Investigating Effects of Multimodal Topic-continuance Recognition on Human-Robot Interviewing Interaction

This study's long-term goal is the development of a communication robot as a partner that can keep talking about specific things about which the user would like to talk and in which they are interested. To achieve this goal, we developed an interviewer robot that adapts topics based on the user's multimodal attitudes. The robot, utilizing the Japanese GPT-NeoX-3.6, selects questions based on the estimated topic continuance level. We regard the topic continuance level as the degree of the user's speaking willingness (willingness to continue the current topic). This paper aims to validate the multimodal topic continuance recognition model and its adaptive question selection strategy. First, we trained the model on the "Hazumi" dialog corpus, which includes user multimodal behavior in human-virtual agent interactions. Second, 10 participants were interviewed with the robot equipped with the trained model. After the interviews, we asked the participants if the topic continuance/change by the robot was appropriate and validated the estimation accuracy.