Online Adaptation of Language Model on Speech Dialogue System

摘要

The performance of the conventional N-gram falls remarkably at different task. And sufficient dialog corpus for study cannot go into a hand easily. we examine the technique of on-line adaptation of the language model in a speech dialog system by using the history of a dialog. Adaptation is performed based on the PPM method and language probability is updated by maximum likelihood estimation from a dialog history. By using mixed probability with the conventional non-adaptive N-gram, especially when there were few dialog histories, the performance was able to be compensated. Moreover, quicker adaptation was performed by using the utterance not only a user but by the side of a system as a dialog history.