Pragmatic interpretation and dialogue management in spoken-language systems

This paper describes the design of pragmatic-interpretation and dialogue-management modules in an automatic enquiry system that can be consulted through spoken natural language over the telephone. The system is designed around a central multi-level data structure representing the discourse that has unfolded during the dialogue. At the highest level of this discourse representation the information exchange is represented as a series of information-state changes or updates. Several conditions in the information state itself give rise to actions of the dialogue manager. The dialogue manager is designed to achieve the user's goal in a manner that is understandable to the user, efficient and correct. This is not a trivial problem because natural language and, in particular, speech understanding lead to many uncertainties. To deal with uncertain information, we have designed feedback and verification mechanisms and means for contextual understanding, underspecification and pragmatic inferencing.