[3R5-OS-13c] OS-13

[3R5-OS-13c-04] Learning utterance timings including proactive utterances during dialogue

〇Junya Nakanishi¹, Ryo Miyoshi², Yuki Okafuji³, Jun Baba³, Yuichiro Yoshikawa¹, Hiroshi Ishiguro¹ (1. Osaka university, 2. Chukyo university, 3. CyberAgent Inc.)

Keywords:utterance timing, proactive utterance, transformer

A computer system provides dialogue services basically with a passive process in which a user makes a request through text or voice, and a system responds to the request by generating an appropriate response. On the other hand, ``proactive dialogue services'' that not only responds to user requests but also creates opportunities to proactively provide information and services by estimating user needs and situations is also important to uncover opportunities to provide services that meet users' hidden desires. One of the difficulties of active dialogue services is determining the timing of proactive utterances. It is desirable to observe the user's activity and not disturb or offend him or her. The purpose of this study was to construct a model for estimating the timing of utterances, including proactive utterances, during a dialogue. For this purpose, we considered the situation of being in a dialogue and the appropriate timing of utterances in that situation, and based on that, we created and evaluated an estimation model. The overall accuracy was over 80\%, but the accuracy of the timing of proactive utterances was low, and issues that need to be considered became clear.