Gradual probabilistic DFA learning with caching for conversational agents
Masayuki Okamoto · Systems and Computers in Japan · 2004
Abstract This paper proposes a method of reducing the cost of gradually constructing task‐oriented conversational agents with FSMs by collecting example dialogues. Probabilistic‐DFA learning algorithms with the state merging method are available for the FSM‐based dialogue model. However, these algorithms must learn the whole model again as often as the example data increase. We proposed a learning technique which decreases the recomputation cost by caching the history of the state merging process. Simulation and actual use of a conversational agent with 110 tour‐guide dialogues showed that the method which recomputes only compatibility‐changed states reduced the total compatibility‐checking cost by 33.0% without significantly changing the quality of learned models. © 2004 Wiley Periodicals, Inc. Syst Comp Jpn, 35(7): 24–32, 2004; Published online in Wiley InterScience ( www.interscience.wiley.com ). DOI 10.1002/scj.10653