Modelling dialogue: Challenges and Aproaches

Modelling dialogue, that is, designing formal systems that reproduce aspects of natural conversation, is a challenging task. Not only must a dialogue model be able to handle most (if not all) of the linguistic phenomena that make monological discourse hard to model (e.g., anaphora, presuppositions, implicatures), there are also additional requirements: unlike (written) text, dialogue unfolds in time, and timing matters; and unlike monologue, a dialogue is an interaction, and is shaped by the interests and capabilities of all dialogue participants (DPs), while it unfolds. Describing these challenges is one aim of this overview paper; the other is reviewing some of the extant approaches to adressing them in dialogue models. In the very general sense of the term as used above, all dialogue systems, i.e., computer systems that interact via natural language with users, are models of dialogue. Traditionally, however, the term has been mostly employed for those modules within dialogue systems that handle what may be called content-flow aspects of dialogue, organising the coherence of utterances and cooperation on the task. Other, more interactional aspects such as the timing of utterances or cooperation on the dialogue itself have been dealt with by other modules, often in a less principled manner. As we argue, observable interactions between these phenomena indicate that a better strategy might be to combine models of these aspects. This, then, is the structure of the paper: in Section 2 we review some of the phenomena observable in human-human dialogue that make it a challenging subject for study, such as the organisiation of whose turn it is to speak, strategies for recovery from communication problems, and regularities in the “flow” of the dialogue. In Section 3 we review approaches that have been taken to modelling these phenomena. We close with a brief discussion of shortcomings of current models and implementations and possible ways to overcome them.