two of the classic puzzles in the analysis of the attitudes can be pro tably viewed ..... In order to reason about the information states of Philip and Martin we must ...
Integrating diverse information resources into dialogue updates? Robin Cooper Goteborg University, S-412 98 Goteborg, Sweden
Abstract. Barwise and Perry (1983) treat meaning in terms of a relation between various information resources such as utterance events, background resource situations and visual scenes. They use this view to give an analysis of attitude relations such as believe and know . This can give us techniques for analyzing cases which need to integrate information sources based on dierent modalities. Here we point out that two of the classic puzzles in the analysis of the attitudes can be pro tably viewed from the perspective of dialogue updates involving multimodal reasoning. We sketch a proposal for a formal analysis of multimodal information states which builds on a revision of the development of Barwise and Perry's ideas in Cooper (1996) and Cooper and Ginzburg (1996). It is inspired by a recent treatment of records and dependent record types by Betarte and Tasistro (forthcoming) within the framework of Martin-Lof type theory.
1 Introduction This paper is concerned with the analysis of multimodal information states, that is, information states which integrate information from distinct resources exploiting dierent modalities. We shall argue that two of the classical philosophical puzzles which have been explored in the philosophy of language can pro tably be viewed from the perspective of dialogue update and seem crucially to involve reasoning about multimodal information states. The trouble with talking about classical philosophical puzzles in connection with research in dialogue or multimodal communication is that the examples can appear to be arcane and irrelevant to the normal cases of communication with which we are concerned. For each of the two puzzles we discuss we will attempt to give an example which suggests that solutions to the puzzle are relevant for the understanding of the kind of everyday dialogues we might nd in out dialogue corpora. The view of informationstates that we are presenting goes back to some of the earliest work in situation semantics by Barwise and Perry (1983). From the beginning Barwise and Perry emphasized the importance of integrating various information sources when they analyzed meaning in terms of a relation between a number of dierent situations. Something that has not been emphasized enough in subsequent attempts to make formal sense out of Barwise and Perry's ideas is the fact that this idea of integrating information from various situations is of relevance for the current interest for interpretation in a multi-modal setting. This paper is an attempt to point out the connection between these ideas and multi-modality. We shall rst give an intuitive sketch of the role of information states in dialogue and their relationship to the analysis of attitudes like believe and know . We shall then describe ?
1st Dec., 1997. This paper appears in Proceedings of the Second International Conference on Cooperative Multimodal Communication, Tilburg, 28-30 January 1998, ed. by H. Bunt, R.-J. Beun, T. Borghuis, L. Kievit and M. Verlinden, Department of Computational Linguistics, Tilburg. I am grateful to Bengt Nordstrom, Gustavo Betarte and other members of the type theory group at Gothenburg for discussion of dependent record types, although they are in no way responsible for what I have done (or failed to do) with them. Work on this paper has been supported by S-DIME (Swedish dialogue move engine), NUTEK/HSFR Language Technology project F305/97 and Maisie (Multimodal access to public information services) (KFB project 1997-0856-059).
the view of information states presented in Cooper (1996) and Cooper and Ginzburg (1996) in connection with the analysis of the attitudes and explain why it seems desirable to recast this view in terms of records and dependent record types as suggested in Cooper (forthcoming a) and Cooper (forthcoming b). We shall then give an account of records and dependent record types as we view them followed by an account of multimodal information states. Finally, we shall discuss what might be involved in de ning updates for such information states.
2 Information states and attitudes in dialogue The notion of information update has become standard in a number of modern semantic theories including work by Irene Heim (Heim, 1982), discourse representation theory (DRT) (Kamp and Reyle, 1993), and dynamic semantics as it has been developed in Amsterdam and elsewhere (Groenendijk and Stokhof, 1991, Dekker, 1993). For a recent paper bringing these approaches together see van Eijck and Kamp (1997). The classical view of information update as presented in these papers is of update in a text or monologue which can be represented diagramatically as in (1). (1) Information state 1 + utterance Information state 2 Here we think of the semantics of a discourse in terms of a progression of information states. Each utterance (or written contribution in the case of a text) brings about a change in the information state, normally by adding information. The meaning of a sentence is considered to be its information change potential, a mapping from information states to new information states. More recently, some researchers have begun to take more seriously the fact that natural language dialogues involve more than one agent and that there are speci c phenomena that become apparent when you consider semantics from the perspective of dialogue. Some examples of work in this area are: Ginzburg(1996a, 1996b), Dekker and van Rooy (1997). We can represent certain aspects of dialogue updates by the diagram in (2). (2)
Information state A1
connections Information state B1
+ utterance by B to A Information state A2
connections Information state B2
Here we consider two agents A and B each in their own information state. As A and B are involved in a conversation together their information states are aligned in some way (represented by the arrows labelled \connections"). Now we suppose that B says something to A. Notice that according to our diagram both A's and B's states change as a consequence of B's utterance. It is not only the case that A gains the information that B wishes to convey, but B also now has information about what B herself said and what consequences that might have for what has been established as agreed in the conversation, or at least accepted for the sake of argument at this point in the dialogue (i.e. the common ground ).
In another paper (Cooper, forthcoming b) I tried to argue that the nature of these information states and the kinds of connections that might exist between them motivated the view that the kind of information states we need for the analysis of dialogue semantics are the same as those needed for the analysis of propositional attitude verbs such as believe and know . In the largely philosophical tradition of work on the attitudes attention is focussed on attitude reports, that is, sentences that report the attitude which an agent has to a given proposition. Examples are given in (3). (3) a. Anna believes that Maria has written a story b. Maria knows that Anna can read While there are many propositional attitudes (e.g. regret , be happy (that) , assume ), attention has been focussed to a large extent on believe and know . These are attitudes which appear to be more basic than the others (for example, in order to regret something you have to believe it or know it) and which provide analytical puzzles which have occupied our minds for the better part of this century. Cooper (forthcoming b) looks at some of these puzzles from the perspective of dialogue phenomena and considers the possibility that the basic form of the puzzle has to do with dialogue update and not with the analysis of attitude reports as such. When we look at some of the oldest of philosophical chestnuts in this area such as Frege's puzzle or Kripke's story of Pierre (both of which we will describe below) from the modern perspective of our interest in dialogue and multimodal interaction, it is striking how the multimodal nature of the puzzle leaps out. I will attempt to argue that this motivates a view of information states building on Barwise and Perry's original ideas in situation semantics which allow for the integration of several dierent information sources. 2.1
Frege's puzzle
What has become known as Frege's puzzle (presented in Frege, 1892) is based on the fact that there are two names for the same heavenly body: Hesperus (\the evening star") and Phosphorus (\the morning star"). In ancient times people did not realize that the star that appeared in the morning was the same as that which appeared in the evening, in fact the planet Venus. Thus it represented an astronomical discovery to discover that Hesperus is identical with Phosphorus. Frege compares the two sentences in (4). (4) a. Hesperus is identical with Hesperus b. Hesperus is identical with Phosphorus Clearly, (4a) does not tell us anything new. It is a sentence that is trivially true, something that we might believe even if we had no idea what Hesperus is. (4b) on the other hand gives us information and it is something that we might believe to be false. In fact, the ancients used to think that it was false. But if we think about the meanings of the sentences in terms of what the words refer to, it is hard to see where the dierence lies since the names Hesperus and Phosphorus refer to the same object, the planet Venus. If we use v to represent Venus, then the logical contribution of both sentences appears to be (5). (5) v = v Frege pointed out that the problem becomes particularly acute when we consider belief reports involving these sentences. (6) a. The ancients believed that Hesperus is identical with Hesperus b. The ancients believed that Hesperus is identical with Phosphorus
(6a) is presumably true, since the ancients were presumably rational beings who believed that things were identical with themselves. (6b), on the other hand, is false since the ancients did not believe that the morning star and the evening star were the same object. How can this be, given that the logical representation of the two sentences in terms of referents appears to be identical? Now let us consider a similar example from dialogue. (7) What is Hesperus? a. # Hesperus b. ? Phosphorus c. ? Venus d. ? a planet e. ? a heavenly body Here we are considering alternative possible answers to a question. Clearly the answer (7a) is inappropriate as an answer to the question (indicated by the `#'). If such a dialogue were to occur, this response would be interpreted as a refusal to answer the question, if anything, even as an insult to the questioner which suggests that the question should never have been asked in the rst place. The remaining possible answers all have question marks in front of them. This is because the appropriateness of these answers depends very much on the information state of the person who asked the question. For example, (7b) is not an appropriate or useful answer if the questioner has never heard of Phosphorus and has no idea what it might refer to. (7c) might be a more useful answer for a questioner who is used to identifying the planet in this way. However, perhaps the questioner does not have a referent for this name either (for example, the questioner might be a child who does not know the names of the planets). In this case an answer like (7d) or (7e) would be more appropriate. This is interesting and important because (7d&e) unlike (7b&c) do not identify a unique referent providing an answer to the question. However, given the information state of the questioner they allow a useful update, an increase in the questioner's information which is not allowed by the other answers since the questioner does not have a referent for those particular names.2 When we consider these dialogue cases and the complexity of the interaction between what is an appropriate answer and the information states of the dialogue participants, we can perceive Frege's original puzzle concerning single sentences as a kind of special case of the more general and complex problem concerned with information update. What is it that determines whether an agent \has a referent" for a term or not or whether the agent believes that two terms refer to two distinct objects? Clearly, this can be related to dierent couplings between the uses of the term to which the agent has been exposed and other experiences. There is a relationship between perceiving Venus in the evening and using the term `Hesperus' and perceiving Venus in the morning and using the term `Phosphorus'. If we are to reason about the information state of an agent who has the problem related to Frege's puzzle we must essentially be reasoning about multimodality, in this case the way in which this agent relates speech events to visual events. So how often do people go around saying things like Hesperus is Phosphorus or even addressing questions such as What is Hesperus? . A striking variant of Frege's puzzle occurred a few years ago when Princess Anne was refused admission to Ascot because the man at the gate did not recognize her as Princess Anne. What can the information state of that man have been like? \She is not Princess Anne", he thought, using the pronoun she to refer to 2
Paul Decker, in personal discussion, has pointed out that the use of what in this question might be interpreted something like \what kind of" and it could be for this reason that the more general answers are preferred. Perhaps the point can be made more strongly by avoiding what and copular verbs. For example, the question Who is coming to dinner? may be most appropriately answered by Two linguists if the questioner is not aware of the names of particular linguists, but believes that they make boring conversation at dinner. Ginzburg(1996b) discusses such examples.
Princess Anne. Engaged in conversation with Princess Anne, he may even have uttered the words \You are not Princess Anne" while, of course, referring to Princess Anne with the pronoun `you'. It would not, and indeed should not, occur to us to attribute denial of the law of self-identity to the man who made this unfortunate mistake. He simply did not recognize Princess Anne, probably never having seen her before in the esh. In order to reason about this man's information state we must relate it to the visual scene presented to him while in conversation with Princess Anne and visual scenes presented to him, while perhaps seeing Princess Anne on television. 2.2
Kripke's Pierre
The second puzzle that we shall discuss was introduced by Kripke (1979). Kripke tells something like the following story. Pierre is a Frenchman, born in Paris, who does not speak English. He sees pictures in a travel brochure of a city called in French \Londres". He thinks it is beautiful. Later in his life, Pierre moves to England, learns English and lives in a city which he comes to know under the name \London". He does not realize that what he earlier saw pictures of under the name \Londres" and what he now knows under the name \London" are one and the same place. He thinks that London, the city he now lives in, is not beautiful. He believes that Londres is beautiful and that London is not beautiful. For him this is no contradiction since he does not realize that the two names have the same reference and he has entirely dierent associations with them, glossy pictures in a travel brochure as opposed to daily life in a noisy and dirty city. (Part of Kripke's point was that Pierre's dierent views of London need not rely on the fact that he has two dierent names for the city. It is enough that he has two dierent sets of associations.) Kripke's puzzle is closely related to Frege's puzzle and may even be viewed as a dierent version of the same puzzle. However, in Kripke's puzzle the belief considered is not about the identity of the two putative objects but about contradictory properties which they have. Is there a dialogue version of Kripke's puzzle? Let us engage in a conversation with Pierre to try and discover the nature of his information state. (8) a. Pierre, is London beautiful? No b. Pierre, is Londres beautiful? Yes Clearly, Pierre is answering truthfully with respect to his beliefs and we are addressing distinct beliefs of Pierre by using the dierent names for the city. It is clear in the dialogue that Pierre's answers are based on his associations with the names \London" and \Londres" and not the questioner's. This is not necessarily the case in a corresponding attitude report such as (9). (9) Pierre believes that Londres is beautiful and London is not beautiful Here \Londres" and \London" could be the names that Pierre would use or they could be names that the speaker of the sentence would use (or in a rich enough context they could be names used by another person). Since it is Pierre's uses of the names that we are interested in the dialogue version perhaps reveals the nature of the puzzle more clearly. This version of Kripke's puzzle shows us that the information states we address when we ask questions of an agent are not simply de ned in terms of facts about referents but must in addition contain information about how these referents are referred to and what kind of background associations the agent has with these referents. Kripke makes it very clear that Pierre's problem is not just linguistic. It has to do with his dierent experiences (visual or otherwise) relating to one and the same referent. If we are to reason about agents with similar problems to Pierre we must be able to reason about the relationship of their information state to experiences in several modalities.
Despite the fact that Kripke's story of Pierre has generated a large literature, it has not, over the years, come to seem a more likely story for all that. One perhaps wonders how often these kinds of misunderstandings occur in ordinary situations. Map Task dialogues (Anderson et al., 1991) are set up in such a way that they should encourage Pierre-like problems. The two participants in the maptask are each presented with slightly dierent maps which do not have precisely the same features and have at least one feature which is named dierently on the two maps. As they cannot see each other's maps they have to engage in conversation to establish if features with dierent names in fact are the same features. Consider the following extract from a dialogue (taken from the Map Task web-page http://www.hcrc.ed.ac.uk/dialogue/public maptask/) one participant has a feature labelled \elm" and the other has a feature labelled \Dutch elm". The discussion concerns, among other things, whether this is the same tree. 167 Martin: And you've ... you've got an elm tree, sort of just to the right of that. 168 Philip: line up? {fg|Oh}, right. 169 Philip: Yeah. {gg|Uh-huh}. 170 Martin: Okay. What we're going to do then is draw a diagonal line from where you are now 171 Philip: {gg|Mmhmm}. 172 Martin: down to a point that goes around the outside of the elm tree. 173 Philip: Right. I've got a popular tourist spot, {fp|erm}, between the Dutch elm and where I am just now. 174 Martin: Do you? 175 Philip: Yeah. So, go straight through that? 176 Martin: Yeah. Go straight through it. 177 Philip: Towards the Dutch elm? 178 Martin: Yeah. 179 Philip: And which side of the Dutch elm should I end up on? 180 Martin: You're going to draw a line around the outside of it, so you should be underneath the Dutch elm. 181 Philip: Right. So I'm drawing a line around the top of the Dutch elm, ending up at ... underneath it. 182 Martin: Yeah. That's right. 183 Philip: Right. Right. 184 Martin: We're going to curve around the outside of the elm. 185 Philip: {gg|Mmhmm}. 186 Martin: Okay? 187 Philip: I've done that. Yeah. 188 Martin: Now. This Dutch elm you have ... if you look across the page to the stile again 189 Philip: {gg|uh-huh}. 190 Martin: is it beneath the stile, or? 191 Philip: Slightly beneath it. The base of the Dutch elm is about maybe half an inch below the bottom of the stile. 192 Martin: {gg|uh-huh}. And away {ab|f} ... Is it about an inch away from the right-hand side of the page? 193 Philip: Yeah. {gg|Uh-huh}. 194 Martin: {fg|Oh}, it must be the same tree. I would think, anyway.
In order to reason about the information states of Philip and Martin we must take into account the relationship of their information states to the visual scenes provided their respective versions of the map and utterances of the phrases elm and Dutch elm .
3 Information states and mental states In Cooper and Ginzburg (1996) a view of information states3 is presented which is based on what we call a triadic theory of belief . We claim in that paper that variants of the triadic view are common to analyses of belief in situation semantics (Barwise and Perry 1983,1985), discourse representation theory (e.g., Kamp, 1990, Asher, 1993), work in the philosophical tradition by, for example, Crimmins (1992) and Richard (1990) and others. The standard Fregean view of attitudes like believe is dyadic4, that is, attitude relations are seen as a relation between agents and propositions as in (10) (10) believe(a; p) In triadic views of the attitudes the proposition p is replaced by two arguments. In the particular triadic view of information states that Cooper and Ginzburg proposed the propositional argument is replaced by what we will now call a mental state m and a context for that mental state c.5 The idea is that the mental state m classi es the internal state of the agent. It corresponds to what Barwise and Perry called a \frame of mind". A mental state has a set of roles which may be linked to objects in the world external to the agent. This linking of mental state roles to external objects is represented by the context c. If the context is appropriate to the mental state, that is, if it provides an object for each of the roles of the mental state, then the mental state is said to have content. Cooper and Ginzburg modelled mental states as situation theoretic types, that is, an object which, when applied to an appropriate assignment to its roles, returns a proposition. Thus the proposition p obtained by applying a mental state m to a context c can be considered as the proposition that the agent would be related to on the dyadic view of attitudes. That is, if (11a) and (11b) are both true then we can conclude (11c). (11) a. believetriadic (a; m; c) b. m(c) = p c. believedyadic (a; p) We call p the content of m with respect to c (or alternatively of the information state < m; c >). The triadic view gives us a ner grain. There may be several information states (pairings of mental states and contexts) which result in the same propositional content. There are in addition information states where the context is inappropriate (that is, does not provide appropriate values for all the roles of the mental state). Such information states do not have a content although they may correspond to perfectly rational states of mind in which agents may nd themselves. Unfortunately, the place in Cooper and Ginzburg's proposal which treated the integration of various information sources such as visual scenes (that is, exactly the place where one could begin to explore a formal treatment of multimodality) involved crucial use of the notion of restricted objects in situation theory, which has raised technical problems in the theory which have never been successfully solved. Partly for these reasons Cooper (forthcoming a) suggests In Cooper and Ginzburg (1996) we used the term \mental state" rather than \information state". This ignores for the purposes of discussion the possibility of a third argument for the time when the belief is held. 5 In Cooper and Ginzburg (1996) the terminology was dierent. What was there called a mental state corresponds quite closely, though not exactly, to what we are calling an information state and what we are calling a mental state was referred to loosely as the \internal state of the agent". 3 4
borrowing some notions from Martin-Lof type theory where the treatment of dependent types has used notions intuitively very similar to those of restrictions in situation theory but without the technical problems that situation theory ran into.6 Apart from solving some technical problems with the treatment of restrictions, dependent types have the advantage that they are inherently dynamic and using them to characterize mental states seems to make it much easier to understand how information states and information update are related.
4 Records and Dependent Record Types The ideas sketched in this section take as their starting point the work of Betarte and Tasistro (forthcoming) which develops within Martin-Lof type theory a notion of record and dependent record type which provides mathematical objects corresponding to contexts in the sense of Martin-Lof type theory. Records are sequences of elds, i.e. pairs of labels and values. Values can themselves be records. In (12) are two examples of records in the notation that we shall use.
phon = \Smith" cat = np 2 3 rel = hire 2 3 6 hirer = a 77 6 b. 66 args = 4 hired = b 5 77 4 ev-time = t 5 pol = 1 These, of course, look very similar to the kind of feature structures that are used in HPSG, for example. One important dierence is that these records are ordered, i.e. they are a sequence of elds (read from top to bottom in our representation). Record types in the formulation of Betarte and Tasistro are sequences of pairs of labels and the types to which the value belongs. Assuming some rather obvious names of types for the sake of illustration here the records in (12) could be said to be of the types in (13). : Word (13) a. phon cat : LexCat 2 3 rel : Rel 2 3 6 hirer : Ind 77 6 6 4 b. 6 args : hired : Ind 5 77 4 ev-time : Time 5 pol : Bool Of course, these types also look like the kind of typed feature-structures which are found in HPSG. One important aspect of the Martin-Lof approach is that the formalism allows you to talk about both types and objects belonging to types. An important aspect of the treatment of record types by Betarte and Tasistro is that a record with additional elds not mentioned in the type will still be of that type. Thus, for example, (14) is also a record of type (13b). (12) a.
2
3
rel = hire 3 2 7 6 hirer = a 7 6 6 77 6 hired = b 6 7 6 (14) 6 args = 4 salary = 100Kecu 5 77 7 6 5 4 ev-time = t pol = 1 6
This observation has been made by Peter Aczel in personal communication.
To this I want to introduce the idea that records can classify other objects. This is not something that is present in Betarte and Tasistro, although it can be seen perhaps as a distant relation of the view of propositions as types in Martin-Lof type theory and also the idea of classi cation introduced by Seligman (1990) and Barwise and Seligman (1997) where objects can be used to classify other objects. The idea is based on the intuitive notion that records give us a way of classifying objects. For example, (12a) is a record which might be regarded as classifying any situation which is an event of uttering the word \Smith". Similarly, the record (12b) can be regarded as classifying situations in which a hires b at time t. I will use this classi cation relation in the construction of record types. Thus records as well as record types may appear to the right of the colon in record types. To save space and improve readability I will use the traditional situation theoretic notation for infons to abbreviate records corresponding to infons7, as in (15). 2
3
rel = hire 2 3 6 hirer = a 77 6 (15) a. 66 args = 4 hired = b 5 77 4 ev-time = t 5 pol = 1 b. hhhire,a; b; tii An example of a record type which exploits the classify relation would be (16)
(16) desc-sit : hhhire,a; b; tii The record (17) would be of this type just in case hhhire,a; b; tii classi es s.
(17) desc-sit = s Record types with such classi ers can be dependent. That is, the representation of the classifying record can contain labels which occur previously in the record type or even references to other records. If r is a record and l is a label in r then we will use r:l to represent the value in the eld headed by l in r. An example of such a dependent record type is in (18), where the labels ref and res can be read as \referent" and \resource situation" respectively.
ref : Ind res : hhnamed,ref,\Smith"ii (19) would be of this type just in case a is of type Ind , i.e. an individual, and s is classi ed by hhnamed,a,\Smith"ii, i.e. a situation which supports this infon. (18)
ref = a res = s Here we see a record serving as an assignment to context roles and the dependent record type as providing a dynamic way of characterizing such contexts. Finally, we will introduce the notion of a proposition de ned in terms of a record.8 If r is a record, then 9(r) classi es True i there is some object which r classi es. For example, 9(hhhire,a; b; tii) classi es True i there is some s such that hhhire,a; b; tii classi es s. That is, the record (20a) is of type (20b) just in case hhhire,a; b; tii classi es some situation. (19)
(20) a. p = 1 b. p : 9(hhhire,a; b; tii) My idea is that all the structured objects found in situation theoretic domains should now be regarded as records, so the record actually is the infon. 8 Propositions themselves are to be regarded as a special kind of record. 7
5 Information States We regard information states as records with two elds, one for a mental state and one for a context as in (21).
mental-state = f (21) context =r Here r is a record and f is a function from records to propositions, i.e. the kind of function that we use to model mental states. An example of such a function is given in (22). 2
3
utt-time : Time 6 ev-time : Time 7 6 7 6 restr 7 : ev-time < utt-time 6 7 6 7 (22) r:6 a : Individual 7 6b 7 : Individual 6 7 4 s1 : hhnamed, a, \Smith", utt-timeii 5 s2 : hhnamed, b, \Jones", utt-timeii (9(hhhire, r.a, r.b, r.ev-timeii)) This is a function with domain records r of type (23) 2
3
utt-time : Time 6 ev-time : Time 7 6 7 6 restr 7 : ev-time < utt-time 6 7 6 7 : Individual (23) 6 a 7 6b 7 : Individual 6 7 4 s1 : hhnamed, a, \Smith", utt-timeii 5 s2 : hhnamed, b, \Jones", utt-timeii to the proposition that the value of label a in r hired the value of label b in r at the time which is the value of the label `ev-time' in r. Notice that s1 and s2 represent potentially separate information resources which provide the background information concerning the naming of the referents. In this example only naming is concerned but this possibility can be exploited to represent information sources involving dierent modalities. For example, we can model Pierre's information state as in (24). (24)
2
3
2
3
t : Time 7 6 7 6 l1 : Individual 7 6 7 6 7 6 7 6 s1 : hhnamed, l1, \Londres", tii 7 6 7 6 6 mental-state = r:6 sv1 : hhdepicted in travel brochure, l1 , tii 7 7 7 6 7 6 7 6 7 6 l2 : Individual 7 6 7 6 6 7 5 4 s2 : hhnamed, l2, \London",tii 6 7 6 7 sv : hh directly perceived by Pierre, l , t ii 2 2 6 7 6 7 ( 9 ( hh pretty, r.l , r.t ii ) ^ :9 ( hh pretty, r.l , r.t ii )) 1 2 6 7 2 3 6 7 t = 1997 6 7 6 l1 = London 7 6 7 6 7 6 7 6 s1 = s 7 6 7 6 7 6 7 6 sv1 = sv 7 6 context 7 = 6 7 6 7 6 l2 = London 7 6 7 6 7 6 7 4 s2 = s 5 4 5 sv2 = sv 0
0
6 Updating Information States
Suppose that agent A wishes to communicate to another agent B a simple fact, Anna hugged Maria, based on a visual scene (a kind of situation) that A perceived and that A decides to communicate this fact by uttering the sentence `Anna hugged Maria'. We might represent A's information state before making the utterance as (25). (Here I am including a third eld for information states with label ego for the agent who is in the information state. I make no strong claims for this being the best way to analyze de se aspects of information states.)
2
(25)
6 6 6 6 6 6 6 6 6 6 6 6 6 6 6 6 6 6 6 6 6 6 6 6 6 6 4
ego
=i
3 2
3
7 x : Ind 7 77 y : Ind 77 s1 : hhnamed, x, \Anna"ii 77 77 mental-state = r: s2 : hhnamed, y, \Maria"ii 77 77 77 ev-time : Time 77 sv : hhhug, x, y, ev-timeii 5 77 s3 : hhsee, i, sv, ev-timeii 77 9 ( hh hug, r.x, r.y, r.ev-timeii) 77 2 3 7 x =a 7 6y 7 7 = m 6 7 7 6 s1 7 7 = sitn 6 7 7 7 context = 66 s2 = sitn 77 7 6 ev-time = t 7 7 6 7 7 4 sv 5 = sitv 5 s3 = sitb 6 6 6 6 6 6 6 6 4
That is A is in a mental state which requires a context which provides a visual scene (perceived by A) where somebody named Anna hugs somebody named Maria. In the information state the mental state is provided with a context which, we will assume, provides appropriate indidivuals and situations. A now uses language to externalize this mental state. This means that various components of A's information state in (25) are linked to utterance events (a kind of situation). We will assume utterance events u1, u2 and u3 for particular utterances by A of Anna, hugged and Maria, respectively. We may represent an information state of A after her utterance as in (26). (This is, of course, a very much oversimpli ed grammatical analysis.)
(26)
2
2
2
3 3
3
phon : Name 6 u1 : 4 ref 7 6 5 7 : Ind 6 7 7 6 6 7 7 6 : hhnamed,3ref, \Anna"ii 7 6 7 6 2 res 6 7 7 6 phon : Verb 6 7 7 6 7 7 6 mental-state = r:6 u2 : 4 ref 5 : Rel 6 7 7 6 6 7 7 6 ev-time : Time 6 7 7 6 2 3 6 7 7 6 phon : Name 6 7 7 6 7 6 4 u3 : 4 ref 5 5 : Ind 7 6 7 6 res : hhnamed, ref, \Maria"ii 7 6 7 6 ( 9 ( hh r:u .ref, r:u .ref, r:u .ref, r:u .ev-time ii )) 2 1 3 2 7 6 2 3 2 3 7 6 phon = \Anna" 7 6 6 u1 = 4 ref 7 7 6 5 = a 6 7 7 6 6 7 7 6 res = sitn 6 7 6 2 37 6 7 7 6 phon = \hugged" 6 7 7 6 6 u2 = 4 ref 7 6 context 57 = hug = 6 7 7 6 6 7 7 6 ev-time = t 6 7 7 6 2 3 6 7 7 6 phon = \Maria" 6 7 7 6 4 u3 = 4 ref 5 5 5 4 =m res = sitn It is the mental state part of this record which is conveyed directly by the utterance. In fact, in a detailed version of this, it would be apparent that the mental state and the meaning of the utterance are represented by the same function from records to propositions. In an ideal world the hearer of the utterance B would have the same information state as that represented in (26). From this B would be able to construct a \non-linguistic" information state which corresponds to the information state (25), except that the value of ego would be dierent and the information about ego's perception would not necessarily be present. However, part of the advantage of this analysis of information states is that it isolates various aspects which can fail in communication. A misunderstanding of the linguistic aspects of the utterance might lead to a mismatch in the mental states. But even if the utterance is completely understood in its linguistic aspects the two agents may not embed the mental state in the same context. For example, B may make dierent assignments to the label in the context, which represents a misunderstanding. B may not have a complete assignment for all the labels in the context which can mean that a referential reading for A is an attributive reading for B. For example, B may obtain from A's utterance the information that A has access to certain resources that B does not have access to. Depending on the purposes of the information exchange it may be sucient for B to know that such resources exist without being able to identify them.
7 Conclusion In this paper we have concisdered two classical puzzles (Frege's puzzle and Kripke's Pierre puzzle) and suggested that they have relevance for the analysis of communication in a multimodal setting. We have sketched an analysis of information states which uses techniques from Martin-Lof's type theory to improve on Cooper and Ginzburg's earlier rendition of Barwise and Perry's proposal using their notion of frame of mind. According to the current proposal an information state is a pairing of a mental state and a context which assigns values to roles associated with the mental state. The mental state is modelled as a function which maps contexts to propositions. In an information state the context may or may not be appropriate to the mental state, i.e. it may or may not supply appropriate objects to all the roles of the mental state. The domain of the function modelling the mental state is characterized by a dependent type of records and it is here that dependencies between various information sources, possibly representing dierent modalities, may be encoded.
References Anderson, A. H., M. Bader, E. G. Bard, E. Boyle, G. Doherty, S. Garrod, S. Isard, J. Kowtko, J. McAllister, J. Miller, C. Sotillo, H. S. Thompson, & R. Weinert (1991) The HCRC Map Task Corpus, Language and Speech , 34, pp. 351{366. Anderson, C. Anthony and Joseph Owens, eds (1990) Propositional Attitudes: the Role of Content in Logic, Language and Mind, CSLI Publications. Asher, Nicholas (1993) Reference to Abstract Objects in Discourse, Kluwer Academic Publishers, Dordrecht. Barwise, Jon and John Perry (1983) Situations and Attitudes , MIT Press. Barwise, Jon and John Perry (1985) Shifting Situations and Shaken Attitudes, Linguistics and Philosophy, Vol. 8 pp. 399{452. Barwise, Jon and Jerry Seligman (1997) Information Flow: the Logic of Distributed Systems, Cambridge Tracts in Theoretical Computer Science, 44, Cambridge University Press. van Benthem, Johan and Alice ter Meulen, eds., (1997) Handbook of Logic and Language, North Holland, Amsterdam and MIT Press, Cambridge, Mass. Betarte, Gustavo and Alvaro Tasistro (forthcoming) Extension of Martin-Lof's type theory with record types and subtyping, Proceedings of the conference \25 Years of Constructive Type Theory" , Oxford University Press. Cooper, Robin (1996) The Attitudes in Discourse Representation Theory and Situation Semantics, in Seligman and Westerstahl (1996), pp. 137{150 Cooper, Robin (forthcoming a) Using Situations to Reason about the Interpretation of Speech Events, Current Issues in Computational Semantics, ed. by Harry Bunt and Reinhard Muskens, Kluwer Academic Publishers, Dordrecht. Cooper, Robin (forthcoming b) Information States, Attitudes and Dialogue, Proceedings of the Second Tbilisi Symposium on Language, Logic and Computation, Tbilisi, September 15{20, 1997, ed. by R. Cooper and T. Gamkrelidze. Cooper, Robin and Jonathan Ginzburg (1996) A Compositional Situation Semantics for Attitude Reports, in Seligman and Westerstahl (1996). Crimmins, Mark (1992) Talk about Beliefs, MIT Press, Cambridge, Mass. Dekker, Paul (1993) Transsentential Meditations: Ups and Downs in Dynamic Semantics, PhD thesis, ILLC/Department of Philosophy, University of Amsterdam Dekker, Paul and Robert van Rooy (1997) Intentional Identity and Information Exchange, paper presented at the Second Tbilisi Symposium on Language, Logic and Computation. van Eijck, Jan and Hans Kamp (1997) Representing Discourse in Context, in van Benthem and ter Meulen (1997). Frege, Gottlob (1892) U ber Sinn und Bedeutung, Zeitschrift fur Philosophie und philosophische Kritik, Vol. 100, pp. 25{50. Translated as On sense and reference, in Geach and Black (1952), pp. 56{78. Geach, P.T. and M. Black, eds (1952) Translations from the Philosophical Writings of Gottlob Frege, Blackwell, Oxford. Ginzburg, Jonathan (1996a) Dynamics and the Semantics of Dialogue, in Seligman and Westerstahl (1996). Ginzburg, Jonathan (1996b) Interrogatives: Questions, Facts and Dialogue in Lappin (1996). Groenendijk, Jeroen and Martin Stokhof (1991) Dynamic Predicate Logic, Linguistics and Philosophy, Vol. 14, pp. 39{100. Heim, Irene (1982) The semantics of de nite and inde nite noun phrases, PhD thesis, University of Massachusetts, Amherst. Kamp, Hans (1990) Prolegomena to a Structural Theory of Belief and other Attitudes, in Anderson and Owens (1990). Kamp, Hans and Uwe Reyle (1993) From Discourse to Logic, Kluwer, Dordrecht. Kripke, Saul (1979) A puzzle about belief, in Margalit, ed, (1979) pp. 239{83. Lappin, Shalom, ed., (1996) The Handbook of Contemporary Semantic Theory, Blackwell, Oxford. Margalit, A., ed (1979) Meaning and Use, Reidel, Dordrecht. Richard, Mark (1990) Propositional Attitudes: An Essay on Thoughts and How we Ascribe them, Cambridge University Press, Cambridge.
Seligman, Jerry (1990) Perspectives: a relativistic approach to the theory of information, Ph.D. thesis, Centre for Cognitive Science, University of Edinburgh Seligman, Jerry and Dag Westerstahl, eds (1996) Logic, Language and Computation, Vol. 1, CSLI Publications, Stanford.