Modelling Simple Dialogs

Introduction

A problem that is currently of some interest in connection with the development of conversational systems is that of dealing with the context in Which the individual utterances of a dialog are embedded. As researchers like Bruce (Bruce (1975) l , Deutsch (Deutsch (1975), and Schank and Abelson (Schank and Abelson (1975)) have pointed out, the plans of the speakers are a crucial element of the context of a dialog.

Knowledge of how speakers communicate intentions via their utterances seems, then, to be Lmportant to the success of a conversational system. 'This problem has been addressed by workers in disciplines other than Artificial Intelligence. The encoding of speaker intention in single utterances has been studied by Searle (Searle(l975)) in his work on 'speech acts'. Commun1cat1on of speaker intention via larger discourse units has been studied by the linguists S1ncla1r and Coulthard (Sinclair and Goulthard(l975)) in their work on the structure of pedagogical dialogs. ·

Our effort involved applying the work of Searle and that of Sinclair and Coulthard to the problem of understanding the speaker intention communicated in utterances of actual didogs from a simple dana in.

The Dialog I<'OOel

We developed and tested a model for about one hundred dialogs which we recorded between passengers and a clerk at an information booth in a train station. By inspection of the dialog transcripts, we discO'.rered a small set of discourse 'acts', each of which indicates a WE'll­ defined speaker intention. For instance, one such 2ct, called an 'ack.heard.ok', indicates that the speaker believes he correctly heard the 1mmed1ately preceding utterance.

In general, the definition of an act may involve preceding utterances, and may express expectations about the acts to follow. We describe these interrelationships in terms of higher-level discourse units, called 'moves', which are themselves interrelated and described in terms of each other. Many of these moves consist in the speaker uttering an act and then listening for a response, to determine whether the act in fact achieved the purpose for which it was intended. our description, then, leans strongly to the view, suggested especially by the work of Bruce, that the structures we observe in our dialogs result from the structures of the interacting plans of the participants in those dialogs. Utterances are intended either to further the speaker's own plan or to prov1de feedback to the other participant on the status of his (inferred) plan.

The system of description can be presented using a 'grammar' which describes the structure of each of our moves in terms of acts and other mOves.

We represented the individual utterances of the dialogs by specifying their illocutionary point and propositional content. Then, using the grCilTD1lar, we developed a canput:er program which generates detailed expectations about both these features of the next utterance in the dialog. The program uses the success or failure of these expectations to determine what role the utterance plays in its dialog (i.e. what act it represents, and Where it falls in the hierarchy of mOves) · The program maintains simple models of both the clerk's knowledge and the passenger's (inferred) knowledge.

Results

. Several test dialogs were translated by hand 1nto the representation expected by the model. The output from the dialog model in each case was a dialog history, with comments appended to each utterance indicating the role it played in the dialog. A brief sample of the output follows (Note that the utterance . representation is not shown here.):

1. P: 'What gate is the train to London at?' 11 Question asked in process of acquiring information.
2. C: 'It's at three--at two fifteen.' // Question asked in process of acquiring information--begins clarification dialog on utterance at index 1.
3. P: 'Yeah.' /I Reply to previous question at index 2.
4. C: 'Gate eight.' // Reply to previous question at index 1.

The model performed well on a set of test dialogs representotive of the variations permitted by the grarrrnar. Some dialogs among those recorded did not fit into any structure described by the grammar, and so were incomprehensible to our model. ~~vertheless, the sort of dialog on which the model could perform successfully seemed to comprise the majority (about two thirds) of the transcribed dialogs; this would seem to indicate thzt our approach to modelling discourse structure holds promise as a basis for future research.