Construction of a Speech Dialogue System Using Multiple Response Generation Models and Consideration of its Dialogue Selection Method

Abstract

Recently, dialogue models have been constructed using a large amount of data to realize naturalness. Next step, integration of such response generation models and task-oriented models will become an issue. In this paper, we describe our speech dialogue system that integrates two dialogue generation models and task-oriented base model. The two models consist of NTT ’s dialogue generation model for chatting, which was created from a large number of dialogue data and our dialogue generation model that responds to nouns spoken by the user. The rule base model consists of set of rules that created from Welfare Ministry ’s Web Q/A information on coronaviruses. We also introduce a dialogue candidate selection method that filters response candidates generated by our system and rescores them. We evaluated the selection method using Likert scale, showing that the reranking mechanism is effective.