Adaptive dialogue strategy based on estimated willingness in interview robot dialogue systems

Abstract

The goal of this research is to develop a communicative robot as a partner that can share what users want to talk about and what they are interested in. To achieve this goal, this study proposes an interviewer robot that adapts topics according to the user's multimodal attitudes. The robot estimates whether a topic change or continuation is appropriate based on the multimodal features of the user during the dialogue, and generates questions based on the estimated results using a large-scale language model. In this paper, we examine the effects of the multimodal topic continuation recognition model and adaptive question generation on the realization of this robot system. First, we trained a topic continuity model using the dialogue corpus "Hazumi," which contains multimodal behaviors of users in a dialogue between a human and a virtual agent. Next, we conducted a dialogue experiment using a robot equipped with the trained model to evaluate the effect of adaptive question generation on the dialogue results.