Extracting Domain Knowledge for Dialogue Model Adaptation

Abstract

Domain shift is a challenging issue in dialogue management. This paper shows how to extract domain knowledge for dialogue model adaptation. The basic semantic concepts are derived from domain corpus by iterative token combination and contextual clustering. Speech act is identified by using semantic clues within an utterance. Frame states summarize current dialogue condition and state transition captures the mental agreement between users and system. Both Bayesian and machine learning approaches are experimented in identification of speech act and prediction of next state. To test the feasibility of this model adaptation approach, four corpora from domains of hospital registration service, telephone inquiring service, railway information service and air traveling information service are adopted. The experimental results demonstrate good portability in different domains.