Spoken dialogue processing method using inductive learning with genetic algorithm

This paper describes a spoken dialogue processing, which includes the learning using spoken dialogue examples. Most of the spoken dialogue systems up to the present are task-oriented, where the processing is based on the prespecified generation rules and database. It is then difficult to handle various topics, such as miscellaneous talks in daily dialogue. In the proposed method, the dialogue between the system and the user is processed as the spoken dialogue example, and the rule is acquired based on the pair of system response and user utterance, through inductive learning using the genetic algorithm. In other words, it is not necessary to prepare the training data beforehand, but the response is composed using the rules acquired from the actual dialogue examples. With this approach, the learning can be executed based on the dynamic data, and the deviation depending on the data is reduced. It is intended in this paper to examine the usefulness of the proposed method. Miscellaneous talks are considered. ELIZA-type system, which is extended to the spoken dialogue, and the proposed method are compared through a comparison experiment and an experiment with multiple examinees. As a result, it is verified that the total ratio of the correct and almost correct responses is improved from 66.3% to 76.1%, and the effective responses can be constructed using the rules acquired from actual dialogue examples. This improvement of 9.8 points indicates that the proposed method is effective in handling miscellaneous talks. © 2004 Wiley Periodicals, Inc. Syst Comp Jpn, 35(12): 67–82, 2004; Published online in Wiley InterScience (www.interscience.wiley.com). DOI 10.1002/scj.10204