In our implementation, the static information states contains, among other things, a set of dialogue plans which contain information about what the system should do in order to achieve its goals. Traditionally [Allen and Perrault1980], it has been assumed that general planners and plan recognizers should be used to produce cooperative behaviour from dialogue systems. On this account, the system is assumed to have access to a library of domain plans, and by recognizing the domain plan of the user, the system can produce cooperative behaviour such as supplying information which the user might need to execute her plan. Our approach is to directly represent ready-made plans for engaging in cooperative dialogue and producing cooperative behaviour (such as answering questions) which indirectly reflect domain knowledge, but obviates the need for dynamic plan construction.
Typically, the system has on the agenda an action to respond to a question. However, the move for answering the question cannot be selected since the system does not yet have the necessary information to answer the question. The system then tries to find a plan which will allow it to answer the question, and this plan will typically be a list of actions to raise questions; once these questions have been raised and the user has answered them, the system can provide an answer to the initial question. This behaviour is similar to that of many natural language database interfaces, but the difference is that the architecture of our system allows us to improve the conversational behaviour of the system simply by adding some new rules, such as the accommodation rules described above.