[3Yin2] Interactive session 1

[3Yin2-01] Construction of a dialogue corpus for utterance generation based on user information and dialogue context

〇Yuiko Tsunomori^1,2, Ryuichiro Higashinaka¹ (1.Nagoya University, 2.NTT DOCOMO, INC.)

Keywords:dialogue system, non-task-oriented dialogue system, utterance generation, user information, personalization

Chat-oriented dialogue systems can improve familiarity by incorporating user information (preferences, experiences, etc.) into system utterances. However, the previous method, which utilizes user information to fill in utterance templates, has a problem in that the utterances lack naturalness. In this research, we address the problem by generating system utterances based on both the dialogue context and the user information. We extracted user information from the Osaka University Multimodal Dialogue Corpus (Hazumi) and collected utterances that take into account both user information and the dialogue context by crowdsourcing in order to construct a dialogue corpus of about 10,000 sets. We then analyzed the collected utterances and constructed baseline models for generating system utterances. The results of automatic and subjective evaluation demonstrated that system utterances based on the user information could be generated while retaining the naturalness of the dialogue context.