JOURNAL OF
SEMANTICS
Volume 1 no. 1, 1982
SWETS & ZEITLINGER BV LISSE
-
THE NETHERLANDS 2000 -
JOURNAL OF SEMANTI...
14 downloads
593 Views
7MB Size
Report
This content was uploaded by our users and we assume good faith they have the permission to share this book. If you own the copyright to this book and it is wrongfully on our website, we offer a simple DMCA procedure to remove your content from our site. Start by pressing the button below!
Report copyright / DMCA form
JOURNAL OF
SEMANTICS
Volume 1 no. 1, 1982
SWETS & ZEITLINGER BV LISSE
-
THE NETHERLANDS 2000 -
JOURNAL OF SEMANTICS AN INTERNATIONAL JOURNAL FOR THE INTERDISCIPLINARY STUDY OF THE SEMANTICS OF NATURAL LANGUAGE
Pieter A.M. Seuren (Nijmegen University)
MANAGING EDITOR:
Peter Bosch (Nijmegen University)
EDITORIAL BOARD:
Leo G.M. Noordman (Nijmegen University)
REVIEW EDITOR
Rob A. van der Sandt (Nijmegen University)
CONSULTING EDITORS: J. Allwood (Univ. Goteborg),
J. Lyons (Sussex Univ.),
M. Arbib (U Mass. Amherst),
W. Marslen- Wil�on
R. Bartsch (Amsterdam Univ.),
J. McCawley (Univ. Chicago),
H.H. Clark (Stanford Univ.),
H. Rieser (Univ. Bielefeld),
Th. T Ballmer (Ruhr Univ. Bochum), J. van Benthem (Groningen Univ.),
(Max Planck lnst. Nijmegen),
B. Richards (Edinburgh Univ.),
G. Fauconnier (Univ. de Vincennes),
R. Rommetveit (Oslo Univ.),
P. Gochet (Univ. de Liege),
H. Schnelle (Ruhr Univ. Bochum),
F. Heny (Groningen Univ.),
J. Searle (Univ. Cal. Berkeley),
J. Hintikka (Univ. Aorida),
R. Stalnaker (Cornell Univ.),
H. Hormann (Ruhr Univ. Bochum),
A. von Stechow (Univ. Konstanz),
G. Hoppenbrouwers (Nijmegen Univ.),
G. Sundholm (Nijmegen Univ.),
St. Jsard (Sussex Univ.),
Ch. Travis (Tilburg Univ ),
Ph. Johnson-Laird (Sussex Univ.),
B. Van Fraassen (Princeton Univ.),
A. Kasher (Tel Aviv Univ.),
E. Keenan (UCLA and Tel Aviv Univ.),
Z. Vendler (UCSDJ.
Y. Wilks {Essex Univ.),
S Kuno (Harvard Univ.),
D. Wilson (UCL).
W. Levelt (Max Planck lnst. Nijmegen),
ADDRESS:
Journal of Semantics, Nijmegen Institute of Semantics, P.O. Box 1454, NL-6501 BL Nijmegen, Holland
Published by the N.I.S.
Foundation, Nijmegen Institute of Semantics, P.O. Box 1454,
NL-6501 BL Nijmegen, Holland
ISSN 0167- 5133
C> by the N.I.S. Foundation
Printed in the Netherlands
CONTENTS page Editorial statement
.
.
.
.
.
.
.
.
.
.
.
.
.
Johan van Benthum and Jan van Eijck The dynam1cs of mterpretatJOn .
.
.
.
•
.
.
.
.
.
.
.
.
•
.
.
.
.
.
.
.
I
. 3
S.C. Garrod and A.J. Sanford The mental representation of discourse m a focussed memory system: Implications for the mterpretatJOn of anaphonc noun phrases . . . • . .
21
S.- Y. Kuroda Indexed predicate calculus .
Susumo Kuno PnnCJples of discourse deletiOn case studies from English, Russ1an and Japanese . . . • .
ISSN
0167
•
SIJJ
.
.
.
•
•
.
•
.
.
.
•
.
•
.
.
•
.
•
43
61
EDITORIAL STATEMENT
It is the central aim of the JOURNAL. .OF SEMANTICS to promote studies in the semantics of natural language. In our century, natural language semantics was, till quite recently, the property of a number of individual disciplines, especially philosophy, linguistics, psychology. What these disciplines offered, in this re�pect, hardly ever went beyond their own boundaries. This is now changing, and the change seems to be rapid. There is a growing awareness that by an interdisciplinary approach solutions are beginning to come into sight for problems that seemed to lie beyond the reach of the respective disciplines in isolation. Moreover, in this climate of cooperation, new problems, or which had resisted exact whose existence was either unknown formulation, are now coming more sharply into focus. Natural language semantics is becoming the common conc:ern of students of philosophy, linguistics, psychology, artificial intelligence, although there still are important differences in method and outlook.
One area where the recent trend toward integration in semantic studies is particularly manifest is discourse phenomena. Philosophers, psychologists, linguists , and students of ar. t ificial intelligence alike seern to feel that the structures and processes involved in the compre hension of texts as orderly accumulations of information·, against JS, vol. l , no. I
Downloaded from jos.oxfordjournals.org by guest on January 1, 2011
In a very literal sense, an autonomous field of studies is constituting itself. We wish to sti mulate this development and promote further integration of the disciplines involved in so far as they concentrate on questions of comprehension and interpretation of linguistic utter ances. We shall welcome contributions, in the form of articles, discus sions or reviews, furtheri ng this aim. This means that the JOU RN AL is intended to be a proper forum not only for more strictly disciplinary studies, as long as they can be read by wider circles of readers, but also for studies on linguistic semantics that go across disciplinary boundaries. It is in the combination of contributions and readership that we hope to cultivate a climate of fruitful interaction and integra tion.
a background of information available through perception or memory, and of mutual knowledge of a "contract" between communicative interactors, are somehow crucial for a better understanding of all kinds ·of semantic phenomena. For this reason, the N IS-Foundation, who publish the JOU RNAL, organized a Colloquium on Discourse Representation in Cleves, September 1 5 - 1 8, 1 9 8 1 .* A number of the papers presented there have resulted in articles, which are distrib uted over the first volume of the JOURNAL. As a matter of policy, we aim at an appropriate balance between the occasional thematic issue and issues containing arbitrary collections of articles and discus sions. This having been said, we can only express our hope and our confi dence that the start of the JOURNAL wtll prove to be propitious. The Editors
2
The Colloquium was made possible by a grant from the Philosophy Faculty of Nijmegen University, which is hereby gratefully acknow ledged. JS, vol. l , no. l
Downloaded from jos.oxfordjournals.org by guest on January 1, 2011
*
THE DYNAMICS OF INTERPRETATION*
Johan van Benthem and Jan van Eijck Abstract
1.
The logic of discourse representation
Two sources of confusion threaten the theory of discourse representa tion: the picture analogy and the careless use of 'semantic tableaus'. Both will be discussed here, in order to create room for a proper enterprise. Next, it will be shown how the current vogue of 'partial models ' (cf. Barwise ( 1 98 1 ), Humberstone ( 1 981)) is related to the first issue. Various suggestions will be made for further connecting research. 3
JS, vol. l , no.l
•
Downloaded from jos.oxfordjournals.org by guest on January 1, 2011
In current semantic theory compositional interpretations are assumed to go from linguistic items to their denotations in some model. This perspective still leaves room for a more dynamical account of how such interpretations are actually created. One natural idea is to assume that each sentence in a discourse is understood through some represen tation, 'mediating' between the language and its models. Thus, the old relation of interpre tation splits up into two new ones, viz. that between linguistic items and their representations, and that between these representations and actual models. Now, at the Cleves conferen ce it was clear that discourse representations are many things to many people. Some view them as syntactic constructs, some as psycho logical ones (yet others prefer to remain confused over this issue). Again, one popular me taphor is that of the partial picture of reality, another that of a procedural recipe for verification. Finally, these representations are supposed to e.rplain such diverse phenomena as anaphora and progressive discourse information. It is not obvious that one coherent notion could do all these jobs. On the other hand, it is not obvious either that one need not try. The purpose of this paper is to clarify some logical issues concerning discourse representa tions, while trying to bring together two of the main themes at the Cleves conference, viz. representation proper and the topic of partial information. General considerations will be found in section 1; section 2 contains applications and illustrations drawn from the two best-devel oped formal paradigms of discourse semantics (cf. Hintikka (1979), Hintikka &. Carlson (1979), Kamp (1981)). It is our contention that more clarity as to the nature and the purpose of discourse representa tion will unite, rather than divide the various currents in this develop ing area.·
VAN B ENTHEM & VAN EIJCK 1 .1
The danger of pictures
Despite the well-known defects of a Wittgensteinian picture theory of language, there is an almost irresistible urge to describe �he division of semantic labour as follows. Each sentence induces a discourse representation, such that truth of the sentence in a model amounts to embeddability of that 'picture' into that model. Now, the term 'embedding' may have various senses. But, even for a quite wide range of such senses, this idea is demonstrably inadequate: Let us assume that our theory assigns, to each sentence S, some discourse model DR(S) such that, for all models M, S is true in M iff DR(S) can be embedded into M. If such an account works at all, it would work for predicate-logical sentences S, one should think. But, a fundamental limitation now reveals itself. Proposition: The only predicate-logical sentences to which the embed ding account applies are (equivalent to) purely existential ones, con structed from (negations of) atomic formulas using only and, or, and there exists. ·
Before one starts protesting, let it be noticed that many of the 'scenarios' in psychological discourse experiments consist ·of such purely existential sequences. ( ' A gentleman entered a shop, and· pro duced a gun ') Another instance is provided by an apparently univer sal counter-example (put forward by Han Reichgelt): A madeus has a horse and Dorothea has a horse, and all these horses bite. A closer look reveals that an equivalent purely existential sentence exists: A madeus has horse that bites, and Dorothea has a horse that bites. •••
We may conclude that a more complex account is needed of the desired relation between discourse representations and actual models, if truth of the original sentence is to be mirrored. And indeed, e.g., Kamp ( 1 98 1 ) has a definition of 'embeddability' which contains essen tially all the machinery of the old Tarski truth definition. There may be interesting intermediate possibilities, however, for the relation between DR(S) and M. For instance, the 'picture' metaphor could also mean something like a 'blurred ' or 'coarse' image, which would be reflected more accurately in the requirement that DR(S) be a homomorphic image of M. (Again, in such cases, model-theoretic preservation results set limits to the faithful rendering of truth.)
4
•
JS, vol.l, no.l
Downloaded from jos.oxfordjournals.org by guest on January 1, 2011
Proof: Observe that if S is true in M and M' extends M, then S is true in M'. (For, if DR(S) is embedded into M, it is, a fortiori, embed ded into the larger M'.) In other words, 5 is 'preserved under exten sions', and by the -los-Tarski theorem of model theory, all such sen tences are logically equivalent to purely existential ones. Q ED.
THE DYNAMICS OF INTERPRETATION But, one might also want to strengthen embeddability to elementary embeddability (which would block the above preservation argument). This possibility will be explored below. Thus, the subject of suitable links between discourse representations and the usual semantic models offers a rich variety of choices to be explored, even though the most simple-minded one has been shown to fail. 1 .2
The ubiquity of semantic tableaus
Since their invention by E. W. Beth and others in 1 955, semantic tab leaus have been put to various uses in logic and philosophy. And indeed, the psychological and mathematical speculations in Beth &: Piaget ( 1 966) foreshadow the present 'discourse representation' ideas to a great extent. The analogy is compelling: tableaus represent existen tial truths by means of discourse referents just as moderno authors would have it. And the same goes for the branching storage of disjunc tions, as well as other connectives.
For example, a universal quantifier is treated differently in the two cases. It will get an abstract 'generic' representation in a struc ture tree; whereas, in a semantic tableau, it becomes a standing instruction to introduce requirements for new individuals in the tab leau. Or, to demonstrate the difference from yet another angle, for the purposes of anaphora, the following sentence is a perfectly ordinary case: Some woman loving no one is loved by all women she does not love. It is only the tableau analysis wich reveals its contradictoriness (which does not prevent it from having anaphoric relations). Nevertheless, semantic tableaus are extremely interesting in the earlier perspective of 'little models' . For one thing, open tableau branches are themselves models for the original se�tence: the analysis has been pushed through completely, while no contradictions occurred. (We are referring here to familiar completeness proofs in terms of semantic tableaus.) More precisely, each open branch in a completed tableau for a sentence S of predicate logic represents a class of JS, vol . l , no. I
.5
Downloaded from jos.oxfordjournals.org by guest on January 1, 2011
It is very tempting, then, to think of semantic tableaus as prime candidates for discourse representations - a tendency which is rein forced by tableau-like terminology in many expositions (cf. Kamp ( 1 98 I )). Nevertheless, one should be extremely careful here, distin guishing between various uses of discourse representations. As will be shown in section 2 of this paper, applications to topics such as anaphora usually necessitate a rather syntactic representation, close to the structure trees of sentences in a dis.course. On the other hand, analyses of discourse information require a kind of 'thinking' interpre tations, combining and comparing the requirements upon models ex pressed by various components of the sentences. This is the area of semantic tableaus proper.
VAN BENTHEM & VAN EIJCK models verifying S, whose domain equals any set of individuals i for which a discourse referent d occurs on the branch, and whose interpre tation verifies atomic form J las on the 'true' side of the branch, while falsifying those on the 'false' side. (There is usually quite some margin here, as not all atomic for_mulas need be decided on the branch.) There is a price to be p aid for this, of course: open tableau branches may be (irremediably) infinite. Are these 'branch models' for a sentence S somehow 'representative' for all models of S? In a sense, they are, and we find to our surprise that there exists a kind of 'embedding' connection after all - circum venting the proposition in section 1.1: Proposition: A predicate logical sentence S is true in a model M if and only if M is an L(S)-elementary extension of some branch model for S. Proof: (L(S) is the sublanguage of the full L consisting of S together
·
A comment is in order here. The preceding theorem is easily extend ed to cover the case where L(S) is the full sublanguage of L generated by the non-logical vocabulary of S. (One has to enrich the tableau for S in some standard fashion, alternating applications of tableau decomposition rules with introduction of Excluded Middle formulas S' or not S ' for all S ' in some fixed enumeration of L(S).) But, the simpler form given here stays closer to the idea of using nothing beyond the components of the represented sentences. Essentially, the above theorem may be found in Kreisel, M ints & Simpson ( 1 97 5), section 1, 1 , and in Pra�itz ( 1 975), section 3. What does such a formal theorem about predicate logic tell us concerning natural language in general? Well, at least we see that the 'dangerous' embedding idea is viable formally, when taken in 6
JS, vol . l , no. l
Downloaded from jos.oxfordjournals.org by guest on January 1, 2011
with all its subformulas.) ' If ' : S is true in all its branch models, and hence in all their L(S) elementary extensions. 'Only if': Suppose that S is true in M. We associate a tableau branch with M by choosing nodes in the complete tableau tree for S, together with an assignment of real individuals in M to d-iscourse referents in the tableau, such that M verifies all formulas on the ' true' side of these nodes, while falsifying all those on the ' false' side. (This can always be done, starting from a single true S in the topnode, by following the tableau-rules, using actual truth (or falsity) in M to make decisions at v -branchings and choices at l-representations.) Now, the set of all individuals in M assigned in this way forms a submodel M' of M which is a branch model for S. Moreover, by the above construction, M is an L(S)-elementary extension of M'. For, occurring on the true (false) side decides truth (falsity) in M, M' in the same way. Q ED.
T H E DYNAMICS OF INTERPR ETATION some appropriate sense. (The succes of quite different forms of tableau embedding, witness Rodenburg ( l 98 I ), only reinforces this point.) Thus, one has a guide-line as to which directions of thought concerning the semantic role of representations are worth exploring - while avoid ing the dead alleys closed off by our first theorem. It remains to be repeated that this result employs infinite representa tions in general. If one insists on finite discourse representations, then, e.g., instead of spelt-out V'l-dependencies, one will have to represent rules (say, as Skolem functions). Thus, the recipe metaphor will be re-instated over the picture idea; as in Hintikka's game-theoreti cal semantics (ct. section 2 of this paper). Let us summarize the kind of enterprise emerging from the previous considerations in the following picture: --
partial discourse discourse representation - model
L---�
1.3
Partial models
Possible world semantics is losing favour with the semantic community. A possible world is an idealized complete state of affairs (or a total state of information concerning such a situation). From various direc tions, 'partial' alternatives are gaining ground (cf. Barwise ( 19 8 1), Humberstone (1981) but the idea is a!rea4:ly found in the early seven ties with Kit Fine's work on relevance logics). Indeed, forcing seman tics for intuitionistic logic has always been a semantics of growing partial information sets (despite the formal analogy with complete world structures). Thus, the idea of using partial models and partial information is in the air. -
At the Cleves conference, a psychologist suggested an even more radical move, pleading for partial individuals. Again, that idea is fore-shadowed already in current interval tense logics: intervals may be thought of as partial temporal individuals, which have not yet made up their minds about the precise points they are going to be (cf. Van Benthem ( 1 982)). We will return to this example below. Now, these ideas may lead us to a road diverging somewhat from the enterprise outlined above. For, a radical representationalist might just as well forget about the 'actual models', and formulate a truth definition directly on the partial discourse models (say, in the context JS, vo!. ! , no. !
7
Downloaded from jos.oxfordjournals.org by guest on January 1, 2011
The first arrow may stand for an algoritmic production, the second for a tableau-like analysis, the third is the beckoning 'picturesque' link between 'representation' and 'reality ' . Whether there are more intermediate stages of representation to be distinguished will depend on intended applications.
VAN BENTHEM & VAN EIJCK of all possible such models). Various clauses are possible here, most of them with a non-classical ring. For instance, negation will probably be treated intuitionistically as 'absence of truth in all possible exten· sions'. Or, to mention a more suggestive case, disjunction need not be distributive any more over partial models: one need only require that a choice between the disjuncts be made 'eventually ' . (Let it be noticed, however, that the terms 'intuitionistic' and 'classical' are rather treacherous in this area: the above clause for disjunction has again classical effects!) For those who dislike this intUJtlOOJStlC turn, there is another road, again suggested by semantic tableaus. When showing that 'true' ( 'false') branch formulas are verified (falsified) in branch models, one is led naturally to think of both truth and falsity as complementa ry, and necessary notions. Thus, one might also start with primitive notions of 'verifying' and ' falsifying' for partial models, which leads to a classical negation. (Cf. Veltman (198 1).)
Theorem: For predicate-logical sentences S, , S21 where S2 is a. sentence in the vocabulary of S , , S, implies 52 if and only if S 2 actually occurs on the true side of each open branch of the S. -tableau.
Proof: 'Only if': Branch models of S, verify S , and hence also �. Therefore, not-S 2 cannot occur on the true side of any open branch - whence s2 does. 'If': If M is an arbitrary model in which S, is true, then - by the earlier theorem - M contains an L(S, )-elementary submodel M' which Now, S 2 occurs on the true side of the is a branch model for S, M'-branch, whence it is true in its L(S, )-elementary extension M. Q ED. •
Again, the message of this formal result is that Seuren's suggestion may at least be viable in some interesting sense for natural language - something which came as a surprise to (at least) these authors. 1 .4
Semantic reunion
The diverging picture which has arisen may now be re-united thus:
8
JS, vol. l , no. l
Downloaded from jos.oxfordjournals.org by guest on January 1, 2011
That these partial discourse models may indeed be all we need to know was also suggested by the colloquium organizer Pieter Seuren, who conjectured that even !ogical consequence may be captured ade quately at this level already. And as it happens, at least at the level of semantic tableaus, this is true. It is true in the trivial sense that S, implies S2 if and only if (S , and not S2 ) has a closed semantic tableau (for predicate logical sentences S, , S 2 ). But, it is also true in a more interesting sense, for tableaus in the extended form men tioned in connection with the preceding result.
THE DYNAMICS OF INTERPRETATION
discourse representation
partial discourse model
As Jaakko Hintikka urged the partiCipants to do, one should look for connections here. For example, there is the 'super-valuation' impulse of striving for connections of this kind: 'S is true in a partial model if and only if it is true in all its com· plete extensions to actual models.'
T o show that t h i s goal does not represent a n idle hope, here are two results from Van Benthem (1982) confirming this message for the case of interval tense logic. Modulo some technical background conditions, ( l ) a tense-logical sentence is (interval-)true at an integer interval if and only if it is (point-hrue at all in tegers w i t h i n that interval, (2) a tense-logical sentence is (interval-)true at an open rational interval if and only if it is (point-hrue at most r a t ionals within that in�erval (i.e., in all of them up to some finite number of exceptions). It is our hope that results like these (and the previous ones) indicate some lines along which integrative research will be done, while avoid ing some of the logical road-blocks awaiting earlier formulations of the nature and pretentions of discourse representations. 2
Two paradigms investigated
Two theories of discourse representation that have left the programma tic stage are those of Jaakko H intikka and Hans Kamp. (Henceforth, familiarity is presupposed with H intikka (1979), H intikka & Carlson (1979), and Kamp (1981).) These will now be analysed in the spirit of the preceding section. Striking formal resemblances come to light, even when empirical predictions may differ. (As both theories share a concern with anaphoric relations, this topic will be the focal exam ple.) It will be seen how this type of theory fits into the earlier meth odological chart, and some morals will be drawn from that awareness. JS, vol. l , no. I
9
Downloaded from jos.oxfordjournals.org by guest on January 1, 2011
Usually, the partial truth definitions and classical truth will not connect up as simply as this. But, the idea motivates a ' target equiva lence' which may well become extremely fruitful for further research: 'S is true in a partial model DM(S) if and only if the set of complete actual models M bearing a suitable 'embedding' relation R to DM(S) is suitably large . '
VAN BENTHEM & VAN EIJCK 2.1
H intikka's games
Pictures and games When it was remarked at the Cleves conference that some discourse representationalists are guided by a picture metaphor, while others view their constructs as procedural recipes for verification, Jaakko Hintikka replied that his game-theoretical semantics combines both view-points: it provides recipes for making pictures . N e v e r t heless, the connection between the rather procedural game-theoretical seman tics and a picture theory is less intimate than this reply suggests. And indeed, the discussion of W ittgenstein in Hintikka (1976) contains no more than an invitation to think of the relation between atomic s e n tences and the facts in a model a s pictural link.
In game theory proper, a strategy is a function from possible game situations to possible game situations, telling a player at every stage what to do next. These strategies are usually finitely representable in a 'game tree', since a player has only finitely many moves available, and the game has a limited length (either inherently, as in poker, or through some stipulation, as in chess). Many theorems of game theory depent vitally on this finiteness of the set of available strate gies for the participants. But, H intikka's strategies that play such a conspicuous role in his theory cannot be finite objects in general. When a language is played relative to an infinite model, the relevant 'Skolem functions' wiU have to encod� infinitely many possible moves. This observation may also explain the absence of any significant applications of game-theoretical results i n s i d e H i n t ikk a ' s theory. The preceding diagnosis does not imply that n o significant mathemat ical theory is possible about infinite strategies. For instance, there exist some deep foundational studies in logic concerning the so-called 'Axiom of Determinateness', which states that all two-person games on the natur-al numbers provide a winning strategy for one of the plavers. But of such notions and results, one finds no trace in game-
10
JS, vol. l , no. l
Downloaded from jos.oxfordjournals.org by guest on January 1, 2011
Talking about analogies, the link between game-theoretical seman tics and the mathematical Theory of Games is also less intimate than Hintikka has it. H is semantic truth/falsity games are extremely simple zero-sum games of a complexity below the treshold where Von Neumann started creatin� the mathematical discipline of game theory. (E.g., the original Mmimax Theorem is already about the existence of 'equilibrium choices' of strategies for two players: a topic beyond the pale of game-theoretical semantics - at least, in its present state.) Moreover, many theorems of game theory proper belong to finite combinatorics. This fact reflects another important difference, which may be illustrated by ·considering the following key notion in any study of games.
THE DYNAMICS OF INTERPRETATION theoretical semantics either. Whichever way one looks at them, 'seman tic games' remains, at best, a suggestive metaphor. Games and truth
In game-theoretical analysis, truth means the existence of some win ning strategy for Myself (against Nature), with respect to a (total) model M. As was noted above, the nature of M may require infinite strategies. Again at the Cleves confence, it was suggested that this problem might be circumvented by withdrawing to the representational level, restricting the domains of universal quantification to already available (finite!) sets of discourse referents. But, such a way-out amounts to doing away with universal quantification altogether. Sen tences with only such pseudo-universal quantifiers reduce to purely existential ones, as was noticed in section 1. I. Hintikka-trees and discourse representations
not A·
•
!A A and B A.
A!
s.
•
•
M / \ ·A B·
JS, vol.l, no.l
.A and B
•
N / \
i f A, then B
not A
/M\ A .s •
i f A, then B
/\ N
A
B
11
Downloaded from jos.oxfordjournals.org by guest on January 1, 2011
Fortunately, no concrete strategies with respect to some actual model need occur in completely informative representations o� the above semantic games. The gist of the game rules can be captured in the following finite tree-format. Put sentences defended by Myself (M) (i.e., attacked by Nature) to the left of a node, sentences defended by Nature (N) (i.e., attacked by Myself) to the right. Mark choices with the name of the player that is to make them. Then, the game instructions build trees according to the scheme
VAN BENTHEM
&
VAN
E IJCK
every X that Y's Z's if a is an X and a Y's, then a Z's some X that Y's Z's
N
M
I
I
M
N
I
a is an X, a Y's and a Z's
I
every X that Y's Z's if a is
an
X and a Y's,
then a Z's-
some X that Y's Z's
a is an X, a Y's and a Z's
Trees and tableaus
The above instructions for negation, as well as those for the quanti fiers bear a close resemblance to Beth-type tableau rules. This goes a long way to explain the almost universal feeling that there must be a close connection between the two. Nonetheless, the above 'Hintik ka-trees' (H-trees, henceforth) that were associated with natural language sentences behave very differently from serT)antic tableaus, as the following example will show. One possible reading of the sentence
Everyone loves himself, and not everyone loves someone.
has the following H-tree:
everyone loves himself, and not everyone loves som everyone loves himself a loves a
� •
•
7� � not everyone loves someone
I
•
M everyone loves someone I N b loves someone I •
12
b loves c
JS, vol.l, no. I
Downloaded from jos.oxfordjournals.org by guest on January 1, 2011
The atomic sentences A arrived at eventually are then to be checked, with the winning convention that: M wins if A occurs to the left (right) and it turns out true (false), N wins if A occurs to the left (right) and it turns out false (true).
THE DYNAMICS OF INTERPRETATION Notice that no strategies are indicated for the players (after all, no specific model M is present). The tree only indicates which player has to move and what his options are. As such; it does not provide the information that this sentence happens to have a winning strategy for Nature in all models. But a Beth tableau for this sentence will 'close', revealing its inconsistency as follows. everyone loves himself, and not everyone loves someone everyone loves himself, not everyone loves someone everyone loves himself
•
I
•
I •
a loves someone
I ! a loves a
.
Thus, in a sense, 'Beth tableaus reflect upon Hintikka trees': they reason about strategies. In this particular example, the tableau teJJs us that the assumption that Myself has a winning strategy would lead to a contradiction: it follows that Nature has a winning strategy in all cases. More concretely, the above H-tree ·opened with two options for her. The Beth tableau tells N that at least one of these will be a winning one (although it does not say which one: that will depend on the particular model M). Nevertheless, why then the persistent tendency to also view Beth tableaus themselves as a kind of game? The reason is that they amount indeed to games of a rather different kind (closer to those envisaged in game theory), studied in the logical 'dialogue theory' of Lorenzen & lorenz ( 1978). This connection will not be pursued here. Summing up, the H-trees are syntactic analysis trees, doing the usual jobs such as determining relative scopes of operators, whilst also providing some information concerning verification of the sen tence. It is the latter feature which makes for their interest in the following application. Anaphora
Hintikka sees a wide range of applications for his game-theoretical semantics: this many-purpose tool treats scope-phenomena in the JS, vol.l, no.l
13
Downloaded from jos.oxfordjournals.org by guest on January 1, 2011
a loves a
everyone loves someone
VAN BENTHEM
&
VAN EIJCK
behaviour of 'any' just as easily as issues in the meaning of knowledge and belief. For purposes of comparison with Kamp's theory, we demon strate its way of handling anaphoric phenomena. Linguists usually distinguish between anaphora at sentence level and anaphora crossing sentence boundaries. Concerning the first kind, Hintikka has not all that much to say: as in most logical seman tics, the !-{..:tree-rules for the quantifiers introduce ordinary bindings. (If anything, an account of the relevant anaphoric possibilities is presupposed here.) In order to account for anaphora in if-then sentences, or across sentences, the H-tree instructions are to be read according to the 'Progression Principle' (Hintikka & Carlson ( 1979)), p�escribing an order of playing from left to right. For instance, the qrdinary propositional rule for if-then, amounting to my choice of attacking the antecedent or defending the consequent, now assumes the following form if A, then B
If a soldier owns a gun, he cleans it.
The succesful N-strategy consisted in producing an example of a soldier with his gun, and these are now available for backward refer ence. Even more spectacularly, If every soldier owns a gun, some soldier cleans it.
may be explained likewise. Nature's strategy produces a gun for every soldier, and this function is triggered by the phrase 'some soldier' to provide a referent for 'it'. This simple story is very attractive, but also very implausible. If the Progression Principle means that one 'subgame' is played for A, after which the players (may) move to B, then Nature's winning strategy cannot be known. For, such a strategy involves typically all N's responses to moves of Myself, which cannot be divulged in a single play. (One would have to play all possible games concerning· A.) Thus, there is a dilemma: either we have a natural course of the game, without full strategies available, or we have the latter without the former. Our point remains the same as before. Having the full strategies 14
JS, vol.l, no. I
Downloaded from jos.oxfordjournals.org by guest on January 1, 2011
The semantic story accompanying the tree now goes as follows. First, we play the A-game. Either M can win, and the game is over, or M ·cannot win. In the latter case, N has divulged a winning strategy for A, which M can use now to his advantage in the second round, when he defends B. This story explains the anaphora in
THE DYNAMICS OF INTERPRETATION available is unrealistic, but also unnecessary. All predictions concern ing anaphoric possibilities in connection with the Progression Principle can be formulated entirely at the level of H-trees: it is enough to know which positions on the left branch are available as anteced ents to which positions on the right branch. More precisely, the rule mi�ht be that an individual in the left branch of the H-tree for an if-then sentence that results from an N-choice after a certain (possibly empty) number of M-choices of individuals, can s e r v e a s a n least antecedent for anaphora occurring i n the right branch after a t that number of M-choices. (This is for purposes o f illustration only. As it stands, the rule is certainly not correct - even discounting pragmatic disturbances.) The strategy story then remains, in the background, as a semantic motivation for these predictions. This is a mere methodical point, of course. The vagueness of the above strategy account is not removed. (E.g., what are suitable triggers? Can a triggered occurrence still exhibit an anaphoric ambiguity?) But then, this is not a paper about anaphora. 2.2
Kamp's discourse structures
Kamp (1981) opens with the promise that his theory may yet provide the missing link between formal semantics and the psychology of linguistic competence: "(. . . ) discourse representations can be regarded as the mental representations which speakers form in response to the verbal inputs they receive. " (p. 282)
Moreqver, a 'radical departure from existing frameworks' is needed, giving rise to the following attractive notion of truth:
"A sentence S, or discourse D, with representation m is tn.te in a model M if and only if M is, compatible with m; and compatibility of M with m, we shall see, can be defined as the existence of a proper embedding of m into M, where a proper embedding is a map from the universe of m into that of M which, roughly speaking, preserves all the properties and relations which m speci fies of its domain. " (p. 278)
We are not qualified to pass judgment upon the psychological ·connection; though it should be remarked that procedural models (as opposed to pictural embedding metaphors) seem to enjoy the ascendancy in current psychology. About the account of truth, more can be said straight-away. Embedding and tn.tth
'Roughly speaking ... '. Upon closer inspection, Kamp's formal defini tion of a 'proper' (or 'verifying') embedding turns out to depend recursively on the complexity of the sentence, introducing quantificaJS, vol.J, no. I
1.5
Downloaded from jos.oxfordjournals.org by guest on January 1, 2011
Model theory and psychology
VAN BENTHEM &: VAN EIJCK tion over such embeddings for each layer of universal quantifiers or if-then constructs. For instance, the following sentente (from Kamp's fragment) If a soldier loves Mary, every widow hates him
will be interpreted eventually as follows. - There exists an embedding f of 'Mary' into the model such that - for every embedding g extending f by assigning a soldier that loves f ('Mary') to 'a soldier' it holds that - every embedding h compatible with g and assigning any widow to 'every widow' will result in that widow hating g ('a soldier'). · Anyone who has taken the trouble of writing out the successive clauses of a Tarski-type truth definition with the full paraphernalia of a s s i g n ments will recognize essentially the same complexity in both cases.
Discourse tableaus
Kamp's 'discourse representation structures' (DRS's) are presented in the familiar tableau-terminology of introducing 'discourse referents'. (Even some of the didactic recommendations are reminiscent of intro ductory logic courses using semantic tableaus.) Still, a DRS is more like a structure tree: as in the Hintikka case, no analysis takes place of the information in the constituents. This point is brought out more clearly by a comparison between Kamp tableaus and Hintikka trees. Instead of losing ourselves in the dreary formalistic details attaching to any explie:it tableau method, let us consider an example, viz. the above soldier-senten��. Its H-tree and its DRS (modulo some technical ities) are given below. The similarities are so obvious that they hardly need spelling out. A technical comparison (not displayed in this paper) will show the following correspondences: to each DRS, there corresponds an H-tree such that the DRS is successfully embeddable in a model M if and only if the H-tree allows for a winning M-strategy with respect to M. not every H-tree is thus derivable from a DRS. The reason behind the second observation is simply that Kamp's frag16
JS, vol. l , no. l
Downloaded from jos.oxfordjournals.org by guest on January 1, 2011
This is not to say that Kamp's truth definition via embeddings is not original. It handles nasty cases of anaphora which Tarski's account does not cover. But the main point here is that 'compatibility' and 'embedding' turn out to introduce essentially the same complexity into the link between discourse representations and actual models that one had in ordinary logical semantics. Admitte�ly, there is a borderline case where the above embedding idea does work out precisely as promised in the introductory quotation. That case occurs when the discourse contains no universal phrases or if-then s e n t e n c e s; i . e . , when it consists of a sequence of purely existential sentences. And we are back at the theorem of section 1 .1.
THE DYNAMICS OF INTERPRETATION ment does not treat negation and disjunction, which allows him to get by with 'one-sided' tableaus, where H-trees are essentially 'two sided' (containing M- and N-roles). If Hintikka's fragment were restrict ed in a similar manner, H-trees could be simplified so as to leave M-defensive-actions only.
?��
if a soldier loves Mary, eve
• a soldier loves Mary N
every widow • hates him N
soldier (u), • and u loves Mary M
if widow (v), • then v
I
�\
•
soldier (u)
H-tree
I
7��
.u loves Mary .widow (v)
v hates him•
a soldier loves Mary soldier (u) u loves Mary
Anaphora
As with Hintikka, the predictions on possible anaphoric relations in if-then sentences can be formulated entirely at the DRS-level. Also analogously, they are motivated by the subsequent interpretation mecha nism. So considerations about available embeddings take over from those about available strategies. And here is where a difference reveals itself. Kamp's anaphoric rule is simply this:
"a pronoun can only be anaphorically related to discourse referents in the same tableau box or in a box related to this through a sequen ce of steps (i} move one box up, (ii) move to a left sister box."
For, precisely these antecedents have. become available �uring the (bottom down) interpretation process. Althou�h Hintikka's rule is not equally well-defined, it will presumably allow all these possibilities. But, it will allow even more, because a winning N-strategy in the left branch of an H-tree for an if-then sentence reaches down to the JS, voi.J, no.!
17
Downloaded from jos.oxfordjournals.org by guest on January 1, 2011
DRS:
VAN BENTHEM & VAN EIJCK very bottom. Thus, Hintikka allows (while Kamp excludes): I( every soldier owns a gun, some soldier clearu it. Thus, various ideas about interpreting the same kind of discourse repre sentations, implying different anaphoric predictions, can live side by side.
2.3
Conclusions
It has been noticed already that the anaphoric predictions supported by the strategy idea and the embedding idea are not the same. It is natural, then, to turn to the linguistic literature for impartial arbitra tion. Unfortunately, however, it turns out that even the best current accounts (Reinhart ( 1 976,1 980)) do not cover the a!;>ove type of sen tence. (The formal analogy between Kamp's 'up-left' rule fo� admissible antecedents and Reinhart's choice of a domain through the c - c o m m a n d relation appears to be accidental, on closer inspection.) Still, an illuminating contrast comes to light between H intikka and the linguists on the one hand, and Kamp on the other. Kamp claims that his theory provides a uniform treatment of aJJ kinds of anaphoric relations. This runs counter to the accepted linguistic division, implicit ly acknowledged by H i ntikka, into ( l ) pronouns that permit a bound variable interpretation, and (2) those that have to be interpreted refer entially (cf. Evans (1980)). Thus, e.g., most linguists consider the follow ing example (in Kamp's fragment) structurally ambiguous: Every soldier who loves a widow who loves him is happy. Either 'him' is bound by 'every soldier ' , or the pronoun refers to an individual mentioned previously in discourse; note that we cannot tell what the H-tree for the sentence looks like before this ambiguity is removed. Kamp, however, leaves it to the DRS to settle the differ ence. So, while Hintikka and Kamp both need an autonomous level of sentence syntax as input for their 'discourse syntax ' , they disagree on the whereabouts of the dividing line between the two.
18
J S , vol. l , no. l
Downloaded from jos.oxfordjournals.org by guest on January 1, 2011
· The above 'rapprochement ' between the theories of Hintikka and Kamp has passed by some obvious differences in their over-all approach. Notably, H intikka only gives some discursive examples of how games are to be associated with surface (but then again, not quite surface) sen tences; whereas Kamp' s achievement resides to a great extent in an al�orithmic production of DRS's for sentences in a certain well defined fragment. On the whole, the greater credit must go to Kamp here; because it remains exceedingly difficult to establish just what is the scope of the game analysis. Nevertheless, Hintikka's looseness may be more natural in the sense that only the order of playing deter mines scope relations, not some pre-given syntactiC analysis. So, one sentence can get several H-trees. But then, such a modification is easily introduced into the Kamp approach as well, say through some Montagovian relation R.
THE DYNAMICS OF INTERPRETAT ION These considerations suggest the following way to fit a Hintikka/ Kamp enterprise into the methodological scheme of section 1 .2 : discourse
discourse representation
consisting of sentences f-structured by 'sentence syntax'
a supplementary structure provided by 'discourse syntax' ('enlightened H-trees' or 'Kamp DRS's')
actual models
'winning M-strategy' 'truthful embedding'
Rijksuniversiteit Groningen Filosofisch Instituut . Westersing�l 1 9 97 1 8 CA GRONINGEN
Note
This paper arose out of a paper read at the Cleves colloquium on Discourse Representation and the ensuing discussions. We would like to thank in particular Barry Richards and Goran Sundholm for their helpful comments. Part of the research for this paper was sponsored by the Netherlands Organisation for the Advancement of Pure Research (ZWO), �rant no. 22-65. *
JS, vol.l, no. l
19
Downloaded from jos.oxfordjournals.org by guest on January 1, 2011
Sceptics might argue that, since all the action seems to occur at the middle level, the semantical part is just a ritual addition. But, as we have seen, this misses a vital point: semantic interpretation procedures turned out to motivate the workings of our discourse repre sentations. What is true, however, is that the semantical part remains rather traditional, in that the usual total models are assumed in the background. Apart from a sympathetic, but cryptic reference by Kamp to Veltman ( 1 98 1 ), there are no signs of a more radical break with traditions by having a partial semantics, closer to the discpurse represen tations themselves. So, the next task for the proponents of a rigorous, but radical theory of discourse representation lies straight ahead.
VAN BENTH E M & VAN EIJCK References
·
20
JS, vol. l , iio. 1
Downloaded from jos.oxfordjournals.org by guest on January 1, 2011
of Philosophy. Barwise, J., 198 1 : Scenes and other situations. Journal 78.7; .:69-397. Beth, E.W. & Piaget, J., 1966: Mathematical Epistemology and Psychology, Reidel, Dordrecht. . Evans, G., 1 980: Pronouns. L inguistic Inquiry, II; 337-362. H intikka, J., 1 976: Language-games. In: Essays on Wittgimstein in Honour of G. H. von Wright, North-Holland, Amsterdam. Pp. 105- 1 25. Hintikka, J., 1979: Quantifiers in natural languages: some logical prob lems. In: H intikka et al. (eds.), Essays on Ma the m a t ical and Philosophical Logic, Reidel, Dordrecht. Pp. 295-314. Hintikka, J . & Carlson, L., 1979: Conditionals, generic quantifiers, and other applications of subgames. In: Avishai Margalit (ed.), Meaning and Use, Reidel, Dordrecht. Pp. 57-92. Humberstone, L, 1 9 8 1 : From worlds to possibilities, Journal of Philosoph ical Logic, 1 0; 3 1 3-340. Kamp, H . , 1 9 8 1 : A theory of truth and semantic representation. In: J. Groenendijk et al. (eds.), Formal Methods in the Study of Language, Mat hematical Centre, Amsterdam, vol. I. Pp. 277-322. Kreisel, G., M ints G.E. & Simpson, S.G., 1 975: The use of abstract language in elementary Metamathematics: some pedagogic examples. In: A. Dold & B. Eckmann (eds.), Logic Colloquium, Springer, Lecture Notes in Mathematics 453, Berlin. Pp. 38- 1 3 1 . Lorenzen, P . & Lorenz, K., 1978: Dialogische Logik, W i s s e n s c h�f t l i c h e Buchgesellschaft, Darmstadt. Prawitz, D., 1 975: Comments on Gentzen-type procedures and the classical notion of truth. In: A. Dold & B. Eckmann (eds.), Proof Theory Symposium, Kiel 1974, Springer, Lectures Notes in Mathematics 500, Berlin. Pp. 290-3 1 9. Reinhart, T., 1976: The Syntactic Domain of Anaphora. Unp u b l i shed Ph. D. disser�ation, M.I.T. Reinhart, T ., 1980: Coreference and bound anaphora: a restatement of the anaphora questions. Un�ublished typescript, Max Pianck-Institut, Nijmegen. Rodenburg, P.; 1 9 8 1 : Intuitionistic correspondence theory. Report, Mathematlsch Instituut, U niversity of Amsterdam. Van Benthem, J .F .A.K., 1 982: The Logic of Time, R e i d e l , D o rdrech t . Veltman, F . , 1 9 8 1 : Data semantics. In: J . Groenendijk et al. ( e d s . ) . , P p . 541-566.
THE MENTAL REPRESENTATION OF DISCOURSE IN A FOCUSSED MEMORY SYSTEM: IMPLICATIONS FOR THE INTERPRETATION OF ANAPHORIC NOUN PHRASES
S.C. Garrod and A.J. Sanford Abstract
1974).
While any attempt at producing a process-model for comprehension
JS, vol. l , no. l
21
Downloaded from jos.oxfordjournals.org by guest on January 1, 2011
To a cognitive psychologist discourse comprehension poses a number of interesting problems both in terms of mental representation and mental operations. In this paper we suggest tha t certain of these prob lems can be brought into clear focus by employing a procedural ap proach to discourse description. In line with this approach a general framework for the mental represe . ntation of discourse is discussed in which distinctions between different types of memory partitions are proposed. It is argued tha t one needs to distinguish both between focussed representations available in immediate working memory and nonfocussed representations available in long-term memory and a lso between representations arising from the asserted information in the discourse and those arising from what is presupposed by it. In the second half of the paper a particular problem of anaphoric reference is discussed within the context of this framework. A general memory search procedure is outlined which contains three parameters for deter mining the search operation. We then attempt to describe certain anaphoric expressions such as personal pronouns and full definite noun phrases in terms of the execution of this search procedure, where distinctions arise from the parameter specification derived from the expressions. The cognitive psychology of discourse is concerned with the nature of the mental processes entailed in understanding what is written or spoken, and the problem of how these processes might be realised in the mind of the understander given the psychological constraints of limited attention and memory which w'e know to obtain. One very attractive li1Je of attack is to view the many and various aspects of a discourse as having an instructional component, in the sense that the reader or listener is being instructed to assemble representations of the elements of discourse in a particular way. An e.rample of this is to be found in a treatment of topic marking within the topic/com ment distinction (Halliday, 1976): topic identification may be thOught of as an instruction to implement a procedure in which the topic con tent is construed as an address in memory to which new (comment) information is to be affixed (e.g. Broadbent, 1973; Haviland &. Clark,
GARROD &: SANFORD inevitably makes use of such a procedural view, it is also sensible to consider a text as having a content, which is more directly interpret able as a set of statements. In the present paper, we shall first consider the question of text content. This immediately raises the problem of how to treat anaphoric reference, which is one of the key contribu tors to text cohesion. Finally, we shall attempt to illustrate how the instructional or procedural aspect of discourse interacts with the con tent aspect by reference to a specific problem of anaphoric reference. Discourse Content
The principal differences between these two approaches lie in the extent to which the mental representation of a discourse matches the form of words making up the discourse, and as a corollary the extent to which 'inferences' in discourse comprehension are made immediately and automatically on encountering each element of it. In fac;t, the second view assigns the establishment of the s i g n i f i c a n c e of a piece of discourse to a very early stage in its processing. We shall ·now illustrate some of the consequences of this. Thus, while the propositional structures of the following two sentences are very similar, they differ quite markedly in their significance: ( 1) The policeman held up his hand and stopped the bus. (2) The wicket-keeper held up his hand and stopped the ball. Thus following (2) with (2') forms a perfectly acceptable piece of discourse, but adding (2') to ( 1 ) does not: (2')
The score was still the same.
This would not be of any consequence if the continuity problem arose from the fact that (2') identified some particular element in (2) which was not present in ( 1), but this does not seem to be the case. The problem comes from the definite NP 'The score' in (2') but this does not identify any single phrase in the prior sentence, and is pretty remote from our understanding of 'wicket-keeper', 'hand ', 22
JS, vol. l , no. )
Downloaded from jos.oxfordjournals.org by guest on January 1, 2011
It is possible to translate a discourse into a set of propositions, concate nated into a connected hierarchical structure (Kintsch, 1 914; Kintsch and Van Dijk, 1 978; Rumelhart, 1 975), and some theorists have assumed that such structures correspond to the mental (memory)representation of the discourse. A somewhat different orientation would start out from the assumption that a discourse is generally organised around settings, arguments, or situations which are already known about to some extent by a person who can read and understand it. Looked at in this way, an important function of the early part of any discourse will be to enable a successful search for a referent situation in the memory of the reader. Sanford and Garrod ( 1 9 8 1 ) have termed such a referent situation a scenario.
THE MENTAL R EPRESENTATION OF DISCOURSE 'stop' or 'ball' in isolation. The origin of the continuity problem seems to be that the NP 'The score' fails to identify anything which has a place in our .knowledge of traffic control while it succeeds in identify.,. ing a necessary component of a game of cricket. This would suggest that any mental · representation arising from the first sentence must in some way incorporate information to the effect that the sentence is 'about' an event in a game of cricket. In other words the sentence functions as a partial description of some situation in which the event being referred to has significance, and the cohesion between the two sentences does not arise from the entities being mentioned in them selves but rather from the situation being described. Let us therefore describe such phrases as 'the score' in (2') as functioning as s i t u a t i ona l anaphors in that they carry back reference to elements which are a necessary part of the previously instantiated situation.
In fact there is a certain amount of evidence which seems to support the view that interpreting situational anaphors does not impose any great load on the processig system. This evidence comes from 'experi ments in which overall sentence comprehension time is used as a meas ure of processing difficulty for that sentence . . In 1 97 4 Haviland and Clark published a paper in which they demomstrated that comprehension time for a sentence containing an anaphoric noun phrase was in part a function of the contextual availability of its antecedent . They com pared contexts like (3) and (4) below for a target sentence such as (5): (3) Mary unpacked the picnic supplies. (4) Mary took some beer from the trunk. (5) The beer was .warm. using a range of such materials they were able to demonstrate that comprehension time for sentences such a� (5) was increased in the case where there was no directly stated antecedent in the context (as in (3)) as compared to cases where the antecedent was directly stated (e.g. (4)). More recently Garrod and Sanford ( 1 98 1 ; in press) have demonstrated JS, vol. l , no. l
23
Downloaded from jos.oxfordjournals.org by guest on January 1, 2011
What is particulary interesting about such situational anaphors is that they might give us a clue as to the nature and availability of the mental representation set up by any antecedent piece of discourse. For instance if it can be shown that interpreting sentences containing such anaphors imposes no greater load on the processing system than the interpretation of sentences containing more straightforward direct antecedent anaphors, then. it must be assumed that unstated information about the situation under discussion is as readily available in the reader or listener's mental representation as information arising directly from the stated discourse. In this way interpretation of anaphoric expressions constitutes a kind of naturalistic exercise in memory retrieval.
GARROD &: SANFORD that such increases in comprehension time do not necessarily occur in the absence of stated antecedents. For instance Garrod and Sanford (in press) showed that sentences containing appropriate situat ! o�al anaphors did not require any more processing than sentences contammg anaphors with directly stated antecedents. The appropriateness of the context was manipulated by using different titles to the passages that a subject would read. As an illustration consider the two passages below: In Court (6) Harry was being questioned (by a lawyer). (7) He had been accused of murder. (8) The la wyer was trying to prove his innocence. Telling a Lie (9) Harry was being questioned (by a lawyer). ( 1 0) He couldn't tell the truth. ( I I ) The lawyer was trying to prove his innocence.
Comparable results can be obtained when contrasting contexts like the following: ( 1 2) Keith drove to London last night. ( 13) Keith took his car to London last night. are followed by a sentence containing an anaphoric reference to a car, i.e. (14) The car kept breaking down.
24
JS, vol. l , no. !
Downloaded from jos.oxfordjournals.org by guest on January 1, 2011
In the passage 6 - 8 a title is used which should indicate· that the passage as a whole is about something happening in court and given such a title a high proportion of subjects expect the presence of a lawyer in this situation. Thus we might predict that a reference to the lawyer in sentence (8) should cause no problems whether or not an antecedent mention occurs in sentence. (6). On the other hand with the second similar passage entitled 'Telling a Lie' no such presuppo sition exists and hence we might expect sentence ( I I) to cause problems for the reader in the absence of an explicitly mentioned antecedent in sentence (9). In fact when the comprehension times were measured for the critical sentences in the two contexts it was found that in the appropriate context (e.g. with the title 'In Court') it made no difference whether the initial sentence contained an antecedent mention (e.g. the phrase 'a lawyer') or not. However, with the inappropriate context (e.g. with title ' Telling a Lie') a subsequent difference in reading time emerged for the critical sentence when no antecedent mention occurred in the text. In other words under certain conditions interpreting 'situational anaphors' imposes no extra load on the process ing system.
THE MENTAL REPRESENTATION OF DISCOURSE In circumstances where the verb severely restricts its instrument, as with drive and vehicle, situational anaphors which refer to this instrument do not seem to require any extra processing over the direct antecedent cases (see Garrod & Sanford, 1 98 1 for a more detailed discussion). Evidence of the sort cited above leads us to conclude that if some element is considered by a high proportion of readers in that community as a necessary component of the situation being portrayed in the prior discourse then it is possible to make direct reference to it in the subsequent discourse without producing any measurable effect in com prehension difficulty for the sentence containing the reference. Memory organisa tion and procedures
In psychology, there is a well-established distinction between two types of memory. The first is a dynamic system of limited capacity which 'holds information pertinent to whatever task is at hand, and which is readily availabl� to the processing system. It has been referred to as short-term working memory (e.g. Baddel�y and Hitch, 1 974). The second corresponds to a more usual use of ' memory' ,.. and is a relatively static store of effectively limitless capacity, in which resides our wealth of knowledge both specific and general. Such a distinction may be seen to have a relevance to discourse processing, in that as a discourse unfolds, we seem to be most aware of the current topic of the discourse and information relevant to it, rather than being equally aware of earlier parts of the discourse. In terms of anaphora, this means that references to entities which are part of the current aspect of a text should be more easily (or quickly) accessed than references to parts which do not correspond to the current topic of discussion. Indeed, experimental work suggests that this is the case (Sanford & Garrod, 1 9 8 1 ; Sanford, Henderson & Garrod, 1 980). Apart from a distinction of this kind, which we shall now call the current focus vs. static memory distinction, it is also necessary to accommodate representations of information not specifically mentioned in a discourse but directly relevant to it, such as the scenarios alluded to in the previous section. To do this, let us first distinguish between asserted information, which is actually given by the text itself, and conceptual information, which corJS; vol. l , no. !
25
Downloaded from jos.oxfordjournals.org by guest on January 1, 2011
Just as anaphoric reference serves as an important basis for text cohesion, so ease or difficulty of reference resolution provides a means for investigating the availability of structures in memory which result from reading discourse. In this section we shall briefly outline a system of memory organisation which reflects differences in availability and provides a framework for considering the more procedural aspects of discourse.
GARROD &: SANFORD responds to the scenario. Now either we could assume a unitary working memory, in which both types of information were intermingled , or we could assume that the two types of information were sufficiently different to correspond to different partitions of working memory. Certainly the two must be distinguished in some way. For instance, it is an easy matter to distinguish between 'Mary dressed the baby' and ' Mary put clothes on the baby ' . Although memory experiments have shown that people are confused to a degree about which of two statements of similar meaning occurred in a text of some length (e.g. Sachs, 1 967), in the short-term such confusions are not so common. So what is said and what is meant can be kept separate. However, another distinction can be made, and that is that a reference can be made to a dependent of an entity which is -explicit (asserted) more easily than it can to an entity whose existence depends upon interpreta tion; consider for instance the following set of sentences ( 1 5) Mary put the clothes on the baby. ( 1 5 ') The material was made of pink wool.
( 1 6) Mary dressed the baby. If ( 1 5) and ( 1 6) led to the same mental representation, then this would not be expected. It is interesting to note that a similar distinction to the one between asserted and conceptual information is also recognised by some of those modelling human memory, and is expressed as the difference between episodic and semantic memory (Tulving, 1 972). Thus while semantic memory is supposed to reflect general knowledge, dissociated from any specific situation in which it was acquired, episodic memory contains knowledge of particular episodes. In the present case, the asserted/conceptual and focused/static distinction yields four memory types, summarised in Table 1 . There are various distinctions now to be made between representa tions in explitit and implicit focus. First of all consider the kind of representations of entities which might be suitable for implicit focus, bearing in mind that it is nothing more than a currently accessible part of semantic memory. If the sentence being represented is Keith was driving down to London, then anaphoric probe experiments show that car is an 'available entity'. However, the implied 'car' is not a specific one on the present' account. It is simply a representation of the fact that part of the definition of drive in this sense is to travel by car. In fact, car in this representation can be thought of as a variable, which can take as a value any specific instance of a car, or even any 'vehicle-like entity'. Thi s is difficult to envisage from the point
_
26
JS, vol. l , no. I
Downloaded from jos.oxfordjournals.org by guest on January 1, 2011
In a reading-time study, ( 1 5' ) was read more rapidly than it was when the antecedent was changed to:
Nature of Memory
Origins of representation
Asserted
Conceptual
Focused
Static
Explicit Focus
Long term Text Memory
Implicit Focus
Long Term Semantic Memory
Table 1 A schematic characterisation of the four memory partitions
In complete contrast, explicit focus seems best represented as tokens for entities which are introduced into the discourse. For instance, introducing 'a car' or 'the car (de novo)' would set up a token for that entity. In this way, Keith drove to London and Keith went to London by car would have different structures, as illustrated in Figure 1 . The main point here is that although i t is possible to find a representa- . tion corresponding to car in both (a) and (b), in (a) it is a variable, , and in (b) it is a token. If it is useful to ·make these distinctions, as we have argued, then one would expect the information in the different partitions to be differentially addressed by various search directives. The force of the present paper is to suggest that such differential addressing finds its linguistic counterpart in the nature of referring expressions. In particular, we shall concentrate on the distinction between Explicit and Implicit focus, and shall begin with the contention that full definite noun phrases (FDNP) and pronouns can be viewed as triggers to imple ment searches of memory, and that they differ in the partitions which they address.
JS, vol. l , no. l
27
Downloaded from jos.oxfordjournals.org by guest on January 1, 2011
of view of declarative representations, but can be readily appreciated from a procedural point of view: the variable car can be looked at as a series of tests which might be applied to any definite noun-phrase in subsequent discourse. If the noun-phrase passes the tests, then i t will stand as an instantiating value for the variable. Such arguments form the centre of most schema-based explanations of comprehension (see, for instance, Norman &: Rumelhart, 197 5).
GARROD &. SANFORD Figure 1 (a) 'Keith drove to London' eith
s
EX PLICIT FOCUS Scenario 1
�
Role l
IMPL ICIT FOCUS Scenario 1 : Driving Role I : Driver : Destination Role 2 etc
� il1s
� Role 2 �
Action
ondon
Role I travels to Role 2 by< Car>
(b) 'Keith went to London by car' eith
''
Role
�� 1
� � �
Role 2
London
Role )
is
IMPL ICIT FOCUS Scenario 2: Human goes Somewhere Traveller Role I Destination Role 2 Mode of travel Role 3 etc Role I travels to Action Role 2 by means of Role 3
Car Anaphoric e:rpressions as processing directives Let us begin by defining a specification for any language string which is to serve as a memory search directive. Such a specification would comprise (a) the domain (s) of memory over which the search i s to take place, (b) the information available in the string which may be used to guide the search, and (c) the type of information being searched for , together with any restrictions on the type. So, if we wanted to define the procedure behind representing a F DN P as a search directive, then our task is to specify (a), (b) and (c). If our specifications are adequate, then all examples of usage in a language should be accommo dated, and a psychological test of the account should yield a positive result. Consider a personal pronoun in this light. One particularly striking thing about pronouns is that they appear strange when used to refer to an implicit antecedent. Thus, 07') is a natural continuation o f 28
JS, vol. l , no. l
Downloaded from jos.oxfordjournals.org by guest on January 1, 2011
t
EXPL ICIT FOCUS Scenario 2
THE MENTAL REPRESENTATION OF DISCOU RSE (17) while (17 ) is not: "
( 1 7) Mary dressed the baby. (17') The clothes were made of pink wool. (17") They were made of pink wool. At first sight, it may appear that one possible problem with 07") is that they could refer to Mary and the baby, although the pragmatics of the rest of the sentence would ultimately rule this out. H o wever, it seems equally odd to use a pronoun to refer to an implied entity of which there will only be one: ( 1 8) Mary won the first round of the mixed tennis championship. ( 1 8') *He was not a very good opponent. ( 1 8") The man was not a very good opponent.
In terms of specifying the retrieve procedure for the personal pro noun it therefore seems sensible to restrict the search domain to that of explicit focus, which would mean that the pronoun 'he' might trigger a procedure of the following kind: RETRIEVE
(a) DOMAIN: Explicit focus. ( b) PARTIAL DESCRIPTION: Singular, Male, ( Human) . (c) RETURN: Matching token identity in explicit focus. But is it possible to formulate a comparable procedure for handling the FDNP? It may be helpful to start by considering certain other contrasts between use of the pronoun and the F DNP. When considering the contrasts between pronouns and FDNPs with respect to situational antecedents we employed a simple substitution procedure and then asked the question, whether the two forms of expression were equivalent under substitution. This method can be extended to look at a number of other cases which would indicate that pronouns behave differently form FDN Ps as anaphoric devices and consideration of these cases suggests that under certain circum stances the two forms of expression take on quite different interpreta tions in the same context. For instance when the antecedent mention is interpreted generically an anaphoric pronoun may be used in circumstances where the equiva lent FDNP is ruled out ( see 19-22 below ) . JS, vol . l , no. l
29
Downloaded from jos.oxfordjournals.org by guest on January 1, 2011
Not only is 'he' unambiguous in this context, but it is also no less specific than the alternative 'the man' in 1 8" , so there is no ambiguity whatsoever, and no new information is being introduced. Considerations such as these lead to the hypothesis that pronouns can only be used anaphorically to refer to explicit representations - i.e. they are not suitable for situational anaphora.
GARROD &: SAN FORD
( 19) (20) (2 1 ) (22)
An animal needs oxygen. It cannot live without water either. *The animal cannot live without water either. An animal cannot live without water either.
In this case the pronoun seems to naturally substitute for the generic indefinite 'an animal' in (22) rather than the definite 'the animal' in (2 1). A somewhat different example of the failure of pronoun/FDNP substitution occurs when the FDNP serves to establish a generic inter pretation of something which has previously been introduced into the context as a specific referent. As with:
(23) Once upon a time there was a cat. (24) Now the cat is renowned to be the laziest of animals. it
Once upon a time there was a cat. Now, they are renowned to be the laziest of animals. but In this case the pronoun is probably cataphoric on the general expression animals at the end of the sentence. Examples such as these and the ones constdered in rela�on to pro nouns and situational anaphora all support the view that the pronoun is very much constrained in its interpretation by the original interpreta tion of the antecedent, yet may take on almost any such original interpretation. With the FDNP on the other hand the interpretation seems to depend more upon the nature of the noun phrase itself and its sentential context than the original interpretation of its potential antecedent. Thus when an antecedent receives a generic interpretation as w i th ( 19) then the pronoun takes on such a generic reading whereas the FDNP cannot in these cases. However when the antecedent is interpreted specifically as in (23) the pronoun seems to require a similar specific interpretation, yet the FDNP may take on a different generic reading if this is forced by the rest of the sentence in which it occurs. At the most general level the pronoun serves primarily as a device for maintaining previous references, while the FDNP may have an additional attributive function allowing it to establish meaning but within the constraints of the current domain of discourse. In this way its interpretation need not rely exclusively on that already established for the antecedent. The distinction between a purely reference mainte nance function versus an establishment of meaning function is in our view reflected in part in the distinction between searching explicit
30
JS, vol. l , no. l
Downloaded from jos.oxfordjournals.org by guest on January 1, 2011
Again it is of course possible for the pronoun to take on a generic reading as with:
THE MENTAL REPRESENTATION OF DISCOURSE focus which contains token representations and implicit focus which represents pragmatic information derived from the prior interpretation of the text, but largely only implied by the text itself.
A more detailed discussion of the nature of construct procedure is beyond the scope of the present paper (but see Sanford & Garrod, 1 98 1 ). However, what we intend to do in the remainder of the paper is to explore the more subtle stylistic distinction between pronouns and the· F DN Ps and see what light these might throw on the details of explicit focus representation. One of the conclusions which we will reach is that explicit focus may be thought of as the repository for structural information arising from both the text as a whole and the sentence under interpretation. Explicit text.
focus and information deriVing from the structure of the
Evidence for the importance of the reference maintenance function of personal pronouns emerges also from studies of the distribution of pronouns or other noun phrases in spontaneous speech. For instance in a recent paper Marslen-Wilson, Levy and Tyler ( 1 982) carried out a very detailed analysis of the circumstances of usage of personal pronouns, zero anaphors and fuller noun phrases in a task where the subject retold a simple comic book story which centred on two main characters. The principle analysis depended upon a breakdown of the story into an �ierarchical structure of distinct events embedded within episodes embedded within the story. This breakdown emerges very JS, vol . l , no. J
31
Downloaded from jos.oxfordjournals.org by guest on January 1, 2011
In terms of formulating an appropriate procedural characterisation for the F DNP this would suggest that it needs to recover a wider range of information from memory than the pronoun and so trigger a more general retrieve procedure operating over the whole focus domain, both explicit and implicit, with the aim of recovering informa tion not only about token representations but more importantly about the pragmatic restriction in implicit focus which may or may not be directly linked to the tokens. At the same time it must be possible for FDNPs to trigger additional procedures whose goal is to construct new elements in the representation on the basis of the information already retrieved. This is necessary both to account for the fact that F DN Ps when used as 'situational anaphors' must enable construction of information in explicit focus to represent the newly established referent, and may also be required when the F DNP is employed to establish a new interpretation of an already introduced referent as with examples like ( 24). The procedural characteristic of pronouns and FDNPs may therefore be distinguished in two main ways ( I) in terms of the differences in restriction of search domain in memory and (2) in terms of the additional construction procedures associated with the interpretation of the F DN P, which are not available for the pronoun.
GARROD &: SANFORD clearly from the nature of the story itself, and turned out to be an exceptionally good predictor of the choice of anaphoric device. For instance if the reference occurred within an utterance which related to the same story, episode and event as one containing t!le antecedent mention then a pronoun or zero anaphor was chosen on 46 occasions out of the 50 observed. Thus in the vast majority of cases when the pronoun was used, it occurred at the most embedded levels of the narrative and functioned to maintain reference within an action sequence. On the other hand in a context which only related to the same overall story as that containing the previous mention there were only 2 out of 8 occasions when the pronoun was used and the incidence of these two could be accounted for in terms of local structural constraints on the utterance in which they occurred.
There is also some recent experimental evidence (Purkiss, unpublished, see Sanford &: Garrod, 1 98 1 ) using the reading time procedure which has a direct bearing on the claim that pronouns . and F DN Ps serve rather different functions in discourse. The materials used in the study have the general form shown in Table 2. In the first sentence of each set, two entities are introduced (e.g. The engineer and the television set). During subsequent sentences, in the 'subject position' materials, reference is made only to the entity desig nated by the object noun phrase. When the final (target) sentence is reached, anaphoric reference is made to the subject position noun phrase in the first sentence. Furthermore, the referring expression can be either a repeat noun-phrase or a pronoun. For the 'object posi tion ' materials, this pattern is changed so that the intermediate sen tences are about the subject of the opening sentence, and the final sentence refers back to the object, again either by a FDNP or a pronoun. The second major variable in the design is · the number of sentences w!lich intervene between antecedent and anaphor. In one condition, the sentences marked with asterisks were included, and in another, excluded. The two major variables correspond to two factors which appear to influence the choice of a pronoun as an anaphor. The first, the 'topicalisation' principle, is exemplified by the apparent ease with which a logically ambiguous pronominal anaphoric reference can be used to refer to an entity introduced as a subject noun phrase in an active sentence - the following illustration is taken from Broadbent 32
JS, vol. I . no. !
Downloaded from jos.oxfordjournals.org by guest on January 1, 2011
This pattern of observation is consistent with the view that the speaker is employing pronouns and zero anaphors simply in order to maintain reference with a minimum of new interpretation, while lexical ly more specific expressions (mainly proper names in Marslen-Wilson et al. 's monologues) are reserved either to introduce characters at the beginning of a narrative or to re-establish them as central actors in a new episode.
THE MENTAL REPRESENT AT ION OF DISCO U RSE Key noWl-phrase in subject positon The engineer repaired the television set. It had been out of order for two weeks. *It was only a few months old. *It was the latest model. He/the engineer took only five minutes to repair it. (TARGET) Had the television set been out of order for five weeks? Key noWl-phrase in object position The mother picked up the baby. She had been ironing all afternoon. *She would not be finished for some time. *She was very tired. The baby/it had been crying nearly all day. (TARGET) Had the mother been sleeping all afternoon? Table 2 Sample of materials used in the pronominal reference study (Purkiss 1978)
(27) The feedpipe lubricates the chain, and it shou ld be adjusted to leave a gap half an inch between itself and the sprocket. Broadbent's study indicated that most people interpret it a s anaphor of the feedpipe and not the chain.
being an
The second major variable corresponds to the principle that pronouns are used to refer to things which have been mentioned recently i n a d i s:.. course. This is exemplified in linguistics by Chafe's ( 1 972) concept of foregroWlding. Subjects read through materials of this type in the self-paced re&ding situation described earlier, and in the analysis attention was paid to the time subjects spent inspecting the final {'target') sentences. The results are shown in Figure 2. Of particular interest here are the results for the object position materials, which clearly demonstrate an ' interac tion between number of intervening sentences and pronoun/FDNP reference form. ·
Provided there is only one intermediate sentence, sentences with a pronoun anaphor are actually read slightly more quickly than those using an FDNP. Indeed, this trend holds over both subject-position conditions. However, when the number of intermediate sentences is larger, the trend is reversed for object-position antecedents: F DNPs are apparently handled more rapidly than pronouns.
JS, vol . l , no. l
33
Downloaded from jos.oxfordjournals.org by guest on January 1, 2011
( 1 97 3).:
GARROD & SANFORD Figure 2 TOPIC
1.7 -; 0 •
� w
:I ;:
i
c w a::
COMMENT
, I
I
/
1.5
/
/ �
1.3
1.1
��
� ,�
1
3
NUMBER OF INTERVENING SENTENCES
3
Mean reading time for target sentences with sentence F D N Ps ; length difference neutralised. So l i d l i n e s : dotted lines: pronoWlS.
A further experiment, the first in a series on plural pronouns (San ford & Garrod, in preparation), supports the dissociation of function viewpoint. Typical mC�.terials are shown in Table 3. It was a fine Saturday morning. John and Mary went into town. She/they/Mary wanted some new clothes. (b) The library was quite fulJ. Linda and Jim could not sit down anywhere. The librarian told hirn/them/Jim to wait.
(a)
Table 3 Sample of materials used in the plural antecedent experiment. 34
JS, vol. l , no. J
Downloaded from jos.oxfordjournals.org by guest on January 1, 2011
These results seem to show that pronominal mappings are most readily made when the antecedent is in the subject position and/or recent. Now it is possible to argue that when a noun phrase is in the subject position, or has been mentioned recently, then it has a 'stronger' representation in explicit focus, and is more readily available as a potential antecedent. However, such an argument cannot explain how pronouns are handled more rapidly than FDNPs in one condition and more slowly in another. To accommodate this, Sanford & Garrod ( 1 98 1 ) suggested that a pronoun m ay search only explicit focus, and that object-position F DN Ps are not represented in explicit focus after a number of intervening sentences, while subject-position ones are. However, as we will see below, this simple strength of representation explanation will need to be modified somewhat if we are to account for certain additi9nal phenomena.
THE MENTAL REPRESENTATION OF DISCOU RSE In this study, a plural topic FDNP in the second sentence is referred to by an anaphor in the third. The pronoun could either be plural, b e i n g codesignative with the entire topic FDNP, singular, being codesignative with one element of the FDN P, or a singular FDNP, being codesignative with the same element of the topic . · A further contrast in conditions is indicated in the materials: the anaphor could be either in the topic or in the comment position of the sentence in which it appears. Now there are many good reasons to suppose that a plural FDNP would be represented as a group in explicit focus, and so would be more readily mapped to a plural pronoun than to a singular one. How ever, Figure 3 shows the situation to be more complex in fact. Figure 3 2.2 'U 2.1 •
!
a:
1.7
�
NAll£
PRO
PLUPRO
Mean reading time for target sentences. Square points: anaphors in the object position of the target' sentence; round points: imaphors in the subject posi tion. Points are joined by lines. {or clarity only. In the figure are shown the reading times for the target sentences containing the various forms of anaphoric expression. The top curve refers to the reading times for those target sentences in which the expression is in syntactic object position while the bottom curve shows the times for the subject position sentences. Of interest here is the interaction between form of expression and its syntactic position in the sentence. For instance, for the plural pronouns it makes no signifi cant difference whether they appear as subject or object of the target sentence but in the case of the singular pronoun there is a considerable and reliable difference in reading time associated with the positioning. It seems that the singular pronoun is only effective in identifying one of the members of the plural antecedent when it occurs in subject position, while the plural pronoun works very well, and the lexically more specific name works moderately well in either position of the sentence. How can we account for this rather extraordinary finding? It may be helpful at this point to consider in a little more detail JS, vol. l , no. l
35
Downloaded from jos.oxfordjournals.org by guest on January 1, 2011
"' 2.0 :I � f:l 1.11 z E c 1.8 "'
GARROD & SANFORD some of the work on usage of pronouns occurring in different syntactic positions of the sentence. In a recent paper Karmiloff-Smith ( I 980) observed that older children always reserved sentence initial pronouns for reference to what she called the thematic subject of the narrative that they were telling where thematic subject corresponded to the central actor in the story. This led her to conclude that pronouns in subject position could possibly be thought of as default expressions for maintenance of thematic subject. A similar conclusion is of course suggested in Marslen-Wilson et al. ( 1 982) observations in which pronouns were employed to maintain the characters in central roles within an already established action sequence whereas more specific lexical items such as proper names were reserved for re-establishing the antece dent in the central role. Given these observations it would seem appro priate to make some additional assumptions about the mechanisms of pronoun resolution and the sort of antecedent information which must be available in the focus memory system.
RETRIEVE (a) DOMAIN: Explicit focus. (b) PARTIAL DESCRIPTION: Male, Singular, H uman (Subject). (c) RETU RN: Matching token identity from the set defined by the variable thematic subject. F rom a processing point of view having retrieve procedures of this sort allows the system to capitalise on the fact that the subject of a sentence in a narrative is usually taken to refer to the thematic subject of the discourse. Thus when a purely reference maintenance device, such as a pronoun, is encountered it is assumf!d that the identity 36
JS, vol. l , no. !
Downloaded from jos.oxfordjournals.org by guest on January 1, 2011
In the first place we might speculate that in addition to the tokens held in explicit focus there might also be certain types of structural information derived from the prior discourse which is also represented. Thus we could find information about the identity of the current thema tic subject or subjects. Since such information can be derived directly from the structure of the previous text it does not seem unreasonable to assume that it is represented in the explicit part of the text represen tation. Furthermore we must assume that the retrieve procedures triggered by pronouns in different syntactic positions within a sentence may be augmented with structural information which would allow them to identify structural variables in explicit focus. Given such assumptions it is possible to differentiate between the kind of search procedure associated with a sentence initial subject pronoun and a pronoun encoun tered somewhere in the rest of the sentence, which is clearly necessary if we are to accommodate Karmiloff-Smith's observation and the results of the reading time experiment reported above. So Jet us assume for the moment that a search procedure triggered by a sentence ir�itial pronoun such as 'he' might be described in term s of the following specifications:
THE MENTAl REPRESEN TATION OF DISCOURSE of the thematic subject is being maintained. The corollary to this is that when a fuller noun phrase is encountered which is primarily not a reference maintenance device but one for establishing reference it would be assumed that a new thematic subject is being established. As a result of this we would expect to find that sentence initial pro nouns were particularly effective in just those circumstances when they identified referents from among the set of things which could be considered as thematic subjects.
let us now return to the specificat ion of the retrieve procedures for the two types of pronoun encountered in the target sentence. When the pronouns were in subject position (See Table ( 3a)) they would both trigger searches for current thematic subject and come up with the two tokens currently allocated. In this way either search would succeed in recovering a matching antecedent. On tl'le other hand when the pronouns were not in subject position the retrieve procedure would not have access to the content of the thematic subject variable since the search would not be directed to this set. The retrieve for the plural pronouns would therefore succeed in finding the token assigned to the group agent whereas with the singular pronoun there would appear to be no syntactically matching singular token readily available, and further time consuming search operations would be needed in order to recover a matching antecedent. Finally, since we have not assumed that structural information should have any effect on the process of interpreting proper names or FDNPs, use of these expressions should not lead to any reading time effects associated with their syntac tic position in the sentence. Clearly this explanation is would need to carry out a results do suggest that we between the types of search noun according to both the that of the pronoun itself.
post hoc and in order to verify it we number of control experiments, yet the need to be able to somehow distinguish procedures employed for resolving a pro structural context of its antecedent and
Finally, it is worth mentioning that the sort of structural augmentaJS, voi. J , no. l
37
Downloaded from jos.oxfordjournals.org by guest on January 1, 2011
This might allow us to put forward a tentative explanation for the rather extraordinary reading time result observed in the experiment reported above. If we first consider how the sentences containing the antecedents might be represented in focus, two types of information would result; first, information relating to the discourse structure to the effect that the two actors (e.g. John and Mary in Table (3a) or linda and Jim in Table (3b)) were both potential thematic subjects, in which case the thematic subject variable would be assigned the two tokens and secondly, that the two actors as a group serve as joint agents in the action being portrayed in the sentence in which they occur, so the tokens as a group would also be mapped into some variable in implicit focus.
GARROD & SANFORD tion of the retrieve procedure which we have been entertaining for pronouns in sentence initial position in discourse may also have its counterpart in the sentence domain itself. It has often been pointed out that pronouns in the subject of a co-ordinate or subordinate clause tend to pick the sentence subject as antecedent. For instance in a sentence like the following:
(25) John hit Bill and then he ran away. the pronoun is usually interpreted as codesignative with John, whereas in a sentence such as (26)
(26) John hit Bill and then Mary shouted at him.
So let us now attempt to summarise the arguments. In the first plac·e we have suggested that it is helpful to characterise the interpreta tion of anaphoric expressions in terms of search procedures which operates on restricted domains corresponding to distinct focus partitions in working memory. On the basis of both previous work in. the general field of cognitive psychology and clarity in describing the representation we have proposed a partition in the focus memory system into two components: one, explicit focus, functioning as a repository for asserted information which may consist in both token representations correspond ing to the various individuals mentioned and structural information arising from the discourse itself; the other, which we termed i m p l i c i t focus, serving to represent i mplicit knowledge-derived information needed to give the discourse significance. Given this distinction in the memory system it was then suggested pronoun resolution could best be described in terms of the execution of retrieval procedures restricted to the explicit partition whereas interpretation of full definite noun phrases could be described as in part arising from the execution of retrieve procedures operating . on both partitions and, in addition, construct procedures. In this way anaphoric pronouns derive their interpretation directly from the previous discourse, whereas F DNPs are interpreted within the broader constraints of our interpretation of the previous discourse and the particular sen tence in which they occur. This reflects the basic distinction between the reference maintenance function of the pronoun and the reference establishment function of the fuller noun phrase. 38
JS, vo!. l , no l
Downloaded from jos.oxfordjournals.org by guest on January 1, 2011
the pronoun seems to select the most recently mentioned antecedent (i.e. Bill). Examples such as these led Caramazza & Gupta ( 1 979) to put forward what they term the parallel function hypothesis for pronoun resolution. Yet again we might assume that these preferences emerge from a restriction on the retrieve procedure triggered by subject posi tion pronouns but in this case within the more local domain of the current sentence.
THE MENTAL R EPRESENTATION OF DISCOURSE General Discussion We have tried to show how a psychologically-based account of reference resolution might be constructed. In natural language, a wide variety of referring expressions are used in a wide variety of circumstances. Even the small number of examples considered in the present paper attests to this variety. The psychological approach to the problem aims to provide a description of the admissible possibilities within the constraints of mental operations. These constraints are presumed to operate in all situations demanding language comprehension, and are presumed to originate in limitations on man as a processor of symbolic information.
One major outcome of the approach is to emphasize the procedural aspects of discourse fragments. Thus it is not sufficient to say that a pronoun and an antecedent noun phrase refer to the same thing in a given case. What is required is a description of the mental opera tions which that pronoun brings about, and which uitimately results in the establishment of coreference. In the present paper we have indicated some of the problems which one encounters in trying to do this, while at the same time hopefully suggesting some solutions. The procedural approach is not merely attractive because it lends itself naturally to computer implementation, but also because it bestows other advantages. For instance, although two different forms of refer ence might be used in a particular situation, the final representation (meaning and significance) may be exactly the same; however, the process leading to that representation would be quite different. An example of this might be the way in which both definite and indefinite FNPs can be interpreted generically. JS, voi. J , no. l
39
Downloaded from jos.oxfordjournals.org by guest on January 1, 2011
The main constraints considered here are those characterising the structure of memory. Almost without exception, psychologists distin guish between working memory (a system oi lim ited capacity, sharing short-term storage and data-manipulation duties) and long-term memo ry. In the first part of the paper, we described how such a distinction gives rise to a major criterion for separating aspects of memory used in text comprehension. Not only psychologists, but also workers in artificial intelligence have found a similar distinction to be useful. Thus Grosz ( 1 977) distinguishes between 'focused' and 'unfocused' information in her discussion of understanding systems which might be implemented in computers. In this case, the utility of the distinction is computational: if references are made to entities, then it is necessa ry to restrict the search domain to manageable proportions. There is also another good reason for supposing that the reference domain must be limited, and that is that many sentences in language are elliptical. It is only possible to use ellipsis if the range of possible referents. is very limited. Indeed, a fully psychological approach would maintain that the possibility of ellipsis only arises because of the constraints which characterise human working memory.
GARROD &: SANFORD Finally, although the discussion here has been restricted in the main to pronouns, the framework put forward is well suited to the N Ps, analysis of other referring expressions, such as indefinite restricted relative clauses, and guantifiers in general. University of Glasgow Adam Smith Building Glasgow, Scotland
References
40
J S , vol. 1 , no. 1
Downloaded from jos.oxfordjournals.org by guest on January 1, 2011
Baddeley, A. &: Hitch, G., 1 974: Working memory. In: G.H. Bower (Ed.), The Psychology of Learning and Motivation. 8; 6 6 7 - 6 7 9 . Broadbent, D.E., 1 973: In Defence o f Empirical Psychology, M e t h u e n , London. Caramazza, A. &: Gupta, S., 1 979: The roles of topicalisation, parallel function and verb semantics in the interpretation of pro nouns. Linguistics, 17; 497-5 1 8. Chafe, W ., 1 972: Discourse structure and human knowledge. In: J .B. Carrol and R.O. Freedle (Eds.), L a n guage Comprehension and Acquisition of Knowledge, W inston, Washington. Pp. 4 1 -70. Garrod, S. &: Sanford, A.J., 198 1 : Bridging inferences and the extended domain of reference. In: J. Long &: A. B a d d eley ( Eds.) H i l l s d a le, N .J. A t tention and Performance IX. L . E . A . , Pp. 3 3 1 -346. Garrod, S. & Sanford, A.J., in press: Topic dependent effects i n language understanding. G.B. Flores d'Arcais, R. Jarvella (Eds.), The Processes of Language Understanding, J . Wiley &: S o n s , Chichester. Grosz, B., 1 977: The · representation and use of focus in dialogue under standing. Technical note 15, SRI International Artificial Intelligence Center. Halliday, M.A.K., 1 967: Notes on transitivity and theme in English, Part 1. Journal of L inguistics 3; 37-8 1 . Haviland, S.E. &: Clark, H.H., 1 974: What 's new? Acquiring new information as a process in comprehension. J o u r n a l of V e r ba l Learning and Verbal Behavior, 13; 5 1 2-52 1 . Karmiloff-Smith, A., 1 980: Psychological processes underlying pro n o m i n a l i sation and non-pronominalisation in children's connected discourse. In: J . Kreiman &: A.E. Ojeda (Eds.), Papers fro'!! the Parasession on Pronouns and Anaphora. Chicago Linguistic Society, Chicago. Pp. 23 1 -250. Kintsch, W., 1 974: The Representation of Meaning in Memory, E r l b a u m , Potomac. Kintsch, W ., &: van Dijk, T.A., 1 978: Toward a model of text comprehen sion and production. Psychological Review, 85; 3 6 3 - 3 9 4 .
THE MENTAL R EPRESENTATION OF DISCOU RSE
JS, v�l. l , no. l
41
Downloaded from jos.oxfordjournals.org by guest on January 1, 2011
Marslen-Wilson, W., Levy, E. &: Tyler, L.K., 1 982: Producing inter pretable discourse: the establishment and maintenance of reference. In: R. Jarvella & W. Klein (Eds.), S p e e c h , Place and Action; Studies in Dei:ris and Related Topics. J. Wiley &: Sons, Chichester. Pp. 339-378. Norman, D., Rumelhart, D.E. &: L.N.R., 1 97 5: Explorations in Cognition, Freeman, San Francisco. Rumelhart, D. E., 1 97 5: Notes on a schema for stories. In: D.G. Bobrow & A. Collins (Eds.), Representing and Understanding Studies in Cognitive Science. Academic Press, New York. Pp.2 1 1 -236. Sachs, J.D.S., 1 967: Recognition memory for syntactic and semantic aspects of connected discourse. Perception & Psychophysics, 2, 437-442. Sanford, A.J. &: Garrod, S., 1 9 8 1 : Und e rs tanding Wri t ten Language; Explorations in Comprehension Beyond the Sentence. J. W iley &: Sons, Chichester. Sanford, A.J., Henderson, R. &: Garrod, S., 1 980: Scenario-shift as a variable in text cohesion. U npublished Report, University of Glasgow. Tulving, E.A., 1972: Episodic and semantic memory. In: E. Tulving & W. Donaldson (Eds.), Organization of Memory, A c a d e m i c Press, New York. Pp. 3 8 1 -403.
INDEXED PREDICATE CALCULUS
S.- Y. Kuroda
Abstract A programme to construct an extension of predicate calculus is pro posed in which predicates and constants are indexed and interpreted with respect to different (mini-)worlds reffered to by . indices. From another perspective the proposed system is an extension of the idea of indexing noun phrases in syntactic representations in generative grammar. Some applications are given. In particular, it is applied to the description of ambiguities in intensional contexts, and a com parison is made with a description recently given by Saarinen.
This view is in a certain obvious sense unrealistic for explicating actual language use in discourse or in conversation as a cognitive activity. In everyday situations, we deal only with a very small chunk of the whole real world. Even if we talk about world politics, our background understanding of the world is fragmented and minuscule. Thus, we might say, at least as a first idealization, that we use the same uninterpreted language (logical representations) in different occasions of discourse and conversation, i.e. a formal system with the same meaning postulates and axioms etc. on different occasions of discourse, but models with respect to which such a formal system is interpreted vary from an occasion of use to another. But, however more realistic this view may sound, this change of view is, one might say, really immaterial, so far as generalities of semantic theory are concerned. For, even though one talks as if logical representations are interpreted with respect to the real world (and JS, vol. I, no. I
43
Downloaded from jos.oxfordjournals.org by guest on January 1, 2011
When the conception of logic is applied to natural language, it seems generally to have been tacitly assumed that logical formulae are inter preted with respect to the 'real ' world; that is, the whole real world is the model w.r.t. which each predicate is evaluated unless it is within intensional contexts. The recent influence of Montague grammar might f urther enhance such a perspective. In fact, each predicate is ihterpre ted not just by the whole real world, but with reference to all possible worlds. Each use of a sentence in discourse is, so to speak, backed up with all possible worlds, if we take what Montague grammarians say literally. 1
S. -Y. KURODA all possible worlds) in general semantic theory, the theory is not made dependent on any particular specific properties of the real 'real world'. The reference made by the theoretical semanticians to the real world would be much like the reference made by logicians to an unspecified set as a model when they formulate general truth conditions. When formal logic is applied on specific occasions, it is applied to various specific sets; but such variability is of no concern to a general theory of logic. L ikewise, reference to the real world and all possible worlds in general semantics, one might say, is only a far;on de parler, a n d w e don't have to be concerned with these concepts in a general theory.
H aving formulated a problem for the theory of discourse, however, will not be engaged in empirical analyses of discourse in a proper sense, in this paper. My present concern is rather this: once one sees the need for such a multiple model discourse theory, the door is open to explore extreme consequences of such an approach. For, one might ask, how small can discourse be, or how small must we assume discourse can be? There cannot really be any natural lower bound of the number of sentences in discourse, just as, conversely, there cannot be any natural upper bound for the length of a sentence. This observation might suggest the following: given the kind of problem we are concern ed with, we cannot separate the sentence from the discourse; i.e. we may not be able' to shield sentence semantics in a fixed-model framework, while we develop a multiple-model discourse theory. Then, we may be. thrown back on the general theory of sentence semantics. W ith this background, I would like to illustrate the usefulness of 44
JS, vol. l , no. !
Downloaded from jos.oxfordjournals.org by guest on January 1, 2011
This comparison, however, is not appropriate in at least two respects. For one thing, the assumption that the fixed world is the' real world has customarily been taken as entailing a specific type of ontological presupposition. Only very recently has this point begun to be subjected to serious criticism. (cf. Saarinen, 1 978; 1 98 1 .) Let me put aside this point for now, however. We shall for the moment be concerned with the other point. Natural discourse or conversation does not proceed like a formal application of logic to mathematics; in natural discourse, as speaker and hearer, we are constantly adjusting ourselves to changing 'contexts', shifts of 'topics ', from one moment to another. Our daily language activity is not made of a sequence of separate chunks of discourse, each with a well-defined universe of topics that can be simulated by a 'mode l ' . If such were the case, in order to simulate natural discourse, we would have to deal with 'model changes' during a stretch of discourse, and a f ixed-model theory of semantics would at least have to be supplemented by a multiple-model discourse theory ' that can deal with the interaction of contexts, or, formally, of models. This means that the theory of discourse in natural language would involve a nontrivial aspect of discourse structure, an aspect which a theory dealing with discourse in formal logic in the standard sense (say, proof theory) does not have to be concerned with.
INDEXED PREDICATE CALCULUS a multiple-model approach in sentence semantics, without motivating it on the study of discourse structures in a proper sense. I shall first present rather simple-minded examples and then later move on to an attempt to relate this approach to descriptive problems of intension al contexts of a familiar type.
•
I agree with Fauconnier that the understanding of sentences involves. mental constructs in terms of which the semantic function of the sentences, at least so far as their extensional aspects are concerned, are to be accounted for. However, Fauconnier's emphatic advocacy of a processing-oriented approach, I believe, is misguided. In my view, processing presupposes structure. According to this view, one might expect that a new approach advocated in the name of 'processing approach' would be significant and interesting just to the extent that it rightly leads to recognize structures of an as yet unrevealed charac ter. If that should be the case one might then investigate all the impli cations of this revelation and endeavour to obtain an adequate structur al account. The initial proposal for a processing approach could then be re-evaluated as a proposal for a processing model. Here is not the place for me to present details of Fauconnier 's approach and direct ly comment on it. However, I will freely borrow examples from his paper. It might also be stated at the outset that the present paper is intend ed only to set out a programme, not as a summary of an accomplished work. In particular, this point may need to be emphasized if the propoJS, vol. l , no. l
4.5
Downloaded from jos.oxfordjournals.org by guest on January 1, 2011
Before proceeding, however, let me just insert a remark on the background of the origin of this paper. Initially, this work was under taken as a sort of respons to G illes Fauconnier 's paper 'Mental spaces In a certain - a discourse-processing approach to natural logic'. 2 perspective we share the same concern. We are both challenging the view apparently prevalent among contemporary philosophers of language and linguistic semanticians under their influence. According to this view, sentences are associated with formal representations of a logical system of some sort such that the meaning of each sentence, at least to the extent that meaning in the extensional sense is concerned (and in recent times even more strongly, as a matter of intensional meaning) is accounted for in terms of referential interpretation (in the sense of formal logic) of the formal system with respect to the whole real world (or in an even more grand scheme, with respect to the class of all possible worlds), which is somehow metaphysically set transcen dentally and absolutely, independent of cognitive structures and psychcr logical processes. Counteracting this standard view, in the above men tioned unpublished paper, Fauconnier advocates a p�rspective in which "there will be no such thing as an abstract discourse-independent seman tic representation, or logical form for a sentence: rather, the sentence is a set of instructions for setting up and referring to the mental constructs which supports the organization of discourse." (p .5)
5.-Y. KU RODA sal is taken to be one for a system of formal logic. No semantic rule in the strict sense of formal logic is formulated. Only plausible indica tions are given to suggest how the proposed system might be formalized as an extension of standard (non-modal and modal) predicate calculus. Let me first consider the following sentence: ( 1 ) Since it was so stuffy in the house, Mary went up to the attic and opened the window.
In contrast ; in indexed predicate calculus (!PC), the relevant aspects of ( l ) can be represented, essentially, as (2)
W ENT-UP
i
(Mary, the attic) & OPEN
j
( Mary, the window)
where i refers to a m ini-subworld around the house and j to a mini subworld of this subworld, say, the attic. J Formally, indexed predicate calculus is an extension of predicate calculus. In addition to the usual vocabulary of predicate calculus, it contains the set I of indices. Each predicate symbol is indexed by an element of I. Instead of having a one-place predicate BLUE, for example, we have an array of indexed one-place predicate symbols, BLUE ; , BLUE ; , Semantically, a model to interpret !PC is a class of 'worlds' , or 'mini-worlds' , W ; , identified by elements of another index set J. !PC resembles possible world semantics in this multi world feature. But the intended function of this multi-world feature in !PC is quite different from that in possible world semantics. A valuation (an interpretation) of !PC determines a function k from I to J . In other words, an index i refers to a world W k ( i ) . •.•
46
JS, vol. l , no. l
Downloaded from jos.oxfordjournals.org by guest on January 1, 2011
If one applies ordinary predicate calculus to definite descriptions in natural language, it may be implicitly understood that the universe of discourse, i .e., the model with which logical expressions are assigned a valuation, is appropriately delimited. The president of France may be referred to as "the president of France" without delimiting the real world as a model for an interpretation of predicate calculus, but this is not generally the case with the use of singular definite nouns. The phrase "the attic" in ( 1 ) is understood to be the attic of a particular house in the context of the discourse in which ( l ) is inter preted with an appropriate mini-subworld of the real world, in which the house referred to by the house in ( I ) is the one and only house to refer to. But this proviso cannot save the direct application of the theory of definite descriptions to ( 1 ). For, the definite descrip tion the window here is most likely understood to be the one and only window of the attic, and obviously not of the house. The house most likely has more windows than just the one in the attic. Hence, we would have to paraphrase the window in ( 1 ) as the window of the attic, syntactically restoring a more abstract represe·ntation.
INDEXED PREDICATE CALCULUS What are 'worlds' ? In the simplest cases, they may simply be consid ered sets. Then, a predicate symbol indexed with i is interpreted in to which i refers; the familiar way with respect to the set W k { i ) under if, for example, P is a one-place predicate, the value of P i We later need a given interpretation is a subset of the set W k { i ) a more elaborate account when we use indexed predicates in intensional contexts and let them be evaluated with respect to a 'world' in our sense. But for the moment, this simple suggestion would suffice for the discussion of the next few examples. •
As another example, consider the following sentence, adapted from Jackendoff ( 1 975), through Fauconnier ( 1 979): (3) A girl with blue eyes has brown eyes.
(4) (Ex) (GIRL(x) & BLUE(x) & BROWN(x)). But (3) is 'factually' contradictory insomuch as no one can have blue eyes and brown eyes simultaneously. In IPC (3) may be rendered as (5) (Ex) (GIRLi (x) & BLUEi (x) & BROWN j (x)) where i and j are intended to be interpreted as referring to different worlds, (the real world, or a mini-subworld of the real world) and the world of the picture in question, respectively. An existential quanti fier binds three occurrences of variable x, two of which serve as argu ments of a one-place predicate indexed by i and the remaining one as the argument of a one-place predicate indexed by j. A natural convention for interpreting such an instance of the existential quantifier would be to assume that the domain of the variable bound by the But i n the ordina quantifier is the 'intersection' of W k{ i ) and W k { j ) As and W k { j ) • ry sense, there is no intersection between W k { i ) in certain versions of modal logic, however, · we assume that cross world identification functions are defined among worlds, without specify ing any metaphysical nature of such functions. Then, ( 5) asserts the existence of an individual who is a girl with blue eyes in the real world (in the relevant mini-subworld) who has an image in the world of the picture with brown eyes. •
Let me formulate the convention indicated above (and an obvious dual of it) for the convenience of later reference: JS, vol. l , no. l
47
Downloaded from jos.oxfordjournals.org by guest on January 1, 2011
In a context in which the speaker is looking at a picture, this sentence may be understood to mean that some girl with blue eyes in the real world is painted as a girl with brown eyes in the picture. Let BLUE and BROWN be one-place predicates to which we assign the meanings 'having blue eyes' and 'having brown eyes' , respectively. If (3) is trans lated directly into ordinary predicate calculus, using these predicates, one would get a form something like
S.-Y. KU RODA (C- 1 ) If an individual variable x occupies a pos1t1on in predicates P i 's, where i ranges over a subset I' of I, and x i s bound by the existen tial (universal) quantifier, the domain of the variable x is the intersection (the union) of W k(i ) , i ranging over I ' . ' Intersection' and 'union' must be determined relative to cross-world identification functions. 4 I will now borrow another example from Fauconnier to indicate that individual constants, in addition to predicates, may be indexed� Consider a movie in which Caesar played by Richard Burton seduces Cleopatra played by Elisabeth Taylor. One might say: (6) In this movie, Richard Burton seduces Cleopatra. Let us represent Richard Burton and Cleopatra by b and c, respectively, and consider the formula where they indicate SEDUCE and the two constants b and c are freely indexed:
Now, the natural convention would be that if a constant a i indexed with j fills a place of a predicate indexed with i, the referent belongs to the 'intersection' of the worlds indicated by i and j. In our example, the seduction takes place in the world of the movie; i indicates the world of movie. Richard Burton is a name for a person in this real 5 As Richard Burton has a role in world; j indicates this real world. the movie, the referent of b i may be found in the 'intersection' of the world of movie and this real world. Likewise, the name Cleopatra belongs . to a historical world, and hence t�e index k of ck indicates this historical world. But since Cleopatra is represented in the movie, the referent of c k may be thought as belonging to the 'intersection' of the world of movie and the historical world. Altogether, then, the above formula is interpretable as true in the desired way. In contrast, if we interpret index i as indicating the present real world (thus, the same as j), the formula is not interpretable 6 , because Cleopatra does not belong to the 'intersection' of the present real world and the historical world. Consider, next, the following Japanese sentence: (8) subete no kyoozyu-ga gakusei-o minna rakudai-saseta. all professor-SUBJ student-OBJ all flunked Given an appropriate context, this sentence may be ambiguous. One reading may be translated as (9) All the professors flunked all the students. 48
JS, vol. l , no. l
Downloaded from jos.oxfordjournals.org by guest on January 1, 2011
(7) SEDUCEi (b i , c k ).
INDEXED PREDICATE CALCULUS and the other as ( 1 0) Every professor f lunked all of his/her students. If one wants to give a sem antic account of this ambiguity of (8) within the ordinary framework of logical representation, one would have to set up two abstract logical representations for (8), essentially encoding the structure of (9) and ( 1 0), respectively. In IPC the ambigui ty is essentially captured as a matter of scope difference, a familiar situation. To see this let an index i refer to a mini-world of the institu tion in question. The reading (9) may be formalized as: (1 I)
(yY) ( S i (y) ::::> (Yx) (P i (x) ::::> Fi (x,y)))
where S, P, and F stand for 'student', 'professor' , and 'flunk', respective ly. Next, consider the reading ( 1 0). For each professor x, we need to specify all the students in a mini-world of this professor. Assume that the index j refers to this mini-world. Then, we might have:
depends to represent this reading. But note that the mini-world W k ( " ) on each professor x. In order to accommodate this dep�ndency, we allow a quantifier to bind a variable at an index position. Instead of ( 1 2), we introduce: (y) :::> lj (x,y ))) ( 1 3) (Vx) (Pi (x) ::::> (Vy) (S f(x) where the intended interpretation of f(x) is the mini-world of each professor x. If we use the notation of restricted quantifiers, ( I I ) and ( 1 3) are replaced by: 0 4)
�Y>s. (Yx\>.
1
1
( 1 5) (� )
P
i
5
F (x,y) i
f(x)
F.
1
(x,y).
Even though the two quantifiers involved in ( 1 5) are both universal, the order of the two quantifiers is relevant, because the first quantifier binds the index of the predicate restricting the second quantifier. In contrast, the order of the two occurrences of the universal quantifier in ( 1 4) is i r relevant. ·
The syntax of indexed predicate calculus, then, must contain one variable function symbols, like f in the above example. The variable position is filled by an individual variable, like x in the above example. JS, vol. l , no. l
49
Downloaded from jos.oxfordjournals.org by guest on January 1, 2011
( 1 2) (\fx) (Pi (x) ::::> ( yy) (Sj (y) :::> F i(x,y)))
S.- Y. KURODA The interpretation given to f(x) is a function from the domain of the variable x (in the above example, W k ( i ) ) to the index set I; that is, for each value of x, f(x) determines a world W k ( f { x ) )" The syntax must also stipulate that if a predicate is in the scope of a quantifier, it may be indexed by a functional index bound by the quantifier. Let me now turn to sentences with an intensional context of a familiar type. Consider: ( 1 6)
Magnus believes that a witch blighted the mare.
Tne familiar technique of scope is customarily used to distinguish different readings of ( 16) in the usual formal logic representation. Thus ,compare ( 1 7)
( !x) (WITCH (x) & (BELIEVE (MAGNUS, BLIGHT (x, the mare)))
( 1 8)
(BELIEVE (MAGNUS, (]x} (WITCH (x) & BLIGHT (x, the mare)))
( 1 9)
(! x) ( WITCH. (x) & BELIEVE. (Magnus, BLIGHT. (x, the mare)))
(20)
BELIEVE. (Magnus, ( h) ( WITCH . (x) & BLIGHT. (x, the mare))).
1
1
l
1
l
l
In (20), the existential quantifier binds two occurrences of variable x, which are both in predicates indexed with j. Hence the domain of the variable is the world indicated by index j. The semantic rule associated with the predicate BELIEVE must stipulate that it is the belief world of the subject of the verb believe, in this case, Magnus, o r perhaps a subworld within it. In contrast, i n ( 1 9) the existential quanti fier binds two occurrences of variable x, one of which is in a predicate indexed with i and the other in a predicate indexed with j. The world indicated by i is pragmatically chosen; an unmarked choice is the real world - or a mini-subworld of it - conceived by the speaker. The domain of variable x is then the 'intersection' of the worlds indicated by indices i and j, i.e., those individuals in the real world who also exist in Magnus' belief world. Before proceeding further let me insert a remark on the formal semantics of indexed predicates. We say that index j indicates Magnus' 'belief world' and predicates indexed with j are interpreted with respect to this world. But in standard logic predicates in intensional contexts are not given a semantic interpretation simply by a set, but by a class of sets, or 'possible wor Ids' , in the technical sense of modal lo�ic. 50
JS, vol. l , no. I
Downloaded from jos.oxfordjournals.org by guest on January 1, 2011
( 1 7) and ( 1 8) are usually taken to represent the so-called specific and the nonspecific readings of ( 1 6), respectively. There is no novelty in indexed logic for the representation of these readings. Different indices, i and j , are assigned to predicates inside and outside the scope of BELIEVE:
INDEXED PREDICATE CALCULUS Incidentally, I am inclined to believe that it would be more appropriate to assume . that predicates in contexts usually taken as nonintensional should be generally interpreted as if they were in intensional contexts; then, indexed predicates are to be interpreted always by means of a class of sets as in epistemic modal logic, and what we here call a (mini-)world is always a class of sets (possible worlds in epistemic modal logic). But this is just a hint and irrelevant to the following discussion. Now consider (21)
( 3 x ) BELIEVE . (MAGNUS, WITCH . (x) & BLIGHT . ( x , the mare)) I
I
I
Both occurrences of variable x are in a predicate indexed with j. Hence, following the general convention introduced earlier, the domain of x is the world indicated by j, i.e. (perhaps, a mini-subworld of) Magnus's belief world. In contrast with (J 9), ( 2 1 ) does not have the existential entailment w.r.t. the real world, as (20) does not. The difference be tween (20) and ( 2 1 ) , in turn, relates to the de-dicto/de-re d i c h o t o m y .
(22)
Magnus believes that Barbara is a witch and that she blighted the mare.
is true. In contrast, this does not follow from de dicto (20); it may be that for no individual x, 'Magnus believes that x is a witcn and that she blighted the mare' is true. Compare ( 2 1 ) with the formula in ordinary predicate calculus obtained by dropping indices from it: (23)
( 3 x) BELIEVE (MAGNUS, WITCH (x)
&
BLIGHT (x, the mare)).
As mentioned earlier, the customary convention associated with logical representations interprets terms in a nonintensional context with respect to the real world (as conceived by the speaker). (23), thus, claims the existence of an individual in the real world, which, according to Magnus's belief, but perhaps not the speaker 's, is a witch. The existence of this individual is 'transparent'; only is her characterization as a witch 'opaque ' . (23) does not represent the reading (2 1 ) of ( 1 6). Ordinary· logic with the customary convention cannot represent this opaque reading (2 1). How can we represent the meaning assigned to (23) by the customary practice in the framework of indexed logic? I do not know whether the English sentence ( 1 6) has a natural reading corresponding to this customary logical form, but I leave this factual question open. If the JS, vol. l , no. !
51
Downloaded from jos.oxfordjournals.org by guest on January 1, 2011
· Thus, ( 2 1 ) is subject to existential instantiation; for some individual, say Barbara, though perhaps unbeknownst to the speaker, the proposition
S.-Y. KU RODA meaning of (23) is not represented by ( 2 1 ), how can it be represented in the indexed logic? We have to bind x in a predicate indexed with i outside BELIEVE, but unlike the representation in ( 1 9) the predicate WITCH cannot serve this purpose in this case. The required predicate can have no conceptual conteut except for the mere existential import w.r.t. the world indicated by index i. A natural solution would be to introduce a 'universal' predicate U, such that for any x, U(x) is true. Such a predicate is redundant in customary logic, but it would be natural and useful for our purpose. Thus, we have (24)
( 3: x) (U. (x) & BELIEVE (MAGNUS, WITCH. (x ) & BLIG H T. (x, the I I l l mare))
to represent the meaning that would customarily be assigned to (23). The domain of x is now again the 'intersection' of the worlds indicated by i and j. Let us now consider the sentence obtained by replacing the proper name Magnus in ( 1 6) by the universally quantified term everyone: Everyone believes that a witch blighted the mare.
a type of sentence discussed by loup ( 1 977) and Fauconnier, following her. We now have two de re opaque readings, one with narrow and the other with wide scope of the existential quantifier w.r.t. the univer sal quantifier: (x) & BL IGHT.
(26)
( Vy) ( ] x) BELIEVE. (y, WITCH .
(27)
( ] x) ( V y) BELIEVE . (y, WITCH . (x) & BLIGH T . (x, the mare)) . I J ( Y) J (Y )
I
J(Y)
J (Y )
(x, the mare))
In ( 26) the existential quantifier is in the scope of the universal quanti fier and the domain of x is dependent on y , i.e., the belief world of y. For each y , x ranges over the belief world of y indicated by the index j(y). In contrast, in (27) the variable x is bound inside the predi cates indexed by j(y), where y ranges over the domain of everyone; hence the domain of x is the 'intersection' of the worlds indicated by j(y)'s, i.e., the common belief of everyone's concerned. If we drop indices from (26) and (27), we get formulae in ordinary logic: (28) ( V y) ( ]x) BELIEVE (y, W ITCH (x) & BLIGHT (x, the mare)) (29) ( ]x) ( Vy) BELIEVE (y, W ITCH (x) & BLIGHT (x, the mare)) How would they be interpreted according to customary convention? Since the existential quantifier outside of the scope of BELIEVE is assumed to carry the existential entailment, these formulae are taken as entailing the existence of an entity x (possibly depending on y for (29)) in the real world, which 'everyone ' believes is a witch and which 52
JS, vol. l , no. l
Downloaded from jos.oxfordjournals.org by guest on January 1, 2011
(25)
INDEXED PREDICATE CALCULUS 'everyone ' believes blighted the mare, although the speaker does not necessarily believe in this characterization of x. These meanings can be represented in indexed logic by means of the universal predicate U introduced earlier: (30) (31)
(x) & BLIGHT (x, ( V y ) ( :!x) (U (x) & BELIEVE (y, WITCH i i j(y) j(y) the mare))) (x) & BLIGHT (x, ( 3 x ) ( V y ) (U (x) & BELIEVE (y, WITCH i i j(y) j(y) the mare))).
In contrast, ordinary logic with the customary convention has no natural way of representing the readings (26) and (27). In comparison with the 'opa q ue' reading (30) and ( 3 1 ) , the correspon ding 'transparent ' readings are represented by (32) (33)
( V y) ( Sx) (WITCH . (x) & BELIEVE . (y, BLIGHT.
In (32) x depends on y and it ranges over the intersection of the worlds indicated by index i and index j(y), i.e. (some subdomain in) the intersec tion of the (speaker's) real world and y's belief world. In (33), in contrast, x does not depend on y and it ranges over (some subdomain in) the common beliefs of the speaker and of the everyone 's concerned. The de dicto reading of (25) is represented by (34)
( V y) BELIEVE. (y, ( 3 x) (WITCH . I
I( y )
(x) & BLIGHT. .
J ( y)
(x, the mare))).
This is the same reading as represented by the ordinary logical form obtained by dropping all the indices from (34): (35)
( V y) BELIEVE (y, ( S x) (WITCH (x) & BLIGHT (x, the mare))).
<;)ur formalism does not admit a de dicto read ing with wide scope of the existential quantifier w .r. t. the universal quantifier. That is, a reading in which the content of y's belief is ( S x) ( WITCH (x) & BLIGHT (x, the mare)) and yet the x is independent of y. Whether such a reading exists for (25), or exactly what such a reading means seems a moot question. One might think of a communal belief of some sort, which a sentence like the following perhaps represents: (36) The villagers believe that a witch blighted the mare. Such a reading could be easily accommodated in IPC, albeit in an ad hoc way, by
JS, vol . l , no. !
53
Downloaded from jos.oxfordjournals.org by guest on January 1, 2011
(x, the mare))) I I I (y ) ( S x) ( V y) (W ITCH . (x) & BELIEVE . (y, BLIGHT . (x, the mare))). I I I (y )
S.Y. KURODA
(37) ( VY) BELIEV E . (y, ( :ix) (WITCH . (x) I
J
&
BLIGHT . (x, the mare))) )
where the index inside BELIEVE is made independent of the variable y and interpreted as indicating the world of communal belief of the villagers. Fauconnier, citing loup ( 1 977) mentions a paradox of formal logic concerning the wide scope 'nonspecific' reading of a sentence of the type (25). This issue originates in Geach 's discussion of intensional identity, though the issue raised by these three scholars may not be totally identical. In fact, the issue Geach raised relates to the scope relation between the existential quantifier and conjunction, rather than between the existential and the universal quantifier, and also, crucially, relates to the anaphoric use of pronouns in natural language, as well as cross-reference between the scope of different types of modalities. His model sentence is:
(38)
Hob thinks a witch blighted Bob's mare, and Nob wonders whether she (the same witch) killed Cob's sow.
H aving said this, I would like to take the same line of approach to the Ioup-Fauconnier wide scope 'nonspecific' reading as Saarinen did for Geach 's intensional identity reading. If I understand Saarinen correctly, this amounts to interpreting the intended 'nonspecific' reading as a de re reading. In terms of the possible world semantics of modality, what differentiates de dicto and de re readings is whether or not a cross-world line of any kind goes through epistemic alternatives; inten sional identity must involve some kind of cross-world line not just through the epistemic alternatives for one individual, but further through the epistemic alternatives of each of the everyone's concerned. What the nature of such cross-world line can be is a metaphysical question which cannot be answered fully within the general theory of formal logic. The customary convention associated with standard logic, however, restricts itself to the particular interpretation of a de re reading, i.e., one that carries the 'existential entailment ', and can only represent wide scope 'specific' readings of a very special sense. Indexed logic escapes from this restriction naturally. Thus, our wide scope de re reading (27) may be taken as accommodat ing loup's and Fauconnier 's wide scope 'nonspecific' reading, if 'specific' is meant strictly to refer to an individual determined by physical/descrip tive cross-world identification, to borrow a ter m from Hintikka and 54
JS, vol. J , no. !
Downloaded from jos.oxfordjournals.org by guest on January 1, 2011
Thus, the potential issue raised by Geach may not be taken simply as a limit case of the problem mentioned by loup and Fauconnier. But here I am not concerned with the full potential of Geach's issue; m�· reference to Geach is merely to _ indicate that the issue of w ide scope 'nonspecific' reading raised by loup and Fauconnier has a historical origin in Geach.
INDEXED PREDICATE CALCULUS Saarinen (cf. Saarinen, 1 98 1 ), and, correlatively, 'nonspecific' leaves room for perspectival or perhaps other cross-world identification. the (Recall loup's illustration of the wide scope nonspecific reading same (witch) blighted everyone's mares, perhaps because of the particular type of blighting that was done ' (p. 243)). For, so far as our formalism goes, we are not committed to any metaphysical import of cross-world identification (in the sense of epistemic modal logic), which formally explicates the sense of de re. ' •..
To conclude, I have perhaps focused too much on logical issues. From a different perspective, !PC may be considered as an extension of the familiar device of indexing noun phrases for indicating corefer ence in formal linguistics. One generates noun phrases with indices freely, and formulates restrictions on coreferentiality. Likewise, we can think of syntax generating freely indexed predicates. Then, it is a role of formal semantics to prescribe certain formal restrictions on coreferentiality or other interpretive conditions. Formal pragmatics, then, may be conceived of as providing graded strategies of further identifying (i.e., coreferencing) indices within the licence semantics allows. This program m e seems to me to be profitable, in particular, for formally separating certain types of semantic vs. pragmatic issues. Many, if not all, of the examples discussed by Fauconnier can, I expect, be recast in formal terms in the perspective of !PC. I will carry out the programm e sketched here systematically in a more extended work on IPC under preparation. Appendix In a recent publication Esa Saarinen ( 1 9 8 1 ) has provided a careful description of ambiguities in intensional contexts. I would like to relate the preceding treatment of ambiguities in intensional contexts to Saarinen 's description. Saarinen discusses (at least) five ways in which quantifiers are ambiguous in intensional contexts. Of these what is relevant to us are his first three, 'the scope ambiguity' , 'an ambiguity in the existential import' , and 'an intermediate reading' .
JS, vol. l , no. l
55
Downloaded from jos.oxfordjournals.org by guest on January 1, 2011
But if a genuine de dicto reading is at issue, the scope dilemma in loup's, and Fauconnier's sense is not resolved in !PC, any more than in standard logic. Our 'solution' to the loup-Fauconnier paradox may, in a sense, be taken as a refusal to admit the existence of the problem. But Fauconnier 's approach does not fare any better. As far as I can see, it cannot distinguish the wide scope opaque de re r e a d i n g , which is represented by (27), and the wide scope de dicto r e a d i n g , i f i t genuinely exists. The advantage of our formalized approach is to help us see separately the limitations, if any, of the formalism of standard logic, on the one hand, and the limitations imposed on it by the cus toma ry convention to employ it, on the other, and if the paradox exists, help us to locate it exactly where it is.
S.-Y . KU RODA The scope ambiguity, in Saarinen's sense in this contex t, concerns the relative scope relation between a quantifier and a modal element. It appears that he uses the traditional paired terms de re I de dicto a n d the modern linguistic paired terms 'specific' I 'nonspecific' interchange ably to differentiate two readings arising from this parameter of ambi guity: de re and 'sp-ecific' to dub the wide scope reading and de dicto and 'nonspecific' to dub the narrow scope reading. I have used the dicho tomy de re I de dicto to indicate this ambiguity. Both in Saarinen's standard logical representations and my indexed representations, this ambiguity is represented (i.e. resolved) in terms of scope relationship. The pair ( 1 7)/( 1 8) in standard logic and the pair ( 1 9)/(20) in indexet:l logic illustrate the ambiguity of a sentence, ( 1 6), along with this para meter.
Saarinen's next parameter of ambiguity, 'the ambiguity in the exist ential import' is a parameter of ambiguity subordinated to the de re read ing. This is the ambiguity illustrated by the pair ( 1 9)/( 2 1 ) in our frame work of indexed logic. It concerns different ways in which occurrences of the variable bound by the (wide scope) quantifier are controlled by indices of predicates. Standard logic with the customary convention cannot distinguish this ambiguity; it fails to represent the reading (2 1 ). Saarinen suggests a 'way out of this conflict', which 'is to allow a systematic ambiguity in the existential import of quantifiers in intensional contexts such that they may or may not involve an existen tial presupposition '. (p. l 3) Thus, Saarinen keeps the standard formalism of logic but departs from the customary practice associated with it. In his article, he does not provide logical formulae , but if I under stand him correctly, the logical representation (23) is itself taken as ambiguous, according to his suggestion. But care must be taken that the two readings that are assigned to (23) in Saarinen 's framework do not together represent the ambiguity between our ( 1 9) and ( 2 1 ). One reading of (23), without the 'existential presupposition ', corresponds to ( 2 1 ) in indexed logic, while the other reading, that with the 'existential . presupposition' corresponds to (24) in indexed logic, as explained earlier. Thus, there is another parameter of ambiguity of ( 1 6), and this illustrates Saarinen 's third ambiguity whereby 'a quantifier phrase splits up', so that its existential import and predicative import are out of and within the scope of an intensional operator, respectively. 56
JS, vol. l , no. l
Downloaded from jos.oxfordjournals.org by guest on January 1, 2011
Incidentally, the term 'scope a mbiguity' and correlatively 'wide scope' and 'narrow scope' reading should be reserved for a general use, whereby any two scope-bearing elements, not necessarily a quanti fier and a modal element, interact to produce an ambiguity. Thus, as far as feasible, it would be advisable to introduce specific terms to refer to different readings arising from a specific case of scope ambiguity.
INDEXED PREDICATE CALCULUS Saarinen does not seem to introduce technical terms but refers to the two readings along with his second parameter · of ambiguity ( 'existentional import') by the 'customary ' and the 'new ' reading of the de re reading. (cf. e.g., p. l 4) I use the familiar terms ' transparent ' (cf. ( 1 9)) and 'opaque' (cf. ( 2 1 )) t o distinguish them. H e seems t o suggest that the new reading arising from this third parameter of ambiguity (his intermediate reading) is 'intermediate between the customary de dicto and de re reading ', but it is, in our conception, better charac terized as intermed iate between transparent and opaque. (24) is, thus, the de re transparent-opaque reading, intermediate between de re t r a n s parent ( 1 9) and d e re opaque ( 2 1 ). Let me summarize here the correspondence between Saarinen's and my own representations and names given to those readings of ( 1 6) discussed above in a diagram form: ( 39)
Kuroda
de re, customary ( I n ( 1 8) --de dicto --de re, new-(23) --intermediate__.-
( 1 9) (20) (2 1 ) (21;)
de de de de
re, transparent dicto re, opaque re, tr�nsparent-opaque.
Sentence (25) introduces still another parameter of ambiguity due to the scope of the universal quantifier relative to the existential. Hence for each type of de re we have a pair of narrow and wide scope readings (de re, transparent ( 32) and (33); de re opaque ( 26) and (27); de re transparent-opaque ( 30) and (3 1 )). University of California, San Diego and Max-Planck-Institut fur Psycholinguistik, N ij megen ·
Notes 1
This paper was prepared for the Colloquium on Discourse Representation held at Cleves in September, 1 98 1 . There is a partial overlap between this paper ·and an earlier paper on the same topic, Kuroda ( 1 98 1 ). I would like to acknowledge the Chicago L inguistic Society for granting permission to incorporate here several paragraphs of the earlier paper verbatim. Thanks are due to E. Engdahl, whose comments on an earlier draft helped me improve the exposition of the paper. JS, vol . l , no. l
57
Downloaded from jos.oxfordjournals.org by guest on January 1, 2011
Saarinen
S.-Y. K URODA
=
=
=
=
58
=
JS, vol. l , no. !
Downloaded from jos.oxfordjournals.org by guest on January 1, 2011
2 This paper was circulated in mimeographed form in 1 979, but is yet unpublished at the time of writing. The main thrust of the paper, however, is summarized in Fauconnier ( 1 98 1 ). 3 1 am not claiming here that all uses of definite nouns can be accounted for directly as definite descriptions with an adequate specification of context, as this example illustrates. I assume that there is at least another use, the anaphoric use of the definite noun, w h i c h cannot be reduced to definite descriptions. The following example, due to J.D. McCawley is to the point. Assume that some dog has already been mentioned in the context of conversation. For example, assume the speaker has said: "My kid got a dog from his uncle yesterday". Now, this sentence might be followed by: (a) The dog is fighting with another one. McCawley contends that, however small one may take the discourse context of this sentence, it contains two dogs and the theory of definite descriptions would fail to account for the use of the dog in (a). I take it that this use of the dog is anaphoric and must be grouped together with anaphoric uses of pronouns, not with those uses of definite nouns that can be accounted for as definite descriptions, with appropriate restriction of context. In this respect, the definite noun must be distin guished from the definite noun phrase in general; for, the definite noun phrase with a restrictive relative clause does not seem to function anaphoricaly. ( With example (a), however, McCawley seems to argue for a different position from that assumed here; he seems to argue that the theory of definite descriptions has no role in accounting for the use of definite nouns.) 4 More exactly, Jet W 1 and W 2 be two worlds. Let us assume first that they are d isjoint. A cross-world identification function f is a (possibly partial) one-to-one function from W1 to W 2 . Let w3 be the union of W 1 and W 2 and let W 4 be the quotient set of W3 by the equiva lence relation R defined as follows: (i) for any x, x Rx; (ii) xRy if f(x) or x f(y). Then, the 'union' in the relevant sense here is y W 4 and the 'intersection' is the subset W5 of W4 consisting of those equivalent classes that contain more than one element of W 1 or w2. Next, assume that W 1 and W2 are not disjoint. Then, let g be a partial function from w 1 to w2 which is the identity function, i.e., the function x. We then impose defined on the intersection of W 1 and W 2 by g(x) on a cross world identification function the condition that it must be an extension of this identity function, that is, f is defined on the g(x) x . We construct W } , w4 , intersection of W 1 and W2 and f(x) and W 5 as above. The 'union' is W 4 , as before. The 'intersectiOn' is defined as the union of W 5 and the intersection, in the ordinary sense, of W 1 and W 2 . Note that f rom the condition imposed on f, it follows that the intersection, in the ordinary sense, of W 1 and W2 may be considered as a subset of W4. In this paper I am not giving examples to motivate the dual part of the convention C- 1 . This convention will be discussed more extensively in the extended treatment of IPC under preparation. Also, in this paper, I am not taking up the problem of negation, falsehood, and
INDEXED PREDICATE CALCULUS possible truth-value gap. It would seem probable that a number of options or alternatives to pursue in this regard are open for !PC just as for (extensions of standard) non-indexed logic. In other words, the problem of negation and truth-value gap concerns another parameter of possible extensions of standard logic orthogonal to the parameter that concerns indexing. 5 By 'this real world' I mean a formal model that is to account for (a relevant part of) the speaker's understanding of what he believes to be this real world. Similar expressions in what follows should be understood with appropriate qualifications of this sort. 6 Or, 'false' depending on how we deal with the problem of negation, falsehood and truth-value gap; ct. note 4. 7 In the context of this presentation, epistemic logic, with possible world semantics, only has the role of explicating the meaning of intensional predicates like BELIEVE giving formal set-theoretic interpre tations to predicates inside BELIEVE. Cross world identification mention ed with respect to epistemic logic belongs to a different level (a meta level, so to speak) from cross world identification associated with the indexing of predicates and constants in !PC.
Fauconnier, G., 1 979: Mental spaces - a discourse-processing approach to natural language logic. Mimeo. Universite de Paris VIII. Fauconnier, G., 1 98 1 : Pragmatic functions and mental spaces. Cllgn i t ion 1 0, 8 5-88. Geach, P. T ., 1 967: Intentional identity. Journal of Philosophy 74. (Reprint ed in P. T. Geach, Logic Matters. Blackwell, Oxford, 1 972. Pp. 1 46- 1 53.) Ioup, G., 1 977: Specificity and interpretation. L inguistics and Philosophy 1 ; 233-245. Jackendoff , R., 1 975: On belief context, Linguistic Inquiry 6; 5 3-93. Kuroda, S.-Y., 1 9 8 1 : Indexed predicate logic, The 1 7th Chicago Linguistic Society Meeting. Pp. 1 56-1 63. Saarinen, E., 1 978: Intentional identity interpreted, L i n g u i s t i c s and Philosophy 2 ; 1 5 1 -224. Saarinen, E., 1 98 1 : Quantifier phrases are (at least) five ways ambiguous. In: F. Heny (Ed.), A m bigui ties in Intensional Contexts. Reidel, Dordrecht. Pp. 1 -45.
JS, vol. 1 , no. 1
59
Downloaded from jos.oxfordjournals.org by guest on January 1, 2011
References
PRINCIPLES OF DISCOURSE DELETION CASE STUDIES FROM ENGL H, RUSSIAN AND JAPANESE*
�
Susumu Kuno
A bstract A syntactically optional constituent in a sentence can be deleted if it is recoverable from the preceding conte.rt. This does not mean, however, that all such constituents are deletable. This paper hypothesizes that there is a pecking order of deletion, which dictates that deletion should proceed from less important to more important informa tion. Evidence is drawn from English, Russian and Japanese in support of this hypothesis. Interaction of this constraint with various syntactic rules in each individual language is examined, and it it hypothesized that unacceptability does not result when the above pecking order of deletion principle is violated due to the structural pressure of the language. Further discourse deletion data from Russian and Japanese are introduced, and principles that control them are formulated and justified. Introduction
In English, Japanese and all other languages that I know, certain con stituents in a sentence ca:n be deleted when the condition of discourse recoverability is met. For example, observe the following exchanges: ( l ) Speaker Speaker ( 2) Speaker Speaker (3) Speaker Speaker
A: B: A: B: A: B:
Did you give any public lectures last year? Yes, I gave a few fJ. Did you find any letters in my mailbo.r? Yes, I found some (J. Can you see Mt. Fuji from where you live ? Yes, I can see it 0 on clear days.
In ( 1 B, 2B, 3B), fhe adverbs last year, in your mailbox and from where I live are missing. Constituents that can be missing under the condition of discourse recoverability are called 'optional' constituents. On the other hand, the object it of (3B) , for example, cannot be missing in spite of the fact that it is perfectly recoverable from context. Constituents that c<�;nnot be missing even when the recoverability condition is satisfied are called 'obligatory' constituents. The object of the verb see is an obligatory constituent. 1
JS, vol. l , no. l
61
Downloaded from jos.oxfordjournals.org by guest on January 1, 2011
1.
S. KUNO The discourse deletion process above seems, however, to apply selec tively. There are contexts in which deletion of optional constituents results in unacceptability: */ is used here to mark sentences that are unacceptable in the specified context, but which might be acceptable given appropriate contexts. (4) (5) (6)
Speaker A: Did you get your Ph.D. last year? Speaker B: a. Yes, I got it last year. b. *I Yes, I got it (J. Speaker A: Did you find this Jetter on the front lawn? Speaker B: a. Yes, I found it there. b. */ Yes, I found it (J. Speaker A: Are you going to run your business from where you live now? Speaker B: a. Yes, I'm going to run it from there because I don't want to leave my husband and move to New York. b. */ Yes, I ' m going to run it (J because I don't want to leave my husband and move to New York.
1.1
Pecking Order of Deletion
Let us examine (4, 5, 6) again. We observe that Speaker A 's questions in these examples can be roughly paraphrased as follows: (7) (4 N (5 N (6 N
= = =
When did you get your Ph.D.? Where did you find this letter? Where are you going to run your business from?
Let us assume that the part of an answer that corresponds to the wh-word in the question represents the most important information in the answer. Then we can assume that (4Bb), (5Bb) and (6Bb) are unacceptable answers because they omit the most important information while retaining less important information. In contrast, the way that . ( 1 A , 2A , 3 A ) are ordinarily inte rpreted, they are not paraphrasable in the same fashion: (8) (1 A) I= When did you give public lectures? (2 A) I= Where did you find letters? (3 A) I= Where can you see Mt. Fuji from? 62
JS, vol. l , no. !
Downloaded from jos.oxfordjournals.org by guest on January 1, 2011
Given the way that each of the above questions is ordinarily interpreted, the (b) sentence is not an appropriate answer. It might be possible to come up with contexts which would make the (A-Bb) exchanges appropri ate, but such contexts would be marked ones and are not immediately obvious. 2 What makes deletion of optional constituents possible in ( I , 2, 3), and inappropriate in (4, 5, 6)? This paper addresses itself to this question, and proposes a constraint on discourse deletion that seems to have a broad scope of application both within individual languages and across languages.
PRIN CIPLES OF DISCOURSE DELETION Rather, these questions have any public lectures, any letters and can their foci, and can be paraphrased in the following way: (9) ( 1 p.J (2 p.J
(J N
= = =
as
How many public lectures did you give last year? How many letters did you find in my mailbox? Is it possible or not possible to see Mt. Fuji from where you Jive?
In the answers ( 1 Bb,2B b, 3Bb), the deleted constituent does corre spond to the focus of the question, and hence, it is not the most impor tant information in the answer. Hence, it is possible to delete it. As discussed in previous papers (Kuno 1 980a, 1 982), these observations have Jed to the following formulation: ( 1 0)
Pecking Order of Deletion Principle: Delete less important informa tion first, and more important information last.
The above formulation embodies three implicit claims. They are:
I will give below a few more examples that must be explained by the Pecking Order of Deletion Principle: ( 1 2) ( 1 3) ( 1 4) ( 1 5) ( 1 6)
Speaker A: Speaker B: Speaker A: Speaker A: Speaker B: Speaker A: Speaker B: Speaker A: Speaker B:
JS, vol . l , no. l
Did you buy a watch in Switzerland? Yes, I bought one � Did you buy this watch in Switzerland? •/Yes, I bought it � Did you publish your dissertation a few years later? Yes, I published it (J. Did you publish your first book while you were still a graduate student? */Yes, I published it (J. Did John come to the meeting? Yes, he came � 63
Downloaded from jos.oxfordjournals.org by guest on January 1, 2011
( 1 1 ) A. The crucial factor that determines the order of deletion is that of "more important/less important" and not "newer/older". In earlier formulations of the principle, as found in Kuno ( 1 97 8a, 1 978b, 1 979a, 1 979b), I erroneously assumed that the crucial factor was that of "newer/older" or " more unpredictable/ more predictable". The formulation given in ( 1 0) is justified in Kuno ( 1 980a, 1 982). B. The order of deletion cannot be accounted . for by a dichotomy between important and unimportant. It needs to be based on relative degrees of importance. c. The Pecking Order of Deletion Principle does not apply when two or more constituents of the same degree of importance are involved. Justification for this claim, as well as for (B), is found in the papers cited above.
S. K UNO ( 1 7)
Speaker A: Did John come by car? Speaker B: */Yes, he came (J.
In ( 1 2), the focus of the question is either a watch or buy a watch. I t i s possible to delete in Switzerland in the answer because it is not the most important information. On the other hand, in ( 1 3}, the focus of the question is in Switzerland, and hence, it is not possible to delete it and retain the less important information in the answer. Similarly, in ( 1 4), the focus of the question is publish, and therefore, it is possible to delete the less important information a few years later. On the other hand, in ( 1 5), the focus of the question is while you were still a graduate student, and hence it is not possible to delete this time adverb in the answer if the other less important information is to be retained. Likewise, in ( 1 6), the focus of the question is come ( qr rather, the sentence has an unmarked interpretation in which come is the focus), and therefore, to the meeting can be deleted. On the other hand, in ( 1 7), by car is the focus and therefore, it is not possible to delete it and retain the nonfocus element came. 3 1 .2
Interaction with Syntactic constraints
( 1 8)
Speaker A: Speaker B:
Did you buy this watch in Switzerland? Yes, I did.
In ( 1 8A) as already mentioned, in Switzerland i s the focus. In ( 1 8B), this focus as well as the nonfocus constituents buy and this V.atch have been deleted. Therefore, there is no violation on the Pecking Order of Deletion Principle involved in the deletion of the latter two constitu ents. What remains to be looked into is whether the retention of I and did has violated the principle. I assume that the auxiliary verb did con veys here the affirmative nature of the answer in contrast to the negative nature of did not, and in this sense, it conveys important information. Since 0 8A) can be answered either with ( 1 8B), or with "Yes, in Switzerland", let us assume that did ( i.e., the affirmative na ture of the answer) and in Switzerland convey equally important informa tion in the answer. On the other hand, the subject of did in ( 1 8b) clearly conveys much less important information than the deleted in Switzerland the Pecking if ( 1 8A) is uttered without any emphatic stress on you. If Order of Deletion Principle is correct, ( 1 8B) should be unacceptable because in Switzerland, which conveys important information, has been deleted while I, which conveys less important information, has been left behind. Why is it that ( I 8B) is perfectly acceptable in spite of this violation of the Pecking Order Principle? I hypothesize that the Pecking Order of · Deletion Principle is sensitive to the distinction between violations which are 'intentional', so to speak, and those which 64
JS, vol. l , no. l
Downloaded from jos.oxfordjournals.org by guest on January 1, 2011
·Thus far, I have avoided discussing the most common discourse deletion pattern in question-answer pairs, namely that of Verb Phrase Deletion. Observe the following discourses:
PRINCIPLES OF DISCOURSE DELETION are 'unintentional '. In the case under discussion, the retention of the subject I, which carries relatively unimportant information in discourse, has been necessitated by a constraint in English which says that a tensed verb must have a surface subject. 4 The moment that the decision was made to leave did behind as a marker that conveys the affirmative nature of the answer, the retention of its subject was automaticaJJy determined by this surface subject constraint. Thus the violation of the Pecking Order of Deletion Principle attributable to retention of the subject is 'unintentional', and therefore, it does not result in unac ceptability. Compare this situation with those crucial cases that we have discussed in the preceding sections. For example: ( 1 9)
Speaker A: Were you born in Tokyo? Speaker B: */Yes, I was born (l.
The decision to delete in Tokyo while retaining bom, which bears much less important information, was not forced by any syntactic constraint of English: it was an ' intentional' decision. Hence, the unacceptability of the sentence.
(20)
Active and Passive Discourse-Rule Violations: S e n t e nces that in volve active avoidable (or intentional) violation of discourse principles are unacceptable. On the other hand, sentences that in volve passive unavoidable (or unintentional) violation of discourse principles go unpenalized and are acceptable.
The above principle can be independently motivated by various phenome na involving interactions of discourse principles and syntactic rules. Justifications of this principle can be found in Kuno ( 1 979b). 1 . 3.
More Data from English
The Pecking Order of Deletion Principle interacts with Gapping in an interesting way. Gapping is a process which is responsible for deriving the structure corresponding to (b) from the one corresponding to (a): (21)
a . John h i t Mary, and Tom hit Jane. b. John hit Mary, and Tom (J Jane.
have noted elsewhere ( Kuno 1 976, pp. 308-3 1 0) that constituents deleted by Gapping must be not only recoverable from the first conjunct, but also be contextuaJJy known. Observe the foJJowing contrasts: (22)
Speaker A:
JS, vol. l , no. l
Who hit who? 65
Downloaded from jos.oxfordjournals.org by guest on January 1, 2011
I have explained above what superficiaJJy appears to be a selective application of the Pecking Order of Deletion Principle as the result of interference from a syntactic constraint. This happens to be a subcase of a more general principle that can be stated as foJJows:
S. KUNO Speaker B: ( 23)
Speaker A: Speaker B:
John a. John b. What did Jane? a. John b. */John
hit Mary, and Tom hit Jane. hit Mary, and Tom (J Jane. John do to Mary, and what did Tom do to hit Mary, and Tom hit Jane, too. hit Mary, and Tom (J Jane, too. 5
The consideration of where the focus of the question lies seems to play a role in the formation of the affirmative-negative questioning pattern of the following sort: ( 24)
a.
Did you or did you not publish the manuscript you showed me last year? m a nuscript you b. Did you publish or did you not publish t h e showed me last year?
It seems that (24a) and (24b) are different in that while the former asks simply whether the statement "you published the manuscript you showed me last year" is true or not, and therefore, allows any appropriate constituent in the statement to be interpreted as the focus of the question, ( 24b) requires that publish be interpreted as the focus of the question. This distinction shows up, at least for some speakers, in the following contrasts in acceptability judgments. (25) (26)
a. Were you or were you not born b. ??Were you bom or were you not a. Did you or did you not buy this b. ?Did you buy or did you not buy
in Tokyo? in Tokyo? perfume in Paris? this perfume in Paris?
bom
The awkwardness, marginality or unacceptability (dep�nding upon speakers) of (25b) seems to be due to the fact that the sequence b or n and not bom requires that the sentence be interpreted as a question with the focus on bom, thus making it pragmatically implausible because 66 ·
JS, vol. l , no. I
Downloaded from jos.oxfordjournals.org by guest on January 1, 2011
In (22), Gapping can apply to delete hit of the second conjunct because it conveys information that is contextually known. (namely, via (22A)). On the other hand, in ( 2 3), Gapping cannot apply to delete hit in spite of the fact that it is recoverable from the first conjunct because it is not contextually known. This fact can be automatically explained by the Pecking Order of Deletion Principle. In (22Ba), Tom and Jane o f t h e second conjunct convey much more important information than hit does. Hence, deletion of hit does not violate the Pecking Order of Deletion Principle. On the other hand, in ( 23Ba), hit represents more important in formation than Tom or Jane does. Hence, deletion of h"it r e s u 1 t s in a violation of the Pecking Order of Deletion Principle. Since this violation has not been forced by any syntactic constraint in English, but has been created by an 'intentional' application of Gapping, an optional rule, the resulting sentence is marked as unacceptable by the Active and Passive Discourse-Rule Violations Principle.
PRINCIPL ES OF DISCOURSE DELETION it can only mean something like " I s getting born what happened to you in Tokyo?". Similarly, ( 26b) seems to be awkward, marginal or unacceptable, depending upon the speaker, because what the sentence implies is "Did you BUY this perfume in Paris, or did you do something else about/for/to it?", an interpretation for which it is difficult to find a pragmatically plausible context.6 2.
Discourse Deletion Phenomena in Russian
Since the Pecking Order of Deletion Principle is such a natural constraint, observing that it is obeyed in various languages is not an interesting task. Rather, discovering apparent counterexamples in languages where the principle is otherwise observed is a more rewarding task, because it is most likely that violations of the principle have been necessitated by some language-particular syntactic constraint. In other words, apparent counterexamples to the Pecking Order of Deletion Principle in a given language might make it possible for us to uncover thus-far unnoticed syntactic constraints. I will illustrate the kind of generalizations that the Pecking Order of Deletion Principle helps us formulate using Russian as an example. 7 2. 1
Pecking Order of Deletion in Russian
In Russian, there are two ways to mark the focus of the question: attaching the interrogative particle li to the right of the focus constJtu ent, which receives a prominent emphatic stress, and prominent emphatic stress alone without li. For example, observe the following, in which con stituents which receive prominent emphatic stress are spelled in capital letters: (1)
a. b.
K U PIL li Ivan gazetu? bought newspaper ' Did Ivan BUY a newspaper?' Ivan K U PIL gazetu? ' Did Ivan BUY a newspaper?'
The constituent which is followed by li is normally placed at sentenceJS, vol. l , no. l
67
Downloaded from jos.oxfordjournals.org by guest on January 1, 2011
The Pecking Order of Deletion Principle that have proposed says that in applying discourse deletion to recoverable elements i � a given sentence, we must proceed from less important to more Important information. This is such a natural constraint that it is difficult to imagine that there could be a language which would not have a similar constraint. In fact, my examination of languages such as Japanese, Korean, Turkish, Thai, German, Swedish, French, Arabic and Hebrew shows that the same constraint that has been established for English also applies to these languages. It is safe to assume that it is at least a near-universal constraint, and perhaps a true lang uage universal . constraint.
S. KUNO initial position, but a focused constituent not marked with li o r d i n a r i 1 y stays in its original position. Similary, observe the following sentences: ( 2)
a. . G AZETU li kupil Ivan? newspaper bought 'Was it a newspaper that Ivan bought?' b. Ivan kupil GAZETU? ' Was it a newspaper that Ivan bought?'
The pattern with li is used mainly in writing and in formal speech en vironments, as in courtroom interrogations, while the pattern without li is used in colloquial speech. The following discourse shows that the Pecking Order of Deletion Principle applies to Russian as well. (3)
V 1 96 0 li godu vy opublikovali vasu pervuju stat'ju? in year you published your first paper 'Did you publish your first paper in 1 960?' Speaker B: a. Da, ja opublikoval ee v 1 960 godu. Yes I published it in year 'Yes, I published it in 1 960. ' b. Da, v 1 960 (godu). */Da, j a opublikoval ee. '*/Yes, I published it.' Speaker A: MNOGO li s tatej vy opublikovali v 1 960 godu? many papers you published in year ' Did you publish many papers in 1 960?' Speaker B: a. Da, ja opublikovaJ MNOGO statej v J 960. godu. yes I published many papers in year 'Yes, I published many papers in 1 960' b. Da, ja opublikoval MNOGO statej. 'Yes, I published many papers.' c. Da, MNOGO (statej). 'Yes, m any (papers). ' d. */Da, v 1 960 (godu). '*/Yes, in 1 960.' ·
The deletion of v 1 960 godu i s disallowed in (3) because v 1 960 godi.L re presents the most important information in the question-answer pair, but it is allowed in (4) because the focus of the question is elsewhere. Similarly, observe the following discourse: (5)
Speaker A: Speaker B:
68
Vy mozete xodit' v universitet peskom? you can go to university on-foot 'Can you go to the university on foot?' a. Da, ja mogu xodit' v universitet peskom. yes I can go to university on-foot 'Yes, I can go to the university on foot. ' J S , vol. 1 , no. l
Downloaded from jos.oxfordjournals.org by guest on January 1, 2011
(4)
Speaker A:
PRINCIPLES OF DISCOU RSE DELETION Da, ja mogu xodit' peskom. 'Yes, I can go on foot. ' c . *IDa, j a mogu xodit'. '*/Yes, I can go.' d . Da, mogu. 'Yes, (I) can . ' b.
As w e have seen in English, the most important information in the question is 'can' . The second most important information is 'on foot' . O n the other hand, 'go' is one of the constituents which represents the least important information. The acceptability of (5Bb, d) and the unacceptability of (5Bc) is exactly as the Pecking Order of Deletion Principle predicts.
(6)
a. b. c.
Ivan kupil gazetu. bought newspaper ' It was a newspaper that Ivan bought.' Ivan gazetu kupil. ' What Ivan did about the newspaper was buy it.' Gazetu kupil Ivan. 'It was Ivan who bought the newspaper.'
In the above sentences, the sentence-final constituent receives the falling intonation that marks a sentence focus. The string up to the focus receives a gradual rising intonation. (6a) can also be used to describe an event neutrally, as in ' Do you know what happened? Ivan bought a newspaper! ' , where the entire sentence is the focus. The second rule that needs to be discussed here is that of Emphatic Focus Fronting. This rule places the focus constituent in sentence initial position, assigns the sentence a distinct high-low intonation, where there is an abrupt fall of intonation right after the fronted focus. We have already seen this rule in operation in the fronting of !i-marked question-foci. Similarly, observe the following exchanges: (7)
Speaker A: Speaker B:
JS, vol. l , no. l
Ivan kupil GAZETU? bought newspaper ' Did Ivan buy a newspaper?' Da, GAZETU on kupil. yes newspaper he bought 69
Downloaded from jos.oxfordjournals.org by guest on January 1, 2011
It is necessary to introduce two other relevant facts in Russian here. First, Russian has a rule, called Rightward Focus Shift, which places the focus of a sentence in the rightmost position in the sentence. The remaining constituents are also arranged in the order of 'from less to more importan t ' information, but I will not concern myself here with this detail. Sentences which are generated by this process receive a gradual rise-fall intonation, where the fall takes place on the focus constituent. For example, observe the following:
S. KUNO (8)
Speaker A: Speaker B:
'Yes, he bought a newspaper. ' Ivan K UPIL gazetu? 'Did Ivan BUY a newspaper?' Da, K UPIL on gazetu. 'Yes, he BOUGHT a newspaper. '
The coexistence of the above two rules in Russian means that the focus of a sentence, if it is to be moved out ot its underlying positon, can be placed either in sentence-final or sentence-initial position. Thus, if we ignore intonation, a given sentence can have two totally different interpretations depending upon whether it is assumed to have been derived by Rightward Focus Shift or Emphatic Focus Fronting. For example, observe the following: (9)
Focus Focus
According to 0. Yokoyama (personal communicatio_n I 98 1 ), Rightward Focus Shift is used predominantly for marking foci in writing, while Emphatic Focus Fronting is used heavily in colloquial speech. 2.2.
Minimal Sentential Answers
The Pecking Order of Deletion Principle interacts with certain syntactic constraints in an interesting way in Russian. F irst, observe the following discourse: ( 1 0) Speaker A: Speaker B:
V
1 960 li godu vy opublikovali vasu pervuju stat 'ju? in year you published your first paper 'Did you publish your first paper in 1 960? ' a. */Da, opublikoval ee v 1 960 godu. 'Yes, ( I) published it in i 960.' b. */Da, ja opublikoval v 1 960 godu. 'Yes, I published (it) in 1 960.'
The total unacceptability of ( J OBa, b) seem s to suggest th at Russian has a surface subject constraint for tensed verbs and a surface object constraint for transitive verbs. Similarly, observe the following discourses: ( 1 1) Speaker A: Speaker B:
70
Cto podaril tebe Ivan? what gave you 'What did Ivan give you?' a. Ivan podaril mne kol'co. gave me ring ' Ivan gave me a ring.' JS, vol. l , no. 1
Downloaded from jos.oxfordjournals.org by guest on January 1, 2011
Annu Kareninu napisal Tolstoj. wrote 'It was Tolstoj who wrote Anna Karenina.' (R ightward (a) Shift) 'It was Anna Karenina that Tolstoj wrote.' (E m p h a t i c (b) Fronting)
PRINCIPLES OF DISCOURSE DELETION b. */Podaril mne kol'co. '(He) gave me a ring.' c. ??Ivan podaril kol'co. ' Ivan gave (me) a ring.' The unacceptability of ( l i Bb) seems to be due to a violation of the Surface Subject Constraint, and that of ( I I Be) seems to be due to a violation of a similar Surface (Indirect) Object Constraint. Similarly, observe the following discourse: ( 1 2) Speaker A: Speaker B:
Kto polozil banany v xolodil'nik? who put bananas i n refrigerator 'Who put bananas in the refrigerato r?' a. Ivan polozil banany v xolodil'nik. put bananas in refrigerator 'Ivan put the bananas in the refrigerator.' b. */Ivan polozil banany (). c. */Ivan polozil () v xolodil'nik.
The moment we assume that Russian has constraints like the ones motivated by ( 1 0) - ( 1 2), we run into difficulty explaining why sentences such as (5Bd) are possible. (5Bd) is particularly puzzling because it would become an unacceptable answer to (5A) if the subject ja 'I' is r e tained. The following examples show that this is not a peculiarity that can be attributed to the fact that mogu 'can' is an auxiliary verb. ( 1 3) Speaker A: Speaker B: ( 1 4) Speaker A: Speaker B: ( 1 5) Speaker A: Speaker B:
Ivan K U PIL gazetu? bought newspaper ' Did Ivan BUY a newspaper?' Da, K U PIL. 'Yes, (he) bought (a newspaper).' IVAN kupil gazetu? ' Did IVAN buy a newspaper?' a. Da, IVAN. b. Da, I V AN kupil. 'Yes, IV AN bought (a newspaper). ' Ivan kupil GAZETU ? ' Did Ivan buy a N EWSPAPER?' a. Da, GAZETU. b. 'Yes, (he) bou�ht a NEWSPAPER.'
( 1 3B) violates both the surface subject and object constraints. ( 1 4B) violates the surface object constraint, and ( 1 5B), the surface subject JS, vol. I , no. I
71
Downloaded from jos.oxfordjournals.org by guest on January 1, 2011
Neither the direct object nor the locative expression of polozil ' p u t ' can undergo deletion. The above results show that Russian has, in addition to a surface subject requirement for tensed verbs, a constraint against deleting any of the strictly subcategorized constituents of verbs.
S. KUNO constraint. When can we violate these constraints, and when can we not? This is not an easy question to answer because these constraints appear to apply at random. We have seen that a surface subject was needed for ( l l Bb), and a surface object for ( l l Bc). But observe the following discourse: ·
( 1 6) Speaker A: Speaker B:
Cto podaril tebe Ivan ( l l A) what gave you 'What did Ivan give you?' a. Kol 'co. ring b. Kol'co podaril. '(He) gave (me) a ring.' =
( 1 6Bb) is a perfectly acceptable answer to 0 6A) in spite of the fact that it has both the subject and the indirect object missing. Similarly, observe the fo!Jowing discourse:
Speaker B:
Kuda ty polozil banany? where you put bananas 'Where did you put the bananas?' a. Ja polozil banany v xolodil 'nik. I put bananas in refrigerator 'I put the bananas in the refrigerator!' b. */(J polozil banany v xolodil'nik.· v xolodil'nik. c. ??Ja polozil (J v xolodil 'nik. d. (J polozil (J e . (J (J 0 v xolodil 'nik. ·
We can assume that ( 1 7Bb) is an unacceptable answer because of a violation of the Surface Subject Constraint, and ( 1 7Bc), because of a violation of the Surface Object Constraint. However, ( 1 7Bd), which violates both of the constraints, is a perfectly acceptable answer to ( 1 7A). The above mystifying phenomenon can be explained if we assume that Russian has the following rule: ( 1 8) The Minimal Sentential Answer Strategy: R e t ain undeleted focus and the verb to produce a minimal sentential answer.
the
For example, in ( 1 7) , the focus of the answer is v xolodil 'nik in the re frigerator ' . The M inimal Sentential Answer Strategy says that in order to produce a minimal sentential answer, the verb polozil ' p u t ' , which is not the focus of the answer, needs to be retained. Hence derives ( 1 7Bd). Let us now assume that the Surface Subject and Object Constraints, 72
JS, vol. l , no. l
Downloaded from jos.oxfordjournals.org by guest on January 1, 2011
( 1 7) Speaker A:
PRINCIPL ES OF DISCOU RSE DELETION and the constraints for other obligatory constituents are relaxed only when the above strategy is adopted. Then we can say that 0 7Bd) is acceptable, in spite of its violation of the Surface Subject and Object Constraints, because the retention of the verb polozil 'put' was permitted in order to produce a minimal sentential answer. On the other hand, the violation of the Surface Subject Constraint in ( 1 7 Bb) is impermissible because it is not a minimal sentential answer: it has retained ba n a n y , which is neither the focus, nor the verb of the pre-deletion sentence. Similarly, the violation of the Surface Object Constraint that ( 1 7Bc) involves is impermissible because it is likewise not a minimal sentential answer: it has retained the subject ja, which is neither the focus nor the verb of the pre-deletion sentence. Hence the unacceptability of ( I 7Bb) and ( 1 7Bc). What happens if the verb of a sentence is the focus? Observe the following discourse: ( 1 9) Speaker /\:
Since the verb cital is the focus of the sentence, it in itself constitutes a minimal sentential answer: it does not drag in any other constituent of the sentence. The violation of the Surface Subject and Object Con straints in ( 1 9Bb) is due to the Minimal Sentential Answer Strategy, and therefore, it does not result in unacceptability. On the other hand, the violation of the Surface Subject Constraint in ( 1 9Bc) cannot be attributed to the Minimal Sentential Answer Strategy: the object t vo j u knigu 'your book ' was not needed in ( 1 9Bc) to produce a minimal senten tial answer. Therefore, the Surface Subject Constraint applies rigidly to this sentence, and marks it as unacceptable. From this point of view, the fact that ( 1 9Bd) is much better than ( 1 9Bc) is interesting and requires explanation. The sentence should be underivable by applica tion of the Minimal Sentential Answer Strategy, and therefore, it should be as unacceptable as ( 1 9Bc). I claim that ( 1 9Bd) is possible only if Speaker B misinterprets A's 9uestion. Assume that Speaker A intended ( 1 9A) as a question with only citaZ 'read' as the focus of the question. This is the unmarked interpretation of the sentence when only cital receives an emphatic stress. Assume, next, that Speaker B has intentionally or unintentionally reinterpreted Speaker A's question as having double foci: i.e., ty 'you ' and cital 'read' . Then, the Minimal Sentential Answer Strategy, applied to ( 1 9Ba), would yield ( 1 9Bd) because ja and citaZ are both foci of the answer. In fact, this answer assigns a JS, vol . l , no. ! .
73
Downloaded from jos.oxfordjournals.org by guest on January 1, 2011
Speaker B:
Ty CIT AL moju knigo? my book you read ' Did you read my book?' a. Da, ja cital tvoju knigu. yes I read your book 'Yes, I read your book. ' b. Da, (J cital (J. c. */Da, (J cital tvoju knigu. d. ?Da, ja cital (J.
S. KUNO slightly contrastive reading to the subject: namely, it implies 'As far as I am concerned, I read (your book)' with the slight implication that someone the speaker knows did not read the book. Thus, this answer can be used as a convenient device for shifting the discourse topic from Speaker B himself to someone else. In other words, speakers can take advantage of the rules under discussion, violate them deliber ately, and produce the desired effect of dragging the conversation into their home ground.
( 20) Speaker A: Speaker B:
Ty CIT AL moju knigu? you read my book ' Did you read my book? ' a. */ Da, fJ cital tvoju knigu. (:: l 9Bc) yes read your book b. ?Da, tvoju knigu fJ cital. B
F inally, the word order o f elements in sentences produced b y applica tion of the Minimal Sentential Answer Strategy requires examination. Observe the following discourses: ( 2 1 ) Speaker A: Speaker B: (22) Speaker A: Speaker B:
(23) Speaker A: Speaker B:
74
Ivan kupil GAZETU? bought newspaper ' Did Ivan buy a N EWSPAPER?' a. Da, kupil GAZETU. b. Da, GAZETU kupil. Cto podariJ tebe Ivan? what gave you ' What
Downloaded from jos.oxfordjournals.org by guest on January 1, 2011
I still need to explain why the contrastive r�ading of the kind that I in mentioned above is possible for ja in ( 1 9Bd), but not for tvoju latigu ( l 9Bc). It seems to be due to the fact that the sentence-initial positon in Russian is a natural position for topics, including contrastive topics. Therefore, ja in ( l 9Bd) is readily interpretable as a contrastive topic, and hence, the interpretation under discussion becomes possible. On the other hand, tvoju latigu does not occupy this topic position, and therefore, it is not reinterpretable as a semi-focus contrastive topic. Hence, the sentence is unacceptable as an answer to ( l 9A). It becomes much more acceptable if tvoju latigu is preposed, and placed in the topic position:
PRINCIPLES OF DISCOU RSE DELETION b.
•(I) put {them) in the refrigerator. ' V XOLODIL 'NIK polozil.
In each of the above, {a) is the unmarked answer. {b) is more marked, and a higher degree of emphasis is placed on the focus in sentence initial position. In any case, the above examples show that the answers produced by application of the Minimal Sentential Answer Strategy can be arranged either in the 'Verb + Focus' order, or the 'Focus + Verb' order. Now, observe the following exchange: { 24) Speaker A: Speaker B:
Kto dal tebe eto kol'co? who gave you this ring ' Who gave you this ring?' a. */Dal IVAN. b. IVAN dal.
The 'Verb + Focus' pattern {24Ba) does not seem to be acceptable. This is in spite of the fact that in a full sentential answer to the question, the verb dol can readily precede the subject Ivan:
Speaker B:
Kto dal tebe eto kol'co? who gave you this ring ' Who gave you this ring?' Dal mne eto kol 'co IV AN. gave me this ring
The above phenomenon can be explained only if we assume that the Minimal Sentential Answer Strategy applies only to structures which maintain the underlying word order of sentence constituents. In (23), for example, the rule applies to a structure which has maintained the underlying 'Subject - Verb - Object - Locative' order: { 26) Speaker B:
Ja polozil 'nik banany v XOLODIL 'NIK.
• 0
• 0
Then, Emphatic Focus Fronting applies optionally, and places the focus locative to sentence-initial position, yielding {23Bb). In contrast, in the case of {25), the Minimal Sentential Answer Strategy applies to the following structure: {27) Speaker B:
IV AN dal mne eto kol 'co.
u I
Emphatic Focus Fronting vacuously applies to IVA N, and yields {24Bb). The fact that {24Ba) is unacceptable can thus be attributed to the fact that the Minimal Sentential Answer Strategy cannot be applied to the structure corresponding to { 25B), which has undergone. Rightward JS, vol. J , no. l .
75
Downloaded from jos.oxfordjournals.org by guest on January 1, 2011
{ 25) Speaker A:
S. KUNO Focus Shift . It also necessitates the assumption that Rightward Focus Shift does not apply to the output of the Minimal Sentential Answer Strategy. I have established above that the Minimal Sentential Answer Strategy applies to structures which maintain the underlying · word order of Russian, and that Emphatic Focus Fronting can apply to the output. of the M inimal Sentential Answer Strategy. This leads to the following order of application of the rules involved: (28)
Surface Subject Constraint, etc.
..
Discourse deletion facts in Russian are extremely complex, and the facts explained above account for only a minuscule portion of the total picture that needs to be elucidated. However, it seems that the Pecking Order of Deletion Principle, coupled with the Minimal Sentential Answer . Principle, has given us a clue for unravelling this extremely complex phenomenon, which I believe has not thus far been the object of any serious systematic analysis. 3.
Discourse Deletion Phenomena in Japanese
Now we look into discourse deletion phenomena in Japanese, a language which has relatively free word order like Russian, and which does not have a surface subject requirement for tensed verbs, nor a ban against deleting constituents that are strictly subcategorized by the verb. 3.1
Pecking Order o f Deletion i n Japanese
It is ordinarily the case in Japanese that a yes-or-no question can be. answered by repeating the main verb of the question or its negative form. For example, observe the following discourses: 76
JS, vol. l , no. l
Downloaded from jos.oxfordjournals.org by guest on January 1, 2011
A model of grammar which does not recognise the basic word order of elements in a sentence and which does not recognize rule ordering would be hard put to account for the facts that I have discussed in this section.
PRINCIPLES OF DISCOURSE DELETION (1)
Speaker A: Speaker B:
( 2)
Speaker A: Speaker B:
(3)
Speaker A: Speaker B:
Kimi wa kinoo Kanda ni ikimasita ka? you yesterday to went Q ' Did you go to Kanda yesterday?' Hai, ikimasita. yes went 'Yes, (I) went (there yesterday).' Kimi wa kono hon o yomimasita ka? you this book read Q ' Did you read this book?' Hai, yomimasita. yes read 'Yes, (I) read (it).' Kimi wa Kankoku ni itta koto ga arimasu ka? , you Korea to went experience have Q 'Have you had the experience of visiting Korea?' Hai, arimasu. yes have ' Yes, (I) have had (it). '
The above strategy of repeating the main verb in answers does not work when the focus of the question is not the verb. Observe the following contrasts: (4)
Speaker A:
Speaker B:
Sensei wa syuusen no tosi ni oumare ni natta no desu teacher end-of-war year in was born that is ka? Q 'Is it the case that you were born the year that the war ended?' a. Hai, syuusen no tosi ni umaremasita. ' Yes, I was born the year the war ended. ' b. */Hai, tJ umaremasita was born */Yes, I was born t). '
The focus o f (4A) is clearly syuusen n o tosi n i 'in the year that the war ended', and not oumare ni natta 'was born (honorific form)'. In (4Bb), 'was born', the main verb of the question (that is, if we ignore no desu) is repeated9, and the sentence is an unacceptable answer to (4A). This fact can be explained automatically if we assume that the Pecking Order of Deletion Principle applies to Japanese, as well. (4Bb) is unacceptable because it involves deletion of more important information and retention of less important information. This assumption is confirmed by the fact that in the following dialogue, it is possible to delete the time adverb:
JS, vol. l , no. l
77
Downloaded from jos.oxfordjournals.org by guest on January 1, 2011
These examples show that the subject and the object are not obligatory constituents in Japanese.
S. KUNO (5)
Speaker A:
Sensei wa syuusen no tosi ni wa moo oumare ni natte teacher end-of-war year in already born ita no desu ka? was that is ' Is it the case that you were already born the year the war ended?' a. Hai, syuusen no tosi ni wa moo umarete imasita. yes end-of-war year in already born was 'Yes, I was �!ready born the year the war ended.' b. Hai, tJ moo umarete imasita. yes already born was 'Yes, I was already born.'
Speaker B:
The focus of (5A) is moo oumare ni natte ita 'was already born'. Syuusen no tosi ni wa is clearly not the focus of the question as can be seen by the fact that it is followed by wa, which marks themes. (5Bb), with the time adverb deleted, is acceptable because it has retained the most important information in the sentence while deleting less important information, in conformity with the Pecking Order of Deletion Principle.
(6)
(7)
Speaker A:
Sensei wa syozyoronbun o 1 960-nen ni syuppan nasatta teacher virgin article year in publish did no desu ka? that is Q ' Is it the case that you published your first paper in 1 960?' Speaker B: */Hai, (syozyoronbun o) syuppansimasita. published yes first article 'Yes, I published (my first article) tl ' Speaker A: Sensei wa 1 960-nen ni wa ronbun o takusan syuppan teacher year in article many publish nasatta no desu ka? · did that is Q ' Is it the case that you published many articles in 1 960?' Hai, tJ takusan syuppansimasita. Speaker B: yes many published ' Yes, I published lots tJ.• ·
Deletion of 1 960-nen ni 'in 1 960' results in unacceptability in (6B), but not in (7B). Incidentally, note that in (6A), 1 960-nen ni, the focus of the question, appears' immediately before the main verb, while in (7 A), the same adverb, which is not the focus of the question, appears earlier in the sentence. marked by the thematic particle wa . . L ikewise, observe the following pair of discourses: 78
JS, vol. l , no. l
Downloaded from jos.oxfordjournals.org by guest on January 1, 2011
I will give below several more pairs of discourses that show the same kind of contrast:
PRINCIPLES OF DISCOURSE DELETION (8)
Speaker A: Speaker B: Speaker A:
(9)
Speaker B : ( 1 0)
Speaker A: Speaker B:
( I I ) Speaker A:
(8A) is normally interpreted as a question with Pari de ' in Paris' a s its focus. (8B) is an unacceptable answer to this question because the main verb, which is not the focus of the question, is retained while the focus has been deleted. On the other hand, (9A) is a question with katta 'bought' and moratta 'received (free of charge)' as foci. Hence,(9B), which has only the focus retained, is an acceptable answer to the question. Likewise, ( l OA) is normally interpreted as a question with respect to whether i t was on foot or by some other means that the addressee went to Kanda. ( 1 OB) is an unacceptable answer to this question because the main verb itta 'went', which has been left unde leted, represents much less important information than aruite 'on foot' which has been deleted. On the other hand, ( l l A) can be interpreted as a question with itta 'went ' as the focus, meaning 'Did you really go to Kanda on foot as you were planning to?' ( l i B) is a legitimate answer to this question because the focus itta has been left undeleted while kinoo 'yesterday', keikaku-doori 'according to the plan', and amite 'o� foot', which represent less important information than itta ' w e n t ' , have been deleted. Next, observe the following exchange: JS, vol. l , no. I
79
Downloaded from jos.oxfordjournals.org by guest on January 1, 2011
Speaker B:
Kono koosui wa Pari de katta no desu ka? in bought that is this perfume ' Is it the case that you bought this perfume in PARIS?' */Hai, (J katta no desu. yes bought that is ' Yes, it is the case that I bought it (J. • Kono koosui wa katta no desu ka, morotta no desu ka? this perfume bought that is Q Q received that is ' Is it the case that you bought this perfume or you got it from somebody? ' (J Katta no desu. bought tha t is ' It is the case that (I) bought (it). ' Kanda made aruite itta no desu ka? to on-foot went that is ' Is it the case that you went as far as Kanda on foot?' */Hai, (J itta no desu.' yes went that is */'Yes, it is the case that I went (there) (J.• Kinoo keikaku-doori, Kanda made aruite itta no desu yesterday as-planned to on-footwent that is ka? Q ' Is it the case that you went to Kanda on foot as scheduled?' Hai, itta no desu. yes went that is 'Yes, it is the case that I went (there) (J. •
S. KUNO ( 1 2) Spe.aker A:
Asoko made aruite iku koto ga dekimasu ka? there to on foot go to possible ' Is it possible to go there on foot?' a. Hai, (asoko m ade) aruite iku koto ga dekimasu. yes there to on foot go to possible ' Yes, it is possible to go (there) on foot.' b. */Hai, () iku koto ga dekimasu. yes go to possible */' Y es, it is possible to go (there) (J.' c. H ai, (J fJ dekimasu. possible yes ' Yes, (it) is possible.'
Speaker B:
( 1 2A) uses deki- 'can ', a free form, to represent potentiality. Japanese has another morpheme, re-I rare- for potentiality . l o Re-/ rare- are bound forms and must always be preceded by verb stems; ( 1 3) ikm i-
'go' 'see '
+ +
re- 'can' rare- 'can' re- 'can'
'can go' ik-em i-rure- 'can see' mi-re- 'can see'
Now, let us use this bound form in place of dekimasu 'be able to (polite . form)' of ( 1 3A). ( 1 4) Speaker A: Speaker B:
Asoko made aruite ik-e-masu ka? there to on-foot go-can-Polite Q 'Can (we) go there on foot?' a. H ai, aruite ik-e-masu. yes, on-foot go-can-Polite ' Yes, (we) can go (there) on foot.' b. Hai, () ik-e-masu. yes go-can Polite */' Yes, (we) can go (there) (J. ' c. * Hai, fJ �re-masu. can-Polite * ' Yes, (we) can.'
What is interesting here is the acceptability of ( 1 4Bb). This answer violates the Pecking Order of Deletion Principle as much as ( l 2Bb) 80
JS, vol . l , no. I
Downloaded from jos.oxfordjournals.org by guest on January 1, 2011
We can account for the above data by assuming that deki- ' c a n ' r epre sents more important information than aruite 'on foot', which, in turn, represents more important information than iku 'go'. ( 1 2Bb) is an unac ceptable answer to 0 2A) because aruite 'on foot' has been deleted in spite of the fact that iku 'go', which represents less important informa tion, has been left behind. ( 1 2Bc), on the other hand, is acceptable because the most important information dekiru 'can' has been left behind while the less important information aruite 'on foot' and iku 'go' has been deleted.
PRINCIPLES OF DISCOU RSE DELETION does because aruite 'on foot ' has been deleted while ik- ' g o ' , which re presents less important information, has been left behind. What distin guishes ( 1 2Bb) from ( 1 4Bb) is the fact that while there is no syntactic constraint in Japanese that forces the retention of iku in ( 1 2Bb), the morphology of the language forces the retention of ik- 'go' in ( 1 4Bb). If re-' 'can ' were a free form, ( 1 4Br) would have been a legitimate answer, just like ( 1 2Bd is. However, the morphology of the language says that re- cannot be used independently, and must be used in conjunction with a stem verb. Thus, the retention of re- in the answer automatically forces the retention of ik-. Therefore, the violation of the Pecking Order of Deletion Principle in ( 1 4Bb) is an 'unintentional' one, and hence, no unacceptability results. 3.2
Partial Discourse Deletion
In 1 . 2, we saw that discourses such as ( 1 5) are acceptable in spite of the fact that Speaker B's response violates the Pecking Order of Dele tion Principle: Did you buy this watch in Switzerland? Yes, I did.
The violation arises from the deletion of the relatively important information in Switzerland and the retention of the less important information I. I attributed the acceptability of ( 1 5B) to the fact that the retention of the subject is forced by the Surface Subject Constraint in English, and hence is not an 'intentional' violation; therefore, the Active and Passive Discourse-Rule Violations Principle does not mark the sentence as unacceptable. Let us investigate the behaviour of responses like ( 1 5B) in Japanese, which does not have a surface subject constraint. First, observe the following discourse: ( 1 6) Speaker A:
Speaker B:
Kimi wa oyoide kono kawa o wataru koto ga dekiru ka? you swimming this river cross to can 'Can you (lit.) cross this river by swimming? 'Can vou swim across this river?' a. Un, dekiru. yes can ' Yes, (I) can.' b. Un, boku wa dekiru. ' Yes, I can. '
On the basis o f what we have observed before, w e can assume that in Speaker A's question, dekiru 'can' represents the most important information, and oyoide 'by swimming' the second most important information. If there is no prominent emphatic stress on kimi 'yo u ' it re presents less important information than the above two. ( 1 6Ba) is the most unmarked response to the question. What happens if we retain the subject in the response, as shown in ( 1 6Bb)? ( 1 6Bb) is an acceptable JS, voi. J , no. l
81
Downloaded from jos.oxfordjournals.org by guest on January 1, 2011
( 1 5) Speaker A: Speaker B:
S. KUNO
The above phenomenon in Japanese is exactly parallel to the phenom enon in Russian that we observed in ( 1 9Bd) of the preceding section. For ease of reference, I· will repeat the relevant discourse below: ( 1 7) Speaker A: Speaker B:
Ty CIT AL moju 'knigu? you read my book ' Did you read my book?' a. Da, ja cital tvoju knigu. yes I read your book ' Yes, I read your book. ' Da, (J cital (J. b. c. */Da, (J cital tvoju knigu. d. ?Da, ja cital (J.
Recall that ( 1 7Bb) is derived by application of . the Minimal Sentential Answer Strategy: as shown in ( 1 7 A), the verb is the focus of the question, and · therefore, the. Minimal Sentential Answer Strategy does not allow any other constituents to be left behind. We noted that ( 1 9Bd) is possible only if ty 'you' in the question is either intentionally or unintentionally reinterpreted as a (secondary) focus of the question, with the resulting implication that someone the speaker knows did not read the book. I further noted that ( 1 9Bc) is considerably less acceptable than ( 1 9Bd), and attributed this to the fact that tvoju latigu 'your book ' does not oc cupy a position in the sentence that is amenable to the required interpre tation as a contrastive topic. Thus, in Russian, shifting the discourse topic is possible when the topic-switch pivot (e.g., ja ' I ' o{ ( 1 7 Bd)) 82
JS, vol. l , no. I
Downloaded from jos.oxfordjournals.org by guest on January 1, 2011
answer to the question, but it means more than what ( 1 6Ba) means. By leaving boku wa un.deleted, Speaker B is implying that there are some people who cannot swim across the river. In other words, Speaker B is using boku wa in a contrastive sense. Let us examine how this contrastive connotation arises, from the point of view of a hearer. It seems that it arises from the attempt on the part of the hearer to interpret ( l 6Bb) as a sentence that has not violated any discourse principles. Namely, if the hearer assumes that boku wa c o n ve y s infor mation that is less important than the deleted oyoide ' b y swimming', then the sentence should be unacceptable because i t violates the Pecking Order of Deletion Principle 'intentionall y '. However, if the hearer assumes that Speaker B has left boku wa u n d e leted for some good reason, namely, to emphasize the fact that the statement applies only to Speaker B himself, and not to others, then, the sentence can be accepted as one that does not involve a violation of the Pecking Order Principle. In this interpretation, boku wa , as a contrastive constit uent, would represent at least as much important information as the deleted oyoide. It seems that this is how this contrastive connotation derives in ( 1 6Bb). This sentence is not a straightforward response to ( 1 6A): Speaker A did not use kimi wa ' y ou ' contrastively. It g ives the solicited information, and says something more about unsolicited, but related information. 1 1
PRINCIPLES OF DISCOURSE DELETION occupies the sentence-initial position. In Japanese, it is possible when w h i c h m a r k s contrastive the topic-switch pivot is followed by wa , themes. We will see later in this section further examples of the use of contrastive wa for shifting the discourse topic. As we saw in ( 1 6Bb), when a deletable constituent that carries a low degree of importance is left undeleted in Japanese, an attempt is made on the part of the hearer to reinterpret it as a contrastive focus except when a very formal dialogue situation is involved as in a court room interrogation or an extremely polite conversation between a teacher and a student. For example, observe the following discourse: ( 1 8) Speaker A: Speaker B:
The above discourse is i n informal colloquial speech 'style, and therefore, it is not possible to treat it as a formal dialogue. 1 2 Among the four re sponses given above, ( 1 8Bd) is the most natural. ( 1 8Ba) is extremely redundant, and sounds as if Speaker B, instead of answering Speaker A ' s question directly, is carrying out a monologue whose content is only partially relevant to the question. ( 1 8Bb) and ( 1 8Bc), both of which involve partial deletion, are unnatural as answers to the question. Pari de and aitu ni cannot be interpreted contrastively because they are not marked with the contrastive marker wa. Similarly, observe the following discourse: ( 1 9) Speaker A: Speaker B:
( 1 9Ba)
Kimi wa ko no hon o yomimasita ka? you this book read 'Have you read this book?' a. Hai, watasi wa sono hon o yomimasita. 'Yes, I have read that book.' b. ??Hai, sono hon o yomimasita. 'Yes, (I) have read that book. ' c. Hai, yomimasita. 'Yes, (I) have read (it).'
sounds !ike either an extremely polite or sarcastic answer,
JS, vol. l , no. l
83
Downloaded from jos.oxfordjournals.org by guest on January 1, 2011
Kimi Pari de Yamada-kun ni atta? you in to met ' Did you meet Yamada in Paris?' a. Un, boku Pari de aitu ni atta yo. yes I in that-fellow met 'Yes, I met him in Paris.' b. ?Un, Pari de aitu ni atta yo. 'Yes, (I) met him in Paris.' c. ?Un, aitu ni atta yo. 'Yes, (I) met him.' d. Un, atta yo. ' Yes, (I) met (him).'
S. KUNO or, part of a monologue on the part of Speaker B . ( 1 9Bb) is extremely unnatural, and native speakers of Japanese would not normally answer ( l 9A) with this sentence. This seems to be due to the fact that sono hon o 'that book', which is not the focus of the question and could have easily been deleted, has been left undeleted. There is no interpreting this expression as a contrastive focus because ' it is not marked with wa. Hence, ( l 9Bb) is extremely difficul t to accept as an answer to ( l 9Ba ). Note that if wa is used, the dialogue · becomes acceptable. ( 20) Speaker A: Speaker B:
K im i wa kono hon o yomimasita ka? 'Have you read this book?' Hai, sono hon wa yomimasita. yes that book read ' Yes, as far as that book is concerned, I have read i t . '
The above observation perhaps explains why there are many negative sentences in Japanese in which deletable constituents show up marked with contrastive wa. For example, observe the following discourse: ( 2 1 ) Speaker A: Speaker B:
K imi wa kono hon o yomimasita ka? you this book read 'Have you read this book?' a. ??lie, sono hon o yomimasen desita. no that book read-not did ' No, I haven't read that book.' b. lie, sono hon wa yomimasen desita. ' No, (I) haven' t read that book.' c. lie, yomimasen desita. ' No, (I) haven't read (it)�'
( 2 1 Ba) is unacceptable for the reason discussed above. ( 2 1 Bc) is the most unmarked answer to the question, but ( 2 1 Bb) is also a perfectly natural and straightforwaf
·
(22) Ban Against Partial Discourse Deletion: If d i scourse deletion of recoverable constituents is to apply, apply it across the board to 84
JS, vol. l , no. !
Downloaded from jos.oxfordjournals.org by guest on January 1, 2011
Speaker B is giving a response which conveys more information than is solicited. It carries with it a negative implication to the effect that there are aiso books that he has not read. In this sense, ( 208) is not a straightforward answer to (20A), but represents Speaker B 's attempt to shift the focus of the present conversation.
PRINCIPLES OF DISCOU RSE DELETION nonfocus constituents. Nonfocus constituents which are left behind by partial discourse deletion will be reinterpreted, if possible, as representing contrastive foci. 1 4 The above constraint, like the Pecking Order of Deletion Principle, interacts with the Active and Pa�sive Discourse-Rule Violations Principle. In case the retention of a nonfocus constituent is 'unintentional' and has been forced by some syntactic constraint of the language, the ban does not apply. For example, compare the following two discourses: (23)
Speaker A:
Speaker B:
(24) Speaker A:
A s has been observed b y Haig ( 1 978), (23Bb) i s a natural and straight forward answer to the question, but ( 23Ba), is not. The latter implies that Speaker B i s going to say something more. The implication here is that he was bom in Osaka, but he was, say, brought up el sewhere. This situation. contrasts interestingly with that of (24). There is no contrastive connotation in umaremasita 'was born' in Speaker B's response. This con trast can be explained in the following way. In (23), Oosaka is the focus. Applying across-the- board discourse deletion except for this focus, as shown in (23Bb), would have produced a perfectly natural answer. Instead, Speaker B has retained umareta(no wa) 'that (I) was born ' in the answer in an apparent violation of the Ban on Partial Discourse Deletion. The hearer of this response, instead of assuming that it is an answer which has violated the Ban, attempts to interpret it as one which does not involve a violation, namely, he assumes that Speaker B has retained umareta(no wa) '(that I) was born' for a good reason. The contrastive interpretation of this expression raises the degree of importance associated with it, and this makes the sentence consistent with the Ban on Partial · Discourse Deletion. Thus, there arises the implication that Speaker B was not brought up in Osaka, but somewhere else. In contrast, in (24B) the retention of umaremasita 'was born' is necessitated by the constraint in Japanese that sentences must end with verbs. Therefore, there is no intentionality involved in the violation that this response involves of the Ban on Partial DisJS, vol. l , no. l
85
Downloaded from jos.oxfordjournals.org by guest on January 1, 2011
Speaker B:
Sensei ga oumare ni natta no wa Oosaka desita ne. that was teacher was-born (honorific) '(Am I right in assuming that) it was in Osaka that you were born?' a. Hai, umareta no wa, Oosaka desu. yes was-born that is 'Yes, (it) is Osaka that I was born in. ' b. Hai, Oosaka desu. 'Yes, (it) is Osaka. ' Sensei wa Oosaka de oumare ni natta n(o) desita ne; teacher in was-born (honorific) was '(Am I right in assuming that) you were born in Osaka?' Hai, Oosaka de umaremasita. in was-born yes 'Yes, I was born in Osaka. '
S. KUNO course Deletion. Hence, there is no need to interpret umaremasita contrastive sense.
in a
Similarly, observe the following discourse: (25) Speaker A:
Speaker B:
(25Ba) i s interpreted ei ther a s an extremely polite and accurate answer, or as a statement in which watakusi ga f u nc t ions as a contras t i ve element. 1 5 (2 5Bb) is unacceptable as an answer to the question. This is because in a normal context, there are no readily available expres sions that tabeta 'ate' can be contrasted with. It would not make sense to say that what he ate was a beefsteak with the implication that what he didn't eat was something else. The unacceptability is due to the fact that tabeta, which is recoverable and is not the focus of the question, has been left undeleted. 1 6 This fact contrasts inter estingly with the perfect naturalness of ( 25Bd), which has the same recoverable, and unimportant tabeta l e f t u ndeleted. This difference lies in the fact that while in (25Bb), there is no syntactic pressure for retaining the subject tabeta no wa 'what (I) ate', the verb-final constraint in Japanese has automatically necessitated the retention of the verb tabemasita 'ate' in (25Bd). Partial discourse deletion in Japanese is in fact a much more complex phenomenon than the data described in this section allow us to show and it will probably be necessary to set up supplementary conditions for the application of the Ban on Partial Discourse Deletion. However, I believe that the spirit of the Ban on Partial Discourse Deletion will remain central to future research in this area of discourse analysis.1 B 4.
Conclusion
In this paper, I have shown that deletion of optional constituents in a sentence in discourse follows a certain principle: namely, it proceeds 86
JS, vol. l , no. !
Downloaded from jos.oxfordjournals.org by guest on January 1, 2011
Sakuban sensei ga ano resutoran de omesiagari ni natta last-night teacher that restaurant at ate (honorific) no wa nan(i) desitakke? that what was Q 'What was it that you ate at that reastaurant last night?' a. Watakusi ga tabeta no wa bihuteki desu. I ate that beefsteak is 'It is a beefsteak that I ate. ' b . ??Tabeta n o w a bihuteki desu. ' It is a beefsteak that (I) ate . ' c. Bihuteki desu. 'It is a beefsteak. ' d . Bihuteki o tabemasita. ate 'I ate a beefsteak. '
PRINCIPLES OF DISCOURSE DELETION from less important information to more important information, and it never applies in the reverse order. I have also shown that this prin ciple, which I have named the Pecking Order of Deletion Principle, interacts with various syntactic constraints in each individual language. Examination of numerous instances of these interactions supports a previously proposed hypothesis, the Active and Passive Discourse Rule Violations Principle, that active (intentional) violations of discourse principles, but not passive (unintentional) violations thereof, result in unacceptable sentences. I have shown that the above principles apply to English, Russian and Japanese. My research shows that they apply to French, German, Swedish, Turkish, Thai, Korean, Arabic and Hebrew, as well. Therefore, it is fairly safe to assume that the Pecking Order of Deletion Principle is at least a near universal, and most likely a true language universal, and that the Active and Passive Discourse-Rule Violations Principle is also a very general principle that applies with syntactic constraints in most (if not all) languages of the world.
The results presented in this paper are only the preliminary results of an attempt to uncover r:egularities in discourse deletion phenomena. It is hoped that this paper becomes a first step along a new avenue of research on the interaction of discourse deletion and syntactic constraints. Harvard University Department of Linguistics Science Center 223 Cambridge Ma 0 2 1 38 Notes * This is an extensively revised version of Sections 1 through 5 and 1 2 of Kuno ( 1 980a). I am greaHy indebted to Linda Shumaker and JS, vol. l , no. I
87
Downloaded from jos.oxfordjournals.org by guest on January 1, 2011
The Pecking Order of Deletion Principle itself might sound too natural and matter-of-course to merit attention, but this principle, taken together with the Active and Passive Discourse-Rule Violatons Principle, give us a means until now unavailable for initiating a systema tic analysis of extremely complex and poorly understood discourse deletion phenomena. These principles have helped us to identify the Minimal Sentential Answer Strategy in Russian and the Ban on Partial Discourse Deletion in Japanese. They have also helped us to uncover parallel strategies in Russian and Japanese whereby a contrastive 'second focus' constituent is retained for the purpose of changing the discourse topic. Application of these strategies results in violation of the Minimal Sentential Answer Strategy and the Ban on Partial Discourse Deletion, but does not result in unacceptability, because the violation is functionally justifiable.
S. KUNO John Whitman, who have read earlier versions of this paper and have given me· numerous invaluable comments. I am also grateful for a number of comments I received from participants of the International Colloquium on Discourse Representation, Cleves, Germany, September 1 5- 1 8, 1 98 1 , at which a preliminary version of this paper was presented. Research represented in the paper has been supported in part by a g rant from the National Science Foundation to Harvard University (Grant No. BNS 76 8 1 7J2). 1 See can have an object missing, as in (i) We see with our eyes. However, this usage is limited to when a generic act of seeing is refer red to, and is banned when a more specific object is involved: (ij) Do you see a book on the table? */Yes, I see (J. 2 The speaker can deliberately avoid giving a complete answer to a question. For example, observe the following exchange: (i) Speaker A: Did you get your Ph.D. last year? Speaker B: Yeah, I got it alright (But a lot of good it did me on the job market.) 3 Naturally, stress shifts the focus of a sentence to the constituent that it i s attached to. Thus, if in Switzerland o f ( 1 2 A ) receives a prominent emphatic stress, it becomes the focus of the question. Thus, the following exchange would be ill-formed: Speaker A: Did you buy a watch in SWITZERL AND? Speaker B: '1/Yes, I bought one (J. Speaker A's question is paraphrasable as 'Was it in Switzerland that you bought a watch?' As predicted by the Pecking Order of Deletion Principle, deleting in Switzerland and retaining the less -. important information bought one results in unacceptability. See Kuno ( 1 982) for more details regarding the role of emphatic stress in changing foci. 4 English allows deletion of the subject of tensed verbs under limited circumstances in an extremely informal spoken or written style. For example, observe the following: (i) Speaker A: What did you do yesterday? Speaker B: Went to see the movies. However, observe the following: (ii) Speaker A: Did you go to see the movies yesterday? Speaker B: a.*/Yes, went (to see the movies). b. *Nes, did. See Kuno ( 1 982) for the exact conditions under which the subject of tensed verbs can be missing. 5 In fact, there i s another reason why (23Bb) i s totally unacceptable. To see this, let us examine the problem of the scope of too: (i) a. John, too, cried. b. John cried, too. (ia) means that John, as well as those mentioned before, cried. The sentence cannot mean that John cried as well as doing those things mentioned before. Let us represent this fact by saying that in (ia), ..•
JS, vol. l , no. I
Downloaded from jos.oxfordjournals.org by guest on January 1, 2011
88
PRINCIPLES OF DISCOURSE DELETION
JS, vol. l , no. !
89
Downloaded from jos.oxfordjournals.org by guest on January 1, 2011
the focus of too is John, and not cried. In contrast, (ib) can mean (A) John cried, as well as doing those things mentioned before, (B) John, as well as those mentioned before, cried, and (C) What happened in addition was that John cried. In other words, the too of (ib) can have either cried, John or the whole sentence as its focus. Let us now observe the following sentences: (ii) John introduced Mary, too, to Jane. It is clear that (ii) cannot mean 'John, as well as those mentioned before, introduced Mary to Jane.' It seems that the sentence can only mean that John introduced Mary, as well as those mentioned before, to Jane. Why is it that in (ib), too can have a noncontiguous constituent (i.e., John) as its focus, while in (ii), it cannot? It seems that the only way to reconcile these two apparently conflicting facts is to assume that (A) too can have as its syntactic focus only a constituent that im mediately precedes it, and that any element within a syntactic focus can function as its semantic focus. In (ii), Mary is the only constituent that immediately precedes too, and therefore, the focus of too is u n a m biguously Mary. On the other hand, in (ib), there are two constituents that immediately precede too: [cried) and [John cried]. In the latter case, either the whole sentence is taken to be the semantic focus of too ( ' What happened in addition was that John cried.') or just John ( ' J ohn, as well as those mentioned before, cried.') or just cried ( t h i s interpretation is indisti11guishable from the one obtained when too has [cried] a s its syn tactic focus). The above constraint is consistent with the generalization by Kim and Whitman ( 1 98 1 ) that the scope-bearing element is associated with the constituent it is i mmediately adjacent to. The above constraint applies to surface sentences. For example, observe the foJJowing sentences: (iii) Speaker A: Did Mary cry? Speaker B: Yes, she cried. Speaker A: Did John cry? Speaker B: a. Yes, he cried, too. b. */Yes, tJ, too. Although he cried of (iiiBa) is completely recoverable from the preceding context, it cannot be deleted. This shows that the focus modified by too must be present in surface sentences. Similarly, observe the following sentences: (iv) Speaker A: Did John hit Mary? Speaker B: Yes, he did. Speaker A: Did he hit Jane? Speaker B: a. Yes, he hit her, too. b. */Yes, he di�, too. c. Yes, he did that, too. (v) Speaker A: Did John hit Mary? Speaker B: Yes, he did. Speaker A: Did Tom hit Mary? Speaker B: a. Yes, he hit her, too. b. Yes, he did, too.
S. KUNO
•••
•••
•••
•••
_ _
90
JS, vol. 1 , no. 1
Downloaded from jos.oxfordjournals.org by guest on January 1, 2011
The unacceptability of (ivBb) as an answer to Speaker A's question is due to the fact that while the intended focus of too is either hed=Ja ne) or hit her, these cannot fall under the scope of too s y n tactically be cause they are missing in the surface sentence. The fact that (ivBc) is acceptable shows that pro-forms such as do that (do it, and do so) can readily fall under the scope of too. The contrast between (ivBb) and (ivBc) shows the crucially different behavior of deletion (or elliptical) patterns and pro-forms vis a vis the problem of the scope of too, and most likely, other quantifier-like expressions. Incidentally, (vBb) is acceptable because the semantic focus of too in this context is he ( w h ich must be stressed), which falls under the syntactic scope of too if we consider he did as a whole to be the constituent immediately preceding too. Returning to \23Bb), we can say that the sentence is unacceptable because too has lost its focus. 6 Such an interpretation becomes easy if this perfume refers not to a specific object that Speaker B possesses, but to a product brand, as in: (i) Speaker A: Did you buy or did you not buy Chane! No. in Paris? Speaker B: Yes, I bought three bottles of Chane! No. 1 � 7 I am indebted to Olga Yokoyama for many of the observations on Russian in this section. The analysis centering around the Minimal Sentential Answer Strategy, however, is mine. See Yokoyama (in prepara tion) regarding the interaction between word order and intonations. 8 Thus, the following exchange is perfectly acceptable: (i) Speaker A: Moju knigu ty CITAL ? my book you read 'As for my book, did you read it?' Speaker B: Da, tvoju knigu, cital. Speaker B 's response has a clear contrastive implication, which fits the i mplication of the question without any difficulty. 9 A question which has its focus not on the main verb, but elsewhere cannot in general be formed by simply adding lea at the end, but re quires the nominalization of the entire clause by no 'that' followed by a copula in its various surface forms (desu is a polite form of da; da is auto matically deleted before lea). The no desu pattern carries with it a pecul , iar semantic content that can be variously described with 'it is that the explanation is that , the fact is that , it happens that , you see. ' See Ktmo ( l 978b, Chapter 1 9) for details. See Kuno ( 1 980b) for the scope problem of ka. 1 0 Re- is used for consonantal stem verbs, and rare- for vocalic stem verbs. 1 1 See Kuno ( 1 982) for the same kind of focus-switching phenomenon in English. 12 There are many syntactic and morphological cues in this discou!"se that indicate that it is extremely informal. The term un 'yes' is used in place of the more formal hai. Bolru is a first person pronoun that is mainly used among close friends. Aitu 'that fellow' is a vulgar expres sion. The verb atta ' met' is an informal form that contrasts with the po-
PRINCIPLES OF DISCOURSE DELETION
JS, vol. l , no. l
91
Downloaded from jos.oxfordjournals.org by guest on January 1, 2011
lite form aimasita and the polite honorific form oaini narimasita, a�d the polite condescending form oai simasita. 1 3 In other words, it is more justifiable to imply as an excuse for not reading a particular book that one has read other books, than to imply, as a comment on the fact that one has read the particular book in question, that one has not read other books. In general, when the expected answer is in the affirmative, a negative answer often triggers an attempt, on the part of the answerer, to imply that the expected state/action has taken place, not for the particular member in question, but for some other related members. 1 4 There is a remarkable similarity between this rule and the Minimal Sentential Answer Strategy in Russian, which we examined in 2.2. I have, however, given a different name to the Japanese constraint for two reasons. First, Japanese has a minimal sentential answer strategy that uses a copulative verb. (i) Minimal Sentential Answer Strategy (Japanese): R e t a i n t h e focus of the answer, and i f i t i s not a verb, add an appropri ate form of the copula to produce a sentential answer. Thus, observe the following exchanges: (ii) Speaker A: Kimi wa kono hon o yomimasita ka? you this book read Q 'Have you read this book?' Speaker B: Hai, yomimasita. 'Yes, (I) have read (it).' Kimi wa DONO HON o yomimasita ka? (iii) Speaker A: you which book read Q ' Which book have you read?' KONO HON desu. Speaker B: book is this 'It is this book . ' I n (ii), the verb i s the focus, and therefore, the retention o f the focus alone produces a sentential answer. In (iii), the object NP is the focus of the question, and therefore, a copulative form (the present-tense polite form) is added to the focus to form a sentential answer. The second reason that I have distinguished between the Russian Minimal Sentential Answer Strategy arid the Japanese Ban on Partial Discourse Deletion is that Russian readily allows partial discourse deletion as long as the Minimal Sentential Answer Stategy has not been applied, and as long as the Pecking Order of Deletion Principle is not violated. Thus, the following exchange is perfectly acceptable: (iv) Speaker A: Ty BRAL segodnja s soboj v skoly zontik? you took today with self to school umbrella 'Did you TAKE an umbrella with you to school today?' Da, ja bral zontik. Speaker B: yes I took umbrella 'Yes, I took an umbrella.' Discourse deletion has applied only partially as to delete segodnja 'today ' , s soboj 'with self' and v skoly ' t o school' but t o retain ja and zontik. How-
S. KUNO
References
Haig, J ., 1 978: Topics in Japanese Grammar (unpublished doctoral disserta tion). Harvard U niversity, Cambridge, Mass. Kim, H .-0. & Whitman, J., 1 98 1 : Scope- bearing elements as indices of syn tactic configurationality. Presented at the Twelfth Annu al Meeting of the North Eastern Linguistics Society. Kuno, S., 1 978a: Two topics on discourse principles. Descriptive and Ap plied Linguistics: Bulletin of the ICU Summer Institue in L inguistics XI, International Christian University, Tokyo, Japan. Pp. 1 -29. Kuno, S., 1 978b: Da111 wa no Bunpoo (Grammar of Discourse). T a i s h u k a n Pub!. Co., Tokyo, Japan. Kuno, S., I 979a: Newness of information and order of deletion. Cahiers Charles V, No. 1 , L 'lnstitut d' Anglais Charles V, Paris. Pp. 2 1 1 -2 2 1 . Kuno, S., 1 979b: 0 the interaction between syntactic rules and discourse principles. In: G. Bedell, E. Kobayashi and M. Muraki (Eds.), Explorations in Linguistics - Papers in Honor of Kazuko Inoue, Kenkyusha, Tokyo. Pp. 279-304. Kuno, S., I 980a: Discourse deletion. In: S. Kuno (Ed.), Harvard Studies in Syntax and Semantics III, Department of Linguistics, Harvard University, Cambridge, Mass. Pp. 1 - 1 44. 92
JS, vol. I , no. I
Downloaded from jos.oxfordjournals.org by guest on January 1, 2011
ever, the answer is perfectly natural, without requiring contrastive 'second focus' interpretation on either ja or zontik. I 5 In main clauses, go-marked subjects cannot have contrastive connotation, but in subordinate clauses, they can. Wa- marked contrastive constituents are allowable in subordinate clauses only when those ele ments with which they are contrasted are also present in the same clauses. 16 Naturally, (25Bb) becomes a perfectly acceptable sentence if it appears in a context which justifies a contrastive interpretation of tabeta 'ate' . For example, observe the following answer (due to John Whitman) to (24A): (i) Speaker B: Tabeta no wa bihuteki da kedo, sore ga kono ate that beefsteak is but it this geri no genin zya nai. Asoko no mizu ga warukatdiarrhea 's cause is not there 's water bad ta no da. that is ' What I ate was beefsteak, but it is not the cause of this diarrhea. It was the case that the water (I drank) there was the problem.' 1 7 There are many other discourse deletion phenomena that interact with the Pecking Order of Deletion Principle, but space does not allow them to be discussed in this paper. For preliminary results con cerning these phenomena, see Kuno ( I 980b, Section 7).
PRIN CIPLES OF DISCOURSE DELETION Kuno, S., I 980b: The scope of the question and negation in some verb final languages. Proceedings of the Sixteenth Annual Meeting of Chicago L inguistic Society. Pp. 1 55 - 1 6� K uno, S., 1 982: Principles of discourse deletion. To be presented at the X lllth International Congress of Linguists, Tokyo, Japan. Yokoyama, T.O., (in preparation): Multi-Aspectual Approach to Russian Word Order.
Downloaded from jos.oxfordjournals.org by guest on January 1, 2011
JS, vol. l , no. !
93