JOURNAL OF SEMANTICS AN I NTERNATIONAL joURNAL FOR THE INTERDISCIPLINARY STUDY OF THE SE MANTICS OF N ATURAL LANGUAGE
MAN AGING EDIT 0R: PETER BoscH (IDM Scientific Cemre, Heidelberg and University of Osnabriick) ASSOCIATE EDITORS: MANFRED KRIFKA (University of Texas, Austin) RoB VAN DER SANDT (University of Nijmegen) REVIEW EDITOR: TIBOR K1ss (IBM Sciemific Centre, Heidelberg) ASSISTANT EDIT 0 R: BIANKA BuscHBECK-WOLF (IDM Scientific Centre, Heidelberg) EDITORIAL BOARD: N. AsHER (University of Texas, Austin) R. BARTSCH (University of Amsterdam) M. BIERWISCH (MPG and Humboldt University Berlin) B. BoGURAEV (Apple Computer Inc) M. Bo RILLO (University of Toulouse) K. BROWN (University of Essex) G. CHIERCHIA (University of Milan) 0. DAHL (University of Stockholm) S. C. GARROD (University of Glasgow) B. GEURTS (University of Osnabriick) M. HERWEG (IDM Scientific Centre, Heidelberg) L. R HoRN (Yale University) ]. jACOBS (University ofWuppertal) P. N. JoHNSON-LAIRD ( Princeton University) H. KAMP (University of Stuttgart)
S. LEVINSON (MPI Nijmegen) S. LOBNER (University ofDiisseldorD SIR joHN LYONS (University of Cambridge) A. MANASTER-RAMER (Wayne State University) ]. McCAWLEY (University of Chicago) M. MoENS (University of Edinburgh) f. J. PELLETIER (University of Alberta) M. PINKAL (University ofSaarbriicken) T. SANFORD (University of Glasgow) R. SCHA (University of Amsterdam) A. VON STECHOW (University ofTiibingen) M. STEEDMAN (University of Pennsylvania) W. WAHLSTER (DFKI, Saarbriicken) B. WEBBER (University of Pennsylvania) H. ZEEVAT (University of Amsterdam) T. E. ZIMMERMANN (University of Stuttgart)
EDITORIAL ADDRESS: Journal of Semantics, c/o Dr P. Bosch, IDM Germany Scientific Centre, Vangerowstr. 18, D-69115 Heidelberg, Germany. Phone: (49-6221-) 59-4251/4437· Telefax: ( 49-6221-) 59-3200. Email:
[email protected] © Oxford University Press All rights reserved; no part of this publication may be reproduced, stored in a retrieval system, or transmitted in any form or by any means, electronic, mechanical, photocopying, recording, or otherwise without either the prior written permission of the Publishers, or a licence permitting restricted copying issued in the Agency
Ltd,
UK by the Copyright Licensing
90 Tottenham Court Road, London W1P 9HE, or in the USA by the Copyright Clearance Center,
222 Rosewood Drive, Danvers, Massachusetts 01923, USA.
journal of Semantics (ISSN 0167
University Press, Oxford,
)IJJ)
is published quarterly in February, May, August and November by Oxford
UK. Annual subscription is
US$130 per year. Journal of Semantics is distributed by Mercury
Airfreight International Limited, 10 Camptown Road, Irvington, New Jersey 07111-110), USA. Second class postage paid at Newark, New Jersey, USA and at additional entry points. US POSTMASTER: send address corrections to journal of Semantics, c/o Mercury Airfreight International Limited, 10 Camptown Road, Irvington, New Jersey 07111-1105, USA.
For subscription information please see inside back cover.
JOURNAL OF SEMANTICS Volume 13 Number 1
CONTENTS R. HORN Exclusive Company: Only an d the Dynamics ofVertical Inference LAURENCE
ALEX LASCARIDES, ANN COPESTAKE AND T ED BRISCOE
Ambiguity and Coherence BART GEURTS
On No
41
W21lltceir @ce CGiruny�ceir 1B3ce Irlliirrn Nce w Yconrlk o
JH[iilla�ry Clhlappellll ainldl Wiilllliiam McG�regor (JEdliitOir§)
'Jflln � CG IT'ttnmmtaur ((])if Tirrnttnllii�nnttnlbiillii1y A 1I'ypollogiicall JP>erspecHve Oinl JBodly JP>airt 'Feirms andl tlhle JP>airt�Wllnolle JRellatiioinl 1995. 15,5 x 23 em. XIII, 931 pages. Cloth DM 398,- I oS 3104,- I sFr 377, ISBN 3-11-012804-7 (Empirical Approaches to Language Typology 14)
Mouton dle Gruyter
This volume presents invited research papers on the topic of
inalienability and the
personal
domain and its
grammatical expression in a variety of unrelated lan guages from Australia, Oceania, Asia, North America, Europe and Africa. A descriptive and typological frame work is used to address a range of current theoretical issues in semantics and syntax, including noun classes, genitives, nominal apposition and "possessor raising" constructions such as datives, noun incorporation and "favourite" constructions, for example, double subjects and double nominatives.
Walter de Gruyter & Co., P.O. Box 30 34 21, D-1 0728 Berlin, Phone: +49-30-260-05-235, Fax: +49-30-260-05-222 Walter de Gruyter Inc., 200 Saw Mill River Road, Hawthorne, NY I 0532, Phone: (914) 747 -0110, Fax: (914) 747-1326 Pl ea s e visit us in the World Wide Web at http:/www.deGruyter.de
Scope of this Journal The JOURNAL OF SEMANTICS publishes articles, notes, discussions, and book reviews in rhe area of narural language semantics. lr is explicirly inrerdisciplinary, in char ir aims ar an integration of philosophical, psychological, and linguistic semantics as well as semantic work done in artificial
imelligence and anrhropology. Contributions muse be of good qualiry (co be judged by ar lease cwo referees) and should relate co questions of comprehension and imerpreration of senrences or rexrs in narural language. The editors welcome nor only papers char cross traditional discipline boundaries, bur also more specialized contributions, provided rhey are accessible co and inreresting for a wider readership. Empirical relevance and formal correcrness are paramoum among rhe criteria of acceptance for publication.
Information for Authors:
Papers for publication should be submitted in 3 copies co che
managing edicor. They should be ryped on A4 (or similar format), one-sided, double-spaced, and
with a wide margin and muse be accompanied by an approx. 200 word summary. Notes and biblio graphical references muse appear ar rhe end of rhe rypescripr. All bibliographical references in che rexr by aurhor's surname and year of publication. Diagrams muse be submirred camera-ready. All papers submirred are subject co anonymous refereeing and are considered for publication only on rhe assumption char rhe paper has neither as a whole or in parr already been published elsewhere nor has elsewhere been submirred or accepted for publication. Unless special arrangemenrs have been made, copyright resrs wirh Oxford Universiry Press. Authors receive 20 offprinrs of their published articles and 10 offprinrs of their published reviews, free of charge. Larger numbers can be supplied ar cost price by advance arrangement.
Subscriptions: The Journal of Semantics is published quanerly.
Institutional: UK and Europe £72.00; USA and Rest of World US$1JO.oo. {Single issues: UK and
Europe £21.00; USA and Rest of World US$38.oo.)
Personal:• UK and Europe £3s-oo; USA and Rest of World US$66.oo. (Single issue: UK and Europe
£w.oo; USA and Rest of World US$19.00.)
• Personal rates apply only when copies are sent to a private address and payment is made by personal cheque/credit card. Prices include postage by surface mail or, for subscribers in rhe USA and Canada by Airfreight or in Japan, Australia, New Zealand and India by Air Speeded Post. Airmail races are available on request.
Back Issues. The current plus cwo back volumes are available from the Oxford Universiry Press,
Walton Srreet, Oxford OX2 6DP. Previous volumes can be obtained from Dawsons Back Issues, Cannon House, Park Farm Road, Folkescone, Kent CT19 sEE, eel +44 (o)1303 8soro1, fax +44
(o)1303 850440. Volumes 1-6 are available from Swers and Zeirlinger,PO Box 830,2160 SZ Lisse, The Netherlands.
Payment is required wirh all orders and subscriptions are accepted and enrered by rhe volume. Pay menr may be made by cheque or Eurocheque (made payable co Oxford Universiry Press), National Girobank (account soo 1056), Credit cards {Access, Visa, American Express, Diners Club), or UNESCO coupons. Please send orders and requests for sample copies co the Journals Subscriptions Deparrmem, Oxford Universiry Press, Walton Srreer, Oxford OX2 6DP, UK. Telex 837330 OXPRES, tel +44 (o)1865 267907, fax +44 (o)186s 267485.
Copyright: Ir is a condition of publication in rheJournal char authors assign copyright co Oxford
Universiry Press. This ensures char requesrs from third parties to reproduce articles are handled
efficiendy and consisrendy and will also allow the article co be as widely disseminated as possible. In
assigning copyright, authors may use their own material in other publications provided char rhe Journal is acknowledged as rhe original place of publication, and Oxford Universiry Press is notified in writing and in advance.
Advertising: Advertisemenrs are welcome and races will be quoted on request. Enquiries should
be addressed to Jane Parker, Oxford Journals Advertising, 19 Whitehouse Road, Oxford OX1 4PA, UK. Tel/fax: + 44 (o)1865 794882. Email:
[email protected].
journal ofSemantics
13: 1-40
© Oxford University Press 1996
Exclusive Company: Only and the Dynamics of Vertical Inference LAURENCE R. HORN
Yale University
Abstract
The goal of this paper is to find out whether for our purposes, we do. 1 I S Y M M E T R I C A L I S M A N D I T S D I S C O N TE N T S The lion's market share i n this industry has always been controlled by the conj unctionalists. For Peter of Spain and his fellow scholastics, ( 1 a) entails the conjunction of (I b) and ( I c) and thus entails each of them singly. ( I ) a. Only God can make a tree. b. God can make a tree. c. No one distinct from God can make a tree. The 'only' particle, so/us or tan tum , is an EXPONIBLE to be expounded or unpacked into 'an affirmative copulative proposition whose first part is the same proposition without only '-the 'PRAEIACENS' or PREJACENT-'and whose second part is a negative proposition denying the predicate of all others apart from the subject' (Tractatus Exponibilium 2I f£, in Mullally 1 945: 1 06-7). Peter's analysis has been co-sponsored or endorsed by a wide range of twelfth-, thirteenth-, fourteenth- and twentieth-century scholars,2 from Burley (c£ Pinborg I 98 1 ; De Ryck 1 98 5), William ofSherwood (c£ O'Donnell 1 94 1 ; Kretzmann I968), Ockham (I 98o: I 3 3), Billingham and Alnwick (c£ De Ryck r982) to Kuroda ( 1 969: 348), Lakoff ( 1 970), Taglicht (r 984), van der Auwerea ( I 98 5), Keenan & Stavi ( 1 986), Krifka (I 992), Moser ( I 992), Burton Roberts ( I 993), and von Fintel ( I 993). A related proposal is elaborated in two papers in this journal by Jay Atlas ( I 99 1 , I993); on Atlas's account, (2a) unpacks
Downloaded from jos.oxfordjournals.org by guest on January 1, 2011
The semantics of only says this: it asserts that no proposition from the set of relevant contrasts C other than the one expressed by irs sister sentence a is true. There is in addition an implicature that a is in fact true. There is an industry devoted to the issue of whether the latter ingredient is an implicature (conversational or conventional), a presupposition, or part of the truth-conditions . . . For our purposes, we don't need to decide. (von Fintel 1994: 133)
2
Only and the Dynamics of Verrical lnference
not simply into the Perrine conjunction in (zb), but into the more complex representation of (zc).
(2) a. Only Muriel voted for Hubert.
b. Fa & -3y(y # a & Fy) (e.g. 'Muriel voted for Hubert and nobody other than Muriel did' c. 3xV'y[(x - y .... Fy) & (Fy-+ y - a)] e.g. 'Exactly one individual, and no one other than Muriel, voted for Hubert'
(3) a. Only Kim can pass the test, and it's possible even she can't. b. #Only Kim can pass the test, [and/butJ it's possible that someone else can. c. #Everyone passed, but [for all I know/it's possible that J someone didn't. d. All the world is queer save thee and me, and even thou art a little queer. (attributed to Robert Owen on separating from his business partner in I 828) c. All the world is queer save me and thee, and sometimes I think thee is a little queer. (attributed to 'an unidentified Quaker speaking to his wife') £ -Only Kim can pass. [ vs. # Kim and only Kim can pass.] -Yes, that's true. %And [what's more/in fact), maybe even she can't. Notice the epistemic qualifier that must in general be present w hen suspending the prejacent: we can't normally say that only Kim can pass the test, and perhaps she can't. But the contrast between (3a, b) is undeniable, the latter as anomalous as any attempt to remove an entailment recognized by the speaker, as in (3c). Further, there are attested citations in w hich an exception is canceled without benefit of an epistemic rider, such as that in (3d), a variant of the Bartlett's familiar quotation in (3e). The exclusive counterpart of these exceptives would be something like 'Only you and I are sane and sometimes I wonder about you'. I find the exchange in (3f) imaginable as well, although intuitions apparently differ here.
Downloaded from jos.oxfordjournals.org by guest on January 1, 2011
Since it is a consequent of his account, one he regards as a virtue, that (I a) entails (I b) as well as ( I c), I'll regard it for now as a variant species of the same symmetricalist genus. But if only sentences have positive entailments, how do we account for the data in (3)-(s), w hich seem to demand an asymmetric theory on w hich the negative exponent as in (I c) is somehow more equal than its positive counterpart ( I b)? Note for starters that the pattern in (3) (see Horn I 969, I 970, I 972) shows that the prejacent can be suspended with the help of an epistemic rider, w hile the exclusionary clause-like any true god-fearing entailment cannot be.3
Laurence R Hom
3
As we shall see in more detail below, only subjects license negative polarity items (such as those boldfaced below) in the predicate, as in a classic example cited by Klima ( I 964: 3 I I ) as evidence for the [+ affective] status of only. (4) Only young writers ever accept suggestions with any sincerity. Further, fronted only phrases (with an adverbial, PP, or NP object in focus) trigger inversion, a property that in modern English is confined to phrases of negative character. The examples in (sc-f) are lifted from Jacobsson (I 95 I : so ss), Jacobson ( I 964: 309-1 0), and Visser (I969: § 1 436).
Smith ( I 922) notes the minimal pair in (se-D, where narrow-scope only fails to license inversion in the latter case, but this is the same point made by Klima (I 964) and others with respect to the classic pairs in (s ) : '
(s ') a. Not even two years ago could you swim there. (-you couldn't swim there even 2 years ago) a'. Not even two years ago, you could swim there. (=you could swim there not even 2 years ago) (Klima I 964: 300) b. In not many years will Christmas fall on Sunday. b'. In not many years Christmas will fall on Sunday. (Klima I 964: 30 I ) c. In no clothes does Robin look attractive. c'. In no clothes, Robin looks attractive. (sex-neutral adaptation of examples from C. Bird cited by Jackendoff I 972: 364) Further, inversion in the w ide-scope contexts exemplified in (sa-e) is obligatory: 'In the course of fairly extensive reading in more recent literature, I have not come across a single example of straight order after only modifying a whole clause' aacobsson I 95 I: 54).
Downloaded from jos.oxfordjournals.org by guest on January 1, 2011
(5) a. Only in Montreal can die-hard film buffs expect to see not only one but three films from Mongolia. (New York Times article on the Montreal Festival des filmes du monde, I 3 August I 995) b. Only in New York can one seek shelter from a storm under takeout mussels in w hite wine sauce. (New York Times Metropolitan Diary, 20 July I 994) c. Only in stories does a dropped glass betray agitation. (Graham Greene) d. Only thus can the corruption of society be checked. e. Only one feature did I notice in the landscape, a large, white villa . . . ( Only a year before he had built himself a mansion in Kensington.
4
Only and the Dynamics of Vertical Inference
In exhibiting the behavioral symptoms of downward monotonic expres sions, only a phrases are for most speakers differentiated from the correspond ing non-monotonic a and only a conjunctions. Thus compare:4 (5 ) a. Tonight and only tonight, the opalescent squid will mate. In a frenzy they seize each other, coup ling again and again . . . (Voice-over from 3-D IMAX movie Into the Deep) b. %Tonight and only tonight will the opalescent squid mate. "
(6) Horn (I 969): only (x �a, Fx) Presupposes: Fa Asserts: -3 y(y # a & Fy) whose silver anniversary the world recently celebrated, (7a) p resupposes rather than entailing (7b), thus preserving the distinction berween the negative asserting only Muriel sentence and its truly conjunctive Muriel and only Muriel counterpart in (7d). (7) a. b. c. d.
Only Muriel voted for Hubert. Muriel voted for Hubert. Nobody distinct from Muriel voted for Hubert. Muriel and only Muriel voted for Hubert.
On this view, the suspension, polarity, and inversion facts of (3)-(5) would follow from the premise that these p roperties are associated with the negative assertion or entailment of the only sentence. A number of subsequent analyses (e.g. Konig I 99I; Barker I993) have essentially followed the same line, which is also echoed in the otherwise rather divergent analyses of negative polarity phenomena in Ladusaw (I 980) and Linebarger (I 98 I , I 987).5 Some asymmetricalists have p ushed the p resuppositional envelope further into the pragmatics, deriving (7b) as a non-truth-conditional aspect of conven tional meaning-a conventional implicatUre or pragmatic p resupposition-or even as a conversational implicatUre. On such accounts, the positive comp onent of only is not part of what is said but part of what is implicated in an utterance like (7a). Thus in Horn (I979), the positive component of meaning contributed by only is taken to be a non-truth-conditional constraint on the appropriate assertion of (7a); the notion appealed to (with some reservations) is that of conventional implicature, a la Grice ( I 97 5) and Kamunen & Peters ( 1 979).6 The same position is endorsed in Rooth ( I 98 5), Krifka (1993), and Fretheim ( I 995). More radically, the positive component can be taken to represent not a
Downloaded from jos.oxfordjournals.org by guest on January 1, 2011
In the light of these diagnostics, we m ight consider casting aside the symmetric co1�unctive theory in favor of one of a range of asymmetric analyses which agree that the relation of ( 1 a) to (I b) is not logical entailment while differing on just what if anything it is. On the analysis in (6),
Laurence
R.
Horn
S
2 M O RE E V ID E N CE F O R A S Y M M E T RY: T H E CASE O F ONL Y IF Before insp ecting these asymmetricalist approaches for flaws, it's worth noting that the distribution and meaning of only if reinforce the conclusion that the positive component of only clauses is not simply entailed, or else A only ifB would be equivalent to A ifand only ifB. As McCawley has observed, A only ifB is better paraphrased by 'Not A if not B' than by 'If A then B' (or by 'B if A'); he points out that (Sa) is equivalent not to (Sb) but rather to (Sc) (McCawley 1 9S I: 49-54; cf also McCawley I 974 and 1 993: S I -S and, independently, Barker I 993)-7 (S) a. You're in danger only if the police start tapping your phone. b. If you're in danger the p olice start tapping your phone. [;£ (Sa)] c. If the police don't start tapping your phone you're not in danger. [= (Sa)] Crucially, as observed by Geis ( I 973), van der Auwera ( I 9S s), Lycan ( I 99 I ), and von Fin tel ( I 994) in addition to McCawley and Barker, only if is compositional, its meaning a function of the meaning of only plus the meaning of if Following Geis and McCawley, we can read only ifB as 'under no circumstances other than B', in the same way that only on Sundays is read as 'on no days other than Sundays', so that (9a) amounts to the negative conditional in (9b). On the related account of Lycan ( 1 99 I : 1 26-7), (9a) is analyzed into (9c).
Downloaded from jos.oxfordjournals.org by guest on January 1, 2011
conventional bur a conversational implicature. McCawley ( I 9S I: 227) is the first to argue explicitly that the use of(7a) conversationally implicates (7b): if the speaker knew that nobody voted for Hubert, it would be a violation of the Quantity maxim (Grice I 975) to assert (7a) (cf Horn I 992 in defense of this view and Atlas I 993 in opp osition). The most asymmetric theory of all would of course be one on which the logical form of(7a) is just that of (7c), with the p ositive comp onent excised. This is in fact Theory G, the line urged by Geach (I 962: I S 7) in Reference and Generality on which no relation whatever obtains between the only expression and its positive counterpart. Thus 'F (some a )' is not deducible from 'F (only a )' either as an entailment or as a non-truth-conditional aspect of conventional force; nor is there any pragmatic inference involved. Geach's argument from logical convenience-'It is formally much more convenient to treat the exclusive p roposition as having precisely the exclusive force of its supposed second [negative] component'-appears to fly in the face of intuition, entailing as it does that "'F (only 8)" will thus be true when "F( )" is true of nothing at all', rendering Only Hilary could reform health care vacuously true.
6
Only and the Dynamics ofVertical Inference
(9) a. I'll go only if you go. b. I won't go if you don't go. c. I'll go only in [the event that/events in which) you go. (- I'll go in no event other than one in which you go) And so, not surprisingly, we have the parallel suspension and cancellation effects via the epistemic riders of the rype of (I o): ( 1 0) a. He goes to church only on Sunday, and not even then if there's a foot ball game on television. b. He'll go to church only if you do, and not even then if . . . c. I'll go only if you do. And maybe not even then.
(I o') a. # He goes to church on Sunday and only on Sunday, and not even then if there's a football game on television. b. # He'll go to church if and only if you do, and not even then if . . . c. # I'll go if and only if you do. And maybe not even then. The inversion and polarity facts pattern accordingly:8 ( I I ) a. b. c. b'. c'.
If you build it, they will come. Only if you build it will they come. If and only if you build it, they will come. Only then will they come. Then and only then, they will come.
(. . .*will they come) (. . .*they will come) (. . .%will they come)9 (. . . *they will come) (. . . %will they come)
Note the contrast with non-negative adverbials (as in ( I I ' )) (I I') Even [if you build it/then), they will come.
(. . .*will they come)
and with overtly conjoined a and only a conditionals (as in ( I I c) and ( I 2 )) . ( I 2) a. Only if George works hard can he ever hope to pass. (Barker I 99 3 : 2 56) b. *If George works hard [he can/can he) ever hope to pass. c. If and only if George works hard [*he canl%can he) ever hope to pass. Curiously, the parallelism between if and only if on the one hand and Muriel and only Muriel on the other is supported by the conjunctionalist van der Auwera ( r 98 5 and personal communication, I 4 June 1 994), who argues that just as 'Horn's old example' (7a) is equivalent to (7d), so too if and only if is essentially just a srylistic/pragmatic/discourse variant of the semantically and truth-conditionally identical only if In the light of the contrasting patterns we have observed here, that conclusion is ruled our and, given the parallelism with
Downloaded from jos.oxfordjournals.org by guest on January 1, 2011
In either case, a switch to an overt conj unction yields a contradiction with the cancellation clause:
Laurence
R.
Horn
7
'ordinary' only sentences, a conjunctionalist account of the latter is equally untena ble. 1 0 By way of summary we can offer a State of the Industry report, with a display of the competing theories on only ranked in order of increasing asymmetry: ( I ) a. Only God can make a tree. b. God can make a tree. c. No one distinct from God can make a tree. Theory C/CA• the conjunction analysis:
( I a) entails the conjunction of ( I b) and
(I c).
On the last four theories, the prejacent ( I b) does not follow from ( r a) by classic entailment; on the last three, it does not represent a truth-conditional inference, and on the latter two it has no semantic status whatsoever.
3 T H E S Y M ME T R I C A L I S T S S T R I KE B A C K : P O L A R I TY A N D M O N O TO N I C I TY In recent work, jay Atlas has subjected the last three approaches to a withering critique on the grounds that we must ofcourse reject any analysis that takes a sentence like Only Hillary could reform health care to be true-and merely misleading-on the grounds that nobody could reform health care. As (I 3a) shows, the putative prejacent implicature staunchly resists cancellation. ( 1 3) a. #Only Hillary could reform health care, and even she couldn't. b. #I love only you, but (maybe) I don't love you either. Of my acknowledgement (Horn 1 992: I 82) that Theory M predicts the semantic coherence of (I 3 b), Atlas retorts ( I 99 3 : 3 I 4 and fn. on 3 I 7): [This) consequence strikes me as so ourrageously counter-intuitive as to be a reductio of the theory. The idea that simultaneously I love only you could be true while I love you is false just
Downloaded from jos.oxfordjournals.org by guest on January 1, 2011
Theory P, the presupposition analysis: ( I a) entails ( I c); ( I a) semantically/ logically presupposes ( I b). If(I b) is false, ( I a) is neither true nor false. Theory PP, the pragmatic presupposition or conventional implicature analysis: as in P except (a) pragmatically presupposes/conventionally implicates ( 1 b). (I b) is non-truth-conditional but semantic condition on the appropriate utterance of ( I a). Theory M, McCawley's conversational implicature analysis: (ra) entails (Ic); a speaker uttering (ra) normally (but not necessarily) conversationally implicates (I b). Theory G, Geach's pure negative analysis: (I a) entails ( r c); no representation of any relation between ( I a) and (I b).
8
Only and the Dynamics of Vertical lnference
seems crazy to me . . . I just have a very hard time with Horn's theory about the truth conditions of I love only you.
Further, Atlas denies that only NP subjects are monotone decreasing (or downward entailing), given the lack of any valid inference from (I 4-a) to ( I 4-b). ( r 4) a. Only Socrates entered the race. b. Only Socrates entered the race early. This observation actually dates back seven centuries to Peter of Spain, who notes (in the Tractatus Syncategoremata, M ullally I 945: 3 3) that no entailment goes through either way in ( r 4 '): a.
b.
(
)
Only Aristotle moves. Only Aristotle runs.
But Peter didn't have polarity or inversion to worry about. How does Atlas account for the downward or negative effects associated with only NP? The suspension facts he dispu tes, along with the putative compositionality of only if The inversion facts he does not acknowledge. And as for polarity licensing, Atlas maintains on the basis of the contrast between (I sa) and the more marginal or unacceptable (I 5 b-d) ( I 5) a. b. c. d.
Only John ever suspected David Alexander. *Only Bill wants Sam to finish the report until Friday. *Only Phil will give Lucy a red cent. ?Only I was all that keen to go to the party.
that-contrast to popular opinion, including that of virtually everyone who has considered the issue, from Klima11 to, amazingly, both Ladusaw and Linebarger- only NP does not in fact license NPis: I do not find the syntactic observation that only triggers Negative Polarity Items sufficiently well-grounded, and so I do not believe that there is yet sufficient ground to take only to be a negative lexical item. (Atlas 1993: 3 13)
In fact, though, popular opinion is correct, as the data in ( I 6) show. ( I 6) a. (Of all her friends,) Only Phil would lift a finger to help Lucy. b. Only your wife gives a hoot about what happens to you. (McCawley I 98 I : 8 3) c. My nose and my lungs are only alive at all because they are part of my body and share its common life. (C. S. Lewis, Mere Christianity, cited in Jacobsson I 95 I ) Only N P is indeed a polarity licenser, although as a non-overt negator and a non-anti-additive quantifier in the sense ofZwarts ( 1986) and van der Wouden ( I 994), not every polarity item will freely occur within its scope. In this respect,
Downloaded from jos.oxfordjournals.org by guest on January 1, 2011
(14')
Laurence R Horn
9
only-initial phrases are quite parallel to another quantifier whose downward monotonic status is widely conceded, NPs of the form jew + CN. The data in (I 7) review the non-discrete nature of polarity licensing; note that both semantic and morphological considerations come into play. 12
(r7) a. John didn't realize that
?#Only Chris slept a wink last night. ?Nobody but Chris slept a wink last night. Nobody slept a wink last night but Chris.
Whether or not it is a full-fledged downward monotonic operator on its own semantic merits, only NP patterns with the downward-entailers and in contrast with the truly non-monotonic NP and only NP, which triggers few if any polarity, inversion, or cancellation effects. And while the intuitions on (I 3) and (I 4) of Peter of Spain and Jay of California cannot be simply dismissed, it's worth reviewing some evidence, especially an elegant argument due to the thirteenth-century logician Walter Burley, suggesting that only phrases may in fact be more downward monotonic than they appear. For Ladusaw ( I 98o: I 6 5-6), the assumed polarity-triggering effect of only phrases follows from their downward mono tonicity, as seen in (I 8): (I 8) a. Only Uohn]F had ever read anything about phrenology. b. Only John walks- Only John walks slowly. Recall that while upward monotonic quantifiers license inferences from subsets to supersets, as in (I 9a), downward monotonic quantifiers license inferences from supersets to subsets, as in (I 9b). Non-monotonic quantifiers, typically analyzable as conjunctions of quantifiers of mixed monotonicity types (e.g. exactly n - at /east n + at most n) license inferences in neither direction.
( I 9)
a. Some Greeks voted for Bill. b. No Greeks voted for Bill.
c, >' •
c. Exactly 756 Greeks voted for Bill. c; •
(zo)
mont: monl:
Some Greeks voted. No Greeks voted. Exactly 756 Greeks voted.
some men, many women, Hillary, most Democrats, all linguists, everyone, the tallest Jo-year-old, {at least} 1 o cats . . . no men,few women, at most 1 o chickens, not every doctor, nobody . . .
Downloaded from jos.oxfordjournals.org by guest on January 1, 2011
you had ever done a thing like that . . . ?*John were here anymore . . . *John had arrived until midnight . . .
b. If
c.
he had swallowed any marbles. ?*Mary loved him anymore. *Ralph arrived until midnight.
10
Only and the Dynamics of Vertical Inference
recognized by Barwise & Cooper ( I 98 I : I 93). quantifiers with the same monotonicity orientation combine with and, excluding but, as seen in (2 1 ),
fu
few violins and no cellos (2 1 ) a man and three women some students and every professor no men and very few women no dogs and hardly anyNPI cats most men and anyFc women while monotonicity-discordant quantifiers most combine with but:
Crucially, only phrases pattern as downward monotonic quantifiers: (2 3) many men (#and/but] only 3 women no men (and/# but] only 3 women 4 C O N V E R S I O N A N D E X I S TE N T I A L I M P O R T But in fact only phrases must be downward monotonic, given the converse relation between only and all, recognized since the medievals: (24) a. Tantum animal est homo convertitur in istam: omnis homo est animal, per is tam regulam: Exclusiva affirmativa convertitur in universalem . JJ Qohn of Holland, in Bos 1 985: 27) b. only (A, B) (only As are Bs) - all Bs are As - B �A The interdefinability of exclusives and universals appears in a variety of tradi tional sources, e.g. Peter ofSpain's Tactatus Exponibilium (Mullally 1 945: 1 06-7), and is exploited, defended, or assumed-often in the form of the subset relation in (24b))-by a bevy of moderns, including Lobner (I 987), Chierchia & McConnell-Ginet ( 1 990: 427), de Mey ( 1 99 I ), johnson-Laird & Byrne ( 1 99 1 : 1 28), Higginbotha m (1 994: 459-63), and von Fintel ( 1 994, 1 995), enoug h of a quorum to challenge the assertion that 'these days, those who adopt this analysis are few and far between' (von fintel I 994: I 3 5). Given this convertibility, to say that only Democrats supported Clinton is to say that all Clinton-supporters were Democrats. But, as has been recognized for a couple of millennia (cf Horn 1 989: Section 1 . 1 .3 for summary and Moravcsik 1 99 1 for a new look), there is an existential inference, generally assumed to hold in the pragmatics, that is characteristically associated with the assertion of a universal. 14 Thus we can infer that (for all the speaker knows) there were indeed Clinton-supporters; otherwise the all-statement would be too uninformative to assert. 1 5
Downloaded from jos.oxfordjournals.org by guest on January 1, 2011
(2o) John (#and/but] no woman few women (#and/but] many men most of the dogs { #and/but] few of the cats (vs. most of the dogs (and/#but] a few of the cats
Laurence
(z s)
Only Democrats supported Clinton
(U?) Democrats supported Clinton
� r-
R.
Horn 1 1
All Clinton-supporters were Dem
rats
F
There were Clinton-supporters
:. There were Democrats who were Clinton-supporters
(26) a. b. c. d.
Only professors are confident. Professors are confident. All professors are confident. Some professors are confident.
Such cases thus provide a difficulty for the conjunctive analysis, even assuming the coherence of an entailment analysis for corresponding stage level predic a tions, as in (z s) or in (26 '). -
(26 ') a. Only professors are ava ilable. b. Professors are available. The point carries over, mutatis mutandis, from bare plurals to definite NPs in the focus of only. When Thomas Wolfe stipulated sixty years ago in his eponymous story that 'only the dead know Brooklyn', he wasn't committing himself to the claim in (27b), which seems to suggest that the whole ghostly crue-in particular Aristotle-knows Brooklyn. (27) a. Only the dead know Brooklyn. b. The dead know Brooklyn. Nor does (27a) enta il the truth of a generic (rather than strictly universal) reading of (27b). Rather, the commitment is merely existential, as is clear from the opening sentence of the narrative: 'Dere's no guy livin' dat knows Brooklyn t'roo and t'roo, because it'd take a lifetime just to find his way aroun' duh f rown' (Wolfe 1 93 5: 9 1 ).
Downloaded from jos.oxfordjournals.org by guest on January 1, 2011
But now we obtain the conclusion that there were indeed Democrats who supported Clinton, which is in fact the strongest positive proposition licensed by (z s). Another way to put the same point is that it's just as true, but just as uncooperative, to assert (zs) if you know that nobody supported Clinton as it is to assert that all John's children are bald on the grounds that John is childless.16 For a conjunctionalist, the unpacking of(zs) yields the indefinite proposition Democrats supported Clinton, which as a stage-level predication (see Carlson 1 977) can be read existentially. But, as von Fintel ( 1 994: 1 39 ff) points out, no such reading is possible when we shift to individual-level predicates. Thus (26a) does not entail (26b), which amounts essentially to the universal claim in (26c). The speaker of (26a) is at most committed to the weaker exitential claim that there are professors who are confident, i.e. to (z6d).
12
Only and the Dynamics of Vertical lnference
Thus, a claim about only + CN is simply not equivalent to a claim about all and only the members of a given set, a fact rendered explicitly in (28a), excerpted from an Internet posting on lactose intolerance: (28) a. Only humans and not even all humans are genetically equipped to drink milk into adulthood. b. Humans are generically equipped to drink milk into adulthood.
(29) a. Only birds have fea thers. b. All birds have feathers. c. All and only birds have feathers. d. All feathered things are birds. In recognizing the mutual convertibility of exclusives and universals, we obtain the equivalences of (3o), preserving monotonicity orienta tions for rhe corresponding argument positions as indicated. 1 8 (3o)
a. only As are Bs
4-1>
all Bs are As 4-1> B !:;;A
b. only (A)(B)
4-1>
every (B)(A)
LL
Jj
DE environments non-DE environments
Finessing the issue of existential import, only behaves like an upward monotone determiner that combines with a focus argument to form a downward monotone quantifier, while every! all is a downward monotone determiner that combines with a restrictor to form an upward monotone quantifier. We can represent the correspondence in (30) schematically by thrce-dimensionalizing the traditional Aristotelian square of opposition:
Downloaded from jos.oxfordjournals.org by guest on January 1, 2011
The corresponding claim in the prejacent, (28b), is not interpreted existentially but universally or at least generically, and again represents a stronger com mitment than is licensed by the assertion of(28a). Of course, an individual-level only predication may be consistent with the corresponding all and only statement, as in (29), but the point is that (29b, c) are not valid consequences of (29a), even though they happen to be true. What follows from (29a) is not (29b) or (29c), but (29d).17
Laurence R. Horn onlyPisS
only non-P isS E
Al!
/
allSisP
noSisP
A!f
0
P also isS oo� ooo-P •S
Iff
not onlyPisS
someSis notP not all SisP
someSisP
But it is far from clear that only is ever actually a determiner. Note that even if it is a determiner when it takes a common noun focus, as in (29a), this analysis fails to extend to the cross-categorial only operator that combines with or modifies full NPs (including proper names, descriptions, and pronouns), VPs, PPs, and adverbs. Note in particular that only combines with the in either order: only the N, the only N. Although the conjoinability of only with all in contexts like (29c) seems to testifY to its Det-hood, all and only may not be a determiner here, as von Fintel ( 1995) has noted: cf. All and only the migratory birds ... The category status of only! CN is of theoretical interest in that if only is a determiner in this frame it stands as the unique counter-example to an otherwise striking generalization that all natural language determiners are conservative (Keenan & Stavi I 986). Conservativity is defined as in (3 I) and illustrated in (3 I '): __
(3 I ) a determiner 0 is CONSERVATIVE iff [D(A)(B)- D(A)(A " B)) (3 I ') a. No ravens are white No ravens are white ravens b. Most owners are greedy Most owners are greedy owners c. Only willows weep of Only willows are weeping willows Only is not conservative because the biconditional in (3 I ' c) is only valid left-to right: only willows are willows that weep but not only willows weep. But if only is never a determiner, always combining with full NPs including those with null determiners as in (29a) and (3 I ' c), the problem for the conservativity universal is dissolved; see Thijsse (I 98 3); Chierchia & McConnell-Ginet ( I 990:
Downloaded from jos.oxfordjournals.org by guest on January 1, 2011
/
I3
14
Only and the Dynamics of Vertical Inference
426 8); van Benthem ( I 99 I : fn. 4); and von Fintel (I 995) for arguments to this effect. De Mey ( I 99 I ), on the other hand, argues that only is a determiner in sentences like (3 I ' c). Even if de Mey is essentially correct, however, there is another take on the question. 1 9 As van Benthem ( I 986: 8) observes, conservativity is motivated by 'the privileged role of the first argument in a determiner statement it "sets the stage" '. But in an only sentence like (3 I ' c), it is not only plus its focus argument that 'sets the stage', as the principle of conversion shows. If (3 I ) represents traditional conservativity, let us say that every natural language determiner is either traditionally conservative or NEO CONSERVATIVE in the sense defined in (32) and illustrated in (32'): -
'
All weepers are willows� All weepers are weeping willows. (by TRAD-CONS)
T H E D O C T R I NE O F V E R T I C A L I N FE RE N CE
s
The scholastic version of monotonicity or vertical (upward and downward) inference can be conceptualized as a set of instructions for permissible moves in climbing up and down the tree of Porphyry, a hierarchical arrangement of categories devised by Porphyry of Tyre (An 300) that collapses the relations of set membership and set inclusion: substance
I
I
\ body I I \
animal
I
I
man
\
\
ass
So< rates
Vertical inference (as in the Alnwick and Billingham treatises in De Rijk I 982, inter alia; cf Sanchez I 994 for elaboration) is captured by the following rules:20 •
e
There is a valid consequence from an inferior to its superior without nega tion or a term having negative force. There is a valid consequence from a superior to its inferior with a preposed negation or distributive term.
Downloaded from jos.oxfordjournals.org by guest on January 1, 2011
(32) a determiner D is NEO-CONSERVATIVE iff [D(A)(B)- D(A n B)(B)] (32 ) Only willows weep� Only F [weeping willows] are weepers. (by NEO-CONS)
Laurence R Horn 1 5
(3 3) a. b. c. d.
Tantum Sor[tes] videt animal. 'Only Socrates sees an animal' 'Only Socrates sees an ass' Tantum Sor[tes] videt asinus. 'Every man is an animal' Omnis homo est animal. 'Every white man is an animal' Omnis albus homo est animal. (Burley, 'De exclusivis': Pinborg I 98 I; De Rijk I 98 s; emphasis added)
For Burley, the oddness of the downward inference from (3 3a) to (3 3b) in the case where the animal Socrates sees is an ox rather than an ass is no more problematic than the well-established downward inference from (3 3c) to (3 3d), which remains valid in the absence of white men. The issue of existential import is to be handled similarly in both cases. And while, as we saw, Petrus Hispanus and Atlas Pomonensis reject the inference from (34a) to (34d) (see ( I 4), ( I 4 ) above), Burley derives it by double conversion: (34a) converts to (34b ), which licenses the downward inference of (34c), which converts back to (3 4d). '
(34) a. Tantum hom movetur. 'Only a man is moving'
�
b. Omne movens est homo. 'Every moving [thing] is a man' ij
d. Tantum homo currir. B roughly means 'If A , then normally B'. CE exploits nonmonotonic validity (f'), which has the following attractive properties:
so Ambiguity and Coherence
We will also on occasions use the notation :J KB(A , B) to mean � (KB 1\ A , B) 1\ -. � (KB, B). In other words, � KB(A , B) means that B nonmonotonically follows from the KB augmented with A, but not from the KB alone. The details of the logic can be found in Asher & Morreau ( I 99 I ) and Asher (I 99 3 b). We have examined in detail elsewhere how to represent different pieces of background knowledge as formulas of CE in DICE (Lascarides & Asher I 99 I , I 993; Lascarides & Oberlander I 993; Asher & Lascarides I 994, I 99Sa), so as to model the pragmatic interpretation of discourse. Here we discuss only those aspects of the representation that are crucial to zeugma. First, consider (I sa). ( I s) a. Rembrandt used a brush. b. The janitor used a brush. Since Rembrandt using a paint brush is the default interpretation of(I sa) in the absence of further context, we can represent this information as a >-rule in CE, which will include some compositional semantic information in the antecedent-such as Rembrandt is the agent-and the information that the brush in question is a paint brush in the consequent. We will call this rule Specializatio�, since it 'specializes' the meaning of (general) brush to paint brush. We assume this rule will also capture the implicatures that the brushes are paint brushes in (z i ) : (2 I ) a. Rembrandt didn't use a brush. b. If Rembrandt used a brush, he will regret it. There will be a similar >-rule which captures the default interpretation of (I 5 b)-call this Specialization/ The detailed specification of these rules does not concern us here, since all that matters for our purposes is their logical relation to the other default rules for pragmatic interpretation. What is
Downloaded from jos.oxfordjournals.org by guest on January 1, 2011
3· F validates Skepticality: in cases of conflicts among defaults where there is no default that has a most specific antecedent relative to the others, F won't verify any conclusions of the defaults. Because of Specificity and Skepticality, CE automatically predicts when knowledge conflict can be resolved and when it can't. 4· F is robust in that if r F ¢ then ¢ will survive as a consequence of the premises r augmented with logically independent information. S· For each deduction A F B there is a corresponding embedded default in the obj ect language (that is, a formula in which one occurrence of the connective > occurs within the scope of another) which links boolean combinations of the formulae A and B, and which is verified to be true. We gloss this embedded default formula as :J (AS, B). So :J (A , B) means A F B. This amounts to a weak deduction theorem. The object language formula :J (A , B) means that A nonmonotonically yields B in the metalanguage.
Alex Lascarides, Ann Copesrake and Ted Briscoe 5 1
• • •
( r, a , {1) > Narration(a, f3) Re; s u l t : (r, a , {1) 1\ cause(ea, ep) > Result( a, f3) E l a b o r a t i o n : (r, a, /3) 1\ Subtype( a, f3) > Elaboration ( a, f3) Na r r a t i o n :
For example, Result states: if f3 is to be attached to a with a rhetorical relation where a is part of the discourse structure r already, and the event described in a caused that described in /3, then normally, the rhetorical relation is Result. In general, axioms in DICE that are for computing rhetorical relations are of this general form-(r, a, {1) 1\ ¢ > R(a, /3)-where ¢ is information about a , f3 and the discourse context r, and R is a particular discourse relation. The way one infers Subtype(a, f3) (glossed as the event condition in a is a subtype of that in f3), or cause(ea, ep) in DICE is explained in detail in Asher & Lascarides (1995a). 3 - 3 A discourse constraint on word interpretation
To explicate how rhetorical structure affects word meaning, we give a detailed analysis of a case which involves the rhetorical relation Elaboration . The rules in Asher & Lascarides ( 1 995a) ensure that Subtype (22a, 22b) holds, because learning to use a paint brush is a subtype oflearning to paint, and the janitor is plausibly a member of school staff
Downloaded from jos.oxfordjournals.org by guest on January 1, 2011
important is that Specializationp, and Specializationc have unrelated antece dents, because one will refer to Rembrandt and the other to janitors. Conse quently, if the rules apply and conflict-i.e. there is some reason why Rembrandt and thejanitor cannot be using different types of brushes-then the conflict will be irresolvable because of Skepticality in CE. DICE also computes the rhetorical relations that connect the DRSS together in SDRT. DRSS for clauses are built in a monotonic fashion via information from syntax, much as in standard DRT. However, in contrast to traditional DRT, one then reasons in DICE about how this DRS attaches to the preceding discourse structure via a rhetorical relation. And unlike the monotonic construction of the DRS representing the clause, the inferences in DICE underlying the choice of rhetorical relation are nonmonotonic, because rhetorical relations are not always linguistically marked but may be inferred on the basis of clues from background knowledge. A further contrast to traditional DRT is that once the semantic representation of the clause currently being processed is built by the grammar, it is evaluated against the model. This is because DICE uses the semantic content of the clause as a clue for inferring the rhetorical relation. In standard DRT, semantic interpretation does not occur until the representation of the whole discourse is built. The default rules Narration, Result and Elaboration in DICE are used for computing rhetorical relations.
52 Ambiguity and Coherence
(22) a. All the school staff learned how to paint. b. Even the janitor learned to use a paint brush. So Specificity in DICE between Narration and Elaboration predicts that Elabora tion (22a, 22b) holds. But inferring this crucially uses the information that the brush in (22b) is a paint brush, since learning to use a (general) brush is not a subtype oflearning to paint. So what about the discourse (23), where there is no explicit stipulation about the brush, and indeed, where knowledge encoded in Specializationc favours cleaning as opposed to paint brushes?
Intuitively, this text is still an elaboration, and we infer that the brush used is a paint brush because of the discourse context. We explain here how this discourse effect on word meaning is encoded formally in DICE, since this discourse effect can also create zeugma. Asher & Lascarides ( 1 995a) provided rules in DICE which encoded lexical sense disambiguation in a discourse context. They argued in favour of a rule they called Lexical Impotence. This rule ensured that one avoids disambigua ting a word according to the clues from within a sentence if it ultimately leads to discourse incoherence. But Lexical Impotence cannot explain the interaction between words and discourse in (23), because if the brush is a cleaning brush (as the clues in the sentence (23b) favour), then a narrative relation can be inferred. Because there is no incoherence, in that a rhetorical connection has been successfully computed, Lexical Impotence will not apply when interpreting (23 )· But intuitively, although (2 3) is coherent when the brush is a cleaning brush, the coherence would be much better if brush is interpreted as referring to a paint brush. We assume that, in general, interpretations of words that give discourses which are only weakly coherent are avoided. This is a generalization of the principle encapsulated in Lexical Impotence, since it will apply to incoherent discourses as the extreme case. So, we want to formalize the principle that we do not make inferences that have an adverse effect on the quality of coherence, even when there are some clues from with the sentences that we should. We can capture this interpretation strategy as a rule in DICE. First, we assume that weak( r) means that the SDRS r is at best only weakly coherent. We will discuss how we might begin to formalise the truth conditions for weak( r) below, in Section 3·4· Let Info(a) and Info(f3) be a gloss for all monotonic information about the constituents a and {3.3 Let KB be the interpreter's knowledge base: this includes the update function ( r, a, {3), the semantic content of r, a and {3, and the rules of DICE. Then the Interpretation Constraint below captures the intuition that
Downloaded from jos.oxfordjournals.org by guest on January 1, 2011
(23) a. All the school stafflearned how to paint. b. The janitor learned to use a brush.
Alex Lascarides, Ann Copesrake and Ted Briscoe 5 3
people try not to infer propositions that lead to weakly coherent or incoherent discourse. It states: if (a) {3 is to be connected to a with a rhetorical relation, and {3 and a are both true, and (b) if the KB that includes not only the update task of {3 to a, but also the information {3 ', nonmonotonically leads to a discourse of only weak coherence or no coherence at all, then normally (c) {3' doesn't hold. •
I n t e rp r e t a t i o n
con s t r a i nt
(a) ( T , a, fJ> 1\ Info(a ) 1\ Info(f3) 1\ (b) :J KB({J', weak( r u {3))) (c) > -.{3'
(24) a. All the school staff learned to paint. b. (And) the janitor learned to use a cleaning brush. Narration has coherence constraints: there must be a distinct common topic, which we will assume in this case is school staff/earn to do something. This is the topic computed by the formal procedure of generalization described by Grover et a/. ( r994), which is adequate for the examples in this paper, although it does not solve the problem of computation of topics in general. But intuitively, although the narrative is found to be coherent, it is only weakly so in this context, since the topic seems strange, as we will discuss in Section 3 ·4· If {3 ' yields a weak discourse, 'J KB({J', weak(ru {3)) holds. S o the Interpretation
Downloaded from jos.oxfordjournals.org by guest on January 1, 2011
This rule captures a dependence between pragmaric inferences from informa tion within the sentence and information from the discourse context. Consider the case when {3' is a pragmatic implicature of {3 in isolation of the discourse context (i.e. Info '({3) > {3 ' is a rule in the KB, where Info '({3) is some subpart of Info(f3)). If assuming {3 ' yields only weak discourse coherence when one attempts to attach the constituents together, then the Interpretation Constraint blocks this pragmatic inference, because its consequent will conflict with that of the default rule Info '({3) > {3', and it is also more specific than this rule. So although {3 ' would follow via DMP from a KB that contained j ust {3, {3' does not follow in the context of T, because the Interpretation Constraint blocks it via Specificity. The Interpretation Constraint models the way in which the discourse (2 3) affects the interpretation of brush . It leads to an inference that the brush in question is a paint brush, thereby overriding the default specified by Specializationc. We will go through this example in detail. Let the DRSS representing the compositional semantic content of(zJa, b) be a and {3 respect ively. Then let {3' be the condition that the brush is a cleaning brush. Specializa tionc can be glossed as Info '({3) > {3 ', and this rule applies. But the condition (b) of the Interpretation Constraint is verified with this {3'. Interpreting the brush as a cleaning brush yields a nonmonotonic inference via DMP in DICE that the rhetorical relation is Na rra tion,just as in (24a, b):
54 Ambiguity and Coherence
Constraint applies, and its consequent is -.fJ '. This conflict with Specializationc, but the Interpretation Constraint is more specific, so fJ is inferred. Thus we infer that the brush is not a cleaning brush. The antecedent of the Interpretation Constraint applies for each fJ ' where the brush is a tooth brush, a hair brush and so on, but does not apply for the fJ ' where the brush i s a paint brush. When this {J' i s added to the KB, then Elaboration(a, fJ) is inferred, j ust as in (22), and the resulting discourse interpretation is perfectly acceptable. So we correctly predict that (2 3 ) is an ela boration, and the brush in question is a paint brush. Thus the Interpretation Constraint extends the Lexical Impotence Rule of Asher & Lascarides (r995a) to allow disambiguation to prefer a particular interpretation if other possibilities lead to a weak but not necessarily incoherent, discourse. In section 4 we will show that the Interpretation Con straint also plays a crucial part in creating zeugma in the absence of multiple lexical entries in the lexicon. -.
'
In the discussion above, we relied on intuition to motivate the assumption that a discourse is weak. Of course, providing formal truth conditions for weak( r) would involve complex reasoning about the semantic content of r and the cognitive states of the participants in the discourse (c£ Grice 1 975), and this is well beyond the scope of this paper. But we can provide a preliminary sketch for the examples we discuss, where it seems that the discourse is weak because the topic is strange. In the example above, for instance, the topic computed when brush was interpreted as cleaning brush was school staff learning to do something. The rule below is intended to capture the intuition that the discourse will be weak if no explanation for this being the topic of conversation can be nonmonotonically deduced from the KB. •
We a k
Di s cou�se
f r om
( J KB (r, y jJ r) 1\ -. J KB (r 1\
S t range
y
Topi cs :
jJ r, 0 1\ Explanation(y jJ r, o))) -> weak( r)
This rule states that if augmenting the interpreter's KB with the SDRS r leads to the nonmonotonic conclusion that y is the topic of r (this is the gloss for the formula y jJ r), and it's not possible to compute from this information and the rest of the interpreter's KB any explanation 0 for why this is the topic, then r is weak. To formalize the computation of the topic of the discourse from the KB-i.e. choosing the y such that J KB ( r, y U r) is true-we can assume the method described by Grover et a/. (1 994). This uses the subsumption relations between entities that are recorded in the KB to compute the generalization of the propositions being connected, and this generalization serves as the topic. Thus
Downloaded from jos.oxfordjournals.org by guest on January 1, 2011
3 ·4 Weak coherence
Alex Lascarides, Ann Copestake and Ted B riscoe 5 5
4 I NT E G R A T I N G T HE C O M P O NE N T S
In the previous section we described a rule that encodes discourse effects on word meaning. We now use it in this section to show how discourse effects create zeugma and the absence of crossed readings, even when there are no MLES. But fi r st, we start with an example that is zeugmatic due to homonymous ambiguity. 4· I Homonyms
Consider sentence ( 1 2) and its logical forms once more.
( 1 2) Some dam busters and bank robbers blew up banks. For the sake ofbrevity, we ignore the collective readings and concentrate on the distributive ones. Two logical forms thus arise because of the two lexical entries for bank: LFground features the predicate ba�ou nd• and LForg features the predicate bankorg· To check the coherence of this sentence, we must resolve this semantic ambiguity.
Downloaded from jos.oxfordjournals.org by guest on January 1, 2011
we will assume that :l KB ( -r, y U -r) holds just in case y i s the generalization of the propositions featured in the SDRS -r . Note that the strangeness of the topic is not absolute; there may be discourses and contexts for which the topic school staff learning to do something would be acceptable. But here we assume that the hearer has no special knowledge about the situation, and that thus there is no apparent explanation for the speaker initially describing the entire schools staff painting, and then weakening the topic by stating that the janitor learned an unrelated activity. The above rule uses the semantics of the rhetorical relation Explanation to define weak discourse. Explanation is defined in Asher ( 1 993a), where he formalizes the proposal of Bromberger ( 1 96?.) that an explanation {3 of a constituent a serves an answer to the question why a? Of course, Weak Dis course from Strange Topics doesn't fully specifY the truth conditional content of weak( -r), since we have said nothing about the content of KB which prevents the interpreter from inferring a reason o for why the topic of the discourse -r is y. But this is a general (and very hard) problem for any formal account of dis course interpretation, and further discussion takes us away from our main con cern in this paper, which is to exploit the logical relationship between the >-rules that encode what to do if -r is weak and the other rules in DICE, rather than providing detailed specifications of exactly when -r is a weak discourse. In essence, we show in this paper that modelling zeugma reduces to this well known open problem, of specifYing the quality of discourse coherence.
s6 Ambiguity and Coherence
4.2 Syntax and pragmatics: a single lexical entry
Now consider an example where there is a single lexical entry, but nevertheless there is zeugma: ( 1 4) c. Rembrandt and our janitor used a brush. Because there is a general sense of brush which subsumes information in the other senses, there is only one logical form for ( I 4c), which features the predicate derived from the vague lexical entry for brush. The interpreter checks the coherence of coordinating the constituents in ( I 4c) together by checking that a rhetorical connection can be computed
Downloaded from jos.oxfordjournals.org by guest on January 1, 2011
To do this, the interpreter assumes that one of the logical forms is true, but not both-i.e. LFground v LForg 1\ -.(LFground 1\ LForg) is part of her KB-and she attempts to draw a nonmonotonic inference via the >-rules in DICE about which logical form is the preferred reading. Two pieces of knowledge are relevant. First, if a bank robber is blowing up an object, which is either a financial institution or a mound of earth, then in the absence of further infor mation we infer that the object in question is a financial institution. Conversely, for dambusters the default is that the object is a mound of earth. These rwo rules can be represented in DICE as >-rules similar to the Specialization rules in Section 3·3· They ensure that although A bank robber blew up a bank and A dam buster blew up a bank are ambiguous sentences at the semantic level, they are dis ambiguated at the pragmatic level. How do these rules affect the pragmatic interpretation of (12)? We are assuming that ( 1 2) is distributive, so the embedding functions that verify the DRSS LFground and LForg permit bank robbers and dam busters to be agents to different events, involving different banks. But even so, neither logical forms permit a mixture ofbanks to be blown up: LFground V LForg 1\ -.(LFagrou nd 1\ LForg) entails that either all the banks involved are financial or all are earth banks. So even though the antecedents of the default rules could be verified by different agents, different banks and different events, the rules are nevertheless dependent. Because the logical forms do not permit bank robbers to blow up financial institutions while the dam busters blow up earth banks, the real world knowledge conflicts. Moreover, it is irresolvable conflict, because the antecedents of the rule are logically unrelated: one is about dam busters and the other about bank robbers. Consequently by Skepticaliry in CE, the interpreter comes to no conclusions about whether bank should be interpreted as bank org or bank grou nd · That is, Skepticaliry produces a zeugmatic effect in ( I 2), because the semantic ambiguity is irresolvable. The interpreter does not begin to check how the coordinated constituents rhetorically link together, because there was no resolution of the truth conditional content of the constituents.
Alex Lascarides, Ann Copesrake and Ted Briscoe 5 7
(25) Rembrandt used a brush. Our janitor used a brush. Using the mechanisms available in SDRT, the interpreter infers in the monotonic component of CE that the relation in ( 1 4c) is Parallel. The task of checking the coherence of the Parallel relation is exactly that for checking the coherence of Parallel in (26) if the interpreter makes the assumption, {J ', that Rembrandt uses a paint brush and the janitor a cleaning brush: (26) ?Rembrandt used a paint brush. Our janitor used a cleaning brush. But although it is logically consistent, (26) is at best a weak discourse, if not incoherent. We assume that this is because Parallel demands a common theme-which in this case could be glossed as people doing something with a brush . In a similar way to the example discussed in Section 3.4, this would be a bad theme, because there is nothing to explain why the speaker intends to talk about people using brushes for very different activities. Context could ameliorate the problem, as in (27), for example. (27) Several artists, two hairdressers and the janitor have been visited by a representative from the ACME wonder brush company. One artist and our janitor now uses ACME brushes. But we assume that ( I 4c) is not uttered in such a context. Thus, assuming fJ ' leads to weak( T u fJ) being true and clause (b) of the Interpretation Constraint is verified. Indeed, the whole antecedent of the Interpretation Constraint is verified. This yields conflict between three default rules in cE: SpecializationP, Specializationc and the Interpretation Constraint. They conflict because the former two rules yield {J' (that Rembrandt uses a paint brush and the janitor a cleaning brush) and the Interpretation Constraint yields -.fJ '. But the conflict between these rules is irresolvable in CE. Hence the interpreter cannot decide which implicature to drop: whether it should be the one that Rembrandt is using a paint brush, or the janitor using a cleaning brush,
Downloaded from jos.oxfordjournals.org by guest on January 1, 2011
between relevant propositions, where these are derived from the syntagmatic representation of the coordinated sentences. The two propositions are formed by joining each property specified in the coordinate structure-in this case, the two generalized quantifiers Rembrandt (A.Q3x(Rembrandt(x) 1\ Q(x))) and our janitor (A.Q3x(janitor(x) 1\ Q(x)))-to the property specified by the rest of the sentence-in this case, the property A.x(x used a brush) given by the VP. So the two propositions to be linked in the analysis of ( 1 4c) correspond to Rembrandt used a brush and our janitor used a brush . Computing the propositions to be connected in this way ensures that we capture the analogy between the coherence of coordination and the coherence of discourse we alluded to earlier. Checking the coherence of ( r 4c) is essentially the same as computing the rhetorical relation between the sentences in (25):
s8 Ambiguity and Coherence
(28) In 1 9 5 0 the typist and the managing director both married teachers. However, our theory correctly predicts that (z8) is not zeugmatic. As in ( r 4c), the relevant constituents must be attached with Parallel in (z8). But even if the managing director marries a female teacher and the typist a male one, the theme (marrying general-i.e. male or female-teachers) is perfectly acceptable. So unlike ( qc), interpreting a and f3 in the way real world knowledge clues would predict does not lead to weak coherence. So the Interpretation Constraint doesn't block the pragmatic implicatures, and DMP on the relevant laws yields the inferences that the managing director married a female teacher, and the typist a male one.4 Sometimes, there is no scope for reinterpreting a constituent in order to improve discourse coherence, because the truth conditional content of the constituents yields weak coherence. Consider (r6b) ( 1 6) b. ?That thesis is orange and unreadable. In this case, using the above method for deriving the proposltlons to be rhetorically connected from the syntagmatic representation of the coordinated sentence, the propositions are that thesis is orange and that thesis is unreadable. The only candidate relation berween them is Parallel, and in order to satisfY its coher-· ence constraints there must be a common theme. As before, we compute this by generalizing on the rwo constituents That thesis is orange and That thesis is unread able. The first constituent predicates the physical aspect of thesis, and the second one predicates its content. Since generalizing over the semantics for an individu ated object and an abstract representation gives something which is basically empty of content, the generalization of the rwo constituents is the thesis has (general) properties. So the common theme constraint is met, but this common theme is very general relative to the semantic content of the constituents being connected, and thus the rhetorical connection is weak in (r6b) and ( r 7b). In this example, the weak rhetorical connection arises when only the
Downloaded from jos.oxfordjournals.org by guest on January 1, 2011
or both. So, although ( qc) is not semantically ambiguous as there is only one · logical form, it is pragmatically ambiguous. This pragmatic ambiguity cannot be resolved and this produces the zeugmatic effect. Ironically although the Interpretation Constraint, in general, improves discourse coherence, here it creates incoherence, because it causes Skepticality in CE in conjunction with other pragmatic knowledge resources. In contrast, in (27), the Interpretation Constraint would not apply: the common theme is acceptable in this discourse context. So the zeugma is ameliorated since there's no irresolvable conflict among the defaults. As in (r 4c), there are real world knowledge clues in (28) that the entities involved probably have different properties (one teacher is male and the other female).
Alex Lascarides, Ann Copestake and Ted Briscoe 59
monotonic or compositional semantics of the constituents are assu �ed. That is, weak( -r v /3) is inferrable from the KB alone. So :l KB(/3 , weak( -r v {3)) doesn't hold for any {3 ' and clause (b) in the antecedent of the Interpretation Constaints doesn't hold. Thus there is no scope for shifting the interpretation of the constituents so that coherence is improved. Now compare ( I 6b) with ( I 6a). '
(I6) a. That thesis has thousands of pages and is unreadable.
4·3 Discourse coherence a nd incremental processing
It is well established that interpretation oflanguage proceeds incrementally (e.g. Marslen-Wilson & Welsh I 978; Frazier I 979; Crain & Steedman I 98 5). Here, we investigate how this affects lexical processing in a discourse context. Consider ( I 4b). (I 4) b. Rembrandt used a brush and so did our janitor. We assume that the interpreter calculates the semantic content and pragmatic implicature of the text as soon as enough information is available to allow an initial decision to be made. In ( I 4b), this means that the interpreter calculates the pragmatic implicatures of Rembrandt used a brush , before parsing and so did ourjanitor.5 So, using a monotonic compositional semantic construction procedure the interpreter builds the DRS a for the first clause Rembrandt used a brush , and a is not only evaluated against the model at this stage, but the pragmatic implicatures of this semantic content are calculated in DICE as well. So brush is specialized to paint brush . Only then does the interpreter begin to parse andso did ourjanitor. Again, the grammar is used to construct the DRS f3 for this clause: z, e', r
(/3)
janitor(z) cto(e', z, c1 ) holds(e', r) . C1 , ?
Downloaded from jos.oxfordjournals.org by guest on January 1, 2011
As before, (I 6a) is coherent only if one can compute a rhetorical connection between the relevant propositions, derived from the syntagmatic representa tion. Just as in ( I6b), the same linguistic clues allow the intepreter to compute a Parallel relation between them, with the same very general common theme. But real world knowledge also allows the interpreter to infer a causal link between the constituents using the default rules specified in Asher & Lascarides ( I 995a), and this in turn yields Result( a, {3). So there is a stronger rhetorical connection between the constituents in (I 6a) than there was in ( I 6b) and thus our theory predicts that ( I 6a) is more acceptable than ( I 6b).
6o Ambiguity and Coherence
( 1 8) John banked the money and then he banked the plane. As before, we calculate the pragmatic interpretation ofJohn banked the money, before parsing the second constituent. Background knowledge allows one to infer in CE that bank in this phrase should be interpreted as bankdeposit (that is, the verb with the financial sense). Now we parse the second constituent. We
Downloaded from jos.oxfordjournals.org by guest on January 1, 2011
The condition c 1 ::::: ? marks the fact that (1 4b) contains an elided clause, which must be resolved via the discourse context. That is, the anaphor c 1 must be resolved, so that we know what the janitor z is doing. Attaching f3 to a with a discourse relation will help resolve it. In this case, coordination constrains the applicable relations to Parallel or Contrast and the use of and selects Parallel. The constraints on Parallel require a partial structural isomorphism (as defined precisely in Asher 1 993a) and a common theme between the linked propositions. Because of the isomorphism constraint, the anaphor c 1 must be resolved to a property which corresponds to the VP information in the first clause, roughly speaking A.x(x use a brush). So the VP ellipsis is resolved as a byproduct of checking the coherence constraints on Parallel, and the elided clause is expanded to our janitor did something, which was use a brush (this is equivalent to ourjanitor used a brush via axioms on the semantics of the auxiliary do). According to our assumption about the process of incremental inter pretation, the interpreters now calculates the pragmatic implicatures of what has been processed so far. Because pragmatic implicature is nonmonotonic in nature, it is important to retract old pragmatic inferences before calculating the new ones in the light of the new information. In this case, the specialization of brush to paint brush that occurred after the first clause is retracted because, just as in ( I 4c), the default rules Specialization1. and Specializationc apply, and if one assumes their consequents then the demands of the coherence constraints on Parallel yields a weak discourse. So the Interpretation Constraint will also apply, with {3 ' being that Rembrandt uses a paint brush and the janitor a cleaning brush. So just as in ( 1 4c), there is irresolvable confl i ct between Specializatio�, Specializationc and the Interpretation Constraint. We cannot decide whether to assume that Rembrandt doesn't use a paint brush, or the janitor doesn't use a cleaning brush, or both. Calculating pragmatic implicature as one processes constituents, rather than doing everything at the end of the discourse, is a risky but necessary strategy, because of limitations of memory, the utility of rapid comprehension, and so forth. Here the risk has not paid off, and reinterpretation has been forced, leading to zeugma. In addition, the effect is more pronounced in ( 1 4b) than it was in ( qc), exactly because reinterpretation as well as irresolvable conflict occurs. Now consider an example involving word repetition.
Alex Lascarides, Ann Copesrake and Ted Briscoe 6 1
must attach it to the first. The similarity in syntactic structure and word forms used prompt the interpreter to infer a Parallel relation (as well as Narration). But if we assume that the senses of the two occurrences of banked in (I 8) are governed by the real world knowledge about banking money and banking planes, then the task of computing a rhetorical relation for (I 8) is essentially identical to that for (29): (29) ?John deposited the money. He turned the plane.
(3o) Rembrandt used a brush and our janitor used a brush. So (3o) and ( I 4c), which are truth conditionally equivalent, are both zeugmatic but for slightly different reasons. (3o) involved reinterpretation where ( 1 4c) did not, but both ultimately yield irresolvable ambiguity. 4·4 The absence ofcrossed readings
In cases where there are MLES, such as forfile and bank, the MLE account of non crossed readings when there is a single occurrence of the word form still stands. (2) John and Bill have each file. (1 3) Texas and Alabama have preservation orders on their most beautiful banks. However, by adding information flow between lexical organization and discourse inference, we can also provide a complementary account of non crossed readings in cases where there is only a single lexical entry, or where there is word repetition. Consider ( I 4d). ( 1 4) d. The grandparents gave the children brushes for Christmas. We have assumed that there is a general lexical entry for brush, which refers to paint brushes, hair brushes, toothbrushes and so on. Because of this, the word brush fails to have a specialized meaning unless something in the context forces one. In ( 1 4d), we do not interpret the sentence specifically as one child getting a paint brush, and the other a floor brush, for example. Rather, the sentence is
Downloaded from jos.oxfordjournals.org by guest on January 1, 2011
At best, (29) is coherent only in some weak sense. Therefore, the Interpretation Constraint applies for ( I 8) when {3 ' stipulates thatJohn bankdeposited the money and ban�urned the plane. The Interpretation Constraint thus blocks both pieces of real world knowledge from applying simultaneously, as it did in ( I 4c). This is because the relevant world knowledge rules have unrelated antecedents, so by Skepticality there is irresolvable conflict. Again, two things contribute to the zeugmatic effect: reinterpretation of the word bank is forced, and there is ultimately irresolvable conflict about how to interpret it. Exactly the same line of reasoning explains the effect in (3o):
62 Ambiguity and Coherence
interpreted as vague or general about what kind of brushes the children received. Here the absence of a crossed reading is epiphenomenal. And with further context, a crossed interpretation can be obtained: ( 1 4) d. The grandparents gave the two children brushes for Christmas. One got a hairbrush and the other a paint brush. Now consider an example where there is word repetition. ( r 9) John had a file and Bill had a file.
(3 r ) John had a dossier. (And) Bill had a tool. So the Interpretation Constraint applies when a crossed reading is assumed, and thus blocks it via DMP. In this way, the interpreter nonmonotonically rules out two of the four logical forms but cannot choose between the remaining two 'non-crossed' logical forms. s
CONCLUSION
We have argued that the view that all cases of zeugma and non-crossed readings are caused by the existence of discrete lexical entries is too simplistic and captures only part of the story. Our complementary account formalizes the pragmatic reasoning underlying the MLE approach to zeugma and non-crossed readings involving homonymous ambiguity, bur also how zeugma and non crossed readings occur in cases involving word repetition and constructional polysemy. Our treatment predicts that crossing readings of words with multiple lexical entries will always result in zeugma. Zeugma can also arise in cases of constructional polysemy but only when the context is such that the discourse would be weak or incoherent if more specific contrasting interpretations are assumed. Thus the test for distinct lexial entries is not the presence ofzeugma alone, but the presence of zeugma even where the discourse would be coherent with contrasting interpretations. For instance, ( ro) is evidence of ambiguity in take, because of the coherence of the discourse even with the different interpretations illustrated by ( roa): (ro) Mr. Pickwick took his hat and his leave. (roa) Mr. Pickwick took his hat and took his leave.
Downloaded from jos.oxfordjournals.org by guest on January 1, 2011
There are four possible logical forms of this sentence, corresponding to the four permutations offile,ool and filedossier· The only candidate rhetorical relation to attach john had a file and Bill had afile is Parallel. If the reading is crossed, then the connection is weak: cf the weak Parallel relation in text (3 r ) :
Alex Lascarides, Ann Copestake and Ted Briscoe 63
ALEX LASCARIDES Centrefor Cognitive Science University ojEdinburgh 2 Buccleuch Place Edinburgh, EHB gLW Scotland, UK e-mail: alex®cogsci.ed.ac.uk ANNE COPESTAKE CSLI
Sta nford University Ventura Hall Stanford, CA 94305 USA e-mail:
[email protected] TED BRISCOE Computer Laboratory Cambridge University Pembroke St. Cambridge CBz 3QG UK e-mail: ejb@ lcl.cam.ac.uk
Received: 20.05.95 Revised version received: 04. 1 2.95
Downloaded from jos.oxfordjournals.org by guest on January 1, 2011
If a word is repeated, crossed readings will not occur if this would result in a weak discourse. Furthermore, if there are conflicting clues in the sentence as to the correct reading, repeated words can produce zeugma. There are other rhetorical effects of word repetition which we have not accounted for, but we assume that these are independent of polysemy. I n all these cases, zeugma is caused by a mixture of lexical organization, grammatical structure, and pragmatics causing Skepticality (or irresolvable conflict) in CE. More specifically, the causes are (a) irresolvable conflict in pragmatic interpretation, created by axioms that link lexical interpretation with the interpreter's knowledge about how to interpret discourse; (b) weak rhetorical connections between the constituents; and (c) reinterpreting incre mental interpretation, which is forced subsequent information in the discourse. These effects can be intermixed, so for instance some examples involve both an effect of reinterpretation and irresolvable conflict. This is a heterogeneous account of zeugma, but there is one important common thread: pragmatic reasoning is central to zeugma, regardless of the presence of MLEs. Our account introduces only one piece of new machinery: the Interpretation Constraint which generalizes the rule Lexical Impotence described by Asher & Lascarides ( r995a). This not only allows us to account for many examples of zeugma but also extends the extent to which discourse coherence constraints can resolve lexical ambiguity.
64 Ambiguity and Coherence
NOTES I
4 We believe that we can provide an account of all the effects attributed by Cruse (I 995) to different degrees of distinctness in readings of words such as brush contrasted to teacher along these lines, without having to make structure distinctions in the lexicon, but we will not discuss this further here. It is not our intention to advance fine grained hypotheses concerning the tem poral organization of human language processing, since these can only be dis tinguished by psycholinguistic experimen tation. Furthermore, the gross (and uncontroversial) assumption that prag matic implicatures are calculated at least on a clause by clause basis is sufficient for our purposes.
REFERENCES Asher, N. ( 1 993a), Reference to Abstract Objects in Discourse, Kluwer Academic Publishers, Dordrecht. Asher, N. (I 993b), 'Extensions for common sense entailment', in IJCAI Workshop on Conditionals (eds), C. Boutilier and J. Del grande, Morgan Kaufmann, Los Altos, CA. Asher, N. & Lascarides, A. ( I 994). 'Intentions and information in discourse', Proceedings ofthe JZnd Annual Meeting ofthe Association of Computational Linguistics (ACL94), Las Cruces, New Mexico. 34-4 1 . Asher, N . & Lascarides, A . ( 1 995a), 'Lexical disambiguation in a discourse context', Journal ojSemantics, 1 2, 69- 1 08. Asher, N. & Lascarides, A. ( 1 995b), 'Metaphor in discourse', Proceedings ojthe AAAl Spring Symposium on The Representation and Acquisition of Lex:ical Knowledge: Polysemy, Ambiguity and Generativity, Stanford, CA, 3-7· Asher, N. & Morreau, M. { 1 99 I ), 'Com-
monsense entailment-a modal theory of non-monotonic reasoning', Proceedings of the 1 2th International joint Conference on Artificial Intelligence (I]CAI-9 1), Sydney, Australia, 3 87-92. Asher, N. & Sablayrolles, P. ( I 995), 'A typo logy and discourse semantics for motion verbs and spatial PPs in French , journal of Semantics, 1 2. Briscoe, E. J., Copestake, A., and Lascarides, A. { 1 995), 'Blocking' in P. St. Dizier & E. Viegas (eds), Computational Lex:ical Seman tics, Cambridge University Press, Cam bridge, 273-302. Bromberger, S. (1 962), 'An approach to expla nation', in Butler, R (ed.), Analytical Philosophy, Oxford University Press, Oxford, 72- 105. Catlin, J-C. & Catlin, J. { I 972), 'Intention ality: a source of ambiguity in English?' Linguistic Inquiry, J. Copestake, A. {1 992), 'The representation of '
Downloaded from jos.oxfordjournals.org by guest on January 1, 2011
Throughout this paper we will use MLE to describe the case where multiple lexical entries exist with distinct semantics. Some theories allow different lexical entries to have the same semantics bur different syntactic properties, but this situation would not give rise to zeugma. 2 Indeed, SpecializationP and Specialization< could be instances of a rule schema-one with Rembrandt substituted in the ante cedent of the scheme, and the other with the janitor. 3 For example, if a is ( I sa), Info ( a) will contain the information that Rembrandt used a general brush, because this is monotonically inferrable from the truth conditional content of a. However, Info(a) will not contain the information that Rembrandt used a paint brush, since this is only nonmonotonically inferrable.
Alex Lascarides, Ann Copestake and Ted Briscoe 6 5
Spring Symposium on the Representation and Acquisition of Lexical Knowledge: Polysemy, Ambiguity and Generativity, Stanford, CA, 75-8 I . Lascarides, A. & Asher, N. { I 99 I ), 'Discourse relations and defeasible knowledge', Proceedings ofthe Z9th Annual Meeting of the Association of Computational Linguistics, Berkeley, CA, 5 5-63. Lascarides, A. & Asher, N. ( I 993), 'Temporal interpretation, discourse relations and commonsense entailment', Linguistics and Philosophy, 16.5, 4 3 7-93 · Lascarides, A. & Copestake, A. ( I 995, forth coming), 'The pragmatics of word meaning', in T. Galloway & M. Simons (ed.), SALT V Proceedings, Cornell Univer sity, Ithaca, NY. Lascarides, A. & Oberlander, J. { I 993). 'Temporal coherence and defensible knowledge', Theoretical Linguistics, 1 9. 1 , I 3 5· Marslen-Wilson, W. & Welsh, A. ( I 978), 'Parsing interactions and lexical access during word recognition in continuous speech', Cognitive Psychology, 10, 29-63. Polanyi, L. { I 98 s). 'A theory of discourse structure and discourse coherence', in W. H. Eilfort, P. D. Kroeberger, and K. L. Peterson (eds), Papers from the General Session at the Twenty-First Regional Meeting of the Chicago Linguistics Society, Chicago, IL. Pustejovsky, J. (I 99 I ), 'The generative lexi con', Computational Linguistics, 17, 409-4 1 . Pustejovsky, J . (I 995), 'Linguistic constraints on type coercion', in P. Sr. Dizier & E. Viegas (ed.), Computational Lexical Seman tics, Cambridge University Press, Cam bridge. Thompson, S. & Mann, W. ( I 987), 'Rhetoric al structure theory: a framework for the analysis of rexts', /PRA Papers in Pragmatics, I, 7 9- I05. Zwicky, A. M. & Sadock, J. M. { I 975). 'Ambiguity rests and how to fail them', in J. P. Kimball (ed.), Syntax and Semantics IV, Academic Press, New York, I -36.
Downloaded from jos.oxfordjournals.org by guest on January 1, 2011
lexical semantic information', doctoral dissertation, University of Sussex, Cog nitive Science Research Paper CSRP 280. Copestake, A. & Briscoe, E. J. ( I 995), 'Semi productive polysemy and sense extension', Journal ofSemantics, 1 2, I 5-67. Crain, s. & Steedman, M. J. s. (I 98 5). 'On nor being led up the garden path: the use of context by the psychological parser', in D. Dowry, L. Kamunen, & A. M. Zwicky (eds), Natural Language Parsing, Cambridge University Press, Cambridge. Cruse, D. A. ( I 986), Lexical Semantics, Cambridge University Press, Cambridge. Cruse, D. A. ( I 995). 'Polysemy and related phenomena from a cognitive linguistic viewpoint' in P. St. Dizier & E. Viegas (eds.), Computational Lexical Semantics, Cambridge University Press, Cambridge, 3 3-49· Deemrer, K. van. ( I 990), 'The ambiguous logic of ambiguity', Proceedings of the First CLIN Meeting, Utrecht, The Netherlands, I 7-32. Frazier, L. (I 979), On Comprehending Sentences: Syntactic Parsing Strategies, Indiana Uni versity Linguistics Club, Bloomington, IN. Grice, H. P. { I 97 5). 'Logic and conversation', in P. Cole & J. L. Morgan (eds), Syntax and Semantics, Vol. 3: Speech Acts, Academic Press, New York. Grover, C., Brew, C., Manandhar, S. & Moens, M. (I 994), 'Priority union and generalization in discourse grammars', Proceedings ofthe Jznd Annual Meeting ofthe Association for Computational Linguistics (ACL94), Las Cruces, New Mexico, I 7-24. Hobbs, J. R. ( I 98 5), 'On the coherence and structure of discourse', Reporr No. CSLI8 5-37. Center for the Study of Language and Information. Kamp, H. & Reyle, U. ( I 993), From Discourse to Lagic, Kluwer Academic Publishers, Dor drechr. Kempson, R. M. ( I 977), Semantic Theory, Cambridge University Press, Cambridge. Lascarides, A. (I 995), 'The pragmatics of word meaning', Proceedings of the AAAI
Journal ofStmantia
I 3 67-86
© Oxford University Press • 996
On No BART G E U R T S
University ofOsnabrock
Abstract
Lexical decomposition is one of the less respectable denizens of the modern semanticist's tool kit. But the term 'decomposition' can be construed in at least two ways. On the one hand, it has a weak construal, which merely implies a commirment to the claim that lexical meanings are definable in terms of a finite, but not necessarily small, inventory of primirives. This commitment is an innocuous one, and is shared by many if not most workers in the field of natural-language semantics. On the other hand, there is a strong construal of 'lexical decomposition', which entails the view that rules of interpretation may be sensitive to the internal structure of a word's meaning, and may operate selectively on only some of its parts. It is this notion which is controversial, and for good reason. For one thing, adopting this strong version of decomposition greatly enhances the predictive force of a semantic theory, and reduces its explanatory power in proportion. For another, there does not seem to be a serious body of evidence to prove that the strong notion of lexical decomposi tion deserves a place in the theory of grammar. Such evidence as is available is restricted to minor areas of the lexicon, and none of it is uncontested. In this paper I want to examine a small part of the quantificational system which would seem to call for a strongly decompositional analysis. Precisely such an analysis has been advocated by all authors on the subject, and it receives support from speakers' intuitions as well as the etymological record. Neverthe less, I shall argue that there is a perfectly viable alternative to the decom positional account, and in fact the explanation that I propose follows directly from some natural, and quite uncontroversial, observations on the nature of quantification.
Downloaded from jos.oxfordjournals.org by guest on January 1, 2011
Since Jacobs ( 1 980) i t has been generally assumed that German kein 'no' requires a decomposi tional analysis. On this analysis, kein means 'not some', which in itself is plausible enough, but furthermore it is claimed chat the negative element of kein may be construed as having scope over an expression chat, in its turn, outscopes the quantil}ring element. ! propose an al ternative to this decompositional theory which explains the same range of data not in terms of scope, but in terms of the kind of objects involved in the interpretation of kein-NPs. Specifically, it is shown that the problematic facts may be accounted for on the assumption that such NPs may refer not only to concrete but also to abstract, or generic, individuals.
68
On No
( I ) a. Ich habe keine Sekretarin. I have no secretary b. Ich rieche keinen Fisch. l smell no fish c. Ich mag keine Feigen. I like no figs In German, the unmarked way of negating a sentence with an indefinite object is by replacing the object with a kein -NP. It is somewhat remarkable that this should be the unmarked option because German nicht 'not' is otherwise quite versatile. For example, although in its normal function as a predicate negation nicht immediately precedes the non-finite verb cluster, if one is present, as in (2a), it may also be prefixed to a definite NP, as in (2b), or a prepositional phrase, as in (2c): (2) a. Dieses Jahr wird der Weihnachtsmann nicht kommen. this year will Santa Claus not come b. Nicht der Kellner hat die Erdbeeren gegessen. not the waiter has the strawberries eaten c. Nicht ohne Widerwille gestand cr. not without reluctance confessed he Occasionally, nicht does turn up in front of an indefinite NP, but such uses are highly marked and must always be interpreted contrastively:
(3) a. Sie hat nicht Bananen gegessen, sondern Mel6nen. b.
she has not bananas eaten but melons Sie hat nicht drei Melonen gegessen, sondern vier. she has not three melons eaten but four
Downloaded from jos.oxfordjournals.org by guest on January 1, 2011
In the following discussion I concentrate my attention on German data, because the facts that I am concerned with show themselves more clearly in German than in standard English. But to some extent the same phenomenon may be observed in English and, presumably, many other languages as well. The counterpart in German to the English determiner no is kein , but kein is more prevalent than no, and is often employed where speakers of English would prefer to use predicate negation. NPs with kein are therefore less restricted in their distribution than their English cousins. They quite freely occur as objects, for example, and it is with this use that l shall be mainly concerned in this paper; some other uses will be considered along the way. Intuitively as well as etymologically, kein is an indefinite determiner which incorporates a negation, its unnegated counterpart being ein 'a'. Like the other determiners, kein inflects for gender, number, and case, but since they are irrelevant to our present purposes, these various endings will simply be ignored.
Barr Geurts 69
It appears, therefore, that there is a negative principle which forbids the juxtaposition of a non-contrastive nicht and an indefinite NP. The examples in (4) confirm this diagnosis.
(4b)-(4d) represent three different ways of negating (4a), and carry more or less the same meaning, namely The umpire is not an idiot'. (4c) is allowed because we have replaced nicht with the more emphatic keineswegs. In (4d) the predicate nominal has been divorced from the negative element via topicalization, as a result of which nicht can be used. These observations reinforce the suspicion that kein is semantically equivalent to nicht + indefinite NP, which combina tion is prevented from materializing as such by some rather superficial quirk of the language. With these preliminaries out of the way, let me now turn to the main subject of this paper, which is exemplified by the following sentence (from Jacobs 1 980: 1 2 5):
( s ) Aile Arzte haben kein Auto. all doctors have no car This sentence has at least two readings, one of which is unproblematic: 1 (6) [all x: doctor x]([no y: car y](x owns y)) On the 'logical' reading represented by (6), ( s ) means that no doctor owns a car. Note that this construal can be represented with the quantifier no. This is not possible with the second construal, which is the problematic ond (7) - [all x: doctor x]([some y: car y](x owns y)) This says that at least some doctors don't own a car. The reading in (7) is not hard to obtain, but the concomitant intonation pattern is marked, stressing aile and, optionally, kein ; as a rule, the 'normal' declarative intonation contour would accompany the reading in (6). In this respect, there is no difference between (s) and (8):
(8 ) Aile Gaste waren nicht betrunken. all guests were not drunk
Downloaded from jos.oxfordjournals.org by guest on January 1, 2011
(4) a. Der Schiedsrichter ist ein Idiot. the umpire is an idiot b. Der Schiedsrichter ist kein Idiot. the umpire is no idiot c. Der Schiedsrichter ist keineswegs ein Idiot. the umpire is not-at-all an idiot d. Ein Idiot ist der Schiedsrichter nicht. an idiot is the umpire not
70
On No
(9) a. Ich suche keine Putzfrau. I seek no cleaning woman b. An diesem Grenziibergang muss man keinen Pass vorzeigen. at this checkpoint must one no passport show By default, both sentences will receive a reading which in all relevant respects is analogous to the one in (7). That is to say, (9a) will be taken to mean that the speaker is not looking for a cleaning woman, while (9b) will convey that it is not required that a passport is produced. In the latter case, this preference correlates with the general tendency of the negation to outscope the modal rniissen 'must'. For example, (r oa) is preferably interpreted as indicated by the bracketing in . ( r ob) rather than (r oc), although the latter reading is available in principle. ( r o) a. Du musst nicht angeln gehen. you must not fish go b. You (must not] go fishing. c. You must (not go fishing]. Thus far I have focused my attention on NPs with kein . Are there other expressions that exhibit the same behaviour? My impression is that, strictly speaking, the answer must be that the matter is not quite clear. Like English, German sports several lexical entries which fuse a negation and an indefinite: niemand 'nobody', nicllls 'nothing', nie 'never', nirgendwo 'nowhere', etc. The first two items on this list, in particular, have been claimed to behave like kein , but for various reasons the evidence is inconclusive. For example, Lerner & Sternefeld (1 984: r 87, r 89) contend that the following favour a split reading:
Downloaded from jos.oxfordjournals.org by guest on January 1, 2011
Generally speaking, this would be a somewhat circumstantial way ofconveying that none of the guests were drunk. But with the appropriate intonation contour, the nor-all reading becomes available.3 On its not-all reading, (8) is construed with nicht having wide scope. This means that, on this reading, there is a mild discrepancy between the sentence's structure and its intended interpretation, but such tensions are nothing out of the ordinary. However, if(s) is to be read as in (7), this tension is exacerbated by the fact that, apparently, the interpretation of kein must be split into two halves, one of which takes scope over the universal quantifier, whereas the other remains in situ . The purpose of this paper is to show that this split construal can be accounted for in a principled fashion, without resorting to lexical decompo sition in the strong sense of the word. Split construals of kein not only occur in admittedly unusual constructions like (s). The following are perfectly normal sentences (which are taken from Jacobs 1 99 1 : 594), and both strongly prefer a split reading-as a matter offacr, in both cases a split reading is the only sensible option:
Bart Geurts 7 1
( I 1 ) a. Ich sehe niemanden kommen. I see nobody come b. Ich darf mit niemandem dariiber reden. I may with nobody about-that speak
( 1 2) Beiden Arzten hat Luise nichts vermacht. to-both doctors has Luise nothing bequeathed Jacobs maintains that, like ( s ), this sentence allows both a reading on which nichts has narrow scope and one on which the negation that it incorporates takes wide scope, while an existential quantifier remains within the scope of heiden Arzten . However, my own findings indicate that native speakers' judge ments don't match Jacobs'. Only some of my informants were able to get a split reading for this sentence at all, and none of them obtained it without having been prompted. In contrast, the reading on which nichts takes narrow scope is unproblematic. To sum up, while speakers' j udgements about ( 1 2) are insecure at best, it is intuitively plausible but hard to prove that the negative NPs in (I I ) have split readings. For these reasons, I will in the first instance confine my attention to split readings with kein , but I will return to examples like (I I a) and ( I I b) in due course. Examples like (5) and (9) are widely held to imply that kein , or at least certain occurrences of kein , should receive a decompositional analysis in the strong sense of the word. This position is implicit in many traditional grammars, which typically issue a warning to the effect that what appears to be an instance of constituent negation may in fact turn out to be just a special form of sentence negation; thus, for example, Behaghel (1 924: 6 5 f£). The first one to explicitly argue in favour of a decompositional analysis (Bech I 9 S S being a notable fore runner) was Jacobs ( I 98o), and his position has since been adopted by Lerner & Sternefeld ( I 984) and Kiss ( I 993), among others; Dahl (I 993) somewhat tentatively suggests that a strongly decompositional analysis is required for the Swedish equivalent of no.
Downloaded from jos.oxfordjournals.org by guest on January 1, 2011
They suggest that ( 1 I a) means 'It is not the case that I see somebody coming', and ( I I b), 'It is not the case that I am allowed to discuss this matter with any body.' On the other hand, if we construe niemanden as having wide scope we get: 'Nobody is such that I see him coming', and 'Nobody is such that I am allowed to discuss this matter with him'. The problem is, of course, that in the first case it makes no semantic difference at all whether we give wide scope to a negative sentence operator or to nobody, while in the second case the difference is so slight that, intuitively speaking, it is negligible. Problems of a rather different nature beset Jacobs' ( I98o: qo) contention that a split construal is available for the following sentence:
72 On No
It is irrelevant to my purposes what the exact details of a decompositional analysis of kein should look like. Jacobs himself employs Montague's quantifying-in method to give the negation wide scope, but obviously the same effect could be obtained with a raising transformation, a modified form of Cooper storage, or, perhaps, with some version of the type-shifting method proposed by Hendriks ( 1 993). The differences between these approaches are considerable, but they don't matter here, and in the following I adopt the transformational metaphor merely for ease of exposition. Formulated in transformational terms, the decompositional analysis comes down to this. We assume that, at some suitable level of representation, the following analysis is associated with ( s ): Of course, this is equivalent to (6), but in this representation kein has been decomposed into a negation operator and an existential quantifier. Further more, we assume that a raising transformation may lift the negation our of its embedded position and associate it with the structure as a whole, thus giving it wide scope. This transformation applies optionally, so either the logical form in ( 1 3) remains as it is, or it is mapped into (7). Thus the two readings of ( s ) are accounted for. Obviously, the same story applies to the pair of examples in (9), and it will also work for cases like (8). This is a strongly decompositional analysis: it is assumed that the meaning of kein is complex structure, part of which may be affected by the raising transfor mation. For reasons indicated at the beginning of this paper, I take it that, ceteris paribus, an alternative account that didn't require this assumption would be more attractive. But are there alternatives to the strongly decompositional analysis of kein ? One option, discussed and criticized by Jacobs ( 1 980), is to let the semantics of kein take care of the split reading. For example, it is possible to translate kein in such a way that the interpretation of kein Auto comes out as follows: ( 1 4) A.JUT(- T( A.x [some y: car y](x R y))) Here R is variable of the type corresponding to the category of transitive verbs, and T is a variable of the type corresponding to NP. (I 4) will combine with the interpretations of lzaben and aile Arzte to produce the reading in (7). So we have set up the interpretation of kein in such a way that it controls the interpretation of the whole sentence, thus enabling it to deposit a negation sign to the left and an existential quantifier to the right of the subject NP's denotation. This strategy is problematic for a number of reasons. First, as Jacobs observes, it implies that kein is multiply ambiguous. For example, whereas in the example given T must be of the NP type, it would have to be of a different type in order to be able to deal with (9b).jacobs apparently thinks (I 98o: I 3 2 f., 1 99 1 : 594) that
Downloaded from jos.oxfordjournals.org by guest on January 1, 2011
( r 3) [all x: doctor x](-[some y: car y](x owns y))
Barr Geurts 7 3
this is enough to discredit the account, but it is not, provided that the ambiguity is systematic. If it is, a type-shifting mechanism might be called up to generate the required readings in a principled manner (see e.g. Hendriks 1 993). A more serious problem, in my view, is the assumption that kein should be ambiguous between a reading which keeps together the negation and the existential quantifier and one which does not. Finally, it should be noted that this account will only work if the type of R in (14) is allowed to vary, too (and concomitant changes are made in the remainder of the meaning of kein ). Thus far we have only seen examples of kein -NPs which occurred as direct objects, but the following may have a split reading, too:
In order to account for this example, R would have to be of the type corresponding to dittansitive verbs. If we put together this point with the previous rwo it should be obvious that this lexicalist approach, even if it can be made to work, is blatantly ad hoc. A more serious alternative to the decomposition analysis might be to assume that, although syntactically it functions as an ordinary article, semantically speaking kein simply means 'not'. This implies that, on one reading at least, kein is not a quantifier and doesn't contain a quantifier either. Consequently, since existential quantification is not part of the meaning of kein, the quantifying force of an NP with kein must come from a different source. This is defensible when such an NP is plural, as the following examples illustrate:
( 1 6) a. Alle Professoren haben Knallfrosche gekauft. all professors have firecrackers bought b. Alle Professoren haben keineswegs Knallfrosche gekauft. c. Alle Professoren haben keine Knallfrosche gekauft. In German, as in English, a bare plural may be construed as containing an inaudible existential quantifier, and in (1 6a) this quantifier lies in the scope of aile Professoren : ( 1 7) (all x: professor x]((some y: firecracker y](x bought y)) In ( 1 6b), this sentence is negated simply by means of the adverbial particle keineswegs 'not at all', which is an emphatic form of nicht (c( (4c)). This sentence has rwo readings: one in which the negation operator is within the scope of the universal quantifier, and one in which it is the other way round.
( I 8) a. [all x: professor x](-(some y: firecracker y](x bought y)) b. -[all x: professor x]((some y: firecracker y](x bought y))
Downloaded from jos.oxfordjournals.org by guest on January 1, 2011
(1 s ) Du musst den Brief keinem Polizisten zeigen. you must the letter to-no policeman show
74 On No
( 1 9)
Ich suche nichts. I seek nothing b. Ich suche niemanden. I seek nobody
a.
( 1 9a) may be an answer to 'What are you looking for?' and (1 9b) to 'Are you looking for an electrician?' and both of these questions may be argued to be ambiguous between a de re and a de dicto interpretation. Moreover, it is remarkable that a similar distinction is not observed with neg-incorporating adverbials like nie 'never' or nirgendwo 'nowhere'. This contrast suggests, too, that there is some truth in the claim that sentences like the above are ambiguous, although the ambiguity is difficult to pin down precisely. To sum up, our first stab at a non-decompositional analysis of negative NPs leaves open one important question and entails that only kein -NPs may have split readings. It is possible that this hypothesis may turn out to be tenable after all, but there is another explanation which I find more promising, because it allows us to maintain that all negative NPs are perfectly ordinary quantifier expressions. It is to this alternative that we now turn. My second, and definitive, proposal covers all sorts of indefinites with
Downloaded from jos.oxfordjournals.org by guest on January 1, 2011
According to the proposal under consideration, exactly the same account applies to ( 16c), at least semantically speaking. The only difference between ( 16b) and ( 1 6c) is a syntactic one: whereas in the former case the negation associates with an adverbial element, in the latter it is represented by an article. Thus what we have dubbed the 'split' reading of (1 6c), i.e. (1 8b), does not involve a split in any way. One potential problem which this analysis faces is that it isn't clear whether it can be extended to singular kein -NPs without too much strain. Plural (or mass) nouns may stand on their own and don't require an article in order to acquire existential import. Singular count nouns, in contrast, hardly ever occur without an article, so we cannot simply assume that in the singular case the existential quantifier will somehow be taken care o£ I am not convinced that this problem is serious enough to dismiss the proposal out of hand, for it might well be argued, I believe, that the article ein 'a' is semantically inane, which would support our hypothesis about kein . But a serious discussion of this matter would take us too far afield. Besides, there is a further problem which I find more worrymg. It is clear that there is no way in which the proposed analysis of kein -NPs could be extended to lexical NPs such as nichts 'nothing' and niemand 'nobody'. Now I have argued above that the evidence on which it has been assumed that these items allow for split readings, too, is not quite convincing, but it can hardly be denied that there is something to these arguments. Consider, e.g., the following examples:
Ban Geuns 7 5
(2o) a. Every featherless bird is now extinct. b. No reptiles are indigenous to the Philippines. c. Many mechanical devices were invented by mistake. In these sentences, it is clear that the quantificational domain of every, many, and no must consist of kinds rather than concrete individuals. Concrete individuals may die but not become extinct; it is species not concrete individuals which may or may not be indigenous to the Philippines; and an invention brings into life a new type of thing not, or not necessarily, a concrete individual. Carlson uses the term 'abstract individual' to refer to whatever it is that these sentences quantify over, and I shall adopt this usage, too. Carlson's examples suffice to prove that abstract individuals must be accepted as a fact of life, but in a way these examples are misleading, too, because they might be mistaken for evidence that quantification over abstract individuals is something special. After all, be extinct, be indigenous to, and invent aren't ordinary predicates. I believe that this impression is simply false: quantification over abstract individuals is rife, but we tend not to notice it for two reasons. First, as a rule abstract talk is not formally distinguishable from concrete talk. Secondly, the statements we make about abstract individuals often entail analogous statements about concrete individuals, and conversely. The following examples illustrate these observations. I point at a copy of Lolita and volunteer the following: (2 1 ) I have read this book.
Downloaded from jos.oxfordjournals.org by guest on January 1, 2011
incorporated negations. According to this proposal these NPs are normal quantifying expressions.4 However, it is well known that the domain of a quan tifier is not always a set of concrete individuals, and whenever quantified expressions like kein N, nichts, niemand, etc. range over something other than concrete individuals, split readings may result. The main attraction of this position is that it requires no special assumptions to account for split readings (such as: lexical decomposition, an exotic semantics which will wrap itself around practically anything, or a mismatch between the syntax and semantics of neg-incorporating items). Split readings are simply an unexpected consequence of the standard system of quantification. In general one conceives of quantificational domains as sets of concrete individuals, but there is nothing in the notion of a quantifier that requires this assumption. A quantifier may range over any set-i.e. any collection of things that can be counted (and some quantifying expressions even seem to apply to non-countables as well). Concrete individuals are our countable entities par excellence, but they are not the only ones to be recognized as such by the metaphysics of natural language. The following examples, which I have taken from Carlson ( 1 977: 438), illustrate this point
76 On No
In uttering this sentence, I may have meant at least two things: that I have read Lolita or that I have read the particular copy of Lolita at which I am pointing. The former reading is entailed by the latter, but not vice versa. If I had the second meaning in mind, I have claimed to stand in the 'have read' relation to a concrete individual, and by implication, to the abstract individual which is instantiated by this particular copy of Lolita . If I had the first meaning in mind, then my claim was merely that I have read the book, not that I have read this particular instance of it. The same type of polysemy may arise when we quan tify over books. In an advertisement, a bookshop boasts: (22) We have more than 1 0,000 books in stock.
(2 3) a. A car stopped in front of the house. b. Thejury was impressed most by a French car-namely, the Citroen ZX. c. A car is a vehicle.
Downloaded from jos.oxfordjournals.org by guest on January 1, 2011
This may mean either that they have more than 1 0,000 copies or more than 1 0,ooo rides, and the sentence's truth may depend on what one counts. In this example, too, one reading entails the other but the two are not equivalent. Abstract individuals are useful to us because they may be, and generally are, instantiated by concrete individuals. But abstract individuals may also, in their turn, instantiate other, more general, abstract individuals: this concrete book instantiates Lolita, which in turn instantiates 'novel by Nabokov'. In general, if we predicate something of an abstract individual, claims about its instances, abstract or concrete, are implied, and conversely. Such implications are systematic, but they depend on the predicates employed. If someone has read a concrete copy of Lolita , then he has read Lolita as well as Nabokov. Coming from the opposite direction, if he has read Nabokov, he has read at least one of Lolita , Ada, Pnin , etc., and at least one copy of either Pnin, Ada, Lolita, and so on. But if he dislikes a particular copy of Lolita , he may actually be fond of 'the book'. Our first specimen of a sentence with a split reading (reproduced below as (26)) was about cars, which is a happy coincidence because the structure of this particular domain is a fairly transparent one. It should be uncontroversial to assume that the abstract individual CAR is instantiated by the models Peugeot 205, 305, . . ., Citroen AX, BX . . ., and so on, which in turn are instantiated by concrete cars. Admittedly, the domain is amenable to a more elaborate classification, but this amount of detail will suffice for our purposes. I shall use the letter M to denote the set of car models. So we have a three-layered hierarchy with the abstract individual CAR at the top, concrete individual cars at the base, and the members of M in the middle, and the same expression may be employed to pick out any of these levels, as the following examples illustrate:
Bart Geurts 77
(23a) is most likely to be used to convey information about a concrete individual; (23b) is about car models, i.e. it picks out one of the members of the set M; and (23c) is about the generic car, i.e. about CAR. Now consider the following sentence: (24) Leo owns a car.
(25) a. Leo doesn't own a car. b. - [some x: car x](Leo owns x) c. [some x: car x](-(Leo owns x)) In (25a) the negation may take scope over the indefinite or conversely, the corresponding logical forms being (2 5b) and (2 5c), respectively. In both representations, the variable x may range either over the set of concrete cars, the set M of car models, or the singleton set {cARj.5 If the negation has wide scope, it doesn't matter which of these sets we choose, but if it has narrow scope, it may make a difference what the indefinite quantifies over. I fin (25c) x ranges over M or the set of concrete cars, we get two equally unlikely readings: 'There is at least one x E M such that Leo is not an x--owner', and: 'There is at least one concrete car that Leo doesn't own.' Either reading may be forced with some help from
Downloaded from jos.oxfordjournals.org by guest on January 1, 2011
Let us suppose that in this example, car may refer to the top element of the car hierarchy, i.e. to CAR, to some x E M, or to a concrete individual. Consequently, (24) may mean that Leo is a car owner, that is an x-owner, x E M, or that there is a concrete car which he owns. Now somebody utters this sentence, and we ask ourselves which interpretation the speaker has in mind. The answer is fairly obvious: it doesn't matter in the least which particular meaning prompted the speaker to say what he said. For instance, if Leo owns a concrete car, then he is a car owner and for at least one x E M, he is an x-owner-and similarly if we take one of the other construals of (24) to be primary. Given the meaning of own and the structure of the domain under consideration, each 'reading' of (24) entails the others. Ownership certainly is a relation between concrete individuals. But if someone owns a given, concrete, object x he is ipsofacto related to any abstract individual instantiated by x, and this relation is also denoted by the verb own. Actually, it is not even necessary to assume that this verb is polysemous between various 'levels of ownership'. Suppose that the relation of instantiation is a partial order on a given domain of individuals, which may be either concrete or abstract. So technically speaking each individual instantiates itself Then we might say that the verb own denotes the relation which is periphrastically expressed by 'own a concrete instance of', from which it would already follow that it doesn't matter whether we construe the indefinite in (24) as being about abstract or concrete cars. Almost the same observations apply if we negate (24):
78 On No
(26) Alle Arzte haben kein Auto. all doctors have no car Assuming that kein is an ordinary quantifier, we would expect to get (at least) two different logical forms for this sentence: (27) a. [all x: doctor x]([no y: car y](x owns y)) b. [no y: car y]([all x: doctor x](x owns y)) Any of the standard methods for scope assignment will produce these two forms. But what do they mean? That question can only be answered after we have fixed the domains of the variables x and y. Since we are only interested in the second variable, let us agree that x must take its value from a contextually given set of (concrete) doctors. As in the previous examples, y may range either over a set of concrete individuals, or over the set M, or over the singleton set consisting of the abstract individual CAR. Now if we construe (26) as (27a), it doesn't matter from which of these three sets y must pick its values. In this respect the example is exactly like (2 sa) on its (2 sb) reading. If for each x there is no concrete car that he owns, then no x is either a y-owner, for any y E M, or a car owner. And so on. However, if we construe (26) as (27b), it does make a difference whether y ranges over a set of concrete cars, the set M, or the set {cAR). Suppose that y ranges over concrete cars. Then we obtain a reading which is weird and should probably be excluded on pragmatic grounds as being too unlikely. If y ranges over M, we get an interpretation which says that for no y E M, all doctors are y-owners. Hence it is not the case that all doctors are Peugeot 205 owners, or that all doctors are Citroen XM owners, etc.
Downloaded from jos.oxfordjournals.org by guest on January 1, 2011
the context or, if that doesn't work, by choosing a different predicate. If, on the other hand, x ranges over {cAR), (2 sa) is construed as 'Leo is not a car owner'-or in other words: (2 sb) and (2 5c) become truth-conditionally equivalent. These observations illustrate an important point. The notion that natural language quantifiers may, and often do, range over abstract individuals is by no means new. But the example in (2 5) suggests that, once we take this fact seriously, it becomes clear that the notion of scope should perhaps be wielded with more care than it has sometimes received. For we have seen that, if we construe x as ranging over the set {cAR), it doesn't matter whether we give the negation wide or narrow scope. Furthermore, the reading on which the quantifier ranges over {cAR} and has wide scope is truth-conditionally equivalent to the reading on which it has narrow scope and ranges over concrete individuals. Such facts urge us to exert some caution in our judgements on scope relations, but it is also in facts like these that, in my view, the key to the riddle of kein lies. Let me now return to the main theme of this paper and to Jacobs' example (26) (� ( s )).
Barr Geurts 79
Apparently, this reading is not readily available in this particular case, but anal ogous examples can be constructed where precisely such a reading is intended. For example, the outcome of the German Ambulance of the Year Contest might be reported as follows: (28)" Mehr als 1 2 Arzte stimmten fur kein Auto. more than 1 2 doctors voted for no car
(29) - [all x: doctor x]([some y: car y](x owns y)) So the split reading of {26) is accounted for despite the fact we have analysed kein as an ordinary quantifier. If we want to make explicit the chain of reasoning running from (27b) to (29), two premisses must be brought out that have hitherto remained implicit. {As a matter of fact, I believe that there are various ways to secure this inference, but it is sufficient for our present purposes if at least one plausible account can be given.) First, we must rule out the possibility that (27b) is made true because the restrictor of no is empty. and therefore we must assume that the abstract individual we have referred to as CAR actually exists, and is in the extension of the predicate car. Secondly, as I have indicated already, we have to make an assumption about the property A y(x owns y), with x standing for an arbitrary doctor. We want this property (and many others) to be projecting in the sense that, if it applies to a concrete individual b, it also applies to all abstract individuals a that are instantiated by b. In short, using � for the instantiation relation and v and u as variables over concrete and abstract individuals, respect ively: (3o) P is a projecting property iff [all u: [some v: v � u](Pv)](Pu) To summarize, we need two premisses: (i) that the predicate ca r applies to the abstract individual CAR, and (ii) that for every doctor, x, A y(x owns y) is a projecting property. Now since y in (27b) ranges over [car), we obtain: (3 1 ) - [all x: doctor x](x owns cAR) The split reading of (26) follows from (3 1 ) and our second premiss: if(32) were true, (32) [all x: doctor x]([some z: car z & z � cAR](x owns z))
Downloaded from jos.oxfordjournals.org by guest on January 1, 2011
In the context indicated, this would very likely be interpreted with kein Auto having wide scope and ranging over M or some subset thereo£ Finally, the variable y in (27b) may range over the set [cAR). Thus construed, (26) claims that for no y E [cAR), it is true that all doctors are y-owners. Which is to say that not all doctors are car owners. Given what we observed earlier this is equivalent to (29) (-(7)), where both x and y range over concrete individuals.
8o
On
No
should be interpreted in terms of a relation between Julius and the abstract individual CAR. Thus construed, we can give the quantifier corresponding to a car wide scope:
Downloaded from jos.oxfordjournals.org by guest on January 1, 2011
(where z ranges over concrete individuals), then (3 1 ) would have to be false. But since it isn't, the negation of (32) must be true, and thus we have obtained the split reading of (26). This analysis can easily be extended so as to deal with Jacobs' example (3 3a) (=(9b)): (3 3) a. An diesem Grenziibergang muss man keinen Pass vorzeigen. at this checkpoint must one no passport show b. o [no x: passport x](x is shown) c. [no x: passport x] o(x is shown) Ignoring the locative PP, which is irrelevant to our purposes, (33b) and (3 3c) are the two logical forms associated with (33a); they parallel (27a) and (27b), respectively. In these representations, the box operator stands for deontic necessity. (33 b) represents one possible set of readings of (33a), but there is no need to discuss these in detail. (3 3c) allows for at least two readings, depending on whether x ranges over concrete passports or over the singleton set inhabited by the abstract individual PASSPORT. The former reading is pragmatically unlikely, but the latter is not. It says, in effect, that at this checkpoint is not (deontically) necessary to be a passport producer-more idiomatically: it is not required that one shows a passport. If we want to account for this in more detail, one additional premiss is needed, which is due to the fact that this is not an extensional context. The premiss is that not only does the abstract individual PASSPORT lie in the extension of the predicate passport, but in addition PASSPORT exists in all worlds that the interpretation of ' o ' needs to have access to, and it is always in the extension of passport (since PASSPORT is an abstract individual, this is not the same as assuming that there are (concrete) passports in all pertinent possible worlds). The second premiss, as in the previous example, is that A. x (x is shown) is a projecting property. With these two premisses, (34a) is derivable from (3 3c), and from (34a), (34b) follows, which is in effect the split reading that we wanted to obtain (as before, z ranges over concrete individuals): (34) a. -D (PASSPORT is shown) b. -o [some z: passport z & z � PASSPORT] (z is shown) Finally, I want to indicate how this proposal might deal with the interaction between kein-NPs and attitude contexts. Given our enriched ontology which comprises abstract as well as concrete individuals, it is plausible to assume that on its so-called de dicto reading, a sentence like (3 s ) Julius is looking for a car.
Barr Geurts 8 1
(36) [some x: car x]Oulius is looking for x) where x ranges over {cAR). But the same structure may be employed to represent the de re reading, as well as an intermediate reading, which it is not generally recognized in this connection, but which is available, as the following thought experiment shows. Suppose that, having uttered (3 5), the speaker is asked whether Julius has any particular car in mind. Here are some of the answers that he might give:
(37a) implies that any car will do, (37b) that he is looking for a particular model, and (37c) that he is looking for a concrete car. The first answer indicates a de dicto reading, the third one a de re reading, and the second answer suggests a meaning that lies somewhere in between. In terms of quantification it is perfectly obvious what this means: in a sense all three readings are de re construals, but the objects that are quantified over are different in each case. Now let us look again atJacobs' example (38a) (-(9a)): (38) a. Ich suche keine Putzfrau. I seek no cleaning woman. b. [no x: cleaning woman x](I seek x) The logical form of (3 8a) that we are interested in is (38b). Supposing that the quantifier may range either over concrete individuals or the abstract individual CLEANING WOMAN, this sentence will have two readings: the former means that there is no concrete cleaning woman that the speaker is looking for, the latter that he is not a cleaning-woman seeker. Which is the reading that we wanted to account for. I have outlined a proposal which allows us to maintain the position that German kein -NPs are unambiguous quantifying expressions, which occasion ally produce interpretative effects that may be unexpected but are not, in fact, something out of the ordinary. Before I proceed to argue that kein-NPs are not normal quantifiers, after all, I want to briefly return to other expressions that have been claimed to give rise to similar effects. We have seen that these effects are somewhat difficult to pin down exactly, and I believe that the analysis I have proposed may help us to see why this should be so. An example like (39a) (=(1 9a)) would be represented as in (39b): (39) a. Ich suche nichts. I seek nothing b. [no x: A x](I seek x)
Downloaded from jos.oxfordjournals.org by guest on January 1, 2011
(37) a. No. b. Yes, a Citroen XM. c. Yes, his neighbour's.
82 On No
(4o) Kein dodo ist ausgestorben. no dodo is extinct Be extinct is a predicate that only applies to abstract individuals (cf (2o) above), and therefore we would be led to expect that at least one of the interpretations of (4o) is that the dodo is not extinct. But this is not what we find: the only possible reading of this sentence is that, among the various species of genus dodo, none is extinct. These problems are caused, I believe, by the assumption that NPs are uniformly construed as generalized quantifiers. This assumption is untenable for quite independent reasons, and if we trade it in for a more refined account ofNP interpretation, the problems mentioned in the previous paragraph dissolve auto matically. If all NPs are treated as generalized quantifiers, three classes ofNPs are lumped together that exhibit clear differences? Adopting Lohner's ( 1 987) terminology, I propose to distinguish between definite, quantificational (in a narrower sense than I have used the term thus far), and indefinite NPs. These classes differ most strikingly with respect to their presuppositional properties, although there are other differences as well. First, definite NPs are presupposi tional expressions tout court: an NP of the form the N triggers the presupposition that there is an N, and it may be argued that this is all there is to the semantics of the definite article. Secondly, quantificational NPs are presupposition triggers, too, but they involve further, non-presuppositional, elements as well. For example, (4 1 a) presupposes that there is a (contextually given) set of elephants, and it asserts that a majority of them were drunk:
Downloaded from jos.oxfordjournals.org by guest on January 1, 2011
Here A stands for a restriction on the values of x which is provided by the context (and which may be empty, of course). We observed earlier that, although intuitively there is some difference between the 'de re' and the 'de dicto' reading of (39a), this difference is hard to make explicit. But meanwhile we have examined several examples which display the same characteristics, and our analysis has given a general explanation of what is going on in cases like this. Like the examples discussed in the foregoing, (39a) exhibits what is sometimes called a primitive ambiguity (e.g. Horn 1 989): the variable x may range either over abstract or concrete individuals, and the interpretation which results in the former case entails the one that results in the latter case but not vice versa. The keystone of my proposal is that abstract individuals are involved in the construal of kein -NPs with split readings.6 However, the actual implementa tion of this idea as presented in the foregoing requires a premiss which is problematic, namely that in split readings, kein N ranges over a singleton set containing just one abstract individual. Not only is this assumption intuitively implausible, it also causes problems with examples like the following:
Bart Geurts 8 3
(41 ) a. Most elephants were drunk. b. Perhaps most elephants were drunk. c. If there were elephants present, then most elephants were drunk.
(42) Walter has two children. However, with the appropriate intonation contour, two children may be used to trigger the presupposition that there is a contextually given set of children. (I suspect that this presupposition isn't triggered by the indefinite itself, and that, accordingly, the indefinite determiner is neither semantically nor prag matically ambiguous, but I will not defend that position here.) For example, (43) Two children are playing in the garden. This sentence has two interpretations. If the indefinite is read without any presuppositions, (43) simply asserts that there are two children playing in the garden. But the sentence may also be construed as presupposing, in addition, that the two children in the garden belong to a contextually given set of individuals. The latter reading is preferred in this particular case, notwithstand ing the fact that indefinite NPs are primarily non-presuppositional expressions, because the indefinite occurs in subject position, and in the vast majority of cases subjects are presupposition triggers (cf. e.g. Prince I 98 I ). Formulated in terms of a discourse semantic theory like DRT (Kamp 1 98 I ; Kamp & Reyle 1 993; Geurts 1995), these observations come down to the following. First, the semantic correlate of a definite NP is a discourse entity which is presupposed, i.e. taken as given. Secondly, the semantic correlate of a quantificational NP is a pair of such entities, one of which is presupposed. For example, the intepretation of(4 ra) involves two collections of elephants, one of which is presupposed and contains the other. Thirdly, an indefinite NP simply introduces a new discourse entity, which is not presupposed, bur which is sometimes signalled to be part of a collection that is presupposed.
Downloaded from jos.oxfordjournals.org by guest on January 1, 2011
Obviously, (4 I a) entails that there were elephants. That this is a presupposi tional inference appears from the fact that it displays the projection behaviour that is characteristic of presuppositions, as (4 I b) and (4 I c) illustrate. In (4 I b), sentence (4r a) is embedded in a non-entailed position, but none the less the matrix sentence licences the inference that there were elphants. In (4 I c), by contrast, this inference is 'blocked', which is precisely what we would expect from a presupposition (see e.g. Geurts I 995 for further discussion). All proper quantifiers behave like this. Thirdly, indefinite NPs may trigger presuppositions, bur their primary use is non-presuppositional. Thus, if a speaker volunteers (42a), he obviously does not presuppose that there are children in anything like the way in which (4 I a) presupposes that there are elephants:
84 On No
(44) a. Ein Dodo ist nicht ausgestorben. a dodo is not extinct b. Ein Dodo ist ein Saugetier. a dodo is a mammal The second factor is that, as we have seen, subject terms in general strongly favour a presuppositional reading. Therefore, the hearer will want to interpret the subject NP of (4o) as meaning, in effect, 'none of a given collection of dodos', and since the predicate only applies to generic entities, this will have to be a collection of generic dodos-or species of dodo. Predictably, the only way to make sense of (44a) must be in the same vein: this sentence can only mean that one of a presupposed collection of dodo species is not extinct. What we end up with, then, is a decompositional analysis in the weak sense of the word. Kein is parsed, in effect, as the semantic collocation of the negative operator and the singular indefinite. But mine is not a strongly decompositional account, because it doesn't require the assumption that certain rules of interpretation are sensitive to the internal structures of the word, and may operate on one of its semantic parts to the exclusion of others.
Downloaded from jos.oxfordjournals.org by guest on January 1, 2011
How does kein (or no, for that matter) fit into this picture? Having originated from the fusion of ein 'a' with a negative element, kein evidently is an indefinite determiner, and therefore it introduces a new discourse entity, albeit within the scope of a negation operator, which may but need not be part of a presupposed collection. Thus construed, we would expect kein to pattern with ein , to a significant extent at least, and that is what we find (as will be illustrated in the following). It remains to be shown that the problems we encountered with the initial version of our theory dissolve when we adopt the analysis of kein that I have outlined in the preceding paragraphs. To begin with, we no longer require the somewhat awkward premiss that in split readings, kein N quantifies over a set that happens to contain just one abstract individual. For, just like the singular indefinite (and unlike an NP with, say, most or some), kein N introduces an indi vidual discourse entity, which may or may not be abstract. Furthermore, it can now be seen that there are two factors that conspire to enforce the reading we observed for (4o). First, in German as in English, generic predicates that only apply to kinds, like be extinct, do not combine with the singular indefinite, although the singular indefinite may be used generically, as (44b) illustrates. That is, (4o) cannot be used to convey that the dodo isn't extinct for the same reason that (44a) cannot convey this message:
Bart Geurts 8 S
Aclrnowledgements
I should like to thank Peter Bosch, Martin Muller, Marc Ronthaler, Wolf Thiimmel, and an anonymous referee for the Journal ofSemantics for their comments on an earlier version of this paper. BART GEURTS Universitiil Osnabrock
Received: 27.1 0.94 Revised version received: 07.03.95
FB7
49069
Osnabruck Germany e-maiI:geurts® hal.cl-ki. uni-osnabrueck.de
I The intended meaning of this notation will not be hard to fathom, but lest any mis understandings should arise, I give here the standard definitions of the quantifiers all, some, and no, which are the only ones that I shall be needing. Let x be a variable that is free in cp and lJ!; then: [all x: cp [(lJ!) is true iff (x: cp }