A Spoken Dialogue System Based on FST and DBN

Natural language understanding module and dialogue management module are important parts of the spoken dialogue system. They directly affect the performance of the whole system. This paper proposes a novel method named action-group finite state transducer (FST) model to cope with the problem of natural language understanding. This model can map user utterances to actions, and extract user’s information according to the matched string. For dialogue management module, we propose dynamic Bayesian network (DBN) model. It can reduce the demands for the corpus compared with Markov decision process (MDP) model. The experiments on the action-group FST model and DBN model show that they significantly outperform the state-of-the-art approaches. A set of subjective tests on the whole system demonstrate that our approaches can satisfy most of the users.