User modelling in adaptive dialogue management

This paper describes an adaptive approach to dialogue management in spoken dialogue systems. The system maintains a user model, in which assumptions about the user’s expectations of the system are recorded. Whenever recognition errors occur, the dialogue manager drops assumptions from the user model, and adapts its behaviour accordingly. The system uses a hierarchical slot structure that allows generation and interpretation of utterances at various levels of generality. This is essential for exploiting mixed-initiative to optimise effectiveness. The hierarchical slot structure is also a source of variation. This is important because it is ineffective to repeat prompts in cases of speech recognition errors, for in such cases users are likely to repeat their responses also, and thus the same speech recognition problems reoccur.