JOURNAL OF SEMANTICS AN INTERNATIONAL jOURNAL FOR THE INTERDISCIPLINARY STUDY OF THE SEMANTICS OF NATURAL LANGUAGE PETER BoscH (IBMScienrific Centre, Heidelberg and Umversiry of Osnabriick) REVIEW EDITOR: T1BOR K1ss (IBMScientific Centre, Heidelberg}
MANAGING EDITOR:
EDITORIAL BOARD: N. ASHER (Universiry of Texas, Ausrin) R. BARTSCH (Universiry of Amsterdam) J. VANBENTHEM (Universiry of Amsterdam) M. BIERWISCH (MPG and Humboldt University Berlin} B. BocuRAEV (Apple Computer Inc) M. BORILLO (University of Toulouse) G. BROWN (University of Cambridge} 0. DAHL (Universiry ofStockholm) S. C. GARROD (Universiry of Glasgow) B. GEURTS (Universiry of Osnabriick) M. HERWEG (IBMScienrific Centre, Heidelberg} L. R HoRN (Yale Universiry) P. N. J oHNSON-LAIRD (Princeton University) H. KAMP (Universiry ofStuttgart) S. LEVINSON (MPI Nijmegen) S. L6BNER (Universiry of Dusseldorf}
SIR J oHN LYONS (Univcrsiry of Cambridge) A. MANASTER-RAMER (WayneStare Universiry) W. MARSLEN-WILSON (MRC, Cambndge) J. McCAWLEY (Universiry of Ch�eago) M. MOENS (Universiry of Edinburgh) F. J. PELLETIER (Universiry of Alberta) M. PINKAL (University ofSaarbriicken) R. A. VAN DER SANDT (Univers�ry of Nijmegen) T. SANFORD (University of Glasgow) R. ScHA (University ofAmsterdam} H.ScHNELLE (Universiry ofBochum) A. VON S TECHOW (University ofTiibingen) M.STEEDMAN (University of Pennsylvania) W. W AHLSTE R (DFKI,Saarbnicken) B. WEBBER (University of Pennsylvania) H. ZEEVAT (Universiry of Amsterdam) T. E. ZIMMERMANN (Umversiry ofSrurrgart)
EDITORIAL ADDRESS: Journal ofSemanrics, c/ o Dr P.Bosch, lBM GermanyScientific Centre, Vangerowsrr. 18, D-69115 Heidelberg, Germany. Phone: (49"-6221-) 59-4251/4483.Telefax: (49"-6221-) 59-3200. Email:
[email protected],
[email protected] New Subscribers to the Journal of Semanrics should apply ro rhe Journals Subscriprion Department, Oxford University Press, Walton Street, Oxford, OX2 6DP. For further informacion see the inside back cover. Volumes 1-6 are available fromSwers and Zeirlinger, POBox 830,216 0 SZ Lisse, The Netherlands. ©Oxford Universiry Press All rights reserved; no parr of this publication may be reproduced, scored in a retrieval system, or transmitted in any form or by any means, electronic, mechanical, photocopying, recording, or orherwtse Without either the prior written permission of the Publishers, or a licence permitting resrncred copymg 1ssued in the UK by the Copyright Licensing
1 9HE, or in the USA by the Copyright Clearance Center, 27
Agency Led, 90 Torrenham Court Road, London W P Congress Street, Salem, Mass o1970.
The Journal of Semantics is published quarterly in February, April, August and November by Oxford University Press Subscripuon is ; I 33·
$119 per year. Second class postage pa1d at Newark NJ and at addmonal mailmg offices. ISSN 0167-
POSTMASTER: send address corrections to The Journal of Semantics, c/o Virgin Mailing and Distribution, Cargo Atlantic, 10 Camptown Road, Irvington NJ 07111-IIOS. USA.
JOURNAL OF SEMANTICS Volume
II
Number 3
SPECIAL ISSUE: SEMANTICS AND PSYCHOLOGY PART I: THE SEMAN TICS OF QUANTIFICATION Guesr Ed1rors: Amhony J. Sanford and Linda M. Moxey
CONTENTS
LINDA M. MoxEY AND ANTHONY J. SANFORD EDITORIAL INTRODUCTION
I
49
ANTHONY J. SANFORD, LINDA M. MOXEY AND KEVIN PATERSON Psychological Studies of Quantifiers
153
ROBERT J . JARVELLA AND LITA LUNDQUIST Scales in the Interpretation of Words, Sentences and Texts
I7
DAviD A. RoUTH On Representations of Quantifiers
I99
STEPHEN E. NEWSTEAD Do Verbs Act as Implicit Quantifiers?
2 15
I
Journal of Logic, Language and Information Editor Peter Gardenfors Section Editors D. Gabbay, Daniele Mundlcl, Stanley Peters, R. Stalnaker, Johan F.A.K. van Benthem This is the official publication of the European Foun dation for Logtc. Language, and Information. The scope of the JOurnal is the logical and computa tional foundations of natural, formal. and program ming languages. as well as the different forms of human and mechanized Inference. It covers the logical, linguistic, and information-theoretic parts of the cognitive sc1ences Examples of main subareas are Intensional Logics including Dynamic Logic; Nonmonotonic Logic and Belief Revision; Constructtve Logics: Complexity Issues in Logic and Ungutstics; Theoretical Prob lems of Logic Programming and Resolution; Cat egorial Grammar and Type Theory: Generalized Quantification: lnformation-Ortented Theories of Semantic Structure like Situation Semantics. Dis course Representation Theory. and Dynamtc Se mantics: Connectionist Models of Logical and Lin gutstic Structures. The emphasts will be on the theoretical aspects or these areas. The purpose of the JOUrnal is to act as a forum for researchers Interested in the theoretical founda tions of the above subJeCts and their interdisci pltnary connections. with an emphasis on general ideas increasing coherence. It is hoped that the JOurnal will meet a need of an tnterdisciplinary com muntty within the cognitiVe sciences whose most tmportant publications have been scattered across JOurnals in different areas so far. The journal IS abstracted andlor mdexed In SOCiofogrcaJ Ab· stracts: Zentralblatt fur Mathematik; tnformabOn Tochnology and the Law· INSPEC
ISSN 0925-8531 Subscription Information 1994, Volume 3 (4 issues) Subscription rate. Oft. 353.00/US$184.50 incl. postage and handlmg
Spectat rate lor members of the European Foundatton tor LogiC. Language and /nlormatton available F01 mon11nformatton. please wnts to the FoLLI Bureau. FWI. UfiiV91'�Jty ofAmsterdam, Pfantage Mutdergraeht 24, t018 TVAmsterdam, The Netherlands PO 8oJt 322. 3300 AH Dordrochl, The Netherlands PO Box 358 Accord Station. Hongham, MA 02018-{)358, USA
1111111111111111
11"'
Journal
Highlight
KWWER ACADEMIC PUBLISHERS
*
Natural Language and Linguistic Theory
Editor Joan Mating Associate Ed1tors Ellen Broselow, Carol J. Georgopoulos, Michael Kenstowlcz. Alec Marantz, Frederick J. Newmeyer
Natural Language and Linguistic Theory provides a
forum for the discussion of theoret1cal research that pays close attention to natural language data. so as to prov1de a channel of communicauon between re searchers of a vanety of pornts of v1ew The JOurnal actively seeks to bridge the gap between descnpt1ve work and work of a highly theoretical, less empirical ly oriented nature. In attempting to strike th1s bal ance, a pnmary goal of the JOUrnal is to encourage work wh1ch makes complex language data acces Sible to those unfamiliar with the language area bemg studied, and work which makes complex theo retical positions more accessible to those working outside the theoretical framework under review.
.
ll
lllll lllllllllll
lllllll lll
lll'"''
Journal
Highlight
- theoretically oriented work on the syntax. seman tics. phonology and the lexicon of natural lan guage; - mterdisciplinary contributions wntten with theo· relical llngUJsls in m1nd; - surveys of recent theoretical developments which facthtate accessibility for a graduate student read ership - reactions/replies to recent papers. - brief notes, comments or observations concerning languages or theoretical po1nts; - letters to the ed1tors; - book reviews of important linguistiCS titles; - spec1al topic issues. The JOurnal is structured to mclude;
Natural Language ana Ungwsl/C Theory Is surveyed tJy MLA lntemat•onal B•bflography. Lingu•siiC & Language BeflaVIOt Ab srraC1S. Current Q>ntenrs/Arts & Humaml•es. CCISoc•al & 88haVJoral Scumces. ASCA Arts & Humamt•es C•ratlon /nd(l)(. Socml SCienCliS C•tB11on /ndtnc, LtngUist•c Abstracts. Lingu•stJc B•bllography IBZIIBR. Computer & lnforrrw10 t n SystemiS Ab stracts Journal Socsolog,ca/ Abstracts. C8mbfldge sc,enr,flc • Technology and the Law AbstraeU. Rttsesrch Alttrt. Informat()(!
Subscription Information ISSN 0167-806x 1994. Volume 12 (4 issues) SubscnpiiOn rate: Dfl. 383.00/US$200.00 incl. post Private rate. 011.156.00/US$59.00 incl. postage and age and handling
handling P.O Bole 322. 3300 AH Oordrechl. The Netnerlanos P.O Boll 358. Accord Stauon, H•ngham. MA 02018-0358. USA
KWWER ACADEMIC PUBLISHERS
Comp utational
Linguistics
Compmational Linguisrirs �� rhc foremost joum�l devoted exclusively to compuurlooal analyses of n;arural language.
lr encompasses AI research in language, linguisrics, and rhe psychology of language processi ng and perfOrmance.
Julia Hirschberg
Wirh each 1ssue prov1ding applied and
Editor
rheo�tical papers. book reviews. technical correspondence, and leners to the editor, the JOurnal prcsenTh a
• YC2d) ln<�innion•I!Urc: S8HIO. AclJ
somulaung forum for the exchange of
Sl� (I() post>gt •ncl hmdl1ngou"'dc
1deas and uc:nds in compmanonal
U.\.A. C.....dun• odd ..Wiuon.l 7�o
language.
G�T Pr
ComputlltionalLinguistiN is the
ch
h
55
H•yw•rd �,.....
JOurnal serving a rapidly expanding
C.mbroc!P'. MA021·12·1399
interdisciplin;uy field and should be:
TEl" 617·BH889 fAX: 1>17-lSII-677')
part of evc:ry research library collecuon.
JOUnul...,rdcrWlm•
R«ommcnd CL to your librarb.n today.
lnclan.J.W wb>tmbmhip rn the .Mooci.ltion li>c C"A>mpu,.ri
Walht (AQ)
Puhfi.h
Bdlcore. MitE W79
Aloocnoon Jill Compuuuorul t.n.,mua
445 South S.-r. So. 1910 Morr""o"'T1·
JSSN ll891·2017. Volu""' 10 fonbcomiJ>&.
NJ 07'>60·1'110 U.S.A.
Linguistics and Philosophy A Journal of Natural Language Syntax, Semantics, Logic, Pragmatics, and Processing Edttor-in-Chlef Gregory N. Carlson
Lingutsttcs
and Philosophy IS a journal for studtes focused on natural language. and is of rnteresr ro practtltoners in the diSCiplines covered '" the t1Ue. Al though the field thus descnbed is so extens1ve that a complete listing of relevant topics is precluded, at least the following specifically fall Wlthtn II: - tradtonal areas In the phtlosophy of language such as meaning, truth, reference, descnption. entail ment. speech acts - traditional areas of linguistics such as syntax. semantics and pragmaucs (when the studies are of sufficient explicitness and generality to be also of philosophical interest) - aspects of artificial intelligence concerned w1th language such as computatJonal linguistics and natural language processing - systems or logic with strong connections to natural language: modal logic, tense logtc, eptstemic logic, mtenstonal logic - philosophical quesllons raJsed by linguisttcs as a sc1ence: hngu1st1cs methodology, the status of lin guistics theories, the nature of linguistic universals - phllosophtcally intereshng problems at the inter section of hngutsrics and other disciplines: lan guage acquis1tion. language and perception, language as a soc1al convention. LlrlgUIStoCS and PtriiOSJOphy as abtlrac!ed/tndexod In Bll/14ttm S•g nalet,que, Commumry OeveloQtrlent ADs tracts, CAJrrent C<>nt�ntsl Arts & HurMnlllfiS:Ants & HU/MMJit$ Cltai•OII lncJOII,I.JngU•SIIC$ & LanQu/lge BMav1or Abstracts; L8/Jguage Teachmg & LlflgviStiCS. MLA lntemattOnltl B•CIIOfJrBPhy. The Ph•losopher's lnd8J( Refire· ttvnyt Zhurnal. 5ocioiogfC4f Abstracts. 18ZIIBR Lingu,stfC$ AC· stracts Lmgu•stlcs BibiiOfJraphy, ZentraiO/att fur Mathemat•lc, RevUfl Phi/OM>ph,q1J8 de LO
ISSN 0165-Q157 Subscription Information 1994, Volume 17 (6 1ssues) Subscnpt1on rate: Ofi.448.00/USS233.50 incl. post
age and hand/mg
Pnvate rate: Dfl. t84.00/USS80.00 tncf. postage and handling PO. Boll 32:2, 3300 AH Doc'dtec;hl. The Netherlands PO Box 358, Accofd Stahon, Htngham. MA 02011Hl358, USA.
"" IIIIIIIIIIIIIJIIIIIIIII'I"II" '' Journal
Highlight
KLUWER ACADEMIC PUBLISHERS
Scope of this Journal
The JOURNAL OF SEMANTICS publishes articles, notes, discussions, and book reviews 111 the area of narural language semancics. lr is explicitly incerdisciplinary, in that it aims at an integration of philosophical, psychological, and linguistic semantics as well as semannc work done in artificial intelligence and anthropology. Concributions must be of good quahry {to be judged by at least rwo referees) and should relate to questions of comprehension and interpretation of sentences or texts in narural language. The editors welcome not only papers that cross traditional discipline boundaries, but also more specialized concributions, provided they are accessible to and inceresting for a wider readership. Empirical relevance and formal correctness are paramount among the cnteria of acceptance for publication. Information for Authors
Papers for publication should be submitted in 3 copies to the managing editor. They should be ryped on A4 {or similar format), one-sided, double-spaced, and with a wide margin and must be accompanied by an approx. 200 word summary. Notes and bibliographical references must appear at the end of the rypescript. All bibliographical references in the text by author's surname and year of publication. Diagrams must be submitted camera-ready. All papers submitted are subject to anonymous refereeing and are considered for publication only on the assumption that the paper has neither as a whole or in part already been published elsewhere nor has elsewhere been submitted or accepted for publication. Unless special arrangements have been made, copyright rests with Oxford Universiry Press. Authors receive 20 offprints of their published articles and 1 0 offprints of their published reviews, free of charge. Larger numbers can be supplied at cost price by advance arrangemenc. Subscriptions
The Journal of Semantics is published quarterly. Institutional: UK and Europe £62.oo; USA and Rest of World US$1 1 9.00. (Single issues: UK and Europe £1 8.oo; USA and Rest of World US$ 3 5.oo.) Personal:• UK and Europe £3 1 .00; USA and Rest of World US$59.00. (Single issue: UK and Europe £9.00; USA and Rest of World US$1 7.oo.) 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 the USA and Canada by Airfreight or in Japan, Australia, New Zealand and India by Air Speeded Post. Airmail rates are available on request. •
Order Information
Payment is required with all orders and subscriptions are accepted and entered by the volume. Payment may be made by cheque or Eurocheque (made payable to Oxford Universiry Press), National Girobank (account 500 1 056), Credit cards {Access, Visa, American Express, Diners Club), or UNESCO coupons. Please send orders and requests for sample copies to theJournals Subscriptions Department, Oxford Universiry Press, Walton Street, Oxford, OX2 6DP, UK. Telex 837330 OXPRES. Advertising
Advertisements are welcome and rates will be quoted on request. Enquiries should be addressed to Jane Parker, Advertising Sales Manager, Oxford Universiry Press, Walton Street, Oxford OX2 6DP, UK. Tel: +44 (o)865 56767 Fax: +44 (o)865 267782.
jollmal ojSrmantrcs 11: q9-151
© N.I.S. Foundanon (199�)
Editorial Introduction: Psychology and Semantics
Downloaded from jos.oxfordjournals.org by guest on January 1, 2011
This issue is devoted to papers of an essentially psychological nature. Psychology relates to semantics in a number of different ways, illustrated by the papers in this issue. Our intention in this introduction is to highlight just some of those ways. The approaches described below are also associated with particular methodological emphases and styles of presentation. First, and at the simplest level, the empirical data which psychology provides serve to put into proper perspective the intuitions which might serve to start a semantic analysis. The paper by Newstead is a good example. Statements which have the form of generalizations, such as Women are interested in clothes , are not usually understood as meaning that all women are interested in (all) clothes . Rather, it is understood that some proportion of women, less than 100 per cent, are interested. Newstead's paper illustrates some of the issues which have a bearing on what it is that determines the proportions which have to hold for the statement to be considered true and appropriate. Similarly, while it is recognized that the proportions of things associated with expressions like many depend upon context, only by experiment can an understanding of the extent and nature of such context effects be determined (Sanford, Maxey & Paterson's paper). Among the more technically refined methods which psychologists have employed to investigate meaning are those of similarity matrix analysis and multidimensional scaling. Routh's paper applies these techniques to the analysis of quantifier meaning. Essentially, the issue is one of determining which properties of quantifiers are utilized in different situations. For instance, if quantified statements are encountered and later a subject tries to recall them, what properties determine the pattern oferrors which will occur? If subjects are asked to judge the similarities of a set of quantifiers, how many dimensions will be required to specify the similarity space? These questions are important for instance, in the design of expert systems, expert data containing quantifiers, frequency expressions, and probability terms (e.g. possible, probable) have been treated as though these expressions differed only in terms of magnitudes on a single dimension (e.g. Fox 1 986; see also Sanford et al. this volume). Routh's results lead to a formal rejection of this assumption. At a somewhat less empirical level a psychological approach to meaning holds that the basis of natural language semantics will best be understood in terms of human communication and human experience. Perhaps the best known approach of this type is LakofPs ( 1 987) treatment of Cognitive Semantics, in which meaning is supposed to be rooted in human experience (see
1 so Editorial Introduction
Downloaded from jos.oxfordjournals.org by guest on January 1, 2011
also Johnson 1 987). The paper by Gibbs et a!. is an example of empirical investigation of an aspect of the cognitive semantic hypothesis. In his paper on the meanings of the term stand, Gibbs utilizes similarity data in order to provide a test of the idea that the various meanings of the term are dependent upon particular recurring bodily experiences. One aspect of the rooting of meaning in experience is seen in the paper of Coventry et a/. on spatial prepositions. Traditionally, the meanings of spatial prepositions have been expressed in terms of objects in Euclidean space. Covenrry et a/. show how space itself should perhaps be defined in terms of functional constraints based on the way in which humans see things as interacting. The experiments show how choice of expression depends upon a number of subtle functional constraints. Also concerned with spatial relational meaning is the paper of Sinha et a/. This exemplifies the value of investigating the acquisition of meaning (especially when different languages are compared). With such a global problem as the psychological roots of meaning, it is difficult to perform direct tests of simple theories, so work in this area has more open experimentation, perhaps akin to 'natural history'. Experimental work here is investigating a large-scale pervasive framework. Psychologists are concerned with meaning as usage, and this angle is exemplified in the papers of Sanford et a/. on quantifiers, and Jarvella & Lundquist on scalarity. According to Sanford et a/., quantifiers are mental oper ators which control those parts of background knowledge and expectation that are brought to bear in interpreting quantifi e d sentences. Similarly, but within the framework of Anscombre & DuCrot (e.g. 1983),Jarvella & Lundquist show how scale polarity is associated with a variety of expressions, and how this influ ences the interpretation of ambiguous anaphoric references. One outcome of these approaches is that some expressions, like focusing particles and quantifi ers, may be thought of as mental operators, operating on memory and control ling attention and patterns of inference, which is a variety of procedural semantics. A final way in which psychology might have an impact on semantics is through the relation of formalisms to natural language semantics. Formal systems serve to explicate necessary mathematical relations between language obj ects. These relations are usually clearest with artificial languages, bur one great hope is that with a suitable system the semantics of natural language(s) may be adequately defined, and seen in perspective. A very successful example of this is the application of Generalized Quantifier Theory to NL quantifiers. Bur that very example serves to illustrate one way in which psychology and more formal approaches to semantics meet. To claim that a quantifier fits a linguistic text which is an analogue of the test of a logical property often requires an intuitive leap: many quantifiers are only intuitively related to the theoretical idealizations. For example, decisions about monotonicity must
Anthony J. Sanford and Linda M. Moxey I 5 I
LINDA M. MOXEY ANTHONY J. SANFORD
RE F ERENCES Anscombre, J.-C. & Ducrot, 0. ( I 9 83), L'A rgumentation dans Ia langue, Mardaga, Liege. Fox,J. (I 986), 'Three arguments for extending the framework of probability', Proceedings of the Workshop on AI and Decision Making , AAAI, Los Angeles, I 985, Elsevier/North Holland, Amsterdam.
Johnson, M. ( I 987), The Body in the Mind, University of Chicago Press, Chicago. Lakoff, G. ( I 987), Women, Fire and Dangerous Things , University of Chicago Press, Chicago.
Downloaded from jos.oxfordjournals.org by guest on January 1, 2011
often be made inmitively and not analytically. From a psychological perspec tive, inruitions are the outcome of processing, so the question becomes one of what kinds of processing underlies them. Sanford et a!. suggest that some of the focusing properties of quantifiers might be the basis of monotonicity judgements. In conclusion, it is often thought that a major role for psychology is to put to the test theories of meaning which derive from linguistic or semantic analysis. By proceduralizing a piece of theory, it can be used to make predictions about processing difficulty or speed, for example. Although 'testing psychological reality' is one potential role of psychology, we suggest that the major use of empirical psychology in semantics is to make clear facts about language in use which formalizations should then accommodate. Ultimately, the question is which will be the simpler explanation of namral meaning, a formalism founded on intuition, or a (formalized) theory founded on language processing as part of human cognition. The papers are presented in two groups, the first having to do with scalarity and quantification, and the second with space and cognitive semantics.
Downloaded from jos.oxfordjournals.org by guest on January 1, 2011
Journal oJSemantrcs 10: 153-170
© N.I.S. Foundation (1<>94)
Psychological Srudies of Quantifiers A N TH O N Y J. S A N FO R D, L IN D A M. MOXE Y, and KEVIN PA TE R S O N University ofGlasgow
Abstract
I
INTR O D U C T I ON
In everyday communication, determiners such as many and Jew, and adverbs such as often and frequently, are commonplace. They also have a central role in formal treatments of semantics, and are both explicitly treated within the formal theory of Generalized Quantifiers (Barwise & Cooper I98 I; Keenan & Stavi I986; Westersthal I989; De Swart I99 I). The use of quantifiers in communication has not been explored in great detail, much analysis being restricted to scalar implicature in the pragmatics literature (Hirschberg I985; Horn I984; Levinson 1983 ). Until relatively recently, the psychological literature on quantifiers was relatively restricted, falling into two main areas. The first is how standard (logical) quantifiers, typically analysed in terms of3 and V, are used in reasoning, and this has been largely restricted to syllogistic reasoning. This work is not reviewed here, but see, for example, Johnson-Laird ( I983) (also Johnson-Laird & Byrne 1991, for an attack on more general problems including double quantification). The second area has been concerned with how nonstandard (nonlogical) quantifiers such as Jew and many map on to scales depicting
Downloaded from jos.oxfordjournals.org by guest on January 1, 2011
In this paper we present a summary review of recent psychological srudies which make a contribution to an understanding of how quantifiers are used. Until relatively recently, the contribution which psychology has made has been somewhat restricted. For example, the approach which has enjoyed the greatest popularity in psychology is explaining quantifiers as expressions which have fuzzy or vague projections on to mental scales of amount. Following Moxey & Sanford ( 1 993a), this view is questioned. Experimental work is summarized showing that quantifiers may be differentiated in terms of the patterns of focus which they produce, which we take as a reflection of the patterns of inference which they induce. Other work suggests that when a speaker uses certain quantifiers it is possible for a listener to draw inferences about what the speaker's prior expectations were, including what the speaker is taken to have believed the listener to expect. These findings are discussed in relation to how quantifiers are selected, and in terrns of a possible psychological basis for certain logico linguistic judgements about quantifiers.
154 Psychological Studies of Quantifiers
2
Q U ANT I F I ER S AND N U MBER S
Given that Many Bricksidefootballfans went to the match , we can ask how many fans that is, or what proportion of fans that is. What are the truth conditions for many expressed in these terms? Variants of this question, that of quantifier-to number mapping, have dominated experimental psychology approaches to natural language quantifiers (NLQs). Such an approach makes sense when it is realized that standard questionnaires often use quantifiers and frequency adverbs to save the respondents from having to use numbers. This leads to problems of how to select a set of quantifiers which is optimally discriminating over a frequency scale (i.e. is an equal interval scale). Of course, people can assign numbers to quantifiers when asked to do so, and within any given context it is possible to produce a set of expressions which roughly form an equal interval scale (Bass, Cascio, & O'Connor 1974; Schriesheim & Schriesheim 1978; Pohl 1981). Early in this research it became obvious that some of the variability in numbers allocated to a quantifier derives from the context in which the quantifiers is used (see Pepper 198 I for a thorough review to that date). As Par ducci ( I 968) pointed out, a student thinking about contraceptive failure would regard s% of the time as often , while missing s% oflectures might be considered to be hardly ever. Pepper & Pryrulak ( 1974) asked subjects to use a scale from I100 to define the meanings of five frequency expressions, which were presented
Downloaded from jos.oxfordjournals.org by guest on January 1, 2011
numbers or proportions. The psychological literature associated with the second approach has done relatively little to further our understanding of how such quantifiers are used, when it is better to choose one expression rather than another, and so on. In the present paper, we draw on work by Maxey & Sanford (e.g. I993a), review some of the standard psychological literature on non standard quantifiers, and then summarize some recent research which blends psychological data with considerations that spring from linguistics and logic. It is appropriate to point out that there is no discussion of the resolution of scope problems with multiply-quantified sentences in this paper, although the issue is of considerable interest to both linguists (e.g. Cooper I983; Ioup I975; Reinhart I983) and psycholinguists (e.g. Carlson & Tannenhaus 1989; Mitcham et a/. 1980 ). Most of the psychological work in this area can be thought of as being concerned with problem-solving rather than natural language under standing in the context of reading or dialogue (e.g. Johnson-Laird & Byrne I99 I), or as encouraging a problem-solving approach on the part of the subjects (Mitcham et a!. 1980). For an approach which is closer to what might happen in normal text comprehension, see Kurtzman & MacDonald ( 1993). The approach taken here is restricted to some properties of single quantifiers.
Anthony J. Sanford, Linda M. Moxey, and Kevin Paterson 1 SS
Downloaded from jos.oxfordjournals.org by guest on January 1, 2011
in six contexts, each of which had a different baseline expectation. (For instance, a shooting in a Hollywood western is high frequency; a sizeable earthquake in California was classed as low frequency.) The higher the baserate, the higher the numbers assigned to the quantifiers, especially to those quantifiers which tend to denote larger amounts (see also Wallsten, Fillenbaum, & Cox 1986). There has been somewhat less investigation of contextual effects on quantifiers of amount than of frequency, although the same intuitive arguments hold and have been given by logicians and linguists (e.g. Altham 1 97 1; Keenan & Stavi 1 986). Maxey & Sanford ( 1993b) asked subjects to indicate what percentage a quantified statement denoted for ten quantifiers in three contexts, where baserate was manipulated in the contexts. The higher the baserate expectation, the higher the numbers assigned, but the effect was restricted to the higher-ranking quantifiers. This study is unique in that a given subject contributed only one reading, so that contrast effects did not influence the results. In a later study, Sanford, Maxey, & Grant (submitted) used ten quantifiers in ten contexts. There was a good linear relation between the interpretation given to each quantifier, and the baserate expectation associated with each context. They argued that the effects of context on the lower-ranking quantifiers are there, but are weaker than on the high-ranking ones. These results are interesting in that they raise problems for the view that the meanings of quantifiers may be represented as (approximate) values on a mental scale, as well as making the search for interval scaling complex and context dependent (Newstead & Collis 1987). They remain to be properly explained, but an attempt at an explanation was made by Clark ( 1991) who argued that a quantifier might be considered as a simple function which takes a value depending on context. For instance, if Few of thefans went to the match , then the interpretation might be given by retrieving or constructing an expectation of what proportion offans normally attend a match, and then taking (say) the roth percentile. This proportion of fans, or less, is then the interpretation offew ofthe fans in the context. In this formulation, few is a function which takes a distribution as an argument, applies a cutoff, and returns a set of values. A similar account, but with other cutoffs, could be applied to other NLQs, but not to all others. Up to now this account has not been tested; it is an interesting idea, none the less, in that it provides a way of separating meaning and interpretation. An approach in which quantifiers are assumed to be mapped on to some sort of internal scale raises a number of questions. First, is the everyday meaning of quantifiers and frequency adverbs captured by scale values? It is of course clear that distinctions can be made amongst quantifiers which are not captured by these values. For example, it does not distinguish between positive or monotone increasing and negative or monotone decreasing expressions. In fact, it does not capture any of the distinctions of generalized quantifier theory. On the other
1 56
Psychological Studies of Quantifiers
Student: 6o% of our class have advanced maths qualifications. Visitor: Is that a lot? This anecdote provides an important clue about usage: whether something is a lot or not depends upon what is expected. When you don't know what to expect, use of a quantifier seems to help out, presumably because it sets up some sort of reference point. So, to substitute a lot for 6o% in our example would be to assert something like in my opinion , given what is relevant to the situation , the number in the class with advanced qualifications is high . Depending on what assumptions may be made about interlocutors sharing a point of view (and hence a reference point), a quantifier may be more or less informative than a number. So the simple view that NLQs are just vague ways of denoting numbers should be discarded. A further piece of relevant data comes from a study by Sanford, Maxey, & McGinley (under review, see also Maxey & Sanford 1 993a) in which recall of a passage containing several quantified statements was tested. In one condition, the statements were made with percentages (such as 17% of the teachers liked the method), while in the other, natural language expressions were used (such as a Jew of the teachers liked the method). It was found that less of the text was recalled under the numbers condition, which was interpreted as showing that numbers are more difficult to process than NLQs. It is therefore somewhat contrary to suppose that NLQs are 'translated' into numbers. It is noteworthy that the problems outlined above re-emerge when generalizations of the type Students likeparties are considered. Truth conditions
Downloaded from jos.oxfordjournals.org by guest on January 1, 2011
hand, it is arguable that the main way in which quantifiers are used is in terms of scale mappings (e.g. Holyoak & Glass 1 978). Unfortunately for such a simple view, Routh (this volume), using a variety of tasks to investigate the perceived similarity of quantifiers, has shown that a single-dimension scaling solution is basically untenable. Other work by Maxey & Sanford (1993b) has shown that quantifiers are not always discrirninably different from each other in terms of scale-values outside of contrastive settings. For instance, when the appropriate techniques are employed the proportions denoted by few, very few, a few, only a few , and not many are not discrirninably different. Thus the mapping account underspecifies those aspects of meaning which are in everyday use. It is therefore important to discover psychologically significant properties of quantifiers outside of the unidimensional framework which are important for their usage. The second question is whether quantified statements are normally translated on to a mental scale of any kind during comprehension. Here the answer is less clear: sometimes as listeners we certainly make an effort to get a numerical specification of a quantified statement. But equally a quantifier can convey more than a proportion:
Anthony J. Sanford, Linda M. Moxey, and Kevin Paterson I 57
regarding such statements are context dependent also (see Newstead, this volume). 3 N E GATIV I T Y , MO NOTO N I C I TY , A N D P RO CE S S I N G One of the obvious distinctions amongst NLQs is that some of them are negative by a variety of tests, such as tag tests ( I ) (Klima 1 964), the application of suspenders of implicature (Horn I 984) which allow the null set {2), and the licensing of negative polarity items in simple declarative sentences (3) {Ladusaw I 979; Zwarts, in press):
From a processing perspective, various properties of negative quantifiers have been established through psychological studies. Just & Carpenter {197 I ) used a picture-sentence verification task, in which sentences of the type quant ofthe dots are red were compared with pictures of I 6 dots, some of which were red and some black. Just & Carpenter showed that patterns of processing varied between quantifiers. Thus they found that when quant was few, the sentence was apparently encoded in terms of the majority of dots. So, given the sentence Few of the dots are red , subjects appear to check whether most of the dots are black. In contrast, for the sentence Afew ifthe dots are red , encoding appeared to be in terms of the red dots. This was established by using a procedure which ensured that sentence encoding preceded picture verification. This difference between the negative (few) and positive (a few) quantifiers was consistent with Clark's ( 1 976; see also Wason 1 965) account of verification, in which subjects encode a sentence in terms of its (pragmatic) presuppositions. Thus Jew ofthe dots are red, taken to deny the presupposition that many was the case, would be coded in terms of things that are true of many, that is, as something like It isfalse to suppose that the number ifdots which are red is many. In contrast, afew, which does not carry such a denial, would not (being coded as something like It is correct to suppose that the number ofdots which are red is small). Moxey & Sanford { I987; Sanford, Moxey, & McGinley, under review) showed that monotone decreasing quantifiers produce different patterns of focus from monotone increasing or nonmonotonic ones. In the case of the statement Some As are Bs, the sets necessary for interpretation include those As which are Bs (at least one), possible As which are not Bs, and possible Bs which are not
Downloaded from jos.oxfordjournals.org by guest on January 1, 2011
(I) Few people think that, do they? * A few people think that, do they? (2) Few people believe that, if any do. *A few people believe that, if any do. {3) Few people believe that any more. *A few people believe that any more.
1 5li Psychological Srudies ofQuannfiers
(also all As, all Bs ). Taking ease of pronominal reference as an index of which of these sets is in focus, it is obvious that (4) is acceptable and (s) is not:
As
(4) Some of the football fans went to the match. They cheered as their team scored. (s) Some of the football fans went to the match. They stayed at home instead. The NLQ some would appear to put focus on those As which are Bs (what Moxey & Sanford termed the Refterence) Set). In contrast, the NLQ few can put focus focus on those As which are not Bs (those who did not attend), what we termed Comp(lement) set reference:
Systematic exploration shows that compset reference is licensed by those NLQs which Barwise & Cooper (I 98 I) classed as monotone decreasing. However, a search for counterexamples will readily demonstrate that compset focusing quantifiers also allow refset references. For instance, to the extent that (6) is acceptable, it exemplifies a refset pattern. In order to assess whether compset focusing occurs outside of invented examples, Moxey & Sanford (I 987) presented subjects with quantified statements, followed by the plural pronoun they, for instance:
(8) Few of the football fans were at the match. They . . . Their task was to continue the they sentence in a way which made sense. The results showed clearly that some quantifiers seemed to license the use of compset reference, while others did not. On the basis of the examples studied to date (including recent work by Sanford, Moxey, & McGinley, under review), a quantifier is compset licensing iff it is (intuitively or analytically) monotone decreasing by the tests suggested by Barwise & Cooper ( I 98 I ), or through the test of licensing the use of a negative polarity item in a declarative sentence (Zwarts, in press). Thus, the following are compset licensing: not all, not quite all, not very many , not many , Jew, veryJew , hardly any , and less than X, where X is a number or a perecentage. The following are not compset licensing: nearly all, many, some , aJew, also more than X, and X where X is a number or a percentage. When compset licensing occurs with a quantifier, it does not mean that only compset continuations occur, simply that they can occur. However, in no case described above is the incidence of compset continuations less than about so%, rising to nearly Ioo% for not all (see Moxey & Sanford I99 3a for further details). In contrast, with those quantifiers classed as not licensing compset, effectively no such references occurred; rather, references were of the refset type. It may be concluded that the most accessible subset is a function of quanti fier (i.e. determiner), with monotone decreasing expressions licensing comple-
Downloaded from jos.oxfordjournals.org by guest on January 1, 2011
(6) Few of the fans went to the match. They cheered as their team scored. (7) Few of the fans went to the match. They stayed at home instead.
Anrhony J. Sanford, Linda M. Moxey, and Kevin Paterson
I
59
ment set reference, and monotone increasing or nonmonotone quantifiers requiring refset reference. We see these results as reflecting where the attention of the listener lies, and in section 4 present some evidence on comprehension to support this view. An analysis of the content of the continuations in the Maxey & Sanford (1 987) study showed an association between type of continuation chosen and quantifier. The analyses led to the conclusion that some quantifiers serve to induce in the listener patterns of reasoning concerned with why the state of affairs being depicted was true, or what were the consequences of that. In the analysis, judges classified each continuation into the following categories:
cause a fight with the opposition fans. of the number who went: e.g. Only a few of the fans went to the match. They had to cheer very loudly so that the team would hear them. Simple. Simple continuation: e.g. A few of the fans went to the match. They saw their team lose yet again. Cons. Consequence
Continuations of the RNT type have a high frequency of co-occurrence with compset continuations, as in the RNT example above. Our interpretation of this is that a sentence of the form Few X do Y leads to the question Why do many X not do Y? Because of this, Maxey & Sanford (1993a) termed quantifiers which produce this pattern comment-generating quantifiers, because they lead to processing which is focused on why the quantified statement holds. Only a Jew is also a comment-generator, with continuations of the RT and cons types predominating and occurring about equally often, as in (9) and ( w): (9) Only a few of the fans went to the match. They thought it was important to show up and give some support. (RT) ( w) Only a few of the fans went to the match. They cheered loudly despite their small numbers. (cons) Thus only a Jew is a comment-generator, but (in these examples) is not exhibiting compset focus. Maxey & Sanford concluded that the interface patterns set up by negative quantifiers in a declarative context are concerned with reasons and con sequences, while positive quantifiers such as a Jew and some give rise to predominantly simple inference patterns. These results have been shown to have consequences for attribution effects (Barton & Sanford 1 990), and for higher-level interpretations, described in section s of the present paper.
Downloaded from jos.oxfordjournals.org by guest on January 1, 2011
RNT. Reason not there: e.g. Few of the fans went to the match. They watched it on TV instead. RT. Reason there: e.g. a few of the fans went to the match. They intended to
16o Psychological Studies of Quantifiers
4 O N -L I NE M E A S URE ME N T O F FO C U S E F F E C T S
Recent research i n which the pattern of eye movements is monitored during reading provides further evidence for focus-biasing effects. The results show that they influence the time-course of interpretation as well as the patterns of continuations typically produced (e.g. Paterson I99J), thus supporting our claim that focus effects are a part of normal interpretation. The study took the pair afew IJew to contrast, a typical material being:
The brackets and bold characters are not part of the text, but are annotations. The text was divided into regions for the purposes of eye-tracking analysis, and each region is shown delineated by brackets with a bold numbered suffix (such as Rr). Thus RI contains all of the text prior to the quantifier, R2 the quantified NP, with the slashes denoting options, and R3 contains the predicate. R4 contains a reference back to the MPs: their presence denotes those MPs who were there, and their absence denotes these who were not there. Rs is the main verb, and R6 is post-verb information. It was predicted that versions of the text having the transitions A few-Their presence and Few-Their absence would cause no difficulties, since the references are to refset and compset respectively (the compatible conditions). !n contrast, the transitions A few-Their absence and Few-Their presence (the incompatible conditions) would create problems (especially so in the former case). In the experiment, subjects read (unannotated) passages like these while coupled to a Stanford Research Institute Dual-Purkinje eye-tracking system. This system allows an accurate record to be made of the position of the eye relative to the text during reading. Our interest was in the pattern of fixations over the entire text as it was read. The tracker enables several aspects of reading behaviour to be measured, and so allows an analysis of how any processing difficulties might arise. First, there is First-Pass reading time, a record of the time spent fixating a region before movement in any direction out of that region, which is regarded as a measure of immediate processing within that region. The second measure was total rime in each region, which is a measure of time spent in all visits to the region. These results are shown in Figure I. Total rime results showed that fixations were longer for the incompatible conditions in regions 4, s. and 6, but not in any other region (Figure I (a)). Thus there is a disruption effect which results in increased reading time, this occurring as early as on the critical reference.
Downloaded from jos.oxfordjournals.org by guest on January 1, 2011
(A public meeting) Local MPs were invited ro rake parr in a public enquiry abour proposals ro build a new nuclear power sration.)R1 (A few/few of rhe MPs)R2 (attended rhe meeting.)RJ (Their presence/ absence)R4 (helped)Rs (rhe meeting run more smoorhly.)R6
Anthony J. Sanford, Linda M. Moxey, and Kevin Paterson 1 6 1
0 (I)
30
----o- A few/presence
�
28
-----ts--
as .s::
26
CJl
§. l!!
�
c: 0
e::J
"0 (I) N as
C)
5l §.
Few/presence ------- Few/absence
24 22 20
2
3
Text regions
4
5
6
50 r---�====�----� ----o- A few/presence -e-- A few/absence -----ts-- Few/presence Few/absence ---· • -
2
3
Text regions
4
5
6
Figure I Time spent fixating each of regions as defined in the text (gaze duration). (a) Upper graph shows first pass durations only in msec per character. (b) Lower graph shows total rime. Critical effects occur after region 3. Legends indicate examples of the sentences contributing to the graphs.
First pass times, a measure of immediacy, showed no systematic effects on region 4, but longer fixations to the incompatible condition on region s. and effects occurring even more strongly on region 6. It appears, therefore, that it is � � � � � ��� � � �� � � � ro � oc compset is detectable (Figure r(b)). Nevertheless, as this region only contains the verb, which is not informative about focus, it is clear that the focus effect must have emanated from the quantifier, a mismatch being triggered in region 4 and resulting in a delay in processing by region S· A further measure which was made was of first-pass mean regressions. This is a measure of backwards movements, and is especially interesting when movement is outside of a region since they can be taken as indicating the seeking of information from an earlier region which might solve a parsing or interpretation problem. Prior to region 4, there are no interesting regression effects. On region 4, there is a reliable tendency for more regressions when the
Downloaded from jos.oxfordjournals.org by guest on January 1, 2011
0
-e-- A few/absence
1 62 Psychological Studies of Quantifiers
5 LEVELS OF I N TERPRETATIO N The previous sections show that the perspective adopted by a reader or listener is a result of the quantifier used. Other patterns of interpretation which go beyond scale-mapping have been reported. Thus Moxey ( 1986; Moxey & Sanford, 1 993b) conjectured that quantifiers could signal various inferences about the speaker's beliefs prior to discovering the facts about something which formed the basis of the quantified statement. For example, when the term not many was used, it might signal that the speaker's prior beliefs were to the effect that he expected more to be the case. This is in fact the same argument as that made by Clark (1 976) and Wason ( 1 965) regarding negation serving to deny presuppositions (see earlier). A pragmatic presupposition depends upon someone making the pre supposition, and from the psychological perspective the question naturally arises as to who might be making that presupposition (i.e. who had a certain expectation that was not matched by the facts). There are many possibilities which may be illustrated with the not many example, including: (a) The speaker exected more to be the case, and (b) the speaker believed the listener would have expected more to be the case. It is clear that these two things are in principle different, as the following explicit statements of (a) and (b) show: (a') I expected more would be the case. (b') I believe you expected more to be the case. Moxey & Sanford (1 993b) examined the possibility that prior expectations of the speaker, and beliefs of the speaker regarding the prior expectations of the listener might be conveyed by comment-generating quantifiers. They used three basic settings which created three different baserate frequency· expecta tions: number of local doctors who are female (low expectation), number of
Downloaded from jos.oxfordjournals.org by guest on January 1, 2011
quantifier is few, regardless of whether the condition is compset or refset. On region 5 , most regressions are due to the afew!compset (absence) condition. Our interpretation of these data is that few , licensing but not demanding compset interpretations, is always slightly ambiguous in terms of focus, so when the cue as to which focus pattern is to be used comes through (region 5), some checking takes place. With afew , however, compset is blocked, so when a compset interpretation becomes clear (by region 6), backtracking to check prior information may take place. In conclusion, the tracking study shows that interpretation is disadvantaged when there is a mismatch between the compset or refset bias of a quantifier and the orientation of subsequent text, which confirms that compset and refset focus effects occur in interpretation.
AnthonyJ. Sanford, Linda M. Moxey, and Kevin Paterson 163
people persuaded by a political speech (medium), and number of people who enjoyed a party (high). A pretest established these as different. Subjects were
presented with a quantified statement, containing one of ten quantifiers, about one of the above numbers. One group of subjects were asked to say what
percentage the quantifier was intended to denote (level 1 ) A second group was asked to say what proportion they thought the speaker had expected before the .
speaker discovered the facts (level
2). A
third group was asked to say what
proportion the speaker thought the lisrener might have expected before the listener heard the quantified statement (level 3). These levels are iHuscraced in Figure 2.
0
0
0
Sheena
Figure
1
Donald
Three levels of interpretation of a quantified statement (made by Sheena) by the
listener (Donald) as examined in the Moxey & Sanford (1993b) srudy. level 1 corresponds to the proportion which the speaker wanes to describe; level 2 is the l s i tener's perception of the speaker's expectations before discovering the facts; and level 3 is the listener's view of the assumptions the speaker is making about the expectations of the listener.
Downloaded from jos.oxfordjournals.org by guest on January 1, 2011
0
164 Psychological Studies of Quantifiers
Downloaded from jos.oxfordjournals.org by guest on January 1, 2011
If there were no effects of quantifier on the second and third questions, then the responses might be expected to be in line with the independently measured baserates, since the speaker was described as a local news reporter, who would therefore have the same expectations as the subjects. However, consider the possibility of containing the pair afew and not many . If aJew does not trigger a denial that more was expected, and not many does trigger such a denial, then with level 2 and/or 3 questions, the percentages reported for not many should be higher than for aJew . Although absolute baserate did influence the answers for all groups, there was a reliable effect of quantifiers as well. Thus the values given suggested that the following quantifiers were associated with the listener believing that the speaker had expected more than turned out to be the case: veryfew, few , not many , and only a few . These quantifiers produced relatively high proportions in response to the question. This is consistent with the idea that these quantifiers presuppose a higher level than is being asserted, and signal denials of the presupposition. Here, though, presupposition is the listener's perception of an expectation on the part of the speaker. In contrast, use of aJew produced low answers to the expectation question. Similar results held for the percentage which the listener inferred that the speaker thought the listener to have expected in advance. The only difference was that for this group,Jew did not seem to signal that more had been expected, a point which Moxey & Sanford believe to be of some potential importance, since it constitutes a major difference between quantifiers which have otherwise similar effects in communication. Perhaps few sounds more formal and aloof in usage than not many because by using it the listener does not engage or accommodate the prior beliefs of the listener (see Moxey & Sanford 1 993a for an extended discussion). In addition to the results with negative quantifiers, the expression quite afew appeared to signal that less had been expected than turned out to be the case, this holding for both levels 2 and 3· The two main discoveries from these studies are first that quantifiers may convey information about the speaker's prior beliefs as well as about the current situation, for instance that the speaker expected more to be the case (not many, Jew) or less to be the case (quite aJew). Secondly, characterizing quantifiers as signalling data about prior beliefs leads to the question of whose beliefs. This may have had a trivial answer (all speaker-listener's prior beliefs) which was not different from the conventional notion of presupposition. As it turned out, there was a strong suggestion that a differentiation between level 2 and level 3 descriptions of prior beliefs must be made in order to explain the com municative functions which quantifiers might have. Of course, whether all of this potential information forms a standard part of interpretation, or whether it is optional, depending upon the particular circumstances of a communication remains to be established.
Amhony J. Sanford , Linda M. Moxey, and Kevin Parerson 165
6 D I S C U S S IO N Psychological research on NLQs has been dominated by rather atheoretical work on scales. More recently there have been investigations of how quantifiers influence the patterns of inference made by llsteners. Given that it is possible to elicit assignments of numbers or proportions to NLQs, it remains to explain how this might be accomplished. We begin the discussion with this issue.
Downloaded from jos.oxfordjournals.org by guest on January 1, 2011
Meaning and Scales . The greatest problem for the view that quantifiers denote numerical values of various sorts is the variability of the mapping from quantifier to scale and vice versa. It was shown that the context of the quantifier, and in particular the number or proportion one might normally expect for a situation, influences the quantifier-to-number mapping. This raises obvious difficulties for making proportions or numbers (however vague) part of the truth conditions for NLQs. Two ways ofincorporating context into the interpretation process have been suggested. The first gives a quantifier a core meaning which is utilized in interpretation. One approach, suggested by Wallsten et al. (1 986), is that each quantifier has a core distribution of values, and that this distribution is combined with a prior distribution which comes from context, perhaps by some process of forming a weighted average. The result is the contextual interpretation. Each quantifier would have a different core distribution. Clark's ( 1 991) suggestion was that a quantifier (such as Jew) might take a distribution of expected prior values and operate on it by cutting off everything below (say) the 10th percentile. Each quantifier would have a different cutoff In this way, the meaning is fixed (everything below the roth percentile of the current distribution), but the current distribution will vary and provides the context to give an interpretation. Neither of these approaches has been thoroughly tested, and both raise the question of how either a distribution or a cutoff could be learned. One possibility might be some sort of prototype learning: by acquiring an NLQ-to-number mapping in contexts where there are expectations of baserate, generalization to other values novel of expectation is possible. It has been shown that feedforward connectionist networks can behave in this way. Whether these approaches are falsifiable remains to be established, but we do suggest that such an approach is inadequate, and offer something more radical as an alternative explanation of how quantifiers interact with contextual information. First, consider the role played by the quantifiers themselves. Experiments have shown that there are basically two groups of quantifiers those which denote small amounts, and those which denote large amounts. Within these groups it is very difficult to differentiate between the expressions simply in terms of the amount denoted. We also know that replacing a
166 Psychological Studies of Quantifiers
Focus and higher-level interpretation . Such a line of reasoning is consistent with the other material reviewed above. Both the focusing property and the comment-
Downloaded from jos.oxfordjournals.org by guest on January 1, 2011
quantifier in a piece of text with a numerical value does not always result in a good paraphrase. Thus for example there are I o seats left and there are a lot ofseats left do not mean the same thing, even when used to describe the same situation, because a lot implies that there are enough for some purpose. Likewise there are I o seats left and there are few seats left do not mean the same thing even when used to describe the same situation, because unless the listener has derailed expect ations about what is normal or desired in the seat situation, I o does not imply that there is not enough for some purpose whereas few does imply this. Given these ideas, we propose that quantifiers provide only very crude numerical information. Basically, an assertion of amount through an NLQ indicates that the amount is large enough to justify making one set of inferences, or small enough to merit another. Secondly, consider the context in which a quantifier is used. The above idea relies on there being a numerical value (say, an expectation or a desired amount) in the context, about which the subject can infer that there is in fact enough, or not enough as the case may be. In experiments where subjects are asked to provide an amount for a quantifier, they are therefore dependent on numerative information being available in the context. For example, there may be a numerical value associated with what constitutes enough in a particular context, or a distribution of likely numerical values for what constitutes enough. If such detailed information is not available, as would be the case for the seats example above, then presumably the subject would have to generalize from this context to an apparently similar one which has more numerical information associated with it. Essentially, the argument is that contexts vary with respect to the sophistica tion of numerical information available within them. They cannot all be represented with normal curves on ratio scales. In addition, even if there is detailed numerical information available, there are likely to be cutoff values associated with the licensing of particular inferences, e.g. if more than 8o% of the people enjoy a party, it's a good party, etc. Furthermore, quantifiers are primarily used to lead the interpreter to make certain inferences, e.g. the party was good. They do this by resolving the issue of whether there are enough or not enough for some purpose. While the kind of account we are aiming for is less well defined than Wallsten's and certainly no more detailed than Clark's, it is based on the view that quantifiers perform a function in the psychological process of language understanding. That is, they lead the interpreter to make inferences which are consistent with the purpose(s) of the producer, rather than simply attempting to describe a state of affairs in a vague way.
Anthony J. Sanford, Linda M. Moxey, and Kevin Paterson 1 67
Downloaded from jos.oxfordjournals.org by guest on January 1, 2011
generation property relate dearly to the control of inference by quantifiers. So, for example, we argue that it may typically matter less to resolve what few students means numerically in some context than to resolve why there were so few (in terms of an explanation of complement-set behaviour). The possibility of higher-order interpretations (section s) is also consistent with the idea that quantifiers express value-judgement information, such as there were more X than I expected , or, there were more X than I thought you would have expected . Psychological work on these problems has only just begun, but it is clear that they are general (Maxey & Sanford I 993a). All of these issues are important for quantifier choice in communicative situations, and are thus fair game for a usable account of meaning. As psychologists, we would frame an account of NLQ meaning in terms of the mental operations which they bring about, and the data outlined above mark a step in this direction. However, this is plainly different from the goal of developing an understanding of quantifier meaning within some more formal semantic (mathematical) framework, such as generalized quantifier theory. Psychologists may take two rather different approaches to the relation of data to formal theory. One approach, which we might term the 'Psychological Reality' approach, takes some aspect of a formal theory and treats it as a prescription for a procedure. Performance tests are then carried out to see if human behaviour follows the procedure. This is perhaps the conventional way of thinking about the relation of psychology and linguistic theories oflanguage. A rather different orientation is to look for a psychological (processing) explanation of the judgements which underlie efforts to fit natural language to a formalism. The basic assumption of this approach is that judgements of grammaticality, and of the semantic or discourse-level acceptability of a piece of language constitute acts of information processing; judgements are not naturally given in any way. In one such case, Maxey & Sanford ( I 993a) noted that many judgements of the monotonicity of NLQs rely entirely on intuition {Barwise & Cooper I 98 I). For instance, consider judgement of whether the following is true: Iffew ofthefans went to the meeting thenfew ofthefans went to the meeting early. Analytically, this is true to the extent that the reading given tofew ofthefans is X (a small number) or less (including zero ), since X or less is analytically monotone decreasing. We suggested that the compset focusing property offew is what leads to the zero interpretation being allowed, because focus on the compset removes attention from the reference set which might then be considered empty (see Maxey & Sanford I 993a: 82-6). Since this is a way of inferring the man dec property ( Q is X or less) we believe this may be the root of judgements of downward monotonicity. A related approach to the explanation of negative polarity items has recently been taken by von Klopp (under review), who takes as her starting point a theory of natural language negation which puts the idea of dissociation
I 68 Psychological Srudies of Quannfiers
Acknowledgements AJS and LM thank The British Academy and ESRC for support; KP is supported by a srudentship from the Carnegie Trust. We are grateful to Robin Cooper and Anna von Klopp for useful discussions. Received: I 2. I 2.9 3 Revised version received: I 6.02.94
Address for correspondence: ANTHONY J. SANFORD Human Communication Research Centre Department ofPsyclzology University ofGlasgow Glasgow G t z 8RT Scotland, UK
RE F E RE N CE S Altham, J. E. J. ( 1 97 1 ), The Logic of Plurality, Methuen, London. Barton, S. B. & Sanford, A J. (I 990), 'The control of attributional patterns by the focusing properties of quantifying expres sions', Journal ofSemantics , 7, 8 I -92. Barwise, J. & Cooper, R. (I 98 I ), 'Generalized quantifiers and narural language', Linguis tics and Philosophy, 4, I 59-2 I 9. Bass, B. M., Cascio, W. F., & O'Connor, E. J. ( I 974). 'Magnitude estimations of fre quency and amount', Journal of Applied Psychology, 59, 3 I 3-20.
Carlson, G. & Tannenhaus, M. (eds) (I 989), Linguistic Structure in Longuage Processing , Kluwer, Dordrecht. Clark, H. H. ( I 976), Semantics and Comprehen sion , Mouton, The Hague. Clark, H. H. ( I 99 I ), 'Words, the world, and their possibilities', in G. R. Lockhead & James R. Pomerantz (eds), The Perception of Structure: Essays in Honor of Wendell R. Garner, American Psychological Associa tion, Washington, DC. Cooper, R. (I 983), Quantification and Syntactic Theory , Reidel, Dordrecht.
Downloaded from jos.oxfordjournals.org by guest on January 1, 2011
between entities at the centre of negation (von Klopp 1993). She uses the compset focus property of quantifiers to explain how they take polarity items. This idea is quite different from the usual (syntactic) accounts, and effectively suggests that licensing negative polarity items is an outcome of focus, rather than negative polarity and focus both being an outcome of monotonicity. In short, our view is that the focusing (and comment-generating) properties ofNLQs reflect their usage . They have these properties because these properties are useful. Properties such as those which characterize generalized quantifiers areformal . The fit ofNLQs to generalized quantifier theory is to be understood in terms of what (psychological processing) properties lead to our judgements of whether the fit is good or not (in terms of the linguistic approximations to the logical formulae). Such an approach is in its infancy, but it does take seriously the very plausible idea that intuitions about language constitute behaviour, and are thus a proper part of psychology.
Anthony J. Sanford, Linda M. Moxey, and Kevin Paterson 169
Downloaded from jos.oxfordjournals.org by guest on January 1, 2011
Hirschberg, J. B. (1985), 'A theory of scalar polarity items', paper presented to the implicature', Technical report MS-CISI 979 Meeting of the Linguistics Society of 8 5-56, Department of Computer and America, Los Angeles. Information Science, Pennsylvania State Levinson, S. C. ( I 983), Pragmatics, Cambridge University, Philadelphia. University Press, Cambridge. Holyoak, K. J. & Glass, A. L. ( 1 978), 'Recogni Mitcham D., Carlin, J., VanDerveer, N., & tion confusions among quantifiers',Journal Loveland, K. (I 980), 'Lexical and structural of Verbal Learning and Verbal Behavior, 17, cues to quantifier scope relations', Journal 249--64. ofPsycholinguistic Research , 9, 367-77. Horn, L. R. ( 1 984), 'Toward a new taxonomy Moxey, L. M. ( 1 986), A psychological investi for pragmatic inference: Q-based and gation of the use and interpretation of R-based implicature', in D. Schifftin (ed.), English quantifiers. Unpublished PhD Meaning, Form, and Use in Context: Linguistic thesis, University of Glasgow. Applications , Georgetown University Moxey, L. M. & Sanford, A.J. ( I 987), 'Quan Round Table I 984, Washington, DC. tifiers and focus', Journal of Semantics , 5, Ioup, G. ( I 975), 'Some universals for quanti I 89-206. fier scope', in J. Kimball (ed.), Syntax and Moxey, L. M. & Sanford, A. J. (I 99 3 a), Com Semantics, vol. 4, Academic Press, New municating Quantities: A Psychological Per York, 37-58. spective , Lawrence Erlbaum Associates Ltd, Johnson-Laird, P. N. (I983), Mental Models , Hove. Cambridge University Press, Cambridge. Moxey, L. M. & Sanford, A.J. ( I 993b), 'Prior Johnson-Laird, P. N. & Byrne, R ( 1 99 I ), expectation and the interpretation of Deduction , Lawrence Erlbaum Associates natural language quantifiers', European Ltd, Hove. journal ofCognitive Psychology, 5, 73-9 1. Just, M. A. & Carpenter, P. A. (1971), 'Com Newstead, S. E. & Collis, J. M. (1 987), prehension of negation with quantifica 'Context and the interpretation of quanti tion', journal of Verbal Learning and Verbal fiers of frequency', Ergonomics, 30, I 447Behavior, 10, 244-5 3· 62. Keenan, E. L. & Stavi, J. ( I 986), 'A semantic Parducci, A. ( 1 968), 'Often is often', American characterization of natural language Psychologist , 23, 828. determiners', Linguistics and Philosophy, 9, Paterson, K. ( 1 993), 'The immediacy of 25 3-3 26. situation-based inferences in quantified Klima, E. S. ( I 964), 'Negation in English', in texts', paper presented at the International J. A. Fodor & J. J. Katz (eds), TheStructure of Conference on the Psychology of Lan Language, Prentice-Hall, Englewood guage and Communication, University of Cliffs, NJ. Glasgow, 3 1 August-3 September. Klopp, A. von ( 1 993) 'Negation: Implications Pepper, S. ( 198 1), 'Problems in the quantifica for Theories of Natural Language Process tion of frequency expressions', in D. Fiske ing', doctoral thesis, Centre for Cognitive (ed.), New Directionsfor Methodology ofSocial Science, University of Edinburgh, Edin and Behavioral Science , vol. 9, Jossey-Bass, burgh. San Francisco, 25-4 1 . Klopp, A. von (under review), 'Polarity items: Pepper, S . & Prytulak, L. S . (1 974), 'Some are the tight questions being asked?' For times frequently means seldom: context Language and Philosophy. effects in the interpretations of quantita Kurtzman, H. S. & MacDonald M. C. ( I 993), tive expressions', Journal of Research in 'Resolution of quantifier scope ambigui Personality, 8, 95- 1 0 1 . ties', Cognition , 48, 243-79. Pohl, N. F . ( I 98 1 ) , 'Scale considerations i n Ladusaw, W. A. ( I 979), 'On the notion of using vague quantifiers', journal of Experi affective in the analysis of negative mental Education , 49, 2 3 5-40.
1 70 Psychological Studies of Quantifiers doctoral thesis, Rijksuniversiteit, Gron ingen. Wallsten, T. S., Fillenbaum, S., & Cox, J. A. ( 1 986), 'Base-rate effects on the interpre tations of probability and frequency expressions', journal of Memory and Lan guage, 25, 5 7 1 -87. Wason, P. C. ( I 965), 'The contexts of plaus ible denial', Journal of Verbal Learning and Verbal Behavior, 4, 7- I 1 . Westersthal, D. ( 1 989), 'Quantifiers in formal and natural languages', in D. Gabby & F. Guenthner (eds.), Handbook ofPhilosophical Logic , vol. 4, Reidel, Dordrecht, 1-3 1 . Zwarts, F. (in press), 'Three types of polarity', in F. Hamm & E. Hinrichs (eds), Semantics .
Downloaded from jos.oxfordjournals.org by guest on January 1, 2011
Reinharr, T. ( 1 983 ), Anaphora and Semantic Interpretation , Croom Helm, London. Sanford, A J., Moxey, L. M., & Grant, G. (submitted), 'Situation-based expectations and the numerical interpretation of quan tifiers'. Sanford, A J., Moxey, L. M., & McGinley, M. (under review), 'Memory phenomena with passages containing quantifiers'. Schriesheim, C. & Schriesheim, J. ( 1 978), 'Development and empirical verification of new response categories to increase the validity of multiple response alternative questionnaires', Educational and Psyclzologi cal Measurement , 34, 877-84. Swart, H. de ( 1 99 I ), 'Adverbs of quantifica tion: a generalized quantifier approach',
© N.l.S. Foundarion (199�)
joumal oJSemanllcs 10: 171-19X
Scales in the Interpretation ofWords, Sentences, and Texts R O B E R T J. JARVE L L A * and L I T A L U N D Q U I ST t • University
ofMontreal and Lund University
tCopenhag en Business School
Abstract
I N T RO D U C T IO N In two recent papers, we have applied a semantic theory of scalarity i n language (Ducrot I 988; Anscombre I 989) to the interpretation of discourse to explore the influence during reading of a specific kind of graded inference, which we have labelled 'scalar inference'. Of special interest in that work is the interpretation of sentences containing scalar adverbs such as more than , less than , almost and only. Consider (I)-(3): (I) George Brown got almost s,ooo voces. (2) George Brown got only s,ooo voces. (3) George Brown got s,ooo votes. Sentences (I) and (2) are oriented in the sense chat they direct the reader cowards one of two conclusions, either of which would be possible with the corresponding neutral sentence (3):
( I ') George Brown got almost s,ooo votes. He is likely to win the election. ( I ") ?George Brown got almost s,ooo voces. He is likely to lose the election. (2') George Brown got only s,ooo voces. He is likely to lose the election. (2") ?George Brown got only s,ooo votes. He is likely to win the election.
Downloaded from jos.oxfordjournals.org by guest on January 1, 2011
Scales implicit in the interpretation of words, sentences, and texts were studied in a series of five experiments in French. In Experiments 1 and 2, readers chose an arrow pomting upwards or downwards as corresponding, or being contrary to, the meaning of scalar expressions and sentences drawn from texts about competitions. A large degree of consensus among these judgments was obtained at both word and sentence levels; in judging sentences, a speed advantage was found when a high level of performance was implied and when a sentence described a leading or winning position. In Experiments 3-5, readers were shown texts which contained similar scalar predications oriented towards a positive or a negative outcome. Sentence orientation was found to affect the interpretation of ambiguous expressions in the texts and also reading time.
1 72
Scalar Interpretation or Words, Sentences and Texts
In work done in Danish and Finnish (Lundquist & Jarvella I994; Lundquist, Jarvella & Hyona I 994), we showed that the orientation ofsentences containing such scalar adverbs activates inferential schemes which, among other things, can help in resolving referential ambiguity. The type of scalar inference which we studied is explained via the notion of topos taken from the same semantic theory. A topos is an inference scheme which involves two scales, as shown in the following canonical form: (the more (the less) an object 0 has the property P, the more (the less) an object 0' will have the property P')
(4') George Brown got almost s,ooo votes. The white-haired corporate lawyer is likely to win the election. (4") George Brown got only s,ooo votes. The white-haired corporate lawyer is likely to win the election. While ( I '), (2 '), and (4') support a co-referential interpretation of George Brown in the first sentence and the subject noun phrase in the second sentence, ( x "), (2 ") , and (4") argue against such an interpretation. The use of the pronoun he in ( I ") and ( 2 ") makes the two sentence pairs unacceptable, whereas in a context like (4") the definite NP can be interpreted as referring to someone other than the individual just mentioned. A disjoint interpretation was, in fact, found for such expressions in our prior work. In addition, the work showed that speed and ease in reading sentences can depend on their orientation. In our previous papers, inspired partly by work in cognitive semantics Gohnson I987; Lakoff I 987; Langacker I 987), we hypothesized that a variety of predications are subjectively oriented upwards or downwards , and that even scalar predications, such as obtained more than x votes/was leading and obtained less than x votes/was trailing, are taken metaphorically as pointing upwards and
Downloaded from jos.oxfordjournals.org by guest on January 1, 2011
In ( I ) above, we take the presence of 'almost' to activate a scale oriented upwards of the form (the more votes X gets . . .), which is consistent with a second scale, available from general background knowledge, (the greater the chance X will win). This inference scheme explains the linguistic acceptability of ( I ') and the unacceptability of ( I ") . In (2), on the other hand, we take the presence of 'only' to activate a scale (the fewer votes X gets . . .), which is oriented downwards and consistent with a second scale (the lesser the chance X will win). This inference scheme explains why (2 ') seems more natural and linguistically acceptable than (2 "). The two scales involved in such inference schemes may be oriented in the same way (the more/the more, or the less/the less), as in ( I '), (2 '), and (4'), or they may be oriented in different ways (the more/the less, or the less/the more), as in ( I ") , (2 "), and in (4"):
Roberr J. Jarvella and Lila Lundquis£ 1 7 3
downwards respectively. The terms almost and only seem to be oriented towards relatively high and low positions on a scale respectively: (s) George Brown got almost s,ooo votes (
=
many). But Bill White got still
MORE.
(6) George Brown got only s,ooo votes (= few). But Bill White got still LESS. Moreover, this implication is present whether a large value is evaluated in the context as being more positive than a smaller one, as in the cases shown above, but also when the converse is true, as in (7) and (8):
The quantity conveyed in (7) is positioned on a scale oriented upwards, and implied to be large and important, that in (8) positioned on a downwards oriented scale, and implied to be insignificant. However, the evaluation has changed from that in (I) and (2). Almost s,ooo votes is evaluated as negative in (7) and only 5, 000 votes as positive in (8). Clearly, the evaluation of almost x or only x turns partly on what a larger predication including them is about. In the present paper, we limit ourselves to cases where directionality and evaluation correspond. In many conventional settings, including the ones studied below, high rather than low values on a scale seem to be associated with more positive evaluations. A general correlation in our life experience between upwardness on the one hand, and quantity and quality on the other, has been argued in cognitive linguistics to underlie a variety of metaphors pervading our use of language (see e.g. Lakoff & Johnson I 98o: 14- I 9). We will be concerned here with directionality as an aspect of perceived meaning at different linguistic levels. Our previous findings concerning scalar inference in text were quite similar in Danish and Finnish, two structurally dissimilar languages. In the present work, done in French, we pursued the question of directionality underlying language-whether it be literal or figurative-in words, in sentences, and in text. It was in French that the theory of argumentation and scales was originally formulated (Anscombre & Ducrot I 98 3), and also where it was applied first to the interpretation of definite descriptions (Lundquist I 987a). In the present paper, we will be concerned with four related topics: (a) the perceived scalar orientation of particular words and expressions; (b) the metaphorical interpretation of sentence orientation; (c) effects of sentence orientation in the process of resolving referential ambiguity; and (d) the consequences of sentence orientation for speed in reading.
Downloaded from jos.oxfordjournals.org by guest on January 1, 2011
(7) George Brown lost almost s,ooo votes. He is likely to be defeated. (8) George Brown lost only s,ooo votes. He will easily gain re-election.
1 74 Scalar Interpretation of Words, Sentences and Texts
EXPE RIMENT I : W O R D S
(9) It's six o'clock. Hurry up! ( w) It's almost six o'clock (= late). Hurry up! (- late) ( I I ) ?It's almost six o'clock (- late). Take your time. (- early) Discussions of the contrasting views on this issue, and whether the phenomena in question should be described linguistically in one way or another, can be found elsewhere (e.g. Wierzbicka I 986; Lundquist I 990; Konig I 99 I ; Lundquist & Jarvella 1 994). Here, we will be interested in showing rather how scalar properties enter into the inferences made. We will begin by illustrating some similarities in directionality perceived for a variety of expressions. The first study we report is one ofjudgments made of the relative orientation of words and multiple-word expressions, including presque (ALMOST) and ne que (ONLY). To approach this issue, we required a task which was focused on and could reflect orientation. There being no such task at hand, we invented one which took as its basis the orientation of the body. In normal erect posture, the gravitational vertical and the surface of the earth (the horizontal) are the central
Downloaded from jos.oxfordjournals.org by guest on January 1, 2011
A rather wide variety of linguistic expressions seem to involve scales in their interpretation. Such expressions include logical quantifiers (e.g. tout, rien ), non logical quantifying expressions (e.g. beaucoup , peu ), and explicitly directional expressions (e.g., en haut, descendre), but also many 'ordinary' predicates (e.g. chaud , froid ; grand, petit) and some adverbs (e.g. plus de, mains de). In holding that the interpretation of such expressions seems to involve a scale, we mean that a dimension can usually be defined in such cases along which different, related expressions can be placed or ordered, or can be used to describe relative direction, position, or strength on that dimension. While there is agreement among linguists as to the usefulness of conceiving of some meanings in terms of scales (Ducrot I 98 8; Anscombre I 989; Horn I 989; Konig I 99 I ), it remains to be shown that non-linguists share these feelings, and the issue of how scalar meanings come into play when language is used has been largely ignored. For some meanings, such as those of the concepts ALMOST and ONLY, there is disagreement among linguists over what their scalar properties are. For example, Sadock ( I 98 I ) arrived at the conclusion that an utterance almostp is equivalent to notp , because had Sender known p , he would have said so and not used the weaker form almost p (Grice's Maxim of Quantity), whereas for Ducrot and Anscombre, the expression almostp activates a scale oriented in the same direction as p because it leads to the same type of conclusion as p does. Ducrot and Anscombre's line ofreasoning is supported by continuation tests, as exemplified in (9)-( I I ):
Robert J. Jarvella and Lira Lundquist 1 75
,
Method Subjects, materials, and procedure
Fifty-seven French-speaking students at the University of Montreal were shown the expressions listed above in a random order. Each expression was judged in tum, and the subject registered her response by pressing one of two buttons housed on the mouse of a personal computer. An arrow pointing upwards and away from the subject was taped onto one mouse button, and an arrow pointing downwards and towards the subject onto a second button. Button presses made and response latencies were recorded.
Downloaded from jos.oxfordjournals.org by guest on January 1, 2011
axes in our perception of what is in front of us. The English expressions which code the directions on these axes are up and down , lift and right . However, an act of pointing is also a very conventional means for indicating direction (see e.g. Bi.ihler 1 934). To approach the directionality underlying words here, we chose to use a nonlinguistic task, and asked subjects to respond to some expressions they were shown by assigning each expression an arrow. In this task, two arrows, one pointing up and away from the body, the other down and towards it, were shown as the only possible alternatives. Use of this dimension for eliciting judgments was dictated partly by convenience, bur the choice made can also be motivated psychologically. The gravitational vertical is, as mentioned, central in our perceptual experience of the world, and is the axis taken from perception and used to define a basic 'more is up, less is down' bodily image schema that has been proposed by cognitive linguists. We do not assume that scalar meanings are usually mapped onto a single dimension, not even those of the limited set of words studied here. In using this task, we were interested in seeing how the directional aspect of different words' meanings would converge. Some of the expressions studied were scalar adverbs (plus de , mains de, presque, ne que), 1 while others were of vertical origin (en haut, en bas , manter, descendre , au-dessus , au-dessous ) were quantitative terms (tout, rien , beaucoup , peu ), indicated polarity (positif, negatif), or were oriented towards a positive or negative result in a metaphorical sense (gagner, perdre ). Our hypothesis in this study was that expressions which describe upward ness, or which represent or intensify a large quantity, a positive polarity or a positive result, would be judged in the same way, and be judged in the opposite way as expressions which describe downwardness, or which represent or intensify a small quantity or a negative polarity or a negative result. In practice, this means that expressions such as en haut , tout, plus de, positif, gagner, etc. should be judged similarly in the task used, and judged in the opposite way to en bas ' rien ' moins de' negatif, perdre ' etc.
1 76 Scalar Interpretation of Words, Sentences and Texts
Design
Subjects were assigned by rotation to six groups based on ( 1) whether the upwards-pointing arrow was taped to the left or the right button on the mouse, (2) whether the subject was told to choose arrows which best corresponded to the meaning of the expressions shown, or which were contrary to their mean ing, and (3) the time allowed in reaching a decision. Four groups of subjects were tested with a 3-second time limit, varying mouse button assignments and the instructions. Two additional groups were tested with an 8-second time limit, varying mouse button assignments but using corresponding-meaning instructions only.
Corresponding meaning down
up
down
0 89
0 8.9
6o 0 95
71'!
ll3 0 61 17
0 94
l'! J 0
95 0 1 00 0
0 6) 0 1 00 0 92
0 61 0 8J 0 94
l'! J 0 94 0 l'! J 0
97 0 20 10
0 95 30 6s
0 ll9 22 28
l'! J 0 17 0
1 00 12
0 88
nla nla
nla nla
up Verticality expressions en haur 97 en bas au-dessus au-dessous monter descendre Quanriry expressions rout nen beaucoup peu positif negatif Scalar adverbs plus de moins de presque ne que Metaphorical uses gagner perdre
Contrary meaning
90 I)
95 5 and polarity 90 0
I I
Note : hems summing m less rhan uniry reflect cases where n.•sponses were not given within the rime lim" imposed.
Downloaded from jos.oxfordjournals.org by guest on January 1, 2011
1 Experiment 1 (arrows task). Percent of cases in which an arrow pointing upwards or downwards was taken to correspond, or be contrary to, a meaning of some French expressions
Table
Robert J. Jarvella and Lira Lundquist 1 77
Results and discussion
Downloaded from jos.oxfordjournals.org by guest on January 1, 2011
Table I summarizes the results obtained for the 3-second groups tested. It can be seen from the table that the distributions ofjudgments obtained with the two instructional sets are symmetrical. The only other effect of instructions found was that contrary-meaning instructions led to slower responses (mean 1.6 vs. 1.3 seconds, F(I,6) - 62. I 8, MSe - 32430, p < o.ooi by pairs of items where all or almost all responses were given before the deadline). By hypothesis, the first and second item in each pair of expressions shown in Table I should behave in this task in different and opposite ways. We expected that the first item (e.g. en haut) would be assigned an 'up'-arrow when judged with instructions to use corresponding meaning, but a 'down'-arrow when contrary-meaning instructions were used, and that the second item in a pair (e.g. en bas) would be assigned a 'down'-arrow with corresponding-meaning instructions but an 'up'-arrow with contrary-meaning instructions. These predictions were satisfied 98% and 97% of the time respectively for responses given before the 3-second deadline, and 88% and 8 3% of the time if failures to register a decision during this interval are counted as going against the hypothesis. Thus, subjects tended to judge the great majority of items as we had expected. Speed in judging the two sets of items did not vary (I. 5 vs. I ·4 seconds, F(I ,6) - 2.03, MSe - 2 3 I I 9, p - 0.204). French ALMOST and ONLY were the most time-consuming of the I 8 expressions judged. On presque and ne que together, less than half the judgments were registered by the 3-second deadline (the data shown in Table I), and, using an 8-second limit, only 7 5% were registered. With extended time to respond, however, a kind of consensus emerged for both terms. Fifteen of I 7 subjects tested with a longer time limit and corresponding-meaning instruc tions assigned a 'down'-arrow to ne que and none assigned it an 'up'-arrow. For the same I 7 subjects, there were eight responses assigning an 'up'-arrow to presque , and just two assigning it a 'down'-arrow. The missing responses, however, show that there was still uncertainty in interpreting presque. This is in line with results reported by Bassano (I99 I : I 59), showing that it is difficult to decide whether presque refers to a satisfactory amount. Finally, the pair of verbs tested which could be assigned directionality only in a metaphorical sense showed the same general pattern as other pairs. With corresponding-meaning instructions, gagner (wiN) was always assigned the arrow pointing upwards, and, except in two cases perdre (wsE) was assigned the arrow oriented downwards. This pattern closely resembles that obtained for the verbs tested which explicitly describe upwards and downwards movement, namely monter and descendre . In general, we interpret the results as showing that, for speakers of French, there is a fairly transparent correspondence between a directional aspect of the meaning of the I 8 expressions studied and
1 71! Scalar Interpretation of Words, Sentences and Texts
that symbolized by an arrow pointing up or down. Scalar expressions coming from a variety of domains were systematically classified as having a meaning oriented in either one direction or the other, and the pattern of judgments across domains converged. E X PE R I MENT
2:
SENTE N C E S
Downloaded from jos.oxfordjournals.org by guest on January 1, 2011
Experiment I was concerned with a directional aspect of the local semantics of certain French words and expressions, i.e. outside of any larger linguistic context. In the present study, we began to examine how scale direction may be combined within and between sentences in French. In this study, we were concerned with a more metaphorical kind of directionality corresponding to the concepts of OBTAINING A GOOD (high) SCORE, OBTAINING A POOR (low) SCORE, LEADING, TRAILING, WINNING, and LOSING. The task introduced in Experiment I was now extended to elicit judgments about full sentences. The sentences were drawn from eight texts written about various kinds of contests (a beauty pageant, a ski race, a dog show, a baking contest, a swimming competition, filling of a job opening, a horse race, and a tennis championship). Each text described two fictitious performers by name (as being favored), a level of performance achieved by each performer, and an apparent consequence of the two performances, in terms of a given performer's position. We have suggested previously (Lundquist & Jarvella 1 994) that a sentence which implies a high level of performance or describes a leading/ winning position will be seen as positive and therefore be understood figurat ively as being oriented upwards, whereas a sentence which implies a low level of performance or describes a trailing/losing position will be seen as negative and therefore understood as being oriented downwards. Patterns of reading times strongly supported a distinction in the processing of sentences of these kinds. The arguments which we were able to offer for the underlying cause being their directionality, however, were indirect. Basically, they were anchored in work done in cognitive semantics and in findings from experiments on sent ence verification (e.g. Clark 1 976). We thus decided to look for some more unequivocal evidence of a link between directionality underlying language and reading performance. This link was sought in the latency needed to perceive directionality in sentences. Three predictions derived from our theory were made here. First, we hypothesized that, in an arrow-assigning situation, subjects would system atically interpret the sentences from the competition settings of interest as figu ratively having a meaning directed upwards or downwards. Second, on the basis of our earlier results, sentences whose figurative meaning can be taken to be directed upwards were predicted to be judged for directionality faster than ones
Robert J. Jarvella and Lira Lundquist 1 79
whose meaning can be taken to be directed downwards. And third, we pre dicted that, in a pair of related, oriented sentences, the second sentence would be read and judged faster if it shared perceived orientation with its antecedent. Method Materials
(a) Le s km de ski nordique (The s km Nordic ski race) (b) Marie Dufour etJosee Aubert etaient deux concurrentes serieuses. (Marie Dufour and Josee Aubert were two serious contenders) This lead-in was followed by two sentence pairs. The predications within each sentence pair were co-oriented or anti-oriented. Co-oriented sentence pairs, such as (c)-(d) and (e)-(�, (c) Sur la longue montee, Marie avait une prise sur la neige presque ideale. (In the long ascent, Marie had a grip on the snow that was almost ideal) (d) Marie etait a ce moment-la nettement devant. (At that moment Marie was clearly ahead) (e) Dans la descente sinueuse vers l'arrivee, Josee n'avait qu'une glisse peu rapide. (In the twisting descent towards the finish,Josee had only a slowish glide) (� Josee perdit la course du jour. Qosee lost the day's race) consisted of an antecedent sentence implying a low level of performance followed by a consequent describing a trailing or losing position, or an antecedent sentence implying a high level of performance followed by a conse quent sentence describing a leading or winning position. The individuals' names within a pair of co-oriented sentences were held constant. Anri-oriented sentence pairs, such as (g)-(h) and (i)-U), (g) Sur la longue montee, Marie n'avait qu'une prise mal assuree sur la neige. (In the long .ascent, Marie had only a poor grip on the snow) (h) Josee etait a ce moment-la nettement devant. (At that moment Josee was clearly ahead) (i) Dans la descente sinueuse vers l'arrivee,Josee avait une glisse presque ideale.
Downloaded from jos.oxfordjournals.org by guest on January 1, 2011
A total of 32 sextuples consisting of a title plus five sentences were constructed, adapted from eight longer narrative texts about competitions described further below. Following the title, the first sentence in a sequence mentioned by name two contestants favored in a competition:
1 80 Scalar Interpretation of Words, Sentences and Texts
(In the twisting descent towards the finish, Josee had a glide which was almost ideal) U) Marie perdit la course du jour. (Marie lost the day's race)
Subjects and procedure
Two competitions in each version, preceded by a practice item about a bicycle race, were presented to so subjects tested previously in Experiment 1. As before, an arrow pointing up and away from the subject was taped to one burton of the mouse and an arrow pointing down and towards the subject was taped to a second mouse button. The subject pressed a third button on the mouse to advance the 'text' to the four sentences of interest, and then judged each sentence shown in turn. All sentences were presented on a single So character line. Antecedent sentences averaged 70 characters in length, consequent sentences 36 characters. Design
Subjects were assigned by rotation to groups, as in Experiment I. To avoid response inhibition, the same mouse buttons were used as before by each subject to indicate decisions, and the instructions remained to choose arrows which best corresponded to the meaning of the expressions shown, or which were contrary to their meaning. The time allowed for registering a decision was 3 or 1 0 seconds.
Downloaded from jos.oxfordjournals.org by guest on January 1, 2011
consisted of an antecedent sentence implying a low level of performance followed by a consequent describing a leading or winning state, or an antecedent sentence implying a high level of performance followed by a conse quent describing a trailing or losing position. To maintain coherence, the indi viduals' names within a pair of anti-oriented sentences were switched. The predicates in antecedent sentences (e.g. get votes , show capacity, have a grip on ) were ones for which directionality and evaluation correspond; as a scale value, or scale strength, increases, the subjective value of a performance should increase, and vice versa. These predicates were combined with plus de , presque and beaucoup de to imply high levels of performance, and with moins de, ne que, and peu de to imply lower levels. The level ofperformance implied in successive antecedent sentences was varied (high-low or low-high), resulting in four versions of each competition.
Robert J. Jarvella and Lira Lundquist 1 8 1
Results and discussion
Judgments made within the time limit allowed are summarized in Table 2. It can be seen there that, for the 3-second groups, largely complementary patterns of response were obtained with the two instructional sets, but also that 3 seconds was not enough time for subjects to read and give responses on long sentences. Allowing 10 seconds to register a response permitted subjects to respond to all sentences, but also revealed a lower degree of agreement on some sentences. Fast subjects evidently assigned arrows to sentences in the way we had predicted more often than slow subjects.
Corresponding meaning 3-sec limit Arrow Performance level implied in antecedent high low Position described in consequent leading/winning trailing/losing
I o-sec limit
Contrary meaning 3-sec limit
up
down
up
down
38 2
0 16
87 34
13 6s
0
I3 0
84
2 71
89 20
II 8o
6 33
47
2
up
down
Note : Items summing to less than umry reflect cases where responses were not gJVen Within the rime limit imposed.
Antecedent sentences which imply a high level of performance, and consequent sentences which describe a leading or winning position were predicted to be assigned an 'up'-arrow when tested with corresponding meaning instructions, but a 'down'-arrow when tested with contrary-meaning instructions. For responses given by the deadline {those shown in Table 2), this prediction was satisfied 99% of the time with corresponding-meaning instructions and the 3-second time limit, 85% of the time with contrary meaning instructions and the 3-second limit, and 88% of the time for corre sponding-meaning instructions and the 1 0-second limit. Similarly, we predicted that antecedent sentences implying a low level of performance, and consequent sentences describing a trailing or losing position, would be assigned a 'down'-arrow when tested with corresponding-meaning instructions, but an
Downloaded from jos.oxfordjournals.org by guest on January 1, 2011
2 Experiment 2 (arrows task). Percent of cases in which an arrow pointing upwards or downwards was taken ro correspond, or be contrary ro, a meaning of sentences drawn from rexrs about competitions
Table
1 82 Scalar Interpretation of Words, Sentences and Texts
'up'-arrow when tested with contrary-meaning instructions. For responses given by the deadline, these predictions were satisfied 96% of the time with both instructional sets using the 3-second limit, and 73% of the time using corresponding-meaning instructions and the r o-second limit. We also hypothesized greater speed in classifying sentences we took to be oriented upwards. It can be seen from Table 2 that more respdlnses were given with the 3-second limit on items theoretically oriented in this way. The 3-second time limit proved to be too short to permit evaluating this hypothesis in much detail, though a 2 X 2 analysis of variance of the latency data for consequent sentences revealed a main effect for both instructions (F(1 ,29) 8.03, MSe 0.145, p < o.o r ) and directionality (F(1 ,29) 1 0.88, MSe 0.052, p o.ooJ). Here, subjects responded faster with corresponding meaning instructions (2.0 vs. 2.3 seconds), and when interpreting expressions hypothetically oriented upwards (2.1 vs. 2.3 seconds). However, the hypothesis could be tested more fully on the data obtained with a ro-second limit, where more than 99% of responses had been given by the deadline. Data for these judgements are summarized in Table J, using a measure of reading time, namely milliseconds per character of text read, which corrects for differences in sentence length. The predicted effect of direc tionality on these latencies was strongly confirmed. Judgments were made faster for sentences which we hypothesized to be oriented upwards rather than downwards (58 vs. 70 ms/character overall, F( r , r 6) 20.6, MSe 870, p < o.oor). And this effect was reliable for antecedents and consequents separately (p < o.oos for each), with magnitudes of about 700 and soo ms respectively for full sentences. Our third hypothesis was not confirmed. Consequent sentences in co oriented pairs were not read systematically faster than consequents in anti oriented sentence pairs, though the difference was in the predicted direction =
=
=
=
=
=
3 Experiment 2 (arrows task). Time used (milliseconds/character) in assigning arrows to sentences theoretically oriented upwards or downwards (10-sec nme limit, corresponding-meaning instructions)
Table
Hypothetical orientation of sentences Anti-oriented
Co-oriented Sentence
upwards
downwards
upwards
downwards
Antecedent Consequent
57 6o
64 70
54 63
65 79
Note : For
anti-oriented pa�rs, the consequent follows an antecedent oriented m the oppoSite direction.
Downloaded from jos.oxfordjournals.org by guest on January 1, 2011
=
Robert J. Jarvella and Lira Lundquist 1 8 3
(65 vs. 7 I ms/character, amounting to some 200 ms per sentence, F(I , I 6) 3-38, MSe 723, p o.o8). The results of this experiment establish the link that we were seeking between speed in reading and the perception of directionality which metaphorically underlies the meaning of sentences in texts. As in the examples below, a high level of performance was interpreted in terms of upwards orientation, a low level of performance in terms of downwards orientation. Siinilarly, high and low relative standings in the contest were interpreted correspondingly: =
=
=
Both antecedent and consequent sentences which were oriented towards positive outcomes were interpreted faster. Since, in antecedent sentences studied, quantity and evaluation corresponded (vs. in a case like George made nearly 1 o errors), both these factors may have contributed to the results obtained here. The parallel results obtained for consequent sentences, however, suggest that evaluation was a stronger determinant in the making ofjudgments.
EXPE R I ME N T 3 : I N FE RE N CE I N TEXT A speed effect that was anticipated but not found to be reliable in Experiment 2 was that of co- vs. anti-orientation. Such an effect has been found previously for both inferences and for sentence reading times. The topos concept as we have elaborated it makes predictions that different inferences about reference will be exploited when readers encounter co- vs. anti-oriented sentence pairs; that for co-oriented pairs, inferences involving one individual will be used (e.g. (the more votes X gets, the more likely X will win)), but that for anti-oriented pairs, inferences involving different actors will be exploited (e.g. (the less votes X gets, the more likely Y will win)). Inference schemes activated by scalar predications were approached in our previous workpartly by asking readers to interpret definite noun phrases in texts. For example, one text studied was about a job opening for an executive secretary, in which two candidates were invited for an interview. Mter a lead-in identifying both candidates by name, there followed two antecedent-consequent sentence pairs. Each consequent sentence began with a definite NP whose referent was ambiguous but which tended to be disambiguated by the sentence's predicate. In one version, the text (in translation) continued: (k) En stenographie anglaise, Sandrine Barra ne put ecrire que 6 5 mots minute.
a
la
Downloaded from jos.oxfordjournals.org by guest on January 1, 2011
(10) George got almost 5,000 votes. He suRpassed all his competitors. (I I) George got only 5,000 votes. He was on the BOTIOM of the list.
1 84 Scalar Interpretation of Words, Sentences and Texts
(In English stenography, Sandra Barra was able to write only 65 words per minute) (l) II semblait que la petite Grenobloise allait obtenir le poste. (It seemed that the small woman from Grenoble was going to get the job) (m) En comptabilite informatique, Laura Ursini montra beaucoup de capacites pour le programme de gestion. (In computer accounting, Laura Orsini showed a great deal of capacity for using the administrative program) (n) La secretaire frisee du L.E.P. Lecourbe obtint le poste. (The curly-haired secretary from L.E.P. Lecourbe won the post)
Method Materials
Four versions of texts studied previously in Danish and Finnish were adapted into French. Each text narrated a competitive event for individuals, in which the chief characters described were invented. The job competition text as adapted is shown in Table 4· Two competitors were identified by name early in each text as being favorites to win, or having the highest previous merits. The central portion of
Downloaded from jos.oxfordjournals.org by guest on January 1, 2011
As in Experiment 2, the predicate in a consequent sentence was co- or anti oriented with that of its antecedent. In the case of anti-predication, as in (k)-(1), the ambiguous NP was interpreted as referring to the individual named in the antecedent no more than 10-25% of the time. In the case of co-predication, as in (m)-(n), the ambiguous NP was interpreted as referring to the individual named in the antecedent about 95% of the time. In addition, referents of the NPs in anti oriented predications were rated as less clear, and texts containing them rated as being less comprehensible. We interpreted these results, and some correspond ing ones for latencies, as suggesting that a 'simple' topos, i.e. an inference scheme whose scales are oriented in the same direction ((the more, the more) or (the less, the less)), is activated first and preferred, but that it will give way to use ofa more complex inference scheme if it leads to incoherence in text. The present experiment was a replication in French of part of that previous work. On the basis of the topos concept and the results of the prior studies, we hypothesized that the interpretation of an ambiguous definite NP as referring or not referring to a just-named individual would be a function of whether the sentence pair containing the two expressions was co- or anti-oriented. Secondly, we predicted that referents for such ambiguous NPs would be judged as being less clear in anti-oriented pairs, and that texts containing anti-oriented pairs would be judged to be less comprehensible.
Robert J. Jarvella and Lira Lundquist 1 85
4 Experiment 3 (reference resolution). One of rhe eight French competition texts srudied
Table
Note: In constructing verstons of texts, rhe first antecedent and first consequent sentence were combined in all four ways. In the second antecedent sentence, the onenrarion chosen Implied was the opposite of that imphed 10 the first antecedent.
the text was then manipulated against this background. Table 4 illustrates how the four versions of passages studied were constrUcted, each based around two antecedent-consequent sentence pairs, which could be both co-oriented, both anti-oriented, or one of each (in both orders). The predications studied were quite heterogeneous and included both quantitative predicates, as in Table 4, and more qualitative predicates, as in (c)-U) above. Antecedent sentences in pairs averaged S s characters in length, consequent sentences 54 characters. Subjects and procedure
Twenty new speakers of French studying at the University of Lyon served as subjects. Each subject read one version of all eight texts, and two texts in each of the four forms studied. The version of each text assigned to a subject was determined by rotation. The sequence of competitions described was the same for all subjects. Subjects were told they were taking part in a study of text understanding, and would be asked to read some short texts and answer a few questions about them. Each text was typed on a separate page, and followed by three questions. The first question asked 'Quel est le degre de comprehensibilite du texte?' and called on the subject to rate the text on a scale of 1 ('peu comprehensible') to s ('tres comprehensible'). The second and third questions called for the subject to
Downloaded from jos.oxfordjournals.org by guest on January 1, 2011
L'enrretien Dans une annonce d'offre d'emploi d'un journal du dimanche, on recherchair une secretaire experimenree pour le directeur d'une enrreprise. Sur les s 7 demandeurs d'emploi, deux furent convoquees pour un enrretien et des tests. Les deux candidates s'appelaienr Sandrine Barra er Laura Ursini. En srenographie anglaise, Sandrine Barra fur capable d'ecrire presque/ne put ecrire que 65 mots a Ia minute. II semblair que Ia petite Grenobloise allair obrenir/perdre le posre. En comptabilire informatique, Laura Ursini monrra beaucoup de/ne montra que peu de capacires pour le programme de gestion. La secreraire frisee du L.E.P. Lecourbe obtint le poste. Apres cet enrretien epuisant, son mari l'invira :i diner dans un des meilleurs restaurants de Paris.
1 86 Scalar Interpretation of Words, Sentences and Texts
indicate who was referred to by two definite NPs in the text, and to rate the clarity of this reference, again on a s-point scale. These judgments were obtained by constructing constituent questions based on each of the conse quent sentences, with the definite NPs replaced by interrogative pronouns (e.g. 'A votre avis, qui obtint le paste?). The instruction 'ecrivez le nom' was added following each question. Results and discussion Reference resolution
Table 5 Experiments 3-4 (reference resolution). Percentage of unified and disjoint readings of a name and ambiguous definite NP in co- and anti-oriented sentence pairs
Study
Hypothetical orientation in sentences
Reading Experiment 3 Experiment 4 with scalar adverbs without scalar adverbs
Co-oriented
Anti-oriented
Unified
Unified
Disjoint
18
78
20 34
71 55
Disjoint
96 85 81
7 10
Nor�· Sums shown are less chan uniry because of cases where another or n o response was given.
Downloaded from jos.oxfordjournals.org by guest on January 1, 2011
In 3 1 2 of 3 20 answers given, subjects identified one of the favored competitors from a narrative by name. These responses were scored as signaling a unified or a disjoint reading: in a unified reading, the definite NP in a consequent was interpreted as referring to the individual named in the just-prior antecedent; in a disjoint reading, it was interpreted as referring to the other individual named earlier in the text. In the few remaining cases here and corresponding ones in Experiment 4 below, subjects copied a definite NP from the text, wrote comments such as )e ne sais pas', 'aucune idee', or 'rrop obscur', or made no response. Table s shows the percentage of unified and disjoint readings obtained for NPs in co- and anti-oriented consequent sentences. It can be seen there that disTributions of responses go quite strongly in one direction or the other. Some 96% of readings given to NPs in co-oriented sentence pairs were unified, whereas 78% of readings given to the same NPs in anti-oriented sentence pairs were disjoint. In the two conditions, subjects thus interpreted the ambiguous expressions in essentially opposite ways. (The two percentages do differ,
Robert J. Jarvella and Lira Lundquist 1 87
however, F(I ,I 9} - 1 3.36, MSe = o.o68, p. < o.oo2.) These findings resemble ones found in the same setting previously: in text context, readers appear to use scalar inference schemes to deduce the referents of expressions which may otherwise be ambiguous. Ratings of clarity and comprehensibility
Table 6 Experiments 3-4 (reference resolution). Mean rated clarity of reference
and comprehensibility for the texts studied
Study Rating
Sentence pairs present in text Two co-oriented patrS
One co- and one anti-oriented pair
Two anti-oriented pair
Experiment 3 clarity comprehensibility Experiment 4
4· • 4·2
3·6 3·7
3· 1 3·3
clarity comprehensibility
3·9 4·0
clarity comprehensibility
3·5 3·5
with scalar adverbs 3·5 3·6 3-4 3-7 without scalar adverbs 2.8 3· 1 2.8 3·0
Downloaded from jos.oxfordjournals.org by guest on January 1, 2011
Subjects also rated referents as being less clear when the ambiguous NPs appeared in anti-oriented sentence pairs (mean = 3.2 vs. 3·9 on the five-point scale, F(I , 1 9) = 24.91 , MSe = 1 .46, p < o.oo1 ). Clarity of reference in texts as a whole was rated lowest when both sentence pairs were anti-oriented, was rated intermediate when one pair was anti-oriented and the other co-oriented, and was rated highest when both sentence pairs were co-oriented (F(2,38} = 10.78, MSe = 1 .8 1 , p < o.oo1). Comprehensibility ratings showed a pattern similar to ratings of clarity (F(2,3 8} = 9·93. MSe = 0.909, p < o.oo 1 ; see Table 6). The results in French for clarity of reference and text comprehensibility also resemble findings obtained for these text factors in the earlier work (Lundquist, Jarvella, & Hyona 1994).
1 88 Scalar Interpretation of Words, Sentences and Texts
EXPE R I MENT 4 : S CALAR ADVERBS A N D I N FE RENCE
·
Method Materials
The experimental material was prepared in two versions, one identical to that used in Experiment 3 and the other with scalar adverbs (plus de , moins de , presque, ne que, etc.) removed from antecedent sentences in each text. Thus, if the Experiment 3 version of a sentence read 'ne put ecrire que ((or) put ecrire presque) 6s mots a la minute', the new version read 'put ecrire 6s mots a la minute', if the prior version read 'marqua des points chez plus de la moitie des juges', the new version read 'marqua des points chez Ia moire des juges', and so on. Because the text was predicted to become less clear without the adverbs, we refer to co- and anti-orientation below in terms of the original, Experiment 3 version of the material. Subjects and procedure
The subjects were 36 new speakers of French, enrolled in courses at the University of Quebec at Montreal. The same instructions and procedure were used as in Experiment 3·
Downloaded from jos.oxfordjournals.org by guest on January 1, 2011
Experiment 3 replicated our previous findings in Danish and Finnish concerning both use of scalar inferences to resolve ambiguity in text, and the effects of co and anti-predication on the perceived coherence of text. The present study was aimed at raking a further step, by asking to what extent the results of Experi ment 3 were actually brought about by the scalar adverbs used to modify predicates in our test sentences there. Previous empirical work (Lundquist 1 987b) does suggest that the presence of expressions like plus de , presque, mains de, and ne que in predicates does have a disambiguating effect. To answer chis question, we performed another study on disambiguation, using two sets of textual material rather than one. One set of this material was that used in Experiment 3· In the other set, the scalar adverbs in test sentences were simply removed. If the adverbs have the disambiguating effect we hypothesize for them, their absence should lead to less uniform unified and disjoint interpretations of NPs, to lower ratings of clarity of reference, and to lower ratings of comprehensibility.
Robert J. Jarvella and Lita Lundquist 1 89
Results and discussion Reference resolution
=
Clarity and comprehensibility
Referents for definite NPs were also rated as being less clear without scalar adverbs in antecedent sentences (mean = 3.1 vs. 3.6 on the 5-point scale, F(1 ,3 5) = 4.3 1 , MSe = 6.40, p < o.o5). Likewise, text comprehensibility was rated as being lower without the adverbs present (mean = 3.2 vs. 3·7 on the corresponding scale, F(1,35) = 8.55, MSe = 2.44, p = o.oo6). As in Experiment 3, there was an effect of number of anti-oriented sentence pairs in texts on both clarity ratings (F(2,70 = 5.66, MSe = 1 .96, p = o.oo5) and on ratings of comprehensibility (F(2,70) = 8.88, MSe = o.8 5, p < o.oo 1). Texts containing two co-oriented pairs were rated as being most clear and comprehensible, while texts originally having two anti-oriented pairs were rated as being least clear and least comprehensible. As can be seen in Table 6, however, with adverbs removed from texts, clarity and comprehensibility ratings fell for all versions of texts smdied. The results of this smdy show that the bare predicates in antecedent sentences in our texts were not generally sufficient for interpreting ambiguous NPs found in them. Scalar adverbs had the predicted disambiguating influence. The effect was statistically reliable in anti-oriented sentence pairs, and the adverbs had a further effect of enhancing clarity of reference and text comprehensibility generally. Like Experiment 3, Experiment 4 supports the notion that inference schemes issuing from use of scalar expressions are used to draw deductions about who is being talked about in text. How readers resolve
Downloaded from jos.oxfordjournals.org by guest on January 1, 2011
In 525 of 576 answers given, subjects identified one of the favored competitors from a narrative by name. As previously, these responses were scored as signaling a unified or a disjoint reading. In Table 5, we summarize the response distributions obtained with and without scalar adverbs. The trends in the data were somewhat less sharp than in Experiment 3, perhaps because the smdents tested here were less advanced. However, removing scalar adverbs from antecedent sentences did lead readers to make fewer (68% vs. 78%) interpretations of ambiguous NPs in consequents which could be explained by orientation (F(1 ,35) 5.1 8, MSe = 0.273, p = 0.029). As shown in Table 5, the absence of adverbs had a neutralizing effect especially in the percentage of disjoint readings (55% vs. 7 1 % responses predicted by orientation, p < o.o1). For sentence pairs which were co-oriented in the original stimulus material, the difference observed in unified readings made without vs. with adverbs (8 1 % vs. 85%) was small and not reliable.
1 90 Scalar Interpretation ofWords, Sentences and Texts
ambiguities seems to tum on whether predications are co- or anti-oriented. The data showed further that readers feel more confident about an interpretation when sentences are oriented in the same direction. E XPE RIMENT
s:
S E NTE N C E READ I N G T I ME
Method Materials
The srimulus material was the set of eight narrative texts written about competitions and studied in Experiment 3· With respect to the sequences studied in Experiment 2, the present texts were different in a number of respects. Each text began with a title and introduction, continued with two pairs of antecedent-consequent sentence pairs, and ended in a conclusion. Both antecedent and consequent sentences were somewhat longer, different names were used for the characters described, and consequent sentences included referentially ambiguous definite NPs. Passages were shown to subjects in a single order, preceded by one practice item. Subjects, procedure, and design
The subjects were 40 speakers ofFrench studying at the University ofMontreal, who had earlier taken part in Experiments 1 and 2. Each subject read one version of each text, two texts containing two co-oriented sentence pairs, two containing two anti-oriented pairs, and four contained one co- and one anti oriented sentence pair. The versions of texts to be read by subjects were assigned by rotation. The experiment was conducted using a personal computer. The text read was advanced by pushing the space bar on the keyboard, and the time between a
Downloaded from jos.oxfordjournals.org by guest on January 1, 2011
In Experiment s, the final study we report, we return co a pair of earlier issues. First, are sentences which are oriented upwards, imply good performances and report posirive results, read and understood faster than ones oriented downwards? This is what our previous . research suggested, as well as Experiment 2. And, second, is a sentence read and understood faster when it is co-oriented with an antecedent than when it is anti-oriented? This is what was found previously in Finnish and Danish, but the results ofExperiment 2 were inconclusive on this issue. A rimed reading task was used to ask these quesrions. Subjects read the narratives studied in Experiment 3 one sentence at a time, advancing the text seen manually.
Roherr J. Jarvella and Lira Lundquist I 9 I
Results and discussion
The two theoretical questions of interest were evaluated from differences in reading times, tested using planned comparisons following a general analysis of variance. To correct for length differences between antecedent and consequent sentences, as in Experiment 2, time spent per character of text read was used as the dependent variable. We also report the size of effects found per sentence read. Orientation and antecedent sentences
Antecedent sentences in text whose meaning we take to be metaphorically directed upwards were read faster than sentences whose meaning is directed downwards. As Table 7 shows, this prediction was confirmed for both of the antecedent sentences separately in texts. For the first antecedent sentence, a difference of almost 10 ms per character was obtained in the direction predicted (F(1 ,39) 1 4.26, MSe 5 34, p < 0.001 ). For the second antecedent sentence, the corresponding difference was about 9 ms per character (F(1 ,39) - 1 1 .63, MSe 5 16, p o.oo2). For full sentences, these effects are on the order of 8oo ms. For the first consequent sentence in texts, which described a leading or a trailing state, there was also a systematic 9 rns per character effect =
=
=
=
Downloaded from jos.oxfordjournals.org by guest on January 1, 2011
reader's key presses was recorded. Subjects were told that the study concerned text comprehension; they would be asked to read a series of short narrative texts sentence by sentence. Texts were to be read normally, but with care, since it would not be possible to look back once a sentence was left. When the reader had understood a sentence, she was to push the space bar to make the next sentence appear. Following each text, a yes/no question of the form 'Est-ce que X gagna?' ('Did X win?') was asked, where X was the name of one of the two favorites in the competitive event that had just been described. The text was displayed in full together with this question, in case the reader wished to check it before responding. The reader answered the question by pushing one of two keys marked 'oui' and 'non'. The contestant named in the question was varied, so that half the questions were plausibly answered in each way. 'Oui' responses were given to 89% of questions asking if the contestant who had been named in a sentence implying a high level of performance had won, while 'non' responses were given to 87% of questions asking if the contestant who had been named in a sentence implying a low level of performance had won. Thus, the text generally appeared to be interpreted by readers using the same kind of inferences supported in Experiment 3·
I 92 Scalar Interpretation of Words, Sentences and Texts
7 Experiment s (reading texts sentence by sentence). Time used (milliseconds/character) to read sentences hypothetically oriented upwards or downwards
Table
Type of sentence
Theoretical orientation upwards 78
87
9I 97
94 III
67
75
88 90
n/a n/a
(soo ms for the sentence as a whole) favoring the sentence with the more positive orientation (F(I ,39) 8.84, MSe 70I , p o.oos). =
=
=
Co- vs. anti-orientation of consequent sentences
Numerically, there was a tendency for co-oriented consequent sentences to be read faster than anti-oriented sentences. The difference amounted to about 6 ms/character overall, equivalent to 300 ms per sentence. As in Experiment 2, however, the difference failed to reach significance overall (F( I ,39) - J.OI, MSe I 9JS, p 0.09). Inspection ofTable 6 reveals that the difference found was limited to the initial consequent sentences encountered in a text, where it was highly significant (I I ms/character, F(I ,39) - 8.s8, MSe - I I 6 I , p o.oo6). In the second consequent, it amounted to just 2 ms/character (F < 1 .0). In the first consequent sentence, the disadvantage of anti-orientation was also larger for a consequent describing a trailing than a leading state (I 7 vs. 6 ms/character). This pattern is reminiscent of one found in Experiment 2. Finally, consequent sentences were relatively more time-consuming to read. As in Experiment 2, there was a general slowing down observed in reading a consequent sentence vs. its antecedent (F(I ,39) - 68.74, MSe - I 368, p < o.oo I ). In Experiment 2, the slowing-down effect amounted to about 8 ms per character, but here it was I 7 ms per character, and thus more than twice as large. We believe that the doubling in size of the effect observed here rests mainly in the process of disambiguating the definite NPs present in consequent sentences. =
=
=
Downloaded from jos.oxfordjournals.org by guest on January 1, 2011
First sentence pair Antecedent Consequent co-oriented with antecedent anti-oriented with antecedent Second sentence pair Antecedent Consequent co-oriented with antecedent anti-oriented with antecedent
downwards
RoberrJ.Jarvella and Lira Lundquist 1 93
GENERAL D I S C U S S I O N
Downloaded from jos.oxfordjournals.org by guest on January 1, 2011
We have suggested previously that scales which underly predications in sentences often are oriented, or directed, upwards or downwards, and that even scalar expressions of a more abstract kind, such as those in obtained almost x points/was winning and obtained only x points/was losing , are taken meta phorically as pointing up and down respectively. In the present paper, we have pursued the issue of scalarity in language by taking up the processing of some directional aspects of meaning on the word and sentence levels, as well as on the text level. The results suggest that directionality in language in terms of oriented scales may intervene in reading on all three linguistic levels. On the word level, we saw in Experiment 1 that subjects attributed a set of French expressions studied with an upwards- or downwards-oriented meaning predicted for them in more than 95% of cases. In general, expressions which describe aboveness, or represent or intensify large scale values or quantities, a positive polarity or a positive result were judged as having a meaning pointing up, while expressions which describe belowness or represent or intensify small scale values or quantities or a negative polarity or result, werejudged as having a meaning pointing down. This interpretation emerged when subjects were instructed to choose an arrow which best symbolized meaning, and when they were instructed to choose an arrow which was contrary to meaning. Expressions which are oriented only in a metaphorical sense were classified as consistently as other expressions in the task. And adverbs which, in Ducrot & Anscombre's theory of scalarity in language, are taken to activate scales pointing upwards on a scale, towards the same conclusion as p, and downwards, towards the conclusion of not p, were interpreted with this directionality. A further metaphoricaVfigurative interpretation of scalarity was observed on the sentence level. In Experiment 2, sentences containing scalar predicates of the type obtains almost I more than x points and predicates describing a leading or a winning state were perceived as being oriented upwards (8 5-99% of cases predicted, depending on experimental treatment), whereas sentences con taining scalar predicates of the type obtains only I less than x points and predicates describing a training-losing state were perceived as being oriented downwards (7 3-96% ofcases predicted). It was also found that sentences we considered to be directed upwards were evaluated faster than downwards-oriented ones (effects on the order of 500-700 ms were obtained). Our hypothesis that a sentence co oriented with an antecedent sentence would be read and evaluated faster than one not sharing its orientation was only weakly supported, however. A 200 ms per sentence tendency observed in this direction was not statistically reliable. In Experiments 3 and 4, concerned with text interpretation, we were able to replicate findings for Danish and Finnish showing that use of scalar predicates
1 94 Scalar Interpretation of Words, Sentences and Texts
(the more (the less) an object 0 has the property P, the more (the less) an object 0' will have the property P') are statements about linguistic as well as extralinguistic knowledge which is recruited when we encounter certain expressions in text Scalar adverbs, in particular, seem systematically to instruct the reader about how to use context
Downloaded from jos.oxfordjournals.org by guest on January 1, 2011
in text gives way to a specific type of scalar inference, which has a profound effect in contexts of referential ambiguity. In Experiment 3, the idea that co oriented scalar predications lead to a unified reading of an ambiguous definite NP in a consequent sentence, but anti-oriented predications lead to a disjoint reading, was confirmed in 96% of cases for co-orientation, and 78% of cases for anti-orientation. The referents of the NPs were also rated as being clearer in co than in anti-oriented predications, and texts containing co-oriented predica tions were rated as being more comprehensible. In Experiment 4, it was further shown chat interpretation of the NPs' referents was more uniform, and ratings of their clarity and of text comprehensibility were higher, when predications in antecedent sentences contained scalar adverbs than when they did not. Finally, in Experiment s, we showed that reading time is a function of the orientation of sentences in text. As found in our previous work in other languages, in French, sentences containing predications of the type obtains almost/more than x points were read faster (here by 8oo ms) than sentences having predications of the type obtains only/less than x points. There was also a significant advantage (of soo ms) found when a consequent sentence read described a leading or winning over a trailing or losing position in a competition. However, again only a nonsignificant tendency was found (but with a magnitude of 300 ms) for a consequent sentence to be read faster when it was co-oriented with the antecedent. Basic scales found in language probably have an intrinsic orientation, founded partly in perception. What may be a bit special about the scales studied in the later experiments reported here is that they are oriented upwards or downwards only in a figurative sense. To explain results found for sentences in terms of verticality, some appeal needs to be made to the nature of metaphor in language interpretation generally. Work in the field of cognitive semantics has identified the origin of metaphor in schemas embodied in our visual and somatic perception. Other work concerned with the consequences for language of our having a forward, upwards-oriented field of vision (e.g. Clark 1 976), has shown that this basic orientation should not be underestimated. 'Up'-scales and 'down'-scales seem to be fairly naturally associated with the quantities and qualities, events and processes, represented in our stimulus material, and used here in reasoning from text. What about the nature of the inferences made? Inferential schemes of the topos type and having the form
Robert J. Jarvella and Lita Lundquist I 95
Downloaded from jos.oxfordjournals.org by guest on January 1, 2011
in order to make inferences, in a way which is predicted by the topos theory {Experiments J, 4, and s). A further linguistic explanation {Konig 1991) may be that, in explicitly focusing on a particular point on a scale, and thereby conveying a direction and an evaluation, scalar adverbs implicitly, via the scope of the sentence, refer to alternatives on the scale, thus giving more detailed instruction about how to draw on background knowledge. To return to our election examples at the outset of this paper, if we are told that George Brown got only s,ooo votes, it is fair to conclude that his chances of being elected were low. Having a scalar means for representing the plausible relations between such events is an advantage for a theory which aims to explain data such as those reported in Experiments J, 4, and S· For our purposes, it does not matter whether the inferences that are activated and exploited are interpreted linguistically as presuppositions, entailments, or implicatures. What we wanted to show is that they are instrumental in the processing of actual texts. The question as to just how such inferences are activated and exploited is a more difficult one to answer. Thejudgments of reference made were made very systematically. Data we have reported both earlier and in the present studies also suggest that the inferences used are activated fairly quickly, resulting in reference resolution taking place while subjects are still reading a text, but at the same time that the inferences may be less than fully automatic, especially in the case of anti-oriented sentence pairs. The understanding of a text describing a competition scenario could perhaps be characterized in the following way. Competition texts are organized around a main theme, winning the context, which is associated with an upwards-oriented scale, but which dominates several subordinated themes developed in different subparts of the contest, themselves correlated with the events reported there. We enyisage that the reader constructs a model of the situation described in the text for each specific subpart of the competition. In this model, the reader situates a performer who is mentioned by name on an upwards- or downwards-oriented scale according to whether the level of performance described is implied to be high or low. This is the product of reading an antecedent sentence such as those studied here. When, in a consequent sentence, a definite NP is used to describe someone as leading or trailing, this event is represented on a second, oriented scale. The reader then examines her model using a competition topos to see whether the person situated on the first scale was placed in a position there which is con sistent with the outcome reported. If position on the two scales corresponds, an ambiguous expression is interpreted as referring to the first individual. This explains why a consequent sentence which is co-oriented with its antecedent led here and previously to a unified reading of a name and a definite NP. If, on the other hand, the position of the person named in an antecedent is inconsistent with the outcome reported in the consequent, to resolve the reference the reader generally checks
196 Scalar Interpretation of Words, Sentences and Texts
Acknowledgements The work reported was supported in part by the Natural Sciences and Engineering Research Council of Canada (Universite de Montreal general fund), the Swedish Research Council in the Humanities and Social Sciences, and by the Joint Committee of the Nordic Social Science Research Councils. We are grateful to Carsten Elbro, Jan Gabrielsson, Catherine Kerbrat Orecchioni, Madeleine St. Pierre, and Suzie Mathieu for assistance at various stages of this work, and to two anonymous readers for their comments on a previous version of this paper.
Addressfor correspondence LITA LUNDQUIST Copenhagen Business School Dalgas have 15 DK-zooo Frederiksberg Denmark
Received: 1 6.1 2.93 Revised version received: 1 0.03.94
Downloaded from jos.oxfordjournals.org by guest on January 1, 2011
to see if another character, mentioned earlier in the text, is in an appropriate position. If she finds one, she interprets the ambiguous expression as referring to that individual instead. This explains why anti-orientation between an antecedent and consequent sentence leads to predominantly disjoint readings of a name and definite NP, and it can account for why a consequent sentence which is anti-oriented with its antecedent can take longer to understand. In the present studies of French, as in the previous ones of Danish and Finnish, we found abundant evidence that reference assignment occurs for ambiguous expressions with the predicted results, and also that sentences oriented upwards are read faster. Previously, consequent sentences were also found systematically to be read faster when they were co-oriented with their antecedents. The present data were somewhat more equivocal on this point. There seemed to be some advantage with co-orientation, but this was over shadowed by a solid effect also favoring reading of consequent sentences that were oriented upwards. In general, the results for latencies also show more points of convergence than divergence. In all three languages, similar effects on reading time emerged for the orientation of individual sentences, and there was a tendency for consequent sentences which were co-oriented, and shared reference, with their antecedents to be read faster. Thus, it is not the presence of effects, but rather their relative magnitude and stability, which varied somewhat. The main effects appear to be quite language-independent. One explanation for the reduced speed advantage found here in reading sentences that are co-oriented with their antecedents may be that written French shows a stronger aversion to making repeated reference in text using the same expression. As a consequence, readers of Fr.ench may be more used to dealing with anaphors whose understanding rests on background knowledge, including that made available by topos. Further research which compares the use of such inference schemes in the process of understanding is motivated.
Robert]. Jarvella and Lira Lundquist I 97
N O TE I
French presque and ne que, as well as plus de and moins de , bur also other similar expressions nor studied here, go under various names, including 'argumentative operators' (Ducrot I 982: I s I ), 'quantity
modifiers' (Ducrot & Anscombre I 983: 20), 'scalar adverbs' Uayez I 98J), and 'intensity modifiers' (Champaud & Bas sana I 987).
RE FERE NCES Johnson, M { I 987), The Body in the Mind: The Bodily Basis of Meaning, Imagination and Reason , University ofChicago Press, Chic ago. Konig, E. ( I 99 I), The Meaning ofFocus Particles: A Comparative Approach , Routledge, Lon don. Lakoff, G. (1987), Women, Fire, and Dangerous Things: What Categories Reveal about the Mind, University of Chicago Press, Chi cago. Lakoff, G. & Johnson, M. {I98o), Metaphors We Live By, University of Chicago Press, Chicago. Langacker, R W. {I 987), FoundationsofCogni tive Grammar, Stanford University Press, Stanford. Lundquist, L. {I987a), 'Programme argumen tatif er desambigulsarion referentielle', Revue Romane, 22, I 6 3-8 1 . Lundquist, L . { I 987b), 'Differents types d'inferences mis en reuvre dans Ia resolu tion de l'ambigiiire referentielle', Psychol ogie Franfaise, JJ, 289-95. Lundquist, L. (I 990), 'Argumentation: Semantics or Pragmatics?' in L. Lundquist & L. Schack-Rasmussen (eds), Pragmatics and Its Manifestations in Language, Copen hagen Studies in Language, no. I J, Arnold Busck, Nyt Nordisk Forlag, Copenhagen. Lundquist, L. & Jarvella, R ]. { I 994). 'Ups and downs in scalar inference', Journal of Semantics, I I, 3 3-s 3· Lundquist, L., Jarvella, R J.. & Hyona, ]. {1994), 'Text, topos, and mental models', co appear in Discourse Processes.
Downloaded from jos.oxfordjournals.org by guest on January 1, 2011
Anscombre, J.-C. {I 98 s). 'Theorie d e !'argu mentation, ropol er structuration argu Revue de Linguistique mentative', Que'bicoise , 18, I 3-6). Anscombre, J.-C. & Ducrot, 0 . {I98 3). L 'Argumentation dans Ia langue, Mardaga, Liege. Bassano, D. {I99I), 'Operareurs et connec teurs argumentatifs: une approche psycho linguistique', Intellectica , I, I I , I 49-91. Biihler, K. {I 934). Sprachtheorie: Die Darstel lungifunktion der Sprache, Fischer, Jena. Champaud, C. & Bassano, D. {I987), 'Argu mentative and informative functions of French intensity modifiers "presque" (almost), "a peine" Gust, barely) and "a peu pres" (about): an experimental study of children and adults', Cahiers de Psychologie Cognitive, 7, 6os-3 1 . Clark, H . H . { I 976), Semantics and Comprehen sion , Mouton, The Hague. Ducrot, 0. ( I 982), 'Notes sur !'argumentation er !'acre d'argumenter', Cahiers de Linguis tique Franfaise, 4, I 43-6 3· Ducrot, 0. (I 988), 'Topol er formes ropiques', Bulletin d'Etudes de Linguistique Franfaise, no. 22, Tokyo. Fauconnier, G. ( I 98 s). Mental Spaces: Aspects of Meaning Construction in Natural Language, MIT Press, Cambridge, MA. Horn, L. R {I 989), A Natural History of Negation , University of Chicago Press, Chicago. Jayez, ]. { I 98J), 'Games, frames, and French cognitive adverbs', Cahiers de Linguistique Franfaise, 5, 247-70.
198 Scalar Interpretation of Words, Sentences and Texts Sadock,J. M. ( 1 98 1 ), 'Almost', in P. Cole (ed.), Radical Pragmatics, Academic Press, New York.
Wierzbicka, A. (1 986), 'Introduction', Special issue on particles, Journal ofPragmatics, 10, 5 1 <)- 34·
Downloaded from jos.oxfordjournals.org by guest on January 1, 2011
journal of&mantics 1 1:
© N.I.S. Foundanon (199{)
1 9')-2 1 .�
On Representations of Quantifiers D A V I D A. R O U T H University ofBristol
Abstract
1 I N T RO D U C T I O N The standard, logical quantifiers all, some and no ( ne) have been the subject of much psychological investigations in recent years (Evans, Newstead, & Byrne 1 993). By comparison, the ubiquitous and much more numerous, non-standard quantifiers (e.g. '!'any, most, Jew , a Jew) have suffered from relative neglect (Moxey & Sanford 1 993a). This is curious since participants in most forms of discourse and communication (e.g. weather forecasts, news reports, political propaganda, scientific reports, and everyday conversations) clearly require a fluent command of the semantic and pragmatic properties of an extensive range of quantifiers. In real life, much can hang on the felicitous use of quantifiers. For example, Edwina Currie MP was forced to resign as a UK Parliamentary Under-Secretary for Health in December 1 988 when she said that 'most egg production' was infected with salmonella (see Currie 1 990). In a Channel 4 TV interview (Dispatches , 2 5 October 1 989), she explained that: 'I did not intend to say "most eggs", and did not say "most eggs".' She added: 'If l had covered myself more carefully I would have said something like "many" or "some" or "a few".'
Downloaded from jos.oxfordjournals.org by guest on January 1, 2011
In an ongoing psychological investigation, using a multi-task, multi-model strategy, the similarity of the meanings of 20 narural language quantifiers was studied by means of a freesort task. The induced proximity measures were modelled using various forms of spatial, multidimensional scaling and non-spatial (discrete feature) rree models. The data were better represented by the latter class of model. It is suggested that there are important components of meaning beyond the continuous, internal, unidimensional scale of quantity (whether amount or proportion) hypothesized by several previous investigators. The utility and implications of the favoured categorical representation require further investigation.
zoo
On Representations of Quantifiers
M UL T I D I M E N S I O N A L S C A L I N G , TRE E S , A N D C L U S TE RS
M U L T I -TAS K , M U L T I - M O D E L STRATE GY Pioneering applications of clustering and multidimensional scaling techniques to semantic data were reported by Miller ( I 967, I 972) and by Fillenbaum & Rapoport ( I 97 I), amongst others. It is now a fairly commonplace idea that these differing methods of analysis tend to focus upon complementary aspects of local and global structure, so that it is at least prudent to analyse proximity data by both means (Kruskal 1 977; Shepard 1 980). However, as Johnson & Tversky (I984} showed in their critical analysis and
Downloaded from jos.oxfordjournals.org by guest on January 1, 2011
For both illustrative and substantive purposes, the work to be reported in the present paper concerns one selected study from a collection of psychological investigations into the meanings of natural language quantifiers. The general aim of this ongoing programme is to investigate people's perception, knowledge, and use of such quantifiers, using a multi-task, multi-model strategy (c£ Johnson & Tversky I 984}, which will be discussed later. By observing people's judgements about the similarity of the meanings of each possible pair of members of a set of objects (e.g. words, phrases, sentences, etc.), it is possible to investigate and represent the structure of the resulting proximity matrix so as to reflect the set's local and global semantic structure. This may be achieved by means of a number of computer-based methods which have been developed to fit a variety of models (Shepard I 98o). In general terms, the commonest models belong to one of two families, either spatial models or discrete network models. Perhaps the most widely used of the former is multidimensional scaling, in which the portrayal of objects as points in Euclidean space can facilitate an understanding of structure in terms of underlying dimensions. In particular, a two-dimensional space portrays the relations between point representations of the objects in the familiar form of a map, so that similarities between objects are inversely related to distances between points. For network models, the commonest form is a tree in which each object is represented by a terminal node (c£ a 'leaP on a tree) and the distance between objects is represented by the shortest path between the relevant nodes (found by traversing 'branches' bearing the 'leaves' of the tree). A hierarchical tree structure lends itself to an interpretation in terms of clusters of objects which may facilitate an understanding of similarity in terms of common and distinctive features.
David A. Routh 201
S C AL I N G A N D REPRE S E N T I N G Q U A NT I F IE R S A considerable body of psychometric work on the meaning of quantifiers has aimed to investigate mappings between quantifiers and (unidimensional) scales. The literature has been well reviewed by Moxey & Sanford (1993a) and need not be considered in derail here. But, for example, respondents have been required to indicate the amounts corresponding to quantifiers as a function of set size (Borges & Sawyers 1 974; Newstead, Pollard, & Riezebos 1 987) or
Downloaded from jos.oxfordjournals.org by guest on January 1, 2011
advocacy of the multi-task, multi-model strategy, it is now understood that matters go deeper than this. These authors studied perceived proximity relations between I 8 risks induced by several tasks. They found that the structure of pairwise similarity judgements was better represented by a tree (discrete feature) model, whereas the structure of their multivariate rating (dimensional evaluation) data was better accounted for by a geometric (spatial) model. Thus, whenever different tasks and contexts are used to study proximity relations between the entities in some particular domain, striking variation in the degree of compatibility beween formal, structural models and empirical proximity measures may well prove to be important and pervasive (Suppes, Krantz, Luce, & Tversky 1 989: chapter 1 4). Moreover, there is also evidence that set-theoretic (feature) representations may be more apt for many datasets from conceptual domains, whilst geometric (spatial) models may be more apt for many datasets from perceptual domains (Pruzansky, Tversky, & Carroll 1 982; Tversky & Hutchinson 1 986). More recently, techniques for eliciting and representing people's knowledge have continued to proliferate (Olson & Biolsi 1 99 1 ; Arabie & Hubert 1 992; van Mechelen, Hampton, Michalski, & Theuns 1 993). There has also been a substantial accumulation of evidence concerning the constructive and contingent character of many human preferences, judgements, and beliefs (Payne, Betttnan, & Johnson 1 992). Direct access to canonical representations by one method or another may be a highly desirable goal for the purposes of cognitive psychology and semantics (c£ Fillenbaum & Rapoport 1 971), but it now seems increasingly unlikely that this will ever be achieved. The unattainability of desirable forms of invariant representation by some direct means need not lead to a counsel of despair. Instead, it seems that we must be more prepared to be responsive to the complexities and limitations of human information-processing (c£ Payne et al. 1 992). All things considered, a multi-task, multi-model strategy allied to a sensitive task analysis would seem to offer a viable methodology to supplant any approach to psychological representation that rests upon the supposed feasibility of some form of direct access or revelation Gohnson & Tversky 1 984).
202 On Representations of Quantifiers
Downloaded from jos.oxfordjournals.org by guest on January 1, 2011
proportions as a function of prior expectations (Moxey & Sanford I 992, I 99 3 b). Contrary to Borges & Sawyers' conclusion, the obtained estimates for the scale values corresponding to quantifiers have proved to be context-sensitive. From the perspective of a multi-task, multi-model strategy, the major shortcoming of such work is that it has involved a univariate rather than a multivariate, dimensional evaluation task. The restriction to a single, if obvious aspect of quantifier meaning is potentially a self-denying ordinance. Indeed, by comparison with the study of risks (see Johnson & Tversky I 984), it is probably fair to claim that the relevant variables required for an adequate multivariate profile of quantifier meaning may not yet have been fully identified. Clearly, the recent interesting work on matters such as the control of attentional focus by quantifiers, prior expectation, and on the communicative function of quantifiers, should help to remedy matters (see Moxey & Sanford, I 993a). Next we should consider the status of another line of work which at first sight may appear to support a unidimensional conception of quantifier meaning, namely the work by Holyoak & Glass (I 978) using a recognition memory paradigm. These authors found that five quantifiers (all , many, some, a few , and none) yielded memory confusion matrices that were well fitted by the particular loglinear model which corresponds to a unidimensional version of Luce's (I963) biased choice model. They suggested that (episodic) memory traces for quantifiers might incorporate an analogue representation of amount. However, in our own work with the same paradigm, using several sets of quantifiers (e.g. all , many , half, several, a few , and none), we have been able to explore certain valuable correspondences between additive tree representations (Sattath & Tversky I 977) and loglinear models (Routh I 990). Without digressing into technical detail, as we shall see later (e.g. Figure 2), each object in an additive tree model is represented by an external node of the tree, object clusters by internal nodes, and the proximities between objects by the lengths of the paths joining them. In the event, our confusion data were more consistent with singular trees (which possess a single internal node) than with linear trees (in which all objects lie on a line). It was found that a singular tree, which arises when each object possesses only unique features, corresponds to Smith's ( I 968) multiplicative similarity model (MSM), a particular version of Luce's (1963) model in which inter-object similarity can be decomposed into contributions from the two objects (see Townsend I 978). 1t was also discovered that Holyoak & Glass's ( I 978) data are well fitted by the MSM model and a singular tree. To the extent that such a paradigm does enable one to capture aspects of the meaning of quantifiers, it appears that we should abandon the idea of a continuous, unidimensional scale of meaning in favour of a more categorical representation.
David A. Routh
203
F RE E S O R T T A S K
Downloaded from jos.oxfordjournals.org by guest on January 1, 2011
The aim of the present study was to construct and compare formal or psychological representations of the relative meanings of an extensive set of commonplace quantifiers, by modelling proximity data elicited in aJreesort task (Miller 1 969; Rosenberg 1 982). When a large set of objects must be judged, the more traditional method of pairwise comparisons can make prohibitive demands on a respondent's time and motivation. By comparison, the freesort method enables a respondent to make judgements about an entire set of objects fairly efficiently. Both tasks share the advantage that neither the investigator nor the respondent need specify any relevant dimensions or features to be used as a basis for making judgements. Respondents are free to select and attend to underlying attributes and dimensions as they see fit. Subsequent cluster or scaling analysis of the aggregate data from a large panel of respondents is a good way of attempting to capture some estimate of the range of attributes or dimensions associated with the objects, uncontaminated by an investigator's preconceptions (Rosenberg 1 982). Useful task analyses of the paired comparison and freesort techniques have been offered, respectively, by Johnson & Tversky (1 984) and Rosenberg (1 982). In the case of both tasks, it seems that respondents do have to compare the objects and tend to make use of subsets of available features. As Rosenberg has noted, a possible disadvantage of 'single shot' freesort tasks is that respondents may be led to ignore or suppress an obvious dimension (or feature contrast). In contrast, a dimensional evaluation task restricts respondents to a fixed set of global dimensions, without necessarily invoking any direct comparisons of the objects (Johnson & Tversky 1 984). For present purposes, a freesort task appeared to be advantageous since by requiring respondents to compare quantifiers it becomes more likely that non-global components of meaning will become more salient. None the less, it is notable that Reyna (1981) reported an apparently successful recovery of a unidimensional representation when using a freesort task to investigate the meanings of modal adjectives (e.g. possible and probable). Technically, it is possible that the low values of stress she reported may have reflected some degree of degeneracy in the solutions, with objects tending to collapse into cluster (c£ Shepard 1 974). Unfortunately, it is not possible to resolve this point with access to the relevant 'Shepard diagrams' (i.e. plots of the rank transformed proxirnities versus the solution distances). Degeneracy would be evident if a Shepard diagram contained a small number of discontinuous steps. The position would have been less uncertain if Reyna had carried out some form ofcomplementary cluster analysis. As indicated earlier, it is a highly desirable practice to use both scaling and clustering (c£ Shepard, 1 980) and this policy was adopted in the present study.
204 On Represenrarions of Quanrifiers
P RE S E N T S T U D Y
Method Respondents
These were I 29 undergraduates from the University ofBristol, participating in a First year Practical Class in Psychology for a course credit. Their median age was I9 and the ratio of male to female respondents was approximately I : 2.
Downloaded from jos.oxfordjournals.org by guest on January 1, 2011
Two other unpublished freesort studies oflarge sets o f quantifiers preceded the present study. One of these involved the same set of 20 quantifiers as those detailed later (see 'Method') and so included the comparatives more and fewer. However, it simply investigated the meanings of isolated quantifier words with and without a partitive 'or, excluding any phrasal frame or reference to class size. The second study excluded the comparatives and investigated the meanings of I 2 isolated quantifiers (all , majority , most, many, lots , several, some, half, a few, minority, few, and none). In neither case did a multidimensional scaling analysis reveal any obvious global, amount, or proportion dimension of meaning. This outcome is important for dealing with the potential objection that the presence of comparatives could have a strong contextual effect, rendering respondents unable to use a strictly unidimensional representation of meaning, even if they wished to. Moreover, it was found that the freesort data were fit rather better by certain tree (discrete feature) models (ADDTREE, Sattath & Tversky I 977; EXTREE, Corter & Tversky I 986) than a multi dimensional scaling (MDS) model (Kruskal & Wish I 978). The present study also investigated the influence of the 'set size being referred to' upon the meanings of the quantifiers (c£ Borges & Sawyers 1 974; Newstead et al. I 987). This had the advantage that analyses based upon multiple proximity matrices could be entertained. For example, the INDSCAL version of an MDS model is applicable to multiple matrices and yields non-rotatable (and often psychologically interpretable) dimensions (unlike the standard MDS analysis of a single matrix). Also, any effects of size ought to be captured by variations in the weights associated with each dimension (Kruskal & Wish I 978; Arabie, Carroll, & DeSarbo I 987). In other words, different panels of respondents might emphasize the same underlying dimensions differently, according to the context provided by set size. It was hypothesized that the tree models would again provide a superior fit to the data, possibly showing some variation as a function of set size, and that yet again no interpretable global, unidimensional representation of quantifier meaning in terms of quantity would be forthcoming.
David A. Routh 205
Materials
Twenty exemplars of quantified phrases were prepared using the following frame: Q ofthe Students in a Class oJN. Within each phrase, Q was replaced by one of the following quantifiers: All , Each , Every one, Half, None, Lots , A lot, Many , Most , Several, Few, A Jew , Some , Any , Three-quarters , A majority , A minority , Fewer, A number, and More . Four sets of phrases were prepared, and within any set N was fixed at one of the four possible values of set size, i.e. 12 , 24 , 4 8 , and 96.
The respondents were randomly allocated to one of four panels of approxi mately equal numbers, with a separate panel for each level of the set size variable. Each person received a relevant, randomized set of 20 typed slips of paper, 6 X 5 em., with each bearing a different exemplar of a quantified phrase. They were instructed to sort them into a freely chosen number of groups on the basis of their similarity of meaning. They were required to keep the exemplars visible throughout, and to number their groups I , 2, 3 and so on when they had finished. Results Preliminary analyses
A co-occurrence matrix was constructed for each set size (panel of respondents) showing the proportion of respondents who had sorted each possible pair of exemplars into the same categories. The complementary proportions were treated as a dissimilarity measure for the purposes of analysis (c£ Miller 1 969). A non-metric version of the INDSCAL model was fit to the four matrices for the different set sizes using the ALSCAL procedure (Takane, Young, & De Leeuw I 977) in its SPSS-X ( 1 986) implementation. The average r2 values (proportions of explained variance) were 0.72, 0.86, and 0.92, respectively, for the 2-, 3-, and 4-dimensional solutions. However, it was clear that none of the dimensions could be interpreted in terms of quantity (amount or proportion). Moreover, the weights for each dimension showed very little variation as a function of set size. At the present time, 'multiple matrix' versions of the tree models (ADDTREE and EXTREE) are not yet available in the literature. But separate analyses of individual matrices showed that obtained structures were essentially invariant over set size. Accordingly, the final analyses were conducted on a new matrix for the data pooled over all respondents. It is these which we shall now consider in detail.
Downloaded from jos.oxfordjournals.org by guest on January 1, 2011
Design and procedure
.w6
On
Representations of Quantifiers
Analyses of pooled data
Table 1 Percentage of linear and monorone variance accounted for by rhe various solutions
r-index X 100
Monotone Linear
Level of measurement Ordinal Interval
Tree
MDS
ADDTREE
EXTREE
ID
2D
JD
96 92
98 97
39 27
8I 62
92 84
The goodness-of-fit measures resulting from the analyses are shown in Table 1 , and indicate that both the EXTREE and 3-dimensional non-metric (ordinal) MDS models fit the data very well. However, it is also clear that the ADDTREE model yields quite a good solution. Both the interval and ordinal one-dimensional MDS solutions were rather poor. In addition, inspection of the Shepard diagrams (plots of the transformed proximities versus solution distances) suggested that they were degenerate. They contained a small number of discontinuous steps indicating that the exemplars had collapsed into a small number of clusters (c£ Shepard I 974)·
Downloaded from jos.oxfordjournals.org by guest on January 1, 2011
Two different scaling programs were used in this phase of the analyses. First, ALSCAL was used to obtain MDS solutions in a Euclidean space of between I and 3 dimensions, both for ordinal and interval levels of measurement. Next, EXTREE (Corter & Tversky I 986) was used to obtain a non-hierarchical (overlapping) clustering solution, as well as a hierarchical ADDTREE solution (Sattath & Tversky I 977). The selection of the 'best' solution requires that a balance be struck between conflicting aspects, such as interpretability, goodness-of-fit, and the number of free parameters used. Two measures of goodness-of-fit were employed: first, the square of the product-moment correlation between the solution and the data, rLZ, was used to measure the proportion of linearly explained variance; and, second, the square of the product-moment correlation between the solution and the 'best' monotone transformation of the data (known as disparities), rM Z, measuring the proportion of monotonically explained variance (c£ Pruzansky et al. I 982). In general, a two-dimensional MDS solution is comparable to ADDTREE in terms of its associated free parameters, 2n versus a maximum of(2n - 3), where n is the number of exemplars. A three-dimensional solution uses more free parameters than EXTREE, 3 n versus a maximum of (2n - 3 + (n/2) ).
David A Routh 207
Multidimensional scaling
1 .5 ,------, e A MINORITY • A FEW
�
FEW R
1 .0 0.5
W
FE
• SOME
ANY •
e A NUMBER •
All •
0.0
EACH •
EVERY ONE •
-0.5 -1.0
NONE
e HALF
sEVERAL
N
::2: i5
•
A LOT LOTS MOST l. .. •MANY MORE A MAJORITY
• THREE-QUARTERS
- 1 .5 '-;;----:-'--::--+.:---;;-'-;o-----:;:-'-=-----=-'-=-----=-'=-----=' -1 .5 0.0 0.5 1 .0 1 .5 2.0 -1.0 -0.5 DIM1 Figure
1
Tree models
Figure 2 displays the additive tree solution (Sattath & Tversky 1977). In this case, the quantifiers are represented by terminal nodes of the tree on the right hand side. The distance between quantifiers is given by the horizontal path length separating them; the vertical segments are purely for cosmetic purposes. It is possible to interpret a tree representation as a hierarchy of clusters and in
Downloaded from jos.oxfordjournals.org by guest on January 1, 2011
Figure 1 displays the two-dimensional Euclidean solution for an ordinal level of measurement (the Shepard plot was S-shaped and non-degenerate). This represents the quantifiers as points in a plane, with the rank order of the interpoint distances approximating the rank order of the proximities. Even if one rotates the axes to facilitate interpretation, it is very clear that the points cannot be projected on to an obvious global dimension corresponding to quantity (whether of amount or proportion). However, some global aspects of quantifier meaning are revealed. There is a left-to-right distinction between the 'fuzzier' quantifiers (e.g. Jew, several, many , most) and the more precise ones (e.g. all, none, each ). Also, there is vertical contrast between relatively small and relatively large quantities. Elimination of the data relating to the comparatives (more and fewer) left the configuration essentially unchanged.
208 On Representations of Quantifiers ALL UNIVERSAL EACH ;--
EVERY ONE NON-ASSERTIVE
ANy
�
NEGATIVE
NO NE THREE-QUARTER s HALF
LOTS ABSOLUTE A LOT
M U LTAL
I
MOST A MAJO RITY MORE
COMPARATIVE
SEVERAL ASSERTIVE EXISTENTIAL
rl
SOME A NU MBER
'---
ABSOLUTE
rl
PAUCAL
A FEW FEW A MINORITY
COMPARATIVE
Figure
FEWER
2
terms of common and unique features (Tversky 1977; Tversky & Sattath 1979; Corter & Tversky 1986). In the latter case, each segment corresponds to a measure of the features that belong to the objects originating from the segment. The terminal segments represent the unique features of each quantifier and higher order, non-terminal segments represent features shared by members of the relevant cluster. It is evident from Figure 2 that most quantifiers are well
Downloaded from jos.oxfordjournals.org by guest on January 1, 2011
MANY
David A Routh 209
Downloaded from jos.oxfordjournals.org by guest on January 1, 2011
defined in terms of unique features, but they also combine into clusters that are fairly interpretable. The labels for the clusters in Figure 2 have been borrowed from Quirk, Greenbaum, Leech, & Svarcvik (1985, Table 6.48). It is intriguing that the clusters appear to be consistent with the categories of indefinite pronouns derived by using traditional techniques. Perhaps chis should occasion no surprise, for our procedure, of consulting a panel of competent users of English about contrasts of meaning, may not be too far removed from traditional linguistic practices. The additive tree solution appears co capture the local or non-global structure of quantifier meaning. Running from the top to the bottom of Figure 2 the major clusters are ordered from those with more precise meanings to those with the vaguer or fuzzier meanings. Also, the vertical arrangement of the major clusters roughly corresponded to an ordering in terms of quantity (the position of none is the major anomaly). This phenomenon, in which the distances of terminal nodes (shown on the right} from the root (on the left} exhibit a leftwards-curving concave pattern, was found in a stronger form by Sattath & Tversky (1977, Figure 8). In the case ofthese authors, it corresponds to an ordering of animals in terms of size. It is suggestive evidence that the small number of adjacent, major clusters do carry crude information concerning their ordinal position. Of course, this is quite different from the sort of evidence that would be required for a global, continuous dimension of amount, graduated in terms of the fine-grained meanings of a myriad of specific quantifiers (c£ Newstead et al. 1 987; Moxey & Sanford 1 993a). Further analysis suggested that the clustering of the quantifiers is not entirely hierarchical. This may be seen from the overlapping clusters present in the extended tree representation (Corter & Tversky 1 986) exhibited in Figure 3· Once again, a non-terminal segment defines a cluster consisting of all the quantifiers that originate from it. However, the extended tree also includes labelled segments (denoted by capital letters, e.g. C). Each such segment represents a unique cluster of quantifiers (e.g. More and Fewer labelled by E) which overlaps the basic additive tree. Once again, the distance between quantifiers is a measure of their distinctive features, and is given by the horizontal path-length between them. Note that the path length now excludes any marked segment that occurs twice, since such a segment is a common feature, not a distinctive feature. Thus the marked segments should be regarded as being retractable springs. For example, More and Fewer (marked by E in Figure 3) are closer in meaning than suggested by Figure 2. The sets of overlapping features are listed at the root of Figure 3· Overall, the extended tree replicates the major hierarchical clusters evident in the additive tree, but also reveals some informative overlapping clusters. It is especially interesting that several can have a meaning close to afew (if we eliminate the segments marked by H, I, and X}, but also a meaning close to many and lots (eliminating the
2 1 o On Representations of Quantifiers NNNNNNNN
I I
---
ALL
1.----- EACH NNNNNNN - EVERY ONE
UUUUUU
------
ANY
r----- NONE
? DDDDDDDDDDDDDDDD
----
THREE-QUARTERS
CCCCCCC
I
----
MANY
.----L_ ccccccc -- LOTS --i DDDDDDDDDDDDDDDDD
I
�'----- A LOT
1 --- MOST .---
_
'--------l I'----- A MAJORITY - MORE
EEEEEEEEEEEEEE CCCCCCC
l
IIIXXXXX HHHHHH --------1
l
----
'------- SOME
u u u uuuu
------
HHHHHHH
,_____
SEVERAL
1 xxxx -x
_ __ _ _ _
111 --l
�
A NUMBER
---
1'------
A FEW
FEW
- ------- A MINORITY '-
Marked features C
EEEEEEEEEEEEEE
-------
FEWER
(MANY, LOTS, SEVERAL)
D
(THREE-QUARTERS, MANY, MOST, LOTS, A LOT, A MAJORITY. MOREl
E
( FEWER, MORE)
H (A FEW, SEVERAL, SOME, A N U M BER) I
(SEVERAL, SOME, A FEW, FEW)
N (EVERYONE, ALL) U (A N U M BER, ANY)
X (SEVERAL, SOME, A FEW, FEW, A M INORITY)
Figure
3
Downloaded from jos.oxfordjournals.org by guest on January 1, 2011
L---------11'-------------- HALF
David A. Routh
2I I
segments marked by C). This is consistent with the observation by Borges Sawyers (I 974) that there may be two distinct meanings of several.
&
D I SCUSSION
I990).
Several caveats are appropriate. First, we do need further investigations using a dimensional evaluation task, with suitable global dimensions. In other words, it must be conceded that we have not as yet explored the full potential of the multi-task, multi-model strategy. There is also the interesting possibility that
Downloaded from jos.oxfordjournals.org by guest on January 1, 2011
The major finding of the present study concerns the fact that there are components of the meaning of quantifiers beyond the global, internal, continuous, unidimensional scale of quantity (amount or proportion) hypo thesized by several previous investigators. Such a conception has been criticised before (see Newstead et a/. I 987; Moxey & Sanford, I 99 3a). Clearly, the evid ence suggests that a tree representation is much more illuminating than an MDS representation. Also, an extended tree representation (Figure 3) appears to reveal some particularly interesting overlapping clusters. Thus, it seems that we should entertain the possibility that people's knowledge of the meaning of quantifiers has an important categorical organization, one which is consonant with a classification based upon traditional linguistic procedures (c£ Quirk et a/. I 98 s). In a complementary fashion, the MDS solution (Figure I) also suggested that the semantic representation of the quantifiers might be viewed in terms of certain 'feature' contrasts. However, both Figure 2 and experience suggest that a high number of such contrasts or dimensions (e.g. s or 6) would be needed to yield a solution as informative as the tree models. An intriguing possibility is that it could prove fruitful to pursue the categorical basis of the meaning of quantifiers in terms of a frame-based semantics (c£ Fillmore & Atkins I992). Such an approach contrasts with theories concerned with semantic fields. According to these authors, 'frame semantics' should be founded upon notions such as 'cognitive frames or knowledge schemata', and the meaning of a word is only understood relative to a number of conceptual prerequisites (background schemata relating to experience, beliefs, and customs, etc.). The psychological approach to investiga tion and modelling pursued in the present paper may have some potential in this context. It is also interesting to note that if one imagines a stripped-down version of Figure 2, focusing upon the quantifiers studied by Holyoak & Glass (I 978), namely all, many , some, aJew and none , and retaining the displayed distances, then we end up with an approximately singular tree. Thus, the results of the present study are consistent with previous findings discussed earlier (Routh
2 I 2 On Representations of Quantifiers
Acknowledgements I am especially grateful to Linda Moxey and Tony Sanford for an advance copy of their monograph to quantifiers. Also, I wish to thank Carole Burgoyne and rwo anonymous referees for very helpful criticisms, and not least Tony for his enthusiastic and extremely patient advice. All the usual disclaimers should be invoked. Received: 10.0 1 .94 Revised version received: I s.o3.94
DAVID A. ROUTH Department ofPsychology University ofBristol 8 Woodland Road Bristol BS8 1 TN UK
RE FERE N C E S Arabie, P., Carroll, J. D., & DeSarbo, W . S. Corter, ]. E. & Tversky, A. (1986), 'Extended ( 1987), Three-way Scaling and Clustering , similariry trees, Psyclzometrika , 51, 429-5 I . Sage Publications, London. Currie, E. ( 1990), Life Lines: Politics and Health 1986- 1 988 , Pan Books Ltd, London. Arabie, P. & Hubert, L.J. ( I992), 'Combina torial data analysis', Annual Review of Evans, J. St. B. T., Newstead, S. E., & Byrne, Psychology, 43, I 69-203. R M. ( 1 993), Human Reasoning: The Psy chology of Deduction , Lawrence Erlbaum Borges, M A. & Sawyers, B. K. (1974), 'Common verbal quantifiers: usage and Associates, Hove. interpretation', Journal ofExperimental Psy Fillenbaum, S. & Rapoport, A. ( 1 97 1 ), Struc tures in 1he Subjective Lexicon , Academic chology , 102, 3 3 5-8. Press, London. Carroll,]. D. ( 1976), 'Spatial, non-spatial, and hybrid models for scaling', Psychometrika , Fillmore, C.J. & Atkins, B. T. (1992), 41, 43 9-63'Toward a frame-based lexicon: the
Downloaded from jos.oxfordjournals.org by guest on January 1, 2011
hybrid forms of representation, part-tree and part-dimension (Carroll 1 976), could be informative. Also, other graph-theoretic approaches such as path finder networks (Schvaneveldt, Durso & Dearholt 1 989) may deserve explora tion. In the case of the present dataset, it can be reported that we have failed to uncover any evidence of a unidimensional network. with quantifiers ordered in terms of corresponding quantities from one end to the other (c£ Schvandeveldt et al. 1 989, Figure 8). Finally, further work must be undertaken so as to clarify the nature of the unique, distinctive, and common features involved in quantifier meanings. As is the case with all products of modelling, it is important that some effort should be devoted to investigating the functional consequences for behaviour and thought of the distinctions captured by the tree representations. It would be particularly valuable if it were possible to clarify what features are so readily recruited by respondents in psychometric studies.
David A. Routh 2 I 3
Psychological Research , Springer-Verlag, London, 279-96. Moxey, L. M. & Sanford, A.J. ( I 993a), Com municating Quantities: A Psychological Pers pective , Lawrence Erlbaum Associates, Hove. Moxey, L. M. & Sanford, A. J. ( I 993b), 'Prior expectation and the interpretation of natural language quantifiers', European journal ofCognitive Psychology, 5, 73-9 1 . Newstead, S. E., Pollard, P., Riezebos, D. ( I 987). 'The effects of set size on the interpretation of quantifiers used in rating scales', Applied Ergonomics, 18, I 78-82. Olson, J. R. & Biolsi, K. J. (I 99 I ), 'Techniques for representing expert knowledge', in K. A. Ericsson & J. Smith (eds), Toward a General Theory of Expertise: Prospects and Limits, Cambridge University Press, Cam bridge, 240-8 5. Payne, J. W., Bettman, J. R., & Johnson, E. J. ( I 992), 'Behavioral decision research: a constructive processing perspective', Annual Review ofPsychology, 43, 87- I 3 I . Pruzansky, S., Tversky, A., & Carroll, J. D. ( I982), 'Spatial versus tree representations of proximity data', Psychometrika , 47, 3-24. Quirk, R., Greenbaum, S., Leech, G., & Svarrvik, J. ( I 98 5), A Comprehensive Gram mar ofthe English Language, Longman, Lon don. Reyna, V. F. (I 98 I), 'The language of possi biliry and probability: effects of negation on meaning', Memory and Cognition , 9, 642-50. Rosenberg, S. (I 982), 'The method of sort ing in multivariate research with applications selected from cognitive psy chology and person perception', in N. Hirschberg & L G. Humphreys (eds), Multivariate Applications in the Social Sciences , Lawrence Erlbaum Associates, New Jersey, I I 7-42. Routh, D. A. ( I 990), 'On the representation of natural language quantifiers', paper pre sented to the 2 I st Meeting of the European Mathematical Psychology Group, BristoL Sattath, S. & Tversky, A. ( I 977). 'Additive similarity trees', Psychometrika , 42, 3 I CJ-45·
Downloaded from jos.oxfordjournals.org by guest on January 1, 2011
semantics of RISK and its neighbours', in A. Lehrer & E. V. Kittay (eds), Frames, Fields, and Contrasts: New Essays in Semantic and Lexical Organization , Lawrence Erl baum Associates, Hove and London, 75I 02. Holyoak, K.J. & Glass, A. L { I 978), 'Recog nition confusions among quantifiers', Journal of Verbal Learning and Verbal Behav ior, 17, 249-64. Johnson, E. J. & Tversky, A. (I 984), 'Repre sentations of perceptions of risks', Journal of Experimental Psychology: General, 1 1 3, 5 5-70. Kruskal,J. {I 977). 'The relationship berween multidimensional scaling and clustering', in J. Van Ryzin (ed.), Classification and Clustering, Academic Press, London, I 744· Kruskal,J. B. & Wish, M. (I 978), Multidimen sional Scaling, Sage Publications, London. Luce, R. D. ( I 963), 'Detection and recogni tion', in R. D. Luce, R. R. Bush, & E. Galanter (eds), Handbook of Mathematical Psychology, vol. I , Wiley, New York, I 0389. Mechelen, I. van, Hampton, J., Michalski, R. S., & Theuns, P. (eds) ( I 993), Categories and Concepts: Theoretical Views and Inductive Data Analysis, Academic Press, London. Miller, G. A. (I 967), 'Psycholinguistic approaches to the study of communi cation', in D. L Arm (ed.), journeys in Science: Small Steps-Great Strides , Univer sity of New Mexico Press, Albuquerque, 22-73· Miller, G. A. ( I 969), 'A psychological method to investigate verbal concepts', journal of Mathematical Psychology, 6, I 69-9 I . Miller, G . A . ( I 972), 'English verbs o f motion: a case study in semantics and lexical memory', in A. W. Melton & E. Martin (eds), Coding Processes in Human Memory, Winston, Washington, DC, 3 3 5-72. Moxey, L M. & Sanford, A. J. ( I 992), 'Con text effects and the communicative func tions of quantifiers: implications for their uses in attitude research', in N. Schwarz & S. Sudman (eds), Context Effects in Social and
2 1 4 On Representations of Quantifiers
Probabilistic Representations, Academic Press, London. Takane, Y., Young, F. W., & De Leeuw, J. ( 1 977), 'Non-metric multidimensional scaling: an alternating least squares method with optimal scaling features', Psychometrika , 42, 7-67. Townsend, J. T. ( 1 978), 'A clarification of some current multiplicative confusion models', Journal ofMathematical Psychology, 18, 2 5-38. Tversky, A. ( 1 977), 'Features of similarity', Psychological Review, 84, 3 27-52. Tversky, A. & Hurchinson, J. W. ( 1 986), 'Nearest netghbour analysis of psycho logical spaces', Psychological Review, 93, 322. . Tversky, A. & S� ttath, S. ( 1 979), 'Preference trees', Psychological Review, 86, 5�2-73.
Downloaded from jos.oxfordjournals.org by guest on January 1, 2011
Schvaneveldt, R. W., Durso, F. T., & Dear holt, D. W. ( 1 989), 'Network structures in proximity data', The Psycholo� ofLearninx and Motivation , 24, 2�9-84. Shepard, R N. ( 1 97�). 'Represenration of structure in similarity data: problems and prospects', Psychometrika , 39, 373-42 1 . Shepard, R N . ( 1 980), 'Multidimensional scaling, tree-fitting, and clustering', Sci ence, 2IO, 39o-8. Smith, ]. E. K. ( 1 968), 'Models of confusion', paper presenred to the Psychonomic Society, St. Louis (cited by Townsend 1 978). SPSS-X ( 1 986), User's Guide, 2nd edition, McGraw-Hill, London. Suppes, P., Krantz, D. H., Luce, R. D., & Tversky, A. ( 1 989), Foundations ofMeasure ment, vol. II, Geometrical, Threshold, and
Journal ofSemalltus
1 1 : 21
s-230
© N.l.S. Foundation ( 1<J9-l)
Do Verbs Act as Implicit Quantifiers? STEPHEN E. NEWSTEAD
University oJPlymouth
Abstract
1 I NTRODUCTION For a number of years now, psychologists have been interested in the possibility that verbs in sentences which are generalizations may carry with them certain assumptions about the quantities associated with the subject and object of the sentences. Consider, for example, the sentence Nurses like accountants . At first sight, it might appear that this is a universal statement, meaning that all nurses like all accountants. However, this is not necessarily the case, since it is by no means clear that the sentence would be rendered false if one or more nurses were found who did not like accountants. In other words, the sentence may imply only that most nurses like accountants. This will be referred to as implicit quantification. It is of considerable interest to know just what quantities are necessary for sentences of this kind to be true, and also what factors influence these quantities. For example, are different quantities implied by different verbs, or by different subjects and objects? Does the context in which the sentence occurs have any effect? Implicit quantification was first investigated in a srudy carried out by Gilson & Abelson ( 1 965) which is now largely neglected. Their study used a novel and unusual technique for srudying people's willingness to accept generalizations
Downloaded from jos.oxfordjournals.org by guest on January 1, 2011
A number of studies suggest that verbs can act as implicit quantifiers on the subjects and objects of the sentences in which they are used; rhus a sentence such as Children like animals has truth conditions which fall short of the universal. Furthermore, it has been claimed that the quantity implied varies as a function of the type of verb used, for example, whether the verb describes an observable event or a subjective state. The present research investigated this effect further by asking people to provide dtrect numerical estimates of the quantity implied. Experiments t and 2 indicated that there are differences berween verb types, bur that these are rather different from those obtained using other techniques. Experiment 3 showed that, with appropriate controls for the subjects and objects, there were no differences berween the verb types. The results are interpreted as indicating that people base their estimates not on the quantity implicit in the verb bur on the expected frequency of the activity described by the sentence. It is suggested that expected frequency, rather than the semantic properties of the verb, may help explain nor only tmplicit quanrificarion bur also implicit causality in verbs.
2 1 6 Do Verbs Act as Implicit Quantifiers?
Downloaded from jos.oxfordjournals.org by guest on January 1, 2011
when these were implied rather than explicitly stated. For example, participants were told that there were just three types of tribe, Southern, Northern and Central, and that two of these liked paintings while the other did not. Participants had to judge whether or not this evidence rendered acceptable the claim that Tribes like paintings . The strongest finding to emerge from this study was that the verb used in the sentence was the prime determinant of people's willingness to accept the generalisation. It was found that the episodic verbs produce, buy , have , and steal led to significantly more acceptances than the more subjective verbs get angry witlz , understand , like, and avoid, irrespective of what were the subjects and objects of the sentence. Subsequent research has replicated this finding many times (e.g. Abelson & Kanouse 1 966; Kanouse & Abelson 1 967; Kanouse 1 972; Klemp 1 974; Podeschi & Wyer 1 976; Wyer & Podeschi 1 97H). What is the explanation for these findings? The main line of research that has been pursued in trying to answer this question is that of categorizing verbs into types according to whether or not they permit generalizations to be made. A number of possibilities have been considered. Gilson & Abelson ( 1 96 5) proposed a distinction between manifest and subjective verbs. Manifest verbs are ones which express unitary, usually observable relationships, while subjective verbs indicate less observable relationships, often ones involving sentiment or feeling. Manifest verbs are the ones which permit generalizations. This has since been supplemented by the addition of a distinction between positive and negative verbs, with positive verbs permitting more inductive generalizations (Kanouse & Abelson 1 967; Kanouse 1 97 2). Writers such as Pee ters ( 197 1) and Lewicka ( 1 986) have claimed that this distinction, which they characterize as one between approach and avoidance, is in fact the principal determinant of generalization. Klemp ( 1 974) explains the findings in terms of three distinctions: polarity (whether the verb is positive or negative); lexical markedness (whether the verb is marked or unmarked); and semantic type (whether the verb is a private one, expressing a relationship known only to the person themselves, or a state one). Given the choice of possible distinctions, it might be thought that it would be a relatively straightforward matter to ascertain which provides the best fit to the data. This is not the case, partly because there are subtle differences in the generalizability of verbs in the various studies, bur primarily because these distinctions are not as hard and fast as they at first sight appear. This is graphically illustrated by Kanouse ( 1 972) who recategorized the verb want as a manifest verb primarily in order to explain the unexpected results that it produced in his study! It is probably fair to say that not one of the distinctions proposed provides a completely adequate explanation of all the findings. It has been widely assumed that the differences between the verbs stem from differences in implicit quantification. It has been suggested (e.g. by Gilson &
Stephen E. Newsread
217
Downloaded from jos.oxfordjournals.org by guest on January 1, 2011
Abelson 1 96 5) that verbs impose subtle constraints on the number o f entities that might be associated with their subjects and objects. For example, the verb produce in the sentence Tribes produce magazines is assumed to be true if only relatively few of the tribes do produce magazines. On the other hand, the verb like in the sentence Tribes like magazines is assumed to suggest a rather larger number. This explains quite readily why produce results in more inductive generalizations than like; in the situation where just one-third of the tribes provide positive instances, the threshold for acceptance ofproduce is more likely to be surpassed than is the threshold for like . However, stated in these terms, implicit quantification is in effect simply a restatement of the findings. What is needed is independent support for the claim that verbs act as implicit quantifiers. Abelson & Kanouse ( 1 966) asked their participants to provide quantifiers for a wide range of statements. They were asked to choose one of six quantifiers (one or two , aJew, some , many, most , and all) to apply to the subject and the object of sentences such as Artists avoid magazines . The quantifiers were numerically encoded, with one or two being scored as 1 and all being scored as 6. As was expected, the verbs have, buy, and produce received lower scores than the verbs like , understand , and avoid ; in other words, it requires less supportive inductive evidence to accept generalizations based on the former group of verbs. However, there are important qualifica tions to this finding. In particular, the differences berween the verbs were found principally with respect to the sentence objects; there was very little difference in generalization berween verbs when only the subjects were considered, and it is possible that this difference was not independently significant. Coupled with the fact that so few verbs were used, this means that these findings must be viewed with some caution. Subsequent research has not served to resolve this issue. Kanouse ( 1 972) confirmed the finding that verb type influenced the quantifier assigned to sentence objects (he did not look at sentence subjects), and also found that sentences were easier to recall when the quantifier used with the verb was consistent with that implicitly suggested. Podeschi & Wyer (1 976) used a res tricted range of quantifiers (some and all) and they too found that different verb types led to different choices of quantifier. They did not break down this effect by subject and object, but they suggest it held for both of these. Klemp (1 974) used a similar task to that employed by Abelson & Kanouse ( 1 966) and Kanouse (1 972) to assess the implicit quantifiers of J2 verbs but found contrary results. There were no differences in the quantifiers selected for either sentence sub jects or sentence objects as a function of the polarity, markedness and semantic type of the verbs. Since these were all factors that affected the tendency of people to accept generalizations, he concluded that verbs do not differ in the quantities that they suggest, at least not in a way that can readily explain per formance in the Gilson & Abelson task.
218
Do Verbs Act as Implicit Quantifiers?
EXPE R I ME N T
I
The aim of this experiment was to investigate whether verbs act as implicit quantifiers. The verbs used were either manifest or subjective and either positive or negative, giving four verb types altogether. It was expected, based on the results obtained by Kanouse & Abelson ( I 967) and others, that manifest verbs would lead to lower implied quantities than subjective ones and that positive verbs would lead to lower quantities than negative. Participants
There were I 6 parnClpants, all of whom were undergraduate psychology students at the University of Plymouth and who participated for course credit. Material
The following verbs were used: Manifest positive: Manifest negative:
buy, use, make, have destroy, harm, damage, abuse
Downloaded from jos.oxfordjournals.org by guest on January 1, 2011
The question, then, of whether verbs play a key role in implicit quantifica tion, with different verbs playing different roles, has not been unequivocally answered, and certainly there is no agreed categorization of verb types. The purpose of the present research was to re-open the question of whether verbs play a controlling role in implicit quantification. As we have seen, the way in which this has been investigated previously has been by asking subjects to indicate which quantifier is the appropriate one to use with given verbs. This may not be the best technique to employ. For one thing, the scoring system used by authors such as Kanouse ( I 972), i.e. scoring the quantifiers I -6, assumes that the quantifiers lie on an equal interval scale, a highly dubious assumption (Newstead & Collis I 987). Further, there are context effects in the inter pretation of quantifiers, such that the same word may be taken to signify different things in different situations (Pepper & Prytulak I 97 4; Moxey & San ford I 99 3). If, for example, the word some means something different when paired with have than when paired with understand, it is inappropriate to draw conclusions based on the frequency with which the quantifier is chosen with the two verbs. In an attempt to overcome these problems, the present study used a more direct measure of implicit quantification in which participants were asked to give numerical estimates of the amounts signified.
Stephen E. Newstead 2 I 9
Subjective positive: love, admire, understand, enjoy Subjective negative: hate, fear, avoid, ignore
The town of Arcadia has a population of exactly 1 000 people. In this booklet you will be presented with a series of statemems describing the inhabitants of this town. Your task is to indicate how many people you think are indicated by each statement. You will first of all be asked to give a smgle whole number indicating the most hkely number of people being referred to. Then you will be asked to give the minimum number of people who could be referred to given the statement presemed; and also the maximum number of people who could be referred to. The statmems will be similar to the following: If Arcadians like children, then . . . Arcadians like children. If you think this statemem is most likely ro refer to 740 Arcadians, then 740 1s the number thar you should emer in the space provided.
Similar instructions were then given about filling in the numbers for the maximum and minimum values that were to be indicated. Participants were asked to use only single whole numbers and to proceed through the items in the order in which they appeared in the questionnaire. The 1 6 test items then appeared on separate pages of the booklet, in a different random order for each participant. On each page there was a statement and spaces in which to place the optimum, minimum and maximum values. Procedure
The participants were run in a single group and were allowed as much time as they required to complete the booklet.
Downloaded from jos.oxfordjournals.org by guest on January 1, 2011
These are the same as were used by Kanouse ( 1 972) with the exception of damage (Kanouse used the verb fight ). The reason for this change was that the verb fight leads to anomalous sentences when used with inanimate objects, as it was both in this study and that of Kanouse. The verbs were combined with 1 6 possible sentence objects: sweets, pottery, bags, television sets, tablecloths, shirts, chairs, clocks, umbrellas, cars, board games, washing machines, magazines, paintings, cameras and bicycles. These 1 6 possible sentence objects were divided arbitrarily into four sets offour, and each set of four was rotated round the four verb types such that each object occurred equally often with each verb type. The subject of the sentences was always the same-the inhabitants of the imaginary town of Arcadia. The sentences were put into a booklet for which the instructions began as follows:
220 Do Verbs Act as Implicit Quantifiers?
Results
The results can be seen in Table 1 , where the mean typical, minimum, and maximum values are given for each verb. On the basis of previous research it would have been expected that manifest verbs would produce lower magnitude estimates than subjective ones. There was a slight trend in this direction on all of the scales, and for the minimum values this difference (2 5.1% vs. 3 1 .0%) approached significance, F( I , 1 4) 3·74. p 0.074. The difference was non significant with typical and maximum values, Fs 2.0 1 and 2. 1 1 respectively, p > 0. 1). =
=
=
1
Mean values, expressed as percentages, given in Experiment 1
Manifest positive Buy Use Have Make Mean Subjective positive Love Admire Understand Enjoy Mean Manifest negative Destroy Harm Damage Abuse Mean Subjective negative Hate Fear Avoid Ignore Mean
Typical value
Minimum value
Maximum value
63.1 65.2 63 · 3 25-9 5 ·P
33·8 35·o 4 1 .4 I 5·5 3 1.4
8p 83·4 85., 45·9 74·4
55·' 62.4 59·8 6o.8 59·7
32· 7 35·4 39·9 38.8 36-7
82.2 78.0 74·8 78·9 78·7
26.9 3 1 .2 36.6 27.6 30. 1
I 5·5 20.3 '9· 3 20.4 ! 8.9
57·0 50-7 59·2 56.0 5 5·7
37·7 p.6 3 5·3 37·7 3 5 ·8
26.7 25.1 2 1 .2 34·0 26.8
60.9 5 5·3 60.9 50·5 56.9
It would also have been expected that positive verbs would produce lower estimates than negative ones, but in fact the difference went significantly in the opposite direction. On the minimum ratings, the mean for positive verbs was 3J.J%, that for negative verbs 22.8%. This difference was statistically significant, F( I , 1 4) 6.54, p < o.os, as it was also with the typical and maximum ratings, =
Downloaded from jos.oxfordjournals.org by guest on January 1, 2011
Table
Stephen E. Newstead
22 I
F( 1 , 1 3) 20.4 1 , p < o.oo 1 and F( 1 , 1 4) 1 1 .09, p < o.o 1 respectively. There was no significant interaction between these two factors on any of the ratings (all Fs < 1 ). These results are difficult to reconcile with previous research. The finding that in the present study positive verbs produced much higher estimates than did negative verbs is directly contradictory to all previous studies (Abelson & Kanouse 1 966; Kanouse 1 972; Podeschi & Wyer 1 976; Wyer & Podeschi 1 978). What is more, essentially the same verbs were used as in one of the earlier studies, that of Kanouse ( 1 972). Previous studies required participants to select appropriate quantifiers while the present experiment used direct estimates of quantity, bur it is far from clear why this change should lead to such different results. One difference between the present and previous studies which might explain the discrepancy is that the present study looked at quantification of subjects, while most earlier studies have looked at quantification and generalization of objects. The only other study to look specifically at generalization over subjects was that ofAbelson & Kanouse ( 1 966), and that too found only a small difference between manifest and subjective verbs (it did not look at the effects of negation). It is possible that implicit quantification does not occur to any great extent over subjects, despite the regularity with which this assumption has been made. In order to investigate this possibility further, a second study was carried out which looked at quantification of both subjects and objects. Furthermore, in order to check that the results were not an artefact of the specific subject term used, a different term was used in Experiment 2. As a further check on the generality of the findings, a slightly different task was used, one requiring participants to give their responses as percentages rather than absolute num bers. �
�
2
This study investigated participants' direct estimates of the quantities implied by verbs in both the subjects and objects of the sentence. Participants
There were 1 6 participants, all of whom were psychology students at the University of Plymouth who took part in the experiment for course credit. None had taken part in Experiment 1 .
Downloaded from jos.oxfordjournals.org by guest on January 1, 2011
E X PE R I M E N T
222 Do
Verbs Act as Implicit Quantifiers?
Materials
People ignore magazines This is most likely to indicate that % of people ignore magazines As a minimum, this indicates that % of people ignore magazines __
__
The test items differed in the verb used, and also as to whether participants were asked to rate the quantity associated with the subject of the sentence (as in the above example) or the object of the sentence. For half of the participants the 1 6 test items involving the rating of the subject appeared first, in random order, followed by the 1 6 items involving the raring of the object; for the remaining participants this order was reversed. Participants were asked to provide both the percentage which the statement was most likely to indicate and the minimum percentage that could be indicated, using a single whole number between 1 and 1 oo. They were not asked to provide a maximum percentage. This omission simplified the task, and lost little since the results obtained in Experiment 1 with this task simply reflected those obtained from asking participants to estimate the most likely and minimum quantities. Results
The findings of this srudy are presented in Table 2. Looking first at the minimum ratings, there was no difference between the overall ratings given to subjects and objects, F( 1 , 1 4) 2.07, p < 0. 1 and just one significant interaction involving this factor, a two way interaction with the positivity/negativity of the verb, F( 1 , 1 4) 5 8 2 , p < o.os. Objects seemed to give higher values than subjects with negative verbs bur not with positive verbs. In other respects, the results were virrually the same as those obtained in Experiment 1 : there was no difference between manifest and subjective verbs, F( I , 1 4) - o.o J , p > . 1 , but positive verbs produced significantly higher ratings than negative ones, F(I , 1 4) 1 4- 3 5. p < 0.0 1 There was no significant interaction between these =
=
=
.
.
,
Downloaded from jos.oxfordjournals.org by guest on January 1, 2011
The same verbs were used in this srudy as had been used in Experiment 1 . The subject term was always people and the object term was always magazines ; these were chosen as fairly neutral terms which made reasonable sense with each of the verbs used in the experiment. Booklets were prepared containing the test items. The instructions to the booklet asked participants to consider only the information contained in the statements, and asked them to assume that they did not know the people to whom the sentences referred, nor anything about what they did to magazines. Each test item was similar to the following example:
Srephen E. Newsread 223
Table
2
Mean values, expressed as percenrages, given in Experimenr 2 Subjecr Typical value
Minimum value
Typical value
Mmimum value
78.6 65.2 74·9 23 · 3 6o.5
4 1.8 26.2 36.6 4·8 27·3
68.3 59·7 56.2 58.2 6o.6
39·3 22.0 1 8.o 41.5 30.2
69.6 46.8 8o. I n8 68.5
J2.6 20.6 4 1 .9 42.8 34·5
6o.8 50.8 84.6 79·4 68.9
22.4 1 7·3 45·I 37·5 30.6
5 1 .9 30.8 46.6 32·3 40·4
25.8 I 1.7 I 9·9 I 0.5 I 7.0
63·4 36.2 48.8 46·3 48.0
3J.I I 3.8 22.7 2 1.8 22.9
3J.I 2I.I J4. I 45·8 33-5
I 6.9 8.6 1 0.4 I 6.2 I 3.0
4 1 .3 2 1 .4 � t ·9 59· I 43· I
I 5.6 I 1.9 28.4 3 1 .4 2 1 .8
two factors, F( I , I 4) 1 .26 p < . I . Visual inspection ofTable 2 reveals that this pattern of results clearly holds both for sentence subjects and for sentence objects. A similar, but not identical, picture emerges for the typical ratings. There was no main effect of subject/object, F( I , J 4) 2,60, p > o. I , nor, in this case, any significant interactions involving this factor. There was no overall difference between manifest and subjective verbs F( I , I 4) o.o I , p > . I , but positive verbs gave significantly higher ratings than negative ones, F( I , J 4) 24.76, p < o.oo t . There was also an interaction between these two factors, F( I , 1 4) S·77• p < o.os, with subjective verbs producing higher ratings than manifest when the verbs were positive but the opposite way round when the verbs were negative. This weak interaction has no easy explanation and has not been found in any previous studies nor in the present Experiment I . �
=
=
=
=
Downloaded from jos.oxfordjournals.org by guest on January 1, 2011
Manifesr positive Buy Use Have Make Mean Subjective posi rive Love Admire Undersrand Enjoy Mean Manifesr negative Destroy Harm Damage Abuse Mean Subjective negative Hare Fear Avoid Ignore Mean
Objecr
22�
Do Verbs Act as Implicit Quantifiers?
EXPE RIMENT
3
The aim of this study was to investigate implicit quantification over the same set of verbs as used in the previous studies but using neutral, abstract material. Participants
There were 25 parnCtpants, all of whom were psychology students at the University ofPlymouth who took part for course credit. None had taken part in either of the two previous experiments. Materials
Booklets were prepared in which the 1 6 verbs were paired at random with letters of the alphabet as subjects and objects. Each test item involved
Downloaded from jos.oxfordjournals.org by guest on January 1, 2011
In general, the findings of this experiment indicate that the pattern of results obtained in Experiment 1 was not due to the fact that the quantification was over subjects, since an almost identical pattern of results emerged with ratings over objects. Furthermore, the use of a different subject term seemed to have minimal effects, as did the change from asking participants to give their estimates using percentages rather than absolute numbers. It seems reasonable to conclude that the results of Experiment 1 are robust, at least using this set of verbs, and very different from those obtained in previous studies using different techniques. There is a slight complication in that two interactions emerged in Experiment 2 which are not easy to explain; however, the fact that they emerged with only one of the types of ratings suggests that they may not be strong effects. Both Experiments 1 and 2 used a restricted number of subjects and objects. Experiment 1 used the subject term Arcadians throughout, while Experiment 2 used only the subject term people . There was more variety with the object terms, at least in Experiment 1 , but it remains a possibility that the obtained effects are attributable in part to the particular subject and object terms selected. It is important to know whether the verb effects obtained are more general than this. Hence in Experiment 3 completely arbitrary, abstract material was used: the subjects and objects were all randomly chosen letters of the alphabet. Ifthe estimates remain the same, this would indicate that the verbs are powerful determinants of quantification, albeit in a rather different way than might have been assumed on the basis of previous research. If, however, the estimates change, this would suggest that the verbs are not the only factor in quantification but that subjects and objects can also have an influence.
Srephen E. Newsread
225
participants giving a typical and minimum estimate of quantity, expressed as a percentage. The following is a typical test item: This is most likely to refer to As a minimum, this refers to
Qs buy Zs % of Qs and % of Qs and
__
__
% of Zs % of Zs
__
__
Method
Participants were run in small groups of varying sizes. They were allowed as much time as they required to complete the test items. Results
The results are summarized in Table 3· With respect to the minimum ratings, an analysis of variance involving the factors of subject/object, manifest/ subjective and positive/negative showed no significant effects on any factor, nor were there any interactions between factors. The three-way interaction between all the factors approached significance, F( 1 ,23) - 4. 1 0, p - 0.05 5, but no other effects were even marginally significant. With the typical ratings, there was just one significant effect, again the three way interaction involving all the factors, F( 1 ,23) < 5·54. p 0.005. The source of this interaction seemed to be a slightly lower rating for the manifest negative verbs when they were the objects of the sentences. It is not easy to devise any plausible explanation as to why this should be the case. The results of this experiment indicate that when abstract terms are used, differences between the verb types, and indeed between the individual verbs as well, tend to disappear. =
GENERAL D I SC U SS I O N Taken together, the results from these three studies point to a very clear conclusion but one which is surprising in the context of previous research: verbs do not dominate quantity assignment. The first two experiments
Downloaded from jos.oxfordjournals.org by guest on January 1, 2011
Participants were told that the letters referred to two sets of entities (people, things or objects) but that they did not know which. They were asked to fill in the percentages that they thought were indicated by each statement. They were told that there were no right and wrong answers, and that it was their personal judgement that was of interest. Each verb occurred just once in the booklet, and the order in which the verbs appeared was varied.
226 Do Verbs Act as Implicit Quantifiers?
Table 3 Mean values percentages given in Experiment 3
Subject Typical value
Minimum value
Typical value
Minimum value
60.4 63.8 73· 1 62.8 6 5.2
36.6 34·0 45·5 36·7 38.o
53·6 63·9 63·9 68.8 62.2
36.8 37·9 42.2 45·2 40. 1
74·4 69.6 52.1 66.8 6q
45·6 43-4 28.6 J7.2 38·7
62.0 59·3 56.7 67.6 6 1 .4
36·9 39·5 3 4.2 40.0 37·7
58.6 65·3 60.5 53·8 59·5
3 1 .6 J7.0 34·9 28.9 33·1
50·7 52.6 5 8.8 44·6 5 1 .7
27. 5 26.6 36·9 22.6 28.4
62.8 58.8 59·0 58.1 59·7
38.1 36.6 3 J.2 32·9 3 5-2
70.2 63.0 59·2 57·7 62.5
48.2 43· 1 3 5 ·4 . p .2 42.0
indicated a small but nonsignificant difference between manifest and subjective verbs, in the direction predicted on the basis of previous research, and a large and highly significant difference between positive and negative verbs but in exactly the opposite direction to what had been predicted. Most dramatically, when completely abstract material was used in Experiment 3, there were no differences whatsoever between verbs of these two types. It seems clear that the subjects and objects are exerting an influence on the ratings given in Experiments 1 and 2. The most likely explanation seems to be that participants were responding on the basis of the expected frequency of the complete scenario depicted by the subject-verb-object combination. It is easy to see how people might respond in accordance with their prior expectancies based on their own experience. Experiment 1 required partici pants to indicate the number of inhabitants of an imaginary town who behaved
Downloaded from jos.oxfordjournals.org by guest on January 1, 2011
Manifest positive Buy Use Have Make Mean Subjective positive Love Admire Understand Enjoy Mean Manifest negative Destroy Harm Damage Abuse Mean Subjective negative Hate Fear Avoid Ignore Mean
Object
Srephen E. Newsread
227
Downloaded from jos.oxfordjournals.org by guest on January 1, 2011
in various ways. It seems likely that participants brought with them to this task expectancies as to how many of the inhabitants of any town would behave, for example, how many of them might buy television sets. And, if this is a fairly typical British town, this proportion is likely to be quite high. Similarly, participants might have expectancies concerning the number of inhabitants of a town who would destroy television sets and this number might intuitively be expected to be somewhat lower. If these expectancies are similar with other entities (such as the number of people who might be expected to buy or destroy chairs, cars, or shirts), this could readily explain the result that was obtained, that buy gave consistently higher ratings than destroy. Subsidiary findings lend some support to this explanation. It is clear from inspection of Tabies 1 and 2 that the verbs in each category are not a homogeneous group. This is perhaps best illustrated with the manifest positive verbs. The verb make seems very different to the other three verbs, consistently producing lower quantities than them. For example, in Experiment 1 the typical value accorded to make was 25.9%, while the lowest value accorded to any of the other manifest positive verbs was 6 J. l %. Perhaps, with the decline of manufacturing industry in the UK, there is an expectancy that very few things are likely to be made by anyone. In addition, there was also some evidence that the quantities implied varied as a function of the object term used. This could only be investigated in Experiment 1 , since this was the only study to use different object terms, and even here each verb was used with only four different objects. Nevertheless, there did seem to be some interesting findings. For example, the typical value chosen for use was 5 1 % for pottery, 8 1 % for cars; the typical value for have was 42% for cameras, 84% for chairs. Intuitively, these differences seem to correspond to differences in expected frequency, since we would probably expect more people to use cars than pottery in a typical town, and more to have chairs than cameras. These differences cannot be properly tested for statistical significance, and it would be wrong to read too much into them, but they are at least suggestive. However, perhaps the most compelling argument for the role of prior expectancies is that the use of abstract material in Experiment 3 completely removed any difference between different verb categories or indeed beween individual verbs. The most plausible explanation for this is surely that this abstract material led to no prior expectations and that in the absence of these there were no differences between verbs. If the present explanation is correct, then the quantities attached to subjects and objects derive from prior expectancies about the scenarios involving the combination of subject, verb, and object in the sentence. The verb might be especially influential in determining the scenario, but it is by no means the only factor.
228 Do Verbs Act as Implicit Quantifiers?
Downloaded from jos.oxfordjournals.org by guest on January 1, 2011
There are instances where the same verb seems to lead to very different expectancies depending upon the subject and object it is used with. Examples of this have been presented already, and there are likely to be even more dramatic differences in the expected frequency attached to the verb make in People make ice cubes (high frequency) as opposed to People make television sets (low frequency). Hence one would expect participants to attach very different quantities to both the subject and object in such sentences. It is of course true that the present studies differed in a number of ways from others which have been used to investigate implicit quantification. The present research used direct numerical estimates of the quantity indicated in the sentence subjects and objects, while previous studies (e.g. Kanouse I 972) have asked participants to assign verbal quantifiers. Since quantifiers can change their meaning as a function of context, while numerical values presumably do not, it seems likely that the present studies are less flawed than previous ones; hence it would be ill advised simply to dismiss the present results as being due to some experimental confound. In fact, it seems likely that studies using quantifiers rather than numerical estimates also reveal more about expected frequencies than about implicit quantification. In other words, the reason that people think that Helen likes dresses means that Helen likes MANY dresses (Kanouse I 972) may derive from prior expectations. It may be that people expect women to have a liking for dresses in general. Alternatively, it may be that the statement is interpreted as meaning that Helen has a greater than expected liking for dresses, and hence that she will like more dresses than the average person. Either way, the quantity derives from prior expectancy rather than any inherent properties of the verb. Helen likes snakes might lead to a very different estimate. This same line of reasoning can also explain other research on implicit quantification which has used the Gilson & Abelson ( I 96 5) technique rather than direct estimates of the quantifier or quantity. It will be remembered that Gilson & Abelson found that, in the light ofjust one-third positive instances, participants would much more readily accept generalizations using verbs such as have than ones using verbs such as understand . The differences between verbs were seemingly consistent, being found across different subject and object terms. According to the expected frequency explanation one might expect that variations in the terms used would lead to changes in expected frequency and hence to variations in generalizability. It is, however, possible that such variations did occur but did not show up in Gilson & Abelson's analysis, since they did not investigate interactions between particular verbs and subjects/ objects, only main effects. Since the verb is obviously an important determinant of the overall scenario it is also possible that the differences between expected frequencies were relatively small. It thus remains a real possibility that expected frequency can provide an explanation for the findings.
Stephen E. Newstead 229
Acknowledgements The author would like to thank Penny Armstead for her assistance in collecting and analysing the data reported m this paper, and Tony Sanford and rwo anonymous referees for their helpful comments on an earlier version of this paper. STEPHEN E. NEWSTEAD Department ofPsychology University ofPlymouth Drake Circus Plymouth PL4 8AA UK
Received: o6.1 2.93 Revised version received: 10.02.94
Downloaded from jos.oxfordjournals.org by guest on January 1, 2011
In recent years, the work of Gilson & Abelson has been followed up principally in research on implicit causality. It has been claimed that some verbs (e.g. hit) lead to the inference that the subject of the sentence is the causal agent, while other verbs (such as like) put the causality in the object. To illustrate, john hitsJim suggests that the John is the cause of the hitting, while John likesJim suggests that Jim is the cause of the liking (see Semin & Fiedler, 1 992, for a recent review). Categorization of verbs into different types has become more and more refined but the divisions bear much in common with the categoriza tion schemes attempted with respect to implicit quantification, as the authors of such schemes readily acknowledge. Nor is this similarity coincidental since, as Kanouse ( 1 972) and Manetti & de Grada ( 1 99 1 ) point out, implicit quantification is one possible explanation of causal attribution. A verb which suggests that the activity is carried out by only a small number of people might lead to a greater tendency to accept the subject as the causal agent, as this will make the subject more distinctive as a potential causal agent. This raises the question, then, as to whether these classificatory systems, too, owe more to prior expectations than to inherent properties of the verbs themselves. Context can undoubtedly have an effect, since the extent to which we believe that Jim is the causal agent in john likesjim will depend in part on how well liked Jim is by other people and also on how inclined John is to like people in general. What is not known, however, is the extent to which implicit causality is determined by such context-dependent expectations rather than by properties of the verbs themselves. The general tenor of the arguments put forward in this paper bears more in common with writers such as Corrigan ( 1988) who have argued that verb effects can vary as a function of context. If implicit causality is indeed primarily determined by the scenario created by the entire sentence, it will be necessary to re-evaluate the extensive research which has focused almost exclusively on the verb.
230 Do Verbs Act as Implicit Quantifiers?
RE F E R E N C E S Maxey, L. M. & Sanford, A. J. (1 993), 'Prior expectation and the imerpretation of natural language quamifiers', Eun,pean Journal ofCoxnitive Psychology, 5, 7 3-9 1 . Newstead, S. E. & Collis, J. M. ( 1 9X7), 'Con text and the imerpretation of quamifiers of frequency', Erxonomics, 30, 1 44 7-62. Peeters, G. ( 1 97 1 ) 'The positive-negative asymmetry: on cognitive consistency and positivity bias', European Journal of Social Psychology, I, 45 5-74. Pepper, S. & Pryrulak, L. S. ( 1 974), 'Some times frequently means seldom: context effects in the interpretation of quantitanve expressions', journal
Downloaded from jos.oxfordjournals.org by guest on January 1, 2011
Abelson, R. P. & Kanouse, D. E. (1 966), 'Subjective acceptance of verbal general isanons', in S. Feldman (ed.), Coxnitive Con sistency: Motivational Antecedents and Behavioural Consequents , Academic Press, New York. Corrigan, R. (1 9HM), 'Who dun it? The influence of actor-panem ammacy and rype of verb in the making of causal anriburions', journal of Memory and Lan xuaxe ' 27, 4 77-65. Gilson, C. & Abelson, R. P. (1 965), 'The subjective use of inductive evidence', Journal ojPersonality and Social Psychology, 2, 3 0 1 -10. Kanouse, D. E. (1 972), 'Verbs as implicit quamifiers', Journal
