Shuji DOSHITA, Masahiro ARAKI, Tatsuya KAWAHARA
Department of Information Science, Kyoto University
Sakyo-ku, Kyoto 606-01, Japan
e-mail: doshita@kuis.kyoto-u.ac.jp
We present a computational model of spoken dialogue. In our model, we pay attention to following three points: (1) the recognition process of user's utterance and the understanding process of user's intention, (2) updating system's belief according to the user's intention and generating appropriate system's response in order to achieve the user's plan, (3) how to utilize the prediction of next utterance in speech recognition that is made from present dialogue situation. In the stage of understanding user's intention, we use conversational space, that represents user's behavior in purpose oriented dialogue. These conversational space is constructed by dialogue corpus. In generating system's intention and response, we use event hierarchy and system's mental state. Event hierarchy represents the hierarchical relations of plans and decomposed relation of plan and actions. By using this knowledge, dialogue system can infer the most appropriate way of achieving plan. System's mental state represents system's belief, that is used for preconditions and effects of plans or actions. Therefore, system's intention is generated by system's belief, plan decomposition and act pair, that extends the idea of utterance pair. System's response utterance is generated from this intention by template base method of sentence generation. The prediction of next utterance is generated by topics management system. Topics manager extracts current topics of dialogue by case frame information and pursues them by cue phrase, that indicates continuation and termination of current topics. Topics manager also use event hierarchy to predict next utterance. From current user's speech act and system's generated intention, minimal cover of event hierarchy is inferred. These covered event and decomposed plans are the predicted topics. If some detriment is appeared in dialogue, priorities are given to these events following detriment resolution method. The prediction of next utterance in speech understanding stage is realized by connecting these predicted events and word concepts of semantic network, that is used keyword-driven speech parser.
Keywords: dialogue model, plan recognition, topics, speech understanding