\olume
4.
�umber i
198->
JOURNAL OF SEMANTICS AN INTERNATIONAL JOURNAL FOR THE INTERDISCIPLINARY STUDY OF THE SEMANTICS OF NATURAL LANGUAGE MANAGING EDITOR:
Peter Bosch
EDITORIAL BOARD:
Simon C. Garrod Leo G.M. Noordman Pieter A.M. Seuren
REVIEW EDITOR:
Rob A. van der Sandt
ASSISTANT EDITOR:
Bart Geurts
CONSULTING EDITORS:
W. Levell (Max Planck Inst. Nijmegen),
J. Allwood (Univ. Goteborg),
J. Lyons (Trinity Hall, Cambridge),
R. Bartsch (Amsterdam Univ.),
W. Marslen-Wilson (Max Planck Inst. Nijmegen),
J. van Benthem (Groningen Univ.),
J. McCawley (Univ. Chicago),
H.E. Brekle (Univ. Regensburg), G. Brown (Univ. of Es�x),
B. Richards (Edinburgh Univ.),
H.H. Oark (Stanford Univ.),
R. Rommetveit (Oslo Univ.),
H.-J. Eikmeyer (Univ. Bielefeld),
H. Schnelle (Ruhr Univ. Bochum),
G. Fauconnier (Univ. de Vincennes),
J. Searle (Univ. Cal. Berkeley),
P. Gochet (Univ. de Liege),
A. von Stechow (Univ. Konstanz),
J. Hintikka {Florida State Univ.),
M. Steedman (Edinburgh Univ.),
St. lsard (Sussex Univ.),
G. Sundholm (Nijmegen Univ.),
D. Israel (SRI, Stanford)
Ch. Travis (Tilburg Univ.),
P.N. Johnson-Laird (MRC Appl. Psych.
B. van Fraassen (Princeton Univ.),
Umt, Cambridge),
Z. Vendler (UCSD),
E. Keenan (UCLA),
Y. Wilks (New Mexico State Univ.),
S. Kuno (Harvard Univ.),
D. Wilson (UCL).
EDITORIAL ADDRESS:
Journal of Semantics, P.O. Box 1454, 6501 BL Nijmegen, The Netherlands
Published by Faris Publications,
P.O. Box () N.I.S.
ISSN
509, 3300 AM
Foundation 0167·5133
Printed In the Netherlands by ICG
Dordrecht, The Netherlands.
Printing
FORIS BOOKS the expert-system type of lexicographic facilities
WORD EXPERT SEMANTICS
used lor building up the knowledge base The
An lnterllngual Knowledge-88Hd
languag�reontrastive nature of the system de
Approach
scribed Is underlined by the ex1enslve use of en 1986
Bart C. Papegaa1j
Bnd. 011.120, Pap. on. 57,-
X+ 254 pp.
ISBN 90 6765 262
ISBN 90 6765 261
3 X
In computational linguistics, stud1es of semantics are usually directed at the interpretation of func tion words, notably quantifiers and connectives. Typically, the problems dealt with in current lor mal or logical semantics are grosso modo invar
Intermediate language based on Esperanto. A final chapter provldas an outlook on future de velopments In computerized lexicography, sell learning sy1ems and other aspects of language technology. The general approach is in line with a trend in the field of machine translation to shift the locus of attention from the grammar formalism to the dic tionary, and beyond, to systems increasingly based on artificial Intelligence.
Iant with respect to language translation. The present work addresses the semantics of
Contents:
content words, in particular the relation between
Introduction
lexical and conceptual fields and how this relation
Part
varies among languages. From a background of
Ambiguity
machine translation and contrastive lexicology,
Chapter 1: Natural Language Processing/Chapter 2: Lexical Ambiguity/Chapter 3: A Concise His to
the author presents a framework lor contextual
1:
Machine Translation and the Problem of
neal Survey/Chapter
resolution of lexical ambiguities. This framework relies on a knowledge base of
4:
MT
&
Understanding -
Contemporary Techniquas/Chapter 5 The Im
lexical universes or word experts: clusters of non
portance of the Dictionary/Chapter 6: The Need
linguistic
lor Integration
information
such
as
hyperonyms,
exemplary slot fillers lor thematic roles, and other
Part II: The Semantic Word Expert System
word associations. The general set-up of these
Chapter 1: General Lay-out of the DLT System/
word experts has partially been derived from the
Chapter 2: The Structure of the Lexical Know ledge Bank/Chapter 3: Disambiguation with
definition-with-examples formula of conventional
SWESIUChapter 4·
diction arias. A central element in the book is the notion of
The Disambiguation Dia
logue
dependency pairs, atomic chunks of ontological
Part Ill: The Semantic Work Bench- A Develop
or contextual knowledge. These are grouped into
ment Tool
lexical universes, which in turn make up a seman
Chapter
tic network of taxonomic relations. The depen dency pair arrangement constitutes a basis lor
1:
Developing SWESIL and the LKB/
Chapter 2: The Semantic Work Bench/Chapter 3: Tests and Results/Chapter 4: Melby Test Results.
IV:
relined pattern matching of alternative interpreta
Part
tions, consecutive to the strictly syntactic parae
Chapter 1: Computer-Lexicography/Chapter 2:
of a sourca language text fragment. The result is a mechanism lor word sense compu
Future Developments
Macrocontext and Discourse Analys1s!Chapter 3: The Self-Improving System/References
tation which is probabilistic and complementary to interactive disambiguation in a semi-automatic
" ... instructive reading lor system designers and
translation system.
students of computational linguistics... "
The book discusses the design philosophy and
functioning of the word expert system, as well as
F. E. Knowles, MA, MSc, Professor Aston University, Birmingham.
FORIS PUBLICATIONS HOLLAND/USA Available at your bookstore or direct from: FOAlS PUBLICATIONS HOLLAND 509 - 3300 AM Dordrecht - tel. (0)78-510454
P.O. Box
P.O. Box
FORIS PUBLICATIONS USA. Inc. 5904 - Providence Rl 02903 tel. (401 )232-2722 -
or
of Langu•ge,
FORIS BOOKS FORTHCOMING
STUDIES IN DISCOURSE REPRESENTATION THEORY AND THE THEORY OF GENERALIZED QUANTIFIERS Jeroen Groenendljk and Martin Stokhof, (eds.)
FOUNDATION S OF PRAGMATICS ANDLEXCIAL SEMANTICS Jeroen GroenendijkandMartin Stokhof(eds) 1986
Bnd. Pap.
±170 pp.
ISBN90 67652644 ISBN 90 67652652
The papers collected In this volume center around two themes: lexical semantics. and pragmatics
1986
Bnd. Pap.
±200pp.
ISBN 90 67652660 ISBN 90 67652679
The papers collected In this volume are contribu tions to the theory of generalized quantifiers and discourse representation theory, two major Inno vative theories In natural language semantics of recent origin. Topics treated Include: a procedural view on generalized quantifiers; expresalblllty and
They deal with theoretical Issues as well as with
constraints on lexical meaning; 'phase' quantifica
descriptive topics. Including contextual dimen
tion; the analysis of aspectual oppositions; the
sions In lexical meaning; the location
relation between referential and quantifier mean Ings of descriptions and Indefinite NPs; type-free
of
events;
composltlonallty In morphology; the eplstemlc meaning of Japanese particles; the status of the
theories for natural language semantics; VP-Inter
meaning postulates; the projection problem of pre
pretatlon In a compositional discourserepresenta
supposition; Informational paradoxesandtheorles
tion theory; and dlscourserepresentatlontheory as
of Information representation In possible world
part of a phenomenological theory of belief.
semantics This book will be of interest to scholars and ad
This book will be of Interest to scholars and ad vanced students in theoretlcal linguistiC/1, philoso
vancsd students in theoretical lingulatics,phlloso
phy of language, and phliosophical loglc.
phy ollanguage, and philosophical logic. Contents: Preface/Johan van Benthem: Semantic Automa
Contents: Preface/Renate Bartsch: Context-dependent In
ta/Edward Keenan: Lexical Freedom and Large
terpretation of Lexical Items/ A/Ice tar Meulen: Lo
Categorles/Sebest/an L6bner: Quantification 8!1 a
cating
Composltio
Major Module of Natural Language Semantics/
nallty and the Syntax of Words/ Ken-lchlro Shirai:
Hank Verkuyl: Non-
Events/Mtchael Moortgat:
Japanse Noun-Phrases and Particles wa and gal
Barbara Partee: Noun Phrase I nterpretatlon and
Thomas Ede Zimmermann: Transparent Adverbs
Type Shilling Prlndples/Raymond Turner: Formal
and Scopelesa Quantiflers/Godehard Link: Pres
Semantics and Type-Free Theorles/Ewan Klein:
pie In Pragmatic Wonderland or: the Projection Problem for Presupposition Revlslted/Fr&d Land
VP Ellipsis In DATheory/HenkZeevat A Treatment of Belief Sentences In Discourse Representation
man: Paradoxes of Elimination.
Theory.
FOAlS PUBLICATIONS HOLLAND/USA Available at your bookstore or direct from: FORIS PUBLICATIONS HOLLAND 509 - 3300 AM Dordrecht- tel. (0)78-510454
P.O. Box
P.O. Box
FORIS PUBLICATIONS USA. Inc. 6904- Providence Rl 02903 -tel. (401 )232-2722
or
SCOPE OF THIS JOURNAL The JOURNAL OF SEMANTICS publishes articles, notes, discussions, and book reviews in the area of natural language semantics. It is explidtly interdisciplinary, in that it aims at an integration of philosophical, psychological, and linguistic semantics as well as semantic work done in artificial intelligence and anthropology. Contribl.tt.ions must be of good quality (to be judged by at least two referees) and should relate to questions of comprehension and interpretation of sentences or texts in natural language. The editors welcome not only papers that cross traditional discipline boundaries, but also more specialized contributions, provided they are accessible to and interesting for a wider readership. Empirical relevance and fonnal correctness are paramount among the criteria of acceptance for publication.
INFORMATION FOR AUTHORS Typescripts for publication should be sent in 3 copies to the managing editor. They should be typed on M (or similar format), one-sided, double spaced, and with a wide margin and must be accompanied by an approx. 200 words summary. Footnotes and bibliographical references must appear at the end of the typescript Diagrams must be submitted camera-ready. All papers submitted are subject to anonymous refereeing. Authors receive 20 offprints of their published articles and 10 offprints of their publish ed reviews, free of charge. 13rger numbers can be supplied at cost price by advance ar rangement Unless special arrangements have been made, copyright rests with the N1S Foundation. PRICES AND CONDmONS OF SUBSCRIPTION The JOURNAL OF SEMANTICS appears in four issues per year of approx. 100 pages each. Subscriptions for private use are available at the reduced rate ofDfl. SO,- per year; the institution rate is Dfl. 120,-. The price for single issues isDfl. 35,- (reduced rate: Oft. 15,-) . All prices exclusive packing and postage. Airmail- and SAL-rates are available on request We regret that no delivery can take place before payment has been received. Subcrlptions not cancelled before October 1st automatically extend to the following year. Placement of orders Implies the consent of the subscriber to these conditions. All orders should be sent to Foris Publications, P.O. Box 509, 3300 AM Dordrecht, The Netherlands.
JOl R '\ \l 01
\olul!"c
4
�ptcmbcr 19 c;;
"\umber l
CONTP•.'lS
ll KK \ Nll'\lll OTO 801 LI
Of R \�\1)
I)'\ VID \1 ( \Rl
209
lrrugtnaUt:n ....nd ftwon .
The 1anguJge ul Soc1al At:t.s
ER
(omnon Sen'i! lnfcren�.:c 11111 ru �gUided 237
,\n.tphor Rc-;olvcr . BARI C•El'RfS
<•encnc.o;
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
. .
[)lfCUHltJn 1-1 l ( KlR f H nng_r\ The •r\
\l.llcn"e" A
l I D '" v.IIcnv;:.., A<:.-.oc.tdtcd (nmm·nt'>un r,,r \a�t-n�c [ 'll. pen�nl a-,; d Ba
1'-o.liJUO
.
. •
.
•
.
.
•
• •
•
.
•
.
•
•
.!51
Barkre\'""• '"' l [ \1MI '\'\ J 11kk • II nt•kka (tn t.•ll..�borcllt un \l.tlh !illk K Jlu'>) Tlu t,amt' of lan�uaf(C' \rutltt'l '" (tamr lhellr�mal 1H"fflOIIIlC • ,mJ II� tfpp.,etllltl [) I
0\>I R
Rohcn �taln.1k�r
Ill\"- R£ f( HC.E 1 I T hlr ill .oun ( mpr h nhfn
lnqum
D11!..
und \\ K nt.-.d'l �trute tu u/
2"1
JoJUnaf of S�mantiC$ 4:
209-222
IMAGINATION AND FICTION
ILKKA NIINILUOTO
ABSTRACT
the syntax and the semantics of imagination sentences. Following Hintik:k:a's treatment of propositional attitudes like knowledge and perception, the propositional construction
gines that p" is taken as the basic form to which other sentenc� (such as imagines an F", "a imagines
b as an
"a ima "a imagines b", "a
F") are reduced through quantifiers ranging over 'world
lines', i.e., functions picking out individuals from the relevant possible worlds or scenes. This intensional analysis is compared and contrasted with
Barwise and Perry's situatiOn semantics. It
is also suggested that the logic of imaginatiOn helps us to understand some peculiarities of fictional discourse. For example, acts of imagination can be directed towards fictional entities (e.g. Donald Duck:, Anna Karenina) as well as real ones. Further, fictional texts,lik:e novels, can
be thought of as occurring within the scope of an imagination operator, relative to the author or T can be viewed as performing an illocutionary act of recommendation of the form: Let ll! imagine that Tl
the reader. The author of a fictional text
Imagination is needed both in the creation and the understanding of fictional texts. For R omantic poets like Poe and Baudelaire, imagination was indeed the 'Queen of the faculties'. • It is no wonder then that imagining as a mental activity has been discussed extensively within philosophy, psychology, and aesthetics. Logicians have failed to make a contribution in this field, how ever. In this paper, I argue that a logic of imagination can be developed as a special case of H intikka's semantics for propositional attitudes} I also suggest that this logic helps us to understand some characteristics and peculiarities of fictional discourse.
I.
A logic of imagination should not be understood as a logic-in-use, i .e., as a system of rules which our mind ought to follow in imagining. Rather it is a systematic framework for studying the syntax and the semantics of sentences containing some form of the term "imagine"- in analogy with the so-called logic of perception which studies sentences constructed with the verbs "to see" or "to perceive" ( cf. H intikka 1 969, 1 975; Niiniluoto 1 979, 1 982). The logic of imagination in this sense is primarily a tool for understanding
Downloaded from jos.oxfordjournals.org by guest on January 1, 2011
This paper employs possible worlds semantics to develop a systematic framework for studying
210
2.
To treat imagination as a propositional attitude means that statements of the form (1)
a imagines that p,
where "a" is an individual name and "p" is a declarative sentence, are taken to be the basic expressions for the logic of imagination. Other similar atti tudes include knowledge, belief, and perception: K aP 'a knows that p' B aP 'a believes that p' S ap 'a sees that p' Let us write laP for the statement ( 1 ). Hintikka's general strategy is to view statements of the form (2)
a
as expressing a relation between the person a and the proposition 11�1 expressed by the sentence p . More precisely, let A� be the class of all possible worlds compatible with what a
-ing in w). Let ll�l be the class of possible worlds in which p is tru e . Then (2) is true in world w if and only if
Downloaded from jos.oxfordjournals.org by guest on January 1, 2011
natural language. I t can hardly be expected t o solve a n y philosophical disputes about the true nature of i magining or any factual disagreements about the psychological laws that imagination in fact satisfies. Still, as a vehicle for conceptual analysis it may enhance our understanding of the characteristics of imaginatio n . In this respect, it is comparable to the phenomenology of imagination (cf. Sartre 1 972; Casey 1 976) which attempts to achieve the same goal by a different method. A phenomenologist tries to give a descriptive account of imagination by introspectively studying selected examples of "intentional" acts of imagi ning, while a logician is primarily analysing language, i .e . , the structure and the meaning of linguistic expressions with the word "imagine". In spite of this methodological difference, there is a close parallelism between the phenome nology and the logic of imagination. Again a similar situation obtains in the case of perception (cf. Hintikka 1 975; Smith and Mcintyre 1 982; Dreyfus 1 982). The intentionality of the acts of imagination is reflected on the level of language as the intensionality of the descriptions of imagination. The differ ent ways in which an act of imagination can be directed towards an object can also be nicely captured in the logic of imagination.
211 (3)
A}r,
3.
It is an immediate consequence of the condition (3) that the logic for I a satisfies the principles (l l) (12) (13)
I a(A:::::> B ) :::::> (Ia A:::::> I a B) I a(A AB) = (Ia A Ala B) I aT, if T is a tautology
While (II) and ( 12) seem quite natural, (13) is clearly problematic. Is it really true to say of each agent a that he or she imagines that Bacon is Shakespeare or Bacon is not Shakespeare? Perhaps our agent a has not ever even heard about these two remarkable historical characters. How could they then play any role in his or her imagination? These questions are related to the famous problem of 'logical omniscience' in epistemic logic: K aT holds for every tautology T. Hintikka (1975) and Rantala ( 1 982) have shown h ow to avoid this undesirable consequence of(3). Thus, Rantala's technique of impossible worlds could be adapted to the logic of I a, so that we get rid of (13). The problem of logical omniscience is one of the motivations for the situation semantics that Barwise and Perry have developed as an alternative to
Downloaded from jos.oxfordjournals.org by guest on January 1, 2011
i.e., p is true in all the q>-worlds in A�,
212 the possible worlds semantics of propositional attitudes (see Barwise 1981; Barwise and Perry 1981). Barwise's basic idea is to analyse perception sen tences with embedded 'naked infinitives' like "Esa saw Ingmar run". Let NI(p) be the naked infinitive form of sentence p. For example. NI(p) is " Ingmar run", when p is "Ingmar runs". Then Barwise's suggestion is as follows: (4)
"a sees NI(p)" is true in world w if there is a scene s in w such that a sees s and s supports the truth of p.
(5)
a sees NI(p)� p .
I n H intikka's approach, seeing is an intensional operator Sa which i s follow ed by a description of the propositional content of a's perception . Therefore, the force of Sa p is something like "a seems to see that p" or "It appears visually to a that p", so that the principle (6)
S a p� p .
i s not valid. For this reason , Hintikka's system is able to treat phenomena related to the intentionality of perception - such as visual illusions and hallucinations (cf. Niiniluoto 1979, 1982). Moreover, if we allow perspectival quantification over events, then the Hintikka-type analysis can also give an adequate treatment of the extensional naked infinitive perception reports. (Cf. Niiniluoto 1982; Saarinen 1983; Vlach 1983; and H igginbotham 1 983.) Barwise's strategy (4) does not work for imagination sentences. If there were naked infinitive imagination reports in English (such as "John i magined Mary run"), then an analysis along the lines of(4) would lead to the undesira ble result (7)
a imagines NI(p) ::> p .
(Cf. Higginbotham 1983:120.) I t is clear that imagination should not satisfy the principle (8)
I ar- � p .
Downloaded from jos.oxfordjournals.org by guest on January 1, 2011
A scene s i n world w is 'a visually perceived situation' in w; a situation i n w is a partial submodel of w, i.e., a configuration of certain objects of w having properties and relations to each other and located in space and time. A situation s supports the truth of only those sentences p whose truth value is determined by the perceived scene. For Barwise, seeing a scene is an extensional relation between a perceiver and a part of the world. Thus, his definition (4) implies
213
Downloaded from jos.oxfordjournals.org by guest on January 1, 2011
To treat imagination within situation semantics, it should not be assumed that imagined scenes are parts of the world- at least not parts of the world external to the mind of the imaginer. However, if we allowed 'imaginary scenes' that have a real mental existence, the relation between the imaginer and his scenes would not be extensional- as the analysis of type (4) requires. Barwise and Perry, in Situations and Attitudes (1983), outline a general theory of attitudes, where 'abstract situations' are used for the 'indirect classification' of psychological states. As this theory is intended to be applic able to such attitudes as belief, perhaps it works also for the case of imagina tion. It seems, however, that this approach would not eliminate the problems that arise from the intentionality of imagination- and that the extensional application of situation semantics has to be complemented with ideas from possible worlds semantics. These points can be illustrated by an example. It also serves to suggest a somewhat modified possible worlds analysis of imagination which shares with situation semantics the virtue that the principle (13) does not hold. Assume I imagine that Ingmar is dancing with a blond girl. According to a simple-minded analysis (criticized by Ryle 1949), this means that I am as it were 'looking with my mind's eye' at a picture or a 'mental image' oflngmar with a blondie- and perhaps even hearing a tune 'with my mind's ear'. In the terms of situation semantics, this might perhaps be expressed by saying that imagining amounts to seeing a mentally existing situation- or an imaginary configuration of some objects with some properties and relations. Apart from the difficulties with the postulation of'my mind's eye', the idea of a movie theatre inside my head showing mental pictures suggests a too passive model of imagination: we are not the observers of our mental images, but we create and recreate them through our acts of imagining. Whether these acts are sensuous or nonsensuous (cf. Casey 1977:43), they are conceptual relative to our 'language of thought'. When we try to describe the content of an act of imagining, we notice that our 'mental images' are more or less unspecific in many of their relevant features. Is Ingmar wearing a dark suit or a tail-coat? Who is the blond girl? Are they dancing a waltz or a tango? What is the time and place of this scene? How many other persons are there in the dim margin area of my picture? Hence, to describe a picture with some unspecific features is equivalent to giving a systematic list of all complete alternatives that this picture allows (Ingmar-in-tail-coat-with-Marilyn-Moroe-dancing-waltz-in-this-room etc.) These complete alternatives correspond to the 'possible worlds' in theclassA :f. I defined in Section 2. It is instructive to compare this situation to the case where I actually see that Ingmar is dancing with a blond girl. Ifl don'tseeclearlywhat Ingmaris wearing,I can observe it by going closer to him. My expectations may thereby get fulfilled, and my anticipations may turn out to be correct or incorrect. In the case of imagination these concepts do not make sense (cf. Casey 1977: 168). Still,ifmy image oflngmar is indefinite with respect to his suit,it is up to me to decide this matter by continuing and sharpening my imaginative activity.
214 It is natural to assume that the complete alternatives in the classesA � s and '
A� 1- corresponding to the exampies of seeing Ingmar and imagining Ingmar,
resj,ectively - are described in the same conceptual framework (cf. Casey 1977:134). Moreover, there is no reason to regard thesealterna tivesascomplete possible worlds where each fact is determined (i.e.,each descriptive sentence is either true or false). For example, it is entirely irrelevant to my imagination about Ingmar's dance whether Ronald Reagan is the President of the United States or not. We may thus assume that these alternatives are what Hintikka (1975, 1983) calls small worlds or what Barwise calls situation-types. Condition (3) can then be replaced by the requirement (9)
Each alternative inA � 1 supports the truth of p. '
Esa imagines thatlngmar is dancing with a blond girl. Esa imagines that lngmar is dancing. . ·.
but blocks inferences like Esa imagines that Ingmar is dancing. Esa imagines that Ingmar is dancing or the moon is cheese. . ·.
4.
It was noted above that imagination should not satisfy the success condition (14)
Ia A:JA.
Ifl imagine that I am Francis Bacon,it does not follow that I am Francis Bacon. Indeed, one might even suggest an anti-success condition for Ia:
For example, in L'Imaginaire (1940) Sartre claims that there is an intimate connection between imagination and nothingness:
"But if! imagine Peter as he mightbeatthismomentin Berlin-orsimplyPeteras he exists at this moment (and not as he was yesterday on leaving me), I grasp a n object which i s n o t a t all given to m e or which i s given to m e simply as being beyond reach. There I grasp nothing, that is, I posit nothingness. In this sense the imaginative consciousness ofPeter in Berlin (what is he doing at this moment? I imagine he is walking in the Kurfilrstendamm,etc.),isverymuchclosertothatof the centaur( whose complete non-existence I proclaim), than the recollection of
Downloaded from jos.oxfordjournals.org by guest on January 1, 2011
This definition licences inferences like
215 Peter as he was the day he left . What iscommon between Peterasan image and the centaur as an image is that they are two aspects of Nothingness." (Sartre 1 972:210-2 1 1 ).
But is it possible to imagine a logical contradiction? If I imagine that Ingmar is both dancingand notdancingatthe same time,thentheremustbe two Ingmars in my field of imagination. This is a violation of the laws of nature (the same physical individual appears in two locations at the :;arne time), but it is not a logical contradiction. And who haseverbeenabletoimaginearoundsquare? As some people have thought that they can do this,it is possible to imagine thatone imagines a contradiction. But still it seems clear that the principle
where 0 is the operator of logicalpossibility, should be accepted. With ( 12),( 17) entails the principle
5.
Let us go back to the example where Esa imagines that Ingmar is dancing with a blond girl. Then a well-defined physical individual, Ingmar, enters Esa's imaginary worlds. This can be expressed by Hintikka's physica/ quantifier(Ex) as follows (Ex)I Esa (x = Ingmar & xdances). Here the variable x ranges over'physical world-lines', i.e., functions that pick out from each imaginary world the same, 'physically cross-identified' indivi-
Downloaded from jos.oxfordjournals.org by guest on January 1, 2011
However, on closer analysis it seems that this passage is not intended to support principle (15) at all. We may agree with Sartre that his image of Peter in Berlin is not'given' to him in the same way ashisperceptionsandmemoriesof Peter-and this image is probably wrong. But, it seems to me,it would be too much to claim that this image must be wrong: it may happenafterall that Peterinfactiswalking in the Kurfilrstendamm when his friend Jean-Paul is imagining that he is doing so. Therefore,(15) as a general principle should be rejected. It is certainly possible to imagine situations and courses of events which are physically impossible. This is shown both by surrealist novels and films and by science fiction, which play with effects thatarisefrom violationsofnaturallaws. If OPh is the operator of physicalpossibility, then the followingprinciple is not valid:
216 dual. On the other hand, the personal identity of the blond girl is unspecific- in one possible world she could be MarilynMonroe, in anotherDoris Day,etc. The function that picks out from each alternative the blond girl in that world is a 'perspectival world-line'-it cross-identifies those girls who play the same role in Esa's field of imagination as the dancing partner of lngmar. ( Similarly, perspectival world-lines in the case of perception identify the individuals who play the same role in the perceiver's visual space.) If ( ffx) is the perspectival quan tijier which ranges over perspectival world-lines, then we can write (3y)IEsa ( y is a blond girl &y dances with lngmar). Hence,
In the logic of perception, the sentence "a sees b" with a direct object construction can be formalized in two radically different ways:
( 10) ( 1 1)
( 3x)Sa( x=b) ( 3x)( x=b &Sa (3y)(y=x)).
Here ( 10) says that something visually appears to a and a identifies it as b. For example, "Macbeth sees the ghost of Duncan" has the form ( 10). A visual illus ion, where a sees bas c, can be expressed as follows: ( 1 2)
( 3x )(x=b &Sa( x=c)).
Thus, ( 1 1 )says in effect that aseesbasexisting, butitdoesnottellhowaidentifies what he sees. As H intikka ( 1975) requires that a's perspectival world-line canbe continued to an individual bin the actualworldonlyifthere isacausalchain from b to a's perception, it is natural toread ( l 1 ) as "a looks atb"(cf. Niiniluoto 1 979, 1 982). The counterparts of( 10), ( 1 1 ), and ( 1 2) in the logic of imagination are ( 13)
( 3x)Ia( x=b)
( 14)
( 3x)(x=b &Ia( 3y)(x=y))
( 15)
( 3x)(x=b &Ia( x=c))
Here ( 1 3) and ( 15)are special cases of the formulas ( 16)
( 3x)Ia F( x)
( 1 7)
( 3x)(x=b &Ia F(x))
Downloaded from jos.oxfordjournals.org by guest on January 1, 2011
( Ex)(3y)IEsa ( x = Ingmar &y is a blondie &x dances with y).
2 17 where F(x) is a formula· with x as a free variable. To illustrate the expressive power of these formulas, let us first consider Sartre's favourite example, where he imagines that Peter is walking on the streets of Berlin. According to Sartre, perceiving Peter and imagining Peter are two different ways in which our consciousness can be related or directed to the same object: "The imaginative consciousness I have of Peter is not a consciousness of the image of Peter: Peter is directly reached; my attention is not d irected on an image, but on an object." (Sanre 1 972:5.)
(17)*
a imagines of b that he is an F
Hence, ( 14) can be read by (14)*
a imagines of b something
or ( 14).. ••a is imagining about b The term "b" in formulas (14) and (17) is outside the scope of the operator Ia, and its occurence is therefore referentially transparent. It also follows from these formulas that (3x)(x=b), i.e., b exists in the actual world. The analogy with the logic of perception is not complete, however, since the cross-identification between the actual world and the imaginary worlds need not take place through causality. This is a crucially important feature of imagination -a point that Sartre emphasizes in his example discussed above. If I imagine Ingmar as Ingmar, i.e., as a person with his familiar appea rance, then my act is of course causally related to my earlier perceptions and memories about him, but not necessarily to his present existence. Further, imagination is 'free' in the sense that I can also imagine what some existing person whom I have never even seen looks like. What is more, I can imagine lngmar as someone else. This situation is expressed by formula ( 15) when b:Fc, which thus can be read by ( 15)
•a imagines b as c.
Here "c" occurs within the scope of Ia, and hence in a referentially opaque position. Assume, for example, that I am following lngmar's lecture, and I amuse
Downloaded from jos.oxfordjournals.org by guest on January 1, 2011
Thus, Sartre imagines of Peter that he is walking in Berlin. Similarly, Esa imagines of lngmar that he is dancing. These sentences can be formalized by (17) - which therefore has the reading
218 myself by imagining that he is Alfred Tarski. To do this, I need not have any picture of Tarski in my mind I just stretch my imagination to think that the lecturer who looks like Ingmar is Tarski. Then it is true to say that I non-sensuously imagine Ingmar as Tarski. There is no causal connection between Ingmar and my perspectival world-line in this case. On the other hand, suppose I sensuously imagine a doll as Marilyn Mon roe. In this case, there is a causal connection between my earlier perceptions and memories of Marilyn and my image of Marilyn, but not between the doll and llJY image. There is also an interesting overlap of my fields of perception and imagination: when I see the doll as a doll, but imagine that it is Marilyn, we have the following situation -
If I imagine a doll as Marilyn , then I also imagine Marilyn in the sense of formula ( 1 3): there is a perspectival world-line which picks out Marilyn from my imaginary worlds. Thus, (13) and ( 1 6) have the natural readings (13)
•a imagines b.
(16)
•a imagines an F.
Here 'b' occurs in a referentially opaque position. Unlike (14) and ( 1 5), formulas ( 1 3) and (1 6) do not entail that b exists. They thus cover two kinds of cases: imagining a thing or person which also has real existence (Ingmar, Marilyn, a horse, etc.), and i magining something 'purely imaginary' or fictitious (the present king of France, Anna Karenina, a unicorn, etc.). Formulas ( 1 3) - ( 1 7) illustrate different meanings that the direct object construction "a imagines b" may have. It is important to note their common feature: they are all defined in terms of the operator I a which takes proposi tional that-clauses. Hence, what Casey ( 1 976) calls imaging is a special case of
imagining-that. 2 6.
Sartre points out that "cases may be cited in which I produce an image of an object which has no real existence outside of myself'. But, he adds, the chimera does not exist even "as an image" (Sartre 1 972:5). On the other hand, Brentano would say that an imagined unicorn - towards which our act of i magining is "directed" - had "intentional inexistence." In his analysis of a " naive hallucination", where Macbeth sees a dagger before him and believes what he sees, Smith (1983) comes to the conclusion that Macbeth's percep tion has "no object". More precisely, he rejects the alternatives that the objects of naive hallucinations could be existent objects, nonexistent objects, objects in possible worlds, sense-data, intentional objects, or objects that exist only insofar as they are perceived.
Downloaded from jos.oxfordjournals.org by guest on January 1, 2011
(3x)(3y)(x=y=the doll & S a(x is a doll) & I a(x is Marilyn)).
219
7. One interesting aspect of imagination has so far been ignored in our discussion. Consider some fictional character, like Walt Disney's Donald Duck or Anna Karenina in Tolstoy's novel. While it is never possible for me to perceive Donald Duck (I can only see pictures and films about him), I can imagine various things about him - e.g., that he enters this room now. Similarly, I cannot see Anna Karenina, but I can imagine what she looks like. In such cases, Donald and Anna not only play a role in my imaginary worlds in the sense of formula (13), but I also seem to be able to imagine of them something in the sense of formula (17). In other words, it seems that acts of imagination can be directed towards fictional entities as well as real ones. This idea leads us directly to the notorious problems within the semantics of fictional terms. Assume I imagine of Anna Karenina that she has green eyes. This cannot be formalized by a sentence of the form (17), i.e., by (18)
(3x)(x=Anna Karenina & I a (x has green eyes))
since (18) entails that Anna Karenina exists in the actual world. We also stumble on the problem of 'intentional identity' (cf. Saarinen 1979): if you and I are both imagining something about Donald Duck, in what sense are our acts directed toward the same object? As Donald does not exist in the actual world, this cannot be formalized by (19)
(3xX3y)(x=y=Donald Duck & I a(x is . . . ) & I a(Y is . . . )).
Downloaded from jos.oxfordjournals.org by guest on January 1, 2011
Smith attributes to Hintikka the view that Macbeth's hallucination is "intentionally related to at most one object in each of the many possible worlds compatible with what is presented in the experience" . But instead of saying that a perception is "directed toward something in each world compa tible with what the subject sees", H intikka (1969, 1975) himself has repeated ly argued that the proper counterparts of 'sense-data', as the objects of perception, are the perceptual world-/ines.3 These 'entities' are intensional in the sense that they do not belong to one world - to the actual world or to any single possible world. There is thus an important difference between saying that perception is multiply directed to objects in several possible worlds (Smith's interpretation) and that perception is directed to a world-line. There are quantifiers that range over world-lines, but still these lines exist no where'. The status of 'mental images' in the logic of imagination is similar to Hintikka's reconstruction of sense data in the logic of perception. To assume that such intensional entities exist in one world would be j ust as objectionable (cf. Ryle 1949, Ch . VIII) as the corresponding existence assumption about sense data.
220 It seems to me that we have to make a distinction between private and public fictional entities. If I now create in my mind some fancy character, no
(20)
Let
us
imagine that T!
As a recom mendation, (20) is neither true nor false; and as an imagination statement it does not commit a to defending the truth of T. This analysis is compatible with the possibility that a nevertheless intends to convey some interesting truths by his story: even if the actual world is not a model of T, the text T may have logical consequences p which are actually true. These truths need not even be consciously thought by a - he may only have an 'intuition' that the story T e ntails something vitally important about the world or about human nature. Thus, in spite of the facts "a imagines that T'' and "T entails p", it need not be the case that "a imagines that p". This follows from our treatment (9) of imagination statements. In writing the text T, the author attempts to describe the propositional content of his imagination. In terms of Section 2, this means that he should
Downloaded from jos.oxfordjournals.org by guest on January 1, 2011
one is able to refer to it or to direct any mental acts towards it. But I can make this private 'entity' public by giving a name to it, by describing or drawing it, and by telling or writing stories about it. This is precisely what Tolstoy did with Anna Karenina and Disney with Donald Duck. After they published their descriptions of Anna and Donald, other people could then refer to these fictional creatures. Hence, two persons can think different thoughts about the same Donald Duck, but on ly in the sense that their mental acts are causally related to the same public descriptions and drawings of Donald. And it is only through these Donald-descriptions in the actual world that my perspectival Donald world-line can be linked with Disney's own Donald-world-line. These remarks suggest the following treatment of fictional discourse. Let a be the author of a fictional text T (such as a short story, a novel, or a play), where T is conceived as a conjunctive sequence of sentences. For simplicity, we assume that T does not contain any non-fictional sentences, i.e., assertions which do not belong to the fictional story of T. Then in publishing the text T its author certainly is not asserting that T is true in the actual world. John Searle (1979) suggests that a is (nondeceptively) pretending to assert that T. However, with explicitly fictional stories it does not seem to me at all plausible to say that the author 'pretends' to assert anything or 'pretends to refer' to something.4 Instead, a is making a sincere attempt to describe something that he has first imagined in his own mind. Indeed, if a is the author of T, then it must be true to say that a imagines that T (cf. Haapala 1984) - or at least that a has imagined that T. In writing and publishing the story T, the author is then transforming his private imaginary characters to public ones. So a is really engaging in the illocutionary act of recommending his readers to share his imagination, i.e., a's act has the following form :
221
Univumy of He/1/nki D�pt. of Philosophy Unioninkatu 40 B 00170 Hel1inkl 1 7 Finland
NOTES • This paper is a reviud version of a paper which appeared under the title "Remarks on the Logic of Imagination" in G. Holmstrllm and A.J.I. Jones (eds.): Action, Logic, and Social Th�ory (:::: Acta Plrilo1ophlca Fennica vol. 38 ( 1 985), pp. 1 8 3-202. I. For my first attempt to develop a logic of imagination, see Niimluoto ( 1 983). 2. It seems to me clear that imagining-how is also reducible to imagining-that. For example, to imagine how it feels to dance with Marilyn Monroe, I imagine that! dance with Marilyn - with feelings. 3. For a treatment of hallucinations within Hintikka's logic of perception, see Niiniluoto ( 1 979). Smith's distinction between "naive", "neutral", and "hip" hallucination would require that the belief operator Ba is combined with S8 . If this is done (cf. Niiniluoto 1979), then we can expres5 the difference between the terms "the dagger now here sensuously before me" and "the dagger
Downloaded from jos.oxfordjournals.org by guest on January 1, 2011
specify the class A�,I of his imaginary worlds. Through his work, he thus "projects"- as Wolterstorff (1980) says- a class of possible worlds. It hardly ever holds, however, that A� I is simply the class Mod(T ) of the models ofT: the author a narrows dow� the class Mod(T) with a set of contextual presuppositions Pa ( cf. Lewis, 1978), which contains at least factual assump tions about the world and semantical conventions about the language of the text T. The class of possible worlds projected by the author a is then Mod(T&Pa). Everything that is true in this class, i.e., follows logically from T &Pa• is true in thefictionT. Everything that is false in this class isfalse in the fictionT. It follows that there is a class of sentences true in some elements of M(T &Pa) and false in others.These sentences are indetermined in the fiction T. A different class of possible worlds Mod(T &Pb) is projected by a reader b who reads the textT with his own presuppositions Pb, where Pb and Pa may be more or Jess similar to each other. In reading about the fictional characters that a made public throughT, the reader b may then use.his own imagination and produce a more or Jess vivid private image of the fictional world describ ed by T. In this way, each reader of a literary work of art will have an active role in 'constituting' for himself the imaginary world that this work speaks about. The communication between the author a and the readers b of a fictional textT is made possible primarily by the shared presuppositions in Pa and Pb which help the reader to understand the intended meaning of the author. Through the public character of the textT, the author and the readers are able to think about the same fictional objects. But, as the presuppositions Pa may be partly unconscious to the author a, an interpreter b may claim, as it often happens, that he understands the text T better than its author.
222 actually now here sensuously before me" (cf. Smith 1983: 108).
4. Searle rejects the view that "fiction contains different illocutionary acts from nonfiction"
(such as 'telling a story'), since this would commit us to the clatm that "words do not have their
normal meanings in works of fiction" (Searle 1979: 64). This is a mistake, since ordinary words may have the same meanings in two different illocutionary acts, e.g., in assertions and questions.
REFERENCES Barwise, J., 1981: Scenes and Other Situations,
Journal of Philosophy 78: 369-691. Journal of Phtlosopy 78: 668-691. Barwise, J. and Perry, J., 1983: Situations and Allitudt!s. MIT Press, Cambridge. Mass. Casey, E.S., 1976: Imagining; A Pht!noml!nological Study, Indiana University Press, Barwise, J. and Perry, J., 1981: Situations and Attitudes.
Bloomington. Mass. Haapala, A., 1984:
Fikuojo 10dt!l!isuus. Helsinki: Helsingin yliopiston yleisen kirjallisuustieteen,
teatteritieteen ja estetiikan laitoksen monistesarja n:o 12.
Higginbotham, J., 1983: The Logic of Perceptual Reports: An Extensional Alternative to Sttuation Semantics.
Journal of Philosophy 80: 100-127. Modt!ls for Modalillt!S. Retdel, Dordrecht. Hintikka, J ., 1975: Thl! lntl!ntions of Intl!ntionality. Reidel, Dordrecht. Hintikka, J., 1983: Situatioll5, Possible Worlds, and Attitudes. Synthi!SI! 54: 153-162. Lewis, D., 1978: Truth in Fiction. A mmcan Philosophical Quartuly 15: 37-46. Niiniluoto, 1., 1979: Knowing that Ones Sees. In: E. Saarinen 1!1 a/. (eels.), Essays in Honour of Jaokko Hin1ikko. Reidel Dordrecht. pp. 249-282. Niiniluoto, 1., 1982: Remarks on the Logic of Perception. In: Niimluoto and Saarinen (1982), pp. Hintikka, J., 1969:
116-129. Niiniluoto, 1., 1983: On the Logtc of Imagination. In: I. Patoluoto 1!1 a/. (eels.),
Vt!xing Qut!stions.
Reports from the Department of Philosophy, University of Helsinki, N:o 3, pp. 23-28. Niiniluoto, I. and Saarinen, E. (eds.), 1982;
lntt!nsionol Logic; Thl!ory and Applications. Acta
Philosophica Fennica vol. 35, Helsinki. Rantala, V., 1982: Impossible Worlds Semantics and Logical Omniscience. In: Niiniluoto and Saarinen
( 1982), pp. 106-115. Tht! ConCI!pt of Mind, Hutchinson, London.
Ryle, G., 1949:
Saannen, E., 1979: Intentional Identity Interpreted. In: E. Saarinen (ed.),
Gamt!-Tht!orl!tica/
St!mantics. Reidel, Dordrecht. pp. 245-327. Saarinen, E., 1982, Propositional Attitudes are not Attitudes Towards Propositions. In: Niiniluoto and Saarinen (1982), pp. 130-162. Saannen, E., 1983: On the Logic of Perception Sentences.
Syntht!SI! 54: 115-128. Tht! Psychology of Imagination. Methuen, London. Searle, J., 1979; Exprt!ssion and Mt!aning. Cambridge University Press. Cambridge. Smith, D.W., 1983: Is This a Dagger I See Before Me? Syntht!sl! 54: 95-124. Smith, D.W. and Mcintyre, R., 1982; Hussu/ and lnlt!ntionality; A Study of Mind, Mt!aning, and Languagl!. Reidel, Dordrecht. Vlach, F., 1983: On Situation Semantics for Perception. Syntht!sl! 54: 129-152. Wolterstorff, N., 1980: Works and Worlds of Art. Oxford University Press. Oxford. Sartre, J.-P., 1972:
Downloaded from jos.oxfordjournals.org by guest on January 1, 2011
Dreyfus, H.L. (ed.), 1982: Husur/,lnlt!ntionality, and Cognitiwl! Sclt!nCI!. MIT Press. Cambridge,
Journal of SemanticJ 4: 223-236
THE LANGUAGE OF SOCIAL ACTS
BO ELE DE RAAD
ABSTRACT
comprehensive set of 1330 interpersonal verbs was selected from a Dutch dictionary. Both the rationale of the selection criteria for this set, and the reliability of these criteria are discussed and the successive steps m the stocktaking-procedures and some adaptations of the origmal set are described. Fmally, the procedure for structuring the domain of verbs
is explained. Suggestions
are made for the use of the resulting method for the description of social acts, consisting of 45 verb scales.
This paper reports about a stocktaking and a structuring of social acts in terms of verbs. The goal of this project is to develop a method for the description of what people do when they are in each other's presence. The project should result in a taxonomy of discourse situations, to be conceived of as key situations or scenarios that can be used for the understanding of interpersonal situations. Burke (1945) termed such key situations "represen tative anecdotes". The attainment of this goal presumes the availability of a standard vocabulary for the description of social acts. The purported method of data-collecting is particularly useful in research areas where the sequential structure of conversational episodes is investigat ed (Argyle 1980, Edmondson 1981, Stech 1979). For dialogues, for example, this aim can be accomplished by judging the actions of the two interlocutors separately.The two sets of data thus obtained can be interrelated by means of canonical correlation analysis.The use of a provisional method according to this procedure based on the present research gave more insight into the patterns of talk-sequences (De Raad 1985). The method may also be relevant for a variety of other purposes. For instance, where research is focussed on how people make inferences from behaviour about personality (Jones & Davis 1965, Buss & Craik 1983) it is possible to obtain descriptions of the behavioural content of certain traits, using the so called "act frequency" approach described by Buss and Craik (1983). Assuming that dispositional assertions are understood as summary statements about behaviour, Buss and Craik asked subjects to name acts that would count as manifestations of a given disposition. It is also possible to
Downloaded from jos.oxfordjournals.org by guest on January 1, 2011
This paper describes the derivation of a taxonomy of interpersonal verbs which are used for the description of social acts. Starting from a theoretical delineation of the domain of social acts, a
224
Downloaded from jos.oxfordjournals.org by guest on January 1, 2011
obtain behavioural consistency descriptions for single acts as well as for the overall act output of persons. Furthermore it is possible to scrutinize subsets of acts for their role structure, thereby chiming in with the more situational approaches ( cf. Price & Bouffard 1974). Yet another application is possible in the study of the emergence of action categories in the child ( cf. Huttenlocher, Smiley, & Charney 1983). For all these different areas of application the vocabulary for the description of the acts should meet the requirements of standardization and representativeness. The background for this research is the interest in problematic interperso nal situations, an example of which is a situation in which a "pe rson lends something to his partner, and the partner views this as a gift ( and feels, therefore, insulted). Such situations are called problematic when the conver sational partners have different perspectives on how to define their common situation (cf. Pearce 1976). The project was inspired by Kouwer's (1973) fundamental notion rendered in Dutch as "gesprek", henceforth translated as "discourse". The semantic definition of "discourse" is that there are at least two interlocutors, and that there is a content, viz., an exchange of perspectives on a common focus of interest. "Discourse perspectives" can be roughly described as social acts by means of which people do something to each other ( Helwig 1958). The term discourse-situation indicates an interest in relatively short-lived concrete interactions. A reason for this interest is that problematic situations mainly appear in shortterm interactions. Relatively long-established relations such as admiration, leadership, or submission are taken into account indirectly, since such relations can only be studied via the observation of concrete behaviour. Kouwer conceived of discourse as an indisputable datum (cf. Hofstee 1974), much in the way that Watzlawick, Beavin, and Jackson (1967) conceived of interaction. Discourse cannot consistently be denied, because already its denial would presuppose dis course: we always start with talking to somebody. Although discourse as a presupposition is an indisputable datum, we can talk about discourse as an object. Necessarily, possible descriptions of this object do not coincide with discourse as a presupposition.Thus, there is an absolute separation between discourse as a presupposition and discourse as an object of discourse. In everyday life this separation becomes apparent when the discourse partners no longer accept their situations as being self-evident or "natural" ( Messin ger, Sampson, & Towne 1962). If the fact that everyday life is taken for granted is itself object of reflection in relation to the original definition of discourse, the possibility of a confrontation of perspectives about the mea ning of the situation is given.That is, the situation becomes one of listening to ..., explaining to ..., agreeing with ..., etc. Thus, the reflection gives rise to possible themes of discourse. Finally, in order to make these themes discussa ble, they have to be described in terms of social acts. The description of the perspectives of both discourse-partners presumes the same standard vocabu lary. The present aim, then, is to supply a vocabulary for the description of tokens of attestations which are implicitly or explicitly issued in discourse
225
Downloaded from jos.oxfordjournals.org by guest on January 1, 2011
situations. These tokens can be referred to as "moves" ( Goffman 1976, Grimshaw 1982), interactional units in which cues of intenton, content, and context are incorporated into a single act that carries the plot the partners have in mind for each other, for example: admit, warn, assert, threaten, deny, or request. The vocabulary aimed at is not restricted to "speech-act verbs", since the aspects of discourse situations to which verbs refer are often not connected with speech.The vocabulary thus constitutes the semantic field of the language of social acts ( cf. Nowakowska 1981; cf. Verschueren 1981). A first approach towards such a vocabulary is Austin's (1962) classifica tion of illocutionary acts. This classification was criticized and extended by Searle (1975). Both Austin's and Searle's classification of illocutionary acts turned out to be untenable, not so much because their classifications comprised only small numbers of verbs, but rather because of the poor definition of their classes, which prevents an enlargement or a completion of the classifications ( Ballmer & Brennenstuhl 1981). There are some other analyses of verb-domains, but these are performed within specific fields of interests, such as verbs of judging ( Fillmore 1 969), verbs of perception ( Scovel 1971), verbs of motion (Talmy 1975}, and verbs of cooking ( Lehrer 1974). A recent large-scale analysis of the lexicon of verbs was performed by Ballmer and Brennenstuhl (1981; cf. also Ballmer 1983, 1984). Although different in method and goal from the present one, their approach offers interesting perspectives towards the building of a thesaurus of thematically organized act schemes, or "representative anecdotes" ( Burke 1945). Their analysis consisted of four steps in which they taxonomized approximately all German verbs. The first step involved an inspection of the entire set of verbs in order to extract so-called "semantic centers". The result of this first step was a rough grouping of verbs according to the analyst's "rather vague intuitions about the semantic content of expressions" ( Ballmer & Brennens tuhl 1981). Although the second and the third step involved more rigorous categorization procedures, mainly on the basis of similarity judgments per formed by the researchers themselves, the resulting classification is probably strongly influenced by the original first grouping. During the fourth step the verbs within the resulting categories were ordered according to their mutual presuppositions, so that act-models could be developed. Procedures such as followed during this last step might be used in future research on the taxo nomization of discourse-situations. Another approach is that of Osgood ( 1970), which was oriented towards the development of procedures for discovering the semantic features of word forms. His approach, which was about interpersonal verbs, is more in line with the present purpose. However, it was based on only a rather small sample of verbs ( 210). The present approach took as its starting point the conception of discour se-perspectives in terms of social acts. A social act is defined as a "move": an interpersonal directed activity which exists from the time it is instigated until
226
Collecting the Verbs Both the collecting of the verbs and the description of the criteria used were carried out by three persons, henceforth called the "research-team". Starting from the assumptions, that (1) everything people find of interest in interpersonal behaviour, is sedimented into language, and (2) what is thought or said about behaviour can best be represented by verbs, all verbs that might be used to refer to interpersonal behaviour in one way or another were collected from a Dutch dictionary ( Koenen & Endepols 1966), and were thereupon subjected to structuring procedures. From the start the research-team had only a vague notion of the definition of a relational verb, prompted by Kouwer's (1973) description of perspectives i n discourse. This original notion was indicated by such phrases as "the verbs should de!.cribe something that is 'social"' and "the verbs should describe what people do to each other". After some provisional criteria for the
Downloaded from jos.oxfordjournals.org by guest on January 1, 2011
it can be said to have been executed ( cf. Fiske 1978). An act thus defines a certain interval of time as an episode ( Harre & Secord 1972, Forgas 1976, Sjoberg 1981). Moreover, a social act should be perceived as a form of intentional behaviour directed towards a goal (Allwood, 1980). Descriptions of interpersonal activities by means of act-describing verbs thus provide the ways actors define their situations ( McHugh 1968, Watzlawick et al. 1967, Jessor 1981). Rommetveit (1981) pointed to the intricacies involved in interpreting acts, and stressed the multiplicity of possible accounts for any given act. Acts can be described at different levels of meaning. Any social act can be connected with a bundle of communicative intentions (cf. All wood 1980). Which level of meaning will become salient is in part contingent upon the interests of the participants involved ( cf. also Osgood 1970). An important criterion for choosing a level of description in this respect is that the ascribed meaning should readily become shared in talk.This means that the vocabulary should comprise virtually all levels of descriptive terms provided that they can be used and usually are used to describe the moves of a single participant in the interaction-situation. A first procedural step in studying the language of social acts involves a rule for the selection of the descriptive terms, i.e., verbs. The criteria for admission of verbs should be formulated. Furthermore, the set of admitted verbs should be reduced in such a way that a list results that exhaustively represents the meanings within the domain, and which is workable for observers. The result of the reduction procedures will have much in common, qua meaning potential, both with what has been written about dimensions of interpersonal behaviour (e.g. , Benjamin 1974, Wiggins 1980) and what has been said about coding procedures for analyzing relational communication ( e.g., Rogers & Farace 1975).
227 selection of the relational verbs from the dictionary had been determined, the actual selection of the verbs took place. The three members of the research team performed this selection independently. They discussed their results. In cases of disagreement, criteria were specified for the admission or the rejec tion of a verb. At the end of the selection procedure the criteria were formulated or reconstructed from the notes that were taken during the discussion sessions. The criteria formulated were rather extensive in descrip tion and were therefore subjected to thorough consideration, This led to a distinction between criteria that were considered to be of practical impor tance, such as 'no obsolete words' ( e.g. scourge), 'no technical terms' (e.g. disinherit), 'no jargon', etc., and others that were considered to be of principal importance. The latter criteria were formulated as follows: The verb presupposes two or more persons (or parties); it does not denote a purely private activity. E.g., advise: meets this criterion; think: does not meet this criterion.
2.
The verb pertains to behaviour about which somebody ( whoever) can make up his mind to do it ( intentionality); it does not concern behaviour that befalls somebody. E.g., blinking( eyes): meets the criterion; blush: does not meet the criterion.
The result of the selection procedure was a set of 1 330 verbs. InTable 1 some examples of verbs are given which were deleted on account of the criteria of principal importance. Criterion
I
rain climb th� pol� stack up (hay) multiply mummify pl�at
Criterion
II
M d�scend�d from g�t mouldy rust through flick�r hiccup hanku afur
Tab/� I. Examplesof verbsdeletedon account of the principal criteria
In order to reduce the set of verbs as much as possible while retaining most semantic differences, the set was scanned for possible synonyms. This resul ted in groups of synonymous verbs. Within these groups those verbs were retained that were most current or that could be judged least synonymous.
Downloaded from jos.oxfordjournals.org by guest on January 1, 2011
I.
228
The application of this procedure resulted in the removal of 344 verbs from the list of 1330. The remaining set of 986 verbs was used as input to the structuring procedure.
Reliability ofthe Principal Criteria
Judges
+ Research
+
Team
-
Total
Tab/� 2:
Total
-
77
18
95
6
883
889
83
901
984
Cross-classtfication of Verbs for Judges and Research-team
On the basis of this judgment task we were also able to estimate the total number of verbs ( meanings), relational and non-relational, in the dictionary used. This dictionary counted some 1350 pages. The total number of verbs can therefore be estimated at 13.5 X 984 = 13,284 verbs.
Structuring the Stock of Verbs The final list of 986 verbs had to be reduced in order to become manageable. Starting-point for the reduction process was the aim to reduce while retaining as many distinguishable meanings as possible. The procedure followed was described by Miller (1969), and is henceforth called the "Miller Procedure". This proce1ure involves analyzing similarity ratings achieved by means of a sorting task.
Downloaded from jos.oxfordjournals.org by guest on January 1, 2011
A balanced sample of 100 pages was selected from the dictionary. All the verbs that appeared on these pages were listed and handed to subjects to be judged on the two criteria. Seventeen subjects took part in this task. Subse quently, they had to judge half the verbs on one criterion, and the other half on the other criterion. Subjects were instructed to indicate whether the verbs met the criteria or not. In case they did not know the meaning of a verb, they had to cross that verb out. The number of times a verb was crossed out was subtracted from the number of times it was marked as a relational verb by the total number of subjects. After this procedure a verb was considered a relational verb if it was passed by more than half of the subjects on both criteria. There was a total of 984 verb-meanings. The research team had selected 95 of these, of which the 17 judges selected 77.The total number of verbs selected by these judges was 83.The results a recross-classified inTable 2.The phi-coefficient is .85, which is sufficiently reliable.
229
General Procedure
Subtasks Subjects: A total of 500 paid subjects took part in the subtasks. All of them were students, mainly from the Social Science departments. Because Miller (1969) suggested that the number of sorters should be about the same as the number of words to be sorted, for the first two samples of 100 verbs each, we started with 100 sorters. However, anticipating the results of the invariance analysis, it was concluded that 50 subjects could also reliably sort 100 verbs. Therefore, in the remaining eight subtasks only 50 subjects took part in each task.
Results: Each of the ten subtasks yielded a different number of factors. In order to retain as many distinguishable meanings as possible, two criteria
Downloaded from jos.oxfordjournals.org by guest on January 1, 2011
The words to be sorted were typed on cards.The cards were given to subjects with the instruction to sort them into non-overlapping groups that have about the same meaning. The subjects were allowed to determine the number of categories they wished to create themselves. Therefore, the number of words per category and the number of categories per subject vary from category to category and from subject to subject. The sorting-information was converted into an n x n matrix, in which n is the total number of verbs in the sorting task (cf. Miller 1969). The cell numbers indicate the number of times the verbs in question were put together in a category. Since the categories were non-overlapping, the diagonal num bers equal the total number of subjects. Dividing the matrix by the number of subjects yielded a matrix of means of cross-products. This last matrix was subjected to raw-score factor analysis by means of Principal Components. Because the total amount of verbs was considered too large for a single sorting task, a two-step procedure was followed. First, the 986 verbs were randomly assigned to ten groups of about 100 verbs each. These groups of verbs were factored separately (subtasks). Second, the resulting factors from each of the ten analyses were again written on cards, represented by the verbs with the highest loadings, and these were again subjected to similarity ratings and factor analysis ( main task). In order to draw conclusions about the consistency of the categorization by subjects, or the interchangeability of subjects, for each of the subtasks two matrices were composed for each half of the subjects. On one such matrix Principal Components Analysis was performed, and on the other Principal Components followed by Varimax rotation. Thereupon, the matrix of load ings, obtained from the first analysis, was rotated to the matrix of loadings obtained from the second analysis, through simultaneous one-sided ortho gonal rotation ( cf.Ten Berge, 1977: 12). Refinding a factor in the replication is an indication of the reliability (invariance) of the factor.
230
were used to determine the number of factors in each of the ten tasks.The first was the eigenvalue > I criterion. The second was the interpretability of the factors . InTable 3 some results are presented.The coefficients of congruence indicate the in variance of the factors.Those above .80 may be considered reli able, those between .71 and .79 as moderately reliable, and those below .71 as unreliable.
Main Task
Subjects and Procedure: In this task 150 subjects took part. Most of them were students from the Social Science departments. The subjects sorted the 226 factors according to the Miller Procedure. The mean number of categories was 24.7 with a standard deviation of 14.6. The sorting results were subjected to factor-analysis.
Sam pte• Means Sd's Factors•• % of var.
18.7 9. 1 25 63
2
3
4
5
6
7
8
9
10
1 8.4 9.4 25 63
1 7.6 9.2 27 66
1 3. 7 7.5 22 63
24.4 1 2.8 28 62
1 8.4 1 0.0 27
19.5 9.4 25
64
64
1 8.2 1 0.9 23 65
1 9.0 9.0 25 65
23.4 1 2. 2 27 62
Coefficients of Congruence (CC) CC>.79 . &O>CC>.70 . 7 l>CC
16 2 7
14 4 7
12 5 10
10 4 8
14 3 II
7 4 16
II 4 10
12 2 9
13 2 10
10 4 13
Number of Factors used•••
23
25
24
19
25
23
23
17
23
24
(•) (••) (•••)
The number of verbs in the samples I to 6 was 1 00; for the samples 7 to 10 the numbers were 98, 96, 96, 96 respectively. The total number offactors was 254 See "Main Task"
Table 3: Mean number of categories, standard deviations of number of categories, number of principal components with Eigenvalues> I , percentage of variance persubjects,andcoefficientsof congruence
Downloaded from jos.oxfordjournals.org by guest on January 1, 2011
Not all of the factors obtained in the subtasks were useful for the main task. In the first place only those factors were considered that had at least two verbs with a loading > .40. In the second place, those verbs were deleted that had a higher loading on another factor. After this, factors in which at least two verbs were left with loadings > .40 were retained. These factors were repre sented on cards by the 2, 3, or 4 highest loading remaining verbs. The result was that of the 254 original factors, 226 were retained.
23 1
Results: Principal Components Analysis yielded 1 5 factors meeting the eigen value criterion of 2.5. These factors were rotated according to Varimax. The total amount of variance explained by these factors was 4 1 %. The 15 factors are to be conceived of as the basis for the required method of data-collecting. Interpretation of the factors based on the loadings which exceeded .40 are provided in Table 4 (in capitals). From each of the 15 factors three groups of verbs were selected, each group consisting of three verbs. The list of 45 groups of three verbs each ( cf. Table 4) constitutes the method for the description of social acts. Only those verbs were incorporated into that list that are both most consistent with and most homogeneous with respect to the central meanings ( interpretations) of the factors.
The 15 Factors that came about on the basis of similarity in meaning can be regarded as a two-way account. On the one hand, they tell us what semantic 'junctions' are encoded in the language system. These semantic junctions represent the meaning potential for encoding in language the 'moves' in discourse situations. On the other hand, these semantic junctions represent a set of options available to the social agent, belonging to the repertoire of human beings. They indicate what agents can do in discourse-situations. Since the verbs were selected on the basis of the two characteristics "social" ( two or more persons presupposed) and "intentionality", the set of options can be called the "dramatist potential" of social agents. It should be noted that the list does not contain verbs denoting 'internal' acts such as hope, desire, think,feel, know, doubt. It is particularly the "social" criterion that causes an important difference between the present structure and other lexicons of verbs or structures ( e.g. Kiesler 1 983; e.g. Ballmer & Brennenstuhl 1 98 1 ). The list reflects fairly well both in number and in substance the array of categories which make up the so-called "interpersonal circle" ( Kiesler 1983; cf. Benjamin 1 974), notwithstanding the fact that this 'circle' reflects a personological orientation. The circle represents a two-dimensional space reflecting the joint action of two basic interpersonal dimensions, namely, Control (Dominance versus Submission) and Affiliation ( Love versus Hate). On the circular continuum 16 categories are located at polar ends represent ing semantic opposites. For Dominance and Submission there are synonyms or near-synonyms in the factors III (Authority) and XIII (Docility), respecti vely. For Love and Hate there are synonyms or near-synonyms in the factors V and VII ( Love and Protection) and in the factors I and X ( Verbal and Physical Aggression). The other factors can be considered blends of the two dimensions. Sociability (II), Emotional Demand ( IV}, Argumentation ( VI), and Edification ( XI) can be located between Dominance and Love; Obscurity ( VIII), Aloofness ( XII), and Restoration ( IX) between Hate and Submission;
Downloaded from jos.oxfordjournals.org by guest on January 1, 2011
Discussion
232
VERBAL A GGRESSION
I call a p. nam�s. iflnllt, curs� 2 blow a p. up, snap at a p., abus� 3 sw�ar at, storm and rag�. scold
II
SOCIABJUTY
I grut, �nurtain, agru 2 hav� a chat, g�t into touch wuh, join 3 call on, associate with, mak� acquaintanc�
Ill
A UTHORITY
impau. summon, d�mand 2 send for, prns a p., dicta/� 3 ordain, hold r�sponsibl� for, influ�nce
IV
EMOTIONA L DEMAND
I whine, lament, cha11er 2 brag, whim�r. put on airs 3 complain, sulk, hav� one's way
v
L O VE
I fondle, embrac�. ogle 2 car�ss, coddl�. seduce 3 hug, kiss, make lov� to
VI
ARGUMENTA TION
argu�. pos�. m�ntion 2 matntain, b�ar witnns to, propagatt 3 emphasiu, object, confirm
Vll
PROTECTION
I look afttr, prottct, support 2 comfort, tak� pity on, put heart into a p. 3 assist, help, r�=re
Vlll
OBSC URITY
I obscure, distort, �mbezzl� 2 secr�l�. devise, trifl� 3 disguise, spirit away, kup back
IX
RESTORA TION
I r�stor�. r�dum, pay 2 buy off, fork out, �xcuu 3 do p�nance, ransom, ind�mnify
X
PHYSICAL A G GRESSION
torture, molest, hit 2 kidnap, torment, overpower 3 injure, auack, defy
XI
EDIFICA TION
tnstruct, giv� lessons, advise 2 sketch, tnform, show 3 coach, illustratt, explatn
X/I
ALO OFNESS
I s�c/ude o.s., be non-commiued, escape 2 abandon, dissociate from, keep aloof 3 kup to o.s., stand mutt, pipe down
X/II
DOCILITY
I acquit o.s. wdl, put up with, capitulatt 2 climb down, control o.s., take as one's due 3 comply with, swallow, give a lurch
Downloaded from jos.oxfordjournals.org by guest on January 1, 2011
I
233 XIV
COMPETITION
I com�ll!. ddi�ra/1!, stand up to 2 play a trump, outbid, bet 3 chall�ng�. disput�. bargain
XV
A PPRAISAL
I �xtol, pralu, back a p. 2 comm�nd, applaud. chur 3 congratulat�. crack up, pay homag� to
� ) Translation by the author
Tab/� 4: Factor Labels and Rating Scales for Social Acts•
Downloaded from jos.oxfordjournals.org by guest on January 1, 2011
Competition ( XIV) between Hate and Dominance, and Appraisal ( XV) bet ween Love and Submission. Interpersonal behaviour can be described by different kinds of terms, at different levels of abstraction, and from different theoretical perspectives. In pursuing a circumplex structure, Wiggins (1979, 1980), for example, restrict ed himself to the interpersonal trait domain, thereby choosing for a rather abstract level of categories within the field of person-cognition ( cf. Kiesler 1983). Bochner et a!. (1977) derived a circumplexically ordered taxonomy of interpersonal behaviour, on the basis of self- and other-ratings on Lorr & McNair's (1965) Interpersonal Behavior Inventory. This inventory consists of statements that have a more concrete reference in the situation, such as "making startling remarks that attract attention", or "showing no irritation or anger even when justified". Statements such as these closely resemble the results of Buss & Craik's ( 1 983) act-nomination procedure. The emphasis in the present research on verbs represents an intermediate level of abstraction between the two just mentioned. Subsets of verbs, e.g., those comprised by the factors I and X may on the one hand represent the trait-level of description ( e.g., aggression), and may on the other hand refer to certain contexts. Verbs carry presuppositional aspects that refer to their use or their contextual relevance ( cf. Ballmer & Brennenstuhl 1981). Fillmore {1969) for example, paid attention to the presuppositional aspects of verbs of judging such as accuse, blame, and criticize. He attempted to describe the "role structure" of these verbs. The list obtained can be used to make explicit the purport ( illocutionary force) of the interpersonal activities, or the 'moves' in discourse-situations. In rating tasks judges have to indicate for each group of three verbs, to what extent the verbs together, in their common meaning, are applicable for describing the moves in question. The scores on the three individual scales of a factor should be summed in order to get more reliable measures. The empirical reason for using groups of verbs instead of single verbs to define the individual rating-scales was drawn from Kilkowsky (1976) and Hofstee et a!. (1981).These authors provided indications that, for personality judgments, ratings on groups of adjectives were more consistent than ratings on single adjectives. A theoretical argument is that the use of groups of terms prevents
234
Dept. for Personality PJychology and the Psychology ofPerJonallty UniverJityofGroningen GroteMarkt J/132, 9712 H V Groningen, The Netherlands
REFERENCES
1 980: An analysis of communicative action. I n M. Brenner (ed.) 1 980: The analysis of social situations. I n M . Brenner (ed.). Austin, J.L., 1 962: How to Do ThingJ With Words. Oxford University Press, Oxford. Ballmer, T., 1 983: Semantic structures of texts a n d discourses. Journal of Semantics 2:221 -252. Ballmer, T., 1 984: The position of argumentation i n the framework of a text linguistics, speech act theory, and lexicology. Journal of Pragmatics 8: 9-29. Ballmer, T. and Brennenstuhl, W., 1 98 1 : Speech A ct ClaJsification: A Study in the Lexical analysis of English Activity Verbs. Springer, Berlin. Benjamin, L.S., 1974: Structural analysis of social behavior. Psychological Review 8 1 : 392-425. Bochner, A.P., Kaminski, E.P., and Fitzpatrick, M . A ., 1 977: The conceptual domain of inter pe rsonal behavior: A factor-analytic study. Human Communication Research 3:291 -302. Bolinger, D., 1 980: Language - the Laaded Weapon: The Use and Abuse of Language Today. A l l wood, J ., A rgyle, M . ,
Longman, London.
(ed.), 1980: The Structure of A ction. Blackwell . Oxford. 1 945: A Grammar of Motives. Prentice-Hall. Englewood-Oiffs, N.J. Buss, D.M. and Craik, K . H ., 1 983: The act frequency approach to personality. Psychological Review 90: 105- 1 26. Buss, D. M . and Craik, K . H ., 1984: Acts, dispositions, and personality. I n B.A. Maher & W . B . Maher (eds.), ProgresJ in Experimental PerJonality Research, vol. 1 3. Academic Press, New
Brenner, M. Burke, K.,
York.
De Raad, B., 1 985: Person-talk in Everyday Life: Pragmatics of Utterancu about Personality. UnpubliShed doctoral dissenation, Groningen.
Downloaded from jos.oxfordjournals.org by guest on January 1, 2011
judges from communicating stereotype meanings that are associated with single terms (cf. Bolinger 1980). The judges need not be qualified on their meta-linguistic awareness, since the verbs in question are selected for ordina ry language users. Moreover, experience shows that ordinary users have no problem with the meaning of the verbs. The most typical use for which the method of data-collecting was develo ped is in the area of act-sequences, involving an agent and an addressee.The participants can describe their actions according to the instruction above. If the sequences are restricted to two-tum interactions the agents-addressee contiguities can, for example, be studied by performing canonical correlation analysis (e.g. De Raad 1986). In cases of two or more turns the sequences can, for example, be analysed by studying the transition probabilities of one move to another, representing so-called Markov chains (e.g. Argyle 1980). Another possible use of the method can be in the personological area. Subjects can be asked to judge the frequency of actions of their own or of others in terms of the verbs. Examples are the fields of dispositional assessment (e.g. Buss & Craik 1 984) and interpersonal assessment (e.g. Wiggins 1980).
235 De Raad, B . , 1986: Pragmatics of utterances about personality: Everyday life scenarios.
Deventer. Hofstee, W . K . B . , Brokken, F.B. and Land, H . , 1 98 1 : Constructie van een standaard-persoon lijkheids-eigenschappen-hjst (SPEL). N�derlands Tijdschnft voor de Psychologi� 34: 443-452. Huttenlocher, J . , Smiley, P. and Charney, R . , 1 983: Emergence of action categories in the child: Evidence from verb meanings. Psychological Revi� 90: 72-93. Jessor, R . , 1 98 1 : The perceived environment in psychological theory and research. I n D. Magnusson (ed.). Jones, E. E . and Davis, K . E . , 1 965: From acts to dispositioas: The attribution process in person perception. In L . Berkowitz (ed.), Advances in Experim�ntal Sr;cial Psychology, vol. 2. Academic Press, New York. Kiesler, D. J. , 1 983: The 1 982 mterpersonal circle: A taxonomy for complementarity in human transactions. Psychological R�vt�w 90: 1 85-2 1 4. Kilkowsky, J . M . , 1 976: An �mpirica/ inv�stigation of the consist�ncy ofself-ratings Unpublished doctoral dissertation, University of Minnesota. Koenen, M.J. and Endepols, J., 1 966: V�rk/arend handwoordmboek der n�derlantbe taal. Wolters, Groningen. Kouwer, B.J . , 1973: Exisuntie/� psycho/ogi�: Gronds/agen van h�t psycho/ogisch� g�sprek. Boom, Meppel. Lehrer, A., 1 974: Semantic Fi�lds and Lextca/ Structur�. North-Holland, Amsterdam. Lorr, M. and McNair, D . M . , 1965: Expansion of the interpersonal behavior Circle. Journal of
Personality and Social Psychology 2: 823-830. Magnusson, D. (ed.), 198 1 : Toward a Psychology of Situations: An Jnuractionist P�rspective. N . J. Erlbaum, H illsdale. McHugh, P., 1 968: Defining the Situation: Th� Orgamzation of Meaning in Social Interaction. Bobbs-Merrill, New York. Messinger, H . , Sampson, H . and Towne, R . D . , 1 962: Life as theater: Some notes on the dramaturgical approach to social reality. Sociom�try 25: 98- 1 1 0. Miller, G . A . , 1 969: A psychological method to investigate verbal concepts. Journal of Math�
matical Psychology 6: 1 69- 1 9 1 . Nowakowska, M . , 1 98 1 : Structure o f situation and action: Some remarks o n formal theory of action. In D . Magnusson (ed.). Osgood, C.E., 1970: I nterpersonal verbs and interpersonal behavior. I n J . L. Cowan (ed.),
Studi�s in Thought and Limguage. University of Arizona Press . Arizona.
Pearce, W . P . , 1 976: The coordinated management of meaning: A rules-based theory of inter personal communication. In G . R . Miller (ed.), Explorations in Interpersonal Communication. Sage, Beverly Hills, Calif. Price, R. H. and Bouffard, D.L., 1974: Behavioral appropriateness and situational constraint as di mensions of social behavior. Journal of Personality and Social Psychology 30: 5 1 9-586.
Downloaded from jos.oxfordjournals.org by guest on January 1, 2011
In A. Angleitner, G . L . M . Van Heck, and A. Fumham (ed.s.), European Personality Psycho logy: R�cent Th�or�ttca/ and Empirical D�v�/opm�n/5 II. Swets and Zeitlinger, Lisse. Edmondson, W., 1 98 1 : Spok�n Discourse: A Mod�/ for Analysis. Longman, London. Fillmore, C.J., 1 969: Verbs of judg1ng: An exercise in semantic description. Pap�rs in Linguistics 1 : 9 1 - 1 1 7. Fiske, D . W . , 1 978: Straugt�s for Personality R�uarch: Th� Observation versus Jnurpretation of B�havior Jossey-Bass, San Francisco. Forgas, J . P . , 1 976: The perception of social episodes: Categorical and dimensional represen tations in two different social milieus. Journal of Personality and Social Psychology 34: 1 99-209. Goffman, E . , 1 976: Replies and responses. I.Anguag� tn Soci�ty 5: 257-3 1 3. Grimshaw, A . D . , 1 982: Comprehensive discourse analysis: An instance of p rofessional peer interaction. I.Anguag� in Society I I : 1 5-47. Harre, R. and Secord, P.F., 1 972: Th� Explanation of Social B�havior. Blackwell, Oxford. Helwig, P., l958: Dramaturgied�s m�nschlichen Lebens. Klett, Stuttgart. Hofstee, W . K . B . , 1 974: Psycho/ogisch� uitsprak�n over puson�n. Van Loghum Slaterus,
236 Rogers, L.E. and Farace, R . V. 1 975: Analysis of relational communication i n dyads: New measurement procedures. Human Communication R�s�arch I : 222-239. Rommetveit, R., 1 98 1 : On meanings of situation and social control of such meaning in human communication. In D . Magnusson (ed.). Scovel, T. S., 1 97 1 : A look-see at some verbs of perception. Languag� uarning 2 1 : 7 5-84. Searle, J.R., 1 975: A taxonomy of illocutionary acts. I n K. Gunderson (ed.), Languag�. Mind and
Knowl�dg�. University of Minnesota Press, Minneapolis. SjOberg, L., 1 981: Life situations and episodes as a basis for situational influence on action. In D. Magnusson (ed.). Stech, E . L . , 1 979: A grammar of conversation with a quantitative emp1rical test. Human
Communication R�s�arch 5: 1 58- 1 70. Tal my, L . , 1 975: Semantics and syntax of motion. In J.P. Kimball (ed.), Syntax and Semantics, Vol 4. Seminar Press, New York. Ten Berge, J . M .F., 1 977: Optimizing Factorial Invarianc�. Unpublished doctoral dissertation, University of Groningen. Watzlawick, P., Beavin, J . H . and Jackson, D .D . 1967: Pragmatics of Human Communication. Norton , New York. Wiggins, J.S., 1 979: A psychological taxonomy of trait-descriptive terms: The i nterpersonal domai n. Journal of Personality and Soctal Psychology 37: 395-4 1 2. Wiggins, J. S., 1 980: Circumplex models of interpersonal behavior. In L. Wheeler (ed.), Review of
�rsonality and social psychology Vol I. Sage, Beverly Hills, Calif.
Downloaded from jos.oxfordjournals.org by guest on January 1, 2011
Verschueren, J., 1 98 1 : Problems of lexical semantics. Lingua 53: 3 1 7-35 1 .
Jour/Ul/ of St!mantics
4: 237-246
COMMON SENSE INFERENCE IN A FOCUS-GUIDED ANAPHOR RESOLVER DAVID M . CARTER
ABSTRACT
and local focussing as heavily as possible; it does not rely on the presence of large amounts of world or domain knowledge, which are notoriously hard to process accurately. The ideas reported are implemented in a program called SPAR (Shallow Processing Anaphor Resolver), which resolves a naphoric ambiguities in simple English stories and generates senten ce-by-sentence paraphrases that show what interpretations have been selected. To resolve anaphors, SPAR combines and develops several existmg techniques, most notably Sidner's theory of local focussing and Wilks' 'preference semantics' theory of semantics and common sense inference. Consideration of the need to resolve several anaphors in the same sentence results in Sidner's framework being modified and extended to allow focus-based processing to interact more flexibly with processing based on other types of knowledge. Wilks' treatment of common sense inference
is extended to incorporate a Wider range of types of mference without jeopardizing its
uniformity and simplicity. In the absence o f large quantities of world knowledge, successful anaphor resolution is seen to depend on the coordination of predictions made by system components exploiting various knowledge sources. Such coordination normally allows ana phors to be resolved correctly even when no single source of predictions
is sufficient on its own.
I . A SHALLOW PROCESSING HYPOTHESIS
A computer system that resolves anaphors in non-specialised texts such as stories must be able to take account of many different kinds of knowledge. Some of the most important are: syntactic knowledge about agreement and configurational constraints on coreference; semantic knowledge, especially of word meanings; local and global focussing; and common sense knowledge of the world. Of these, the last is by far the hardest to represent and apply, because of its complexity, diversity and quantity. Present-day systems capa ble of text understanding can therefore normally only cope with a rather limited range of texts: limited in domain, and/or in the type of reasoning required. However, it seems plausible that a considerately written text ( i.e. one that is perspicuous and unambiguous, in accordance with Grice's maxim of man ner) will normally be constructed in such a way that constraints on interpreta tion derived from different kinds of knowledge will tend not to conflict but
Downloaded from jos.oxfordjournals.org by guest on January 1, 2011
This paper describes an investigation of the feasibility of resolving anaphors in natural language texts by means of a 'shallow processing' approach which exploits knowledge of syntax, semantics
238
A story processing system which exploits linguistic knowledge, particularly knowledge about local focussing, as heavily as possible, and has access only to limited quantities of world knowledge, which it invokes only when absolutely necessary, can usually choose an ap propriate antecedent for an anaph or even in cases where the common sense inference mechanism by itself cannot do so. This paper deals with the problems of coordinating the application of syntax , word sense semantics, and especially local focussing and world knowledge, in a shallow processing anaphor resolver. Such coordination turns out to be crucial to accurate performance. The approach adopted is a synthesis of Sidner's (1979) work on local focus and Wilks' ( l975a, l 975b, 1977) prefer ence semantics, both of which will now be reviewed.
2. BACKGROUND
In Sidner's theory, pronoun interpretation ( PI) rules are applied to each pronoun in a sentence independently of the others. The rules suggest candi date antecedents, normally one at a time, according to the contents of a set of focus registers which have been set during processing of earlier sentences. If a candidate agrees syntactically with the pronoun, it is temporarily bound to it, and inference is invoked using semantic and common sense knowledge. In Sidner's words, "the complexity of the inferencing is constrained to asking for confirmation of the sentence predication, thereby eliminating combina torial search for antecedents and non-terminating inferencing." ( Sidner 1979:75). This confirmation consists of "proving that the sentence with the pronoun replaced by the cospecified noun phrase is consistent with other knowledge." (p. 150). In other words, as long as the suggested candidate does not give rise to a contradiction, it is accepted. Sometimes, however, the rules suggest two or more candidates at once.
Downloaded from jos.oxfordjournals.org by guest on January 1, 2011
rather to confirm one another and work together to guide the reader towards correct interpretations. Moreover, because language has some degree of redundancy, the same information may often be contained in more than one constraint or prediction. For example, when resolving a pronoun, we might expect that the most focussed possible referent will usually also be the one that common sense reasoning suggests is most plausible. If so, it may be possible to resolve most instances of at least some kinds of ambiguity by adopting a 'shallow processing' approach in which linguistic knowledge ( syntactic, semantic and focussing) is exploited as fully as possi ble, while world knowledge is present only in very limited quantities and is exploited, by common sense inference ( CS I), only when linguistic knowledge alone is insufficient to resolve an ambiguity. Specifically, this research sets out to test the following Shallow Processing Hypothesis:
239
3. COORDINATED A N A PHOR RESOLUTION: THE SPAR SYSTEM
My S PA R ( Shallow Processing Anaphor Resolver) system ( Carter 1984, 1 985) is a Sidnerian anaphor resolver which uses Wilksian semantics and CSI to do the work of Sidner's 'normal mode' and 'special mode' inference respectively. It is applied to initial sentence interpretations produced by Boguraev's (1979) English analyser, a system that employs syntactic knowl edge encoded as an augmented transition network and a modified form of Wilksian semantics. This analyser resolves most word sense and structural ambiguities (producing alternative interpretations if any are unresolvable in
Downloaded from jos.oxfordjournals.org by guest on January 1, 2011
When this happens, inference is invoked in a different mode to decide which candidate ( if any) is most plausible. Although Sidner viewed this mode of inference as 'special', it is in practice required quite often. Full noun phrase anaphors are also resolved using focus-based rules, which however only require 'normal mode' (single-candidate) inference and not 'special mode' . Sidner did not detail the functioning ofeither mode of the inference mechanism, although she viewed doing so as "an important aspect of ongoing research in artificial intelligence" (p. 1 50), and gave a number of examples of what the mechanism should be capable of. Sidner's theory is incomplete in two important respects. Firstly, the PI rules do not specify how or when candidates from the same sentence as the pronoun should be considered. Secondly, no attention is given to any possi ble interaction between the applications of the rules to different anaphors in a sentence. Both problems need to be tackled in a practical system. In Wilks' preference semantics theory, focussing was assigned a very minor role. Instead, common sense inference ( CSI) rules were invoked whenever a pronoun could not be resolved by semantic pattern matching alone. Each rule expressed a ( usually causal) relationship in terms of two linked semantic patterns consisting of general semantic primitives. They were used to construct (hopefully consistent) chains connecting the representation of a sentence containing problem pronouns with those of sentences containing candidate antecedents. A completed chain would bind one or more pro nouns. Thus in Wilks' framework, semantic matching was used to assess the plausibility of single candidates and therefore roughly corresponded to Sid ner's proposed 'normal mode' of inference. CSI, on the other hand, was used to select one candidate out of many for each problem pronoun, as in Sidner's 'special mode' of inference. Wilks' semantics and CSI mechanisms, although more modest in their aims than those Sidner assumed, were specified in . enough detail to allow implementation to be begun. They thus provide a way to get an implementation for non-specialised texts of Sidner's theory off the ground; and they are sufficiently simple, flexible and general to be a suitable basis for a shallow processing system.
240
Downloaded from jos.oxfordjournals.org by guest on January 1, 2011
the absenceofextrasententialcontext)butdoesnottackleanaphoricambiguity. The interpretations produced arerelativelyshallowcase-organiseddependency structures with word senses defined by Wilksian 'formulas', trees constructed from a set of a hundred of so semantic primitives. SPA R resolves the anaphors in the dependency structures, in the process resolving any remaining word sense and structural ambiguity. When each sentence has been fully processed, a paraphrase is produced which makes clear what ambiguity resolution decisions have been made. The system acts on the dependency structure( s) for a sentence in a story as follows. In preparation for PI rule application, the semantic formula for each word sense in a dependency structure is matched with the surrounding parts of the structure. This provides a measure of 'semantic density' ( strong agreement is a ground for preferring that reading over others) and also constraints the semantic ranges of pronouns. For example, the formula for drink specifies a liquid object, so in the sentence He drank it, it would be restricted to match only a liquid antecedent. Next, PI rules are applied to each pronoun in a sentence; other focus-based rules are applied to full noun phrase anaphors. Applications to different anaphors are independent except where one anaphor is a constituent of another. Although the PI and full N P rules themselves closely resemble Sidner's, there are important differences in the way they are applied. These differences are a consequence partly of adopting a shallow processing ap proach and partly of extending Sidner's theory to consider intrasentential candidate antecedents and interactions between different anaphors in the same sentence. In the PI rule application, if there is only a single candidate antecedent suggested by the PI rules, it is first matched with the pronoun, using their Wilksian semantic formulas and any restrictions imposed in the first stage of processing.This matching corresponds roughly to invoking Sidner's 'normal mode' inference, since most contradictions resulting from temporary bind ings take the form of semantic clashes. If the match succeeds, a firm 'predic tion' that the pronoun and candidate corefer is returned by the PI rules and put aside until later. Otherwise the rules go on to suggest further candidates. If, instead of a single candidate, the PI rules suggest several candidates at once, each one is matched semantically with the pronoun. If several survive, CSI is not invoked immediately, as in Sidner's original framework; rather, alternative predictions are returned, to be adjudicated later. CSI is postponed for two reasons. Firstly, as we will see below, it is often possible, by compa ring predictions for different anaphors, to avoid CSI altogether; and second ly, because an inference chain often binds several anaphors at a time. It is therefore both simpler and more efficient to invoke CSI for all unresolved anaphors in a sentence together than for each one as it arises. Consideration of intrasentential candidate antecedents is not achieved in the obvious way by inserting extra PI rules, because the rules are already rather complex and because there seem in any case to be very few reliable
241
(i)
I took my dog to the vet on Friday. He bit him in the hand.
(trivially altered from Sidner ( 1979)), the PI rules aided by the semantic matcher, predict that he could be either the dog or the vet, while him can only be the vet (since hand is defined as part of a person, not of a dog). The MGC constraint prevents he and him coreferring; SPAR therefore reasons that since him must be the vet, he cannot be. Thus by not invoking CSI early on when the PI rules suggested two plausible candidates for he, we have managed to avoid CSI altogether. If configurational constraints detect a clash between two firm predictions (i.e. predictions without alternatives), the PI rules are re-applied to the anaphors involved in order to find further plausible candidates. In practice, this is very seldom necessary. CSI is only invoked if some anaphors remain unresolved even after confi gurational constraints have been applied. As Sidner recommended, the pre dictions of the PI rules for those anaphors are used to constrain both the starting points for inference and the directions in which it proceeds. We are only interested in inference chains which bind anaphors to candidates predicted by the PI rules; other chains represent a clash between focussing and CSI, and when this occurs, the CSI prediction is rejected. If Sidner's theory is correct, this would be the correct course of action even if the CSI mechanism were a powerful one; the fact that it is not makes the focus prediction even more obviously preferable. If the PI rules, configurational constraints and CSI between them finally resolve all anaphors, no more need be done. However, because the CSI mechanism has only limited amounts of knowledge available, it will often not find all the chains that a more powerful mechanism might construct, and so some anaphors may remain unresolved. In such cases, up to three weaker heuristics are applied which, although there are many counterexamples,
Downloaded from jos.oxfordjournals.org by guest on January 1, 2011
preferences statable in focussing terms either between intrasentential candi dates or between intrasentential and contextual candidates. Instead, before the PI rules are applied to each pronoun, the focus registers are temporarily augmented with intrasentential candidates, ordered approximately as speci fied by Hobbs' (1976) syntactic algorith m 1 ; the PI rules then pick them up j ust as they do contextual candidates. An important consequence of this is that fewer of the PI rules' suggestions consist of a single candidate; it becomes more common for several candidates to be suggested together. However, as stated above, CSI (i.e. 'special mode' inference) is not invoked immediately when this happens. When the PI rules have been applied to each anaphor in the sentence, configurational constraints (specifically, approximations to Reinhart's (1983) c-<:ommand and MGC constraints2) are used to rule out as many inconsistent predictions as possible. Sometimes this removes the need for CSI. For example, when processing the short text (i),
242
4. AN EXAMPLE
As envisaged (and to a small degree implemented) by Wilks, CSI rules expressed weak generalisations about causality. In Wilks (1977) a two-fold 'taxonomy of causation' is proposed: CSI rules should be indexed so that when a problem pronoun occurs in a sentence describing a deliberate action, an explanation (i.e. pronoun-binding chain) is sought by looking for the motives of the actor; otherwise one should look for the causes of the event. For Wilks, a further subdivision of rule types is only justified if it too corresponds to a procedural distinction . This twofold taxonomy of causal knowledge has been sufficient for the texts processed so far by SPAR, but it has become apparent that more is required than the kind of causal rules Wilks considered. Each of the following types of knowledge needs to be treated specially. Firstly, since the semantic formulas for noun senses often describe a typical function or use of an object, it is important to be able to exploit such knowledge when appropriate. (This i s analogous to the use of 'scripts' in some text understanders, although a script is normally far more detailed than a semantic formula). Secondly, the scarcity and shallowness of world knowledge means that there should be some shallow rules to recognise associative rather than straightforwardly causal relationships. Thirdly, 'static' rules, describing states of affairs such as 'themes' (Schank and Abelson 1977) which occur so commonly that they are their own j ustification, are sometimes needed. Inference chains ending with such rules do not need a specific connection to the story context. Finally, S PA R's inference mechanism is extended to reason about an actor's plans and goals by simulating, in a l imited way, the actor's own inference processes. Such sim ulation means that very little specifically goal-based knowledge is required; all the knowledge used for event-level inference can also be used for goal-level inference. All thest· factors come into play when the first three sentences of the story given below are processed. This processing clearly shows the importance of
Downloaded from jos.oxfordjournals.org by guest on January 1, 2011
usually seem to arrive at correct interpretations when they apply; when they do not apply, other, still weaker preferences associated with Sidner's PI rules are used. The first and most important heuristic is that repetitions should be prefer red. If a pronoun and one of its remaining candidates have the same role in two semantically similar events in the story, that candidate is preferred. For example, if a character has performed a particular action, it is arguably more l i kely that he will perform such an action again than that somebody else will. The second and third heuristics prefer, respectively, interpretations in which the discourse focus (as defined by Sidner's rules) remains unchanged and those in which pronouns are c-commanded by phrases that corefer with them. The usefulness of these heuristics will be seen in the following example. (For full details of all three heuristics, see Carter 1985).
243 coordinating focus and common sense knowledge. CSI is less useful for the rest of the story; instead, the repetition heuristic is sufficient to resolve difficult anaphors. (1) (2) (3) (4) (5)
In sentence ( I ), CSI is needed to choose between the candidates John and Bill for both pronouns. It makes the correct choices, using static rules containing the information that people tend to make promises about their own deliber ate actions rather than other people's, and that people tend to want their own possessions to work. In (2), his is resolved without CSI , since the PI rules and semantic matching are able to recommend John firmly. CSI is invoked to choose between John , the actor focus, and Bill , the potential actor focus, a s candidates for he. It makes use of the formula for garage, which says that a garage is a place where people mend things, and decides that he is taking it to the garage so that someone can mend it. Since both John and Bill can be expected to want the car to work (John because he has made a promise, and Bill because he owns the car), alternative chains are returned which predict both referents for he (and bind it to the car). The third weak heuristic (preferring pronouns to be c-commanded by coreferring phrases) correctly selects John as the referent. Thus whereas in ( 1) , CSI provided the discriminatory power, here in (2) it was either not required to do so, or it failed toprovideaconclusiveresult. In the latter case, the choice ofantecedent wasdeterminedonlybya post-CSI weak heuristic. In (3), the PI rules find that both occurrences of he and the him are ambiguous between John and his friend. However, the first he is definitely resolved as John because configurational constraints forbid the alternative. They also forbid the second he and him to corefer, but since both pronouns are still ambiguous, no alternatives can be ruled out. CSI is therefore in voked. Rather than reasoning, as a more powerful inferencer might, that him is John because John is likely to want tools to mend the car, it simply binds him to the first he using a shallower, more general CSI rule stating that people are more likely to want to possess things themselves than to want other people to possess them. Since we already know that the first he is John, him is now resolved, and so the second he can be identified with the friend by applying configurational constraints again . Thus in this sentence, focussing, CSI and syntax were all vital to anaphor resolution.
Downloaded from jos.oxfordjournals.org by guest on January 1, 2011
( 6) ( 7) ( 8) (9)
John promised Bill that he would mend his car. He took it to his friend's garage . He tried to persuade his friend that he should lend him some tools. His friend said that he was not allowed to lend tools. John asked his friend to suggest someone from whom he could borrow tools. His friend did not answer. Fulfilling his promises was important to John. He was angry. He left.
244
In the remainder of the story, the CSI mechanism fails to find any chains on the occasions when it is called. However, the repetition heuristic enables the correct decisions to be made. It recognises that since sentence (3) mentions the friend lending John tools, he in (4) is the friend and he in (5) is John ( given the obvious semantic relationship between borrowing and lending). It also realises that his in (7) is John and not the friend, since John was mentioned as making a promise in ( 1 ) and the friend is not associated with any promises. The paraphrase for the story ( which reads rather unnaturally but should make the reference resolutions clear) is:
(4')
(5 ')
(6' ) ( 7') (8 ') (9')
John promised Bill that John would repair Bill's car. ToJohn'sfriend's garage, John conveyed the car. John attempted to convince John's friend that that friend should loan John some repair implements. John's friend said that that friend was not allowed to loan any repair implements. John requested John's friend to recommend someone who John could borrow some repair implements from. John's friend did not answer. Discharging John's promises was urgent to John. John was angry. John departed.
5. SOME I NTERDISC I PLINARY COMPA RISONS
Work on anaphora in theoretical linguistics has been mainly concerned not with identifying the process by which anaphors are resolved, but with charac terising the kinds of relationships that can hold between anaphors and antecedents and stating structural ( typically surface syntactic) restrictions on intrasentential coreference. As explained above, structural restrictions are only one of the sources of k nowledge used by S PA R. However, it is clear from the linguistic research that the range of possible anaphoric relationships is far broader than that catered for by S PA R and, indeed, most other anaphor resolving programs. Quantification, genericity and intensional contexts give rise to complications which SPA R treats in only the most rudimentary way. Some improvements in this area, perhaps using the treatment of Kamp ( 1 98 1 ), would clearly be needed if SPA R were to be used in any practical application; but dealing with such phenomena would not necessarily be the most immediately effective way to test further the Shallow Processing Hypothesis. In contrast to theoretical linguists, psycholinguists have paid considerable attention to the processes of anaphor resolution. However, it should be stressed th.1t SPA R is not intended to provide a model of the way people resolve anaphors; its goal is the engineering one of effective anaphor resolu-
Downloaded from jos.oxfordjournals.org by guest on January 1, 2011
(1 ' ) (2') ( 3')
245
6. RESU LTS AND CONCLUSIONS
In all , SPAR has processed over sixty stories on a variety of topics. These stories are of two categories. The first category consists of forty texts , of two or three sentences each, that were specially written (or selected) to test or illustrate various facets of SPAR's processing. All of the 65 pronouns in these texts are correctly resolved. The second category consists of over twenty stories written by people with little or no knowledge of SPAR's workings; many were originally written for other language-processing systems. These stories are on average nine sen tences long; the longest has twenty-three sentences. Of the 242 pronouns in them , 226 (93%) are resolved correctly; this figure could be increased to 232 (96%) by the implementation of an error recovery procedure described in Carter ( 1 985). The contribution of CSI to this performance is indicated by the fact that inference chains were used to determine the resolution of only 29 ( 1 2%) of the 242 pronouns, in all cases correctly. A large numbe r of other chains were formed but either only confirmed resolution decisions already made or were rejected as incompatible with the predictions of other components of the system. The simplicity of the texts processed relative to 'real' texts means that only tentative conclusions can be drawn from these results. However, it appears that the effective use of the limited quantities of knowledge available to a shallow processing system depends critically on the way in which compo-
Downloaded from jos.oxfordjournals.org by guest on January 1, 2011
tion by computer. Indeed, the shallow processing approach is motivated by the observation that computers, unlike people, have great difficulty i n using world knowledge. Even so, some interesting parallels exist between SPAR's operation and some theories of human anaphor resolution. For example, Sanford and Garrod ( 1 98 1 : 1 33- 145) use various experimental results to argue for a model in which topicalised and recently-mentioned entities are held, with different degrees of prominence, in working memory. When a pronoun is encountered, working memory is searched first for a referent, with candidates being assessed on the basis of syntactic and lexical knowledge. I nference using more general world knowledge need only take place if a referent cannot be unambi guously identified. This strategy is roughly similar to that of SPA R , in which common sense inference is only invoked if focussing, syntax and word-sense semantics are inadequate (although in fact some of what Sanford and Garrod class as lexical knowledge would be represented in SPAR by CSI rules). Interestingly enough , Sanford and Garrod further argue (pp . 196- 1 98) that discourses requiring frequent general inference for anaphor resolution are often "inconsiderate" and to some degree abnormal. It is on a "considera teness" assumption like this that the Shallow Processing Hypothesis is based.
246 nents using different kinds of knowledge are coordinated. Linguistic knowl edge (including syntax, semantics and focus) should be fully exploited to resolve as many ambiguities as possible, while world knowledge, which is i nevitably incomplete and less reliable, should be applied as late as possible, and then only when linguistic knowledge alone cannot resolve all ambigui ties. The performance of SPAR provides some evidence that a shallow processing approach may enable us to resolve the anaphors in non-speciali sed texts surprisingly accurately. U111 versity of Cambridgt! Compult!r Laboratory Corn Exchangt! Strut Cambridgt! CB2 3QG, U.K.
I.
Hobbs' algorithm resolved definite pronouns by selecting candidate antecedents from
surface syntactic trees for sentences in a text. The first candidate to agree syntactically with the pronoun was accepted. Although SPAR does not have direct access to surface syntactic repre sentations, the semantic structures provided by Boguraev's analyser are sufficiently shallow to allow the necessary surface syntactic relations to be inferred quite accurately. 2. These constraints, which apply to surface syntactic structures, may be stated in a slightly simplified form as follows. ( I ) A node N P I c-commands NP2 if the lowest branching node dominating N P I also dominates N P2 (A branching node IS a node with more than one daughter). I f N P I c-commands NP2, and N P2 is not a pronoun, then NPI and NP2 do not corefer. (2) If N P I is a non-reflexive pronoun, and NP2 is dominated by N P I 's minimal governing calt!gory (the lowest S or N P node dommating N P I) , then N P I and NP2 do not corefer.
REFERENCES Boguraev, B.K., 1 979: Automatic Rt!solution of Linguistic Ambiguuies, TR- 1 1 , University of Cambridge Computer Laboratory. Carter, D . M . , 1 984: An Approach to General Machine Translation based on Preference Semantics and Local Focussing. Procudings of ECAI-84. Carter, D . M . , 1985: A Shallow Proassing Approach to Anaphor Rt!so/ution, TR-88, University of Cambridge Computer Laboratory. Hobbs, J . R . , 1976: Pronoun &solution, Research Report 76- 1 , Department of Computer Science, City College, City University of New York. Kamp, H . , 1 98 1 : A Theory of Truth and Semantic Representation. In: Groenendijk and Janssen (eds), Formal Mt!thods in the Study of Languagt!. Groningen: Amsterdam. pp. 1 -4 1 . Reinhart, T. , 1 983: Anaphora and Semantic Jnurprt!tation, Croom Helm, London - Canberra. Sanford, A.J. and Garrod, S.C., 1 98 1 : Understanding Writlt!n Language, Wiley, Chichester. Schank, R.C. and Abelson, R.P., 1 977: Scripts, Plans, Goals and Understanding, Lawrence Earlbaum Associates, Hillsdale, N.J. Sidner, C . L . , 1979: Towards a Computational Theory of Dejinilt! Anaphora Comprehension in English Discourse, TR-537, M . I .T. Artificial Intelligence Laboratory. Wilks, Y . A . , 1975a: An Intelligent Analyzer and Understander of English, Communications of the A . C. M. 1 8:264-274. Wtlks, Y.A., 1 975b: A Preferential, Pattern-Seeking Semantics For Natural Language Inference, A rtificial Intelligence 6: 53-74. Wilks, Y.A., 1 977: What sort of Taxonomy of Causation do Understanding? Cognitive Science I :235-264.
we
need for Natural Language
Downloaded from jos.oxfordjournals.org by guest on January 1, 2011
NOTES
Journal of &mantics 4: 247-255 GENERICS BART GEURTS
ABSTRACT
There is a fairly general tendency to analyse generic statements as referring fact that, thus far, none of these analyses have proven to be satisfactory, there always remains the problem of explaining what a kind is. I propose , instead, to do away with kinds altogether, and to regard generics as ex pressing stereotypical assumptions. Although, at the moment , I have just the broad outlines of a theory to offer, this approach seems to fit the available data quite well. Moreover, it obviates the need to assume that generic descriptions are referential expressions .
1.
Intuitively speaking, generics are statements about kinds. Thus,
( 1)-(3)
are said to be generic, or at least to have a generic reading:
(1)
Beavers b uild darns.
(2)
The dodo is extinct.
(3)
A rabbit
has four legs.
The subject terms of these sentences seem to refer to kinds. It remains to be seen , however, whether this pre-theoretical notion can be upheld . Oearly, this depends crucially on the explication of what a kind is, and this paper largely deals with the latter problem. It is proposed that we replace the common-sense notion of kinds by the theoretical notion of stereotypes. This proposal will be seen to provide a plausible explanation for the ways in which generic sentences function. At the same time, it allows us to consistently deny that generic descriptions refer, and with that to dodge awkward ontological issues . In the following, we shall only consider non- complex sentences with a
Downloaded from jos.oxfordjournals.org by guest on January 1, 2011
to the entities that are commonly identified as "kinds" . Apart from the
248
(4)
• A dodo is extinct.
On the other hand, if a generic sentence explicitly refers to a member of the kind in question, thus emphasizing the distributive nature of the predicate, then an indefmite but not a defmite singular NP can be subject. Cf. (5)
a. A rabbit, such as Nancy, has four legs. b .•The rabbit, such as Nancy, has four legs.
All in all, the dichotomy between distnbutive and non-distributive generics
seems to be justifiable. This is important , since the theory we are about to sketch applies solely to the first class . The reason for this restriction will soon become apparent. At least some occurrences of generic NPs are readily paraphrasable as quantified expressions. For instance , it seems that the sense of
(6)
a.
A
sparrow is a bird .
can be expressed equally well by saying that all sparrows are birds. There fore, it is tempting to assume that, in general , generics can be treated as
Downloaded from jos.oxfordjournals.org by guest on January 1, 2011
subject term that is either a definite or indefinite singular NP or a bare plural. Although these forms have attracted by far the most attention in the literature, the class of generics is much wider, and comprises also , e.g., sentences figuring mass terms or abstract nouns. So this restriction does not reflect a principled decision on our part. To begin with, we want to distinguish two kinds of generics, which we shall term distributive and non-distributive, respectively. If somebody utters ( 1) we may normally conclude that the kind denoted by the subject term is presupposed to have at least one member to which the predicate applies. There is a limited number of predicates, however, that block such distributive inferences, the reason being simply that they reject subject terms referring to individuals (cf. Carlson 1978 : 47-49 ; Quine 1960 : 1 22-123). Predicates belonging to this group are be extinct, be common and be rare . Accordingly, whereas we shall classify ( 1 ) and (3) as dis tributive generics, (2) represents the non-distnbutive class. The distinction between distributive and non-distnbutive generics is supported by further observation. For instance, it appears that, in this particular respect, definite and indefmite singular NPs are more or less in complementary distnbution. On the one hand , if the predicate is a non distributive one, the subject NP can be defmite but not indefmite . (2) is acceptable, (4) is not :
249 implicitly quantified sentences. But they cannot. To see why this is so , consider the following sentences : (6)
b . A sparrow has feathers. c. A sparrow lays eggs. d. A sparrow does not lay eggs.
2. Let us consider the following, fictitious, situation. The giraffe is in im minent danger of dying out. To be more exact, three specimens are still alive, and only one of them is fitted with the usual four legs . The other two each have , say, three legs . This state of affairs has been caused by recent and rather dramatic developments. One could imagine that the African countries have unanimously declared war on the US , that there upon the Americans have decided to spray the African continent with Agent Orange , etc. What matters is that the situation is relatively new. Now given this situation it is stated that
(7)
A giraffe has four legs.
True or false? If we had to give an answer, we would probably judge (7) true, although it is likely that we would hedge . Compare this with any plausible quantified 'translation' of (7). Unless we settle for a paraphrase
Downloaded from jos.oxfordjournals.org by guest on January 1, 2011
The quantifier presumed to underly (6a) was seemingly equivalent to the universal quantifier. But (6b) could still be true in a world in which not all sparrows happened to be feathered. As far as this sentence is concerned, the generic quantifier rather seems to have the force of "most " . But again, this interpretation is too strong for (6c) . Here , the hidden quantifier can at best be taken to mean something like "roughly half of all". Now we could postulate that the quantifier underlying a generic sentence must always have this weak interpretation. But this would lead to incorrect predictions for (6d) and counter-intuitive interpretations for (6a) and (6b) . To show that an approach along these lines cannot be but ill advised we would have to push this discussion somewhat further .1 But we are not trying to disprove any of the previously proposed theories. This brief survey was intended, rather, to reveal an important characteristic o f generic sentences, namely that , when w e try t o translate generics as quantified sentences, all sorts of quantifiers are bound to tum up . Generics seem to have what Carlson ( 1 978 : 43) has called "fluctuating truth con ditions".
250
Downloaded from jos.oxfordjournals.org by guest on January 1, 2011
with some, which would be decidedly too weak, every quantified counter part of (7) obviously would be false. Now the circumstances take a tum for the better. The last four-legged giraffe passes away, but the remaining three-legged giraffes beget children. their children beget children, etc. - and they all have three legs. Would (7) still be true then? It is far from likely that we would stick to it. This example seems to suggest that verifying a generic statement is not altogether a straightforward matter - at least compared with the verification o f quantified sentences. It is as though generics possessed a certain kind of inertia: their truth value is not just a function of the state of affairs at a given moment. There are reasons to believe , however, that even this conclusion grants too much. I suspect that the traditional notion of verification (or falsification) is unsuited to deal with generics. This sus picion is strengthened by various observations, which indicate that generics typically occur in situations where the question of their truth or correct ness is unlikely to become an issue . It is sometimes suggested (e.g. by Dahl 1 975) that generics divide into two separate classes, descriptive and normative. Although I consider this distinction to be of minor importance from a semantical point of view, the observation on which it is based is highly relevant. It is indeed a conspicuous trait of generics that they regularly crop up in sermons, political addresses and debates, advertisements - in short, discourses involving norms and values. Moreover - this is so obvious that one tends to overlook it - generics o ccur quite frequently in learning situations . A related b ut, at least as far as I know, hitherto unnoticed point is that the substance of a generic is seldom new. Usually, it is taken for granted that tigers are striped, that a giraffe has four legs, and so on. Extensive portions of the background knowledge ordinarily left implicit in linguistic interactions are expressed most naturally by generic sentences. I do not know how often and under what circumstances speakers actually exploit this possibility. This issue would have to be decided statistically. But it seems obvious to me that if speakers want to state explicitly the bridging assumptions (Haviland and Clark 1974) linking their utterances, they often can resort to generics. Summarizing these observations, we may conclude that generics differ from ordinary factual statements in that they are typically used to express general assumptions , laws, maxims, principles, prejudices or rules of thumb . It is by no means a characteristic of generics that they, so to speak, con front the world directly. The idea that there are sentences which are not immediately affected by experience is not new. As is well known, Quine (1960) defends the position that the theoretical sentences which form part of our scientific theories only can be tested because they are linked with observation
25 1 sentences. In this respect, generic sentences and theoretical sentences are alike (of course, theoretical sentences often are generic). It is not the generic sentences themselves that are put to the empirical test (if they are testable at all) , but the observation sentences to which they are related . Taking these considerations into account, I propose a general ex planation that, b riefly put, amounts to the following. A generic sentence is, from the speaker's point of view , the expression of a stereotypical ass umption about a kind , and , with respect to the hearer, a request to revise, update, or simply retrieve certain stereotypical assumptions.
3.
by now : the production and comprehension of natural language is best explained in terms of cognitive structures which are constructed and ex tended in the interplay between speaker and hearer. Following Bosch
( I 983),
we will call these structures context models. 2 The interpretation
of (sequences of) n atural-language utterances requires various kinds of b ackground knowledge (including grammatical knowled-g e). We take it that this b ackground knowledge is organized, at least for a good part, in stereotypes. The central component of a stereotype associated with a given property p is a set of properties that are assigned by default to any individual to which p is said to apply. This is not to say that stereotypical assumptions can never tum out to be wrong. But as soon as we have sufficient reason for doing so, we will simply withdraw them. In other words, stereotypical assumptions are upheld as long as they are not counter indicated . Interpretative processes exploit b ackground knowledge in order to construct a context model. As it is not clear from the outset what kind of role generics could play in these process es , let us tentatively suppose that they are not directly involved in the building of a context model, but operate , instead, in the domain of b ackground knowledge . (Eventually , we shall have to modify this picture somewhat.) We assume that speakers' information about a given kind K is embodied in a stereotype, which for our purposes can be identified with a property set
P. 3
Now, we define the
generic interpretation of an utterance as a function that takes the stereo
P associated with a given kind K, and maps it onto p' . I n illustration, we take u p sentence ( I) again :
type
(I)
Beavers build d arns .
It has been noted b y , e .g . ,
D ahl ( I975 : 1 10-1 1 I)
that this sentence i s in
Downloaded from jos.oxfordjournals.org by guest on January 1, 2011
The b asic assumption of the theory I want to outline should be familiar
252 fact ambiguous. In the fust and most natural interpretation, it ascribes to beavers the property of typically building dams. But (1) can also be read as claiming that it is a typical property of dams that they have been built by beavers. Either interpretation can be forced by accenting either beavers or dams, or by clefting and/or passivization (cf. Seuren 1985 : 295 -304). In French, furthermore, this difference is brought out by a choice between the defmite and the indefmite article (Van Langendonck 1980: 1 090). Cf. (8)
a. Les castors construisent des digues . b . Le s digues sont construites par des castors .
(9)
Dutchmen are good sailors.4
( 1 0)
The blackbird lays two or three eggs .
According to the theory that I have indicated, (9) would result in the addition of the property "is a good sailor" to the stereotype associated with "Dutchman", and the interpretation of ( 1 0) would append the property "lays two or three eggs" to the stereotype associated with "blackbird". But it could hardly be called reasonable to assume by default that the Dutchman whom I met just now will turn out to be a good sailor, or that this blackbird here will promptly produce two or three eggs . The first answer that comes to hand is that we give the properties to be added to the stereotypes associated with "Dutchman" and ''blackbird", respectively, the form of a conditional. In the first case , this would yield something like "if x is a sailor, then x is a good sailor". In the second case, we could have , say, "if x is female , then x lays two or three eggs". However, if we opt for this solution, we bear the burden of having to specify the antecedents of these conditionals in a general and systematic way. The problem is nicely exposed by Ziff ( l972 : 1 28-129) : One says 'A cheetah can outrun a man'. We put it to a test : a man and a cheetah are turned loose in a field. The man lopes away while the cheetah sits lazily in the sun. But 'can' is not 'will ' . The cheetah
can
Downloaded from jos.oxfordjournals.org by guest on January 1, 2011
Within the framework proposed, the treatment of a sentence like ( 1 ) is quite straightforward. If (1) is understood as a claim about a certain kind of animal, it results in the addition of the property "builds dams" to the stereotype asso ciated with "beaver". On the other hand , if it is about dams, then the property "built by beavers" is added to the stereotype associated with "dam". On the face of it, this account would appear unable to deal with sen tences like
253 even if it did not. We try again. This time we force the cheetah to run by beating it ; but the man easily outruns the cheetah : the cat en cumbered with an awkward two-hundred-pound weight moves slug gishly. An encumbered cheetah is a cheetah . And
I said a cheetah
can
out
run a man. Should I have said not that a cheetah can but, more cautiously, that some cheetahs
can outrun a
man, namely unencumbered cheetahs?
But what about an unencumbered cheetah whose feet have been bound since
birth?
If we make provision for foot-bound unencumbered
cheetahs shall we not also have to provide a place for three-legged cheetahs, drugged cheetahs, cheetahs forced to run after being force
than can outrun a man'?
fed and so on and on? Would it be better never mind which - cheetahs
simply to say 'Some -
cifying more and more conditions, without being able to decide where we should stop. Consequently, I believe that we must look for another, more satisfactory solution.
In the literature on generics, it has become obligatory to remark that it takes more than one counter-example to refute a generic statement (see , e .g., Carlson
1 97 8 : 34-35 ; Dahl
1 97 5 : 1 02 - 1 03). This is undoubtedly
correct. However, this trite remark is often taken as fmplying that, in general, a generic statement is contradicted, though not necessarily refuted, by any member of the kind denoted by the subject term to which the predicate does not apply . And this is patently absurd . It is just not true that every Dutch non-sailor contradicts (9), or that any male blackbird would contradict ( 1 0) . Only Dutch sailors who are incompetent can force us to raise doubts about (9), and the correctness o f ( 1 0) becomes question able as soon as we discover b lackbirds that lay five, fifteen or fifty eggs at a time . These cases are, for some reason, relevant, and only of them we can rightly claim that they constitute counter-examples to the sentences in question.
P consists of properties which P applies, unless there
In the foregoing, it was said that a stereotype that can be assigned to any individual i to
are indications that i does not have these properties. But , as Ziff has reminded us, "can" is not "will" . Quite evidently, it would be useless, and even troublesome , to always make all the inferences one can make . It only makes sense to resort to default assumptions if one has to - if, that is, they are relevant one way or the other. The stereotypical expectation that Dutchmen are good sailors is not activated until we have ascertained that the Dutchman before us is a sailor, and something similar goes for the stereotypical assumption expressed by ( 1 0) . Therefore , we must reconsider our notion of what a stereotypical property is . We shall say that a stereo typical property is a property p that is assigned to an individual i as soon as p is relevant, and if there are no indications that i is not-p.5
revision should take care of sentences like (9) and ( 1 0).
This
Downloaded from jos.oxfordjournals.org by guest on January 1, 2011
Ziff's argument shows , in a word , that we could go on ad inrmitum spe
254 By now, it must have become clear that we propose to defer the problem concerning generic sentences. In fact , we contend that an elucidation of notions like "stereotype" and "relevance" would at the same time explain the seemingly erratic behaviour of generics. Nonetheless, our proposal still
is
worthwile , since a semantic theory will have to explicate these notions
anyway.
4. We have decided to consider generics as functions that operate on stereo presupposed in the interpretation of utterances . Yet presumably, this b ackground knowledge is relatively stable . Not on any account is it likely to change with every generic statement that is being processed . So the picture we have sketched
is
too simple . We cannot assume that generics
operate directly on b ackground knowledge . On the other hand , 'generic information' should not be allowed to mix with the contents of the context model under construction - at least as long as we take this notion in a narrow sen se . For we have argued that generics differ from other declaratives in matters of verification and truth. The most convenient solution for this problem probably would be to open a separate compartment (either within our outside our context models) to house the information that generics express. Perhaps, this compartment could also accomodate other kinds of information . It possible ,
for instance, that
certain
context-specific inferences
is
could
fmd room here. But right now I do not intend to take a stand on this issue. I merely want to note that the general approach to generics ad vocated here requires a reconsideration of our overall model of natural language interpretation. Now we can turn back to our initial question. Do generic NPs refer? It
is
widely held that they do. First of all , this notion is inherent in our
common-sense
understanding of generic descriptions. Carlson ( 19 78)
declares this common-sense notion sound semantic theory by stating that generic NPs are on a par with proper names : as these refer to individuals , generic NPs refer to kinds . However, considerations of economy urge us to look askance at this position , which, for all its intuitive appeal, has rather drastic implications for the ontology of natural language . But , fortunately , these implications can be avoided. If we may assume that our proposal is
in principle correct, then there is no need to regard generic descriptions as referential expressions. We analysed generics (roughly) as asserting that "if you should encounter such and such individual, then you may reason ably assume so and so" (some generics actually exhibit something very
Downloaded from jos.oxfordjournals.org by guest on January 1, 2011
types, and stereotypes form part of the background knowledge ordinarily
255 much like this structure
in
their surface form) . The success o f such rules
of thumb admittedly depends on regularities
in
the world. But it is not
necessary to postulate entities for them to refer to . University of Nijmegen Department of Philosophy and Psychological Laboratory P. O. Box 9108
6500 HK Nijmegen.
NOTES See Carlson
Bosch
3.
(1978 : 3 3-56) for a
much more elaborate critique.
The over-all view outlined in this paragraph is worked out i n more detail by
( 1 9 8 3 : 65-78).
And , presumably , only
for o u r present purposes. I t is highly unlikely that,
event ually , stereotypes will turn out to bear more than a superficial resemblance to unstructured objects like sets.
4.
Carlson
5.
I feel that, at least for the time being, we must be quite literal-minded about the
( 1 9 7 8 : 44)
quotes this example fro m the Port-Royal logic.
processes involved in storing and retrieving predicates. The predicate that is stored after processing a generic sentence is identical to the predicate that is retrieved to make inferences about a given individual. It follows from this that, as it stands, my account does not apply to non-distributive generics. Loosening up the envisaged mechanism would of course be possible, but at present I see no sufficient (i.e. in dependent) reasons for doing so.
REFERENCES Bosch, P.,
1 983 : Agreement and Anaphora. Academic Press, London-New York. 1 9 7 8 : Reference to Kinds in English. Indiana University Linguistics
Carlson , G.N., Club. Dahl,
6., 1975 :
On generics. In : E . L . Keenan (ed.), Formal Semantics of NatW'al
Language. Cambridge University Press, Cambridge. Haviland, S.E. and Clark, H . H . ,
1 97 4 :
What's new? Acquiring new information
process in comprehension. JoUJ'nal of Verbal L earning and Verbal Behavior
as
13.
a
1 960 : Word and Object. MIT Press, Cambridge , Mass. 1 98 5 : Discourse Semantics. Basil Blackwell, Oxford. Van Langendonck, W., 1980: On a narrow conception of genericness. Linguistics 1 8 . Ziff, P . , 1 97 2 : Understanding Understanding. Cornell University Press, Ithaca-London. Quine, W.V.,
Seuren, P.A.M.,
Downloaded from jos.oxfordjournals.org by guest on January 1, 2011
1. 2.
Journal of Semantics 4: 257-263
D ISCUSSION VALENCES LTD. vs VALEN CES ASSOCIATED Comments on Heringer's Association Experiment as a Basis for Valence Theory•
H. ECKERT
ABSTRACT
that associations between verbs and question words can serve as a basis for valence theory. The results of his association experiment, however, do not permit us to mfe r dependency relations, to distinguish clearly between optional and obligatory elements, to specify quantitative valence, or to distmgmsh between elements that are grammatically and semanucally implied by the verb as opposed to merely contextual elements. I should also like to argue that the questions that are satd
to impose themselves upon the speakers need not necessa rily do so because of the semantic
power of the verb, and that the values for certain question words are partly influenced by the test method, where each question word affects the values of the subsequent ones. I feel that while the association experiment Yields supporting evidence for valences i n a number of cases it cannot claim to functiOn as a basis for valence theory.
In his article on "The Verb and its Semantic Power: Association as a Basis for Valence Theory" Heringer ( 1 985) gives a detailed account of the carefully compiled and computed data of his association experiment that is concerned "with syntagmatic associative relations between verbs and questions." (83) (Page references without author and year refer to Heringer ( 1 985)). I should like to deal here with two aspects of his article: fi rstly with his account of existing valence theories, whose explanatory power he considers very limited, and secondly, with the validity of his claim that association can function as a basis for valence theory. Heringer concentrates on one particular aspect of valence, i.e. the distinc tion between complements and supplements, which - in an earlier article - he described as "the basic issue of valence theory, because it concerns its explanatory power in general" (Heringer 1 984:35)1• I would agree with Heringer that "valence theory has not succeeded in justifying syntactically the difference between complements and supplements . . . " (80), but I feel that the syntactically based distinction should be rejected for the right reasons. Summing up this theory, Heringer claims, "Complements are those NPs that are obligatory whereas supplements are those that are facultative. This was the leading idea" (80). None of the leading linguists in this field that I happen • cf. H.J. Heringer: "The Verb and its Semantic Power", this Journal, vol. 4.2: 79-99 ( 1 985).
Downloaded from jos.oxfordjournals.org by guest on January 1, 2011
Heringer claims that the value of existing theories of valence is limited as they have failed to gtve a clear account of the crucial distmction between complements and supplements. H e maintains
258
"My argument is that the entire way of looking at this was misguided because it does not bear reference to what is really the basis of this distinction. What is amazing about this historical discussion is, after all, that one constantly presupposes a distinction which one wishes to establish in the first place. Or put differently, we obviously have a concept of the distincuon without having a definition."
This critique aims at the status of definitions and test methods. There are, basically, two different procedures open to us. We either have a preconceived idea of what we are trying to define, and our tests, methods or operations yield results that may or may not fit this definition. According to Heringer, this is the procedure adopted by previous valence models, which he claims h ave all failed. Alternatively one can devise an experiment without any preconceived ideas and postulate that if the method yields two types of results then one type will be complements, the other supplements by definition. As I am not absolutely certain which of the two procedures Heringer would claim for his experiment, I shall consider both possibilities. The baci�ground for the association experiment is the assumption that (a) the "semantic power of the verb, in the sense intended here [my emphasis],
Downloaded from jos.oxfordjournals.org by guest on January 1, 2011
to be familiar with (Helbig, Schumacher, Engel, Rail, Emons, Fink, Tarvai nen, Allerton, to mention but a few, and, of course, Heringer himself) have subscribed to this view. What caused the endless discussions in this area, however, was precisely the difficulty of distinguishing between optional complements and optional supplements.2 Heringer also gives an account of more recent developments, and here again I wish to point out a few terminological short cuts. With regard to the 'new', semantically based theory he says, "complements are arguments of the verb, and supplements are additional predications . . . " (80). Arguments may be represented by complements, but they are not identical to them and there is no one-to-one correspondence. We are, in fact, dealing with three different levels of analysis here: arguments, their corresponding cases, and comple ments. Some arguments cannot be represented as complements in the actual sentence: e.g. the stolen things with the German verb besteh/en ('steaf), or food with G. dinieren('dine'), a relation often referred to as 'incorporation'. Other arguments may be represented twice in one and the same syntactic structure: Er unterrichtete uns /davon/, /daB er nicht kame/ (literally: 'He informed us /of it/ /that he would not come/.') And some surface structure elements do not have a corresponding argument.3 I do not wish to overemphasize this point because Heringer himself is fully aware of all these problems, as is clear from his other writings, but it is important to remind the reader of the fact that the association experiment must be measured against a far more refined model of valence than the one outlined in Heringer ( 1 985). In an earlier account of previous attempts to distinguish between comple ments and supplements Heringer ( 1 984:35) says,
259
Downloaded from jos.oxfordjournals.org by guest on January 1, 2011
consists in its power of requiring or evoking certain nominal complements. The communicative background for this is that certain questions impose themselves upon speakers, given a particular verb." (82), and that (b) "The degree of probability with which these questions arise reflects the degree of necessity of the complement." (83) If this is supposed to be a test criterion then the elements referred to here are complements by definition. This would be a new definition of complements as a result of the association experiment, and any discrepancy between this class of complement and previously estab lished ones might be of terminological or historical interest, or one could raise the question of material adequacy (i.e. was it wise to choose these traditional labels for these data?), but the discrepancies would not invalidate the distinction drawn here. I suspect, however, that Heringer finds himself in a similar position to those linguists that he criticizes in his article ( 1 984) quoted above. He, too, seems to have a preconceived idea of what complements are (see pp. 81 and 82), and seems to "presuppose a distinction which one wishes to establish in the first place." He then sets out to check this by an association experiment. But the results leave him with a nagging feeling that they have not always quite captured what they were meant to. There are numerous examples in his article where he does not say, 'Contrary to expectations X is a complement because of its position in our diagrams', but where he argues that X is a complement in spite of its position in the diagram. In other words, every time things 'go wrong' they are explained in terms of other theories and methods. The main result of Heringer's association experi ment is that "The degree of probability with which these questions arise reflects the degree of necessity of the complement." Let us consider a number of aspects that might have had an influence on the subjects' associations and thus the choice of question words that were evoked, and let us examine if these aspects are all relevant to verb valence. If valence specifications of the verb are to serve any purpose we must distinguish between those elements that are grammatically, semantically or logically dependent on the verb, and those elements that are not and that simply occur in a given utterance. Lachen ('laugh'), e.g., is not inherently interpersonal, i.e. the lexical frame of arguments and cases does not logically imply a second person. It is possible, however, to think of a context in which a second person is involved, and this is obviously what a number of test persons did when they asked mit wem? ('who with'). If we compare Er /acht gerne ('He likes to laugh') and Er verhandelt gerne ('He likes to negotiate'), we see that mit seiner Freundin ('with his girl friend') may be added to both utterances. But there is a fundamental difference between these two additions as far as the two German verbal lexemes are concerned, because in the latter case (as opposed to the former) mit jemandem ('with somebody') is always understood even if it is not mentioned. In some theories this criterion is seen as the basis for the distinction between complement and supplement•. Heringer would argue, I suppose, that the different status of mit wem in
260
Downloaded from jos.oxfordjournals.org by guest on January 1, 2011
each case would be reflected by the varying distances between the preposi tional phrases and the corresponding verb. He claims that the distinction between complement and supplement is gradual, "since semantic boundaries are not sharp but fuzzy boundaries" (8 1 ). If valence and case grammar can put forward cogent reasons for a watertight distinction between - at least some - lexemes that are inherently interpersonal and others that are not, then I do not see any advantage in blurring that distinction. I fee l that compared with other theories5 the association method is lacking in explanatory power in this respect. The values of some question words might be due to cultural aspects rather than to verb-specific and lexical meanings. With gehorchen ('obey') the question word warum? ('why') has the second highest frequency and good values for rank and latency. Let us assume, for the sake of the argument, that the same experiment had been conducted with one hundred recruits of the Prussian army in 1 9 1 0. It is by no means certain that warum? would have scored equally well. Different results that are due to a change of attitudes are compatible with the assumption that the inherent meaning of the lexeme gehorchen has not changed. In his analysis of kaufen ('buy') and verkaufen ('self) Heringer says, "Verbs of the buying-scene are characterized by four semantic roles: buyer, seller, goods and payment" (90). His diagrams (Figures I and 2, p. 89) for these two verbs do not contain this information. For kaufen, the first question word that may - but need not - be associated with payment is womit? ('what with') and is in tenth position, preceded by wem? ('who for'), which in the case of kaufen, as opposed to verkaufen, does not correspond to any of the four semantic roles postulated for the buying-scene. Heringer, too, seems to feel that if there is such a core of semantic roles they should evoke question words and these should be close to the centre of the diagram, i.e. close to the verb. He himself mentions "The embarrassing fact that the question von wem? ['who from'] does not appear in the evaluated range . . . " (90) but explains this in terms of suppletion, because he argues that the seller (with kaufen) "ap pears in the disguised form of wo ['where']". This explanation does not seem satisfactory: in some of the other cases that Heringer refers to as suppletion and partial suppletion question words corresponding to identical semantic roles do occur within the evaluated range. (Cf. schulden: was/wieviel ('owe: what/how much') and lachen: woriiber/warum ('laugh: what about/why') p.90n.) Moreover, wo? does not necessarily and exclusively ask for the seller: "Wo hast du das gekauft?" - "In England." - "Und von wem?" ("Where did you buy that?" - " In England." - " And who from?") Or consider: "Ich hab' das von Erwin gekauft." - "Ja, aber wo und wann denn?" ("I bought this from Erwin." - "Yes, but where and when?") The different positions of wo? with kaufen and verkaufen need a lot of explanation. There is one interesting comment on this in connection with the distinction between complements and supplements. "The reason [for the more central position of wo? in the diagram of kaufen] is that wo? is not purely
26 1
Downloaded from jos.oxfordjournals.org by guest on January 1, 2011
a question for a supplement, but may also ask for the seller" (90). This implies that according to Heringer wo? corresponds to a supplement if it simply asks about the place of a transaction, but corresponds to a complement if it asks about the seller, who represents one of the four semantic roles of the verb. The prototype of a complement seems to be an element that corresponds to a semantic role in the case frame of a lexeme. If this is accepted these four semantic roles should form the inner circles in the verb satellite diagram, which, however, is not true of kaufen and verkaufen. Suppletion, co-occurrence and co-reference are related concepts in this model and pose a number of problems. " . . . it was necessary to pool the responses in niches. E.g. the responses warum?, wesha/b?, weswegen?, aus we/chem Grund? wegen was? (why, for what reason, because of what?) were pooled under the lead word warum? The idea is that these questions aim at the same kind of answer" (84). For bezah/en ('pay') the question words wem? and an wen? (referring to the receiver of the payment) are put in the same group, not because of the test or as a test result but because of prior and independent semantic considerations. If the very question was whether both wem? and an wen? represented complements or supplements, the association experiment with the application of pooling would not solve this problem. A different problem arises with some other kinds of suppletion. "The questions was? and wievien in connection with schu/den aim at the same slot, so do the questions woriiber? and warum? in connection with /achen' (90). Here Heringer does not resort to pooling "since wievien in the case of schulden is a potential specification of was?, whereas woriiber? and warum? in connection with lachen are partially competing questions" (90). There are good reasons for and against pooling here and our choice is important because it affects the proximity of the question words or the niche to the verb, which is the main test result. Whatever we decide on, however, we will have neglected some aspects, which will distort the satellite diagram. This is a problem that this method has imposed on itself and that does not arise with other methods. The experiment is partly based on the assumption "that certain questions impose themselves upon the speakers, given a particular verb" (82). This may be true, but not necessarily for the same reasons for all speakers and not necessarily because of the semantic power of the verb. Most German speakers have been exposed to at least nine years of traditional grammar, the introduc tion of new vocabulary and so on. If you are used to patterns such as jemandem etwas geben ('give somebody something') and jemandem die Tiir offnen ('open the door for somebody'), the corresponding questions may impose themselves on you , but this does not prove that offnen is an interper sonal verb. Heringer himself tries to explain the unexpected, relatively low frequency of wer? ('who') in terms of the infinitive's function, which "is just to mention the verb without any subject" (86). Assuming that the infinitival citation form is the reason for the value of wer?, it implies that the results of this
2 62
kaufen: warum? 4. 1 ; wie? 1 5. 3 verkaufen: warum? 8.5; wie? 7 . 3 I should like to argue that the values differ not because the semantic power that attracts warum? to verkaufen is less, but because other question words for verkaufen push warum? further to the periphery. This is a consequence of the test method, which necessitates a linear sequence of question words, where each word affects the values of the subsequent ones. In other theories q uestion words asking about the reason of the transaction could be assigned the same status for both lexemes. Another aspect that has played an important role in valence theory and in the distinction between complement and supplement, but which is not re vealed in the satellite diagrams, is the distinction between optional and oblig atory elements. Even if (because of the often smooth transitions) it is not possible to deduce the line of demarcation between these two types of e lements, it would be reassuring if the obligatory ones preceded the optional ones. In my own dialect I would accept Was schulde ich? (' What do I owe [you]?'), but not • Wem schu/de ich? (' Who do I owe [something]?'). In Herin ger's diagram the order is wer?, wem?, was? A linear sequence in two-dimen sional diagrams cannot represent the complexities of quantitative valence: some verbs require the filling of a particular slot, but do not require a specified element (or in this case: question word). With some other verbs the occurrence of element A is dependent on element B . Since the subjects of this e xperiment were not free to choose new contexts or start new sentences for every question word, dependency rules cannot be traced, nor can they be represented in this type of diagram.
Downloaded from jos.oxfordjournals.org by guest on January 1, 2011
expe riment are influenced by factors other than the semantic power of the verb. Once again I feel that if the experiment is to serve as the basis for valence theory, the values Of wer? should provide us with data and criteria by which other theories might be judged. Instead, other theories are used to 'explain a way' values that do not seem to fit. Moreover, the explanations have a ring of circularity. If we say, 'In view of the importance of the subject the value of wer? seems relatively low', we are told that this is because of the infinitive's function. If we ask, 'Since the infinitive serves as the basis for the associa tions, should the values of wer? not be even lower than they actually are?', the a nswer is that this is, indeed, remarkable and "surely an indication for the paramount importance of the subject" (86). The verb satellite diagrams specify the sequence and distances of question words. "They are based on the ratio .1.0..¥- and display the semantic distance of the questions to the verb" (86). Let us consider the ratio of wie? and warum? for kaufen and verkaufen. As these two verbs are related to the same scene (in Fillmore's sense) there does not seem to be any reason for assuming that the semantic distance for the 'how' a nd 'why' of these transactions should be entirely different for each of the two verbs in question, and yet the values are as follows:
263
It is, of course, possible to argue that it is not necessarily the semantic power of the verb that makes the elements in its context obligatory or optional, and that the association experiment was not designed in the first place to reveal quantitative valence and dependency relations. The admission that the notion of 'necessity of the complement' does not include obligato riness would, however, weaken the claim that this method can serve as the basis for valence models. Plidagogische Hocluclrule Flen.Wurg Milrwiker Str. 77 D-2390 flensburg Fed. Rep. Germany
I The quotes from Heringer ( 1 984) are my translations of his anicle, which appeared in German. 2 Cf. , e.g. M. Rail, U . Engel, D. Rail ( 1 977 :40): " I . Aile auf verbale Subklassen beschrl!nkten Satzglieder sind Erg1!nzungen. 2. Aile obligatonschen Satzglieder sind Ergl!nzungen." See also my review of this book (Ecken 1 979). 3 These are JUSt three of a number of possibilities. For a more detailed account see Ecken ( 1 985). 4 For illustrations and funher references see Ecken ( 1 977) and ( 1 979) .. 5 Cf., for instance, Fink's ( 1 977) semantically-based valence concept.
REFERENCES Ecken, H., 1 977: Lexical Field Analym. ISKO, Hamburg. Ecken, H., 1 979: Review of Rail, M., Engel, U , Rail, D. ( 1 977). IRA L. XVI I/3, 271-273. Ecken, H., 1985: Practical and theoretical aspects of lexical entries for verbs of communication. In: G. Hoppenbrouwers, P. Seuren, A. Weijters (eds.), Meaning and the Lexicon. Foris, Dordrecht. Fink, S . R . , 1 977: Aspects of a pedagogical grammar based on case grammar and valence theory. Niemeyer, TObingen. Heringer, H . J . , 1 984: Neues von der Verbszene. In: Sprache der Gegenwart (Jahrbuclr 1983 des Instltuts fiir deutsclre Sprache), Band LX; 34-64. Schwann, Dusseldorf. Heringer, H .J., 1 985: The Verb and its Semantic Power: Association as a Basis for Valence Theory. Journal of Semantics 4: 79-99. Rail, M . , Engel, U., Rail, D., 1 977: Dependenz- Verb- Grammatik fiJr Deutsch als Fremdsprache. Julius Groos Verlag, Heidelberg.
Downloaded from jos.oxfordjournals.org by guest on January 1, 2011
NOTES
Journal of s�mantics 4: 265-270
BOOK REVIEW Jaakko Hintikka (in collaboration with Jack Kulas),
The Game ofLanguage. Studies in Game-Theoretical Semantics and its Applications. (Synthese Lan guage Library 22). Reidel, Dordrecht, 1 983. Pp. xii+324, Dfl. 1 1 5.00 (cloth). JAN LEMMENS
.
.
.
Downloaded from jos.oxfordjournals.org by guest on January 1, 2011
The book is a collection of ten essays. Two of them, viz. Seman tical games, subgames, and functional interpretations and Any problems - no problems, appear here for the first time in print. The first essay, Game-theoretical semantics: insights and prospects, origi nally appeared in the Notre Dame Journal ofFormal Logic 23 ( 1 982), but has been revised and expanded for this book. It gives a short introduction to Game-Theoretical Semantics (GTS) and a perfunctory survey of what Hin tikka deems GTS has accomplished and promises to accomplish. Hintikka suggests a complementation of GTS by the 'functional-lexical approach to syntax'. The second essay, Semantical games and transcendental arguments, origi nally appeared in E.M. Barth and J.L. Martens ( 1 982). Here Hintikka tries to relate the philosophical justification of GTS to Kant's philosophy of mathe matics, space, and time. According to Hintikka, Kant contended that our ways of reasoning about existence must be grounded in those human activi ties through which we come to know the existence of individuals. What are these activities? (Kant located them - mistakenly, according to Hintikka - in sense perception.) They are the activities of seeking and finding. These rule-governed activities (verification processes) are captured, according to Hintikka, in his semantical games and these are therefore the correct ap proach to logical inference and logical semantics. Certain related views go astray at this point according to Hintikka. He contends that Dummett and Prawitz interpret verification - their favoured key concept for a theory of meaning - as a method of formal proof. The truth of a sentence can, Hintikka says, seldom be established by means of formal proof procedures. Here Hintikka is mistaken that according to Dummett (formal) proof is the central notion in a theory of meaning for a language; cf. following quotation: "Such a (intuitionistic) theory of meaning generalizes readily to the non-mathemati cal case. Proof is the sole means which exists in mathematics for establishing a statement as true: the required general notion is, therefore, that of verifica tion. On this account, an understanding of a statement consists in a capacity to recognize whatever is counted as verifying it " (Dummett 1976. My italics). There is no suggestion that the notion of verification must be restricted to formal proof procedures, on the contrary. Admittedly, the notion of recognizing what establishes a statement as true is not without its problems (as Wittgenstein has taught us), but it is hard to see that GTS can eventually
266
too. The last part of this chapter is devoted to an application of subgames to discourse semantics. This paper seems a little unbalanced and some improve ment on the presentation of the central issue would be welcome. The fourth essay, Any problems - no problems, by Hintikka and Kulas, deals with the problem of negation in GTS. Several game rules for GTS i nvolve an operator neg which gives for any given sentence S its contradictory (semantical negation) neg(S). First a set of syntactical rules is formulated, but in some cases they do not give the right result. Next, a kind of argument is presented to the effect that there can be no set of complete rules for negation
Downloaded from jos.oxfordjournals.org by guest on January 1, 2011
circumvent these problems. GTS stops short at (interpreted) 'atomic' sen tences. What is the relation between the truth or falsity of these atomic sentences and our knowledge of them? The idea that we recognize the truth or falsity of these sentences forces itself upon us. The third essay, Seman tical games, subgames, and functional interpreta tions, can be read as a sequel to Conditions, generic quantifiers, and other applications of subgames (in Saarinen 1979). It is divided into two parts. The fi rst links GTS for first-order logic to some developments in logic and in the foundations of mathematics via the notion of a subgame and functional i nterpretations. The idea of a subgame was introduced in the latter essay. The functional interpretation of a first-order sentence is a translation of it into a h igher-order language using functions and functionals. A game rule for GTS can then be expressed by a translation rule. For example, if -S means that any attempt to verify S will provide a way of verifying -S, this can be expressed by the translation: ( 3:cD)(s)[-S'(s.�s))], assuming that the translation of S is: ( 3:s)(�)[S'(s.�)), where s in S'(s.�) is Myselfs strategy of verifying S and � is Nature's strategy. Here the functional cD expresses the strategy used by Myself and this strategy depends on Nature's strategy s in an earlier subgame. Note that negation has the effect of role-switching, i.e. Myself assumes Nature's role and Nature Myselfs role. The second part of the paper investigates what new light can be thrown on natural language phenomena by the subgame idea. The GTS treatment of anaphora is considered, but not essentially furthered. There is a promise of a more formal approach to the treatment of a naphora where an anaphoric pronoun is processed by applying a game rule to it (the book promised is presumably H intikka and Kulas, ( 1 985). Hintikka discusses a problem noted by James Higginbotham ( 1 982), who remarks that an anaphoric reading of them in: If everybody else consults two doctors, I'll consult them too, gives two doctors wide scope . This creates a difficulty for GTS, because (i) the consequent is considered, if at all, after the antecedent game is played, and (ii) the antecedent is ambiguous between the reading giving two doctors wide scope and the reading giving everybody else wide scope. Hintikka seems to deny that the former reading exists. His own reading, however, creates a related problem, since the interpretation of everybody else depends on i nformation present only in the consequent, given such examples as If everybody else consults two docters, John will consult them
267
Downloaded from jos.oxfordjournals.org by guest on January 1, 2011
in English . It uses Barwise's proposition that if S is a sentence of a first-order language with partially ordered ('branching') quantifiers (FPO) and if its negation -S i s logically equivalent to an FPO sentence, then S is logically equivalent to some first-order sentence, i .e. S can be reduced to a sentence involving no branching quantifiers (Barwise, 1 979). If there are English sentences which, when translated into FPO, cannot be reduced to a first-order sentence (and there seem to be such sentences}, then their negation cannot be expressed in FPO . Therefore, to the extent that English is required to stay within the bounds of FPO, there is no hope of giving a set of complete rules for negation in English. In order to cope with this problem, the authors introduce the SPS principle which says that semantics parallels syntax (or that semantics and syntax go together). Although the SPS principle plays an important role, the authors do not attempt to find a more precise formulation of it. Stil l , they claim that the principle yields "quite definite predictions". One consequence of the SPS principle is, according to the authors, the following. When a quantifier phrase occurs in subject position in a sentence S, there are two different ways of forming the contradictory of S: (i) negating the quantifier phrase by prefixing a "not" to it (term negation), or (ii) negating the verb of the main clause of S (regular negation). The SPS principle 'predicts' that term negation is acceptable only for quantifier words whose game rules do not have priority over the negation rule in semantical games. The authors note that an application of a game rule might result in neg(X), where X is not acceptable. This presents a problem for GTS, because "speakers are not likely to associate any interpretation to an unac ceptable string" (p. 97). If X = Dick has dated any girl (example of the authors), X is unacceptable according to the any-thesis. But the any-thesis (which says that in an otherwise grammatical context any is acceptable if and only if substitution of every for any results in a sentence not identical in meaning) presupposes the possibility of interpreting an unacceptable sen tence. So it is not clear why the semantic interpretation of neg cannot guide us here and recourse has to be taken to the SPS principle. The SPS principle suggests that, in case X is unacceptable, neg should be expressed by a clause-initial syntactical constituent. Thus, neg(X)=Diqk hasn't dated any girl, for X as above. If X is acceptable, neg(X) functions differently. This has, as the authors note, an important consequence in view of the alleged fact that neither the set of acceptable sentences nor the set of unacceptable sentences is recursively enumerable (Hintikka 1 979): the operation of forming the contra dictory of a given sentence in English is not recursive. A further non-recursi vity result is obtained in connection with sentences of the form "b knows whether X". The fifth paper, Temporal discourse and semantical games, originally ap peared in Linguistics and Philosophy 5 ( 1 982). Game rules (G. past) and (G. past perf) are formulated and discussed. These rules involve moments of time (necessitating a change in the underlying model), chosen by one of the players. Depending on which player picks a time, a 'universal' - or 'existen-
268
the United States is the most important political leader in the world. The seventh paper,"/s", semantical games, and semantical relativity, origi nally appeared in the Journal ofPhilosophical Logic 8 ( 1 979). It denies that is is multiply ambiguous in the sense of the Frege-Russell tradition. Moreover, the Fregean trichotomy is claimed to be demonstrably false, on the assump tion that GTS is the one correct theory. The claim depends on the thesis that substituting a term in a sentence containing is for another term does not affect the meaning of is. What is the criterion of synonymy here? There are no game rule for is, so such an approach is excluded. Hintikka takes his recourse to the notion of "to be a descendent of the same occurrence or': "Our rules for semantical games show that the allegedly different occurrences of 'is' can sometimes be descendents of one and the same occurrence of 'is', with which they are therefore synonymous." (p. 1 78). For example, an application of (G. an(n)) to Jack is a boy results, say, in the sentence: Jack is John Jr. and John Jr. is a boy. The occurrences of is in the latter sentence are descendents of the same occurrence of is, viz. the occurrence of is in the former sentence. It is hard to see that this argument shows that there is no difference in meaning between the three occurrences of is in the above sentences. It seems to me that there is a difference between semantic phenomena and how they are account ed for (say, by some semantic representation). What the above argument would show is that such an explanation does not arise at the level of applica tion of the game rules. On this account it will count against GTS if no explanation of the different meanings of is is forthcoming. Furthermore it contradicts Hintik ka's claim that the Fregean trichotomy is relative to one particular semantic theory.
Downloaded from jos.oxfordjournals.org by guest on January 1, 2011
tial-quantifier' - reading results. Readings are multiplied since no strong ordering principles apply. The interplay between (G . and) and (G. past) seems to explain phenomena which compelled Davidson to introduce an ontology of events. Two weak principles are introduced in order to exclude less plausible readings: (i) the left-right order of verbs involved in a sentence should match the temporal order of the moments chosen, and (ii) the number of independent choices governed by one and the same game rule applied to phrases in one and the same sentence should be kept at a minimum. ((G. past) may be applied at the same time to the main verb of more than one clause in a sentence). Further game rules for temporal adverbs and temporal connec tives are discussed. The sixth paper, Definite descriptions in game-theoretical semantics also by Hintikka and Kulas, covers the same ground as Hintikka and Kulas ( 1 982). A game rule, (G. anaphoric the), is form ulated which suggests an obvious generalization to all anaphoric expressions and seems to yield a formal treatment of Geach's 'donkey sentence ' . A restriction on this rule is formulat ed, the exclusion phenomenon, which forbids the reference of an anaphoric the-phrase to an individual introduced by a word or phrase occurring in the same clause. The formulation of this phenomenon would however seem to need some modification in view of such examples as: The current President of
269
Downloaded from jos.oxfordjournals.org by guest on January 1, 2011
The eighth paper, Semantical games and Aristotelian categories, appeared under the title "Seman tical games, the alleged ambiguity of 'is', and Aristote lian categories." in Synthese 54 ( 1 983), and was revised for publication in this book. The first part of it deals with the interpretation of quantifier phrases in English . The problem is this: How do players know from which (sub)domain to select an individual? It is suggested that the range of the quantifier-expres sions (the subdomain from which the players must select an object) is shown by wh-words, as e.g. in: Some boy who loves soccer is a hooligan. There is, according to Hintikka, a one-to-one correspondence between wh-words and the widest range a quantifier in English can have. In the above example, who indicates that the widest range of some is the class of people. When, however, in the above example, we remove the relative clause, we get the sentence Some boy is a hooligan, and here the players must rely on the meaning of "boy" in order to decide the maximal quantifier range. If such an indicator is missing, as, for instance in Someone is a hooligan, the quantifier word must amplified to provide the missing information about what the domain for the players' choice is. Hintikka comes to the conclusion that natural languages have only many-sorted quantification. The second part of the paper is devoted to a discussion of the multiple ambiguity of is and Aristotelian categories. Hintikka holds that "game-theo retical semantics implies a version of Aristotle's doctrine of categories, incorporating all of the most important - and most puzzling - general theoretical features of it" (p. 2 1 9). The ninth paper, On the any-thesis and the methodology of Linguistics originally appeared in Linguistics and Philosophy 4 ( 1980). A thesis corres ponding to the any-thesis, the ever-thesis, is formulated: "'ever' is unaccept able in an otherwise grammatical context if an exchange of 'ever' for 'always' yield a semantically equivalent sentence" (p. 234). This, as well as resolutions of apparent counterexamples to the any-thesis, is presented as further evi dence for the any-thesis. Next, Chomsky's ( 1 980) objections against Hintik ka's any-thesis and its alleged methodological consequences are discussed. The last essay, Theories of truth and learnable languages, first appeared in S. Kanger and S. Ohman ( 1 98 1 }, and has been expanded for republication in this book. It presents putative counterexamples to the principle of composi tionality and it is argued that the rejection of the principle of compositionali ty does not imply the rejection of the assumption of a paralellism between syntactical and semantical rules. Failure of compositionality would threaten a Davidsonian approach to natural languages but not GTS, which is offered as a viable alternative. The last part of this chapter is devoted to dissuading the reader from trying to save the principle of compositionality by using a higher-order logic (as in Montague semantics). The book contains a good index and the relevant literature is included in a bibliography. Hintikka and his associates have shown that GTS is a promising approach to natural languages. Not only can GTS handle some of the most puzzling
270 aspects of natural languages, but GTS has also set a new paradigm for their investigation . The book is an account of work in progress; new and stim ulat ing idea's are introduced, objections answered and refinements of earlier ideas offered. A wide range of topics is dealt with, logical, philosophical, as well as linguistic ones. One cannot but admire Hintikka's vast learning and creativity. One might wonder, however, what the authors' aim is in publish ing the book, apart from updating and collecting a number of papers already previously published. Unfortunately, most of these papers bear the mark of being stages of work in progress and what is missing is a more systematic and more detailed account of GTS. Such a book would certainly enhance GTS' status and would make a better assessment of GTS' merits possible.
REFERENCES E.M. Barth and J.L. Martens (eds), 1 982: A rgum�ntation: Approach�s to Th�ory Formation, John Benjamins, Amsterdam. J. Barwise , 1 979: On branching quantifiers in English, Journal of Philosophical Logic 8. N . Chomsky, 1980: Rul�s and R�pr�stntations, Columbia Umversity, New York. Michael Dummett, What is a theory of meaning? I I . In: G. Evans and J. McDowell (eds), 1976:
Truth and M�aning. Clarendon Press , Oxford. J. Higginbotham, 1 982: Comments on Hmtikka's paper. In: Notre Dame Journal ofFormal Logic 23.2. J . Hintikka, 1 979: Quantifiers in natural languages: some logical problems. In: Saarinen (ed.) ( 1 979). J. Hmtikka and J. Kulas, 1982: Russell vindicated: towards a general theory of definite descrip tions, Journal of s�mantics I . J . Hintikka and J . Kulas, 1 985: Anaphora and D�finit� Descriprion:r. Two Applicarion:r of Gam� theor�tical Semantics. Reidel, Dordrecht. S . Kanger and S. Ohman (eds), 1 98 1 : Philosophy and Grammar. Reidel, Dordrecht. Esa Saannen (ed) 1 979: Gam�-Th�or�tical S�mantics: E:rsay:r on S�mantics by Hintikka, Carlson, P�acock�. Rantala and Saarin�n. Reidel, Dordrecht.
Downloaded from jos.oxfordjournals.org by guest on January 1, 2011
D�pt. of Phtlosophy Univ�rsity of Nijm�g�n Thomas van A qumostraat 3 6500 HK Nijm�g�n. N�thulands
Journal of Semantics 4: 27 1 -274
BOOK REVIEW Robert Stalnaker, Inquiry. MIT Press, Cambridge Mass./London, 1 984. Pp. xii + 1 87. CLOTH $26.95. D.E. OVER
Downloaded from jos.oxfordjournals.org by guest on January 1, 2011
Stalnaker's book will be welcomed by all those who know his papers on propositions, the propositional attitudes, possible worlds, and conditionals. For this book continues the impressive development of the views so well expressed in those papers. Stalnaker begins by explaining the real foundation of his views. This is that the concept of a possible world is not only of technical utility in formal semantics, but that it is the concept which should be used to account for intentionality or representation. He labels his theory of intentionality or mental representation 'the pragmatic picture'. In this picture beings with propositional attitudes are seen essentially as agents who decided which action to perform out of a range of alternative possibilities. These agents have desires and beliefs - they prefer some possibilities to others and think that some actions but not others would tend to bring about preferred possibilities. On this basis, Stalnaker argues that the objects of the propositional attitudes are possible states of the world, and so that propositions ought to be taken as simply sets of such states. Beliefs and desires, according to this picture, are ascribed content in order to explain action, but how are beliefs and desires to be distinguished from each other? Stalnaker claims, ' We believe that P just because we are in a state that, under optimal conditionals, we are in only if P, and under optimal conditions, we are in that state because P, or because of something that entails P.'(p . 1 8) . He summarizes this analysis by saying that belief is a 'backward-looking' attitude, covered by what he calls 'fidelity conditions', and an instance of a propositional relation he calls 'indication'. Beliefs have a fixed content and are distinguished from desires' . . . because of their presumed causal connections with the world.' (p. l 9) Like his papers Stalnaker's book is admirably concise, but h is analysis of belief is one of the topics on which I feel he should have said more. H is causal or reliability analysis of belief seems to give us the following picture. As 'backward-looking' states, beliefs are formed by a functional system in volving our sense organs and reasoning powers. Given adequate light, a good point of view, sufficient concentration, and other 'optimal conditions', a fact that P will cause this system to be in a state which indicates that P; and if this state is also a part of a further functional system which combines it with desire to explain possible action, then the agent possessing these systems has the belief that P. But this analysis, as it stands, seems to be heir to some of the problems of early analyses of knowledge in terms of causation and reliable mechanism, which in a simple form are inadequate even for knowledge of
272
.
Downloaded from jos.oxfordjournals.org by guest on January 1, 2011
empirical matters of fact. Our knowledge or belief under optimal conditions that all ravens are black is not caused by the fact that all ravens are black. We can know or believe under optimal conditions (on the basis of general, non-constructive reasoning) that the shortest spy is a spy, although we have no causal contact at all with the shortest spy. Certainly Stalnaker says too l ittle about optimal conditions, and he should really have tried to answer some questions about how some people can acquire irrational or totally bizarre beliefs. Is it always possible to define optimal conditions for the systems or sub-systems which produce such beliefs? Are they always condi tions under which such systems or sub-systems would produce rational beliefs? Perhaps he should have tied belief to the requirement of a jusitifica tion and then given a reliability analysis of what constitutes a good justifica tion. He does presuppose a causal and reliability analysis of knowledge or of justification elsewhere in the book. For example, in his chapter on 'Possible Worlds' he asks how we can know any modal facts if possible worlds are as David Lewis describes them and we have no causal contact with them. This is a difficult question for Lewis, but it could have been made even more so by a more open and extended discussion of what it presupposes. An earlier version of this chapter was published as Stalnaker's well known and influen tial paper on 'Possible Worlds', and in it he repeats and expands his good criticisms of two types of view of possible worlds represented by Lewis and by Robert Adams. He also continues to make his own view of possible worlds sound very natural and attractive, but in the end he just does not say enough in a positive vein to counter Lewis's reply, to be found in his new book On the Plurality of Worlds, that there is something 'magical' about Stalnaker's account of possible worlds. Stalnaker argues for what he calls a moderate realism about possible worlds. He sees possible worlds as 'ways things might have been' and as 'elements of our actual world'. (p. 50) Yet he also holds that they are abstract objects which need not be taken as 'among the ultimate furniture of the world'. (p 5 7) They are abstracted with a minimal structure as primitives in a theory of rational agents, and beyond that their nature is left open, according to this view. But consider a false belief, say, that Ortcutt is a spy. The content of this belief will contain only these abstract elements, and Lewis demands to k now how such things can represent Ortcutt as a spy. At places in this book, Stalnaker does get close to the topic of de re belief, such as when he considers beliefs that are 'sensitive to facts' about an object. (p.68) But unfortunately he does not develop here the interesting points on this topic made in some of his more recent papers. Such a discussion would have clarified his notion of a possible world and have possibly answered Lewis's charge about 'magical' representation. An investigation of the concept of being sensitive to facts about an object would also have gone well with an open discussion of the analysis of knowledge or of justified belief which Stalnaker seems to favour. Another important topic which he does examine at length is that of the
273
Downloaded from jos.oxfordjournals.org by guest on January 1, 2011
difference between his pragmatic picture of intentionality and what he calls 'the linguistic picture'. In this contrasting point of view, the fundamental notion is that of the language of thought, which is used to try account for the intentional or representational nature of thought. Hartry Field is a well known supporter of this general picture, and Stalnaker strongly criticizes his position in some detail. Stalnaker, for instance, points out how hopeless it seems to be to try to explain a denotation relation for the language of thought without appealing to some intentional thoughts. A causal theory of reference, of the sort inspired by Kripke's work, does not seem capable of doing this job. As Stalnaker recalls, Kripke himself stresses that he cannot reduce reference to causation, because he makes use of the notion of intending to preserve the reference of a term, whatever the causal history of its use. How could the linguistic picture account for the representational character of an intention to preserve reference? It appears that any attempt to do so would be circular, and thus Stalnaker concludes that the pragmatic picture is correct - that mental representation, as analysed by possible worlds, is fundamental and should be used to explain linguistic representation. Extreme versions of the pragmatic and linguistic pictures are apparently irreconcilable, but Stalnaker realizes that moderate versions of these two views approach each other. He accepts that there must be ' . . . some form of internal representation in any creature that is correctly said to have beliefs and desires.' (p. 22) He wisely remains agnostic about the exact nature of these internal representations; but if any system of such representations is said to be a language, then he is willing to say that in that sense the pragmatic picture requires the existence of a language of thought. However, he argues that there would still be an i mportant difference between his position and that of the linguistic picture. He states that for the pragmatic picture, but not for the linguistic, ' . . . the form in which beliefs and desires are represented is not essential to their content.' (p.23) Now this is yet another point where further clarification is needed. Two different internal representations could have j ust one content under both pictures. Under the pragmatic picture, this content would be a set of possible worlds; while under the linguistic picture, the content could be a single 'cognitive role'. For most supporters of the linguis tic picture seem to agree that 'syntactically' separate 'sentences' in the lan guage of thought could have the same causal or functional powers in the mental system, and this is often what is called the same 'cognitive role'. Even so, Stalnaker could still argue that separate cognitive roles could have the same content in his sense, and that for that matter the same cognitive role could have different contents in his sense in different contexts. Some philosophers who employ the concept of a cognitive role might have a need for some notion of a possible world, but this would have to be a purely 'internal' or 'narrow' concept, defined in terms of something like possibly sensory patterns (as in one of Quine's well known analyses of possible worlds). Of course, Stalnaker's notion is not at all like that - it is an 'external' or 'wide' concept of possible worlds. His position even allows different social
274
Philosophy Dil•ision, Forster Building, Sunderland Polytechnic. Sund�rland, SRI JSD, England.
Downloaded from jos.oxfordjournals.org by guest on January 1, 2011
contexts to determine different relevant sets of possible worlds and different fidelity conditions for agents, although the internal representations and cognitive roles may remain unchanged in the various contexts. With its emphasis on explaining action, the pragmatic picture is, in sum, a wide one, while the linguistic picture is a narrow one - to use 'wide' and 'narrow' in the way Putnam has taught us. There are indeed philosophers who try to com bine the wide and the narrow in a single 'dual-aspect' theory, but Stalnaker would presumably reply to them that the wide, pragmatic picture must always be taken as the more basic. There are standard objections to an analysis in terms of possible worlds of beliefs with necessary or i mpossible propositions as contents, but Stalnaker updates his well known response to these objections in a very interesting and important chapter on 'The Problem of Deduction'. And he there replies to the other standard objection to this type of analysis that we do not believe all the deductive consequences of what we believe. His two pronged counter-at tack on all these objections is, first, to re-state his point that an agent may fail to grasp that a necessary or impossible proposition is expressed by a given sentence; and second, to argue that an agent can have separate belief sub-sys tems, and that the function of deduction is to put these sub-systems into larger, more coherent systems. Unfortunately, it is impossible to discuss here the many deep issues raised in this chapter (including that of a 'qualified consequence condition' of great interest for justified belief). But I will j ust note that his belief sub-systems could only be individuated in general by (internal or external) linguistic criteria, and that this implies some further relationship between his picture and the linguistic one. The second part of Stalnaker's book is a defence of his theory of conditio nals (developed with Richmond H. Thomason). He tries to provide a founda tion for his theory in Chapter 6, 'Conditional Belief, and to defend it in some detail (particularly against Lewis's objections) in Chapter 7, 'Conditional Propositions' (the last part of which is based on his paper, 'A Defense of Conditional Excluded Middle'). In the final, Chapter 8, ' Realism about Counterfactuals', Stalnaker argues for what he calls a m odest realism about counterfactuals against points made .by Mackie, van Fraassen, and Dum mett . This part of the book is required reading for anyone interested in conditionals, but there is no space to say anything more about it here. I may have given the false impression above that I find Stalnaker's writing altogether too tense, but actually I think that the book as a whole is a model of how to write on its difficult topics without wasting words. And it is a measure of Stalnaker's success that he so often stimulates the reader to further thought.
lo1U71al of &mantics 4: 275-278 B O O K REVIEW
T. van Dijk and W. Kintsch, Strategies of Discourse Comprehension. Acade mic Press, New York, 1 983. Pp. xi + 4 1 2. CLOTH $38.50. HAN REICH GELT
Downloaded from jos.oxfordjournals.org by guest on January 1, 2011
The aim of the book is to present "a broadly based, general, but coherent approach to the investigation of discourse phenomena, grounded on the notion that discourse understanding is strategic" (p. ix). The basic assump tion underlying the model is the idea that discourse processing is a strategic process in which a mental representation is constructed of the discourse with the goal of interpreting the text. I n this process both information from the discourse, information from the context , and general background informa tion about the events described in the discourse can be used. Given these general assumptions about discourse processing, the authors are faced with two tasks: - first, they will have to make clear what exactly they mean by a strategic process. Second, they will have to make clear what mental represen tations of discourse look like. Taking the second question first, the theory assumes that a verbal input is decoded into a list of atomic propositions which are then organized, on the basis of some knowledge structure, into a coherent text base. The text base is the semantic representation of the input discourse in episodic memory, and is itself used to construct a macrostructure which represents the most essential information in the input discourse, i .e. the gist of the text. In parallel with the text base, a situation model is constructed which integrates the comprehen der's existing world knowledge with the information derived from the text that is being processed. Thus, the result of the process of text comprehension is a multilevel processing record including traces of the actual linguistic input, of the meaning of the text, and of the effect the text had on the comprehen der's world knowledge. The notion of strategy is discussed in detail in chapter 3, but given the title of the book, it is not surprising that the main body of the book is devoted to strategies and their role in discourse comprehension. In chapter I, the authors distinguish between 6 types of strategy, which are discussed in more detail in chapter 4-9. First, there are propositional strategies discussed in chapter 4, which are used to construct the propositions which form the foundation of the text base. Propositions are constructed on the basis of word meanings, activated from semantic memory, and syntactic structures o f clauses. The construction of a text base has been successful as soon a s it meets a number of criteria, one of which is local coherence. The seond set of strategies are the local coherence strategies which are used to establish meaningful connections between successive sentences in a discourse.They are discussed in chapter 5. The third type of strategy is the subject of chapter 6. They are the so-called macrostrategies which can be used recursively to arrive
276
Downloaded from jos.oxfordjournals.org by guest on January 1, 2011
at a macrostructure of the text, which is the theoretical counterpart of the notion of the gist of a text. A fourth type of strategy is postulated because of the observation that a lot of discourse exhibits a conventional structure. The schematic strategies are used for manipulating such superstructures of a text. They are further discussed in chapter 7. Although the model is mainly concerned with discourse comprehension, the authors also include produc tion strategies in their list and they discuss them in chapter 8. The last major type of discourse processing strategy are strategies for use of knowledge. The authors recognise the importance of world knowledge in production and comprehension and discuss a number of strategies relevant for use of world knowledge. Chapter 9 is devoted to this type of strategy. Chapter 2 contains a review of the literature on discourse comprehension, which the authors admit is highly selective and biased to their point of view. The final chapter of the book, chapter 10, is an attempt at showing how all the processes which are discussed in the previous chapter can take place in real time and it introduces a cognitive model of the language comprehender. In some of the chapters, there are sample analyses of a text fro m Newsweek printed on the inside of the cover of the book which are intended to i llustrate the strategies discussed in that chapter. The main criticism which can be levelled against the book is that the strategies which are given as examples are almost never clearly specified. Similarly a lot of concepts and a lot of terminology is introduced, but almost always without a clear definition. Take for example the notion of a proposi tional representation. In chapter 4 the authors discuss some strategies for deriving propositional representations from sentential surface structures. However, nowhere is there a definition of what propositions look like. There is a somewhat confused discussion about the notion of proposition in the philosophical and linguistic literature. It is for example claimed that accor ding to Montague "a proposition, then, is indeed a conceptual structure, a function, which is "exemplified" (in Camap's terms) by its instantiations, that is, values, in some possible world", which is blatantly untrue - in the realist tradition in logic in which Montague's work is firmly rooted, mathe matical concepts, of which the notion of proposition is an example, really exist out there in the world. On the basis of this discussion then, a proposition is abstractly defined as "an intensional unit, corresponding to the meaning of a sentence in linguistic theory, and tQ the conceptual representation of a sentence in a cognitive model of language". This characterization of the notion of a proposition is somewhat vague to say the least. When it comes to deciding on a representation format for propositions, the authors reject at this stage of development logical systems because "the adequate logical representation even of atomic propositions has become so complex. " So, they use the representation in Fig. I as an example representation of the sentence: Yesterday, John inadvertently gave the old book to Peter in the library.
277 A CT I O N
l
P R E D g i ve(a,b,c)(g)
L
M O D i nadverte ntl y(g)
ARG =
AG E N T J oh n PROP
a
OBJ E CT boo k ( b)
L
M O D o l d ( b)
GOAL Peter
=
c
T I M E yeste rd a y PLACE i n the l i brary
The main problem with the above representation is that the authors never specify what the construction rules are, and in the absence of such rules, the whole exercise becomes pointless. If one of your main points is that the first step in the construction of a mental representation of discourse is transfor ming an input sentence into a list of atomic propositions, and you fail to specify what these atomic propositions look like, then the point of the enterprise becomes somewhat unclear. Obviously, in the absence of a gram matical specification of the representation format, there is also no formal semantics, a fact which the authors admit. Given the unorthodox representa tion format, it is not clear to me how one would give the language a formal semantics. However, the need for a formal semantics and for inference rules becomes clear in chapter 6, in which the notion of a macrostructure is introduced. A macrostructure is "an abstract semantic description of the global content, and hence the global coherence of discourse". They are defined by macro rules, of which three examples are defined: deletion (given a sequence of propositions, delete each proposition that is not a presupposition of another proposition in the sequence), generalization (given a sequence of proposi tions, substitute the sequence by a proposition that is entailed by each of the propositions in the sequence), construction (given a sequence of propositions, replace it by a proposition that is entailed by the joint set of propositions in the sequence). Note that the last two rules use the notion of entailment. But given the absence of a specification of inference rules the above rules are completely void. And if we use a standard definition of valid inference, then the last two rules lead to the conclusion that any set of tautologies is a possible -
Downloaded from jos.oxfordjournals.org by guest on January 1, 2011
CIRC
278
Ctntr� for Cog11itiv� Sci�fla ci D�partm�flt of Artificial lflt�llig�"c� University of Edi11burgh Edinburgh Scotland
Downloaded from jos.oxfordjournals.org by guest on January 1, 2011
macrostructure. Also, given the fact that macrorules are recursive and given the assumption that no two propositions can presuppose each other, then first rule entails that the empty proposition is an acceptable macrostructure as well. Finally, I would like to remark is that the style in at least certain parts of the book is rather long-winded. With a more concise style, it probably would have been easier and certainly quicker to spot the fundamental weaknesses in the book. Summarizing, the main problem with the book is the introduction of a lot of terminology without clear and unambiguous definitions, and in the ab sence of such definitions the book becomes very obscure. Also, the fact that the strategies are not clearly specified means that the sample analyses of the Newsweek text, potentially one of the redeeming factors of the book, become hard to check for the non-initiated, and would have to be accepted as an act of faith.
