Knowledge sources in spoken dialogue systems

We describe how the architecture of the modularized linlin dialogue manager can be augmented to handle the cases where the user's initial information request is underspeci ed, and where therefore the system needs to ask the user about the missing pieces of information. In many existing spoken dialogue systems this has been handled by a so-called task model, which has been embedded in the dialogue move management. The present paper makes two points. First, that the term 'task' is ambiguous, since it can refer to the user's underlying reasons for engaging in the dialogue, to the user's information seeking activity, and to the system's actions when preparing for providing the user with an answer to the posed question. Second, we argue that this aspect of dialogue management should be managed by a separate module of the system, both because our previous work has shown that it is not necessary in all cases, and because it is conceptually di erent from general interaction management knowledge. For these reasons we advocate using a separate knowledge module called an Information Speci cation Form for managing these cases.