Introduction to the Discussion Section

As soon as the aim of natural language systems extends beyond the mere processing of isolated sentences or interaction with the user through isolated questionanswer pairs, it becomes necessary for the system to maintain some sort of record of both the user's and the system's previous dialog contributions. In its simplest form, such a record would contain the objects being mentioned, and the recency of them being mentioned (the notion of "object" is used here in a broad sense, also denoting actions, events, states, etc.). Such a record may be exploited, for example, in resolving or generating anaphora, which in general may only refer to those objects that have been mentioned recently. Also the choice between definite and indefinite articles depends, among other factors, upon whether or not the objects being referred to have already been introduced in the previous dialog. A slightly improved dialog record would not only contain the objects that have been mentioned, but also the (conceptual representation of) the dialog contributions themselves. This can be exploited, for example, in the analysis of argumentative structures in discourse, that is, for determining why one argument supports some other argument. More advanced forms would include representations of the current and previous dialog focuses, themes, dialog substructures, and dialog rules employed by the system, etc. The notion of "discourse model" has recently found wide acceptance for referring to this representation of the structure of a discourse. However, when the user mentions an object in a dialog contribution, the system can assume that the user knows (or believes in the existence of) the mentioned object or the properties being attributed to it. These assumptions about the user must be taken into account in generating cooperative dialog contributions and certainly form part of the user model. For instance, from the UNIX commands the user mentions, the help system of Chin (1988) can draw assumptions about his/her level of expertise. Or, when the user asks, "Who is the present king of France?", a cooperative system might check the correctness of the user's belief that can be inferred from his/her question. Conversely, when the system mentions some fact, it can also assume that, from now on, the user knows this fact and regard it as being common knowledge, for instance within explanations. Thus there seems to be some relationship between user models and discourse models, at least in view of the fact that entries that should be part of the discourse model might also possibly be contained in the user model in a similar form. At the user modeling meeting in Maria Laach, a talk by Ethel Schuster about this relationship provoked a lengthy discussion. Opinions that were propounded included that the discourse model and the user model are separate, but related to each other; that the discourse model is part of the user model; that the user model is part of the discourse model; that user model and discourse model are two sides of the same coin; and also that a discussion on the relationship doesn't make any sense since, as soon as the two concepts have been precisely defined, the problem will disappear. It was agreed, however, that the discussion should be continued in written form in order to better pinpoint what issues are controversial and which ones are undisputed with respect to the question at hand. The discussion at the Maria Laach workshop was recorded on tape, so that transcripts could be prepared and distributed among the participants. Also, the first draft of each written contribution was circulated, so as to give the discussants the chance to respond to the rebuttals of their colleagues. Moreover, in order to filter out weaker contributions, all papers were subjected to an internal rating by all participants and reviewed by the editorial board of Computational Linguistics. In order to increase the coherency of the discussion and to avoid redundant terminological definitions, two papers were chosen to form a common point of reference, namely Schuster's contribution and the paper by Grosz and Sidner (1986). The authors were told to assume that readers are familiar with these two works and to refer to them whenever possible in order to avoid redundancy. The papers should not serve as a target of repeated critique (as, for example, in the BBS journal), but form a common basis for the discussion. Moreover, contributors were asked to define, as precisely as possible, in what sense they used the terms user model and discourse model. Most discussants regard the user model as containing " the system's beliefs about its users" (Schuster). Kobsa and Wahlster require that, in addition, "the user model be separable by the system from the rest of the system's knowledge" (Kobsa). And Cohen claims that " the user model must