An adaptive domain knowledge manager for dialogue systems

This paper describes the recent effort to improve our Domain Knowledge Manager (DKM) that is part of a mixed-initiative task based Spoken Dialogue System (SDS) architecture, namely to interact within an ambient intelligence scenario. Machine-learning applied to SDS dialogue management strategy design is a growing research area. Training of such strategies can be done using human users or using corpora of human computer dialogue. However, the size of the state space grows exponentially according to the state variables taken into account, making the task of learning dialogue strategies for large-scale SDS very difficult. To address that problem, we propose a divide to conquer approach, assuming that practical dialogue and domain-independent hypothesis are true. In this context, we have considered a clear separation between linguistic dependent and domain dependent knowledge, which allows reducing the complexity of SDS typical components, specially the Dialoguer Manager (DM). Our contribution enables domain portability issues, proposing an adaptive DKM to simplify DM's clarification dialogs. DKM learns, through trial-and-error, from the interaction with DM suggesting a set of best task-device pairs to accomplish a request and watching the user's confirmation. This adaptive DKM has been tested in our domain simulator.