Extended person-machine interface

ARTIFICIAL INTELLIGENCE

157

Extended Person-Machine Interface* Rachel Reichman-Adar D e p a r t m e n t o f Electrical E n g i n e e r i n g a n d C o m p u t e r Science, Institute for Cognitive Science, University o f California, L a Jolla, C A 92093, U . S . A .

R e c o m m e n d e d by Terry W i n o g r a d

ABSTRACT Current computer natural language systems lack discourse capabilities. They mainly treat utterances in isolation and are not able to track the coherence of extended dialogue. In general, however, human communication and task performance are not accomplished in this manner. Useful interface systems should incorporate extended conversational power. In this paper, we present an abstract computational ~ystem for extended person-machine interface. The system is theoretically motivated by an investigation into the human conversational system. The computer system incorporates many of the rules used by individuals in everyday interactions. Use of these rules enable implicit communication between conversants about their underlying models of an interaction. They function in place of explicit meta-communication about the structure of an ongoing exchange. Such communication is necessary since building compatible models of an interaction is necessary for effective interaction. Providing a comparable system of rules for the computer is then mandatory if we expect it to be able to integrate subsequent requests and exchanges to earlier ones.

1. Introduction This paper presents an abstract computational system for naturally occurring dialogues. It illustrates that effective p e r s o n - m a c h i n e interfaces have to be sensitive to the underlying rule systems governing p e r s o n - p e r s o n communication. In the paper, we explicate many aspects of the human communication system and present an abstract computational system based on the findings. Current c o m p u t e r natural language systems mainly treat utterances in isolation and function mostly as information retrievers. H u m a n communication and task performance, however, are rarely accomplished in this form. Useful *The work was supported in part by the Office of Naval Research, Department of the Navy, under Contract No. N110014-77-C-0371, and by the National Science Foundation, under Contract No. IST-81200018. Artificial Intelligence 22 (1984) 157-218

0(X)4-3702/84/$3.011 © 1984, Elsevier Science Publishers B.V. (North-Holland)

158

R, REICHMAN

interface systems, therefore, will have to be able to engage in e x t e n d e d discourse with users. T o do so necessitates c o m p u t e r s u n d e r s t a n d i n g the relation of s u b s e q u e n t u t t e r a n c e s (requests) to p r e c e d i n g ones. H u m a n c o m m u n i c a t i o n c o n t a i n s little explicit m e t a - c o m m u n i c a t i o n . Instead, m e t a - c o m m u n i c a t i o n is c o n v e y e d by a sophisticated a n d subtle use of the linguistic system. This m a i n l y non-explicit m e t a - c o m m u n i c a t i o n e n a b l e s conv e r s a n t s to discern their c o - c o n v e r s a n t s ' u n d e r l y i n g m o d e l s of an i n t e r a c t i o n . B u i l d i n g c o m p a t i b l e m o d e l s of an i n t e r a c t i o n is intrinsic to effective comm u n i c a t i o n . If the c o m p u t e r is to e n g a g e in el~ective c o m m u n i c a t i o n it too will have to be sensitive to, and aware of, these subtle linguistic m e c h a n i s m s . T h e following excerpt, t a k e n from a s i m u l a t e d i n t e r c h a n g e b e t w e e n a user and c o m p u t e r system at the C o o r d i n a t e d Science L a b o r a t o r y at the U n i v e r s i t y of Illinois, exemplifies the type of discourse p h e n o m e n a that sophisticated c o m p u t e r interface systems will have to deal with. T h e system's d a t a base is on aircraft m a l f u n c t i o n s , a n d users are asked to use the system to d e t e r m i n e m a l f u n c t i o n p a t t e r n s [ 17, 591. Excerpt 1 W~ How many hours did work centers spend on down-time in 197l'?

S: Year 1971 1971 1971 /971

Work Center Down-Time PE3 21 829 AF3 3 757 AC2 22 733 PEI 1 847

U: Can you determine if there was a high concentration of one type of malfunction for the two work centers PE3 and AC2? But first, I would like to see that above work center information for 1972. Year 1972 1972 1972 1972

Work Center Down-Time PE3 15 327 AF3 912 AC2 1 176 PEI 15 926

Do you want to see the data on PE3 and AC2? U: Yes, for 1971 and 1972.

Points of interest in Excerpt 1 (1) T h e clue words ' B u t first', in U ' s u t t e r a n c e serve as a m e t a - c o m m u n i c a t i o n to the c o m p u t e r : " p u t my first o r d e r o n hold, r e t u r n to it later, a n d set u p the following digression s e q u e n c e " . In p e r s o n - p e r s o n c o m m u n i c a t i o n we find a b u n d a n t use of such clue words replacing e x t e n d e d m e t a - c o m m u n i c a t i o n a b o u t the flow of an o n g o i n g exchange.

EXTENDED MACHINE-PERSON INTERFACE

159

(2) A computer system performing as S does must take the initiative in discourse engagement, and be able to recognize and carry out any outstanding discourse expectations placed on it. (3) In S's return to U's suspended order, S uses the unmodified noun phrase, ~the data', to refer to the 1971 "data on a high concentration of one type of m a l f u n c t i o n . . . " . This unmodified noun phrase causes no semantic ambiguity despite the intervening presentation of the 1972 data. For a computer system to recognize that this reference is unambiguous and maxim-abiding it would have to know, as S does, that after a digression the discourse state returns to the one prior to the digression. Thus, the intervening 1972 discourse is unavailable for referent resolution and the antecedent is unique. (4) The user must specifically amend his initial order to get data on 1972 included in the suspended request, despite the fact that 1972 was the last topic of discourse. This too is explained by the fact that resumption of an interrupted subject removes all intervening discourse from the discourse context considered in subsequent discussion. Natural human communication abounds with the above type of topic shifting and mingling. Our linguistic system has e m b e d d e d in it specific mechanisms which facilitate this type of interaction. Simple words like 'But, first', 'By the way', and 'Anyway' enable users to implicitly communicate to a co-conversant their underlying structural decomposition of the discourse without the need for extended discussion like, " O k a y , now, before we continue with this topic of . . . that you are talking about, I want to briefly digress onto this topic o f . . . " . Simply using the clue words, 'By the way', can replace all of this explicit and detailed meta-communication (cf. [38, 46, 54]). Even more subtle meta-communication is conveyed by a conversant's use or non-use of a pronominal. For example, given that a speaker is talking about "Mary and politics", and s/he then shifts onto the orthogonal topic of "Mary and her new job", rather than having to explicitly say to a co-conversant: " I ' m shifting the topic now to something else about Mary", the speaker can just reintroduce Mary by name in the new topic. This shift to a nonpronominal serves as a signal to the listener that a shift in discourse structure has occurred (cf. [47] for numerous examples of this form). It is on the basis of linguistic mechanisms like the above that co-conversants are able to build compatible models of an interaction. A human--computer interface system similarly will have to provide users a method for such meta-communication. It has to be sensitive to the human rule system in order to (i) understand a conversant's message and structural analysis of an exchange, and (ii) not miscue a user out of ignorance. For example, while utterances in violation of discourse pronominalization rules (like, " D o n ' t use a pronoun when shifting onto a new, orthogonal topic of discourse") are not visibly ungrammatical like the starred sentences of sen-

160

R. REICHMAN

tential linguistics, they are ungrammatical in that they miscue a listener about the underlying structure of the discourse. Thus, while the listener might be able to retrieve the referent intended, his or understanding of the referent's role in the overall discourse flow may be flawed. In general, then, the linguistic mechanisms and discourse rules embodied in our system, though vital for smooth conversational flow, are not fast-hard rules like the sentential rules of syntax. The rules reflect what usually occurs in natural conversations. Of course there are violations of such rules much as there are the following types of conversational sequences: "But what does this have to do with what we were talking about'?" Such sequences, we would claim, often result from violation of the linguistic and structural rules underlying effective, smooth conversational flow which we present in the paper. Grice states that the conversational process is composed of a set of "stages" with the execution of each 'conversational move' taking us to another stage of the discourse [18]. Describing this process, Grice notes that at any given point, it seems fair to say that "some possible conversational moves would be excluded as conversationally unsuitable" (p, 45). (}rice proposes a set of conversational maxims, and defines "'inappropriate" conversational moves as the set of moves that would violate at least one maxim. Such maxims include:

Quantity: " M a k e your contribution as informative as required (for the current purposes of the exchange) and do not make your contribution more informative than is required." Relation: "Be relevant", take into account the "different kinds and foci of relevance", "how these shift in the course of a talk-exchange", and "allow for the fact that subjects of conversation are legitimately changed". Manner: " H o w what is said should be said ~'. e.g., "'avoid obscurity of expression, aw~id ambiguity, be brief, and be orderly". These maxims capture some of the dynamic constraints placed on conversants in dialogue. Using these maxims, one might assume we could get a computer to engage in discourse with people. However, given the current status of these maxims this is impossible: (1) What is a discourse context'? (2) What does it mean for something to be in the "foci" of this discourse context, i.e., what influences does the subject of the discourse foci have on the discourse's succeeding semantic development and on the surface linguistic tools used to express developments'? (3) How does a conversational move affect the preceding discourse context? (4) How does the discourse context determine what otherwise would be an ambiguous reference or an overly informative utterance? (5) What is 'relevance', i.e., can relevance be formally defined as a set of semantic or logical relationships between utterances'? (6) Given a more exact specification of such maxims, how can they be integrated to yield a system capable of discourse modeling as an active process')

EXTENDED MACHINE-PERSON INTERFACE

161

One can consider this work to be a formalization of constructs and operations that are needed to transform Grice's general maxims into a well-defined set of rules. The rules are stated in terms of identifiable discourse structures and formalized semantic and logical relationships between the utterances contained in these structures. They are presented in the form of an 'abstract' discourse processing module: abstract, in the sense that the system is not yet implemented. The processing module is written in a predicate calculus-like language with terms like 'Instance', 'Imply', 'Infer', and 'Generalize' considered primitives of the system. In an implemented version, these terms would correspond to function calls onto other modules of the full computerized system. So, for example, the statement, " S p e a k e r Express A Instance B", in our module, should be seen as standing for at least two function calls: (i) a call onto some semantics module which, having access to some structured world knowledge data base, given A (input from the discourse module) could retrieve some proposition B (from the data base) which was an instance of it; and (ii) a call onto some syntactic module, which, having access to a lexicon and the like, given a content-filled proposition of the form ' A Instance B' (input from the discourse module after return from the semantics module) could generate a 'wellconstructed' language utterance for this message. A semantics module of this form, however, does not currently exist, though, it and other such modules needed in a full computerized natural language system are the center of much A1 natural language research. Our purpose here is not to specify a full and working computerized discourse system: rather, we wish to highlight and characterize one m a j o r module of such a future system. The module to be presented is written as a 'generator' of discourse. Generation, here, refers to characterization of some of the m a j o r structural, semantic, and logical criteria that the message to be generated must meet in order to be a 'coherent' and 'successful' continuation of an ongoing exchange. The analysis captured, however, is equally applicable to the interpretive process. Its major features include expectations, cues, and segmentation, which we believe are fundamental features of both generation and interpretation. With minor transformations, the system can be transformed into a 'parser'. In such a transformation we would replace all generation calls using the 'Express' function to calls onto 'Interpret'. Then, for example, where in 'Express' it is stated that a pronominal form can be used for a set of specific entities, within the ' I n t e r p r e t ' function it would be noted that pronominal references should be resolved in terms of this same set of identified entities. Analogously, where in 'Express' we generate a particular clue word which is to precede a forthcoming message, the recognition of that particular clue word by 'Interpret' would tell 'Interpret' to functionally parse the forthcoming utterances in a certain way; the appearance of a different clue word would result in a different functional interpretation. While we feel many aspects of our module are used by individuals in actual

162

R. RIE:;I(HMAN

discourse engagemenL we are not claiming that they use all of its particulars We are of the position, however, that a similar, analogous system is used by conversants. We posit therefore that in our module we capture some of the main features underlying a conversant's discourse processing. The module should be viewed as an approximate simulation of the types of decisions and structures used by conversants in discourse engagement.

2. Overview of the Context Space Approach A most remarkable feature of natural conversation, which we do not experience as r e m a r k a b l e since we manage it so easily, is that topics are typically developed, suspended, and resumed without need of explanation or even comment. With every utterance, a listener must grasp the relevance of what is being said to what was said b e f o r e - - w h e t h e r , for instance, the current statements are said in illustration, in support, in agreement, in disagreement. are tangential, represent a shift of topic, or represent a return to an earlier subject of discourse. Without the ability to grasp the role served by a speaker's utterance in a conversation, a listener is lost, confused, and forced to ask, "'But what does this have to do with what we were talking a b o u t ? " Such questioning illustrates a clear presupposition that operates in discourse communication: we expect speakers to select and order their c o m m e n t s so that the relation of the remarks to preceding discourse can be easily understood. Since conversations typically consist of several subconversations and are proliferated with many unexpected twists and turns (topic shiftings) without extended meta-communication, it must be the case that conversants have some underlying mechanisms which enable them to follow these twists and turns so that at any given point the conversants relate the current utterances to a same piece of preceding discourse. Effective, efficient communication occurs because participants share implicit knowledge about the shifting of conversational reference frames and even about shifting between different levels in a conversation (when, for example, talk about a subject is suspended for discussion of the conversation itself). A major aspect of discourse processing then is to track a conversation's underlying structure and to be able at all times to identify that limited preceding section of talk, which need not be a linearly preceding section, which is currently governing the generation and interpretation of what comes next. We refer to the talk which is currently being generated and that limited preceding section of talk to which it is related as a current discourse reference frame. Discourse reference frame identification requires that we be able to partition an ongoing exchange into a hierarchical organization of distinct, but related and linked, discourse units [20, 34, 35, 46, 61]. At any given point of interaction these different units have varying foreground-background influences on the ongoing exchange. Tracking these different influences, and having them comparable with a co-conversant's assignments, enables conversants to know to

EXTENDED MACHINE-PERSON INTERFACE

163

what particular preceding section of talk a co-conversant's utterances are related. Given identification of this preceding section of talk, a listener then must be able to understand the role of the speaker's utterances. Close examination of naturally occurring conversations reveals that there is a standard set of relational discourse roles that utterances play. Knowing this set of roles, and choosing those most likely to occur in the current reference frame, allows a listener to integrate a speaker's utterances with preceding ones without undue difficulty. There are two major discourse roles that an utterance can serve: it can be an embellishment or continuation of what has gone on before, or it can begin a new communicative act serving a new discourse role. An utterance of the latter form creates a shift in the discourse by establishing a different reference frame from what preceded it. These types of communicative acts are called conversational moves and discourse units are differentiated from one another by the conversational moves they respectively fulfill. A m o n g the types of conversational moves we can observe are: presenting a claim; explaining a claim; giving support to a claim; challenging a claim; shifting a topic; and resuming a preceding section of discourse (cf., [13, 23, 32, 47, 51, 60)]. In illustration of some of the basic concepts presented thus far, and how these concepts govern the choice of a speaker's surface linguistic forms, consider Excerpt 2 taken from an informal conversation between friends. Excerpt 2 A:

i. 2. 3. 4.

I think if you're going to marry s o m e o n e in the Hindu tradition, you have to--Well, you, They say you give m o n e y to the family, to the girl, but in essence you actually buy her.

B:

5. 6. 7. 8.

It's the same in the Western tradition. You know, you see these greasy fat millionaires going around with film stars, right? They've essentially bought them by their status (?money).

C:

9. No, but, there, the w o m a n is selling herself. 10. In these societies, the woman isn't selling herself, 11. her parents are selling her.

Notice that on Line 9, C uses the far deictic "there" to refer to the Western tradition last mentioned on Line 5, whereas she uses the close deictic "these" on Line 10 to refer to the Hindu tradition last mentioned on Line 1. These reference choices cannot be explained by appeal to external reference criteria since the conversation was taped in Switzerland and in terms of a locative or emotive use of deictics, Western society is the closer rather than the Hindu society. Instead, we must look at the structural organization of this discourse to

164

R. REICHMAN

understand the import of such references.

Underlying structure : Lines 1-4:

Context space C l - - t h e initiating space

Lines 5-8:

Context space C 2 - - t h e analogy space (Marked by clue words "It's the same")

Lines 9-11: Context space C 3 - - t h e challenge space (Marked by clue words "No, but") At any given point of discussion, a discourse reference frame is composed of at most two discourse units (henceforth to be called context spaces): the ground context space which is the context space containing that limited preceding section of talk in relation to which subsequent utterances are being generated; and the figure context space which is the context space being developed by the utterances currently being generated, i.e., by the current conversational move. We then have in our system the rule that close discourse deictics and pronominalization can only be used for entities in one of these two context spaces. In the above passage, there are three context spaces. On Line 9, the figure space is context space C3--as it is the context space currently being developed. The ground space, however, is C1 and not C2, though, C2 is the linearly preceding space. The ground space is the space which is most highly activated at the time when a figure is being developed. C1 is this most influential space since a shift in local topic by analogy cannot usurp the more foreground role of the initiating space of the analogy, unless the analogy is accepted. Thus, had thc analogy been accepted, the conversational topic could have legitimately shifted perhaps to the selling status of women the world o v e r - - r e m o v i n g any special emphasis from the particular situation in the Hindu tradition. The clue words, "No, but", however, signal rejection of the analogy. Not accepted, then the main topic remains with the Hindu situation. Hence, the non-correlation between deictic choice and linear positioning in the discourse flow. This example brings into relief the efficacy of a conversant's use of such simple words like "No, but `+, which can convey an amazing amount of structural discourse information to a co-conversant. "No, but", here, not only conveys rejection of the analogy, but, given this rejection, we then have the additional implicit, rule-governed information about what is the current topic of discourse and how pronouns and deictic references should be resolved. Slight changes in a clue word can significantly affect what structural and/or semantic information we pick up about the subsequent discourse flow. For example, the clue words "No, but", signal that a direct challenge is about to take place. Similarly, however, the speaker could have chosen to indirectly challenge the analogy given. In such a case, the clue words, "Yes, but", would have been mandated. The particular structural affects and resulting linguistic

EXTENDED MACHINE-PERSON INTERFACE

165

options, however, would have been the same here since both indirect and direct challenges reflect non-acceptance. 2.1. Context spaces

A context space carries along with it a lot of information other than the specific set of utterances contained in the context space. An analysis of a context space must include a specification of what roles its various parts play in the overall discourse structure, and any relevant features of the space that may interact with 'well-formedness' or 'maxim-abiding' constraints. Some of the major features of a context space are: (1) an internal representation of the set of functionally-related utterances said to lie in the context space; (2) a marker reflecting the influential status of the context space at any given point in the discourse; (3) a pointer to the preceding discourse context in relation to which this context space was developed; (4) a specification of just what type of relation was involved. To understand our notion of discourse structure it will be useful to consider its analog in the constituent structuring of single sentences. It is common knowledge that sentences can be decomposed into noun and verb phrases, each of which in turn is further decomposed into its constituent parts. A noun phrase, for example, is said to be composed of a head noun and its modifiers, while a verb phrase, is said to be composed of a main verb and its modifiers. Now, in the interpretation process of a single sentence, given that a noun has just been processed and a relative clause is begun, a listener has no difficulty in recognizing that the relative clause is to be interpreted in the context of this preceding noun. Or, having processed a verb, a listener has no difficulty in recognizing that a succeeding adverb is to be interpreted in relation to this preceding verb. Discourse utterances can similarly be parsed into constituents, which are processed in the context of co-constituents. In sentences, some constituents are independently interpretable; that is, they can appear without the accompaniment of other constituents. For example, we can have a verb without an adverb and a noun without a relative clause. Some constituents, however, necessitate the presence of other constituents for their interpretation and 'legality' of presence. We cannot, for instance, have a relative clause without having the noun that it modifies, nor can we have an adverb without the verb that it modifiesJ Similarly, in discourse, some types of constituents can appear in isolation, while the appearance and interpretation of others are contingent on the presence of some other constituent. For example, we cannot have purported

~Excluding elliptical utterances and the like.

166

R. REICHMAN

statements of support without the presence of a claim to be supported; nor can we have an explanation without something to be explained. A major aspect of discourse processing is the identification of discourse constituents and the distinguishing of dependent and non-dependent constiluents. Given a dependent constituent, we know that this should be interpreted in direct relation to its associated independent constituent, which we will treat as a superordinate constituent. As we shall see, these hierarchical relations between discourse constituents, in which some are subordinate to others, greatly aid conversants in identifying the reference frame for any given utterance in a conversation. Since discourse constituents are created by the taking of conversational moves, it should be clear that, as mentioned, shifts in the discourse reference frame result from the taking of these moves. However, though these shifts and conversational moves occur in s e q u e n c e - - t h e flow of discourse is finear--a participant's model of a conversation is not simply a listing of these sequential moves, and the structure of a discourse is not defined by this temporal sequence. Rather, one's model of a discourse consists of the packaging of the utterances of these moves into the hierarchically structured context space constituents of the discourse. Thus, we identify the structure of a discourse by the identification of its context spaces and the relations between them, A m a j o r non-dependent constituent in a discourse is what we refer to as an issue context space. Issue spaces act as 'topic-setters', and generally contain assertions that a given state-of-affairs is true or false, good or bad, possible or not possible, and so on. Such an assertion usually becomes a topic of discourse. Developing this topic consists of, for instance, supporting the assertion, challenging the assertion, discussing contingency factors of the assertion. D e v e l o p m e n t s such as these create new context spaces that are subordinates of the issue context space. Proper interpretation of a subordinate necessitates identification of the issue space of which it is a subordinate. In this sense, issue context spaces are the discourse analogs of sentential 'head nouns' and "main verbs'. They can stand alone but their 'modifiers' cannot. They tell us what the topic of discourse is; explain why new utterances are being generated; and provide the context in light of which sub-parts of the discourse can be processed.

2.2. Focused processing We have said that discourse utterances are generated and interpreted in the context of other discourse units, context spaces, and that conversants must be able at any point in a conversation to identify the reference frame for any given utterance. H o w is it that participants determine what reference f r a m e - discourse c o n t e x t - - i s currently relevant? To begin to answer this question, let us return to the analogy between

EXTENDED MACHINE-PERSON INTERFACE

167

discourse interpretation and sentence interpretation. Given the utterance, " T h e boy with the telescope saw the girl on the hill", the prepositional phrase "with the telescope" clearly modifies "the boy". Given the utterance "the boy saw the girl on the hill with the telescope", however, it is not clear which noun the prepositional phrase modifies. Nonetheless, it is accepted that the phrase modifies only one of the nouns, either the boy or the girl, and not both of them. Similarly, in discourse, thematic and relational development of a subordinate context space is in direct modification of only one other context space. Though at times it may be unclear which preceding space is being 'modified', it is quite clear to all that only one of the spaces is meant to be modified. The processing of discourse is thus highly focused. Depending on the discourse flow, different preceding context spaces are deemed the most influential at a given point in the discourse, and from the set of potential "combiners', the context space chosen is the one of most influential status. A context space's influential status distinguishes between the foreground versus background roles that it may play at different points in a discourse. Only a context space in a foreground role has a 'direct' relation to the way that succeeding discourse utterances are generated and interpreted. The focused processing of discourse is evident in the fact that not only do entire context spaces vary in influential status in the course of a conversation's development, but in addition, the different discourse entities mentioned in the set of utterances contained within a single space vary in influential status as well. Discerning the point of a speaker's utterances requires identification of those entities mentioned by the speaker which play a major role in what s/he is saying. For instance, if a speaker relates an episode that took place in a park between Dorothy and a poodle, in order to illustrate that Dorothy is funny, the park in which the episode took place is not an entity in focus for the speaker and is not part of the point being conveyed. A mentioned entity's focus level is thus often predictable from the currently relevant discourse context and the type of conversational move being processed. As mentioned earlier, certain surface linguistic forms, such as pronominal reference, in fact reflect and evidence these distinctions in different entities' focus levels and one's current reference frame (cf., [7, 19, 20, 27, 28, 29, 34, 37, 46, 56]). If a speaker's surface linguistic forms correspond to a listener's predictions of which elements are in focus and what is the current reference frame, then the conversation will flow smoothly and the listener will feel that she or he correctly understands the speaker's statements. If, however, the speaker's surface linguistic forms conflict with the listener's prediction, the listener will be at a loss, and need either to revise his or her discourse model and/or to request clarification. An important part of discourse processing, then, are the predictions listeners make about the probable surface forms in a speaker's utterances.

168

R. REICHMAN

2.3. Discourse expectations

An important mechanism that aids listeners in their focused processing of discourse is reference to a set of 'discourse expectations', which derive from the fact that the presence of one kind of discourse constituent is often predictive of constituents to follow. In the case of sentential processing, for example, we know that given the presence of a transitive verb, a noun phrase functioning as an object of this verb will follow. Similarly, in discourse, given a challenge of a conversant's claim, we expect the following utterances to reflect either an a b a n d o n m e n t of the claim in faw)r of a new one, a counterchallenge, or an offering of additional support. Given that discussion of a particular subject is interrupted by a digression, we expect the initial subject to be subsequently resumed and completed. Thus, analogous to Grice's notion that some conversational moves are 'unsuitable' at a given 'stage' of discourse, it is also the case that at particular stages of discourse, some conversational moves are 'expected' and 'most appropriate'. Not only does a conversational move set up particular expectations of what type of move will follow, it also specifies the particular preceding context space in relation to which such a subsequent conversational move will be taken and often the speaker expected to take the move as well. For instance, it is the specific context space interrupted that will be restimed; it is the specific context space serving as a challenge that will be counterchallenged (i.e., the slot fillers of the space will be referenced in the counterchallenge). In our introductory example, it is the first request which will be returned to. and the return can be done by the computer since the user has already given this instruction to the computer by his "But, first", clue words. As a conversation progresses, each conversational move thus constrains continuing options of conversational development and sets up expectations of the context relevant for interpretation of succeeding utterances. It is delineation of these conversational moves and their standardized effects on one's model of the underlying discourse structure (via reassignment of focus levels and status assignments and the setting up of discourse expectations) that drives our abstract computational module for discourse generation and interpretation. We assume that these same conventions are driving the discourse processing of all individuals in our shared language community. 2.4. Conversational moves

We have said that interpreting discourse requires a listener to understand the function that utterances serve in the development of discourse topics. These different functions constitute distinct types of conversational moves, and discourse development is a weave of these various moves. Much of the implicit knowledge speakers and listeners share is knowledge of the particular corn-

EXTENDED MACHINE-PERSON INTERFACE

169

p o n e n t s of various c o n v e r s a t i o n a l m o v e s - - w h a t kinds of u t t e r a n c e s must be m a d e in o r d e r to fulfill various discourse functions. Such implicit k n o w l e d g e then guides a listener's i n f e r e n t i a l process for 'filling in the b l a n k s ' . In acc o r d a n c e with their discourse functions, c o n v e r s a t i o n a l m o v e s have specific criteria for d e v e l o p m e n t . It is k n o w l e d g e of these criteria that e n a b l e listeners to d e l i n e a t e the b o u n d a r i e s of context space structures, a n d to identify a c u r r e n t r e f e r e n c e frame in relation to which u t t e r a n c e s are to be i n t e r p r e t e d . Formally, c o n v e r s a t i o n a l moves r e p r e s e n t the various kinds of s e m a n t i c a n d logical relations that can hold b e t w e e n u t t e r a n c e s of a discourse. T h e form a l i z a t i o n s are high level a n d abstract, a n d can be c o m p a r e d in their c o n t e n t free form to the f o r m a l i z a t i o n s of a formal logic system. T h e c h a r a c t e r i z a t i o n s include explication of the implicit c o n n e c t i o n s b e t w e e n u t t e r a n c e s s t a n d i n g in a particular type of relation. For example, let's briefly consider the system's characterization of a support c o n v e r s a t i o n a l m o v e : C o m m o n forms of s u p p o r t entail citing an authority for the claim u n d e r discussion, a n d / o r stating some state-of-affairs F such that there exists some generic s u p p o r t - p r i n c i p l e P of the form " A Implies B", such that F is an i n s t a n c e of A (or not B), C is an i n s t a n c e of B (or not A), a n d using some rule of i n f e r e n c e such as M o d u s - P o n e n s (or M o d u s - T o l l e n s ) , C follows from F. Consider, for example, R ' s s u p p o r t u t t e r a n c e s on Lines 1-14 of the following excerpt t a k e n from n a t u r a l discourse. 2 Excerpt 3

R:

1. Except however, John and 1 just saw this two hour TV 2. show

M: 3. Uh hum, R:

4. 5. 6. 7.

where they showed--it was an excellent French TV documentary--and they showed that, in fact, the aggressive nature of the child is not really that much influenced by his environment.

M: 8. How did they show that? R:

9. They showed that by filming kids in kindergarten,

M: 10. Uh hum, R: 11. showing his behavior among other children, M: 12. And then? R: 13. and showed him ten years later acting the same way, 14. towards, urn, 2In her support, R relies on the accepted procedure that one can generalize from a study sampling random individuals of a population. Of course, one possible form of challenge against such a support is for an opponent to demand evidence that the subjects of the study are really typical of the general case.

R. REICHMAN

170

R's authority citation includes the source of the claim (a study), rcasons for believing this source (excellent French TV documentary), and her access to this source (seeing a presentation on TV). Her substantive support can be analyzed as follows (using the abbreviations. C (claim), F (fact of support), P (principle). M (mappings between P and C), M' (mappings between P and F), and I.R (logic rule)): - ( 7 = Not(The aggressive nature of the child is influenced by his environment)--Not(A); - F = Not(A child's aggressive behavior in kindergarten changed over ten years}--Not(B); - P = If (A) a person's social interactive behavior is environmentally influenced. Then (B) that person's social interactive behavior will change over time: - M = ((person's social interactive behavior, child's aggressive behavior)) - M' = ((person's social interactive behavior, child's aggressive behavior), (over time, ten year duration period from kindergarten)); - LR - Modus-Tollens. In our system we have defined a set of subcategorizations for the support move. For example, corresponding to the above type of Modus-Tollens support move, we have the following characterization: "Speaker Express F S.T. P * - Infer some generic principle of the form If A Then B S.T. F Instance Not(ThenPart(P)) & HEAD-CCS.Claim.State-of-Affairs Instance IfPart(P)" HEAD-CCS points to the current ground context space; Claim is a slot of this ground context space; and State-of-Affairs is a subslot of Claim and it contains the initial claim to be supported.

Notice that our system merely specifies the characterization. Actual generation of F necessitates a very sophisticated semantics module capable of retrieving from a data base an F and P which meet the requirements of this characterization. Now, corresponding to the different types of conversational moves specified in the system, we have also characterized different types of context spaces. For example, we differentiate between issue and support context spaces by characterizing support spaces as having (among others) the following slots: Source, Method, Credential, Access, F, P, M, and M'. Thus, the system treats implicit c o m p o n e n t s of a m o v e as important, and as much a part of the discourse, as those verbally expressed. These c o m p o n e n t s are necessary parts enabling a context s p a c e - - v i a its contained u t t e r a n c e s - - t o fulfill a given discourse role.

EXTENDED MACHINE-PERSON INTERFACE

171

3. Some Specifics of the Context Space Theory 3.1. Context spaces As we have noted, the categorizations of context space types and the component slots specified for different types mirror the different units and p h e n o m e n a found in a discourse world. For example, corresponding to the fact that the presence of a certain type of discourse constituent is often dependent on the presence of some other independent constituent, distinct spaces have been defined for independent and dependent constituents, such that dependent constituent spaces fill slots of independent constituent spaces. Additionally, corresponding to the fact that the analysis of some types of conversational moves entails the explication of standardized implicit components, a context space containing the utterances serving such a move will have slots to contain such implicit components. All context spaces have the following slots, which are obligatorily filled (with possible values depending on the type of space) for the completion of a context space: (1) Type: This slot specifies the context space's category name. (There are ten types of context spaces described below.) (2) Derivation: This slot specifies whether the substantive claims within the context space were explicitly stated by the speaker, or whether they were implicit and thus inferred by the system. (3) Goal: This slot identifies the function s e r v e d - - t h e conversational move p e r f o r m e d - - v i a the context space (e.g., support, clarification). (4) Contextual Function: This slot is a structured slot consisting of the following two subcomponents: - M e t h o d : This slot contains the particular method used to perform the conversational move of the space (e.g., Flat-Denial, Modus-Ponens, Analogy). - C o - r e l a t o r : This slot contains specific reference to the context space in relation to which the conversational move is being made. (5) Speakers: This slot contains a list of persons who have generated the utterances lying in the context space. (6) Status: This slot contains specification of the foreground/background role of the context space at a given point in the discourse. (7) Focus: This slot specifies the influential status of individual elements within the context space. It is composed of four subslots, each of which contains constituents of the utterances contained in the space: - H i g h : This slot notes the constituent in 'high' focus. -Medium: This slot lists all constituents in ' m e d i u m ' focus. - L o w : This slot lists all constituents in 'low' focus. - Z e r o : This slot lists all constituents of 'zero' focus. Before describing the major types of context spaces and the particular

172

R. REICHMAN

additional slots specific to them, it will be useful to understand mort" precisely the information contained in the slots we have called 'Status' and 'Focus'.

3.2. Status assignments We have stated that at any given point a discourse reference f r a m e - - r e l e v a n t discourse context--is composed of at most two context spaces, and further that it is in direct relation to these two context spaces that succeeding utterances arc generated and interpreted. One should not infer, however, that once in the background, a context space cannot reappear in the foreground. In addition, though a context space may be in the background, this is not synonymous with its having no influence on the continuing flow of the conversation. For example, unexpected digressions may occur in a discourse, suspending the instantiation of a particular discourse expectation or completion of a context space under discussion. While the digression temporarily usurps the foreground role, and in the process relegates the interrupted context space to the background, the succeeding conversational development is still constrained and dependent upon the now background context space. For instance, it would be inappropriate to begin a new Issue context space after the digression context space has been completed. To use Grice's terms, such a conversational move would be "unsuitable at this stage of the discourse". Instead, what is d e m a n d e d is that one return to, and resume discussion of, the interrupted background context space. Mirroring the different roles that a context space may play, all context spaces have an associated Status slot, which is used to: (i) determine whether or not this context space will be accessed in our focused processing of succeeding discourse utterances; (ii) determine possible effects succeeding conversational development can have on this context space, i.e., reflect any outstanding discourse expectations which refer to it; (iii) determine the type of referring expressions to elements of the utterances contained in this context space that would be both unambiguous and sufficiently informative: (iv) determine the clue words needed for re-activation. The following seven types of conversational influence, which we feel is the closed set of status' needed for any discourse, have been distinguished: (1) Active: The context space in which the utterances currently being stated are placed. There can only be one active context space at a given point in the conversation. (2) Controlling: The context space in direct relation to which an Active context space is being d e v e l o p e d - - a l l semantic and logical generation rules can only access slot fillers of this space. There can only be one controlling context space at a given point in the conversation. (3) Controlling*: A previously Controlling context space whose Controlling

EXTENDED MACHINE-PERSON INTERFACE

173

status is temporarily usurped by subordination to the Active context space which was created by an analogy or by an interruption. Since analogy and interruption both only cause a temporary or local shift in topic with the obligation of resuming the initiating topic, the Controlling context space must be uniquely identified so that it can be reassigned a Controlling status when its subordinate space is re-activated. A Controlling* status gives us this unique identification. (4) Open: A previously Active context space that was interrupted before completion of its corresponding conversational move. (5) Generating: A context space in a background role for which we have some outstanding discourse expectations of reforegrounding: reforegrounding will depend on the subsequent discourse flow. For example, in debates where argumentation of a given claim leads to some subargumentation, the outcome of the subargument will affect whether or not we resume discussion of the initiating claim. A context space containing an initiating claim of some subargument then has a Generating status value while the subargument is argued. (6) Closed: A context space in a background role for which there is no particular expectation of resumption--all criteria for completion of the conversational move it serves have been met. (7) Superceded: A context space in a background role for which we have no expectation of resumption but whose topic of discourse has not been completed. Rather, the topic has been modified and is being developed in a subsequent context space. Modification usually occurs by replacing its claim slot filler, in another space, by a claim which is either a 'finer-restatement' or a 'further-generalization' of the initial claim. This too most often occurs in debates (but is not limited to them) where an opponent's challenge of one's claim can lead one to giving up on the scope of the claim, but not to give up on the claim in its entirety. Of these seven, the two most prominent are the Controlling and Active context spaces; since, at any given point in a conversation, utterances are generated and interpreted only in direct relation to the utterances contained in these two context spaces. 3.3. Focus levels As noted, the focus level slot is used to specify the influential status of entities mentioned in a speaker's utterances which are contained within a context space. We have mentioned, as well, that a particular entity's focus level governs one's choice of surface form to refer to that entity. Corresponding to the different means we have to reference elements, we have different possible focus level assignments: a pronominal reference reflects a high focus level; reference by name, a medium focus level; reference by description, a low focus level; and implicit reference, a zero focus level. In general, the discourse role played by the conversational move of which a discourse entity is a part governs its initial

174

R. RE1CHMAN

focus level assignment; and the conversational moves following its mention, through their effects on the status of the space in which it is contained, governs its updated reassignments. (For a complete specification of these rules see [47]) In illustration of the basic slot-value description of a context space, consider the essential specification for context space C2 in the H i n d u - W e s t conversation presented earlier (Excerpt 2). Context space C2: Lines 5-8.

"It's the same in the Western tradition. You know, you see these greasy fat millionaires going around with the film stars, right? They've essentially bought them by their status."

Type: Issue. Derivation: Explicit. Goal: Generalization. Contextual Function: Method: Analogy. Co-relator: C1 (the current controlling context space). Speakers: B. Status: Active. Focus: It [buying and selling of women] = high focus. Western tradition - medium focus. 3 . 4 . T y p e s of c o n t e x t s p a c e s

There are basically two m a j o r types of context spaces composing a discourse world: issue and non-issue spaces. Issue spaces, we have said, are the 'topic setters' of discourse, the independent constituents in light of which all subordinate context spaces are generated and interpreted. 3.4.1. Issue context spaces All issue spaces have the following three additional slots: (1) Claim: This is a structured slot composed of two parts: - State-of-Affairs: This slot specifies the proposition of the utterances in the context space. - ' T y p e ' Predicate: This slot specifies the modality of the proposition (true, not true; good, bad; etc.). (2) Topic: This slot specifies the topic being set by the issue space (generally a generic formulation of the state of affairs specified in the space). (3) Support CS: This slot contains a list of all the context spaces developed in support of the claim of the issue space. In an issue context space, the truth or necessity, or the goodness, or the appropriateness of a particular state-of-affairs is discussed. Mirroring these different types of assertions about a state-of-affairs, we have defined three types of issue spaces: In an epistemic issue context space, the truth or necessity of a given state of

EXTENDED MACHINE-PERSON INTERFACE

175

affairs is discussed. The 'type' of predicate associated with the claim is an 'epistemic' predicate, the values of which are: true, not true, necessary, or not necessary. In an evaluative issue context space, the goodness of a given state-of-affairs is discussed. The type of predicate associated with the claim is an 'evaluative' predicate, the possible values of which are pairs consisting of a particular evaluative adjective and a corresponding positive or negative measure. For example, (crazy, negative) is a possible value for this slot. If a speaker negatively evaluates a state-of-affairs under discussion, the discourse system usually infers that the speaker is advocating a change in the state of affairs. In this case we also provide an 'Advocated-Inference' slot to note this inferred advocation of a change. In a deontic issue context space, a speaker questions the appropriateness or fairness of a given state of affairs. The type of predicate associated with the claim is 'deontic', the possible values of which are "Why should?" or "Why shouldn't" Ix be the case]. These three types of context spaces are not subcategories of issue spaces but exist at the same (highest) level of categorization. Any one of these three types can however be further subdivided into debative or non-debative issue spaces. Non-debative issue spaces are ordinary issue context spaces; debative issue spaces have the following four additional slot definitions: (1) Protagonists: This slot replaces the usual 'Speakers' slot associated with a context space, and it contains a list of conversants advocating the contents of the claim slot. (2) Antagonists: This slot contains a list of conversants attacking the assertion advocated by the protagonists. (3) CounterClaims: This slot contains a list of context spaces developed in challenge to the claim of the protagonists. (4) CounterSupports: This slot contains a list of context spaces developed in challenge of the supports of the claim under attack. In addition to the specific slots defining particular types of issue spaces, spaces developed by certain methods may require additional slots to reflect the components of those methods. For instance, the module's explication of any conversational move using an analogy as its method of development entails inferring implicit components of an analogy (such as mappings between the two domains under discussion); thus any issue context space developed by analogy will have the following additional slots: (1) Abstract: This slot contains the generic proposition P of which the initiating and analogous claims are instances, and also a listing of all generic propositions relevant to the conversational move served by the analogy (explicitly stated or implicit ones which are causally related to stated propositions). (2) Relations: This slot contains a list of the relations used in the propositions of the Abstract slot.

176

R. R E I C H M A N

(3) Mappings: This slot contains a list of mappings between constituents of the generic proposition P onto the constituents of the propositions in the analogous and initiating context spaces. 3.4.2. Non-issue context spaces The second category of context spaces encompasses comment, narrative, and two types of support context spaces. Comment and narrative context spaces: In our context space partitioning of the following excerpt, J's utterances (Lines 1-7) would form a narrative space, while both R and M's utterances serve as comments on this narrative: J:

1. 2. 3. 4. 5. 6. 7.

This happened last--in September of last year. Uh, a guy, another prisoner in death row, he was, uh, he was about to die o f - - h e was to be executed within two days, okay? This is in Alabama, it's a Southern state. And, uh, he would refuse to eat or something like that. Well, I think it was a little longer, but, so they force fed him until--up to his execution.

R: 8. Isn't that stupid? M: 9. That's disgusting.

The distinction between c o m m e n t and issue context spaces reflects the fact that a c o m m e n t space is not a topic setter and does not necessarily introduce a proposition to be evaluated. If a speaker's c o m m e n t does b e c o m e the next issue of discussion, then the c o m m e n t is put into the claim slot of a new issue space. If the c o m m e n t does not b e c o m e a matter of discussion, it is appended to an 'Old C o m m e n t s ' list in a currently Active context space. A c o m m e n t context space is a t e m p o r a r y subconstituent of another kind of space that is currently Active. The following additional slots are specific to narrative spaces: (1) Orientation Section: The opening of a narrative usually consists of an introduction to the participants in the narrative and the time and place at which the episode occurred. Corresponding to this the Orientation slot of a narrative space is composed of the following subslots: Actors: This slot contains a list of lists, where each list is composed of two parts: an actor identifier and a list of predications specified about this actor. Time: This slot specifies the time of the episode and any particulars stated about this time period (e.g., "It was a ghastly period in world history"). - L o c a t i o n : This slot specifies the location of the episode and any predications given this location (e.g., " I t ' s a Southern state"). (2) 0Events: A list of events that occurred in the episode. The events are usually ordered in linear time and causality. (3) HEvents: A list of contrastive events that could have, but did not, occur in the episode. -

-

EXTENDED MACHINE-PERSON INTERFACE

177

Support context spaces: There are two major types of support context spaces: Narrative-support and Non-narrative-support spaces. The former consists of the telling of a story in which one event, led up to by preceding told events, is an instance of the claim being supported. The latter consists of the uttering of a particular state of affairs which, using some generic principle of support and some formal rule of inference, leads one to conclude the claim being supported. Both forms of support can be accompanied by naming some authority for the information cited; or an authority citation may be the sole support for a claim, and constitute a support conversational move in itself. All support context spaces have the following special slots: (1) Support-Fact: This slot either contains some propositional representation of the event of a 'Support-Narrative' that is an instance of the claim being supported, or, for non-narrative-supports, it contains a propositional representation of the state-of-affairs that implies (by reliance on some generic principle and rule of inference) the validity of the claim being supported. (2) Mappings: For non-narrative support context spaces this slot will be filled by the mappings of elements of the generic principle of support onto the support statements, and the mappings of the elements of the principle onto the claim being supported. For narrative-support context spaces it will be filled by the mappings of the elements of the claim being supported and its instance of support. (3) Authority: This is a structured slot consisting of the following four parts: -Source: This slot specifies the person or group cited as an authority. -Method: This slot specifies the origin of the authority's information. -Credentials: This slot specifies the authority's credentials. -Access: This slot specifies the speaker's access to the authority's information (e.g., "saw on TV"). An analysis of substantive non-narrative support utterances entails the explication of many implicit components. As illustrated previously, the discourse system analyzes such support moves by identifying some principle of support and setting up mappings between this principle, the stated fact of support, and the claim being supported. 3 Reflecting this analysis, context spaces containing utterances filling a non-narrative support discourse role have the following additional slot definitions: (1) Support-Statement: This slot is a propositional representation of the actual utterances stated in support. The slot's value will usually be identical either to the slot value of 'Support-Fact' or 'Principle-of-Support'.

3Actually, the formalization of this move allows for the fact that sometimes a conversant will explicitly state the generic principle underlying the support while leaving the particular state-ofaffairs unstated and implicit. For example, it is quite acceptable for a speaker to support a claim that "Alaska is depressing" by saying that "All cold places are depressing", leaving one to infer (or retrieve from knowledge) that "Alaska is a cold place".

178

R. REICHMAN

(2) Principle-of-Support: This slot contains the generic principle underlying the support, and is subdivided into 'IF' and 'Then' parts. In exemplification of a complete context space specification for a portion of discourse, consider the specification of the genetics-environment dialogue presented earlier (Excerpt 3), which contains an Issue context space and a support context space related to it: Context space C1 : Type: Epistemic Debative Issue. Derivation: Explicit. Goal: Challenge. Contextual function: Method: Exclusive-Or (Either A or B), Co-relator: [precedes excerpt]. Protagonists: R Antagonists: Empty [filled later in the discourse], Status: Controlling. Focus: Aggressive nature of the child = high focus. Environment = low focus. Claim: State-of-Affairs: The aggressive nature of the child is influenced by his environment. 'Type' Predicate: Not True. Topic: The influence of the environment on a childs's behavior, Support CS: C2. CounterClaims: Empty [not later filled]. CounterSupports: Empty [later filled]. Context space C2. Type: Non-Narrative Support. Derivation: Explicit. Goal: Support. Contextual Function: Method: Modus-Tollens. Co-relator: C1. Speakers: R. Status: Active. Focus: Child/kids = high focus. His behavior = medium focus. TV shows/documentary = low focus, Kindergarten/other children = low focus. Support-fact: A child's social interactive behavior ten years after kindergarten was identical to his social interactive behavior in kindergarten. Authority: Source: Study. Method:Investigative filming of kids over time. Credentials: Excellent French TV documentary, Access: R & J watched on TV. Support-Statement: A child's interactive behavior ten years after kindergarten was identical to his social interactive behavior in kindergarten,

EXTENDED MACHINE-PERSON INTERFACE

179

Principle-of-Support: IfPart: One's behavior is influenced by one's environment. ThenPart: One's behavior will change over time. Mappings: IfMappings: ((child, one)(aggressive nature, behavior)). ThenMappings: ((child, one) (ten year duration period from kindergarten, over time)).

In addition to the context space types we have so far described, other discourse phenomena--descriptions, small talk, question and answer pairs, for i n s t a n c e - - m a y ultimately be incorporated into context space structures for the purposes of discourse analysis. The limited set of context space types thus far characterized by and incorporated into our computational system, however, have been found to be adequate for describing a wide range of discourse forms--narratives, debates, therapy sessions, instruction, etc. 3.5. C o n v e r s a t i o n a l

moves

To engage in maxim-abiding discourse it is not sufficient to merely know the standard set of relational discourse roles that appear in discourse. Equally as important is knowing how, when, and where to use these roles. As noted by Grice, at particular 'stages' of discourse some moves are 'suitable' while others are not. In our characterization of the standard set of conversational moves making up many discourse forms we therefore have also characterized, for each move, its requisite discourse environment. We call these characterizations the preconditions of a conversational move. For example, given the conversational move of returning to an initiating claim of some analogy, a precondition for performing this m o v e in discourse is that there have been some preceding analogy. Grice also observed that the execution of a conversational move takes us to 'another stage' of discourse. This results from the effects that a conversational move has on the underlying discourse structure. For example, the effect of interrupting development of an Active context space is to set up the discourse expectation that after the interruption there will be a return and completion of the interrupted space. Or, the effect of citing a second support for some claim is to close discussion of the first support and render its context space a background discourse role. All conversational moves characterized in our system then have an associated set of effects that cause updating of our discourse constructs whenever they are performed. For instance, in the interruption case, the context space interrupted is assigned an Open status assignment and it is taken out of the reference frame to be used during generation and interpretation of the digression context space. Later in the paper we present a simulation of the discourse module processing a piece of discourse. In this simulation we describe and explain in some detail our notion of a conversational m o v e ' s preconditions and effects. For the

180

TABI.E

R

REI('HMAN

1

Conversational

move

Suppor~

('lue word(s) Because . . .

Like...

:

when. So ~. . lake

Restatement and/or conclusion o f point being supported Interruption Return to previously interrupted context space Indirect challenge Direct challenge Subargument concession Prior logical abstraction Further development

Incidentally

Anyway...

; By the way...

In a n y c a s e . . .

Yes/Right b u t . . . : Except, however (No) B u t . . . All right/Okay b u t . . . But look/listen/you s e e . . . Now

.

m o m e n t , therefore, we simply present a listing (Table 1) of some of the major conversational m o v e categories identified in the work with their associated clue word signals. (For a complete listing of these categories, their individual subcategorizations, and all associated preconditions and effects, see [46, 47].) 4. The Discourse Module 4.1. ATN formalism Our discourse module is based on an extensive study of naturally occurring dialogues. What we present, however, is not simply a snapshot analysis of a given discourse's underlying structure. Rather, our aim is to present the outline of a data processing system which characterizes major aspects of discourse engagement as an ongoing, dynamic process between co-conversants. Our system characterizations have therefore been incorporated into the A T N formalism [63]. An A T N is simultaneously a 'grammatical', 'static' description of a set of rules underlying a language, and at the same time a 'dynamic', "process model' for actual generation/interpretation of this language. In sentence grammars, the basic unit of analysis is the word, and complete transition of a sentence A T N entails the production or interpretation of a single English sentence. Now, in the discourse system, the unit of analysis is a context space, and this may contain a set of utterances, a single utterance, or a portion of one. Transition of the discourse A T N entails the production or parsing of a single conversational move. Arc tests specify the developmental options available at any point in a discourse--what moves are viable in the given discourse context. A complete discourse is facilitated by many cycles through the different network paths. The following briefly summarizes the context space module's use of A T N states, arcs, registers, and actions.

EXTENDED MACHINE-PERSON INTERFACE

181

(1) States represent the places where we test to see if the current discourse environment facilitates a given thematic development, and where we push to subprocesses that must take place before the current conversational move can be continued or completed. Being in a given state roughly characterizes the decisions and actions performed thus far on a single conversational move. (2) Arcs represent the conversational moves available in a discourse or the c o m p o n e n t actions of such moves. Just as in sentence A T N ' s we have arcs for the successive constituents of a phrase, here too we have arcs for the constituent utterances and decision moves of a discourse. (3) Registers are used for two purposes: (i) they represent the basic elements (slots) of a 'deep structure' analysis of conversational material, much as the registers used in a sentence A T N represent the basic constituents into which an utterance is broken down; and (ii) they are used to track dynamic aspects of the discourse (such as its focus) and to constrain the traversal of paths through the A T N at various points in the discourse. (The register contents often denote specifics of the preceding discourse environment and thus influence the appropriateness of subsequent conversational moves.) (4) Actions on arcs fall into two categories: (i) setting registers and constructing and updating context space constructs; and (ii) generating the clue words associated with a given conversational move, and characterizing the utterances filling its m o d e of development. To engage in maxim-abiding discourse, where utterances are semantically and organizationally connected to one another, discourse tracking is mandatory: preceding conversational development constrains and sets up predictions for subsequent development. Each cycle through the network corresponds to a single conversational move. Within a single cycle then, the module records all relevant effects of the move being processed. In this manner, on a subsequent traversal, it can access these noted effects and thereby correctly model the different dependencies between preceding and succeeding discourse utterances.

4.1.1. Registers A T N s use registers as a means of passing values between states, and as a means of blocking inappropriate arc transitions. On any arc between two states, any n u m b e r of register tests may be made to determine whether the current context warrants or requires such a transition. The context space module makes extensive use of registers for transition testing. In general, these tests are more complex than those found in sentence A T N ' s and they usually include testing of logical relationships between parts of the utterances being generated and the slot fillers of context spaces in a current relevant discourse context. We have found the following set of registers to suffice for tracking the information needed by a discourse module. (1) The expectation register: The expectation register is a structured register

182

R. REICHMAN

that contains the following four parts: - Function: A state in the ATN which corresponds to a conversational move category. -Speaker: The conversational participant expected to carry out this discourse move. - Context: That context space which will most likely be active or controlling during the processing of this conversational move. Associated-Constraints: Specifies the constraints that will have to be met when this expectation is fulfilled. (2) The Expectation-List register: The Expectation-List register contains a list of yet unfulfilled discourse expectations. (3) The CCS register: The CCS register points to the Active context space in the current discourse environment, which is the context space under current (or last) discussion. (4) The H E A D - C C S register: The H E A D - C C S points to the Controlling context space in the current discourse environment, i.e., to the context space in direct relation to which the Active context is being (or was) developed. (5) The Discourse-Mode register: We distinguish between two discourse modes: the discussion mode and the debate mode. The distinction is made because certain moves are more appropriate in one mode rather than in the other. (6) The Speaker register: The Speaker register denotes the conversant whose conversational move is currently being processed. (7) The Participant-List register: The Participant-List register contains the names of all the conversants of the discussion except the one currently holding the floor (i.e., except for the conversant named in the Speaker register). It is updated each time a new speaker is chosen. (8) The Sides register: The Sides register contains lists of the participants on each side of an argument during a debate. The Sides register is continually updated and allows for participants to change sides and play a 'devil's advocate' role. (9) The Future-Defender register: In the midst of a debate, before the module reassigns the Speaker register to an antagonist, it records the name of the conversant about to be challenged in the Future-Defender register, so that it can set up the discourse expectation that this challenged conversant will respond with a counterchallenge. (lO)The Domain-Constraints register: This register contains all the points already conceded by conversants within an argument, to ensure that these 'flawed' points are not re-used. -

4.2. Flow description----design criteria 4.2.1. Organizational, constructive, and productive states There are three major modes of processing performed in discourse generation: (i) finer and finer categorization of the type of message to be generated; (ii)

EXTENDED MACHINE-PERSON INTERFACE

183

updating our discourse mental models in preparation for generation of such remarks; and (iii) actual production. Correspondingly, the ATN states in our module (via their associated action on arcs) are of three types: (1) organizational, wherein tests and decisions are performed in finer categorization of the forthcoming conversational move category; (2) constructive, wherein the updating actions associated with a given move, which include register assignments and the creating and updating of context spaces, are performed; (3) productive, wherein actions produce the message to be generated (or recognize it in recognition mode). Organizational routines are high level, traffic directing; constructive and productive routines can be considered low level, or the terminal stations to which traffic is directed: they can only be reached via transition of an organizational state. Organizational states first push to constructive states and only subsequently push to productive ones. In this manner, any necessary updating of the discourse model is performed before a productive state is reached. 4 This is necessary since all productive actions are written in terms of a current relevant discourse context. This context usually must be updated before message formalization. 4.2.2. Discourse expectations Capturing the highly expectation driven nature of discourse processing, the context space module begins each new conversational move cycle by first checking in its discourse model to see whether or not there are any outstanding discourse expectations. When viewed as a generator, the module's choice of fulfilling one discourse expectation over others is optional and random. In its current stage of development, there is no routine to call at this point which will decide whether an expectation is mandatory and which one of a set should be chosen. Criteria of prominence and (non-)optionality will be added at a later point in its development. 5 When viewed as a parser or simulator of existing dialogues, the module's choice of discourse expectation is dictated by the flow of discourse being analyzed. When a discourse expectation is chosen, the expectation is taken out of the Expectation-List register and put into the Expectation register, and control is passed to the state predicted in the discourse expectation. Currently, discourse expectations are set up for resuming an interrupted context space; returning to the initiating subject of an analogy; further-challenging an opponent's claim or support in a debate; and counterchallenging an opponent's challenge. q n recognition mode, of course, the updatings performed along arc transitions to a final state are tentative until we reach such an end state. Conceptually, one should think of the paths as being tested in parallel until one path (or several) show success, while others fail due to precondition tests encountered along the way. 5Expectations involving context spaces whose statuses are either Open or Controlling* should be considered nonoptional; expectations involving generating status context spaces, optional.

184

R. REICHMAN

4.2.3. When to choose next speaker O u r module's organizational schema has been constructed in terms of those aspects of the preceding conversation felt to be most constraining on succeeding conversational moves. These aspects, in general, do not usually hinge upon speaker shifts, and, therefore, perhaps contrary to expectation, we do not begin a new conversational move with a choice of next speaker. First paths in the A T N have to do with deciding what type of conversational m o v e can. and will be, executed, rather than focusing on who will perform this move. In addition, there are some conversational moves where it is fell that a particular conversant is more likely than another to be the executor. For these moves we could not choose a speaker at random without introducing complications into the A T N ' s formulation. For example, in debate mode, a speaker will usually counterchallenge an o p p o n e n t ' s claim, rather than counterchallenging a claim made by the speaker him/herself or by a m e m b e r of the same side. The state responsible for counterchallenges, currently only accessed via execution of a discourse expectation, is the Challenge-Choice~Step 1 state. Let's say that in its start state, the module begins by selecting person A as next speaker, and A is known to be on the same side of the argument as the last speaker, person B. For such a case, we would somehow have to restrict the module's next transition so that it did not view going to the ChallengeChoice/Step 1 state (which would result in A's challenging B's statements) as a likely (or frequent) path to be followed. Rather than complicating the context space design by having to direct its flow of control (i.e., possible transitions) in terms of these expectations that it has for a particular speaker, the module delays immediate speaker selection in the start state and uses the discourse expectation register as a means of facilitating such traffic dependencies. For example, whenever a speaker A's claim is challenged by speaker B, the module sets up the discourse expectation that A will respond with a counterchallenge to B's challenge. By forestalling the choice of next speaker and beginning instead with the choice of a conversational m o v e (such as executing a discourse expectation) the module can then determine the most likely executor of this conversational move (i.e., by having the accessed routine use the expected speaker information contained in the discourse expectation). This design choice of not beginning a conversational m o v e with next speaker selection seems to be a good psychological model of actual discourse production. If we consider the point in the A T N where a next speaker is chosen as that point in time when a conversant actually decides to speak, then it seems most probable that by the time a conversant has made this decision, s/he has at least p e r f o r m e d enough initial processing to have an idea in mind of what s/he will say. That is, one would expect that conversants first consider their reaction to a previous speaker's statements (in the context of the discourse flow in general) and based on that reaction, decide whether or not they will perform a

EXTENDED MACHINE-PERSON INTERFACE

185

conversational move and if so what form it will take (e.g., getting back to a previously interrupted subject, agreeing with and therefore supporting the preceding speaker's statements, or not agreeing with and therefore challenging the preceding speaker's statements). It seems less likely, that, in general, without any prior processing, conversants just decide that it is their turn to speak and "speak they will". The type of processing performed in our module before it selects a next speaker is precisely deciding upon such high level conversational move decisions. 4.2.4. When we generate clue words As mentioned, the conversational moves presented earlier were basically category names representing sets of subcategorizations. Now, while each conversational move has its own associated set of preconditions and effects, conversational moves within a given category will usually have a number of such preconditions and effects in common with each other. For example, all subcategorizations of a substantive direct challenge can be prefaced with the clue words "No, but", while all indirect challenges can be prefaced by "Yes, but". Generation of the assertions constituting any one of these challenge forms, of course, can only occur after final subcategorization of the type of direct or indirect challenge to be cited. However, once having chosen the substantive direct challenge form, rather than the indirect form, we already know what clue words to generate. In the ATN, we would generate the clue words at this point. This choice of clue word generation before full subcategorization does not simply represent a more parsimonious design in the ATN. Rather, we feel that it also allows a more approximate modeling of a real person in dialogue. In natural discourse we often find the following prosodics: clue word, pause, substantive speaking. We feel the type of processing performed by the discourse module in this pause period--further subcategorization of a particular move and corresponding discourse model updating--can reasonably be behind a person's pause period as well.

4.3. Illustrative example To illustrate the operation of the discourse ATN and exemplify some parts of our conversational move taxonomy, let us follow the simple case of the module's simulated generation of C's analogy rejection in Excerpt 2. We begin by repeating the excerpt, which we follow with a citation of some of the data structures and states accessed in the simulation. Lastly, an English description of the simulation is given. Excerpt 2 A:

1. I think if you're going to marry someone in the Hindu 2. tradition, you have to--Well y o u - - T h e y say you give

18(3

R. R E I C H M A N 3. m o n e y to the family, to the girl, but in essence, you 4. actually buy her. B:

(':

5. 6. 7. 8.

lt's the same in the Western tradition. You know, you see these greasy fat millionaires going around with film stars, right'? They've essentially bought them by their status (?money).

9. No, but, there, the w o m a n is selling herself. In these 1{I. societies, the woman isn't selling herself, her parents 11. are selling her.

4.3.1. The discourse model 4.3.1.1. The context spaces By Line 9, there are two context spaces in the discourse model: context space C1 corresponding to the conversational move on Lines 1-4 of the excerpt; and context space C2 corresponding to the move on Lines 5-8. 6 Context space C1 is an Epistemic-Issue context space (TYPE: Epi), whose claim is that it is true that in the Hindu tradition women are sold as wives by their families. Context space C2 presents an analogous claim to the claim presented in context space C1. Therefore, during discussion of C2, C1 has a Controlling status assignment. C l ' s being in a Controlling status reflects that the utterances of C2 are said in direct relation to those of C1, and that C1 dictates what constitutes appropriate development of C2 (e.g., discussing w o m e n ' s wear in the West would currently be inappropriate). Context space C2 is an Analogous-Epistemic-Issue context space (TYPE: Analogy-Epi). Its claim is that it is true that the situation in the West, vis&-vis the selling of women, is analogous to the situation of woman selling in the Hindu tradition. The following is the abstract proposition of which the claims of both the initiating and analogous context spaces are instances ('$' representing a generic class): In location:$cultural-traditions m e n buy object:$women for a given function:$escort from a given source:$people.

The formal definitions of C I and C2 are listed below. (The slot names of these frame structures are capitalized.) Context space C1.

TYPE: Epi. MODE: Explicit. MODALITY: Epistemic. 6Though, in the above, distinct conversational m o v e s happen to coincide with speaker shifts, in general, there is no a priori connection between these two p h e n o m e n a .

EXTENDED M A C H I N E - P E R S O N INTERFACE

187

CLAIM: STATE-OF-AFFAIRS: In the Hindu tradition, you buy a girl as a wife from her family, EPISTEMIC-PREDICATE: True. TOPIC: Buying a girl as a wife in the Hindu tradition. SPEAKER: A. STATUS: Controlling. FOCUS: HIGH: Hindu tradition. MEDIUM: (Girl, Man, Girl's Family). Context space C2.

TYPE: Analogy-Epi. MODE: Explicit. MODALITY: Epistemic. GOAL: Pre-Generalization. CONTEXTUAL-FUNCTION: METHOD: Analogy. CO-RELATOR: C1. CLAIM: STATE-OF-AFFAIRS: In the Western tradition, fat greasy millionaires buy film stars as girlfriends by spending money on them. EPISTEMIC-PREDICATE: True. TOPIC: Buying a woman as a girlfriend in the Western tradition. ABSTRACT: RELATIONS: (Buy). PROPOSITION: (In Location:$cultural-traditions man buys object:$woman with money for function:$escort from source:$people). MAPPINGS: ((location :$cultural-traditions, Western tradition, Hindu tradition) (object:$woman, film star, girl) (function:$escort, companion, wife) (source:$people, film star, girl's family) (buy object:$woman with money for function:$escort, buy film star with money as a companion, buy girl with money as a wife) (buy object:$woman with money for function:$escort from source:$people, buy film star with money as a companion from film star, buy girl with money as a wife from girl's family) (In Iocation:$cultural-traditions man buys object:$woman with money for function:$escort from source:$people, In western tradition, greasy fat millionaires buy film stars with money as companions from film stars, In Hindu tradition, man buys a girl with money as a wife from girl's family)). SPEAKER: B. STATUS: Active. FOCUS: HIGH: Western Tradition. Medium: (Film star, Man).

188

R. REICHMAN

4.3.1.2. The registers Current non-nil registers are: - CCS = C2; the current Active context space is C2. - H E A D - C C S = C1; the current Controlling context space is C1. -Discourse-Mode = Discussion; we are not in the midst of discussing disputing claims. - S p e a k e r = B; B is the last conversant to generate discourse, and he is still holding the floor. -Participant-List = (A, C, D); D has spoken earlier in this conversation. 4.3.2. States accessed in the simulation Figs. l and 2 represent that portion of the discourse A T N involved in simulation of this piece of dialogue. N u m b e r s on arcs correspond to transition tests that must be met for the transition to be maxim-abiding. Explication of these tests is given in Appendix A. For brevity, only push actions are noted on transition arcs. In the non-pictorial representation of Appendix B we list all actions associated with each state. Infix notation is used for multi-argument predicates, with spaces delimiting the elements (e.g., Expectation A P P E N D ExpectationList), and should be read as English-like sentences; prefix notation is used for single-argument predicates (e.g., NOT(X)); the selection of a slot of a construct is indicated by ' n a m e of construct' '.' ' n a m e of slot' (e.g., Expectation.Context); and indices are notated by { }'s (e.g., CCS.CounterClaim{I}). In Appendix C, a trace of the module in its simulation of Lines 9-11 of the excerpt is given. States in an A T N are really iust place holders: it is via tests and actions along state transitions that processing gets performed. Therefore, in our general overview description of those states of the A T N accessed in generation of Excerpt 2, we focus on these tests and actions. (1) Produce-Next-Move. Produce-Next-Move is the A T N ' s start state, where the paths of most conversational moves begin. In high-level conceptual terms, the new conversational m o v e can be related to the preceding discourse in one of five ways, as captured by choices A - E of this state: -Expectation: execute a m o v e predicted and/or required by preceding conversational development (e.g., answer a question, return to an interrupted context space, return to the initiating context space of an analogy after the analogy is completed, etc.); -Developmental1: have the m o v e be a developmental continuation of the preceding discourse but one not necessarily expected or mandatory (e.g., begin a debate, support, comment-on, question, further-develop, restate, etc.): -Developmental2: same as D e v e l o p m e n t a l l with exclusion of starting a debate once already in debate mode; - Nondevelopmental : have the move be a nondevelopmental continuation of the preceding discourse (e.g., restart the discussion of a closed context space, abstract onto a topic logically prior to the one currently being discussed, attempt

EXTENDED MACHINE-PERSON INTERFACE

189

• LOGICAL ABSTRACTION

NOT(21)'~ 1 ~'21 ACT: d I J i ~ ~ ACT

:

COMMENT- ) CHOICE

(SUPPORTCHOICE1

JUMP

TEST: 13 ACT: JUMP

JbLMP

JUMP CHOICE2 TEST: 15 ACT: JUMP ,

ACT: JUMP

ACT: JUMP UTTERANCE

/

'i~'j ~ . ~ ...... _L~ TEST: 14

ACT: CHOOSE EXPECTATION GO EXPECTATION. CONTEXT

]

JUMP

TEST:

~,~"

( DEVELOPMENT-~ \

I

/

ACT: Jt.rMY

ST.

ACT: JUMP •

CHOICE1

]ACT: PUSH SI{IFTSHIFT-SPEAKER-EXPECTATIOn__ SPE?dCERS;J IE2 ACT: PUSH

.I

SHIFT~SPEAKER-J~ SAMESI

TEST: 6

TEST: NOT 16 CHALLENGEINDIRECTLY

TEST: 17 ACT : JUMP

TEST: 16 ACT: JUMP

TEST: NOT 16 ACT: PUSH CHALLENGEDIRECTLY

CHALLENGE~"~ ACT : JUMP

FIG. 1. Pictorial ATN representation•

to interject something tangentially related to the current topic of discourse, etc.); -

Transitional: generate a metastatement that will explicitly state the connection between the next conversational m o v e and what went on before (e.g., "Speaking of discrimination, I wanted to tell you this story").

190

R. REICHMAN

ACT: (POP) ~

I

ACT" JU~ TO

I

y CHALLENGE~-N suPPORT- ) X ~

P

/

AC~: .:>~

{ SUBAaG_U~Nr-'~

k.~NCESSIO~J

PRINCIPLE )

[TEST: 8

///

I [ [ / 1._.....~ ~ - - - -

ACT: ( P O P )

t

/

I

!LsL:, S H

U

ACT: ,:UYA' PR
~

/

~

TEST: !9

~ACT: _JUMP]0 ~ FURTHER-

rEST: 5, NOT(6) "~ SUPPORTACT: JU>~ CHALLENGE3

C~LLENGE-~ I-

\

I

ACT:

\

i TEST: NOT(2) NOT(3) . / NOT(4) / ACT: J U i ~ /

ACT: PUSH

/

/

/

:

\ TF2Z;T: \ N~37(7)

/

\

ACT- SdIU?- /

\

\

CHALLENGE/ /

ANALOGY- I / t

MAPPINGS

TEST: "7

\

ACT: P'=~ [ .RELEVAL;T- I

• -

FACTOR~/

SUPPORTREJECTION

ACT: (POP) : (POP) FIG. 2. Pictorial A T N representation (continuation of Fig. 1).

(2) S t a r t - D e b a t e . W h i l e the s a m e set of m e c h a n i s m s and structural d y n a m i c s s e e m to occur in m o s t forms of e x p o s i t o r y text, w e have found it necessary to distinguish b e t w e e n the d e b a t e and n o n - d e b a t e m o d e in discourse. For a debate, w e m u s t record a n u m b e r of additional features of the transaction;

EXTENDED MACHINE-PERSON INTERFACE

191

mainly the fact that here conversants partition themselves onto opposing sides. In reflection, the module begins a debate by first choosing a next speaker, and then putting this speaker on one side of the debate and the preceding speaker on the other side. These initial sides are recorded in the Sides register, which is updated whenever a conversant changes sides in a debate or an unaligned conversant takes a side. (3) Challenge-Choice/Step 1. The Challenge-Choice/Step 1 state is responsible for counterchallenges: it can only be reached via execution of a discourse expectation in the environment of an ongoing debate. Associated with the discourse expectation that leads us to this state is the name of the conversant whose counterchallenge is expected. In choosing next speaker this expected speaker information is used in one of two ways: (i) push to the Shift-Speaker-Expectation state, where the expected speaker is selected as next speaker; (ii) push to the Shift-Speaker-Same side state, where: -a conversant already known to be on the same side of the argument as the expected speaker is selected as next speaker; - a previously uncommitted speaker is selected as next speaker. This conversant is appended onto the expected speaker's side of the debate (i.e., updating of the Sides register); - a conversant from the other side switches sides and is selected as next speaker. This allows the last speaker to hold the floor and allows for 'sincere' switching and 'devil advocate' playing. This change in a conversant's position is recorded in the Sides register. In addition, as we usually find conversants counterchallenging an opponent's challenge, the module here also records that the preceding speaker is a likely Future-Defender of the claim about to be challenged. 7 (4) Challenge-Choice/Step 2. There are three major subcategorizations of a challenge conversational move: - T h e direct challenge form: A conversant either emotively undermines an opponent's claim; demands that the opponent support a claim or unsubstantiated dismissal-challenge; or, makes an assertion which directly opposes the opponent's. - T h e indirect form: A conversant does not verbalize explicit disagreement with an o p p o n e n t ' s claim, but rather, cites a claim, which if true, implies that the o p p o n e n t ' s claim is either not true, irrelevant, or only one skewed version of the truth. - T h e support form: A conversant responds to an opponent's demand that s/he 7This is appropriate even for a speaker countering her or his own preceding claim, because in such situations a speaker usually equivocates between the two positions and frequently follows the "switched side position" with a "But, on the other hand", returning to the first position put forward.

192

R. REICHMAN

support a claim. As this m o v e occurs in response to an o p p o n e n t ' s challenge (of the d e m a n d support category), it is treated as a counterchallenge much as are other responses to an opponent in a debate. A conversant is always free to choose a direct or indirect form of challenge except in the case where s/he is expected to respond to an o p p o n e n t ' s preceding d e m a n d for a support. In cases of an outstanding support demand, the module, following maxim-abiding discourse development, limits subsequent transition to the support state. In the Challenge-Choice/Step 2 state then we get our first subcategorization of the Challenge conversational move. Actions associated with this move are to append the current speaker onto the Antagonist list of the current Active context space, and to create the discourse expectation that the current speaker may, at a later point in time, further challenge the F u t u r e - D e f e n d e r ' s preceding claim. Creating further challenge expectations enables the module to cleanly model discourse pops incurred either when a conversant fixes some flawed subargument, or when a team m e m b e r follows a speaker's subargument challenge with additional argumentation against the o p p o n e n t ' s argument that led to current subargumentation. As noted, however, conversational moves also have an associated set of tests. For example, the module does not create further challenge expectations under the following discourse conditions: - T h e current speaker is challenging an o p p o n e n t ' s preceding fiat dismissal of her/his argument in the active space. One only has two means of countering a flat dismissal: (i) accept the dismissal and concede the subargument: or (ii) demand the o p p o n e n t to replace the unsubstantive challenge with a substantive one. There is no r o o m here for further-challenges. - T h e current speaker is responding to an o p p o n e n t ' s d e m a n d that s/he support an earlier claim. In this case, the challenge is really just supporting one's own claim in the Active space. Setting up a further-challenge of this context space by its protagonist is inappropriate. - T h e current speaker is responding to an o p p o n e n t ' s d e m a n d that s/he give a substantive challenge against the Active space in place of her/his earlier unsubstantive challenge. In this case, the further-challenge expectation set up when the conversant first did the challenge with the fiat dismissal is still on the Expectation-List register. The challenge currently being processed is the counterchallenge expectation to the support demand. A new further challenge need not be set up. (5) Challenge-Directly/Step 1. There are many ways of directly attacking an o p p o n e n t ' s argument. Within a given discourse context, however, only some of these types of attack are appropriate. Corresponding to each possible form of attack, a distinct arc transition emanates from this state. The transition tests on the arcs capture the context sensitive features of these different developmental options. The three main conceptual subcategorizations of this move are:

EXTENDED MACHINE-PERSON INTERFACE

193

-Emotive~questioning challenges: keeps the burden of proof on the opposition; it undermines an opponent's position but it does not cite any evidence or specific claims against it; for example, a "So what" type of challenge (Choice A, B, and C); - S u b a r g u m e n t concession: conceding a flaw in one's argument, followed either by ending all debate (i.e., total concession) or by a fixing of the flawed argument (Choice D); -Substantive challenges: entails involving oneself into the validity of the opponent's argument, usually by citation of some contrary claim and/or evidence against it; for example, denying the truth of the opponent's claim or support fact (Choice E). Choosing any one of these direct challenge forms, and their many forms of subcategorization, is constrained by preceding conversational development. For example, given no preceding challenge, a subargument concession is inappropriate. In general, a conversant can either challenge an opponent's claim (i.e., issue context space), or s/he can challenge an opponent's support for a claim (i.e., support context space). The module's first action along the substantive challenge form is to decide whether to attack an opponent's claim or support of such. The following briefly summarizes the two cases where updating actions have to be performed here: (i) Current active space is an issue context space. If we choose to challenge a preceding support of an Active issue space, we must reassign the Active issue space to Generating; reassign its high focus element to a medium focus; and re-activate the last support context space of this Issue space. (ii) Current active space is a support context space. If we choose to challenge an issue context space and one of its support spaces is currently Active, we must close this Active support space; reassign its focus level assignments to zero; and re-activate its to be challenged issue space which is currently pointed to by H E A D - C C S . If the substantive challenge form (Choice E) is chosen, the generation of clue words specific to this move are "No, but". However, in counterchallenge of an opponent's argument that entailed citation of a state of affairs that did not occur, the clue words "Well, of course", or "But, of course", can be used as well. An opponent claims a state-of-affairs A did not occur (a Modus-Tollens form of challenge) when the other opponents' claim, if true, would have included A's occurring. Claiming A does not occur, then, argues against the initial position. To counter such an argument, a defender of the original position, must illustrate that this nonoccurrence of A, in fact, does not violate predictions of the initial argument, and that, on the contrary, its non-occurrence is quite consistent and expected by the initial position (i.e., the 'unexpected' event is not unexpected at all). Hence, the use of "Well, of course", or the like, in such counter-argumentation. Lastly, in the case of a further-

194

R. REI('HMAN

challenge, the clue words "And, furthermore", or the like+ are generated instead of "No, but", or ++Well, of course `+ Choosing a substantive challenge is choosing a general subcategory of the Challenge move. In the module, this subcategory is then divided into a number of other conversational subcategories. Each further subcategory has additional tests and updating actions. For example, the 'Challenge-Support-Specifics" subcategory (Choice E/A) is only appropriate as a counterchallenge to specifics of an opponent's preceding support context space: the +Range-Application' subcategory (Choice E/C) is only appropriate as a counterchalleng~ to :t preceding 'New-Factor' type of challenge: while the 'Challenge-Analog~Mappings" subcategory (Choice E/F) is only appropriate against a preceding analogy. If appropriate and the Challenge-Analogy-Mappings subcategory is chosen, the module performs a number of additional updating actions to prepare its discourse model for subsequent message characterization. Its first action on this transition arc is to set the Analogous register to the analogous context space about to be challenged (i.e., the current Active space). This register is set to ensure that during the analogy challenge, the initiating space of the ~malogy will retain its Controlling status, and the analogous context space will be assigned a Generating status. This reflects that, of the two spaces to be cited in the forthcoming analogy challenge, the initiating space is the one of higher influential status. Before transition to any substantive subcategory, the module pushes to a constructive routine where a new issue context space is constructed to hold the forthcoming challenging remarks. CCS is assigned to this space, whose following slots are set: Goal: Challenge. Mode: Explicit. Co-Relator: The context space being challenged. Status: Active.

After this discourse model updating, control is passed to the state specified along the transition chosen---one of the subchoices of Choice E. (6) Challenge-Analogy-Mappings/Step 1. An analogy rejection usually rests on some non-identity of relations between the correspondent domains, where this non-identity either undermines the conversational move which the analogy was meant to fulfill or undermines the very relation being claimed constant between the two domains. The module's formalization of these two forms of analogy rejection is in terms of one's being able to infer a set of enabling relations between the relations underlying the analogy and slot fillers of its Mappings slots (cf., [48] for a full discussion of the module's analysis and characterization of an analogy move and rejections thereof). After message characterization, the module puts

EXTENDED MACHINE-PERSON INTERFACE

195

a propositional representation of the generated message (returned from the semantics and syntactic modules) into the Claim slot of the Active context space. (7) Challenge-Directly/Step 2. All forms of the substantive challenge entail creation of a new challenge issue context space. This new issue space is inserted into the CounterClaim or CounterSupport slot of the opponent's DebativeIssue space. (8) Challenge-Choice/Step 3. After generation of the challenging remarks the discourse module sets up the discourse expectation that the challenged conversant will counterchallenge the challenge just given (i.e., (ChallengeChoice, Future-Defender, CCS) is appended onto the Expectation-List register). 4.3.3. The

trace

To simulate generation of Lines 9-11 of Excerpt 2, the module begins in its start state, 'Produce-Next-Move'. Since there are no outstanding discourse expectations at this point, and the discourse mode is discussion, the module can, among other options, choose to further develop discussion of the woman selling situation in the West; generalize onto the issue of women being sold the world over; totally leave discussion of the selling of women; resume discussion of the Hindu situation; or dispute the analogy cited. Simulating the discussion as it occurred, the module chooses the latter (i.e., transition to Start-Debate). The module begins the debate by setting the 'Discourse-Mode' register to 'Debate'. Then the conversant about to be challenged, B, is assigned to the 'Future-Defender' register, and he is put on one side of the 'Sides' register. Next, C is chosen as next speaker; she is deleted from the 'Participant-List' register; B is added onto this register; the 'Speaker' register is assigned to C; and C is put onto the other side of the 'Sides' register. Control then passes to the ChallengeChoice/Step 2 state where the module subcategorizes the forthcoming challenge conversational move. First a test is made to determine whether a further-challenge expectation is warranted by the current discourse environment. Since C herself just began the debate, all such tests are met and the module sets up the expectation that C (the current speaker) may at some later point further challenge the analogy just cited (the claims contained in the still Active, about to be challenged, context space): (Further-Challenge, C, C2) is appended onto the Expectation-List register. Since C is challenging a debative-issue context space and not one of its support constituents, and she has just begun the debate and is therefore not responding to a demand for a support of some preceding claim, the module appends C onto the Antagonist-List of the current Active context space, C2. At this point, transition tests to the direct and indirect challenge states are met and simulating C's move, the module chooses the direct challenge subcategory.

196

R. REICHMAN

There are a number of subcategories of the direct challenge move. In the current discourse environment, a number of these subcategories are applicable, others are not. For example, the 'Demand-Support-Rejection' (Choice B) is inappropriate since C's move begins the debate and therefore the preceding speaker could not have just flatly rejected an earlier argument by C (i.e., CCS.Comment.Contextual-Function.Method could not equal 'Flat-Rejection'). Modeling C, in this example, the module chooses Choice E - - t h e substantive challenge form. Since all substantive challenges entail generation of some assertion, a new context space is created to hold the forthcoming challenging utterances. The module now constructs context space C3, fills its Goal, Mode, Co-Relator. and Status slots, and sets the 'CCS' register to point to it. Next, the module is ready to generate the clue words which accompany a direct challenge conversational m o v e - - h e r e , "No, but", is generated. Then, the module further subcategorizes the substantive direct challenge form. Each subcategorization choice has its own set of tests. In this discourse environment a number of options are available; modeling C, the module chooses the analogy rejection move (Choice E/F). Formally, C's rejection takes the form of asserting a non-correspondence of relation involving slot fillers of the mappings set up in the analogy: in the Western case there is an I D E N T I T Y relation between the role fillers of the $Source and $Object case slots of the B U Y relation, whereas, in the Hindu case this relation does not hold. This non-correspondence between the domains is relevant to current discussion because ownership by the sellee is a precondition for a legal buying: it E N A B L E S a BUY; this ownership relation is true in the Western case but it is not true in the Hindu case. Since this relation is vital to the relation being claimed constant between the two domains (all C A U S E and E N A B L E relations are considered vital), the noncorrespondence undermines the applicability of the analogy. Using its earlier explication of the implicit connections between B's analogous utterances and A's initiating ones (via slot fillers of C2, the analogous space), the module formulates the analogy rejection. As C2, the space just challenged, is a debativeqssue space, C3 is put into its CounterClaims slot and B's expected response is noted by the (ChallengeChoice, B, C3) expectation appended onto the 'Expectation-List' register. This completes this conversational cycle and the module returns to its start state, Produce-Next-Move, for the forthcoming conversational move. Let's now briefly compare the context space approach to discourse processing with some 'other current theories. 5. Other Approaches to Discourse Processing O n e c o m m o n approach to discourse has been to characterize and differentiate b e t w e e n different discourse forms. Each such c a t e g o r i z e d form is treated as a

EXTENDED MACHINE-PERSON INTERFACE

197

distinct discourse genre or schemata with its own special and predictable structure. Various analysts have generated a profusion of genre classifications, based on widely varying criteria for distinction, and delineated different conventions applicable to each type. Some classifications are based on the purpose of a set of utterances, and result in classifications like 'narratives', 'arguments ", and 'descriptions'. Other classifications are based on the style of presentation and result in distinctions like ~poetry' or 'verse'. Still others are based on content, and result in classifications like 'drama', 'fiction', and 'riddles'. Whatever the classification scheme, this approach assumes that genre dictates structure, that shared rules of discourse development depend on these various possible 'structures' (e.g., [5, 22, 24, 31, 36, 44, 45, 49, 60]). We agree that in general we have different expectations of what someone telling a story and someone making an argument will say, and that we have different expectations about the order in which the different kinds of elements will be arranged in these different discourse forms. Our analysis of discourse must therefore incorporate these distinctions. However, the ~genre dictates structure' approach leaves some fundamental questions about discourse unanswered. Since it is clear that in any ordinary discourse or conversation, speakers use a number of different 'genres'--stories, descriptions, arguments, for instance--what conventions govern the development of the whole conversation that includes these subparts? An adequate theory of discourse needs to specify the conventions or rules that apply across as well as within discourse forms: rules that describe our discourse processing regardless of particular discourse developments. An approach that has attempted to include such general rules of discourse is that of the ethnomethodologists, which focuses on the universal and contextfree conversational phenomenon of 'turn taking' [50]. This focus has resulted in descriptions of rules that define 'legal' points of transfer from one speaker to the next. In this view, conversations develop as speakers take turns completing conventional paired forms such as greeting/greeting; offer/acceptance or refusal; question/answer; and so on [25, 26, 53]. While this scheme has the advantage of identifying some general conversational rules which pertain to the function of different conversational elements, the rules are less linguistic conventions than simply conventions of general interpersonal behavior. More important, because the emphasis is on paired forms in discourse, the analysis is mainly restricted to linear conversational development. It has no way to explain (except as exceptions to the rule) the large proportion of discourse that is nonlinear and nonsequential: the frequent shiftings, suspensions, and resumptions of topics that we find in natural dialogues. Accounting for these must take a central role in any adequate theory of discourse structure. More strictly linguistic attempts to show how conversants follow each other through discourse have taken two approaches. In the first, analysts attempt to explain how discourse 'hangs together' or 'coheres' by attempting to identify

198

R. REICHMAN

surface linguistic cohesion devices. Halliday and Hasan's work, "'Cohesion in English", exemplifies this approach [21]. Halliday and Hasan diw)rce their claims about the surface linguistic evidence for text 'cohesion' from any notion that a text may have a rich, underlying deep structure. The linguistic claims, then, though general in form and specific to language use, fall short of the type of analysis attempted here. As illustrated, one of our main conjectures is that 'lexical cohesion' can only be properly understood and identified given a rich and careful description of a text's underlying deep structure. Halliday and Hasan's evidence of lexical cohesion, on the other hand, include the observations that (i) in 'cohesive' text we find an abundant appearance of related words in close proximity (e.g., key and door appearing within short distance of each other), and (ii) we find pronouns following nouns. Most 'ties' like these. however, are mere artifacts of the talk content and are found in the single sentence as well. In the second discourse-linguistic approach, of which we are a parl, identifying a conversation's 'deep structure' in terms of structural relations between discourse elements, precedes linguistic analysis. Among this class of researchers we find the works of Grosz [20 l, Hobbs [23], Linde [34], Linde and Gougen [35 I, Longrace and Levinsohn [37], and Weiner [61]. Let's briefly consider the works of Linde and Grosz. Both researchers, studying task oriented dialogues, show that we can parse a discourse into a tree-like structure where each node of the tree represents a distinct discourse element. They then demonstrate how given this structure we can explain and even predict the appearance of many 'surprising' surface linguistic forms found in a discourse. Grosz, for example, studying the transcripts of experts teaching laymen how to assemble a water pump, shows that in the water pump world distinct discourse elements corresponded to the particular subtasks involved in such an assembly. Grosz demonstrated, then, for example, that when conversants complete discussion of some subtask and return to the larger task in which the subtask is embedded, they will often use a pronominal form to reference something in this larger task even in those cases where a seeming contender for this pronominal is found in the linearly closer discussion of the subtask. Grosz notes that it is the underlying hierarchical structure of the dialogue which explains why no ambiguity arises--after completion of a subdiscussion its elements are no longer available for pronominal reference. Here, you notice we have a truly discourse-linguistic phenomena: not something explainable by, or an artifact of, the external world of talk. Despite the fruitful indications of this type of structural linguistic approach, the currently most influential analyses of discourse processing have abandoned 'the linguistic approach' as fundamentally wrongheaded and misdirected. These analyses proceed from the assumption that it is not specific linguistic conventions that guide discourse development and interpretation; rather they assert that it is the 'world knowledge' conversants share that accounts for

EXTENDED MACHINE-PERSON INTERFACE

199

mutual understanding (e.g., [1, 6, 12, 16, 30, 41, 42, 43, 57, 62]). Morgan and Sellner ]40], for instance, in a review of linguistic approaches, conclude that "as far as we can see, there is no evidence for cohesion as a linguistic property, other than as an epiphenomenon of content". The world knowledge school's alternative explanation of discourse processing is based on 'speech act theory' [2]. Here, like in the context space approach, discourse utterances are interpreted functionally. The source of functionality differs in the two systems, however. In the context space approach, functionality stems from a set of conventional rules underlying discourse engagement; in a speech act system, functionality stems primarily from the 'working of the external world' and the underlying psychological states of conversants. In the world knowledge view, the primary, prominent feature governing discourse processing is a listener's identification of a speaker's intention. It is this knowledge, quite external to the discourse, that allows conversants to interpret each other's meaning and follow the discourse flow. If conversants apprehend each other's intentions, they can on the basis of these share compatible models of the current topic, focus, and role of discourse elements. As Spiro [58] states: "When comprehending utterances in conversations, the speaker's intent is all that is attempted to be communicated. If that intent is apprehended, the goal of communication is met and the communicative act is 'understood'." There are, however, important difficulties inherent in this approach to discourse analysis. While the determination of a speaker's intent or goal is certainly a part of grasping meaning, it is of utmost importance to distinguish between a conversant's intention for an utterance and the communicative effect of that utterance in the conversational context. Though not intending to inform, a speaker may yet inform; though not intending to persuade, a speaker may persuade; though not intending to support, an antagonist's remarks may be shown to support the position the antagonist intended to dispute. This last case highlights the fact that not only can an utterance have an effect or force independent of a speaker's intent, but even when a listener knows that the force of an utterance does not necessarily correspond to a speaker's intent, the effect of that utterance may remain the same. An analysis by Clark and Carlson [10] of a single line of Shakespearean dialogue, Othello's "Come, Desdemona", illustrates the rigor with which the world knowledge theorists apply their method to explain the processing of verbal communications: Othello intends to get D e s d e m o n a to understand that he wants her to go with him. He gets her to understand this by getting her to recognize his intention to get her to understand this.

It is easy to see how this kind of purported explanation could be applied to the analysis of extended conversation. But even in such a simple case, if a

200

R. R E I C H M A N

communicator's goal could not be inferred until a message were grasped, and if a message could not be grasped until the communicator's goal was inferred, prospects for communication would not be good. There is really another issue here: mutual knowledge. We would argue that the world knowledge theorists' focus on intentional accounts of discourse has become so radical as to obscure the obvious: it is often the case that understanding necessitates some mutual, or privy knowledge between conversants. For example, consider the following scenario and analysis presented by Clark and Carlson [10]. Scenario 1. 2. 3. 4.

Ann and Barbara went to both the Tate and British Museums. Barbara had discussed with Charles her going to the Tate Museum. Barbara does not know what museum plans Ann and Charles have discussed. Charles to Barbara in front of Ann: "'Did the two of you go to the museum'?'"

Analysis "Charles could expect Barbara to recognize that the museum he was referring to was the Tate, since it was the only museum they had discussed together; that is, he could expect her to pick out the Tare by means of her recognition of his intention,"

We agree with Clark's analysis up until the last phrase "her recognition of his intention". Substitute "mutual knowledge" here and we have a winner (cf., [111). In sum, while a speaker's intent may well be reflected by a communication, grasping that intent cannot be a necessary precondition for understanding. A 'pure' world knowledge theory is simply unable to account for how communicative goals are in fact apprehended. Indeed, we would rather claim that there are a standard set of conventional discourse roles that utterances play, and that a speaker's intent may or may not necessarily be reflected in these conventional roles. Usually, of course, there will be a correspondence, stemming, in all probability, from individuals, over time, acquiring standardized procedures for ways in which to express underlying intentions, beliefs, and desires (similar to Lewis' [33] and Bennett's [4] analysis in discussing 'convention'). If the convention hypothesis is correct, then we can understand how, rather than speaker intent identification driving interpretation of the speaker's utterances, it is the interpretation of those utterances through conventionalized conversational move characterizations which enables us to hypothesize the speaker's underlying psychological state. Some world knowledge theorists allude to this convention hypothesis and inadequacy of an intentional account alone. Searle [55], for instance, points out that our use of language does not vary arbitrarily because we decide to intend new meanings; that "meaning is more than a matter of intention, it is also a

EXTENDED MACHINE-PERSON INTERFACE

201

matter of convention": "In the performance of an illocutionary act the speaker intends to produce a certain effect by means of getting the hearer to recognize his intention to produce that effect, and furthermore, if he is using words literally, he intends this recognition to be achieved in virtue of the fact that the rules for using the expressions he utters associate the expressions with the production of that effect. It is this combination of elements which we shall need to express in our analysis." In general, however, Searle does not explicate these rules of linguistic convention which must operate to associate particular expressions with the production of particular effects. We hope to have provided such a system.

6. Conclusion This paper has reported on some basic issues involved in the design of a person-machine interface capable of extended dialogue with a user. At the conceptual level, I have illustrated that such a system, at minimum, has to be able to: (l) characterize the functionality of discourse utterances; (2) characterize the different implicit components of different types of conversational moves; (3) constrain inappropriate conversational development; (4) track the effects of preceding conversational moves on the discourse context; (5) predict and execute most likely forms of subsequent development. A prototype of a discourse module has been presented with delineation of some of the data structures, organizational control, and conversational move taxonomy needed in such a module. We also illustrated that a static taxonomy of semantic and logical criteria for conversational moves was insufficient. We showed that, in addition, one must also be able to delineate the requisite discourse environment needed for each move, such that a conversant's performance of this move, at a particular stage of discourse, would be appropriate. Thus, in our system, for each move characterized, we have also specified a number of preconditions represented by arc transition tests in the ATN. In addition, we recognized that the taking of each conversational move takes us into another stage of discourse. We have reflected this by also associating a set of effects with each conversational move defined. These effects are represented by arc transition actions in the ATN. As noted earlier, the rules embodied in the context space module are independent of particular speaker/listener beliefs, world knowledge, attitudes, and personal motivations. In actual discourse, however, use of these rules are intertwined with any given speaker's world knowledge and beliefs. In this respect, the context space discourse system is analogous to sentence grammars

202

R. REICHMAN

(e.g., [8[). For example, given that a speaker has the following phrase structure rules:

S ~ NP VP NP--* D E T N VP-+V s/he will never be able to generate the utterance, " T h e machine is broken", if s/he does not have the word "machine" in her lexicon. Similarly, if R had not known any studies supporting the genetic side of the genetic-environmental controversy, she could not have successfully presented a study supporting her side of the argument (Excerpt 3), This would not mean that she did not have this possible form of argumentation as a rule in her discourse system. She simply would not have had any appropriate propositions stored in memory that could fulfill the abstract constraints set on this path. Thus, this work should not be seen as a replacement theory for frames (Minsky [39]), scripts (Schank and Abelson [52]), schematas (Bartlett ]3]), or semantic network theory (Collins and Loftus [14], Collins and Quillian [15]). Rather, it supplements these theories by illustrating that there is a characterizable set of independent discourse processing rules shared between conversants of a given language community. These shared rules are a major enabling factor to our being able to effectively communicate with people we've never met before. The rules form, direct, constrain, and impinge upon the structures we use in conversing. The language module proposed here shares many of the cognitive structures used by other systems of human cognitive processing. For example, the earlier cited rule that the interpretation and/or generation of a context space is done only in direct relation to a single preceding space corresponds to one's general selective attention and focused processing way of accomplishing many sorts of tasks. The following are but some discourse phenomena reflecting discourse focused processing: (1) There seems to be only one current discourse topic at a given point in time. (2) While from any given point in the discourse there are a number of open lines of future development, typically, only one line of functional and thematic development is chosen. It is probably safe to conjecture that an attempt at simultaneous option fulfillment would be perceived either as an ambiguity or go unnoticed. (3) We seem to continuously update the status of preceding context spaces so as to highlight a specific set of utterances while unhighlighting another set, as reflected by our use of such clue words as " B u t " , " A n y w a y " , "Incidentally", and corresponding continued pronominalization or nonpronominalization to

EXTENDED MACHINE-PERSON INTERFACE

203

elements of these preceding context spaces. This unhighlighting phenomenon is further reflected in the fact that after following one option of development, other options previously available, no longer seem viable. (4) Focused processing is further mirrored in the fact that intersentential pronominalization seems to be limited to a specific single entity considered to be in the highest focus in a relevant discourse context. In general, simultaneous achievement of more than one discourse function by a given set of utterances (i.e., by having them simultaneously directly related to varying discourse topics and serving varying conversational moves), is considered to be a 'creative' use of language. We even have special descriptive terms for such occasions: allegory, pun, play on words, double entendre, etc. These terms reflect that here we have distinct genres, perceivably different both in form and frequency from the ordinary. Even in these creative discourse forms, it is questionable if we can simultaneously process these utterances in terms of their multi-functionality. Consider for example the allegory. I think one might agree that the processing of the 'story' of the allegory is qualitatively different than our processing of an ordinary narrative in discourse. To actually listen to the events specified as a 'real' story, probably would entail losing hold of the allegory being told. At minimum, it probably necessitates some conscious 'shifting' between the two modes of interpretation, similar to the "shifting" needed to 'simultaneously' see the two views of a Necker cube. Other correspondents between the language module and general cognitive abilities and modes of thinking are: expectation driven processing; cue driven processing; segmentation; reference frame identification; hierarchical organization. In addition, the underlying primitives of the system, instantiation, implication, presupposition, and focus, are cognitive functions and functional abstractions that humans employ for varied reasons. Within the context space grammar we have specific points at which calls are made to such structures. As Chomsky suggests (Chomsky [9]) it is yet an open question as to what should fall under the language faculty and hence appear in some form in its representations of meaning. I would claim that many of these non-language-specific operations are used in the domain of the language faculty; they form the basic operations in terms of which we can formalize the abstract language rules governing conversational speech. This sharing of cognitive structures between separate systems seems quite reasonable, since one would expect the mind to be equipped with a basic set of cognitive operations. The modularity involved, then, is the specific orderings of such operations in the fulfillment of different tasks.

204

R. REICHMAN

We repeat however that no claims are being made about the psychological reality of the specific system presented here; though, it does seem that an individual's discourse processor would be structured in terms of the basic analytic apparatus presented (context spaces, context space relationships, status and focus level assignments). This work demonstrates that tracking a conversation's changing 'relevant contexts'--reference f r a m e - - a n d the changing constraints and demands placed on conversants is a basic, mandatory component of such a discourse process, and that such tracking can effectively be performed in terms of these specified structures. However, one can only speculate as to how this process is actually accomplished in the mind. At its highest level of characterization, the context space theory addresses the most general and abstract features of discourse structure. At this level of characterization, which ignores features specific to particular thematic discourse developments, a context space is merely an abstract structure with slots for (1) a propositional representation of the discourse utterances said to lie in the space; (2) the conversational move fulfilled by the space; (3) a propositional representation of the implicit components needed for the space to fulfill a move; (4) the influential Status of the space at any given point in a discourse: (5) links to preceding discourse context spaces to which this space has some relation, and a specification of the relation; (6) links to subconstituent context spaces; (7) focus level assignments to constituents of the utterances said to lie in the space. If we use this level of characterization, all discourse utterances, whether of narrative, technical, discussion, explanation, description, or argument form, share the same structure and are governed by a single set of discourse rules. The following synopsises some of these general rules: (1) A conversation is a series of conversational moves, each move corresponding to a particular functional relation between the utterances generated and utterances contained in a preceding context space (if any) in the foreground of discourse. (2) Utterances in a single context space serve the same conversational move. (3) All conversational moves have an associated set of preconditions and effects which are derived from, and operate on, one's model of the underlying discourse structure. (4) During development of a subconstituent space, the superordinate space is assigned a Controlling status assignment. (5) At any given point, one's current reference frame for subsequent discussion is the Active and Controlling space pair.

EXTENDED MACHINE-PERSON INTERFACE

205

(6) Closing a context space results in all of its contained entities being given zero focus level assignments. (7) Putting a context space into a Generating status, results in a reassignment of its high focus constituent to a medium focus. (8) A context space having an Open status assignment, reflects that it is expected and required that the space be resumed and completed after completion of the digressive context space. (9) Closed context spaces can be ignored in the discourse popping process. (10) Discourse popping results in the closing of all context spaces developed during intervening discussion (i.e., those not already closed). (11) Generation and interpretation rules only access context spaces in the current reference frame. (12) Linguistic markers like "but", signal that an Active context space should be Closed, and that a Closed space is to be re-activated; clue words like "By the way", signal that an Active space should be reassigned an Open status. (13) Linguistic markers like "now", and "like", signal that a subordinate context space is about to be developed. (14) Close discourse deictics can only be used for constituents in a Controlling or Active context space. (15) Pronominal reference can only be used to refer to an entity in high focus in a Controlling or Active context space (subject to some intersentential syntactic exceptions). Using rules like the above enable conversants to build compatible models of ongoing interaction despite the many twists and turns found in everyday discourse. It is our belief that future computer natural language systems hoping to engage in extended dialogue will need a system based on similar principles. Appendix A. Tests Applicable in Transitions (1) Expecting support of an unsubstantiated fiat rejection. Discourse expectation predicts that the antagonist will actually engage into the validity of the protagonist's position. This expectation results from a protagonist's demand that the antagonist back-up her or his preceding emotive flat rejection or irrelevance rejection of the protagonist's earlier argument. Expectation.Associated-Constraints = 'Supply-Support-Rejection'

(2) After denial of truth. The speaker is counterchallenging an opponent's preceding challenge of the truth of the current speaker's preceding claim. For Some I, CCS.ContextuaI-Function.Method{I} = 'Deny-Truth' (3) A f t e r

unsubstantiated

flat

rejection.

Current

speaker

is counter-

challenging a "So what" or "The X which you speak of has nothing to do with anything" type of challenge.

206

R. REICHMAN CCS.Comment.ContextuaI-Function.Method

=

'Flat-Rejection'

(4) After apply-expansion. The speaker is counterchallenging aK: opponent's claim that her/his argument does support the scope of relevance claimed by the speaker. For Some I, CCS.ContextuaI-Function M e t h o d { I } = 'Apply-Expansion'

(5) After a preceding challenge. The speaker is responding to a previous challenge, i.e., is counterchallenging. CCS.Goal 'Challenge' Or (CCS.Goal - 'Fix-Claim ~ And For Some I, (CCS.ContextuaI-Function .Co-Relator{I} .Goal 'Challenge') Or (CCS.Goal = 'Support' And (HEAD-CCS.Goal = 'Challenge' Or (HEAD-CCS.Goai = 'Fix-Claim' And For Some I, H EAD-CCS.ContextuaI-Function.Co-Relator{I}.Goal = 'Challenge')))).

(6) Further-Challenge. This is a second challenge to an opponent's claim, where this second challenge may have been immediately preceded by clue words like "All right, but", i.e., a subargument concession. Type-Further-Challenge = 'Challenge'

(7) No preceding propositional support. Opponent has never given propositional support to her or his preceding claim. CCS.Goal NE 'Support' Or CCS.Support-Fact = Nil Or CCS.SupportCS = Nil Or CCS.Suppc.rtCS.Support-Fact = Nil

(8) Some existing support of preceding claim. Some support was given to the claim being challenged, at minimum, the citation of some authority for the truth of the claim.

EXTENDED MACHINE-PERSON INTERFACE

207

CCS.Goal = 'Support' Or CCS.SupportCS NE Nil

(9) After expansion challenge. Current speaker is counterchallenging a "Y is what is important here" type of challenge. For Some l, CCS.ContextuaI-Function.Method{I} Irrelevance-By-Expansion' -

(10) After an epistemic claim. The modality of the claim being challenged is true/not true, or necessary/not necessary. CCS.Modality - 'Epistemic'

(11) After a New-Factor challenge. Counterchallenging a counterchallenge that implied that the preceding claim was not valid due to an attribute of the class or entity under discussion. For Some I, CCS.ContextuaI-Function.Method{I} 'New-Factor' -

(12) After an analogy. Responding or arguing with an analogy. For Some I, CCS.ContextuaI-Function.Method{I} 'Analogy' -

(13) Evaluative claim under discussion. CCS.Modality = 'Evaluative'

(14) Some outstanding discourse expectations. Expectation-List NE Nil

(15) In debate mode. Discourse-Mode = 'Debate'

(16) Expecting a demanded support for a preceding claim. Expectation.Associated-Constraints = 'Supply-Support-Claim' (17) S u p p o r t i n g a p r e c e d i n g c o u n t e r c l a i m . Type-Further-Challenge = 'Support' (18) S u b o r d i n a t i n g

to a Controlling

Re-Enter NE Nil (19) C o n c e d e d a s u p p o r t as f l a w e d . Rejected-Space.Goal = 'Support'

c o n t e x t space.

2118

R. REICHMAN

(2/)) Some authority for the claim has been cited. CCS.Authority NE Nil Or HEAD-CCS.SupportCS.Authority NE Nil

(21) Just generated support for an implicit challenge. HEAD-CCS.MOde = 'Implicit'

Appendix B. Complete Listing of Module Portion Involved in Trace Produce- Next-Move

Choose (A: If Not(Expectation-List - Nil) Then Expectation ,-Choose(Expectation-List) Go Expectation.Function B: If Discourse-Mode NE 'Debate' Then Go Developmental-Choice/Step 1 C: If Discourse-Mode = 'Debate' Then Go Developmental-Choice/Step2 D: Go NonDevelopmentaI-Choice E: Go Transitional-Utterance) Developmental-Choice~Step 1 :

Choose (A: Go Start-Debate/Step 1 B: Go Step 2) Step 2:

Choose(A: B: C: D: E: F: G:

Go Further-Development-Choice Go Pre-Generalization-Choice Go Generalization-Choice Go Negative/Positive-Evaluative-Choice Go Explain-Choice Go Question-Choice Go Comment-Choice)

Start-Debate~Step 1 :

Discourse- Mode~---'Debate' Future-Defender*-Speaker side{I} .-(Speaker) Choose (A: Push Shift-Speakers B: Go Step 2) Step 2:

side {2}.-(Speaker) Sides*--(Side{1 },Side{2}) Go Challenge-Choice/Step 2

EXTENDED MACHINE-PERSON INTERFACE

Shift-Speakers: oldspeaker*-Speaker Speaker*Choose(Participant-List) oldspeaker APPEND Participant-List (POP) Challenge-Choice~Step 1 : Futu re-Defender*--Speaker Choose(A: Push Shift-Speaker-Expectation B: Knownside*-Expectation.Speaker Push Shift-Speaker-Same Side) Step 2: If Not(CCS.Comment.ContextuaI-Function.Method = 'Flat-Rejection') And (Expectation .Associated-Constraints NE 'Supply-Support-Claim' OR 'Supply-Support-Rejection') Then (Further-Challenge,Speaker,CCS) APPEND Expectation-List If Expectation.Associated-Constraints = 'Supply-Support-Claim' Then Expectation.Associated-Constraints*-Nil Go Support-Of Else IF CCS.Goal NE 'Support' Then if Not(Speaker MEMBER CCS.Antagonists) Then Speaker APPEND CCS.Antagonists Else If For Some I, Not(Speaker MEMBER

CCS.ContextuaI-Function.Co-Relator{i}.Antagonists) S.T. CCS.ContextuaI-Function.Method{I} NE 'lnference-Of'/'Derived-From' Then Speaker APPEND CCS.ContextuaI-Function.Co-Relator{I}.Antagonists Choose (A: Push Challenge-Directly/Step 1 B: Push Challenge-Indirectly) Step 3: Expectation,-[ Function*Challenge-Choice Speaker*-Future-Defender Context*-CCS] Expectation APPEND Expectation-List

209

21[)

R. R E I C H M A N

Expectation*--Nil Type- Further-Challenge*-Nil Go Produce-Next-Move Challenge-Directly~Step

1:

Choose( A: If Not(Expectation.Associated-Constraints 'Su pply-Su pport- Rejection') And Not(For Some I, CCS.ContextuaI-Function.Method{I} = 'Deny-Truth') And Not(CCS.Comment.ContextuaI-Fu nction Method 'Flat-Rejection') And Not(For Some I, CCS.ContextuaI-Function.Method{lt 'Apply-Expansion') Then Choose(l: Go Emotive-Flat-Relection 2: Push Irrelevance-Rejection If * EQ Nil THEN (POP)) B: If Not(Expectation.Associated-Constraints = 'Supply-Support-Rejection') And CCS.Comment.ContextuaI-FunctionMethod = 'Flat-Rejection' Then Go Demand-Support-Rejection C: If Not(Expectation.Associated-Constraints = 'Supply-Support-Rejection') And Not(CCS.Comment.ContextuaI-Function.Method 'Flat-Rejection') And (CCS.Goal NE 'Support' Or CCS.Support-Fact- Nil Or CCS.SupportCS = Nil Or CCS.SupportCS.Support-Fact = Nil) Then Go Demand-Support-Claim D: If Type-Further-Challenge EQ Nil And (CCS.Goal = 'Challenge' Or (CCS.Goal = 'Fix-Claim' And For Some I, CCS.ContextuaI-Function.Co-Relatodl}.Goal = 'Challenge') Or (CCS.Goal = 'Support' -

-

EXTENDED MACHINE-PERSON INTERFACE

And (Head-CCS.Goal = 'Challenge' Or (HEAD-CCS.Goal - 'Fix-Claim' And For Some I, HEAD-CCS.ContextuaI-Fu nction.Co-Relator{I}.Goal 'Challenge')))) Then Go Concede-Subargument E: If CCS.SupportCS NE Nil Then Choose(l: No-Op 2: space*-CCS Push Generating-Space CCS*-Last(CCS.SupportCS) CCS.State*-'Active') Else IF CCS.Goal = 'Support' Then Choose(3: No-Op 4: space*-CCS Push Close-Space CCS,~-HEAD-CCS CCS.State*-'Active' HEAD-CCS*-Nil) If Type-Further-Challenge EQ Nil Then If CCS.Method NE One-of('Modus-Tollens', 'Modus-Tollendo-Tollens*', 'Modus-Tollendo-Tollens') Then Speaker EXPRESS ("No") ("But") Else Speaker EXPRESS ("Well") ("Of course") ("But") -

Choose ( A: If CCS.Goal = 'Support' & CCS.ContextuaI-Function.Method NE 'Analogy' Then Push New-Challenge-Epi Push Challenge-Support-Specifics B: Push New-Challenge-Epi Push New-Factor C: Push New-Challenge-Epi Push Range-Application D: If For Some I, CCS.ContextuaI-Function Method{I} = 'New-Factor' Then Push New-Challenge-Epi Push Irrelevant-Factor E: Push New-Challenge-Epi Push Deny-Truth F: If For Some I, CCS.ContextuaI-Function Method{I} = 'Analogy'

211

212

R. RE1CHMAN

Then Analogous-Space~--CCS Push New-Challenge-Epi Push Challenge-Analogy-Mappings))

Step 2 (For Some I,

(ccs APPEND (IF HEAD-CCS.Goal = 'Support' & HEAD-CCS.ContextuaI-FunctionMethod NE "Analogy" Then HEAD-CCS.ContextuaI-Function.Co-Relator{I} .CounterSupports Else CCS.ContextuaI-Function.Co-Relator{I} .CounterClaims) S.T. CCS.ContextuaI-Function.Method{I} NE 'Derived-From' Or 'Inference-Of') (POP)

Generating-Space: space.States-'Generating' space.Focus.High APPEND space.Focus.Medium space.Focus.High~--Nil (POP)

Close-Space: space.State*-'Closed' space.Focus.Zero*-(space.Focus.Zero, space.Focus.Low, space.Focus.Medium, space. Focus.High) space.Focus.High,--space.Focus.Medium*-space.Focus.Low-~--Nil (POP)

New- Challenge-Epi : rcontextual-fu nction-~-[ co-relator ~CCS] rgoal ~'Challenge' rmode *-'Explicit' Push Construct-EPI nc~,--* Go Update-Support-Analogy-Challenge

Update-Support-Analogy-Challenge: IF Analogous-Space NE Nil Then space-,--Analogous-Space push Generating-Space Else (If Re-Enter = True Then H EAD-CCS.State*--'Controlling* ' Else space*-HEAD-CCS push Generating-Space

EXTENDED MACHINE-PERSON INTERFACE

CCS.State*-'Controlling' HEAD-CCS*-CCS CCS*-nc CCS.State*-'Active' (POP)

Challenge-Analogy-Mappings~Step 1: Choose( A: Speaker EXPRESS Propositions A,B S.T. For Some I,J, For Some Relation, R{K+ 1} A ASSERTS True(Analogous-Space.Mappings.X{I} R{K+ 1} Analgous-Space.Mappings.X{J}) And B ASSERTS NotTrue(Analogous-Space.Mappings.Y{I} R{K+ 1} Analogous-Space.Mappings.Y{J}) S.T. True(Analogous-Space.Mappings.Z{I} R{K+ 1} Analogous-Space.Mappings.Z{J}) Enable/Cause/Important-To Analogous-Space.Abstract.Relations{K}) B: Speaker EXPRESS Proposition A S.T. For Some I,J,K,L A Predicate Q on Analogous-Space.Mappings.X{J} S.T. Possible(Infer Relation Set, (R{K+ 1}... R{K+ L})) S.T. Q Implies 1. True(R{K+ 1}... R{K+ L}) In X{Max} & 2. Not(True(R{K + 1}... R{K + L})) In Y{Max} WHERE 1 & 2 Imply Not(Possible(Analogous.Space.Goal))

Step 2: SETSLOTS(CCS) Contextual-Function*-[ *-Challenge-Analogy-Mappings] Claim *-[ State-Of-Affairs~-A,B Epistemic-Predicate*-True] (POP)

A p p e n d i x C. T h e T r a c e

ENTERING: PRODUCE-NEXT-MOVE/STEP 1 CHOOSING B: Discourse-Mode = 'Discussion' NE 'Debate' THEREFORE going developmental-choice/step 1

213

214 ENTERING: DEVELOPMENTAL-CHOICE/STEP 1 CHOOSING A: go start-debate/Step 1 ENTERING: START-DEBATE/Step 1 Discourse-Mode,,-'Debate' Future-Defender~B side{l} *-(B) CHOOSING A: [pushing to SHIFT-SPEAKERS STEP 1 : oldspeaker*-B SPEAKER ~--C Participant-List~--(B,A,D) RETURNING TO: START-DEBATE/Step 2 side2,--(C) Sides~--((B),(C)) go challenge-choice/step 2 ENTERING: CHALLENGE-CHOICE/STEP 2 C2.Comment Nil & Expectation = Nil THEREFORE Expectation-List*-((Further-Challenge, C, C2)) Expectation = Nil & C2.Goal NE 'Support' & C2.Antagonists = Nil THEREFORE C2.Antagonists*-(C) CHOOSING A: pushing to CHALLENGE-DIRECTLY/Step t STEP 1 : CHOOSING G: [Type-Further-Challenge- nil & C2.ContextuaI-FunctionMethod NE 'Modus-Tollens' THEREFORE

C EXPRESS: 'No, but' * * * * * * * * * * * * * * * * * * * * * * *

C2.SupportCS = Nil & C2.Goal NE 'Support' THEREFORE No-Op C2.ContextuaI-Function.Method = 'Analogy' THEREFORE CHOOSING F [analogous-space*-C2 pushing to NEW-CHALLENGE-EPI

R. R E I C H M A N

EXTENDED MACHINE-PERSON INTERFACE Step 1 : rcontextual-fu nction*-[ co-relator ~ C 2 ] rgoal*-'Challenge' rmode*-'Explicit' pushing to CONSTRUCT-EPI ridentifier~C3 setslots(c3) RETURNING TO: NEW-CHALLENGE-EPI returned register value: ~ = C3 nc~--C3 going update-support-analogy-challenge ENTERING: UPDATE-SUPPORT-ANALOGY-CHALLENGE Analogous-Space = C2 NE Nil THEREFORE space*-C2 pushing to GENERATING-SPACE C2.State*-'Generating' C2.Focus.Medium *-(Film star, greasy fat millionaire, Western Tradition) C2.Focus.High.-Nil RETURNING TO: UPDATE-SUPPORT-ANALOGY-CHALLENGE CCS*-C3 C3.State*-'Active' RETURNING TO: CHALLENGE-DIRECTLY/STEP 1 pushing to CHALLENGE-ANALOGY-MAPPINGS STEP 1 : C EXPRESS: "There the woman is selling herself. = A In these societies the woman isn't selling herself, her parents are selling her." - B because: 1. For I - 2 , J - 4 , K = 1, 2. Z { 2 } - object:$woman X{2} = film star Y{2} = girl Z{4} = source: $people X{4} = film star Y{4} girl's family Relation, R{K+ 1} = IDENTITY 3. A = "There the woman is selling herself" ASSERTS True(X{2} R{K + 1} X{4})

215

216

R. RI~,IICHMAN

3. B

"In these societies the woman isn't selling herself, her parents are selling her" ASSERTS NotTrue[Y{2} R{K + 1} Yt4}) 4. True(Z{2} IDENTITY Z{4}) ENABLES BUY Step 2: setslots(c3) Contextual- Function~[ Function ,--Challenge-Analogy-Mappings] Claim *-[ State-of-Affairs ~-A,B Epistemic-Predicate~True] RETURNING TO: CHALLENGE-DIRECTLY/STEP 2 HEAD-CCS = C1 & C1 .Goal NE 'Support' THEREFORE C2.Cou nterClaim,--(C3) RETURNING TO: CHALLENGE-CHOICE/STEP 3 Expectation ,-[ Function *-Challenge-Choice Speaker ,--B Context ,--C3] Expectation-List~--((Challenge-Choice, B, C3), (Further-Challenge, C, C2)) Expectation *--Nil go produce-next-move

ACKNOWLEDGMENT I'd like to thank Bill Woods, Brad Goodman, and Cheri Adrian for reading and commenting on drafts of this work. REFERENCES 1. Allen, J., A plan-based approach to speech act recognition, Ph.D. Thesis, Universily of Toronto, Ont., 1979. 2. Austin, J.L., How to Do Things with Words (Oxford University Press, Fair Lawn, N J, 1962). 3. Bartlett, F.C., Remembering (Cambridge University Press, New York, 1932). 4. Bennett, J., Linguistic Behavior (Cambridge University Press, New York, 1976). 5. Brewer, W., Literary theory, rhetoric, and stylistics: Implications for psychology, in: R. Spiro, B. Bruce and W. Brewer (Eds.), Theoretical Issues in Reading Comprehension (Erlbaum, Hillsdale, N J, 1980) 221-237. 6. Bruce, B., Analysis of interacting plans as a guide to the understanding of story structure, Poetics 9 (1980) 295-311. 7. Chafe, W., Givenness, contrastiveness, definiteness, subjects, and topics, in: C. Li (Ed.), Subject and Topic (Academic Press, New York, 1976). 8. Chomsky, N., Aspects" of a Theory of Syntax (MIT, Cambridge, MA, 1965). 9. Chomsky, N., Rules and representations, The Behavioral and Brain Sciences 3 (1980) 1-61. 10. Clark, H. and Carlson, T., Hearers and speech acts, Language 58 (1982) 2. 11. Clark, H. and Marshall, C., Definite reference and mutual knowledge, in: A K . Joshi, I. Sag and B. Webber (Eds.), Linguistic Structure and Discourse Setting (Cambridge University Press. New York, 1981).

EXTENDED MACHINE-PERSON INTERFACE

217

12. Cohen, P. and Perrault R., Elements of a plan-based theory of speech acts, Cognitive Sci. 3 (1979) 177-212. 13 Cohen, R., Understanding arguments, Proc. CSCSL Canadian Society for Computational Studies of Intelligence, 1980. 14. Collins, A. and Loftus, E.F., A spreading-activation theory of semantic processing, Psychol. Rev. 82 (1975) 407-428. 15. Collins, A. and Quillian, M., Retrieval time from semantic memory, Jr. Verbal Learning and Verbal Behavior 8 (1979) 240-247. 16. Dore, J., Children's illocutionary acts, in: R. Freedle (Ed.), Discourse Production and Comprehension (Ablex, Norwood, N J, 1977) 227-244. 17. Goodman, B., A model for a natural language data base system, R-798, University of Illinois, Urbana, IL, 1977. 18. Grice, H.P., Logic and conversation, in: P. Cole and J. Morgan (Eds.), Syntax and semantics (Academic Press, New York, 1975) 41-58. 19. Grimes, J., Topic levels, in: Theoretical Issues in Natural Language Processing 2 Association for Computational Linguistics, 1978. 20. Grosz, B., The representation and use of focus in dialogue understanding, Doctoral Dissertation, University of California at Berkeley, CA, 1977. 21. Halliday, M.A.K. and Hasan, R., Cohesion in English (Longman, London, 1976). 22. Hasan, R., Text in the systemic-functional model, in: W. Dressier (Ed.), Current Trends in Textlinguistics (De Gruyter, Berlin, 1977). 23. Hobbs, J., A computational approach to discourse analysis, City University of New York, 1976. 24. Hymes, D., Language in Culture and Society (Harper & Row, New York, 1964). 25. Jefferson, G., Sequential aspects of storytelling in conversation, in: J. Schenkein (Ed.), Studies in the Organization of Conversational Interaction (Academic Press, New York, 1978). 26. Jefferson, G. and Schenkein, J., Some sequential negotiations in conversation, in: J. Schenkein (Ed.). Studies in the Organization of Conversational Interaction (Academic Press, New York, 1978) 134-155. 27. Karmiloff-Smith, A., Language as a formal problem-space for children, N I A S Child Language Conference: Beyond description in child language, 1979. 28. Kuno, S., Empathy and syntax, in: S. Kuno (Ed.), Harvard Studies in Syntax and Semantics (Harvard University Press, Cambridge 1975) 1-73. 29. Kuno, S., Generative discourse analysis in America, in: W. Dressier (Ed.), Current Trends in Textlinguistics (De Gruyter, Berlin, 1977). 3(I. Labov, W. and Fanshel, D., Therapeutic Discourse (Academic Press, New York, 1977). 31. Levinsohn, S.H., Progression and digression in Inga (Quechuan) discourse, Forum Linguisticum 1 (1976) 122-147. 32. Levy, D., Communicative goals and strategies: Between discourse and syntax, in: T. Givon (Ed.), Syntax and Semantics 12 (Academic Press, New York, 1979). 33. Lewis, D.K., Convention (Cambridge University Press, New York, 1969). 34. Linde, C., Information structures in discourse, Ph.D. Thesis, Columbia University, New York, 1974. 35. Linde, C. and Gougen, J., Structure of planning discourse, J. Social Biol. Struct. 1 (1978) 219-251. 36. Longrace, R., Hierarchy and Universality of Discourse Constituents in New Guinea Languages: Discussion (Georgetown University Press, Washington, DC. 1972). 37. Longrace, R. and Levinsohn, S., Field analysis of discourse, in: W. Dressier (Ed.), Current Trends in Textlinguistics (De Gruyter, Berlin, 1977). 38. Meyer, B.J.F., Following the author's top-level structure, in: R. Tierney, J. Mitchell and P. Anders (Eds.), Understanding Reader's Understanding (Erlbaum, Hillsdale, NJ, to appear). 39. Minsky, M., A framework for representing knowledge, in: P. Winston (Ed.), The Psychology of Computer Vision (McGraw-Hill, New York, 1975). 40. Morgan, J. and Sellner, M., Discourse and linguistic theory, in: R. Spiro, B. Bruce and W. Brewer (Eds.), Theoretical Issues in Reading Comprehension (Erlbaum, Hillsdale, N J, 1980) 165-201.

218

R. REICHMAN

41. Newman, D., Children's understanding of strategic interaction, Ph.D. Thesis~ C t v t~nivcrsity of New York, 1980. 42. Power, R., The organization of purposeful dialogues, Linguistics 17 (1979) 107-15~ 43. Pratt, M.L., Towards a Speech Act Theory qf l, iterarv Discourse (Indiana Unb
R e c e i v e d J u l y 1982; revised version received A u g u s t 1983