A cannot be extracted (although its
subconstituent NP can be).
The same patterns hold in RNR structures.
(73) a. He might have spoken to t1 about herself1 and certainly wrote to t1 about herself1 [the angry candidate whose
interview went so badly]1.

< previous page page_125 next page >


If you like this book, buy it!
< previous page page_126 next page >
Page 126
b.*He might have spoken t1 about herself1 and certainly wrote t1 about herself1 [to the angry candidate whose
interview went so badly]1.
Therefore, if (72) is to cover all the facts, RNR constructions must on the basis of this evidence also be regarded as
extractions.
4.2.2.6 The Second Reflexive Constraint
Certain instances of reflexive pronouns not corresponding to independent logical elements are traditionally referred to
as inherent. Such reflexives, found in (74), are distinct from the ordinary reflexives in (75).
(74) a. Amanda perjured herself.
b. Errol devoted himself to his sheep.
c. Lois never exerts herself.
(75) a. Amanda criticized herself.
b. Errol described himself to the caller.
c. Lois works herself very hard.
Notably, unlike ordinary reflexives, inherent reflexives cannot be L-extracted.
(76) a. Herself1, Lois criticized t1/described t1 to the caller/works t1 very hard.
b.*Herself1, Lois perjured t1/devoted t1 to her sheep/never exerts t1.
(77) a. It was herself1 that Lois criticized t1/described t1 to the caller/works t1 very hard.
b.*It was herself1 that Lois perjured t1/devoted t1 to her sheep/never exerts t1.
The same restriction distinguishes inherent from ordinary reflexives as RNR pivots.
(78) a. Lois may have criticized t1 and should have criticized t1 herself1.
b.*Lois may have perjured/exerted t1 and should have perjured/exerted t1 herself1.
Arguably, then, the constraint in question holds for extraction in general, a notion that must, again, cover both L-
extractions and RNR.
4.2.2.7 Stranded Prepositions
Ross (1967) observes that, like L-extractions, English RNR can strand prepositions. ((79a) is from Ross 1967, 141, and
(79d) from Rodman, 1972, 79.)

< previous page page_126 next page >


If you like this book, buy it!
< previous page page_127 next page >
Page 127
(79) a. I am confident of t1 and my boss depends on t1 [a successful outing ...]1.
b. Frank talked to t1 and Martin talked about t1 [the visitors from Andromeda]1.
c. Ellen argued for t1 and Gwen argued against t1 [the proposal to fire all the workers]1.
d. I skirted around t1 but Guinevere ran into t1 [the bush behind the girls' gym]1.
This situation does not in itself argue for the unity of RNR and L-extraction, since the facts would follow as well from
the sort of structures posited in McCawley's proposal about RNR.
However, an argument against the latter resides in the fact that many of the conditions on preposition stranding under
RNR are identical to those on preposition stranding under L-extractions. Rodman (1972) notes that different cases have
to be distinguished with respect to L-extractions.
(80) Typology of preposition stranding under L-extractions
Instances in which stranding is (a) required, (b) blocked, (c) optional.
Case (80a) is illustrated in (81) and (82).
(81) a. They scoffed at somebody.
b. Who(m)1 did they scoff at t1?
c.*[At whom]1 did they scoff t1?
(82) a. She made light of the problems facing the PTA.
b. the problems which1 she made light of t1
c.*the problems [of which]1 she made light t1
Other forms given by Rodman (1972, 82, 84) as also requiring stranding include account for, insist on, and keep track
of Construction of relevant examples is left to the reader.
Case (80b) is illustrated in (83) and (84).
(83) a. Jerome tickled Marsha in that way.
b.*[What way]1 did Jerome tickle Marsha in t1?
c. [In what way]1 did Jerome tickle Marsha t1?
(84) a. Ernie did it for someone else's sake.
b.*[Whose sake]1 did Ernie do that for t1?
c. [For whose sake]1 did Ernie do that t1?
Finally, case (80c) is exemplified in (85).

< previous page page_127 next page >


If you like this book, buy it!
< previous page page_128 next page >
Page 128
(85) a. He discovered the troll under that bridge.
b. [What bridge]1 did he discover the troll under t1?
c. [Under what bridge]1 did he discover the troll t1?
Notably, the typology in (80) holds for corresponding RNR forms. Cases that require, preclude, or optionally allow
stranding under L-extraction respectively require, preclude, or optionally allow it under RNR.
(86) a.*Jane could have scoffed t1 and should have scoffed t1 [at that idea]1.
b. Jane could have scoffed at t1 and should have scoffed at t1 [that idea]1.
c.*They could have made light t1 and should have made light t1 [of the difficulties in question]1.
d. They could have made light of t1 and should have made light of t1 [the difficulties in question]1.
e.*Jerome may have tickled Marsha in t1 and certainly should have tickled her in t1 [the way that I told you]1.
f. Jerome may have tickled Marsha t1 and certainly should have tickled her t1 [in the way that I told you]1.
g.*Ernie may have done it for t1 and certainly should have have done it for t1 [somebody else's sake]1.
h. Ernie may have done it t1 and certainly should have done it t1 [for somebody else's sake]1.
i. They might have discovered the troll under t1 and should have discovered the troll under t1 [the bridge near the
falls]1.
j. They might have discovered the troll t1 and should have discovered the troll t1 [under the bridge near the falls]1.
Certain cases are particularly damaging to McCawley's conception of RNR. In (86a,c) RNR gaps corresponding to PPs
are ill formed, just as the corresponding L-extraction structures are. If RNR is an extraction phenomenon, this could
follow from a uniform constraint of the schematic form (87), independently motivated by, for example, (81c) and (82c).
(87) PPs of type X cannot be extracted.
But in McCawley's terms, (87) would not predict the ungrammaticality of (86a,c) since internal to his proposal, nothing
has been extracted. Similarly, focus on (86e,g), where RNR gaps corresponding to heads of PPs are disallowed, just as
in the corresponding L-extractions. If RNR is an

< previous page page_128 next page >


If you like this book, buy it!
< previous page page_129 next page >
Page 129
extraction phenomenon, this could follow from a uniform restriction of schematically the form (88), also independently
motivated by, for example, (83b) and (84b).
(88) Heads of PPs of type Y cannot be extracted.
But in McCawley's terms, (88) could not cover (86e,g), which would not involve extraction. More precisely, in terms of
multiple-mother-node structures like (4), such cases would not actually contain stranded prepositions. For, as (3) makes
explicit, under the nonextraction view, RNR does not affect constituency relations. Thus, the parallelism between, for
example, (84b) and (86g) would be accidental.
4.2.2.8 Exceptives
As in section 2.2.10, I refer to phrases like those italicized in (89) as exceptives.
(89) a. Terry invited nobody but Bob.
b. They spoke to everybody other than Franklin.
c. The nurse watched everything except that operation.
d. They talked to everyone except to Wanda.
Exceptives form islands for L-extractions.
(90) a.*It was Bob1 that Terry invited nobody but t1.
b.*Who1 did they speak to everybody other than t1?
c.*[That operation]1, I am sure that the nurse watched everything except t1.
d.*It was [to Wanda]1 that they talked to everyone except t1.
And exceptives equally resist the formation of RNR structures.
(91) a.*Terry may have invited nobody but t1 and should have invited nobody but t1 [Bob and his crew]1.
b.*They may have spoken to everybody other than t1 and should have spoken to everybody other than t1 Franklin1.
c.*The nurse wanted to watch everything except t1 and did watch everything except t1 [the most difficult and lengthy
operation of the day]1.
d.*Harry should have talked to everyone except t1 and probably did talk to everyone except t1 [to Wanda]1.
Once more, under McCawley's proposal in (3), the ungrammaticality of (91) is anomalous and fails to be accounted for
by the constraint that blocks the L-extractions in (90).

< previous page page_129 next page >


If you like this book, buy it!
< previous page page_130 next page >
Page 130
4.2.2.9 Another Preposition-Stranding Fact
McCawley's position is that analyzing RNR as a nonextraction phenomenon manifesting multiple-mother surface
structures explains otherwise anomalous proper-ties of RNR cases. The overall claim is that sharp contrasts exist
between RNR properties and L-extraction properties. This yields incorrect predictions relating to another aspect of the
interaction of preposition stranding and RNR. Consider (92).
(92) a. Ted offered apples to h and actually gave peaches to t1 [the lovely young secretary]1.
b. Ted offered apples t1 yesterday and actually gave peaches t1 today [to the lovely young secretary]1.
c. Who1 did Ted offer apples to t1 yesterday and actually give peaches to t1 today?
d.*Who1 did Ted offer apples t2 yesterday and actually give peaches t2 today [to t1]2?
Under McCawley's conception of RNR, an RNR pivot is a surface constituent of all the preceding conjuncts. Under that
view, the L-extracted NP in (92d) has in effect been extracted out of dative to phrases like those in (92c). The multiple-
mother-node view then provides no account for the ungrammaticality of (92d) and its contrast with (92c). More
generally, this contrast illustrates a clear way in which an RNR pivot fails to act as if it were a surface constituent of n
preceding conjuncts.
Under an extraction view of RNR, the restriction manifest in (92d) can presumably be identified in part with the
independently motivated RSIC mentioned in section 4.2.1.2.6. If McCawley's view is rejected, one of the two possible
analyses of the RNR/L-extraction interaction discussed in that section would seem to reduce the principle operative in
(92d) to the one operative in non-RNR cases like (93d), and hence independently motivated; see (93e).
(93) a. I am looking forward to both events with great anticipation.
b. I am looking forward t1 with great anticipation [to both events]1.
c. [Which events]1 are you looking forward to t1 with great anticipation?
d.*[Which events]1 are you looking forward t2 with great anticipation [to t1]2?
e.*[What kind of gorillas]1 does Mary sell t1 to tourists [pictures of t1]2?

< previous page page_130 next page >


If you like this book, buy it!
< previous page page_131 next page >
Page 131
That is, the relevant principle would Block the analysis in which L-extraction operates on a predefined RNR structure,
since RNR pivots, like CXS ones, were claimed to be (locked) islands.
It remains to explain why (92b) is also ill formed (in contrast to, e.g., (19)) on the other analysis, under which RNR
targets are L-extraction remnants. Evidently, to complete the account, in the specific case of PP constructions with
stranded Ps, some principle would have to block RNR from taking that sort of L-extraction remnant as a target, even
though in general, L-extraction remnants were claimed to be allowed as RNR targets. I believe what is involved relates
to the nonstandard account of preposition stranding under extraction sketched in section 4.2.1.3.3. Under that proposal,
such ''remnant" PPs are not actually directly formed by extraction. Thus, it is not unexpected that the freedom for RNR
targets to be L-extraction remnants would not extend to them. But a serious development of this idea is not possible
here.
4.2.2.10 Defective Paradigms
Kayne (1984, 5) observes the following contrast:
(94) a. John, who1 I assure you t1 to be the best ,...
b.*I assure you John to be the best, ....
That is, in this case the complement subject can be L-extracted but cannot appear in the surface object position. In
Postal 1993c I argue that the restriction in (94b) is a special case of the Derived Object Constraint (Postal 1974),
defined on the basis of cases like (95a,b).
(95) a. Franks, they alleged t1 to be a pimp.
b,*They alleged Frank to be a pimp.
Notably, in Postal 1993c I observe, as neither I (Postal 1974) nor Kayne (1984) did earlier, that the extraction required
to save (94a) and (95a) from the constraint in (94b) and (95b) can also be that associated with RNR.
(96) a. They might have alleged t1 to be pimps and probably did allege t1 to be pimps [all of the Parisians who the CIA
hired in Nice]1.
b. I can assure you t1 to be one of the world's ten best cars and hereby do assure you t1 to be one of the world's ten best
cars [the 1992 model De Soto that you see standing in front of you]1.
Since this phenomenon both groups RNR with L-extractions and differentiates RNR pivots from unextracted NP
positions, it provides an additional argument that RNR is a true extraction.

< previous page page_131 next page >


If you like this book, buy it!
< previous page page_132 next page >
Page 132
4.2.2.11 Quantificational Phrases
The phrases containing NPs and PPs associated with, and specifying the class of elements quantified over by,
quantifiers like all and none are islands for L-extractions.
(97) a. They interrogated none of the victims.
b. They claimed you interrogated all (of) the perpetrators.
c.*[Which victims]1 did they interrogate none of t1?
d.*[Of which victims]1 did they interrogate none t1?
e.*the perpetrators who1 they interrogated all (of) t1
f.*the perpetrators [of whom]1 they interrogated all t1
Parallel constraints hold for RNR.
(98) a.*He should have interrogated all h and did interrogate all t1
[(of) the victims of the terrorist attack]1.
b.*I told him not to interrogate any of t1 and he didn't interrogate any of t1 [the perpetrators of the bloody bomb
attack]1.
Again, a view that radically distinguishes the nature of RNR from that of L-extractions evidently cannot capture the
parallelism between (97c,d,e f) and (98a,b).
4.2.2.12 Adjectival Complements
Adjectives of the class evil, nice, wonderful, and so on, permit neither their complement PPs nor the head NPs of those
PPs to be L-extracted (see Stowell 1991).
(99) a. It was evil/nice of Marilyn to do that.
b.*the woman [of whom]1 it was evil/nice t1 to do that
c.*the woman who1 it was evil/nice of t1 to do that
d.*Marilyn1 it was evil/nice of h to do that.
(100) a. That was nice/wonderful of Marilyn.
b.*[Of whom]1 was that nice/wonderful t1?
c.*Who1 was that nice/wonderful of t1?
Notably, parallel restrictions hold for RNR.
(101) a.*It might have been nice of h to do that and probably was nice of t1 to do that [the woman who they are
thinking of hiring]1.
b.*It might have been nice t1 to do that and probably was nice t1 to do that [of the woman who they are thinking of
hiring]1.
(102) a.*That may have been wonderful of t1 and probably was wonderful of t1 [the person who I had just met in the
park]1.

< previous page page_132 next page >


If you like this book, buy it!
< previous page page_133 next page >
Page 133
b.*That may have been wonderful t1 and probably was wonderful t1 [of the person who I had just met in the park]1.
The parallel behavior of RNR and L-extractions with respect to these adjectival complements again supports the unity
of the two types of phenomena.
4.2.2.13 Parasitic Gags
A striking argument for the unity of RNR and L-extractions is based on P-gaps. A view now standardly maintained in
different frameworks is that P-gaps are "licensed" only by extractions, not by passivization, raising to object, and so on
(see, e.g., Chomsky 1982, 1986a; Pesetsky 1982; Engdahl 1983, 1984, 1985; Sag 1983; Kayne 1984; Gazdar et al.
1984, 1985; Kiss 1985; Browning 1987a,b; Hukari and Levine 1987; Koster 1987; Pollard and Sag 1994). That is:
(103) The Parasitic Gap Constraint
The real gap "licensing" a P-gap is an extraction gap (meeting other conditions).
But for (103) to be considered viable, RNR must be analyzed as an extraction phenomenon, since RNR gaps "license"
P-gaps. 20
(104) a. Jerome fired [that secretary]1 after finding*(her1) drunk.
b. [Which secretary]1 did Jerome fire t1 after finding t1 drunk?
c. Jerome fired t1 after finding t1 drunk and Bill hired t1 after finding t1 sober [the tall young woman standing over
there]1.
(105) a. The boss warned/informed Jane1 that he would fire*(her1).
b. It was Jane who1 the boss warned/informed t1 that he would fire t1.
c. The boss warned t~ that he would fire t1 and the police informed t1 that they would arrest t1 [all those who were
involved in the embezzling]1.
(106) a. Greg decided to buy that model car1 after reading about it1/*t1.
b. Greg decided to buy t1 after reading about t1 and Gail agreed to lease t1 before test driving t1 [that new model
electric car which actually doesn't work]1.
That RNR "licenses" P-gaps supports the unity of RNR and L-extractions and yields a sharp objection to the view of
RNR endorsed by McCawley (1982, 1987, 1988) and Levine (1985). That view denies the

< previous page page_133 next page >


If you like this book, buy it!
< previous page page_134 next page >
Page 134
existence of "real" gaps in RNR cases and thus offers no reason why, for example, (106b) ought to be different from
(107). 21
(107)*Greg decided to buy [that new model electric car which actually doesn't work]1 after reading about t1.
That is, if RNR is not an extraction phenomenon, its ability to "license" P-gaps is anomalous. See also section
4.3.4.2.22
4.2.2.14 Interwoven Dependencies
Serious problems for a nonextraction, multiple-mother-node view of RNR are linked to generalization (108).
(108) In general, a standard ATB instance of L-extraction has an interwoven dependency correspondent.
Interwoven dependency refers to the sort of pattern found in English respectively constructions, in which there are
overlapping dependencies of unbounded length at unlimited distances. I touched on these in passing in section 4.2.2.11.
Claim (108) states that, for example, corresponding to (109a) is a well-formed (109b).
(109) a. [Which woman]1 did Fred date t1 and Bob marry t1?
b. [[Which nurse]1 and [which hostess]2]3 did Fred date t1 and Bob marry t1, respectively?
Significantly, the analog of (108) holds for RNR. Alongside McCawley's example (4a), then, one finds (110).
(110) a. John loves t1 and Mary hates t2 [[oysters]1 and [clams]2]3, respectively.
b. Marsha argued for t1 on Tuesday and Louise argued against t2 on Thursday [[communism]1 and [fascism]2]3,
respectively.
One could, I think, hardly deny that (110a,b) are RNR structures. In any event, I argue against such a view immediately
below. Notably, though, it is impossible to provide a representation for (110a,b) anything like McCawley's (4b), The
structure of, for example, (110a) contains a whole constituent, the coordinate NP, not found in its closest nonreduced
correspondent (say, (111))that is, oysters and clams.
(111) John loves oysters and Mary hates clams.
Moreover, each conjunct of (111) enters into constituency relations not found in an unreduced correspondent, that is,
relations to the coordinate

< previous page page_134 next page >


If you like this book, buy it!
< previous page page_135 next page >
Page 135
phrase oysters and clams. Evidently, then, the multiple-mother-node view of RNR has no application to cases like
(110), undermining the idea that RNR forms have the same constituent structures as their presumed more redundant
sources. 23
I am sympathetic to a reaction that it is unfair to criticize McCawley's viewindeed, almost any viewfor a failure to
properly account for interwoven dependency structures. For it is true that almost no account of grammatical structure
and coordination in particular has had anything much serious to say about them. There is, one might argue, something
of an implicit conspiracy of silence governing contemporary grammatical discourse about coordination whereby
proposals are evaluated essentially in isolation from the implications of interwoven dependencies; see section 4.3.
McCawley (1988, 536-540) offers various descriptive comments about respectively constructions but does not consider
how these relate to his view of RNR. Ultimately, though, a serious account of grammatical structure cannot be limited
to views that are incompatible with interwoven dependencies.
The descriptive problems raised by interwoven dependencies are sufficiently severe for most views of grammar that
some linguists might wish to claim that the phenomenon is somehow fundamentally distinct from other coordinate
structures. But the tenability of such a doctrine can be attacked without appeal to anything beyond the constraints
already cited as grounds for the unity of RNR and L-extractions, for such properties systematically manifest themselves
in interwoven dependencies as in ordinary ATB cases.
For example, the Indirect Object Constraint, shown earlier to hold for both simple L-extractions and RNR, also
constrains both ATB L-extractions and the interwoven dependency type.
(112) a. [Which nurse]1 did Ernest sell cocaine to t1 and George sell heroin to t1?
b.*[Which nurse]1 did Ernest sell t1 cocaine and George sell t1 heroin?
c. [[Which nurse]1 and [which hostess]2]3 did Ernest sell cocaine to t1, and George sell heroin to t2, respectively?
d.*[[Which nurse]1 and [which hostess]2]3 did Ernest sell t1 cocaine and George sell t2 heroin, respectively?
Similarly; the Indirect Object Constraint holds for the interwoven dependency form of RNR.

< previous page page_135 next page >


If you like this book, buy it!
< previous page page_136 next page >
Page 136
(113) a. Ernest sold cocaine t1 and George sold heroin t1 [[to the first nurse]1 and [to the second dental assistant]2]3,
respectively.
b.*Ernest sold t1 cocaine and George sold2 heroin [[the first nurse]1 and [the second hostess]2]3, respectively.
The same points can be made with the Genitive Constraint; I leave the construction of relevant examples to the
interested reader.
It should be shown, however, that P-gaps are equally "licensed" by the interwoven dependency forms of both L-
extractions and RNR. Compare (105) with (114).
(114) a. [[Which secretary]1 and [which programmer]2]3 did Jerome respectively fire t1 after finding t1 drunk and hire
t2 after finding t2 sober?
b. Jerome fired t1 after finding t1 drunk, and Bill hired t2 after finding t2 sober [[this tall young woman]1 and [that tall
young man]2]3, respectively.
These facts indicate, I believe, that simple L-extractions, ATB L-extractions, interwoven dependency L-extractions,
regular (ATB) instances of RNR, and interwoven dependency varieties of RNR are all instances of a single overall
phenomenon. If, as I suspect, no current view of grammatical structure has provided a way of representing the
commonality of these constructions, this is less a reason for doubting their unity than it is evidence of the inadequacy
of the views in question.
4.2.2.15 Additive Coordinations
Most views of coordination concentrate on what I will call distributive coordination, illustrated in many earlier
examples and in (115).
(115) [How many frogs]1 did Greg capture t1 and Lucille train t1?
Here the phrase how many frogs semantically distributes in the same way across each of the conjuncts. A distinct type
of coordination illustrated in the previous section is exemplified in (116).
(116) [How many frogs]1 and [how many toads]2 did respectively Greg capture t1 and Lucille train t2?
In such interwoven coordinations, each conjunct of some phrase distributes at least semantically only to specific
conjuncts of others.
There are still other types of coordination. One, which I refer to as additive coordination, is illustrated by (117).

< previous page page_136 next page >


If you like this book, buy it!
< previous page page_137 next page >
Page 137
(117) Greg captured t1 and Lucille trained t1 [312 frogs]1 between them.
In their superficial syntax, additive coordinations resemble distributive ones and contrast with interwoven ones.
Semantically, however, additive structures resemble interwoven structures in that (117) is not, for example, at all
equivalent to (118).
(118) Greg captured 312 frogs and Lucille trained 312 frogs.
A multiple-mother-node conception of RNR has no more application to additive cases than to interwoven ones. On
such an analysis, (117) would apparently have to have a structure in which both the verbs captured and trained have a
direct object of the form 312 frogs. Under any general conditions linking form and meaning, this will yield the wrong
claim that (117) and (118) are paraphrases.
This criticism is really only a variant of the brief antitransformational argument given by Gazdar (1981, 180) on the
basis of earlier observations by Jackendoff (1977, 192) and Abbott (1976, 642). This involved cases like (119), also
manifesting "relational" RNR pivots that fail to distribute semantically.
(119) a. John hummed t1 and Mary sang t1 [the same tune]1.
b. The Red Sox beat t1 and the Giants were beaten by t1 [different teams]1.
Again, a multiple-mother-node analysis of such cases would seem to lead to structures that fail to relate to semantic
form in the correct way. None of McCawley's presentations of his nonextraction view of RNR deal with cases like
(119).
4.2.3 Summary
In sections 4.2.2.1-4.2.2.15 I have documented that numerous characteristics of L-extractions also turn out to be
properties of RNR. The number and diverse nature of the relevant restrictions indicate that a view of RNR that fails to
analyze it in terms very similar to those of L-extractions will miss extensive generalizations subsuming the two. 24
(Further arguments to this effect are found in appendix B.) There are, to be sure, certain real contrasts between RNR
and L-extractions. For example, whereas in general each L-extraction is attested in noncoordinate, ATB, and
interwoven dependency varieties, noncoordinate instances of RNR are not clearly attested. Also, as first observed by
Wexler and Culicover (1980, 301),

< previous page page_137 next page >


If you like this book, buy it!
< previous page page_138 next page >
Page 138
RNR is evidently not governed by what has been called Subjacency. Further, it has a much greater freedom of target
constituent than any English L-extraction. Despite these differences, the extensive similarities between RNR and L-
extractions would appear to mandate a common analysis for the two. This conclusion, although at odds with that of
Wexler and Culicover (1980, 303) and with recent claims by McCawley (1982, 1987), Levine (1985), Ojeda (1987),
and Kayne (1994), is consistent with most past writing on RNR. For example, in works like Ross 1967, Postal 1974,
and Williams 1990, both RNR and L-extractions are taken to be (or at least involve) transformational movements. In
Gazdar 1981, both L-extractions and RNR are taken to instantiate the Slash category mechanism; see section 4.3. The
arguments I have given can thus be taken to strengthen what has been a persistent if not unanimous view: namely,
although unique in certain ways, RNR is an extraction phenomenon.
4.3 Right Node Raising and Slash Category Approaches to Extraction
4.3.1 Background
More than a decade ago, work in Generalized Phrase Structure Grammar (GPSG) introduced a then novel mechanism
for describing extractions, so-called Slash categories. These provide a key element in the GPSG at-tempt to overcome
the view, certainly dominant in 1980 and arguably still so today, that various natural language properties show the
inadequacy of phrase structure grammars and motivate transformational grammars. One claim was that by appealing to
(among other things) Slash categories, phrase structure grammars could provide adequate accounts of extractions.
The basic idea underlying Slash categories is rather simple. If the set of basic category symbols of a phrase structure
grammar is VN, then, as Gazdar (1981, 159) observes, one can define a set of derived categories as in (120).

The intended interpretation of such elements is specified in (121).


(121) "[A] node labeled a/b will dominate subtrees identical to those that can be dominated by a, except that somewhere
in every canonical subtree of the a/b type there will occur a node of the form b/b dominating a resumptive pronoun, a
phonologically null dummy

< previous page page_138 next page >


If you like this book, buy it!
< previous page page_139 next page >
Page 139
element, or the empty string, and every node linking (a/b and b/b will be of the s/b form. Intuitively, then, a/b labels a
node of type a which dominates material containing a hole of type b (i.e. an extraction site on a movement analysis).
So, for example, S/NP is a sentence which has an NP missing somewhere." (Gazdar 1981, 159)
The diverse developments of GPSG since its earliest variants (see Gazdar et al. 1985) maintain the basic idea in (121);
moreover, it is adopted in related frameworks, such as Head-Driven Phrase Structure Grammar (HPSG) (see Pollard
and Sag 1987, 1994). 25 The distribution of Slash categories has been taken to be controlled by various general
principles, such as the Foot Feature Principle and Head Feature Convention (Gazdar et al. 1985), or the Nonlocal
Feature Principle (Pollard and Sag 1994). But these developments in general supplement and extend the conception in
(121), without deviating greatly from its essential assumptions.
My goal in this section is to argue, chiefly on the basis of properties of RNR, that, as characterized in (121), Slash
categories cannot properly describe natural language extractions. To facilitate discussion, I introduce formally a relation
between nodes in (GPSG-style) trees (note, not a relation between symbols referencing nodes in GPSG rules,
metarules, etc.). I assume an indexing of nodes (equivalently, identification of nodes and indices). Moreover, I allow for
the possibility (instantiated later) of taking Slash to be one of several related Slash-like features, by recognizing SLASH
as a set of features with the key properties of the original Slash. I use a/b to represent indifferently members of SLASH.
In general, then, Slash category patterns will be represented as in (122).
(122) Slash structures with indexed nodes

Unlike in standard GPSG, which utilizes notations like VP/NP, here Slash categories are marked with indices, which
uniquely pick out the relevant category labels. This is equivalent to taking the values of the primitive Slash feature to
be node indices. Although a mere convenience at this point, such an indexing has in effect already been argued to be
necessary; see discussion of Pesetsky's (1982) argument below.

< previous page page_139 next page >


If you like this book, buy it!
< previous page page_140 next page >
Page 140
With this background, one can define a binary relation over nodes represented by the verb exfiltrate (and its
derivatives).
(123) Definition
Let i, j, be nodes in a tree T such that j does not directly dominate a trace. Then: i exfiltrates j if and only if .
Informally, then, node i exfiltrates node j if and only if j does not directly dominate a trace and is of the category .
For example, in (124) the exfiltration pairs are all and only those in (125). ((124) is Gazdar's (1982, 171) example
(8.8b) with certain simplifying label changes.)
(124)

(125) (2,3), (2,4), (2,5), (2,6)


It seems correct, then, to say that a fundamental claim of work adopting the Slash mechanism is (126).
(126) Extractions are correctly reconstructed as exfiltrations.
I will argue, though, that (126) cannot be maintained.
4.3.2 Parallel Exfiltrations
A certain amount of discussion in work assuming Slash categories asks whether there are sentences in which more than
one node exfiltrates a given node and considers implications of the answer (see Gazdar 1982, 177-178; Maling and
Zaenen 1982, 252-256; Chung and McCloskey 1983; Gazdar et al. 1985, 81-82; Ojeda 1987, 274-275; Pollard and Sag
1987, 76). I refer to such cases as parallel exfiltrations (P-exfiltrations).

< previous page page_140 next page >


If you like this book, buy it!
< previous page page_141 next page >
Page 141
Maling and Zaenen (1982) provide (127) as an English instance of this phenomenon. ((127a) is their (60b) with
irrelevant simplifications; (127b) is their (61) with my indexing.)
(127) a. [Into the wastebasket]1 Hilary put t2 t1 and Rob dropped t2 t1 [their autographed copies of Syntactic
Structures]2.

This is evidently an RNR structure, but it also involves L-extraction of the PP. Here the category corresponding. to into
the wastebasket and that corresponding to their autographed copies of Syntactic Structures have both seemingly
exfiltrated several different constituents, for example, those indexed 3, 4, 5, 6, 7, and 8.
Despite the arguable existence of P-exfiltrations, the version of GPSG in Gazdar et al. 1985 is formulated so as to
disallow them. 26 Imposition of this constraint is known by its developers to be a flaw but claimed (pp. 81-82) to be
part of a reasonable research strategy. However, Gazdar et al. leave the impression that the constraint is viable at least
for English. This wrongly ignores cases like (127) as well as those like (128) involving extraction from embedded
questions. Recognizing that P-exfiltrations exist in English, Pollard and Sag (1987, 1994) have designed the HPSG
framework to allow them via Maling and Zaenen's (1982) suggestion that the value of Slash be allowed to be a set of
category indices, rather than the single category of GPSG work.

< previous page page_141 next page >


If you like this book, buy it!
< previous page page_142 next page >
Page 142
A strong anti-Slash category argument was formulated by Pesetsky (1982, 554-556) on the basis of a specific type of P-
exfiltration. Since the argument has remained both unpublished and unreferenced in the Slash category literature, it is
worth going over. 27 First, Pesetsky observes that P-exfiltration is possible from the VPs of embedded questions.
(128) a book which1 I know who2 [VP to talk to t2 about t1]
Minimally, in Slash terms, the VP involves P-exfiltration of two NPs. Next, Pesetsky argues as follows:
(129) ''Worse yet, the distinction between a single slash category and a double slash category must be sensitive to the
indices on the gaps. Thus, if a category contains two gaps with different indices, they must both be represented in the
slash notation...."
Pesetsky's support for these conclusions is the interesting paradigm in (130).
(130) a.*a book that1 I know who2 to [VP/NP talk to Bill about t1] and [VP/NP-NP persuade t2 to buy t1]
b.*a book that1 I know who1 to [VP/NP talk to t2 about Mary] and [VP/NP-NP persuade t2 to buy t1]
c. a book that1 I know who2 to [VP/NP talk to t2 about t1] and [VP/NP-NP persuade t2 to read t1]
Moreover, he notes that a category containing two gaps with the same index cannot be treated as involving double
Slash categories under pain of failing to rule out (131).
(131)*a book that1 I know who2 to [VP/NP talk to the author of t1 about t1] and [VP/NP-NP persuade t2 to buy t1]
Finally, Pesetsky remarks that the Slash category labels must bear the relevant indices, to block (132).
(132)*a book that1 I know who2 to [VP/NP{2} talk to t2] and [VP/NP{1} buy t1]
Here, if the representation involves only Slash categories of the form "/NP", the ungrammaticality of (132) will not be
marked as in effect a violation of the Head Feature Convention.
As a critique of the specific formulation of Slash categories in GPSG work, Pesetsky's arguments are, I think,
essentially correct. They show that P-exfiltrations are a worse problem for standard Slash category approaches

< previous page page_142 next page >


If you like this book, buy it!
< previous page page_143 next page >
Page 143
than the published literature indicates. It would seem that a Slash category approach must, minimally, allow Slash to
take as values a set of categories, as Maling and Zaenen (1982) contemplate, but further must recognize that the values
of Slash must be distinguishable by indices or their equivalent. Such a view is proposed without reference to Pesetsky
1982 by Pollard and Sag (1987, 1994) and is implicit in the notation adopted above. But as nothing has shown the
modifications just indicated to be untenable, it would clearly be mistaken to view Pesetsky's argument as a definitive
refutation of the Slash idea. See section 4.3.4.5.
4.3.3 Quasi Exfiltrations
A more fundamental objection to Slash categories can be founded on a formal property of this device that manifests
itself in certain instances of multiple extractions logically distinct from P-exfiltrations. To my knowledge, such cases
have been previously discussed in Slash terms in a brief note by Piera (1985), in a passing remark by Levine (1985,
494-495), and by Hukari and Levine (1989, 1991). The phenomenon is defined abstractly not by distinct exfiltrations
from the same constituent, but by a kind of "recursive" exfiltration, that is, exfiltration of an element a from a
constituent b that itself exfiltrates a constituent c. To facilitate discussion, I appeal to the logical ancestral of Exfiltrate
to define a slightly narrower relation called Quasi Exfiltrate (Q-Exfiltrate; see Carnap 1958, 146-148; Quine 1958, 215-
221). Element a Q-exfiltrates b if and only if it is ancestrally related to b under Exfiltrate (i.e., when there is a sequence
of exfiltration relations linking a and b) but a does not exfiltrate b.
Schematically, instantiations of Q-exfiltrations would look like (133) (note that left/right constituent order is irrelevant).

< previous page page_143 next page >


If you like this book, buy it!
< previous page page_144 next page >
Page 144
Here node 3 exfiltrates nodes 2, 5, and 9 and node 5 exfiltrates nodes 4 and 6. Hence, node 3 Q-exfiltrates all and only
nodes 4 and 6.
Internal to frameworks adopting the Slash mechanism, it is a factual question whether there exist natural language
sentences whose correct analyses involve instances of Q-exfiltration. If the answer were negative, the (only) relevant
issue internal to such views would be what universal constraints guarantee the nonexistence of Q-exfiltration. GPSG
work has by and large assumed that Q-exfiltration does not exist. Gazdar et al. (1985) adopt the restriction in (134), call
it the No Recursion Constraint, which clearly bans it. A somewhat different formulation is given by Gazdar et al.
(1988) (see also Hukari and Levine 1989, 1991).
(134) "[T]he value set that p associates with given category-value feature can only contain categories in which does not
already appear. That is, C can only be in the value set of if is not in the domain of C, or in the domain of any C"
contained in C, at any level of embedding." (Gazdar et al. 1985, 36)
Utilizing the term UDC (unbounded dependency construction), Hukari and Levine comment on the No Recursion
Constraint as follows:
(135) "Formally, by blocking the arbitrary recursive specification of categories, it restricts the set of categories to a
finite number and so contributes to the context-free equivalence of GPSG.... Empirically, it effectively rules out the ill-
formed examples in (110).
(110) a. I decided that pictures of Terry, I could do without.
b.*Who did you decide that pictures of you could do without?
c.*Robin is the person who(m) I decided that pictures of I could do without." (Hukari and Levine 1991, 133)
Despite their approval of the No Recursion Constraint, Hukari and Levine (1991, 134) claim that English does manifest
gap-within-filler constructions, namely, in what I have called object-raising or object deletion structures. For these
authors, such constructions would have otherwise been described with Slash categories, as in Gazdar et al. 1985, 150-
152. Their example is (136).
(136) Napoleon is one hero1 I find [stories about t1]2 just too boring to listen to t2.
Under a standard Slash analysis of such cases, the filler category for the t2

< previous page page_144 next page >


If you like this book, buy it!
< previous page page_145 next page >
Page 145
gap in (136), which is [stories about t1], itself contains a gap whose filler category is one hero1.
Hukari and Levine's solution to this state of affairs maintains the No Recursion Constraint and describes (136) by
reanalyzing the original GPSG feature Slash into two distinct features, Gap and Slash, which share many, but not all,
properties. Object-raising and object deletion structures are then described with Gap, and (as in past GPSG work) L-
extractions like questioning are described with Slash; a distinct proposal is found in Pollard and Sag 1994, chap. 4. I am
not here concerned with the adequacy of Hukari and Levine's suggestion. However, although postulating what is in
effect a class of distinct but related Slash-like features all, for example, obeying the Foot Feature Principlewould
maintain the consequences of the No Recursion Constraint for generative capacity, it would obscure the question of
whether facts like those related to their (110) are derived. Although Hukari and Levine's Gap feature cannot be used to
describe the starred cases of their (110), from here it is not a great step to positing a feature distinct from either Slash or
Gap that could be soused.
4.3.4 English Quasi Exfiltrations
4.3.4.1 Comments
Pretty clearly, the No Recursion Constraint will have to be abandoned in any event if primitive Slash features are main-
rained. Hukari and Levine themselves hint that other languages, especially Scandinavian languages, allow analogs of
the starred cases of their (110). Piera (1985) gives apparently parallel Spanish examples. Further, what must be
regarded as Q-exfiltrations in Slash category terms can be found in English, namely, among RNR constructions. In fact,
several different kinds of Q-exfiltrations can be attested among English RNR constructions. After documenting this in
detail, I examine how Slash approaches to extractions fare with such structures.
4.3.4.2 L-Extraction/Right-Node-Raising Interactions
Consider (137) (from Wexler and Culicover 1980, 300).
(137) a. Who1 does Mary buy t2 and Bill sell t2 [pictures of t1]2?
b. I wonder who1 Mary buys t2 and Bill sells t2 [pictures of t1]2.
Examples like these are also accepted and discussed by McCawley (1982, 101, n. 11) and Ojeda (1987, 262). The
former states generalization (138).

< previous page page_145 next page >


If you like this book, buy it!
< previous page page_146 next page >
Page 146
(138) "While the whole conjoined structure to which RNR applies is an island, the Coordinate Structure Constraint ...
does not rule out extraction from the constituent that RNR 'raises',..."
See also (139) and (140).
(139) a. [Which official]1 did they say Bob suspected t2 and Frank proved t2 [that Sally bribed t1]2?
b. It is Nigel who1 Carol learned when t2 and Edith learned where t2 [they could successfully interview t1]2.
c. the dentist who1 Bob expected the police to assert t2 and Frank expected the lawyers to deny t2 [that you had dated
t1]2
(140) a. [Those cities]1, Frank may t2 and Glen probably will t2 [try to visit t1]2.
b. the drug which1 George wanted t2 and Frank needed t2 [to try t1]2
The examples in (137), (139), and (140) manifest typical features of uncontroversial RNR cases, and there seems to be
no known alternative to viewing them as true instances of RNR. Moreover, although these examples involve
interactions of the same phenomena found in (133), L-extractions and RNR constructions, they are not P-exfiltrations,
since no individual constituent manifests multiple gaps. Rather, in Slash terms, each L-extracted NP must be taken as
extracted from a constituent that is itself an RNR pivot. 28 Hence, (137), (139), and (140) indicate that, in Slash
category terms, proposition (141) must be taken as true of English.
(141) A ((sub)constituent of an) RNR pivot can itself be an L-extraction gap.
Given Wexler and Culicover's (137) and McCawley's (1982) and Ojeda's (1987) discussions, the factual basis of (141)
was in effect observed independently more than a decade ago.29
To link (141) to explicit Slash category work, one can focus on Gazdar's (1981) proposal (maintained in Sag et al. 1985
and Gazdar et al. 1985, 125) that RNR is characterized via the Slash mechanism. The specific idea was that RNR
constructions are described by the extremely general schema in (142) (from Gazdar 1981, 178; semantic aspects
irrelevantly represented schematically as X).

If (142) represents the correct treatment of RNR internal to Slash

< previous page page_146 next page >


If you like this book, buy it!
< previous page page_147 next page >
Page 147
approaches, then (137), (139), (140), and the like, clearly instantiate Q-exfiltration. After considering the implications
of this assumption, I examine (in section 4.3.4.4) the feasibility within Slash approaches to extractions of rejecting a
Slash description of RNR structures.
Oversimplifying in irrelevant ways, standard Slash category posits about L-extractions plus an instantiation of Gazdar's
proposal (142) would lead to, for example, (139a) being assigned the tree in (143), which, for simplicity, omits
structure related to the presence of they say. In (143) I represent L-extractions by the Slash notation "" and RNR
extractions by the Slash notation " ". Precise meaning is given to this distinction just below.

Several issues arise from such structures. First, they raise the same problem for the No Recursion Constraint as those
linked to (136) discussed by Hukari and Levine (1989, 1991). However, this problem admits a solution parallel to the
one those authors proposed. Their new Slash feature (differentiated from their Gap feature) could be further subdivided
into features that can be called L(eft)-Slash and R(ight)-Slash, as partially foreshadowed by Maling and Zaenen (1982,
256, 278, n. 19). 30 L-Slash would serve to describe L-extractions, and R-Slash would, among other things, replace
Slash in analogs of (142). Since SLASH was taken to be a set of features with the properties of the original Slash, L-
Slash and R-Slash (but not Hukari and Levine's (1991) feature Gap) are among its members. I use the notations ab
and for these categories, respectively; hence, (143) can be interpreted as incorporating this new distinction. Recall

< previous page page_147 next page >


If you like this book, buy it!
< previous page page_148 next page >
Page 148
that represents indifferently members of SLASH, hence any of the categories defined by these new features.
Recognition of R-Slash and L-Slash parallels Hukari and Levine's (1991) Gap proposal, costs very little, and minimally
saves the No Recursion Constraint (135) from cases like (137), (139), and (140).
In other ways also, structures like (143) do not seem to adversely affect basic GPSG assumptions. In particular, the
SLASH category on node 4 is not instantiated and thus need not appear on its mother node 3 to satisfy the Foot Feature
Principle. The latter requires that the value of a foot feature F instantiated on a mother node be identical to the
unification of the instantiated F features on all its daughter nodes. R-Slash and L-Slash must of course be categorized
as foot features. If " did appear on node 3, a violation of the Foot Feature Principle would ensue. Specification "2"
must appear on node 3, since that feature is instantiated on node 5.
So far, then, no grave objections to a refined version of Slash category frameworks distinguishing L-Slash from R-
Slash have been derived from structures like (143). Nonetheless, such a structure has a somewhat suspicious property.
NP2, the L-extracted constituent, is not related in any established syntactic way to node 4, 6, 7, 8, 9, 10, or 11. This
state of affairs is suspicious in that it contrasts with the state of affairs in, for example, the corresponding but simpler
(144).
(144) a. [Which official]1 did they say Bob suspected that Sally bribed t1?

< previous page page_148 next page >


If you like this book, buy it!
< previous page page_149 next page >
Page 149
In (144b) the counterpart of NP2 of (143) is syntactically related to the S-bar node 5 (defining the complement of
suspected), to the S node 12 daughter of that, and to the VP node 13 daughter of that. More precisely, in (144b) the L-
Slash specification corresponding to the extracted phrase which official occurs on nodes 5, 12, and 13, whereas in (143)
the analogous specification does not appear on their counterparts. In my terms, the L-extracted constituent in (144b)
exfiltrates nodes 3, 8, 5, 12, and 13. But in (143) the extracted constituent exfiltrates nodes 3, 5, 12, and 13 but not 4, 6,
7, 8, 9, 10, or 11. Rather, it Q-exfiltrates the latter.
Given the formal contrast between Q-exfiltration in cases like (143) and exfiltration in those like (144b), frameworks
embodying the Slash mechanism as so far understood, that is, as characterized in (121), in effect make a general
prediction. Because there is no reason why constraints restricting or banning exfiltration from a constituent K1 with
property P should in general also hold for those that Q-exfiltrate a K2 with P, nothing forces a correlation between
constraints in exfiltrations and those in Q-exfiltrations. More specifically, a lack of correlations in these two classes of
cases follows from the fact that for exfiltrated constituents, restrictions are imposed in Slash terms by controlling the
relations between Slash categories and the nodes intervening between them and their binders. But in the counterpart Q-
exfiltration structures, like (144b), the relevant relations do not exist.
The problems that can in principle arise from this state of affairs can be illustrated by focusing on island constraints.
Gazdar (1982, 174-178) proposes in effect that island constraints not derivable from broader principles can in general
be stated as specific restrictions on membership in Slash categories. He suggests that in effect the general form of a
non-derivative island constraint in this framework is along the lines of (145).
(145)*[a/b X]
"A constituent meeting the characterization X cannot belong to a category ab."
Let it be granted for argument that device (145) is adequate to block standard extractions from islands, like (146).
(146)*[With whom]1 does she know men who want to consult t1?
It nonetheless clearly fails for Q-exfiltration structures like (147a), as the structure in (147b) shows. For simplicity (and
irrelevantly), this structure gives expansions only for the leftmost relative clause.

< previous page page_149 next page >


If you like this book, buy it!
< previous page page_150 next page >
Page 150
(147) a.*[With whom]1 does she know men who want to t2 and women who need to t2 [consult t1]2?

No instance of (145) blocks Q-exfiltrations from islands like that in (147b) because the distribution of Slash categories
fails to yield a violation of any restriction schematized by (145). In (147b) one can assume that the island-defining node
in the leftmost relative clause is 7 or possibly 8. But neither of these is, or can be, marked with "2". That is, the
extracted PP only Q-

< previous page page_150 next page >


If you like this book, buy it!
< previous page page_151 next page >
Page 151
exfiltrates those nodes; it does not exfiltrate them. So cases like (147a) wrongly seem to be allowed in the treatments of
extractions in Slash category terms considered so far.
The failure of the suggested Slash category approach to island restrictions to block a Q-exfiltration case like (147b)
exemplifies a general inadequacy. Contrary to what extant Slash treatments imply, (148) holds.
(148) If exfiltration of type c of constituents from nodes of type n is subject to constraint K on n, then Q-exfiltration of
type c from nodes of type n is in general also subject to K.
Despite its scope and power, (148), like the earlier documentation that RNR pivots (and their subparts) are subject to L-
extraction, is not a novel claim. Rather, it has also been in the literature for more than a decade. The basic
generalization, proposed by Wexler and Culicover (1980), was quoted in (15). Wexler and Culicover's general
conclusion is essentially (148), once one abstracts from their particular (and here irrelevant) transformational views of
extraction and RNR. The evidence they give for it is in effect the pair in (149).
(149) a.*Who1 does Bill know a man who sells pictures of t1?
b.*Who1 does Mary buy t2 and Bill know a man who sells t2 [pictures of t1]2?
They note further that reversing the conjuncts so that the relative clause is in the first makes no difference. Wexler and
Culicover's argument based on (149) is rehearsed approvingly by McCawley (1982, 100; 1988, 530-532) and Ojeda
(1987, 261-262). McCawley aptly sums up the original argument as follows:
(150). "[W]hile RNR can move a constituent 'out of' an island, that constituent remains as immune to [left] extraction as
if it remained within the island." (McCawley 1982, 100)
Levine makes a parallel point.
(151) "[A]lthough RNRaising elements out of islands produces perfectly good sentences, certain extractions from these
Raised elements ... are instead forbidden, as in (2).
(2)*[To whom]1 did John give a briefcase t1 and Harry know someone who had given a set of steak knives t1 [t1]2?"
(Levine 1985, 492-493)

< previous page page_151 next page >


If you like this book, buy it!
< previous page page_152 next page >
Page 152
Levine's initial claim was supported by the grammaticality of (152).
(152) John gave a briefcase t~ and Harry knows someone who had given a set of steak knives t1 [to Bill]1.
In numerous cases, then, constituents a that a category K has Q-exfiltrated behave as if K exfiltrated them, but the
GPSG structures provide no SLASH features on a to describe this. For example, in imposing an island violation, the
phrase a man who sells in (149b) behaves as if the questioned NP who had exfiltrated from it, just as in (149a). But in
(149b), unlike (149a), the phrase a man who sells will not be marked with the relevant SLASH specification (''a/NP" in
standard GPSG terms). Similarly, in imposing an island violation, the constituent someone who had given a set of steak
knives in Levine's example (152) behaves as if the L-extracted phrase to whom had exfiltrated it, just as it has
exfiltrated the parallel phrase in (153).
(153)*[To whom]1 did Harry know someone who had given a set of steak knives t1?
So far I have supported (148) by appealing to the fact that island constraints manifest themselves in Q-exfiltrations
corresponding to standard island-violating exfiltrations. Claim (148) can also be supported by appealing to the
principles governing P-gaps. That Slash assumptions make proper predictions about these has been an important claim
of work in the Slash tradition (see, e.g., Sag 1983; Gazdar et al. 1984, 1985; Hukari and Levine 1987; Pollard and Sag
1994). To paraphrase Gazdar et al.'s (1985, 150) remarks about respectively structures, P-gaps have been "a standard
component of the pedagogic commercial for" Slash categories. Notably, though, extant Slash approaches to P-gaps fail
for certain Q-exfiltration structures. Examples (154)-(156) illustrate P-gaps induced in Q-exfiltration cases. Each (a)
example represents a standard P-gap structure not involving RNR; the corresponding (b) example contains a related
RNR form.
(154) a. the woman who1 your hugging t1 led t1 to suspect that you loved t1
b. the woman who1 your hugging t1 led to suspect t2 and your kissing led t1 to believe firmly t2 [that you loved t1]2

< previous page page_152 next page >


If you like this book, buy it!
< previous page page_153 next page >
Page 153
(155) a. It was Lucille1 that I convinced friends of h to try to hire t1.
b. It was Lucille1 that I convinced friends of t1 t2 and Ted convinced relatives of t1 t2 [to try to hire t1]2.
(156) a. the woman who1 your kissing t1 may amuse t1
b. the woman who1 your kissing t1 may t2 and should t2 [amuse t1]2
Such examples illustrate that, in general, the P-gap phenomenon exists in Q-exfiltration structures corresponding to
exfiltration structures in which it is found.
In standard Slash category terms, the complement in (156a) would have the irrelevantly simplified structure in (157).

The extra or P-gap after kissing is allowed in (157) basically because of the existence and position of the gap after
amuse. More precisely, the foot feature "2" on node 5 unifies with that on node 4 as required by the Foot Feature
Principle. Further, since Slash is also a head feature, "2" also occurs on the head daughter of node 3, the VP node 5,
as required by the Head Feature Convention. Hence, (157) satisfies the relevant GPSG principles. If the object of
amuse were not a Slash category, this would not be true.

< previous page page_153 next page >


If you like this book, buy it!
< previous page page_154 next page >
Page 154
Now consider the structure of (156b), shown in (158).

Here, node 2 actually exfiltrates only nodes 3 and 5; yet this suffices to induce the P-gap inside nodes 4 and 6. But
representation (158) is, of course, not a legal Slash category structure. It violates the Foot Feature Principle since "2"
is instantiated on node 6 but not on its mother 4. Suppose (158) is modified just so "2" occurs on 4; call the result (
158). There are several implications. Least importantly, node 4 in ( 158) then formally instantiates P-exfiltration,
although it is not a genuine P-exfiltration case. More seriously, the principle that blocks the non-P-gap version of
(156a) must somehow allow ( 158) while still. blocking (159).
(159)*the woman who1 your kissing h may amuse Gail
In Gazdar et al. 1985 the principle blocking (159) is the Head Feature Convention. But that also blocks ( 158). To
preserve the Head Feature Convention, one could modify ( 158) to ( 158), in which "2" also occurs on nodes 7, 10,
11, 13, and 14. But the latter nodes do not dominate any trace bound to the extracted NP. The basic nature of Slash
categories as specified by Gazdar in (121) would therefore have to be revised. 31
Minimally, then, the properties of Q-exfiltrations raise serious problems for Slash category views of extractions.
Specifically, the fact that Q-exfiltrations behave like exfiltrations (principle (148)) is not captured by current approaches
to the Slash mechanism. This is so because the

< previous page page_154 next page >


If you like this book, buy it!
< previous page page_155 next page >
Page 155
principles controlling the distribution of Slash features have been designed for structures in which there is no need to
generalize to the notion Q-exfiltration or its equivalents. But in (158), to account for the P-gap located at node 12, it is
necessary to represent the fact that nodes 14, 13, 11, 10, 7, and 4 behave like the nodes 7, 5, and 3 in (157)that is,
behave as if they belonged to the category "2". Actually allowing this will involve minimally abandoning (121), and
abandoning or modifying both the Head Feature Convention and the Foot Feature Principle.
4.3.4.3 Pure Right-Node-Raising Interactions
Nothing in the previous section directly challenges the No Recursion Constraint, given the bifurcation into L-Slash and
R-Slash. However, the evidence given there for the existence of English Q-exfiltrations depended on examples in
which RNR interacts with L-extractions. In fact, there is an even stronger English exemplification of Q-exfiltration
based on RNR alone. Levine (1985) in effect already noticed that an RNR pivot can itself be a target for additional
RNR. Hence, there are Q-exfiltration structures whose only extractions represent RNR. An irrelevantly simplified
example of the type cited by Levine is (160).
(160) John gave silver t1 and Harry gave gold t1 [to the mother t2]1 and I gave platinum [to the father t2]3 [of the
famous quintuplets]2.
Here the PP of the famous quintuplets is an RNR pivot linked among other things to an apparent gap inside a distinct
RNR pivot to the mother t2. A different example of the same phenomenon is (161).
(161) Harry looked for t1 and Jerome found t1 [photos of t2]1 and Louise produced sketches of t2 [the controversial
candidate]2.
A third case not limited to NP extraction is (162).
(162) Frank reported to Louise t1 and Mike admitted to Marion t1 [that Tony could t2]1 and I believe that he should t2
[hire more workers]2.
These examples seem to leave little doubt of the existence of Q-exfiltrations based exclusively on RNR. 33
It is a striking historical oddity, then, that Levine (1985) had, without recognizing it, discovered grave difficulties for
Slash approaches, graver, for instance, than those that Hukari and Levine (1989, 1991) treat by

< previous page page_155 next page >


If you like this book, buy it!
< previous page page_156 next page >
Page 156
postulating the feature Gap. Given the 1985 observation that RNR is in effect "autorecursive," a primitive feature Gap
cannot save the No Recursion Constraint, for as Levine (1985, 494) rightly stresses, "... this iteration of RNR is
apparently bounded only by processing limits" which is to say, not bounded at all. The existence of unbounded, pure
RNR Q-exfiltrations determines that even recognizing Gap, R-Slash, and L-Slash will not suffice, and, moreover, that
no finite expansion of categories can. A theory with primitive Slash features must apparently recognize infinitely many
categories (modulo the possibility dismissed in section 4.3.4.4). Further, an "autorecursive" RNR construction evidently
raises all the problems for Slash approaches discussed in section 4.3.4.2 on the basis of L-extraction/RNR interactions.
Levine's implicit conclusion that cases like (160)-(162) are irrelevant to the Slash mechanism must have been based on
his (1985, 493) view that RNR is not an extraction phenomenon, hence not to be described via the Slash mechanism.
This view is contemplated in an explicit GPSG context by McCloskey.
(163) "Second, it suggests that whatever type of rule or process is involved in the derivation of Right Node Raising
structures, it is different in kind from extraction as normally understood. In particular, for Generalized Phrase Structure
Grammar, it suggests that perhaps the Slash Category apparatus is not involved in Right Node Raising at all."
(McCloskey 1986, 185)
Under that assumption, (160)-(162) are obviously not problematic for the theory of Slash categories. But the denial that
RNR is an extraction phenomenon was argued to be untenable in section 4.2; see also the following section.
4.3.4.4 Non-Slash Description of Right Node Raising
My discussion of the difficulties that English Q-exfiltrations involving RNR constructions raise for Slash approaches
has been based on the view that RNR constructions are, internal to the overall approach, properly described (as in
Gazdar 1981) through some appeal to the Slash mechanism. I have modified this only by recognizing R-Slash and L-
Slash. Consider more closely the feasibility, within Slash approaches to extractions, of rejecting Slash descriptions of
RNR constructions.
To begin with, focus on framework-internal issues. First, without Slash specifications, there is no evident way in which
the categories of RNR

< previous page page_156 next page >


If you like this book, buy it!
< previous page page_157 next page >
Page 157
pivots can be properly linked to the categorial possibilities of phrases otherwise occurring in the positions of RNR gaps.
Second, the virtues of the Slash approach to the interaction of extraction and coordination originally spelled out by
Gazdar (198l) are lost for RNR cases if RNR is not a Slash phenomenon, for all the relevant propertiessensitivity to the
CSC, the apparent ATB exceptions to it, and so onhold of RNR cases just as much as of L-extractions. Third, the fact
that RNR "licenses" P-gaps is anomalous in Slash terms if RNR is not an instantiation of Slash possibilities. Internal
considerations alone, then, do not seem to leave much possibility with/n Slash frameworks of rejecting a Slash
description of RNR constructions.
Moreover, regardless of particular Slash category assumptions, all the evidence of section 4.2 linking RNR to
extractions is relevant. If RNR is not treated, internal to Slash accounts, via the Slash mechanism but L-extractions are,
then all the arguments aimed at McCawley's account of RNR would, in effect, attack this hypothetical approach to
RNR. Overall, then, within such approaches, the general conclusion of section 4.2 that RNR is an extraction
phenomenonwould appear to be inescapable. This creates a conundrum for later versions of GPSG because of the
factors noted by McCloskey (1986), mentioned in note l, which would apparently preclude a Slash account of RNR.
4.3.4.5 Nonprimitive Slash Properties
Since I am neither competent in, nor an enthusiast of, the sort of phrase-structural approaches to grammar in which
Slash categories play a role, it would be inappropriate for me to attempt to consider in detail how best to approach the
above-mentioned issues within that framework. However, I make an observation connected both to the problems that
have been uncovered and to a perceptible major redundancy in Slash category representations.
The basic generalization relevant to my discussion of Q-exfiltrations would seem clear at an informal level. A viable
theory must minimally be able to mark with the equivalents of Slash features not only those nodes that an extracted
element exfiltrates but also those that it Q-exfiltrates. For example, the fact that node 4 in (158) behaves as if the NP
node 2 exfiltrated it should follow from the fact that node 2 exfiltrates the VP node 5, which in turn exfiltrates node 4.
In short, it seems that what principles like the Head Feature Convention and the Foot Feature Principle must be
understood as referring to are not Slash features as so far conceptualized but somewhat more abstract properties.

< previous page page_157 next page >


If you like this book, buy it!
< previous page page_158 next page >
Page 158
This should raise the question of whether the extant view of such features is not, even internal to the phrase-structural
assumptions underlying it, partly mistaken. Conceptualizations of the Slash mechanism so far have posited a primitive
syntactic feature Slash, in order to characterize certain nodes as having certain properties. But there is, a priori, great
redundancy in such features. Consider (157); there is an evident sense in which, given the position of the NP node 2
and the trace that it binds, the existence of "2" on nodes 3, 5, and 7 is redundant. One can make this intuition precise
along the following lines. Assume, in line with previous assumptions, that nonterminal nodes in phrase structure trees
are drawn from a class of nodes (indices) . Take traces to be pairs of the form [t, j] (written tj), where t is
some designated terminal symbol and . One can then eliminate all members of SLASH as primitive syntactic
features and attempt to assign nodes to the equivalents of Slash categories via a relational definition along the lines of
(164).
(164) Definition
Let m, n be nodes in some tree. Then m Slashes n if and only if
a. there exists a node o that is a sister of m, and
b. o dominates n, and
c. there exists a trace tm which is dominated by n.
Note that (164) defines an extensionally slightly broader notion than that captured by "exfiltrate" in that a node that
immediately dominates a trace can be slashed. One could exclude such nodes by an additional narrower relation "strict
slash," which excludes that case. Definition (164) takes advantage of the posit of traces and the fixed positional
relations of an extraction binder to such traces to predict in effect that certain nodes have the Slash property. Although
the definition is relational, nodes can be taken to have the relevant properties via some equivalent of lambda
abstraction. The overall definitional approach permits elimination of Slash as a primitive syntactic feature with the
concomitant possibility of eliminating specification of its values as individual categories (GPSG), sets of categories,
sets of indices (HPSG), and so on. Individual trees are enormously simplified.
Moreover, problems like those cited by Pesetsky (1982) and discussed earlier in connection with (128)-(132) are in
significant part addressed without special statement, for nothing in (164) prevents a single node n from being slashed by
more than one other node, as long as conditions (164a-c) hold. Further, Pesetsky's claim that the Slash mechanism

< previous page page_158 next page >


If you like this book, buy it!
< previous page page_159 next page >
Page 159
requires a proliferation of categories, whatever its original force, is directly circumvented, for a definitional approach to
the property Slash as in (164) requires no extra categories at all. A definitional approach also undermines Pesetsky's
subcategorization argument against Slash (see note 27). Finally, since a defined notion of Slash yields no increase of
categories compared with a fixed set of basic categories not including Slash categories, the "autorecursive" property of
RNR does not lead to any violation of the No Recursion Constraint.
The virtue of a nonprimitive view of Slash can be seen clearly in an-other area when it is extended, via appeal to the
ancestral of the relation defined in (164), for that relation, call it Semi-Slash, can provide a basis for the properties of
what were previously called Q-exfiltrations. Return to (158), where one wants the fact that nodes 3, 5, and 8 are marked
"2" to play a role in "licensing" the P-gap inside node 6, as the parallel slashing facts do in the exfiltration case in
(157). But there is now a basis for that if the relevant principle, whatever it is, refers not to the feature that NP node 2
slashes (e.g., node 11) but to the feature that it semi-slashes (node 11). This it does, since 2 slashes 5, which in turn
slashes 11.
Thus, a definitional approach to the Slash property may offer hope that principles like the Foot Feature Principle can be
reformulated to take account of the defined property Semi-Slash, ultimately subsuming cases like (158) under the
principles relevant for (157), as is evidently mandatory for any adequate account.
Of course, obvious problems stand in the path of developing a definitional, nonsyntactic feature approach to the
intuitive Slash idea. The principles that refer to Slash refer to other features of the syntactically primitive type. There
may be certain technical issues in subsuming a defined property under such principles although, for instance, I see no
reason why the class of foot features (defined by list in Gazdar et al. 1985, 80) could not be revised to be a class of
properties, including Semi-Slash.
A central problem would be to formulate the principles governing coordination properly. Pollard and Sag (1994) argue
that the approach to this domain taken in Gazdar et al. 1985 cannot stand. 34 The latter appeals heavily to the Head
Feature Convention and the view that coordinate constituents are multiheaded, both of which Pollard and Sag abandon.
Whereas Gazdar et al.'s approach has the virtue of not requiring a special principle for coordinate structures, Pollard
and Sag's (1994) proposal appeals to condition (165).

< previous page page_159 next page >


If you like this book, buy it!
< previous page page_160 next page >
Page 160
(165) Coordination Principle (weak version)
In a coordinate structure, the CATEGORY and NONLOCAL value of each conjunct daughter is subsumed by (is an
extension of) that of the mother. (Pollard and Sag 1994, 203)
Making something like (165) compatible with a defined approach to Slash like that of (164) would minimally require
subsuming the defined concept under a replacement for Pollard and Sag's notion "nonlocal value." There could also be
serious issues about how phrase structure rules or their equivalents can be formulated properly if all Slash categories
are defined. I will not pursue these matters.
To conclude, though, the evident virtues of a definitional approach to a notion like Slash suggest that it merits study
internal to any framework that adopts the Slash idea. 35
4.3.5 Slash Categories and Interwoven Dependencies
Even if the problems that Q-exfiltrations raise for Slash frameworks can be solved by some appeal to defined Slash
categories and their ancestrals, or in some other way, the Slash mechanism has other deficiencies in effect uncovered in
earlier discussion. I briefly make this point explicit.
In section 4.2.2.11 I argued that interwoven dependency structures provide an argument against McCawley's multiple-
mother-node view of the surface structures of RNR cases. Interwoven dependencies also seem to derail contemporary
formulations of Slash category approaches to extractions. Compare:
(166) a. [Which pilot and which sailor]1 will Joan invite t1 and Greta entertain t1?
b. [[Which pilot]1 and [which sailor]2]3 will (respectively) Joan invite t1 and Greta entertain t2 (respectively)?
As far as I can see, there are two different Slash approaches to cases like (166b). The first, or syntactically
nondistributive, approach would claim that apart from the word respectively (166b) has essentially the same constituent
structure as (166a), a familiar ATB form in which the exfiltrated phrase binds Slash categories in each conjunct. The
second, or syntactically distributive, approach would recognize partially different structures for (166a) and (166b). In
this view, the extracted coordinate phrase in the latter would not bind either trace. Rather, some novel devices would
have to permit the individual conjuncts to bind these traces

< previous page page_160 next page >


If you like this book, buy it!
< previous page page_161 next page >
Page 161
in the appropriate way. Under the nondistributive approach, the sharp differences between pairs like (166a,b) would
then be attributed somehow to purely "semantic" principles. This tack might be thought to offer some hope of
preserving the standard Slash approach to cases like (166b) or at least something like my definitional reconstruction of
it.
However, this is unlikely. Consider:
(167) [[Which man]1 and [which woman]2]3 did respectively the doctor talk to t1 about himself1, and the lawyer talk
to t2 about herself2?
Under the nondistributive view, traces t1 and t2 would both be bound by the bracketed extractee, NP3. This means in
effect that each of the reflexives would take that large NP as antecedent. But (168) shows that the independently
motivated principles controlling agreement of reflexive NPs with their antecedents normally determine plural
agreement with such an antecedent.
(168) [[Which man]1 and [which woman]2]3 did you talk to t3 about themselves3/*himself3/*herself3?
Further, if each reflexive in (168) takes the conjunctive NP3 as antecedent, how can one antecedent determine different
agreements in the two cases? Thus, a nondistributive approach to interwoven dependencies requires ad hoc
complications of reflexive agreement principles.
A parallel point holds for the distinct kind of agreement in (169).
(169) a. I am going out of my/*your/*our mind.
b. You are going out of your/*my/*our mind.
c. You and I are going out of our/*my/*your mind(s).
The agreement patterns in (169) are maintained in interwoven dependencies.
(170) a. He wants you and me to respectively go out of your mind and (go) out of my mind.
b. He wants you and me to respectively remain sane and go out of my/*our mind(s).
Again the nondistributive approach would appear to impose ad hoc agreement complications. Arguments like those just
sketched will presumably be constructible in many languages on the basis of a variety of grammatical agreement
phenomena. This line of argument suggests that Slash accounts of extractions cannot adopt the nondistributive
approach.

< previous page page_161 next page >


If you like this book, buy it!
< previous page page_162 next page >
Page 162
The alternative is to develop some way to permit individual conjuncts of coordinate phrases to bind all and only the
appropriate traces.
How serious is the problem of interwoven dependencies for Slash frameworks? An appropriate answer might be based
on the observation that interwoven dependencies pose essentially the same problem for Slash approaches that ordinary
coordination is rightly claimed to pose for transformational grammar.
(171) ''[T]ransformational grammar has never been able to capture such a unitary notion of coordination, for reasons
that were endemic to the framework." (Gazdar et al. 1985, 169)
The problem is illustrated by, for example, (172).
(172) a. Kim and Sandy met.
b. Kim sang and was accompanied by Sandy.
The authors observe that (172a) could only have been generated by base rules but that (172b) "had to be derived in a
completely different way, via a transformation of Conjunction Reduction, in any grammar that handled passive
constructions transformationally" (1985, 169). They then point out that analogous examples can be constructed for
"almost every transformation ever proposed" (1985, 170). Their claim is that generalizations are lost in that
Conjunction Reduction "had to be formulated in such a way as to produce structures that were isomorphic to those that
would have been produced if everything had been base generated in the first place."
To the extent that this line of argument against transformational grammar is correct, an equally correct analog based on
interwoven dependencies attacks Slash approaches as so far characterized, for in general each pair like (172a,b) has an
interwoven dependency correspondent.
(173) a. Kim and Sandy sang and danced, respectively.
b. Kim and Ruth sang and were accompanied by Sandy, respectively.
Just as Gazdar et al.'s (1985) argument indicates that transformational approaches fail to capture the generalizations
revealed in (172), so Slash approaches fail to capture those holding across (172) and (173). Notably, Gazdar et al.
(1985) provide no account of such examples and mention interwoven dependencies only as part of a criticism of the
transformational literature.

< previous page page_162 next page >


If you like this book, buy it!
< previous page page_163 next page >
Page 163
(174) "There is an analogy here with the respectively construction, which has similarly been a standard component of
the pedagogic commercial for TG [transformational grammar] despite the fact that no remotely adequate
transformational analysis of the construction was ever proposed.... "(Gazdar et al. 1985, 150)
But a theory that can treat the expressions of (172) in the same terms nonetheless fails if it cannot apply those terms to
(173). Thus, interwoven dependencies may pose roughly the same sort of challenge for Slash approaches as ordinary
coordinations like (172b) do for transformational ones. Such a challenge can be avoided only under pain of maintaining
a theory that denies that interwoven dependency structures occur in natural languages. 36

< previous page page_163 next page >


If you like this book, buy it!
< previous page page_i next page >
Page i

Three Investigations of Extraction

< previous page page_i next page >


If you like this book, buy it!
< previous page page_165 next page >
Page 165

Appendix A
Mistaking Selective Islands for Nonislands
A.1 Nonuniqueness of the Key Error
In section 3.2 I argued that the fundamental reason why A-Ss are not counter-examples to the CSC is that, despite the
attested extractions from them, all the conjuncts of A-Ss are islands. I showed that those conjuncts permitting
extractions are selective islands. Therefore, I claimed, Lakoff's argument from A-Ss involves a conflation of selective
islands with nonislands.
Given the nature of the constraints on selective island extraction, partially discussed in section 3.2, and the fact that the
extraction-facilitating RPs are invisible, mistaking selective islands for nonislands is extremely easy. I suspect that few
who have studied English extraction phenomena in any detail have fully avoided this mistake. I briefly survey other
discussions illustrating what is, I suggest, the same misstep, some leading to unfounded theoretical conclusions partially
comparable to that drawn from A-Ss in Lakoff 1986.
A.2 A Putative Case of "Reanalysis"
Chomsky (1977b, 127) discusses the grammatical extraction in (1).
(1) What1 did he [make a claim] that John saw t1?
Such examples seemingly violate Ross's Complex NP Constraint (CNPC) or equivalents. But Chomsky suggests in
effect that no violations exist because of a supposed "reanalysis," which yields the CNPC-consistent representation in
(1).
Motivation for this proposal vanishes when one observes that (1) instantiates selective island extraction, obeying the
constraints in (27) of chapter 3. Extraction of a broad range of non-NPs is blocked.
(2) a. [How long]1 did she (*make a) claim that Gregor dated Samantha t1?
b.*the way [in which]1 she made a claim that he earned his fortune t1
Extraction sites that are antipronominal contexts (ACs) are barred.
(3) a. What1 did she (*make a) claim that there was t1 in the safe?
b. [How much thought]1 did she (*make a) claim that Ernest gave t1 to the problem?
And grammatical extractions from the contexts Chomsky discusses cannot be A2-extractions (see (58b) of chapter 3).

< previous page page_165 next page >


If you like this book, buy it!
< previous page page_166 next page >
Page 166
(4) a. Marylou dated more officers than (what1) (*Ralph made a claim that) Lucille dated t1.
b. Marylou called [whatever officers]1 (*Ralph made a claim that) she should have called t1.
Even if spelled out precisely (which it never has been), Chomsky's "reanalysis" proposal would have no way to draw
the relevant distinctions. Like Lakoff's anti-CSC inferences from A-Ss, Chomsky's postulation of "reanalysis" to
account for (1) appears to reflect a failure to recognize selective islands. 1 See Baltin and Postal 1996 for criticisms of
other "reanalysis" proposals.
A.3 Preserving the Right Selective Island Constraint
It has long been widely assumed that the NP or PP extractees in complex NP shift (CXS) structures are islands. This
restriction is a key piece of evidence in work on the Freezing Principle (see Culicover and Wexler 1977, 21; Wexler
and Culicover 1980, 27-28; Culicover 1982, 335-336. Let us call the principle at work the Right Selective Island
Constraint (RSIC). The reasons for the specification selective appear in the discussion.
Kayne (1985, 1994) in effect disputes the island character of such constituents, citing (5a) and (5b) (from 1985, 131,
and 1994, 74, respectively) as well-formed CXS counterexamples to the RSIC.
(5) a. Mary is the only girl who1 I dared mention t2 to John [the possibility of him going out with t1]2.
b. the problem which2 I explained t1 to John [only part of t2]1
Although I agree that (5a,b) are well formed, such cases no more counter-exemplify the RSIC than Lakoff's A-S cases
counterexemplify the CSC. Like Lakoff's cases, (5a,b) are object NP extractions; and it is easily verified that such cases
represent selective island extraction, hence are largely limited to NP extraction satisfying (27b-e) of chapter 3. As a
result, the analog of (5a) in which the preposition with is moved immediately in front of who is sharply ungrammatical.
See also the non-NP extraction in (6).
(6)*[How long]1 did they try to mention t2 to John [the possibility of him remaining there t1]2?
Similarly, extraction from an AC inside a right island is impossible.
(7) a.*[What color]1 did they try to mention t2 to John [the possibility of him painting his yacht t1]2?
b.*What1 they tried to mention t2 to John [the possibility of him being able to tell t1]2 was her IQ.
Moreover, like other selective extractions, that from right extractee contexts can-not be an A2-extraction.
(8) a.*The doctor violated more guidelines than (what1) I dared mention t1 to Graham [the possibility of him violating
t2]1.
b.*The more problems Mary thinks up the more problems (that2) I will have to explain t1 to John [only part of t2]1.

< previous page page_166 next page >


If you like this book, buy it!
< previous page page_167 next page >
Page 167
So, Kayne's criticism of the RSlC logically parallels Lakoff's putative refutation of the CSC from A-Ss. Both depend
critically on a false assumption that a (selective) island is a nonisland. This view in turn hinges on an erroneous leap
from the grammaticality of object NP extractions from a constituent to a conclusion that such a constituent is a
nonisland.
A.4 A Denial That Certain Complex NPs Are Islands
Pollard and Sag (1994, 206) cite (9a,b) as grammatical examples.
(9) a. [Which rebel leader]1 would you favor a proposal that the CIA assassinate t1?
b. [Which Middle East country]1 did you hear rumors that we had infiltrated t1?
These authors infer from these object NP extractions that, contrary to Ross 1967 and much subsequent work, no
grammatical principle prevents extraction from such complex NPs; that is, there is no analog of the CNPC. But (9a,b)
represent essentially the same constructions that Chomsky (1977b) focuses on and that I have discussed in section A.2.
These examples are grammatical because they are selective extractions, satisfying (27) of chapter 3. Compare (10a,b),
which do not respect these conditions.
(10) a.*[How long]1 would you favor a proposal that the CIA keep him out of sight t1?
b.*[What color]1 did they hear rumors that he dyed his beard t1?
Further, (11) shows that, like other selective extractions, those discussed by Chomsky and by Pollard and Sag cannot be
A2-extractions.
(11)*They investigated [whatever rebel leaders]1 he favored a proposal that the CIA assassinate t1.
Again, then, a theoretical conclusion is not sound because it mistakes particular selective islands for nonislands.
Examples like (9) cannot justify the view that a grammatical framework can do without some principle that entails the
chief con-sequences of the CNPC.
A.5 The Island Status of Irrealis if Complements
Pullum (1987, 264) concludes that the irrealis if complements he studies are non-islands, unlike conditional if clauses,
on the basis of contrasts like the one in (12).
(12) a.*[Which commitment]1 will Joe quit if we cannot keep t1?
b. [Which commitment]1 would it be useful if we kept t1?
This is misleading in a now familiar way since the contrast does not show that (12a) involves an island and (12b) a
nonisland. The latter could be shown only by attesting grammatical extractions from irrealis if complements that violate
the conditions on selective island extraction. However, there seem to be no such cases.
(13) a.*Who1 would Frank prefer it if I believed t1 did it?
b.*Clever1 though the boss would prefer it if the new manager was t1,...
c.*[What kind of vampire]1 would you prefer it if he turned into t1?

< previous page page_167 next page >


If you like this book, buy it!
< previous page page_168 next page >
Page 168
Thus, at most (12) indicates that irrealis complements are selective islands whereas conditional clauses headed by if are
absolute islands, that is, do not permit the kind of control allowing some NP object extraction from islands. This
conclusion is consistent with nonextraction evidence of the type earlier seen to involve island sensitivity for example,
that relating to "affective" elements and multiple interrogatives.
(14) a. Nobody would prefer it if he fired someone/*anyone.
b.*Who would prefer it if she played how long?
A.6 Deep Extraction
Deane (1988, 1991) discusses what he calls deep extraction, which is extraction across more than one NP node. As he
observes, although it is predicted in (e.g.) frameworks incorporating some variant of Chomsky's Subjacency Condition
that such extraction should not exist, fairly well formed examples of the phenomenon can be attested, such as (15a,b)
(from Deane 1988, 100).
(15) a. [Which tribes]1 have you written articles about your fieldwork with t1?
b. There are certain books that1 atheists experience discomfort with the contents of t1.
But an assumption that (15a,b) threaten general conditions on extraction from deeply embedded constituents has the
same character as Lakoff's (1986) view that extractions from A-Ss threaten the CSC.
The gap-containing NPs in (15) are arguably islands since extractions from them satisfy the conditions on selective
island extraction. Like Lakoff's (1986) examples, all of Deane's (1988, 1991) relevant examples instantiate NP object
ex-traction. When non-NPs are extracted, the results contrast. Compare, for example, (15a) and (16).
(16)*[With which tribes]1 have you written articles about your fieldwork t1?
Of course, (16) is irrelevantly grammatical on a reading where the extracted phrase is not an object of fieldwork.
Compare also:
(17) a. [Which machine]1 did you write articles about a proposal to redesign t1 in that way?
b.*HOW1/[In what way]1 did you write articles about a proposal to redesign that machine t1?
Again, though, there is an acceptable reading where the extractee is not linked to a gap inside the NP headed by
articles.
Grammatical cases of deep extraction also respect the AC, reflexive, and passivizability conditions on selective island
extraction.
(18) a. It was that tower that1 they wrote an article about a proposal to (have Frank) paint t1 green.
b.*It was that color that1 that they wrote an article about a proposal to (have Frank) paint that tower t1.
c. What1 did they (*write an article about a proposal to) have her become t1?

< previous page page_168 next page >


If you like this book, buy it!
< previous page page_169 next page >
Page 169
d.*It was herself that1 they wrote an article about a proposal to have Frank describe Jean1 to t1.
e.*It was those rocks which1 they wrote an article about a proposal to let Frank feel t1 move.
Further, deep extractions of type A2 are impossible.
(19)*Armand has painted more towers than (what1) they wrote an article about a proposal to have Frank repair t1.
Finally, nonextraction evidence confirms that the NP contexts involved in deep extraction are islands.
(20) a. Nobody believed that they wrote an article about a proposal to fire someone/*anyone.
b. Who wants to (*write an article about a proposal to) fast how long?
Moreover, even one-level extractions from noun complements betray signs of being selective (see Koster 1987, 199).
(21) a. Who1 did Ethel discuss the possibility of talking to t1?
b.*[To whom]1 did Ethel discuss the possibility of talking t1?
c.*Slim1 though Ethel discussed the possibility of becoming t1,...
d.*[What color]1 did Ethel discuss the possibility of tinting her hair t1?
e. Nobody discussed the possibility of (*ever) lying.
f. Nobody discussed the possibility of eating (*a bite).
g.*They never mentioned the possibility that Ethel1, we might hire t1.
Overall, then, despite examples like (15), it may be possible to say simply, at least for English, that NP complements
are islands, the attested extractions from them falling largely within the domain of selective island extraction.
A.7 Apparent Support for a Novel Treatment of Finite Subjects
A significant parallel to Lakoff's conclusion from A-S extractions is provided by Chung and McC1oskey (1983)
(hereafter C&M). These authors observe that the GPSG treatment of certain finite subjects differs in a fundamental way
from the GPSG treatment of extractions in not involving Slash categories, discussed further in chapter 4. They then
argue that this unusual treatment is supported by certain asymmetrical extraction data. Specifically, they claim that an
additional extraction from a structure with a "missing" finite subject is grammatical or much better than multiple
extractions in general.
At issue are examples such as (22a,b) (from C&M 1983, 708-709).
(22) a. Isn't that the song that1 Paul and Stevie were the only ones [who wanted to record t1]?
b. This is a paper that1 we really need to find [someone who understands t1].
C&M argue that the GPSG treatment of finite-clause subjects explains why (22a,b) avoid the ungrammaticality
otherwise typical of multiple extractions (to the left) from a single English constituent; see, for example, (23) (from
C&M 1983, 709). The explanation is that the GPSG multiple-extraction constraint bans distinct Slash features on a
single node. Since the subjects (who in (22)) are described

< previous page page_169 next page >


If you like this book, buy it!
< previous page page_170 next page >
Page 170
in GPSG terms without Slash categories, the multiple-extraction constraint is not invoked. But in (23) neither gap is a
finite subject; hence, the ban on multiple extractions would be violated.
(23)*Isn't that the song that1: Paul and Stevie were the only ones that2 George would let t2 record t1?
C&M's insightful observations about cases like (22) and (23) are a model of marshaling intricate facts in favor of
complex theoretical conclusions. Nonetheless, the argument turns out to be unsound, the conclusion incorrect for now
familiar reasons. The constituents in (22) and the other examples cited by C&M that manifest a "missing" subject not
treated in GPSG terms as involving a Slash category permit one other extraction only as selective island extraction. All
the "secondary" extractions are NP object extractions satisfying (27) of chapter 3 and other conditions on selective
islands.
Consider parallels to (22) when the requirements on selective island extraction are not met.
(24) a.*[How long]1 was that the song that Paul and Stevie were the only ones who were willing to sing t1?
b.*Isn't that the song [with which]1 Paul and Stevie were the only ones who were capable of entertaining the children
t1?
c.*Isn't that the song that1 Paul and Stevie were the only ones who knew there was h on the hit parade?
d.*Isn't that the color which1 Paul and Stevie were the only ones who painted their yacht t1?
The constructions treated by C&M thus preclude non-NP extractions (see (24a,b)) and bar extractions from ACs (see
(24c,d)).
The reflexive constraint on selective island extraction also holds.
(25) It was herself that1 Paul and Stevie (*were the only ones who) agreed to talk to Joan1 about t1.
As does the passivizability restriction:
(26) Isn't that the rock which1 Paul and Stevie (*were the only ones who) felt t1 shake?
Finally, extractions from contexts like that in (22) cannot be A2-extractions.
(27) Frank pirated more songs than (what1) Paul and Stevie (*were the only ones who) wanted to record t1.
Nonextraction evidence also indicates that the contexts are islands.
(28) a. Nobody understood that Paul and Stevie were the only ones who (*ever) sang some/*any song/*a note.
b.*Who reported that Paul and Stevie were the only ones who sang how long?
Since (22a,b) represent selective island extractions, the difference between (22) and (23) in no way supports the GPSG
assumptions at issue. Those principles draw no distinction between the (mostly NP) extractions that satisfy the con-

< previous page page_170 next page >


If you like this book, buy it!
< previous page page_171 next page >
Page 171
ditions on selective island extraction and all other extractions and thus would wrongly predict that all of (24)-(27) are as
well formed as (22). The difference between ordinary and selective island extraction then undermines the apparent
support for the GPSG-theoretical principles C&M try to derive from facts like (22).
A.8 Apparent Support for a "Vacuous Movement Hypothesis"
Chomsky (1986a, 51) treats the phenomena discussed by C&M in terms of what he calls the Vacuous Movement
Hypothesis (VMH), stating, "The full range of facts follows from the VMH." Although it is not easy to determine what
the VMH even claims, it includes the idea that wh phrase movement is not obligatory at surface structure. This
assumption is supposed to interact with other ideas of Chomsky's distinguishing subjects from nonsubjects to account
for the contrast between, for example, (22b) and (29).
(29)*This is a paper that1 we really need to find someone that we can intimidate with t1.
Chomsky sketches the logic of the approach as follows:
(30) "In [(22b)] the VMH permits movement of the relative clause operator from t first to the specifier position of CP,
then to its final position, yielding only the very weak CNPC effect typical with someone as the head of the NP; ... In
[(29)] the corresponding derivation is impossible since the specifier position of CP is occupied by the fronted object of
intimidate, and the expression is less acceptable." (Chomsky 1986a, 51)
Obscurity aside, Chomsky's appeal to the VMH may fare somewhat better than C&M's account because, independently
of the VMH, his system has devices (e.g., his Empty Category Principle (ECP)) that differentiate subjects from objects
and objects from adjuncts. This offers a way of partially capturing the selective nature of the extraction in (22b).
Ultimately, however, Chomsky's approach also fails, because the asymmetries built into the ECP do not nearly
reconstruct the constraints on selective island extraction. Although the ECP might be claimed to entail condition (27b)
of chapter 3 and possibly, in combination with other assumptions, also (27a), it fails to predict in particular that analogs
of (22b) are ungrammatical if the object position is an AC, or if the extracted object is a reflexive or an inherently
unpassivizable NP.
For AC facts, compare (22b) with (31a-c).
(31) a.*That is a paper that1 we really need to find someone who thinks that there is t1 in that issue 2
b.*That is a color that1 we really need to find someone who painted their yacht t1.
c.*That is the amount of thought that1 we really need to find someone who gave those proposals t1.
As far as I can tell, (31a-c) would satisfy those of Chomsky's conditions that block (29) as much as (22b) does. They
are nevertheless unacceptable because the

< previous page page_171 next page >


If you like this book, buy it!
< previous page page_172 next page >
Page 172
extraction sites represent some of the ACs illustrated in (34) of chapter 3 and thus violate condition (27c) of chapter 3
on selective island extraction.
The relevance of the selective island reflexive and passivizability conditions (27d,e) of chapter 3 is seen in (32) and
(33).
(32)*It is herself that1 we really need to find someone who talks to Jane1 about t1.
(33)*That is a rock that1 we really need to find someone who felt t1 move.
The latter extraction manifests the passivization constraint seen in (40b) of chapter 3. Finally, Chomsky's approach fails
to determine that extraction from the con-texts at issue cannot be an A2-extraction.
(34) a. I read more papers than (what1) he needs to (*find someone who) read t1.
b. She will read [whatever papers]1 he needs to (*find someone who) read t1.
Since Chomsky's system fails to block any of (31)-(34) or to relate their violations to the more general selective island
conditions that determine them, ultimately his account is at best only marginally superior to the one C&M propose. Just
like Lakoff's anti-CSC argument from A-Ss, both fail in not recognizing that certain NP extractions instantiate
extractions from (selective) islands. 3,4
A.9 Implications
An ultimate methodological moral to be drawn from the cases of misidentified selective island extraction that have been
considered is something like the following. Given that selective extraction is possible out of many islands, to find
genuine counterexamples to claims that some constituent is an island, one must argue that the extraction fails the
conditions on selective island extraction. This requires showing in particular that a full range of non-NPs (not just
selected PPs) extract. Further, it must be documented that the NP extractions from the constituent at issue are free of the
constraints holding for selective island NP extraction, those partially represented by (27) in chapter 3 and the claim that
A2-extractions are impossible. Naturally, claims of islandhood are much stronger if buttressed by nonextraction
evidence such as that involving "affective" elements.
One theoretical inference is that no conclusion about extractions, islands, and so on, can be regarded as sound without
consideration of the differences between nonislands and selective islands. Attempts to support or attack theoretical
assumptions on the limited basis of extractions that are not sensitive to this difference run the risk of failing for the
same reasons documented in chapter 3. Another theoretical point is that, although obviously many problems remain,
true English exceptions to various classically proposed island conditions are far fewer in number and type than has
seemed to be the case.

< previous page page_172 next page >


If you like this book, buy it!
< previous page page_173 next page >
Page 173

Appendix B
Additional Arguments That Right Node Raising Is an Extraction
B.1 Remark
In this appendix I briefly disscuss several arguments that were not present in the original version of chapter 4 but that
appear to undermine a position like the one McCawley and Levine advocate and/or to support the view that RNR is an
extraction.
B.2 Strange Plural Right-Node-Raising Pivots
A further observation contrary to a view like McCawley's, which takes an RNR extractee to be in situ in all the
conjuncts, is provided by examples like (1a,b).
(1) a. The pilot claimed that the first nurse was a spy and the sailor proved that the second nurse was a spy.
b. The pilot claimed that the first nurse t1 and the sailor proved that the second nurse t1*[was a spy]1 /[were spies]1.
Compare (1 b) with (2).
(2) The pilot claimed that the first nurse was t1 and the sailor proved that the second nurse was t1 [a spy]1/*[spies]1.
Under the in-situ view, the conjoined clauses in (1b) involve singular subjects and plural predicates of a sort otherwise
unattested in English.
(3) The first nurse was a spy/*were spies.
Although examples like (1b) raise problems for any approach, at least under an extraction view there is the possibility
of seeing were spies in (1b) as some sort of realization of an n-ad of ATB extracted singulars. But under McCawley's
and Levine's proposals, there would appear to be no solution other than to require conjuncts of the unattested form in
(3), somehow limited to RNR cases.
B.3 Inverse Copula Constructions
There is another way in which RNR constructions involving NP extractees behave like L-extractions of NPs. For
reasons that need not concern us, the subject NP in a certain class of inverse copula constructions (see, e.g., Heycock
and Kroch 1996) cannot be extracted.

< previous page page_173 next page >


If you like this book, buy it!
< previous page page_174 next page >
Page 174
(4) Canonical copula cases
a You are the person most likely to win the election.
b. It is you who1 t1 are the person most likely to win the election.
c. Who1 t1 is the person most likely to win the election?
(5) Inverse copula cases
a. The person most likely to win the election is you.
b.*It is the person most likely to win the election who1 t* is you.
c.*Who1 t1 is you?
The same facts hold when such clauses are embedded as complements to verbs that permit raising to object; this point
is relevant since it means that the subjects are potentially targets of RNR, which finite subjects are not.
(6) a. I believe you to be the person most likely to win.
b. I believe the person most likely to win to be you.
c. It is you who1 I believe t1 to be the person most likely to win.
d.*It is the person most likely to win who1 I believe t1 to be you.
Significantly, then, whereas the raised subject of a canonical copula can be an RNR target just as it can be the target of
an L-extraction, the raised subject of an inverse copula can be neither.
(7) a. Ted believed t1 to be the persons most likely to win and Mike proved t1 to be the persons most likely to win [you
and two of your eight brothers]1.
b.*Ted believed t1 to be you (and two of your eight brothers) and Mike proved t1 to be you (and two of your eight
brothers) [the person(s) most likely to win]1.
Thus, in yet another respect RNR shares constraints with L-extractions, arguing that it is an extraction.
B.4 The Unextractability of Right-Dislocated Phrases
The right dislocation construction is illustrated in (8b,d).
(8) a. Sheila hates Bob.
b. Sheila hates him very much, Bob.
c. Sheila gave the diamonds to Ernie.
d. Sheila gave them to Ernie, the diamonds.
Now, a priori, it would be possible for right-dislocated phrases to be the targets of L-extractions. But in fact this is
impossible.
(9) a.*It was Bob who1 Sheila hated (1him) very much t1.
b.*[Which of the diamonds]1 did Sheila show (*them) to Ernie t1?
c.*the diamonds, which1 Sheila showed (*them) to Ernie t1,...
But exactly the same is true for their participation in RNR structures.
(10) a. Sheila hated (*him) very much t1 but Gladys loved (*him) very much t1 [the guy you just saw in the elevator]1.
b. Sheila may have shown (*them) to Ernie h and she certainly showed (*them) to Carl t1 [the rare sixty-carat
diamonds]1.

< previous page page_174 next page >


If you like this book, buy it!
< previous page page_175 next page >
Page 175

Appendix C
Reaction to Referee Comments
In this appendix I comment briefly on certain criticisms from one of two MIT Press referees, hereafter called MITR. I
address the selected comments in an order that seems to me to maximally connect them to the issues of this book.
With respect to the argument in section 4.2.1.2.5, MITR states:
(1) ''The argument against the extraction analysis of RNR given in section 4.2.1.2.5. is extremely powerful, much
stronger than ANY of the arguments for the extraction analysis of RNR that Postal gives. The argument, originally due
to Wexler and Culicover (1980), is based on the possibility of RNR into islands, illustrated by Mary buys and Bill
knows a man who sells, pictures of Elvis Presley. If RNR were an extraction operation, such constructions would be
expected to violate locality constraints on movement, i.e., Subjacency. [The construction in question should violate the
Complex NP Constraint.] On the base-generation analysis, on the other hand, the grammaticality of such constructions
can be straightforwardly accounted for. Postal claims that this argument for the superiority of the base-generation
analysis of RNR is very weak and puts it aside simply by exempting the RNR movement from locality conditions on
movement. According to Postal, in contrast to other extraction operations, the RNR extraction operation is simply not
subject to locality constraints on movement. (Postal actually suggests that all rightward extractions are exempted from
locality conditions on extraction. However, other rightward extractions clearly obey these conditions)."
My reactions to this criticism are as follows. First, MITR provides no real support for the subjective claim that the
argument against an extraction analysis is "much stronger than ANY of the arguments for the extraction analysis." But
he or she does try to give grounds for calculating the "strengths" at issue; I return to this below. Second, MITR
criticizes my suggestion that right extractions do not obey the relevant locality conditions by claiming that "other
rightward extractions" do obey them. Elsewhere MITR makes clear that he or she is referring here to phenomena like
extraposition of relative clause, extraposition of complements of head nouns, and extraposition of prepositional phrases.
My response would be that these phenomena are not extractions but fall into an overall class of bounded raisings,
which inherently involve an element taking on constituency within an immediately containing constituent.

< previous page page_175 next page >


If you like this book, buy it!
< previous page page_176 next page >
Page 176
MITR continues:
(2) "This is a VERY bad move. In a number of frameworks, for example the GB theory, it is simply not possible to do
this. All instances of movement in this framework are considered to be applications of the general operation Move a,
and the constraints on movement are stated as constraints on Move a. In such a framework it is simply not possible to
take one particular application of Move a and exempt it from locality constraints on movement, which the author does
not appear to realize."
The work on which MITR was commenting was, it is clear, not written in the GB framework. Thus, it makes no more
sense for MITR to cite as a criticism a sup-posed incompatibility between my suggestion and GB assumptions than it
would for a critic of work written in the GB framework to reject some proposed analysis because it was incompatible
with an assumption of, for example, the Head-Driven Phrase Structure Grammar. Clearly, one interpretation of a
supported analysis A truly incompatible with framework F takes A as counterevidence to F.
However, the key here is the concept "truly incompatible." Setting aside the factivity of the verb realize in the last line
of MITR's remarks, it is true that I do not realize that it is "simply not possible" in the GB framework to exclude RNR
from constraints holding for L-extractions. What MITR gives no sign of recognizing is that the claim amounts to an
assertion of a theorem and that to be serious, any claimed theorem needs to be associated with either a displayed proof
or a reference to one, neither of which is given (see Pullum 1983, 447, and section 3.5). What MITR has asserted in
effect is that GB consists of (among other things) a set of axioms AX such that the addition to AX of a further axiom B
equivalent to my claim that RNR is an extraction not subject to, for instance, Subjacency permits derivation of a
contradiction. Without a proof, though, MITR's claim amounts to another instantiation of the empty bluff parodied
under the name Phantom Theorem Move in Postal 1988.
Incidentally, I find it implausible that an actual proof could be constructed. Subjacency is stated in GB terms as roughly
a claim that movements cannot cross more than one barrier, often interpreted as locality conditions on the "licensing" of
empty categories. Surely, restricting this claim to a subclass of movements would not be fundamentally incompatible
with basic GB ideas. For instance, one could distinguish left traces from right traces and limit only the former in the
relevant way. In any event, the burden of proof is always on anyone who claims something is a theorem.
With respect to the argument of section 4.2.1.2.7, MITR states:
(3) "The argument against the extraction analysis discussed in section 4.2.1.2.7. is also very strong, much stronger than
Postal cares to admit. It is based on the fact that certain elements that are otherwise completely immobile, i.e., never
undergo movement, can undergo RNR. Most instances of RNR of otherwise immobile elements violate constraints on
movement or licensing of traces. For example, in the GB framework [(4a)] should be ruled out by the ECP under the
extraction analysis of RNR, on par with [(4b)], because t1 violates the ECP. This, and more generally the fact that
otherwise completely immobile

< previous page page_176 next page >


If you like this book, buy it!
< previous page page_177 next page >
Page 177
elements can undergo RNR, raises a very serious problem for the extraction analysis of RNR, which should be dealt
with and not simply dismissed, which Postal does.
[(4)] a. John believes [CP that t1] and Peter claims [CP that t1] [IP Mary will get a job]
b.*Who does John believe [CP that t1 will get a job]"
Like the earlier criticism, though, this one lacks a discernible logic. First, it again makes no sense to criticize a work not
assuming framework F for including an analysis that violates some principle of F (here the ECP). Second, recent work
internal to GB assumptions (e.g., Culicover 1991, 1993a,b) indicates that the prospects for explicating even facts like
(4b) via the ECP are bleak in the face of grammatical examples like (5).
(5) [Which woman]1 does John believe [that under those conditions t1 will testify against the sheriff]?
That is, there is no real reason to think consistency with the ECP is a reasonable condition to impose on an analysis.
Third, the fact that RNR targets include a range of constituents not subject to L-extraction does not in itself lead to a
conclusion that RNR and L-extractions fail to fall under a general concept of "extraction" any more than the (much)
less sharp differences between, for example, restrictive relative and nonrestrictive relative L-extractions show that these
fail to fall under a common concept. MITR's claim amounts to saying, "Look how different RNR and L-extractions
are." But this alone is not determinative, given that, as argued in chapter 4, they also are strikingly similar in a variety
of ways. Fourth, MITR's claim that RNR's greater target constituent freedom than L-extractions' "raises a very serious
problem" for an extraction analysis is empty if it refers to the ECP claim and unsupported if it refers to anything else.
MITR claims that works written after chapter 4in particular, Kayne 1994 and Bokovic * 1995contain arguments
favoring what is called a base-generation analysis of RNR as opposed to an extraction analysis. Under the particular
analysis that MITR apparently advocates, an RNR (NP) pivot is taken to be in situ in object position in the last
conjunct, and to have determined in some unspecified way the nonappearance of analogs in previous conjuncts.
Support for this base-generation view is supposedly provided by facts about negative polarity item licensing cited by
Kayne (1994). In particular:
(6) a. Mary bought, but John didn't buy any books about linguistics.
b.*Mary didn't buy, but John did buy any books about linguistics.
Although initially perhaps slightly seductive, this line of argument does not hold up when one considers what range of
RNR constructions it could apply to. In general, it is certainly not viable to claim that even an RNR NP pivot is in situ
in object position in the last conjunct. This is shown by examples no more complex than (7).
(7) a. Mary may have given t1 to Fred today and she certainly gave t1 to Louise yesterday [a large box of chocolates]1.

< previous page page_177 next page >


If you like this book, buy it!
< previous page page_178 next page >
Page 178
b. Ernie may believe t1 to be a conservative and certainly believes t1 to be a Republican [the excited critic of the
current welfare system]1.
c. Amanda may have talked to t1 about taxes and certainly talked to t1 about fur prices [the elegant banker from Buenos
Aires]1.
With respect to certain cases like (7a), MITR refers to Boskovic * 1995, claiming that Boskovic argues that in a case
like this, CXS is involved in the nonobject position of the RNR pivot, consistent with a base-generation analysis. But
this fails for (7c), since, as observed in note 18 of chapter 4, Ross long ago pointed out that objects of prepositions
cannot be CXS targets. With respect to cases like (7b), MITR claims that Boskovic 1995 contains a base-generation
analysis consistent with such cases. Not having seen that unpublished work, I cannot comment in detail. MITR
indicates, though, that Boskovic claims there that these cases do not involve RNR but rather ATB application of CXS.
As is well known, though, CXS (unlike RNR) is in general upward-bounded. Uncontroversial cases of CXS do not
permit a shifted heavy NP to appear further from the underlying site than the end of the nearest containing VP (see
Nakajima 1989; Postal 1993c). Therefore, regardless of the analysis of (7b), Boskovic's* approach would have no
application to cases like (8c), which contrast with genuine CXS structures like (8b).
(8) a. Mary said [(that) she believed Jack to be innocent in a very loud voice].
b.*Mary said [(that) she believed t1 to be innocent in a very loud voice] [the guy who was dripping blood in the
library]1.
c. [Mary said [(that) she believed t1 to be innocent in a very loud voice] [and Jack said (that) he believed t1 to be guilty
in a low whisper] [the guy who was dripping blood in the library]1].
This is because in (8c) the shifted NP must be regarded as adjoined at the end of a (conjoined) S constituent that is not
the end of the smallest containing VP, impossible in CXS structures, as (8b) shows.
Second, the base-generation/in-situ view is of course entirely impotent for those RNR structures that involve what are
called interwoven dependencies in section 4.2.2.14.
(9) Mary bought (from Gwen) and Frank stole (from Mike) a diamond and an emerald, respectively.
It would hardly be desirable to construct an analysis for such cases in which a diamond and an emerald forms the base
object of stole and determines deletion of some (what?) object in the first conjunct.
I think (8) and (9) show clearly that the view that RNR NP pivots are in situ in object positions and determine some
kind of right-to-left deletion has no real application to English RNR. That being so, Kayne's interesting observation in
(6) remains unexplained. Moreover, the exceptional plural RNR pivot cases in appendix B. could not exist if Kayne's
conceptualization of RNR as backward deletion was correct.
MITR also claims that Kayne's proposal explains certain facts about failures of RNR in Dutch. As I know nothing
about that language, I cannot comment. But no matter what virtues such a proposal might have for Dutch, they cannot
eliminate the grave problems it faces with the English facts cited here.

< previous page page_178 next page >


If you like this book, buy it!
< previous page page_179 next page >
Page 179
Consider next the question of "strength" of arguments for and against extraction views of RNR. MITR claims and tries
to justify the view that the arguments in sections 4.2.2.3-4.2.2.7 and 4.2.2.12 that RNR is a type of extraction are all
weak because they are either "based on a descriptive generalization that has no theoretical explanation" or "based on ill-
understood phenomena." MIRT then claims that "[i]n most cases it is possible to state the relevant descriptive
generalizations in such a way that the facts that motivated the generalizations can still he covered and the base-
generation analysis of RNR can be maintained." To exemplify this, MITR proposes, as an alternative to my claim in
section 4.2.2.3 that an indirect object (IO) cannot be extracted, the alternative shown in (10).
(10) An IO (an NP in indirect object position) cannot be phonologically null.
Given (10), MITR evidently believes there is no basis in IO facts for choosing an extraction analysis of RNR over, for
example, Kayne's proposal, because that too would provide for null IOs in relevant RNR cases.
One problem with this criticism, though, is that in most frameworks including those that MITR favors, (10) is simply
wrong; by contrast, my noextraction claim about IOs suffers from no known counterexamples. (10) is false in any
framework that recognizes null objects in passives, since it then says all IO passives like (62a) of chapter 4 and (11) are
ill formed, when they are in fact perfect.
(11) Marlene1 was sent t1 several invitations by possible suitors.
On the other hand, the extraction condition in section 4.2.2.3 survives cases like (11) under the weak assumption that
passives do not involve extraction. Hence, although it may be true that one or more of the generalizations underlying
my arguments could be adequately reformulated in a way that would not choose between an extraction view and, say,
Kayne's deletion view, it would take much more than MITR's failed attack on the IO argument to impose that
conclusion.
MITR's suggestion to replace conditions generalizing over both L-extraction and R-extraction with descriptive
statements like (10) raises another point. Even if this were possible in a descriptively adequate way, unlike the situation
with the IO argument, it would hardly support the denial that RNR and L-extractions fall under the same category of
extraction. It would still remain true that RNR and L-extractions would share a wide range of behavior. Otherwise, it
would he open to those denying their commonality to differentiate the null elements (e.g., traces) associated with L-
extractions from those (e.g., putative nontraces) involved with RNR. In short, the commonality would show up as
common treatment of potentially distinguishable null elements.
MITR cites the following paradigm as evidence for the base-generation view:
(12) a. John is, and the women want to he, doctors.
b. The women are, and John is, a doctor.
c.*The women are, and John wants to be, doctors.
d.*John is, and the women want to he, a doctor.
Perhaps there is a natural language with facts like these. But it is certainly not mine or that of my wife. For us, there is
no difference between the pairs (12a,d) and (12b,c), all four of which are hopelessly ungrammatical.

< previous page page_179 next page >


If you like this book, buy it!
< previous page page_180 next page >
Page 180
With respect to the argument of section 4.2.2.2, MITR raises the question of whether the CSC is a constraint on
extraction, noting that this is in particular denied in Munn 1993. MITR states that the Conjunct Constraint "is
particularly suspect as a constraint on extraction" and notes as evidence that "it is even obeyed by VP ellipsis, which
clearly does not involve extraction." MITR illustrates with the following example from Sag 1976:
(13)*I couldn't lift this rock but I know a boy who can GAP and bend a crowbar too.
I am not as certain as MITR that analyzing VP deletion as involving some sort of extraction is out of the question.
Favoring such an approach would be the fact that there are many similarities between VP deletion and the
uncontroversial VP extraction in (14).
(14) They said he can bend a crowbar and [bend a crowbar]1 he can t1.
For instance, both are impossible unless the ultimately empty VP site follows a restricted set of elements including
auxiliaries.
(15) They said he let Mary drink bourbon and
a.1he did let her GAP.
b.*[drink bourbon]1 he did let her t1.
Taking the VP deletion construction to include some kind of extraction of an empty VP is thus not beyond question.
However, the well-known failure of VP deletion to obey conditions like Subjacency is obviously a serious problem.
In any event, in 1997 1 am inclined to think MITR's basic point is correct here. The class of restrictions that Ross
sought to codify in the CSC do hold for a larger class of phenomena than are covered by the notion of extraction. For
instance, they arguably hold for the NP-internal genitive phenomenon.
(16) a. the interrogation of (that friend of) Jane('s) (and Louise) by the attorney
b. (that friend of) Jane's interrogation (*(that friend of) and Louise) by the attorney
If this is correct, then MITR is partially right to conclude that the argument in section 4.2.2.2 fails to support the
extraction view of RNR. I hedge here with partially for the following reason. Even though the CSC arguably holds for
nonextraction structures, it is still necessary to account for the class of structures for which it does hold in such a way
that they share enough structure to make this account viable. This condition may well still refute particular attempts to
regard RNR as a nonextraction phenomenon because the alternatives do not share any-thing of relevance with
extraction constructions or others subject to the CSC. I foresee that this is the case for proposals like that mentioned in
passing in Kayne 1994 to the effect that RNR structures involve left-to-right constituent deletion. I would conjecture
that in the framework of Johnson and Postal 1980 and derivative work, the CSC would end up holding of all and only
constructions involving successor arcs.

< previous page page_180 next page >


If you like this book, buy it!
< previous page page_181 next page >
Page 181

Notes
Chapter 1
1. The gap/coindexing notation in (1) and throughout this book is a descriptive device representing no commitment to
the linguistic reality of either traces or coindexing.
2. A probable exception is exclamatory extraction, whose ability to yield P-gaps is very questionable.
(i) [Which grapes]1 did she buy t1 without tasting pg1?
(ii) [What awful grapes]1 she bought t1 (?*without tasting pg1)!
See Obenauer 1992 for discussion of the failure of French exclamatory extraction to yield P-gaps.
3. This work was supported by NSF grant SBR-9510984 (to Paul M. Postal and Mark R. Baltin).
4. The contrast between A2-extractions and others with respect to the extraction of "backward" controllers of
complement subjects raises other theoretical issues. If I have understood the proposals correctly, the contrast reveals the
inadequacy of accounts of the so-called PRO gate phenomena related to weak crossover facts. The PRO gate effect,
noted by Higginbotham (1980, 1983), involves the failure of certain controlled elements in environments E to manifest
the same crossover violations as pronouns in situ in E do. The problem is that accounts of this effect, like that of
Demirdache (1991, 87-89), seem to offer no basis for the unacceptability of PRO gate cases with A2-extractions.
5. That P-gaps to the right of their "licensing" gaps are compatible with A2-extraction of the latter is indicated by
examples like these:
(i) a. Jacqueline met more candidates than Arnold interviewed t1 without hiring pg1.
b. I saw the same guy that you convinced t1 that they would interview pg1.
Evidently, the contrast between, for example, (15b,c) and (ia,b) raises an important issue for any theory of P-gaps.
6. Given note 2, it is evidently questionable whether (15a) provides independent support for the claim at issue.

< previous page page_181 next page >


If you like this book, buy it!
< previous page page_182 next page >
Page 182
7. The relation in question is and must be described here entirely informally. However, it has a simple formalization in
the Metagraph Grammar framework developed in Johnson and Postal 1980, Postal 1985, 1986a,b, 1989, 1990a,b, 1992,
1996, and other related works. Namely, to say that a specific RP, RPx, links to a particular extractee, Ex, is to say
minimally the following:
(i) RPx heads an arc B that is a replacer of an arc A1 headed by Ex hence, A1 and B have the same R(elational)-
sign and are neighbors (have the same tail node).
(ii) A1 is a foreign predecessor of an arc A2 and seconds B.
(iii) A2 erases A1 and has a remote successor whose R-sign is in the class of Overlay R-signs.
8. As with the issue discussed in note 7, although the notion of controlled RP is left entirely informal here, it can, I
believe, be precisely reconstructed in the framework mentioned in that note. Roughly, I would assume there is an
Overlay R-sign, Control. Then to specify that an RP is controlled is to specify that it must head a (successor) arc with
that R-sign. I would assume that it is mandated by linguistic law that each Control-arc is erased, such erasure in part
defining the control relation. Another aspect of this relation is that the arcs involved must, minimally, be related by the
anaphoric pairing relation.
In these terms, what needs to be said is that with the exception of the few marginal instances of noncontrolled RPs,
English extraction RPs must all head arcs with the R-sign Control, guaranteeing their extraction and their invisibility.
9. That the invisibility of what are taken here to be English controlled RPs is linked to their extraction rather than, for
example, merely to the possibility that certain in-situ pronouns are nonphonetic is argued by the existence of invisible
RPs that are completely insensitive to all island boundaries. As noted in Postal and Baltin 1994, this is the case with,
for example, certain instances of the RPs associated with French left dislocation. Although in general this construction
involves visible surface RPs occurring in situ, in certain cases a prepositional object RP is null. But this null RP can be
separated from the dislocated phrase by locked islands, as in (i).
(i) [Ce mur]1, on veut aider les gens qui craignent de se faire tirer dessus that wall are want to help the people who
are afraid of getting shot at par les terroristes embusqus derrire t1. by the terrorists hidden behind
That the islands here are locked is shown by the fact that A-extractions (e.g., question extraction, restrictive relative
extraction) from the position of the gap in (i) are systematically ill formed. Thus, it is important that any approach to
selective islands not confuse the sort of merely null RPs in (i) with English controlled secondary RPs. The latter,
beyond being invisible, show signs of being such only in connection with their having been extracted, which is what
underlies my claim that they fall under a notion of control.
10. Of course, another logical possibility is to take (32a,b) as topicalizations of non-RPs, the topic extractees then being
subject to either left dislocation or right dislocation. But this alternative appeals to something I do not know to be inde-

< previous page page_182 next page >


If you like this book, buy it!
< previous page page_183 next page >
Page 183
pendently attestable, namely, one L-extraction type taking as its extractee the extractee of a different L-extraction
type.
11. Another domain relevant to supporting the idea that elements can selectively extract from subjects involves P-gaps,
whose presence inside islands not containing the "licensing" gap is, in my view, due to extraction from flexible
unlocked islands. However, space considerations preclude delving into the domain of P-gaps here.
Chapter 2
1. That the relevant NPs here after as and into are PNs is argued in Emonds 1985, chap. 6.
2. See Emonds 1979 for a detailed analysis embodying the claim that the wh form in an appositive is a definite
pronoun.
3. After writing this, though, I became aware of strong evidence that nonrestrictive relative extraction involves an RP
and that this extraction cannot be reduced to A-extraction of a definite pronoun (represented by the wh form): namely,
even nonrestrictive relative extractions in which the wh form is embedded in a larger extracted phrase (so-called pied-
piping cases in the sense of Ross 1967) are excluded from antipronominal contexts. Thus, (ib,c) are no better than (ia).
(i) a.*He likes green1, which1 I painted my house t1.
b.*He likes Jane, [whose1 favorite color]2 I painted my house t2</sb..
c.*They favor that color, [no discussion of which1]2 there was t2.
The ill-formedness of cases like (ib,c) argues that there are pronouns in the positions of t2 and t2, RPs clearly
distinct from the non-RP pronouns represented by the forms whose and which.
4. With respect to the plausibility of (50c), much independent evidence indicates that (visible) RPs can extract (see
McCloskey 1979, 94-97; Sells 1984, 91-94; Koster 1987, 63). Sells documents that Hebrew relative clause RPs
optionally L-extract; they can then appear in the extraction site, in any Comp between that and the ultimate extraction
locus, or in the latter.
5. The kind of extraction at issue is referred to in GB terms as long extraction, for framework-internal reasons not
relevant here.
Chapter 3
1. Although Ross initially defined the CSC exclusively in terms of movement rules, he later generalized the constraints
to cover other types of generative rules as well.
(i) "Variables in chopping rules, feature-changing rules, and unidirectional rules of deletion cannot cross island
boundaries; variables in other rules can." (Ross 1967, 289)
This generalization becomes marginally relevant below, where evidence is cited that certain scope, multiple-
interrogation, and control phenomena obey the CSC.

< previous page page_183 next page >


If you like this book, buy it!
< previous page page_184 next page >
Page 184
2. Like most discussions of coordination, this remark, ignores what I refer to elsewhere (including chapters 1 and 4) as
interwoven dependencies, represented by respectively constructions in English. For example, to true coordinate cases
like (4) there systematically correspond cases like (i).
(i) [Which car]1, [which van]2, and [which motorcycle]3 did respectively Sally buy t1, Marilyn borrow t2, and
Lucille wreck t3?
This is in a sense ATB extractionin which, however, the gaps in the associated conjuncts correspond in an ordered
way to conjuncts of a single apparent extractee. See Postal, in preparation b, for further discussion.
3. Remarkably, in the generative literature of the past twenty-five years closely linked to Chomsky's work, the CSC
tends to go unmentioned. I find no citation or reference to it in, for example, Chomsky 1972, 1975, 1977a,b, 1980,
1981, 1982, 1986a,b, 1988, 1995, McCloskey 1988, Cinque 1990, Rizzi 1990, Haegeman 1991, Manzini 1992, Lasnik
and Saito 1992, Kayne 1984, 1994, Mller 1995, or Brody 1996. Lasnik and Uriagereka (1988, 168) devote one short
footnote to it, specifying that it is "powerful." Koster (1987, 358-359) mentions it only in passing. Exceptionally in this
tradition, Van Riemsdijk and Williams (1986) discuss it extensively, also referencing Pesetsky 1982.
Napoli (1993, 401, 409) breaks with tradition and makes explicit the Chomskyan program's failure to subsume the
CSC:
(i) "Notice that while Subjacency accounts for the [Complex NP Constraint], the [Specified Subject Condition], the
Subject Condition, and the wh-islands, it cannot account for the ungrammaticality of movement out of coordinate
structures and out of adverbial clauses. Thus Subjacency is an improvement over the earlier approach of individual
constraints and conditions because it captures a generalization that the earlier approach missed. But we still would like
to see a motivated account for the islandhood of coordinate structures and of adverbial clauses...."(1993, 401)
Most striking is the lack of reference to the CSC in Chomsky 1977b. This work (p. 86) attempted to characterize
wh movement" in terms of its observation of principles, including Chomsky's own Wh-Island Constraint and Ross's
CNPC. But Lakoff's (1986) objections aside, surely respecting the CSC is at least as characteristic of the phenomena at
issue as respecting either of the other constraints, which are clouded by problems, including what I have called selective
island extraction in section 3.2.2.
Goodall (1987) attempts to integrate the CSC into post-1973 Chomskyan frameworks; however, this attempt
appeals to highly radical notions (e.g., that apparently coordinate constituents are not even constituents) and sharp
departures from standard Chomskyan assumptions. Space considerations preclude an adequate discussion here.
4. Thus, extraction is possible out of Belauan sentential subjects, relative clauses, embedded interrogative clauses, and
adverbial clauses, all of whose analogs in English preclude ordinary extraction (but not necessarily selective
extraction).

< previous page page_184 next page >


If you like this book, buy it!
< previous page page_185 next page >
Page 185
5. More accurately, Goldsmith's discussion involves a certain equivocation. Initially he states that ''[t]he immediate
subject of this paper is a set of sentence-types which form an exception to the Coordinate Structure Constraint of Ross
(1967)..." (p. 134). But at the end of the article he informally proposes a kind of "reanalysis" of the relevant coordinate
structures into noncoordinate ones (see his (20) and (23)). Under such an analysis, there would presumably be no true
counterexamples to the CSC.
6. One cannot say that the "consequent" conjunct is entailed by the preceding conjunct(s). Rather, these conjunct(s)
partially designate members of a premise set including unexpressed premises, which jointly entail the consequent. In
(6a) one unexpressed premise is that de Gaulle is a Frenchman.
7. Whether mixed cases involving D-Ss are possible as well is unclear. At issue are cases like (i), whose status is
somewhat cloudy.
(i) Harry went to the store, bought three pizzas, ate them, and therefore is not hungry.
I suspect that to render (i) fully acceptable, an occurrence of and should precede ate. But this yields a binary VP
structure, whose initial part happens itself to be a linear structure (analyzable as an A-S, on one reading).
8. Given the fact, explicit in (10), that all putative counterexamples to the CSC involve extraction from VPs, Lakoff's
general rejection of the CSC is unwarranted even if, contrary to what I argue in what follows, the VP examples
represent genuine counterexamples. The strongest conclusion his data would justify is a restriction of coordinate
islands to non-VPs.
9. In fairness, I should indicate that in a series of personal communications in 1993 and 1996, Lakoff rejects the claim
that (12) reconstructs the logic of his argument against the CSC. This rejection is based on the (correct) observation that
(12) fails to make explicit various Cognitive Grammar assumptions, which Lakoff believes are central to the argument.
But I do not agree. Since what is at issue in the present work is the viability of the CSC and not anything about
Cognitive Grammar, there is no way that aspects of the latter can be relevant to the narrow concern of this chapter. At
issue is simply whether English includes well-formed sentences that are inconsistent with the CSC, a factual question
on which Cognitive Grammar assumptions can have no bearing.
10. There might seem to be a well-grounded principle justifying Lakoff's inference from partial ATB extraction.
(i) If a single extractee E is linked to each of a nonunary set G of gaps, then each of G occurs in a distinct
coordinate conjunct.
Claim (i) might be taken to reconstruct Lakoff's (1986, 157) flat statement: "Only conjunctions permit across-the-
board extraction." But a principle like (i) is dubious. As Hudson (1976) observes, the extraction associated with right
node raising is not limited to coordinate structures; see Postal 1994 and (ii).

< previous page page_185 next page >


If you like this book, buy it!
< previous page page_186 next page >
Page 186
(ii) Anyone who plans to get engaged to t1 should confer with someone who has dated t1 [that sort of person]1.
Examples like (ii) refute (i), barring either an extreme view of coordination such as that suggested by Williams
(1990), who contemplates taking such cases to be coordinate, or a rejection of the claim that right node raising is an
extraction. The latter issue is the subject of chapter 4.
11. That the quantifier both differentiates true coordinations from the kinds of structures discussed by Lakoff seems to
have first been observed by Schmerling (1975), as discussed by Na and Huck (1992).
12. In characterizing the semantics of A-Ss, Lakoff (1986, 153) states that their conjuncts represent "normal
conventionalized expectations." But conventionalized appears incorrect, as I see no difference among the versions of
(i), which differ with respect to satisfying the condition.
(i) the baby clothes which1 Greg went to the store/airport/brothel/farm and bought t1
13. Principle (25) has not gone unchallenged. Cinque (1975, 1977) and Hirschbhler (1974, 1975) argue for Italian and
French, respectively, that certain apparent RP-linked topicalization constructions are subject to island constraints.
However, in Postal 1991 I counter that data of the sort these authors consider are consistent with (25) under a
nonstandard analysis of the relevant constructions. Although the arguments for this analysis have many weaknesses,
rejecting (25) on the basis of the Romance facts is at least debatable.
14. Statement (26) subsumes the general principle that all pronouns are surface forms as well as the weaker entailment
that all RPs are phonetically realized. The evidence against either variant is massive. It can be divided into what one
could call straightforward evidence and more indirect evidence. The former involves patterns in which certain contexts
X contain visible pronoun occurrences, with a restricted subset of X permitting invisible pronouns. One thinks of the
so-called null subject phenomenon, for example, especially in languages like Spanish, where null subjects are optional.
My 1963 doctoral dissertation (see Postal 1979) was largely devoted to justifying the positing of such invisible
pronouns as both subjects and objects. Parallel phenomena are common in the world's languages.
By more indirect evidence I refer to arguments showing that particular, often subtle and idiosyncratic constraints on
pronouns manifest themselves in a variety of contexts, as a result of the presence of invisible pronouns in those
contexts. In Postal 1970 I advanced various arguments for the presence of invisible pronouns in English control
structures. A different example is that the pronoun constraint found in (ia,b), which contain visible pronouns, manifests
itself in the short examples (iia,b), which do not.
(i) a. She was born in [that building]1 but I wasn't born in it1.
b.*She was born in [Albania]1 but I wasn't born in it1.
(ii) a. That building1 is too horrible for her to be born in (it1).
b.*Albania1 is too horrible for her to be born in (it1).

< previous page page_186 next page >


If you like this book, buy it!
< previous page page_187 next page >
Page 187
For extensive arguments of this sort, see chapters 1 and 2, the rest of this chapter, and Postal 1993a, 1994, in
preparation a,b.
15. The issue of PP extraction from selective islands is extremely complicated. Ultimately, one must, I believe,
distinguish both different types of PPs and different types of selective islands. Selective islands may well form a
hierarchy, some being more resistant to PP extraction than others. At the same time, some PPs are more extractable
from selective islands than others. This is particularly clear in the case of relatively weak selective islands, like those
formed by negativesso-called inner islands, first studied by Ross (1984). Compare (i) and (ii) (the latter from Ross
1984, 260).
(i) a. He did(n't) pour beer into the jars for Sally.
b. [Into which jars]1 did(n't) he pour beer t1 for Sally?
c. [For whom]1 did(*n't) he pour beer into the jars t1?
(ii) It was [with this stiletto]1 that they (*never) stabbed the lasagna t1.
Some selective islands, including those involving relative clauses with quantifier heads, seem to ban all PP
extraction.
(iii) a. They arrested everyone who poured beer into the jars.
b. It was [those jars]1 that they arrested everyone [who poured beer into t1].
c*It was [into those jars]1 that they arrested everyone [who poured beer t1].
d.*It was [with this stiletto]1 that they arrested everyone [who stabbed the lasagna t1].
In some cases PP extraction from selective islands is both speaker-dependent and dependent on type of preposition and
construction. For example, Hornstein and Lightfoot (1991, 389) specify that (iv)(v), involving extraction from an
interrogative selective island, have the same status.
(iv) Who1 did John wonder whether Bill talked to t1?
(v) [To whom]1 did John wonder whether Bill talked t1?
But for me, (v) is as impossible as either version of (vi), whereas Lightfoot (personal communication, January
1992) makes the distinction given.
(vi) the guy [to whom]1 they asked John whether she talked/*mattered t1
The issues linked to PP extraction from selective islands also involve such superficially prepositionless adverbial
forms as when, where, why, and how. It is impossible to extract the latter two from any selective islands (see (vii)), and
impossible to extract any of them from the strongest selective islands (see (viii)). (All of (vii) are good, but only when
the adverbials are construed with arrest.)
(vii)*When1/*Where1/*Why1/*How1 did they arrest everyone [who protested t1]?.
(viii) ?When1/Where1/*Why1/*/How1do you regret [you hugged her t1]?
Although the principles underlying PP extraction from selective islands are important and mysterious, the issue is
fortunately marginal to the defense of the CSC. See note 26.

< previous page page_187 next page >


If you like this book, buy it!
< previous page page_188 next page >
Page 188
16. Relevant to the hedge is the question of whether complement clauses are extractable from selective islands. I find a
difference between, for example, the short and long versions of (i).
(i) [That (*he was a psychopath)]1, they never asked me whether I believed t1.
17. More precisely, S cannot be the highest finite subject position in the island. Extraction of nonhighest finite subjects
is less degraded (see Rizzi 1990; Lasnik and Saito 1992).
The discovery that various constituents allowing certain NP extractions bar finite-subject extraction really marks the
beginning of the recognition of selective islands. The earliest observations seem to have been made in the late 1970s by
Richard Kayne, who spoke of "subject-object asymmetries" (see, e.g., Kayne 1984, 4). The next step seems to have
been the recognition by Koster (1978b) and Huang (1982), among others, that the contexts discussed by Kayne also
preclude the extraction of adjuncts.
18. Although (27d) is stated in full generality, all the evidence in the text involves antecedence by nonsubjects. This is
the case for which the facts seem clearest, and for present purposes it would not matter if one reformulated (27d) to
limit it to these.
19. As stated, (27e) is rather obscure, but its obscurity is not directly pertinent to the current argument. For present
purposes, (27e) can be taken as nothing but shorthand for a listing of specific restrictions including those in the text
given by ostension. Theoretical reconstruction of (27e) is attempted in Postal 1990b, 1993a, 1994, but not very
successfully.
20. The idea that selective island extraction depends on invisible RPs seems to originate in the work of Hans-Georg
Obenauer and later Guglielmo Cinque (see Obenauer 1984, 1985, 1986; Cinque 1990). For general discussion, see
Koster 1987 and chapters I and 2. Neither Cinque nor Obenauer seems to have explicitly related the postulation of RPs
in selective island extractions to Ross's (1967) principle (25). The earliest proposal associating invisible pronouns with
extractions was made by Perlmutter (1972), who, however, posited them in all (at least NP) extractions, thus failing to
draw the distinction made in (58) below between A-extractions and B-extractions.
The view that extraction from selective islands depends on RPs combines with the observation in note 15 that some
PP extraction is permitted from (some) selective islands to require postulation of some invisible RPs corresponding to
PPs, as in Obenauer 1984, 1985. In this regard, it is worth investigating whether the fact that when/where are
sometimes extractable from selective islands whereas how/why are not is linked to the existence of then/there and the
absence of corresponding anaphoric elements for how/why.
21. This claim was refined in chapter 2, to refer to a division of pronouns into two types and the grammatical level(s) at
which the constraints defining ACs are stated.
22. A now well-known fact (see, e.g., Cinque 1990, 109, and chapter 1) is that an extraction site in a selective island y
cannot be separated from the boundaries of y by (many) other island boundaries. For example, compare (ia) and (ib).

< previous page page_188 next page >


If you like this book, buy it!
< previous page page_189 next page >
Page 189
(i) a. the stuff which1 I would prefer it if Jane bought t1
b.*the stuff which1 I would prefer it if Jane, who bought t1, delivered it to us
In (ib) the extraction site is separated from the selective island boundary (marked by "") by the distinct island
boundary (marked by ""). The ungrammaticality of such cases potentially raises an objection against appealing to RPs
for cases like (ia) since it might appear that Ross's view that RPs "save'' structures from island violations, which I have
adopted, would predict no difference between (ia) and (ib). But, as discussed in chapter 1, a solution to this problem lies
in the invisibility of the RPs posited for selective island extraction, if, as claimed there, this involves a control relation
that determines extraction of the relevant RPs: namely, it can be claimed that in order for RPs to be controlled, they
must themselves in general be extracted. In (ia) what was called the primary RP in chapter I must extract to the same
constituent as which. Since in doing so it crosses an island boundary, it must link to a secondary RP. Thus, what is
wrong with (ib) is that the extraction of the RP to the corresponding position crosses the inner island boundary, but
unlike the extraction of which, does not involve its own RP. Presumably a principle to the effect that (at least certain
classes of) extracted RPs cannot have (distinct) associated RPs in their extraction sites would guarantee this result.
23. The only work I am aware of that recognizes that A-Ss are selective islands is Culicover 1990. Culicover observes
independently that PPs, adverbs, and color-NP expressions cannot be extracted from A-Ss; he provides the examples in
(i) (1990, 14), though I have changed his "??" prefixes to "*"s.
(i) a.*[How fast]1 did Mary go to her friend's house and drive the car t1?
b.*[At which bar]1 did John go to Paris and sing "The Marseillaise" t1?
c.*[What color]1 did Mary come home and paint the fence t1?
Despite having noticed the selective island character of A-Ss, Culicover (1990, 14) concludes, without citing Lakoff
1986, that such constructions are incompatible with the CSC: "The fact that it is possible to extract out of one conjunct
shows that there can he no general constraint against such nonparallel extraction contrary to the prediction of the
Coordinate Structure Constraint." That is, Culicover does not recognize that the selective island status of A-Ss justifies
rejecting a premise like (12e).
24. Lakoff states a principle designed, among other things, to guarantee that extractions from A-Ss must involve a gap
in the final conjunct.
(i) "Only scenarios of Type B permit there to be no extraction from the final conjunct." (Lakoff 1986, 154)
This is poorly framed, because, taken literally, it wrongly blocks all A-S cases with no extractions at all. Moreover,
it does not seem to capture the essential contrasts, failing to fully cover the asymmetries in extraction possibilities from
A-S conjuncts. Although it accounts for (e.g.) (iia) versus (iib), it leaves (iia) versus (e.g.) (iii) or (49c,d), which satisfy
it, unexplained.

< previous page page_189 next page >


If you like this book, buy it!
< previous page page_190 next page >
Page 190
(ii) a. the wines which1 Steve got in your car, drove to the store, bought t1, and drank t1
b.*the wines which1 Steve got in your car, drove to the store, bought t1 and went home
(iii)*the car which1 Steve got in t1, drove to the store, bought wines, drank them, and drove t1 into a river
25. Na and Huck (1992, 260) support the asymmetry of extraction from different conjuncts of A-Ss with data that are,
ultimately, partly irrelevant. Specifically, in their example sets (46) and (47), designed to illustrate the impossibility of
extraction from the first conjuncts of binary A-Ss, three of the four starred examples involve non-NP extraction. These
fail to differentiate the claim that these conjuncts are absolute islands from the claim that they are selective islands.
Their fourth starred example is, however, an instance of NP extraction.
26. As with other selective islands mentioned in note 15, A-Ss allow certain types of PP extraction.
(i) a. [Into what jars]1 did Melissa rush downstairs and pour beer t1?
b.*[In what way]1 did Melissa rush in feverishly and pour beer t1?
c. Where1 did Melissa pick up the phone and arrange to meet Greg t1?
d.*Why1 did Melissa pick up the phone for that reason and arrange to meet Greg t1?
My casual impression is that A-Ss fall somewhere toward the weaker end of the selective island hierarchy with
respect to permitting PP extraction.
27. That some speakers find embedded topicalization possible is noted by Lasnik and Saito (1992, 92-104).
28. In the face of (i) and (ii), this principle may require an exclusionary ATB limitation partially parallel to Ross's (3)
and the principle mentioned in note 32.
(i) a. Mark knows that Mozart, I like and that Beethoven, I adore.
b.*Mark knows that I like Mozart and that Beethoven, I adore.
c.*Mark knows that Mozart, I like and that I adore Beethoven.
(ii) a. Harry, Joan respects and Zeus, she believes that I worship.
b.*Joan respects Harry and she believes that Zeus, I worship.
c.*Harry, Joan respects and she believes that I worship Zeus.
Alternatively, it might be correct to restrict principle (73) to noncoordinate islands and to develop a distinct
approach to (i) and (ii).
29. Pesetsky (1982, 618-619) makes the important observation that in coordinate structures, in-situ wh forms behave
significantly like extractions.
(i) a.*Which article proves your theorem and defends which theory?
b. Which article proves which theorem and defends which theory?
As he notes, (ib) is "exactly analogous to the ATB extractions of" the CSC. If, as I believe, the parallelism between
(i) and ATB extraction cases is genuine and systematic, then a proper theory of extractions and in-situ wh forms must
have significantly parallel features. This may be possible in certain approaches, for ex-

< previous page page_190 next page >


If you like this book, buy it!
< previous page page_191 next page >
Page 191
ample, in GB, where such theories have been sketched on the basis of "movement in Logical Form" (see Huang
1982; Pesetsky 1987; Lasnik and Saito 1992). In other, more superficial frameworks, though, it is difficult to see any
way to capture the parallelisms.
30. There seems to be speaker variation here. Although Lakoff (1986, 165) says flatly that subordinate clauses like that
in (i) do not permit extraction, I find (i) essentially acceptable (see Chomsky 1982, 72; Koster 1987, 158; Pollard and
Sag 1994, 191).
(i) Who1 did he sit there while referring to t1?
Note that (i) satisfies the conditions for selective island extraction and is evidently vastly superior to, for example,
(99a,b), which do not.
31. Compare (107d), involving extraction from an AC, with the parallel but grammatical (i), involving extraction from a
context permitting pronouns.
(i) [Which ear]1 did Mike go color-blind and still paint t1 green?
32. My suggestion that (108a-c) are probably grammatical conflicts with claims made by Na and Huck (1992). They
propose principle (I) and indicate that it correctly blocks (iia,b).
(i) The Condition on Asymmetric Conjunction
"In any asymmetrical conjunction, if extraction is performed on a secondary conjunct, it must be performed across-
the-board." (Na and Huck 1992, 159)
(ii) a.*What1 can we (destroy many lakes and) not arouse t1?
b.*[What kind of hangover]1 can you (drink a lot and) not get t1 the next morning?
However, although (iia,b) are ungrammatical, this remains true even when the parenthesized material is not present.
Hence, what is wrong with (iia,b) may well involve factors independent of extraction from B-conjuncts.
33. Actually, the claim in the text is at best only true of the and type of conjunction (see Ross 1967, 275-277).
Disjunctions contrast, for reasons that need not concern us here. In this regard, compare (109b) with (i).
(i) Nobody said that Sally met a doctor here or charmed a lawyer anywhere.
34. As in note 33, the claim is not true of disjunctive cases. Compare (113a) with (i).
(i) Nobody thought he lifted a finger or worried about Bob.
35. Na and Huck (1992, 270, n. 2) remark, "The ... constraint against the movement of conjuncts, is inviolable; thus all
of the exceptions to the CSC that Ross (and others) observed are exceptions to the second part of the CSC." How-ever,
it might be argued that the Conjunct Constraint, as informally stated, is mildly equivocal in that it is unclear whether it
is intended to block (ic), (id), or both.
(i) a. Helen is tall and slim.
b. [tall and slim]1 though Helen is t1
c*slim1 though Helen is tall and t1
d.*[and slim]1 though Helen is tall.t1

< previous page page_191 next page >


If you like this book, buy it!
< previous page page_192 next page >
Page 192
36. Here and below I represent the position of VP anaphora with angled brackets cosubscripted with an antecedent
constituent placed in the same brackets. No theoretical claims attach to this notation.
37. As would be expected under the adjunct view of B-conjuncts defended in section 3.3, there are B-S examples
parallel to (154).
(i) Sam agreed to VP fire a gun1, because his father once did VP1 and still got hired.
Chapter 4
1. A quite different type of incompatibility between later Slash assumptions and RNR is noted by McCloskey (1986,
183-184).
(i) "Given the framework developed by Gazdar, Right Node Raising was simply the expected result of the
interaction between Rightward Displacement and the general schema for coordination. More recent developments in the
theory ... make that analysis unavailable .... This is because many of the classic instances of Right Node Raising ...
would involve a violation of the condition ... that the domain of metarule application should be restricted to the set of
rules that introduce lexical categories." McCloskey suggests overcoming this problem by rejecting the view that RNR is
an extraction phenomenon, precisely the position I will argue against in section 4.2.
2. Ojeda (1987) seems to advance a view in which L-extractions are conceptualized in roughly the way McCawley
conceives of RNR; space considerations preclude discussing this.
3. The general validity of structures involving multiple mothers is not at issue. The framework developed in Johnson
and Postal 1980 and Postal 1982, 1985, 1986a,b, 1989, 1990a,b, 1992, 1996, exploits multiple mothers far more
extensively than McCawley contemplates. In question here is how this idea relates to RNR and ATB phenomena in
general and whether, as McCawley claims, surface representations manifest multiple mothers, a view denied in the
framework just cited.
4. Note though that in Gazdar's (1981) terms, RNR and extraposed relatives are described by the same schema; see
(142) below.
5. There is reason to doubt the truth of (15) if, as I believe, the following data are correct:
(i) the person who1 Clara talked about certain carvings of t1 to Lucille
(ii)*the person who1 Clara talked about t2 to Lucille and Mike talked about t2 to Jack [certain carvings of t1]2
Although I do not understand this difference, if real, it shows that RNR pivots do not always match their
correspondents in non-RNR cases in what Wexler and Culicover call "analyzability." But the degree of failure so far is
limited. That is, (15) should properly be regarded as a biconditional that says, schematically:
(iii) Let P be a non-RNR pivot position and R an RNR pivot position corresponding to P. Then

< previous page page_192 next page >


If you like this book, buy it!
< previous page page_193 next page >
Page 193
a. if constituent C in P can L-extract, then constituent C in R can L-extract; and
b. if constituent C in R can L-extract, then constituent C in P can L-extract.
The pair (i)-(ii) only falsities (iiia). This is, however, an argument against McCawley's multiple-mother-node
approach to RNR pivots.
6. In early work like Postal 1974, RNR is taken to be some kind of transformational operation on distinct pieces of
conjuncts that meet some identity condition. The operation somehow reduces these conjuncts to a single RNR pivot.
Clearly, then, an analog of the possibilities in (17) arises in such a framework. The RNR transformation could operate
on the output of the L-extraction transformation, or conversely.
7. For a very different approach to interactions between RNR and L-extraction, see Oehrle 1990.
8. Such a conclusion is drawn by Oehrle (1990, 412) from the fact that "RNR may involve sequences of constituents ...
of a sort never found in leftward ex-traction," an observation he notes is due to Abbott (1976). Oehrle cites contrasts
like this:
(i) Algernon didn't hand t1 he threw t1 [the cucumber sandwiches at Cecily]1.
(ii)*[The cucumber sandwiches at Cecily]1, Algernon didn't hand t1 he threw t1.
Such contrasts are indeed troubling, as the grammaticality of (i) for those who accept it has no straightforward
description in an extraction approach to RNR. Any attempt to appeal to reiterated RNR on single constituents runs into
the question of what guarantees a word order in the pivot position parallel to that allowed in non-RNR cases. Although
I have no serious analysis, I note that the phenomenon in (i) may also contrast with standard RNR cases. For instance, I
find a difference between (iiia) and (iiib).
(iii) a. Bob didn't mail t1 to Jane (yesterday) but he did fax t1 to Clara (today) [threats against the president]1.
b. Bob didn't mail t1 (*yesterday) but he did fax t1 (*today) [threats against the president to Jane]1.
9. Manifestly, in a specific range of cases, some extractions are more general than others. For instance, the negative
fronting construction is found in many environments where topicalization is impossible.
(i) a. [No other colors]1 did he think they had ever painted their car t1.
b.*[Some other color]1, he thought they had painted their car t1.
(ii) a. [Nothing of the sort]1 did they discuss t1/name him t1.
b. That1, they certainly discussed t1/*named him t1.
In the terms introduced in chapter 2, this is because topicalization is a B-extraction and thus incompatible with ACs
like those in (i) and (ii), whereas negative fronting is an A-extraction and thus insensitive to ACs.
10. Final here is of course a technical term. Oversimplifying somewhat, a final arc is one not erased by an arc with the
same tail node.

< previous page page_193 next page >


If you like this book, buy it!
< previous page page_194 next page >
Page 194
11. My proposal regarding that clauses in apparent subject position obviously shares many features with the earlier
transformational proposal of Koster (1978b), including the claims that the that clause is not a surface subject and that
there is an invisible pronominal subject. There are also sharp differences; only in my view is the pronominal an RP and
the that clause a subject in more abstract representations.
12. In Postal 1986a, sec. 3.2, I motivate a constraint on a distinct class of complement-taking verbs, including feel and
hold, that is somewhat the opposite of (32). With these verbs, the extraposition structure is obligatory. As expected
given the assumptions in the text, such verbs preclude that clauses from being passive subjects, object-raising subjects,
or topics, but allow them to be extraposed in passives and to be RNR pivots.
(i) a. He feels that foreigners are spies.
b.*That foreigners are spies was felt by Sidney.
c.*[That foreigners are spies]1 is easy to feel t1.
d.1[That foreigners are spies]1, no one still feels t1.
e. It was felt that foreigners are spies.
f. Frank may feel t1 and probably does feel t1 [that foreigners are spies]1.
13. Such an RP would be expected to free NP topicalization from island constraints, which is, of course, not the case.
As discussed in chapters 1 and 2, though, this can be argued to follow because the RP is also extracted.
14. In Postal 1986a, chap. 6, I argue that the stranded prepositions of English pseudopassives are a function of
advancement and invisible RPs (see also Postal 1991, 1996).
15. Beyond the parochial claim that what appear in English to be extractions of PP heads are in fact not, it could be
suggested that PP head extraction is universally banned.
16. Negatively, 10-arcs must apparently not be subject to any relational uniqueness condition of the sort assumed in
Relational Grammar to govern relations like 1, 2, 3, and so on. Given the assumptions made here, the clause whose
predicate is talk would, for instance, have two final 10-arcs.
(i) It was Marvin who1 that2 was hard to talk about t2 to t1.
One of these would have who as head, the other that.
17. These appear not to include the idiolect of Steedman (1985), who cites as grammatical sentences that violate the
constraint in question and that I find impossible (e.g., his (19b)).
18. It might be assumed that the ungrammaticality of (63c) is a function of the recurrent claim (e.g., Postal 1974, 126;
Grosu 1976; McCawley 1982, 99, 1987, 186, 1988, 528) that RNR gaps must correspond to final constituents on right
branches. But this formulation of various restrictions on RNR may be question-able given such cases as (i) and (ii).
(i) Melvin offered t1 to the Belgians and Jerome offered t1 to the Dutch [several tons of rotted frankfurters]1.

< previous page page_194 next page >


If you like this book, buy it!
< previous page page_195 next page >
Page 195
(ii) He wanted to drive t1 mad and did drive t1 mad [the recipient of his earlier vows of undying love]1.
It is sometimes suggested that, despite appearances, the rightmost condition is met in such cases because of
preceding CXS (see, e.g., Oehrle 1990, 412). But that claim seems to fail for cases like (iii), where Ross's (1967, 139)
observation that PP objects cannot be CXS targets would then block the perfectly grammatical RNR structure (iiib).
(iii) a.*I talked to t1 about love [the tall woman in the black dress]1.
b. Mike may have talked to t1 about love and certainly talked to t1 about marriage [the tall woman in the black
dress]1.
It might be noted that several one-object verbs subject to the Indirect Object Constraint would meet the rightmost
condition but are still incompatible with RNR.
(iv) a. They wrote Kenneth.
b.*Who1 did they write t1?
c.*They might have written t1 and should have written t1 [the author of the article]1.
(v) a. Trees surrounded the barn.
b.*[Which barn]1 did trees surround t1?
c.*Trees might have surrounded t1 and probably did surround t1 [all of those barns]1.
19. This conclusion faces an apparent problem in Ross's (1967, 40) observation that IOs are incompatible with CXS.
(i)*I loaned t1 my binoculars [a man who was watching the race]1.
To subsume this restriction as well as those noted in the text under (65) would appear to require taking CXS to be
an extraction phenomenon, despite its generally assumed "bounded" character. This position is in effect accepted by
Gazdar (1981, 176), who states that CXS is not really "bounded." I agree essentially with Gazdar that CXS differs
from, for example, RNR basically only in the way any simple instance of an L-extraction L differs from an ATB
instance of L.
There are, of course, certain other problems with this view, some cited by Williams (1990)for example, the fact
(Ross 1967, 139) that CXS cannot strand prepositions, whereas RNR, as discussed in section 4.2.2.7, can. But such
facts are not decisive.
20. A similar-sounding claim is made by Williams (1990); however, his remarks are linked exclusively to sentences
whose analysis as RNR constructions is highly controversial, in particular, cases without any standard coordination.
21. This objection has particular force in Levine's case, for he has adopted the overall GPSG framework, in which P-
gaps are described in a particular way involving the distribution of so-called Slash categories (see section 4.3). Under
Levine's (1985) view, RNR constructions could contain no Slash categories, wrongly entailing that P-gaps are not
"licensed" in RNR cases (see section 4.3.4.3).

< previous page page_195 next page >


If you like this book, buy it!
< previous page page_196 next page >
Page 196
22. Since writing the original version of this chapter, I have concluded in Postal 1994 that the view taken in section
4.2.2.13 that cases like (106b) involve P-gaps "licensed" by RNR is incorrect, on the grounds that such examples do not
in fact contain P-gaps. However, since it is not possible to consider the motivations for such a view here, I have left the
present study as is and refer the interested reader to my previously published but subsequently written paper. In defense
of this decision, I believe that it is a highly controversial and nonstandard view to reject a P-gap analysis of cases like
(106b), once these are noticed. In any event, even if that controversial conclusion is correct, I do not think it
substantively alters the present conclusion that RNR is a true extraction, for the P-gap evidence for this conclusion is
only one type among many.
23. I am not claiming that interwoven dependencies undermine the viability of the idea of multiple mother nodes. As in
note 4, the issue is only the way this idea is exploited in McCawley's view of RNR.
24. Another argument linking RNR to L-extractions could be based on reflexivization facts. The complexity and
subtlety of the situation precludes a detailed discussion; but I should mention the following facts. It is well known that
in certain cases L-extractions "expand" the class of possible reflexive form antecedents. This is generally true in
"picture noun" structures and, for certain speakers (including myself) for simple reflexives as well (see Lasnik and
Saito 1992, 110-111).
(i) a.*[My brother]1 thinks that Gail stole a picture of himself1.
b.*[My brother]1 thinks that Gail will vote for himself1.
(ii) a. It was a picture of himself1 which2 [my brother]1 thinks that Gail stole t2.
b. Himself1, [my brother]1 thinks Gail will vote for t1.
I find that a similar pattern emerges in RNR expressions.
(iii) a. [My brother]1 thinks that Gail might have stolen t2 and probably did steal t2 [a picture of himself1 ]2.
b. [My brother]1 thinks that Gail might vote for t1, and probably will vote for t1 himself1.
25. An important revision with relevance to later discussion is proposed by Sells (1986) on the basis of facts about RPs.
See note 32.
26. This is chiefly due to the authors' assumption that the Slash feature has as its value a single category, a view
rejected in Pollard and Sag 1987, 1994.
27. Pesetsky (1982, 555) also offers more general criticisms of Slash approaches that have, I think, minimal force. He
cites a claim by Noam Chomsky that the Slash notation "allows a rather wild proliferation of categorial labels, without
explaining the basic properties of Move a." Pesetsky adds that Slash categories fail to behave like other categories in
that, for example, no verb subcategorizes for NP/NP or any other Slash category.
The relevance of the claim about category proliferation is at best obscure. To be sure, recognition of Slash
categories based on an underlying vocabulary M yields many more categories than are present in M alone. Recognition
of X-bar catego-

< previous page page_196 next page >


If you like this book, buy it!
< previous page page_197 next page >
Page 197
ries over a basic vocabulary N also yields many more categories than are present in N alone; but Chomsky and
Pesetsky have had no qualms about that expansion. Is this more than a double standard? The remark about "basic
properties of Move a" is unclear. Pesetsky could not have meant to beg the question of whether transformations play a
role in natural language grammars. So he must mean that Slash category work has not specified what would show up as
constraints on transformations in transformational work. That assertion is at best only partially true; consider the body
of Slash results on Ross's (1967) CSC, a restriction that, notably, receives no explicit account in standard "Move a"
works (see note 3 of chapter 3).
Pesetsky's subcategorization point has more substance, but I believe that its premise is false. It is obviously so in
the framework of Gazdar et al., (1985) where, for example, object-raising predicates like difficult subcategorize for a
Slash category. However, this is not definitive since, as discussed below, Hukari and Levine (1991) divide the older
Slash into two features, Slash and Gap, and difficult would then subcategorize only for Gap categories.
A more telling argument involves the usage of the verb please found in (ia). As examples (ic,d) illustrate, this usage
is not possible in structures that do not involve L-extraction of the verb's object. Arguably, then, in Slash terms the
illustrated usage of please requires just the sort of subcategorization for a Slash category that Pesetsky claims is
unattested.
(i) a. Louise can do whatever1 she pleases t1.
b. Louise can do anything1 she pleases t1.
c.*Louise pleases lots of stuff/that/everything I please.
d.*Louise doesn't please anything.
28. This conclusion is not a logically necessary entailment of the examples cited and is, I believe, ultimately not true. In
more abstract frameworks, it is possible to view the RNR phenomenon as taking as arguments structures already
incorporating L-extractions. See the brief discussion of RNR targets as L-extraction remnants in section 4.2.1.2.6.
29. Claim (141) is, however, contradicted by the remark in (i).
(i) "Nevertheless we cannot leftward-extract the target of RNR (or, for that matter, any subconstituent of it ..."
(Oehrle 1990, 424)
The evidence given for (i) seems to consist only of (ii) and (iii) without the bracketed material, added here for
clarity.
(ii)*the painter who1 Sal knows a man who hates t1 and Bill knows a woman who admires tl
(iii)*[Whose work]1 do you know a man who likes t1 and plan to meet a clone who hates t1 ?
These are taken to be relevant to the more general claim via, I infer, the following implicit logic. Oehrle states (pp.
411-412) that L-extractions cannot extract pieces of conjoined relative clauses, supporting this statement by citing (iii),
which he stars. However, he also claims that RNR out of the same contexts is fine, sup-porting this claim by citing (iv).

< previous page page_197 next page >


If you like this book, buy it!
< previous page page_198 next page >
Page 198
(iv) I know a man who likes t1 and (hope to meet) a clone who hates t1 [the work of Reynard]1.
Given (iv), if an RNR pivot could be L-extracted, (ii) and (iii) could be formed indirectly without direct L-
extraction into relative clauses.
However, Oehrle's implicit argument is problematic at best. Although (ii) is fairly bad, for me (iii) seems
grammatical, hardly different in quality from (iv). This is consistent with judgments offered by Steedman (1989, 220),
who cites (v) with only a question mark.
(v) ?[Which kind of semantics]1 will there always be some people who like t1 and some people who dislike t1?
Moreover, studies of extraction from indirect questions have pointed out that relativization out of such questions is
considerably better than questioning from them (see Cinque 1990, 18). Notably, (ii) involves relativization from
something extracted via RNR out of relatives. Possibly, then, the unacceptability effect in (ii) is at least partly the effect
noted with extraction from indirect questions, that is, an independent degradation resulting from intermingling repeated
instances of the same construction type. This might explain why (for me) (iii) is better than (ii). If that were the source
of the ill-formedness of (ii), one would predict that other nonrelative extractions from RNR pivots linked to gaps in
relative clauses would yield greatly superior results. This seems true, as I find (via,b) passably well formed.
(vi) a. Edward1, Calvin dated a girl who tried to befriend t1 and plans to date a woman who failed to seduce t1.
b. [No theory]1 did Ernie interview any natives who accepted t1 or contact any foreigners who rejected t1.
In my terms, the reason these L-extractions from restrictive relative clauses are well formed is that such constituents
form selective islands; hence, analogs where the conditions on selective island extraction fail are not good.
Ultimately, the facts discussed in regard to Oehrle's claim do not seem critical for the larger questions at issue, in
the face of the clear well-formedness of, for example, (137), (139), and (140); see also section 4.4.
30. The revised theory can be presumed to contain statements requiring that the filler of the former type of category
appear to its left and the filler of the latter type to its right.
31. In Gazdar et al.'s (1985) framework vast work is done by the Head Feature Convention, with Slash taken to be both
a head and a foot feature. In particular, extraction from subjects is blocked by this principle, which plays a key role in
the description of P-gaps. Noting that other languages allow extractions from subjects, Pollard and Sag (1994, chap. 4)
abandon the Head Feature Convention and make additional changes in previous GPSG approaches to coordination.
Crucially for present purposes, they propose (i) to block extraction from English subjects.
(i) The Subject Condition
A lexical head's SUBCAT list may contain a slashed subject only if it also contains another slashed element.

< previous page page_198 next page >


If you like this book, buy it!
< previous page page_199 next page >
Page 199
(i) is supposed to account for well-known contrasts between P-gap structures like (iib) and simple subject
extractions like (iia).
(ii) That was the rebel leader who1 rivals of t1 shot a.*the British consul/b. t1.
However, (i) will fail for Q-exfiltration structures like the (b) cases of (154)-(156) for roughly the same reasons that
the Head Feature Convention does. So (i) will wrongly block all these well-formed examples. The reason is that
''another slashed element" will not occur in the right place in such Q-exfiltrations.
32. Sells (1986) proposes allowing Slash categories that fail to dominate gaps. Thus, to handle certain serious problems
involving RPs in Hebrew, Irish, and so on, he would allow, for example, a category NP/NP to dominate a non-trace that
is an RP. In this approach, the linkage between extracted form and RP is not part of the syntax. It is unclear whether
such ideas could have any application to the problems raised by Q-exfiltrations. Space considerations preclude real
discussion; but I would suggest that the problems that motivate Sells's suggestion might profitably be approached via a
proposal similar to that made in section 4.2.1.3.2, that is, one in which apparent monolithic extractions are decomposed
into advancements or demotions that leave RPs plus extractions.
33. An additional piece of evidence for the extraction character of RNR and an indication of the severity of the
problems raised by Q-exfiltration cases for Slash approaches would appear if the rightmost extractee in multiple RNR
cases like (160) and (161) could "license" P-gaps.
(i) Frank didn't admit to t1 that he could t2 nor deny to t1 that he should t2 [hire, train, and deal with t1 as an
equal]2 nor did Glen reach any agreement with t1 [that angry middle-aged person]1.
Despite its complexity and the perceptual difficulties inherent in repeated RNR, (i) seems grammaticalin fact, even
better than (ii), where the P-gaps are licensed by L-extractions from RNR pivots.
(ii) [Which angry person]1 did Frank admit to t1 that he could t2 but deny to t1 that he should t2 [hire t1, train t1, or
deal with t1 as an equal]2, and Glen fail to reach any agreement with t1 ?
The problem with this argument is the fact cited in note 22 that, as I argue in Postal 1994, no evidence supports the
claim that cases like (i) contain P-gaps.
34. Because it predicts that extraction from subjects is universally banned, thus seemingly precluding the existence of
(e.g. Swedish and in fact English (see (44a) of chapter 1).
35. A definitional account of Gazdar et al.'s (1985) foot feature WH might also be feasible.
36. Of course, the challenge also holds for other frameworks, which (as previously noted) in general have nothing to
say about the issue.
Appendix A
1. Chomsky (1977b, 127) also proposes "reanalysis" to account for extractions like that in (i).

< previous page page_199 next page >


If you like this book, buy it!
< previous page page_200 next page >
Page 200
(i) What1 did he [have an opportunity] to do t1?
Although (i) of course satisfies the conditions on selective island extraction, a claim that the infinitival complement
in such examples forms a selective island seems untenable. Although some extractions that violate conditions on
selective extraction (e.g., (iia,b)) are indeed ungrammatical, others (e.g., (iiia-c)) seem acceptable.
(ii) a.*Slim1 though he had an opportunity to become t1,...
b.*the color which1 she had an opportunity to tint her hair t1
(iii) a. [How long]1 did they have an opportunity to work there t1?
b. Lucille met more senators than (what1) I ever had an opportunity to meet t1.
c. Marylou dated [whatever officers]1 she had an opportunity to date t1.
Moreover, nonextraction facts point to the existence of a structure in which the complement is not part of an island.
(iv) a. Nobody had an opportunity to eat anything/a bite.
b. I need to know who had an opportunity to work how long.
Thus, appeal to selective islands appears wrong for (i), which leaves open the question of what the right analysis is.
2. Actually, (31 a) may be irrelevant; it could well be blocked by the same constraint that blocks (i).
(i)*That is a paper that1 we really need to find someone who believes (that) he understands t1.
This seems to satisfy the conditions on selective island extraction as well as those proposed by Chomsky.
3. The critique of Chomsky's (1986a) approach to facts like (22b) and (29) has more general significance. A very large
GB literature now exists (see, e.g., Cinque 1990; Rizzi 1990; Frampton 1991; Lasnik and Saito 1992; Manzini 1992)
that adopts essential features of this approach, though often with significant modifications, additions, or deletions. It
would be a useful exercise to consider whether these variant GB systems fail for reasons parallel to those undermining
Chomsky's (1986a) account.
4. Recognition that the cases discussed by C&M (1983) and Chomsky (1986a) involve selective island extraction does
not as such offer a solution to the problems they raise. In particular, nothing I have said accounts for the difference
between, for example, (22b) and (29). However, the relevance of selective islands suggests that such differences relate
to the conditions that "license" the kind of control characterizing selective island extraction.

< previous page page_200 next page >


If you like this book, buy it!
< previous page page_201 next page >
Page 201

References
Abbott, Barbara. 1976. Right node raising as a test for constituenthood. Linguistic Inquiry 7, 639-642.
Aoun, Joseph, Norbert Hornstein, and Dominique Sportiche. 1980. Some aspects of wide scope quantification. Journal
of Linguistic Research 1, 69-95.
Authier, J.-Marc. 1992. Iterated CPs and embedded topicalization. Linguistic Inquiry 23, 329-336.
Baker, C. L. 1970. Double negatives. Linguistic Inquiry 1, 169-186.
Baitin, Mark, and Paul M. Postal. 1996. More on reanalysis hypotheses. Linguistic Inquiry 27, 127-145.
Borer, Hagit. 1984. Restrictive relatives in Modern Hebrew. Natural Language & Linguistic Theory 2, 219-260.
Boskovic *, Zeljko*. 1995. On right-node base generation. Ms., University of Connecticut, Storrs.
Brame, Michael K. 1978. Base generated syntax. Seattle, Wash.: Noit Amrofer.
Bresnan, Joan W. 1974. The position of certain clause-particles in phrase structure. Linguistic Inquiry 5, 614-619.
Bresnan, Joan W. 1975. Comparative deletion and constraints on transformations. Linguistic Analysis 1, 25-74.
Bresnan, Joan W. 1977. Variables in the theory of transformations. In Peter Culicover, Thomas Wasow, and Adrian
Akmajian, eds., Formal syntax. New York: Academic Press.
Bresnan, Joan W., ed. 1982. The mental representation of grammatical relations. Cambridge, Mass.: MIT Press.
Brody, Michael. 1996. Lexico-Logical Form. Cambridge, Mass.: MIT Press.
Browning, M. A. 1987a. Null operator constructions. Doctoral dissertation, MIT, Cambridge, Mass.
Browning, M. A. 1987b. Null operators and their antecedents. In Joyce McDonough and Bernadette Plunkett, eds.,
Proceedings of NELS 17. GLSA, University of Massachusetts, Amherst.

< previous page page_201 next page >


If you like this book, buy it!
< previous page page_202 next page >
Page 202
Carnap, Rudolph. 1958. Introduction to symbolic logic and applications. New York: Dover.
Chomsky, Noam. 1972. Language and mind. New York: Harcourt, Brace & World.
Chomsky, Noam. 1975. Reflections on language. New York: Pantheon.
Chomsky, Noam. 1977a. Essays on form and interpretation. New York: North-Holland.
Chomsky, Noam. 1977b. On wh-movement. In Peter Culicover, Thomas Wasow, and Adrian Akmajian, eds., Formal
syntax. New York: Academic Press.
Chomsky, Noam. 1980. Rules and representations. New York: Columbia University Press.
Chomsky, Noam. 1981. Lectures on government and binding. Dordrecht: Foris.
Chomsky, Noam. 1982. Some concepts and consequences of the theory of government and binding. Cambridge, Mass.:
MIT Press.
Chomsky, Noam. 1986a. Barriers. Cambridge, Mass.: MIT Press.
Chomsky, Noam. 1986b. Knowledge of language. New York: Praeger.
Chomsky, Noam. 1988. Language and problems of knowledge. Cambridge, Mass.: MIT Press.
Chomsky, Noam. 1995. The Minimalist Program. Cambridge, Mass.: MIT Press.
Chomsky, Noam, and Howard Lasnik. 1977. Filters and control. Linguistic Inquiry 8, 425-504.
Chung, Sandra, and James McCloskey. 1983. On the interpretation of certain island facts in GPSG. Linguistic Inquiry
14, 704-713.
Cinque, Guglielmo. 1975. The Shadow Pronoun Hypothesis and "chopping" rules in Romance. Linguistic Inquiry 6,
140-145.
Cinque, Guglielmo. 1977. The movement nature of left dislocation. Linguistic Inquiry 8, 397-411.
Cinque, Guglielmo. 1990. Types of -dependencies. Cambridge, Mass.: MIT Press.
Culicover, Peter W. 1982. Syntax. New York: Academic Press.
Culicover, Peter W. 1990. Strange extractions. MS., Center for Cognitive Science, The Ohio State University,
Columbus.
Culicover, Peter W. 1991. Polarity, inversion, and focus in English. In Germn F. Westphal, Benjamin Ao, and Hee-
Rahk Chae, eds., ESCOL '91. Department of Linguistics, The Ohio State University, Columbus.
Culicover, Peter W. 1993a. The adverb effect: Evidence against ECP accounts of the that-t effect. In Amy J. Schafer,
ed., NELS 23. GLSA, University of Massachusetts, Amherst.
Culicover, Peter W. 1993b. Evidence against ECP accounts of the that-t effect. Linguistic Inquiry 24, 557-561.

< previous page page_202 next page >


If you like this book, buy it!
< previous page page_203 next page >
Page 203
Culicover, Peter W., and Ken Wexler. 1977. Some syntactic implications of a theory of language learnability. In Peter
W. Culicover, Thomas Wasow, and Adrian Akmajian, eds., Formal syntax. New York: Academic Press.
Deane, Paul D. 1988. Which NPs are there unusual possibilities for extraction from? In Lynn MacLoed, Gary Larson,
and Diane Brentari, eds., CLS 24. Chicago Linguistic Society, University of Chicago, Chicago, Ill.
Deane, Paul D. 1991. Limits to attention: A cognitive theory of island phenomena. Cognitive Linguistics 2, 1-63.
Demirdache, Hamida Khadiga. 1991. Resumptive chains in restrictive relatives, appositives and dislocation structures.
Doctoral dissertation, MIT, Cambridge, Mass.
Dowty, David, and Pauline Jacobson. 1988. Agreement as a semantic phenomenon. In Joyce Powers and Kenneth de
Jong, eds., ESCOL '88. Department of Linguistics, The Ohio State University, Columbus.
Emonds, Joseph. 1979. Appositive relatives have no properties. Linguistic Inquiry 10, 211-243.
Emonds, Joseph. 1985. A unified theory of syntactic categories. Dordrecht: Foris.
Engdahl, Elisabet. 1983. Parasitic gaps. Linguistics and Philosophy 6, 5-34.
Engdahl, Elisabet. 1984. Why some empty subjects don't license parasitic gaps. In Mark Cobler, Susannah MacKaye,
and Michael T. Westcoat, eds., Proceedings of the West Coast Conference on Formal Linguistics, vol. 3. The Stanford
Linguistics Association, Stanford University, Stanford, Calif.
Engdahl, Elisabet. 1985. Parasitic gaps, resumptive pronouns, and subject extractions. Linguistics 23, 3-44.
Fiengo, Robert, C.-T. James Huang, Howard Lasnik, and Tanya Reinhart. 1988. The syntax of Wh-in-situ. In Hagit
Borer, ed., Proceedings of the Seventh West Coast Conference on Formal Linguistics. Stanford, Calif.: CSLI
Publications. [Distributed by Cambridge University Press.]
Fiengo, Robert, and Robert May. 1994. Indices and identity. Cambridge, Mass.: MIT Press.
Fillmore, Charles. 1965. Indirect object constructions in English and the ordering of transformations The Hague:
Mouton.
Frampton, John. 1991. Relativized Minimality: A review. The Linguistic Review, 8 1-46.
Gazdar, Gerald. 1981. Unbounded dependencies and coordinate structure. Linguistic Inquiry 12, 155-184.
Gazdar, Gerald. 1982. Phrase structure grammar. In Pauline Jacobson and Geoffrey K. Pullum, eds., The nature of
syntactic representation. Dordrecht: D. Reidel.
Gazdar, Gerald, Ewan Klein, Geoffrey K. Pullum, and Ivan Sag. 1984. Foot features and parasitic gaps. In Wim de
Geest and Yvan Putseys, eds., Sentential complementation. Dordrecht: Foris.

< previous page page_203 next page >


If you like this book, buy it!
< previous page page_204 next page >
Page 204
Gazdar, Gerald, Ewan Klein, Geoffrey K. Pullum, and Ivan Sag. 1985. Generalized Phrase Structure Grammar.
Oxford: Blackwell.
Gazdar, Gerald, Geoffrey K. Pullum, Ewan Klein, Robert Carpenter, Thomas E. Hukari, and Robert D. Levine. 1988.
Category structures. Computational Linguistics 14, 1-19.
Georgopoulos, Carol. 1983. Trace and resumptive pronouns in Palauan. In Amy Chukerman, Mitchell Marks, and John
F. Richardson, eds., CLS 19. Chicago Linguistic Society, University of Chicago, Chicago, Ill.
Georgopoulos, Carol. 1984a. A case for Subjacency without movement. In Mark Cobler, Susannah MacKaye, and
Michael T. Westcoat, eds., Proceedings of the West Coast Conference on Formal Linguistics, vol. 3. Stanford
Linguistics Association, Stanford University, Stanford, Calif.
Georgopoulos, Carol. 1984b. On Belauan islands: A study in agreement morphology. In Claudia Brugman and Monica
Macaulay, eds, Proceedings of the Tenth Annual Meeting of the Berkeley Linguistics Society. Berkeley Linguistics
Society, University of California, Berkeley.
Georgopoulos, Carol. 1985a. The syntax of variable binding in Palauan. Doctoral dissertation, University of California,
San Diego.
Georgopoulos, Carol. 1985b. Variables in Palauan syntax. Natural Language & Linguistic Theory 3, 59-94.
Goldsmith, John. 1985. A principled exception to the Coordinate Structure Constraint. In William H. Eilfort, Paul D.
Kroeber, and Karen L. Peterson, eds., CLS 21. Part 1, Papers from the General Session. Chicago Linguistic Society,
University of Chicago, Chicago, Ill.
Goodall, Grant. 1987. Parallel structures in syntax. Cambridge: Cambridge University Press.
Grimshaw, Jane. 1982. Subcategorization and grammatical relations. In Annie Zaenen, ed., Subjects and other subjects:
Proceedings of the Harvard Conference on the Representation of Grammatical Relations. Indiana University
Linguistics Club, Bloomington.
Grosu, Alexander. 1972. The strategic content of island constraints. Doctoral dissertation, The Ohio State University,
Columbus.
Grosu, Alexander. 1973. On the nonunitary nature of the Coordinate Structure Constraint. Linguistic Inquiry 4, 88-92.
Grosu, Alexander. 1976. A note on subject raising to object and right node raising. Linguistic Inquiry 7, 642-645.
Grosu, Alexander, and Sandra A. Thompson. 1977. Constraints on the distribution of NP clauses. Language 53, 104-
151.
Haegeman, Liliane. 1991. Introduction to Government and Binding Theory. Oxford: Blackwell.
Hankamer, Jorge. 1973. Unacceptable ambiguity. Linguistic Inquiry 4, 17-68.

< previous page page_204 next page >


If you like this book, buy it!
< previous page page_205 next page >
Page 205
Heim, Irene. 1988. Where does the Definiteness Restriction apply? Evidence from the definiteness of variables. In Eric
J. Reuland and Alice G. B. ter Meulen, eds., The representation of (in)definiteness. Cambridge, Mass.: MIT Press.
Heycock, Caroline, and Anthony Kroch. 1996. Pseudocleft connectivity: Implications for the LF interface level.
Edinburgh Occasional Papers in Linguistics 96-1. Department of Linguistics, University of Edinburgh.
Higginbotham, James. 1980. Pronouns and bound variables. Linguistic Inquiry 11, 679-708.
Higginbotham, James. 1983. Logical Form, binding, and nominals. Linguistic Inquiry 14, 395-420.
Higgins, F. R. 1973. On J. E. Emonds' analysis of extraposition. In John Kimball, ed., Syntax and semantics, vol. 2.
New York: Academic Press.
Hirschbhler, Paul. 1974. La dislocation gauche comme construction basique en franais. Actes du colloque franco-
allemand de grammaire transformationelle 1, 9-17.
Hirschbhler, Paul. 1975. On the source of lefthand NPs in French. Linguistic Inquiry 6, 155-165.
Hooper, Joan, and Sandra Thompson. 1973. On the applicability of root transformations. Linguistic Inquiry 4, 465-497.
Hornstein, Norbert, and David Lightfoot. 1991. On the nature of lexical government. In Robert Freidin, ed., Principles
and parameters in comparative grammar. Cambridge, Mass.: MIT Press.
Huang, C.-T. James. 1982. Logical relations in Chinese and the theory of gram-mar. Doctoral dissertation, MIT,
Cambridge, Mass.
Hudson, R. A. 1976. Conjunction reduction, gapping, and right-node raising. Language 52, 535-562
Hukari, Thomas E., and Robert D. Levine. 1987. Parasitic gaps, slash termination and the C-Command Condition.
Natural Language & Linguistic Theory 5, 197-222.
Hukari, Thomas E., and Robert D. Levine. 1989. Category antirecursion: Paradoxical consequences of gap-within-filler
constructions. In E. Jane Fee and Katherine Hunt, eds., Proceedings of the Eighth West Coast Conference on Formal
Linguistics. Stanford, Calif.: CSLI Publications. [Distributed by Cambridge University Press.]
Hukari, Thomas E., and Robert D. Levine. 1991. On the disunity of unbounded dependency constructions. Natural
Language & Linguistic Theory 9, 97-144.
Jackendoff, Ray. 1977.X *syntax A study of phrase structure. Cambridge, Mass.: MIT Press.
Jacobson, Pauline. 1992. The lexical entailment theory of control and the tough-construction. In Ivan Sag and Anna
Szabolcsi, eds., Lexical matters. Stanford, Calif.: CSLI Publications.

< previous page page_205 next page >


If you like this book, buy it!
< previous page page_206 next page >
Page 206
Johnson, David E., and Paul M. Postal. 1980. Arc Pair Grammar. Princeton, N.J.: Princeton University Press.
Kaplan, Ronald M., and Joan Bresnan. 1982. Lexical-Functional Grammar: A formal system for grammatical
representation. In Joan Bresnan, ed., The mental representation of grammatical relations. Cambridge, Mass.: MIT
Press.
Kaplan, Ronald M., and Annie Zaenen. 1989. Long-distance dependencies, constituent structure, and functional
uncertainty. In Mark R. Baltin and Anthony S. Kroch, eds., Alternative conceptions of phrase structure. Chicago:
University of Chicago Press.
Kayne, Richard S. 1984. Connectedness and binary branching. Dordrecht: Foils.
Kayne, Richard S. 1985. Principles of particle constructions. In Jacqueline Guron, Hans-Georg Obenauer, and Jean-
Yves Pollock, eds., Grammatical representation. Dordrecht: Foils.
Kayne, Richard S. 1994. The antisymmetry of syntax. Cambridge, Mass.: MIT Press.
Kiss, Katalin . 1985. Parasitic chains. The Linguistic Review 5, 41-74.
Koster, Jan. 1978a. Locality principles in syntax. Dordrecht: Foils.
Koster, Jan. 1978b. Why subject sentences don't exist. In S. Jay Keyser, ed., Recent transformational studies in
European languages. Cambridge, Mass.: MIT Press.
Koster, Jan. 1987. Domains and dynasties. Dordrecht: Foris.
Kroch, Anthony S. 1989. Asymmetries in long-distance extraction in a tree-adjoining grammar. In Mark R. Baltin and
Anthony S. Kroch, eds, Alternative conceptions of phrase structure. Chicago: University of Chicago Press.
Kuno, Susumu. 1973. Constraints on internal clauses and sentential subjects. Linguistic Inquiry 4, 363-385.
Kuno, Susumu, and Ken-Ichi Takami. 1993. Grammar and discourse principles. Chicago: University of Chicago Press.
Kuroda, S.-Y. 1968. Review of Charles Fillmore, Indirect object constructions in English and the ordering of
transformations. Language 44, 374-378.
Lakoff, George. 1986. Frame semantic control of the Coordinate Structure Constraint. In Anne M. Farley, Peter T.
Farley, and Karl-Erik McCullough, eds., Papers from the Parasession on Pragmatics and Grammatical Theory.
Chicago Linguistic Society, University of Chicago, Chicago, Ill.
Langendoen, D. Terence, Nancy Kalish-Landon, and John Dore. 1974. Dative questions: A study of the relation of
acceptability to grammaticality of an English sentence type. Cognition 2, 451-478.
Langendoen, D. Terence, and Paul M. Postal. 1984. The vastness of natural language. Oxford: Blackwell.
Langendoen, D. Terence, and Geoffrey K. Pullum. 1977. Preposition stranding in English: A problem and a mystery. In
Samuel E. Fox, Woodford A. Beach, and Shulamith Philosoph, eds., CLS book of squibs. Chicago Linguistic Society,
University of Chicago, Chicago, Ill.

< previous page page_206 next page >


If you like this book, buy it!
< previous page page_207 next page >
Page 207
Lasnik, Howard, and Mamoru Saito. 1992. Move a. Cambridge, Mass.: MIT Press.
Lasnik, Howard, and Juan Uriagereka. 1988. A course in GB syntax. Cambridge, Mass.: MIT Press.
Lawler, John. 1974. Ample negatives. In Michael W. La Galy, Robert A. Fox, and Anthony Bruck, eds., Papers from
the Tenth Regional Meeting, Chicago Linguistic Society. Chicago Linguistic Society, University of Chicago, Chicago,
Ill.
Levine, Robert D. 1984. Against reanalysis rules. Linguistic Analysis 14, 3-29.
Levine, Robert D. 1985. Right node (non-) raising. Linguistic Inquiry 16, 492-497.
Maling, Joan, and Annie Zaenen. 1982. A phrase structure account of Scandinavian extraction phenomena. In Pauline
Jacobson and Geoffrey K. Pullum, eds., The nature of syntactic representation. Dordrecht: D. Reidel.
Manzini, Maria Rita. 1992. Locality A theory and some of its empirical consequences. Cambridge, Mass. MIT Press.
McCawley, James D. 1982. Parentheticals and discontinuous constituent structure. Linguistic Inquiry 13, 91-106.
McCawley, James D. 1987. Some additional evidence for discontinuity. In Geoffrey J. Huck and Almerindo E. Ojeda,
eds., Syntax and semantics 20: Discontinuous constituency. Orlando, Fla.: Academic Press.
McCawley, James D. 1988. The syntactic phenomena of English, volumes I and 2. Chicago: University of Chicago
Press.
McCloskey, James. 1979. Transformational syntax and model theoretic semantics. Dordrecht: D. Reidel.
McCloskey, James. 1986. Right node raising and preposition stranding. Linguistic Inquiry 17, 183-186.
McCloskey, James. 1988. Syntactic theory. In Frederick J. Newmeyer, ed., Linguistics: The Cambridge survey, vol. I.
Cambridge: Cambridge University Press.
Mller, Gereon. 1995. A-bar syntax. Berlin: Mouton de Gruyter.
Mller, Gereon. 1996. A constraint on remnant movement. Natural Language & Linguistic Theory 14, 355-407.
Munn, Alan. 1993. Topics in the syntax and semantics of coordinate structures. Doctoral dissertation, University of
Maryland, College Park.
Na, Younghee, and Geoffrey J. Huck. 1992. On extracting from asymmetrical structures. In Diane Brentari, Gary N.
Larson, and Lynn A. MacLeod, eds., The joy of grammar. Amsterdam: John Benjamins.
Nakajima, Heizo. 1989. Bounding of rightward movements. Linguistic Inquiry 20, 328-334.
Napoli, Donna Jo. 1993. Syntax. Oxford: Oxford University Press.
Obenauer, Hans-Georg. 1984. On the identification of empty categories. The Linguistic Review 4, 153-202.

< previous page page_207 next page >


If you like this book, buy it!
< previous page page_208 next page >
Page 208
Obenauer, Hans-Georg. 1985. Connectedness, variables, and stylistic inversion in French. In Jacqueline Guron, Hans-
Georg Obenauer, and Jean-Yves Pollock, eds., Grammatical representation. Dordrecht: Foils.
Obenauer, Hans-Georg. 1986. Dplacer a et -liage local: Drivations vs. reprsentations. In Mitsou Ronat and Daniel
Couquaux, eds., La grammaire modulaire. Paris: Les ditions de Minuit.
Obenauer, Hans-Georg. 1992. L'interprtation des structures wh et l'accord du participe pass. In Hans-Georg Obenauer
and Anne Zribi-Hertz, eds., Structure de la phrase et thorie du liage. Paris: Presses Universitaires de Vincennes.
Oehrle, Richard T. 1975. The grammatical status of the English dative alternation. Doctoral dissertation, MIT,
Cambridge, Mass.
Oehrle, Richard T. 1990. Categorial frameworks, coordination, and extraction. In Aaron L. Halpern, ed., Proceedings of
the Ninth West Coast Conference on Formal Linguistics. Stanford, Calif.: CSLI Publications. [Distributed by
Cambridge University Press.]
Ojeda, Almerindo E. 1987. Discontinuity, multidominance, and unbounded dependency in Generalized Phrase Structure
Grammar: Some preliminaries. In Geoffrey J. Huck and Almerindo E. Ojeda, eds., Syntax and semantics 20:
Discontinuous constituency. Orlando, Fla.: Academic Press.
Perimutter, David M. 1972. Evidence for shadow pronouns in French relativization. In Paul M. Peranteau, Judith N.
Levi, and Gloria C. Phares, eds., The Chicago which hunt. Chicago Linguistic Society, University of Chicago, Chicago,
Ill.
Pesetsky, David M. 1982. Paths and categories. Doctoral dissertation, MIT, Cambridge, Mass.
Pesetsky, David M. 1987. Wh-in-situ, movement, and unselective binding. In Eric J. Reuland and Alice G. B. ter
Meulen, eds., The representation of (in)definiteness. Cambridge, Mass.: MIT Press.
Piera, Carlos. 1985. Gaps in gaps in GPSG. Linguistic Inquiry 16, 681-683.
Pollard, Carl, and Ivan A. Sag. 1987. Information-based syntax and semantics. Vol. 1, Fundamentals. Stanford, Calif.:
CSLI Publications.
Pollard, Carl, and Ivan A. Sag. 1994. Head-Driven Phrase Structure Grammar. Stanford, Calif.: CSLI Publications.
[Distributed by University of Chicago Press.]
Postal, Paul M. 1970. On coreferential complement subject deletion. Linguistic Inquiry 1,439-500.
Postal, Paul M. 1971. Crossover phenomena. New York: Holt, Rinehart and Winston.
Postal, Paul M. 1974. On raising. Cambridge, Mass.: MIT Press.
Postal, Paul M. 1979. Some syntactic rules in Mohawk. New York: Garland.
Postal, Paul M. 1982. Some Arc Pair Grammar descriptions. In Pauline Jacobson and Geoffrey K. Pullum, eds., The
nature of syntactic representation. Dordrecht: D. Reidel.

< previous page page_208 next page >


If you like this book, buy it!
< previous page page_209 next page >
Page 209
Postal, Paul M. 1985. La dgradation de prdicat et un genre nglig de monte. Recherches Linguistiques 13, 33-68.
Postal, Paul M. 1986a. Studies of passive clauses. Albany, N.Y.: State University of New York Press.
Postal, Paul M. 1986b. Why Irish raising is not anomalous. Natural Language & Linguistic Theory 4, 333-356.
Postal, Paul M. 1988. Advances in linguistic rhetoric. Natural Language & Linguistic Theory 6, 129-137.
Postal, Paul M. 1989. Masked inversion in French. Chicago: University of Chicago Press.
Postal, Paul M. 1990a. French indirect object demotion. In Paul M. Postal and Brian D. Joseph, eds., Studies in
Relational Grammar 3. Chicago: University of Chicago Press.
Postal, Paul M. 1990b. Some unexpected English restrictions. In Katarzyna Dziwirek, Patrick Farrell, and Errapel
Mejias-Bikandi, eds., Grammatical relations: A cross-theoretical perspective. Stanford, Calif.: CSLI Publications.
[Distributed by Cambridge University Press.]
Postal, Paul M. 1990c. Weird extractions and parasitic gaps. Ms., Thomas J. Watson Research Center, IBM, Yorktown
Heights, N.Y.
Postal, Paul M. 1991. An apparent French extraction anomaly. Ms., Thomas J. Watson Research Center, IBM,
Yorktown Heights, N.Y.
Postal, Paul M. 1992. Phantom successors and the French faire par construction. In Diane Brentari, Gary N. Larson,
and Lynn A. MacLeod, eds., The joy of grammar. Amsterdam: John Benjamins.
Postal, Paul M. 1993a. Parasitic gaps and the across-the-board phenomenon. Linguistic Inquiry 24, 735-754.
Postal, Paul M. 1993b. Remarks on weak crossover effects. Linguistic Inquiry 24, 539-556.
Postal, Paul M. 1993c. Some defective paradigms. Linguistic Inquiry 24, 347-364.
Postal, Paul M. 1994. Parasitic and pseudoparasitic gaps. Linguistic Inquiry 25, 63-117.
Postal, Paul M. 1996. A glance at French pseudopassives. In Clifford S. Burgess, Katarzyna Dziwirek, and Donna B.
Gerdts, eds., Grammatical relations: Theoretical approaches to empirical questions. Stanford, Calif.: CSLI
Publications. [Distributed by Cambridge University Press.]
Postal, Paul M. In Preparation a. Grammatical extraction from selective islands.
Postal, Paul M. In preparation b. Islands. In Mark Baltin and Chris Collins, eds., The handbook of syntactic theory.
Oxford: Blackwell.
Postal, Paul M., and Mark Baltin. 1994. Extraction from selective islands. NSF proposal.

< previous page page_209 next page >


If you like this book, buy it!
< previous page page_210 next page >
Page 210
Postal, Paul M., and Geoffrey K. Pullum. 1988. Expletive noun phrases and movement to subcategorized positions.
Linguistic Inquiry 19, 635-670.
Pullum, Geoffrey K. 1983. How many possible human languages are there? Linguistic Inquiry 14, 447-467.
Pullum, Geoffrey K. 1987. Implications of English extraposed irrealis clauses. In Ann Miller and Joyce Powers, eds.,
ESCOL '87. Department of Linguistics, The Ohio State University, Columbus.
Pullum, Geoffrey K., and Gerald Gazdar. 1982. Natural languages and context-free languages. Linguistics and
Philosophy 4, 471-504.
Quine, Willard van Orman. 1958. Mathematical logic. Cambridge, Mass.: Harvard University Press.
Riemsdijk, Henk van, and Edwin Williams. 1986. Introduction to the theory of grammar. Cambridge, Mass.: MIT
Press.
Rizzi, Luigi. 1990. Relativized Minimality. Cambridge, Mass.: MIT Press.
Rodman, Robert. 1972. On the stranding of prepositions in English. In George Bedell, ed., Explorations in syntactic
theory. UCLA Papers in Syntax 2. Department of Linguistics, University of California, Los Angeles.
Ross, John Robert. 1966. Relativization in extraposed clauses (A problem which evidence is presented that help is
needed to solve). In Mathematical linguistics and automatic translation, Report Number NSF-17 to The National
Science Foundation, Harvard University Computation Laboratory, Cambridge, Mass.
Ross, John Robert. 1967. Constraints on variables in syntax. Doctoral dissertation, MIT, Cambridge, Mass. [Published
as Infinite syntax. Norwood N.J.: Ablex (1986). All page references are to this version.]
Ross, John Robert. 1971. Highest island phenomena. Talk handout, March 1971.
Ross, John Robert. 1984. Inner islands. In Claudia Brugman and Monica Macaulay, eds., Proceedings of the Tenth
Annual Meeting of the Berkeley Linguistics Society. Berkeley Linguistics Society, University of California, Berkeley.
Sag, Ivan. 1976. Deletion and Logical Form. Doctoral dissertation, MIT, Cam-bridge, Mass.
Sag, Ivan. 1983. On parasitic gaps. Linguistics and Philosophy 6, 35-45.
Sag, Ivan, Gerald Gazdar, Thomas Wasow, and Steven Weisler. 1985. Coordination and how to distinguish categories.
Natural Language & Linguistic Theory 3, 117-17l.
Sag, Ivan, and Ewan Klein. 1982. The syntax and semantics of English expletive pronoun constructions. In Michael
Barlow, Daniel P. Flickinger, and Ivan A. Sag, eds., Developments in Generalized Phrase Structure Grammar. Stanford
Working Papers in Grammatical Theory 2. Indiana University Linguistics Club, Bloomington.
Schmerling, Susan F. 1970. A note on negative polarity. Ms., University of Illinois, Urbana-Champaign.

< previous page page_210 next page >


If you like this book, buy it!
< previous page page_211 next page >
Page 211
Schmerling, Susan F. 1975. Asymmetric conjunction and rules of conversation. In Peter Cole and Jerry L. Morgan,
eds., Syntax and semantics 3. New York: Academic Press.
Sells, Peter. 1984. Syntax and semantics of resumptive pronouns. Doctoral dissertation, University of Massachusetts,
Amherst.
Sells, Peter. 1986. Resumptive pronouns in Generalized Phrase Structure Gram-mar. In David Lebeaux and Armin
Mester, eds., University of Massachusetts occasional papers in linguistics 10. GLSA, University of Massachusetts,
Amherst.
Steedman, Mark. 1985. Dependency and coordination in the grammar of Dutch and English. Language 61,523-568.
Steedman, Mark. 1988. Combinators and grammars. In Richard T. Oehrle, Emmon Bach, and Deirdre Wheeler, eds.,
Categorial grammars and natural language structures. Dordrecht: D. Reidel.
Steedman, Mark. 1989. Constituency and coordination in a combinatory gram-mar. In Mark R. Baltin and Anthony S.
Kroch, eds., Alternative conceptions of phrase structure. Chicago: University of Chicago Press.
Steedman, Mark. 1996. Surface structure and interpretation. Cambridge, Mass.: MIT Press.
Stowell, Tim. 1991. The alignment of arguments in adjective phrases. In Susan D. Rothstein, ed., Syntax and semantics
25: Perspectives on phrase structure: Heads and licensing. San Diego, Calif.: Academic Press.
Vergnaud, Jean-Roger. 1974. French relative clauses. Doctoral dissertation, MIT, Cambridge, Mass.
Wexler, Kenneth, and Peter W. Culicover. 1980. Formal principles of language acquisition. Cambridge, Mass.: MIT
Press.
Williams, Edwin. 1990. The ATB theory of parasitic caps. The Linguistic Review 6, 265-279.

< previous page page_211 next page >


If you like this book, buy it!
< previous page page_i next page >
Page i

Three Investigations of Extraction

< previous page page_i next page >


If you like this book, buy it!
< previous page page_213 next page >
Page 213

Index
A
Al-extraction, 6-8, 45, 47-49, 68
A2-extraction, 5-9, 45-49, 68, 77, 165, 166, 169, 170, 181n
Abbot, B., 137, 193n
Across-the-board (ATB) extraction, 3, 52, 77, 82, 83, 97, 122, 134-137, 157, 160, 173, 178, 185n, 190n, 195n
Additive coordination, 136, 137
A-extractions, 5, 9, 11, 13, 14, 25-37, 45, 46, 182n, 183n, 188n
Antipronominal context (AC), 4, 10, 19, 32-36, 38-44, 47, 61-63, 65, 67, 68, 76, 80, 115, 166, 168, 170-172, 183n,
188n, 193n
Aoun, J., 72
Authier, J.-M., 70, 71
B
Baker, C. L., 72
Baltin, M., 9, 60, 125, 166, 181n, 182n
Belauan, 184n
B-extractions, 4-8, 10, 11, 14-16, 25-40, 42, 45, 47-49, 68, 112, 115, 188n, 193n
Boskovic *, Z*., 177, 178
Borer, H., 11, 13
Brame, M., 2
Bresnan, J. W., 3, 73, 106, 110, 114
Brody, M., 184n
Browning, M. A., 133
C
Carnap, R., 143
Categorial Grammar, 2
Chomsky, N., 1-3, 23, 36, 37, 73, 123, 125, 133, 165-168, 171, 172, 184n, 191n, 196n, 197n, 199n, 200n
Chopping rule, 36, 38, 59, 183n
Chung, S., 23, 140, 169, 170-172, 200n
Cinque, G., 14, 17, 19, 21, 37, 38, 45, 184n, 186n, 188n, 198n, 200n
Cognitive Grammar, 185n
Complex NP Constraint (CNPC), 165, 167, 171, 175, 184n
Complex NP Shift (CXS), 105, 106, 119, 120, 131, 178, 180, 195n
Condition on Asymmetric Conjunction, 191n
Conjunct Condition, 51, 83-86, 93, 122, 180, 191n
Control/Controlled, 6, 11, 12, 15-17, 23, 38, 65, 75, 189n
Coordinate Structure Constraint (CSC), 22, 23, 51-53, 55-60, 65, 74, 76, 77, 79, 80, 83, 84, 87, 90-95, 121, 122, 146,
157, 165, 166, 168, 180, 182-185n, 187n, 189-191n
Coordination Principle, 160
Copy rule, 36, 59
Crossing dependency, 3
Culicover, P. W., 60, 66, 102, 104, 121, 122, 137, 138, 145, 146, 151, 166, 175, 177, 189n, 192n
D
Deane, P. D., 23, 54, 55, 57, 59, 66, 168
Demirdache, H. K., 181
Dore, J., 122
Dowty, D., 109
Dutch, 178
E
Emonds, J., 31, 112, 183
Empty Category Principle (ECP), 16, 171, 176, 177
Engdahl, E., 133
Exfiltrate, 140, 143, 151, 154, 157, 158
Extraction Control A (EXCA), 15
Extraction Control B (EXCB), 15, 65
F
Farley, P., 33
Fiengo, R., 73
Fillmore, C., 122
Flexible unlocked island, 18-22
Foot Feature Principle, 139, 148, 153-155, 157, 159
Frampton, J., 43

< previous page page_213 next page >


If you like this book, buy it!
< previous page page_214 next page >
Page 214
Freezing Principle, 166
French, 39, 75-77, 87, 88, 93, 181n, 182n, 186n
G
Gazdar, G., 2, 24, 52, 94, 98, 101, 137, 138-141, 144, 146, 147, 149, 152, 156, 157, 159, 162, 163, 192n, 197-199n
Generalized Phrase Structure Grammar (GPSG), 2, 24, 94, 95, 138, 139, 141, 142, 145, 148, 152, 153, 157, 158, 169-
171, 198n
Georgopoulos, C., 52
German, 13
Goldsmith, J., 53, 54, 77, 78, 80, 84, 86, 88, 185n
Goodall, G., 184n
Government-Binding Framework (GB), 1, 176, 177, 183n
Grimshaw, J., 108, 111, 115
Grosu, A., 51, 107, 122, 194n
H
Haegeman, L., 184n
Hankamer, J., 122
Head-Driven Phrase Structure Grammar (HPSG), 2, 24, 139, 141, 158, 176
Head Feature Convention, 139, 142, 153-155, 157, 198n, 199n
Hebrew, 11, 13, 17, 183n, 199n
Helm, I., 47
Heycock, C., 173
Higginbotham, J., 181n
Higgins, F. R., 110-112
Hirschbhler, P., 186n
Hooper, J., 70
Horn, L., 108
Hornstein, N., 72, 187n
Huang, C.-T. J., 43, 188n, 191n
Huck, G. J., 56, 59, 66, 186n, 190n, 191n
Hudson, R., 185n
Hukari, T. E., 109, 133, 143-145, 147, 148, 152, 155, 197n
I
Interwoven dependency construction, 79, 91, 108, 134-137, 161-163, 178
Irish, 11, 13, 17, 199n
Island, 3-6, 8-23, 37-39, 55, 56, 59, 60, 65, 69-75, 77, 80, 81, 88, 93, 95, 101, 104, 129, 146, 149-151, 165, 167-169,
172, 182n, 183n, 186n, 190n
Island constraint. See Island
Island Law, 38, 45
Italian, 186n
J
Jackendoff, R., 137
Jacobson, P., 56, 108, 109
Johnson, D. E., 107, 109, 180, 182n, 192n
K
Kalish-Landon, N., 122
Kaplan, R. M., 2
Kayne, R., 23, 42, 43, 98, 131, 133, 138, 166, 167, 177, 178, 179, 188n
Kiss, K.., 133
Klein, E., 112
Koster, J., 13, 37, 43, 61, 73, 74, 133, 169, 183n, 184n, 188n, 191n, 194n
Kroch, A., 2, 173
Kuno, S., 19, 107, 110
Kuroda, S.-Y., 122
L
Ladusaw, W., 102
Lakoff, G., 22, 23, 52-60, 65, 66, 70, 76, 77, 79, 80, 83, 85, 87, 89, 90, 92, 94, 165-169, 172, 185n, 186n, 189n, 191n
Langendoen, D. T., 3, 52, 122
Lasnik, H., 43, 61, 73, 105, 119, 125, 184n, 188n, 190n, 191n, 196n, 200n
Lawler, J., 83, 84, 86
Left-extraction (L-extraction), 1-4, 7, 11, 24, 25, 97, 99, 103-106, 108, 109, 111, 115, 120, 121, 124, 126-138, 141,145,
146, 151,152, 155, 156, 174, 176, 177, 179, 183n, 196n-198n
Levine, R. D., 2, 24, 98-100, 103, 104, 109, 121, 133, 138, 143, 146-148, 151, 152, 155, 156, 173, 197n
Lexical Functional Grammar (LFG), 2
Lightfoot, D., 187n
Linear structure, 53, 79, 89
Locked island, 12, 16, 17, 21, 182n
M
Maling, J., 140, 141, 143, 147
Manzini, R., 43, 61, 200n
May, R., 102
McCawley, J. D., 24, 92, 98-104, 106-108, 111, 114, 115, 199n-122, 127-130, 134, 135, 137, 138, 145, 146, 151, 157,
160, 173, 192n-194n, 196n
McCloskey, J., 11, 13, 23, 98, 140, 156, 157, 169-172, 184n, 192n, 200n
Metagraph Grammar, 182n
Mller, G., 104, 184n
Munn, A., 180
N
Na, Y., 54-56, 59, 66, 186n, 190n, 191n
Nakajima, H., 178
Napoli, D. J., 184n
No Recursion Constraint, 144, 145, 147, 148, 155, 156, 159
Nonlocal Feature Principle, 139
O
Obenauer, H.-G., 37, 43, 44, 181n, 188n
Object deletion, 69, 70, 123, 124
Object raising, 2, 69, 70, 108, 123, 124

< previous page page_214 next page >


If you like this book, buy it!
< previous page page_215
Page 215
Oehrle, R., 104, 122, 193n, 195n, 197n-199n
Ojeda, A. E., 98, 100, 138, 140, 145, 146, 151, 192n
P
Parallel-exfiltrations (P-exfiltrations), 140-142, 146, 154
Parasitic gaps (P-gaps), 2, 3, 7, 14, 19, 37, 65, 69, 70, 83, 123, 124, 133, 134, 136, 152-155, 157, 159, 181n, 183n, 195n,
196n, 199n
Perlmutter, D. M., 26, 36, 37, 188n
Pesetsky, D., 73, 133, 139, 142, 143, 158, 159, 184n, 190n, 191n, 196n, 197n
Piera, C., 24, 143, 145
Pollard, C., 2, 23, 122, 133, 139-141, 143, 145, 152, 159, 160, 167, 191n, 196n, 198n
Postal, P.M., 1, 7, 9, 13, 14, 16, 52, 60, 83, 107, 109-112, 114, 116, 117, 119, 120, 125, 131, 138, 166, 175-178, 180,
181n, 182n, 185-188n, 192-194n, 196n, 199n
Primary resumptive pronoun (RP), 12, 13, 15-18, 21, 189n
PRO gate phenomenon, 181n
Pullurn, G. K., 3, 23, 43, 61, 94, 107, 167, 176
Q
Quasi-exfiltrations (Q-exfiltrations), 143-145, 149-157, 159, 160, 199n
Quine, w. v. O., 143
R
Relational Grammar, 194n
Remnant, 104, 131
Resumptive pronoun (RP), 4, 5, 7-23, 26, 36, 38, 42, 43, 45, 47-49, 59, 60, 65, 68, 76, 95, 111-119, 138, 165, 182n,
183n, 186n, 188n, 189n, 194n, 199n
Riemsdijk, H. v., 125, 184n
Right node raising (RNR), 24, 97-101, 103-105, 107-109, 111, 113-117, 119-138, 145-147, 151, 152, 155-157, 160,
173-180, 186n, 192n, 193n, 195n-197n, 199n
Right node raising (RNR) pivot, 97, 100, 101, 103, 106, 114, 123, 126, 130, 131, 137, 146, 151, 155, 177, 178, 192-
194n, 197n
Right Selective Island Constraint (RSIC), 105, 106, 166, 167
Rigid unlocked island, 18-22
Rizzi, L., 43, 44, 61, 188n, 200n
Rodman, R., 126, 127
Romance, 186n
Ross, J. R., 1, 5, 8-11, 13, 16, 18, 22, 36-39, 42, 46, 51-53, 55, 56, 58, 59, 65, 70, 76, 93-95, 100, 107, 120, 121, 126,
138, 165, 167, 178, 180, 183n, 188-191n, 195n, 197n
S
Sag, I. A., 2, 23, 112, 122, 133, 139-141, 143, 145, 146, 152, 159, 160, 167, 180, 191n, 196n, 198n
Saito, M., 43, 61, 73, 105, 188n, 190n, 191n, 196n, 200n
Scandinavian, 145
Schmerling, S. F., 59, 70
Secondary resumptive pronoun (RP), 12, 13, 18-21, 189n
Selective island, 5, 6, 8-10, 12, 16, 19, 23, 42-46, 49, 60, 61, 63-71, 73-77, 79-81, 89-91, 93, 94, 165-172, 187-190n,
200n
Sells, P., 13, 183n, 196n, 199n
Shadow pronoun, 36. See also Resumptive pronoun
Slash category, 24, 98, 138, 140, 141-163, 169, 170, 195n, 197n, 199n
Spanish, 186n
Specified Subject Condition, 184n
Sportiche, D., 72
Steedman, M., 2, 198n
Stowell, T., 132
Strong crossover effect, 3
Subjacency, 103, 138, 168, 176. 180, 184n
Subject Condition, 184n
Swedish, 199n
T
Takami, K., 19
Tertiary resumptive pronoun (RP), 12, 13, 18-20
Thompson, S., 70, 107
ransformational Grammar, 1, 2, 37, 138, 162, 163
Tree-Adjoining Grammar, 2
U
Unlocked island, 12-15, 17-22
Uriagereka, J., 184n
V
Vergnaud, J.-R., 31
W
Weak crossover effect, 7
Western Austronesian. See Belauan
Wexler, K., 102, 104, 121, 137, 138, 145, 146, 151, 166, 175, 192n
Wh-Island Constraint, 184n
Williams, E., 125, 138, 184n, 186n
Z
Zaenen, A., 2, 140, 141, 143, 147

< previous page page_215


If you like this book, buy it!

Вам также может понравиться