We propose a learning method of speaker's taste information from natural language dialogs. In this paper, we make favorable value changing situations for detecting good/bad actions for the agent. To find out these situations, we use the knowledge structure acquired from a dialog. In order to acquire the knowledge structure from a dialog, we construct the knowledge from each utterance using Truth Maintenance System while avoiding any contradictions. In order to verify this method, we provide the result of our experiment which attempts to learn new favorable values from a naturally-occurring dialog corpus.

奨励賞受賞