A Conversational Agent for Motivational Interviewing with Summary Prediction

Abstract

Motivational Interviewing (MI) is a counseling approach that aims to elicit clients' intrinsic motivation for change, where generating summary utterances at appropriate timing plays a crucial role. This study proposes a dialogue system for MI that incorporates a summary prediction model trained on MI corpora, together with a summary generation module for generating summaries and a basic response generation module for producing other types of counselor responses. Then, the proposed dialogue system is integrated into a conversational agent with speech input/output capabilities and CG-based nonverbal behaviors. In the system evaluation, we compared the proposed system with three baselines. First, we collected human-agent interaction data under four conditions. We then conducted a perception study on the collected interactions. In addition, we computed evaluation metrics as objective measures by analyzing the interactions based on the MITI (Motivational Interviewing Treatment Integrity) framework. The perception study demonstrated that the proposed system improves the naturalness and motivational impact of summaries. Moreover, the objective measures indicate that the proposed method increases the proportion of complex reflections and reduces MI non-adherent behaviors, suggesting improved adherence to MI principles. These findings suggest that summary timing prediction improves the quality of MI dialogues; however, further improvements are needed in both summary prediction and the agent's nonverbal behavior to better facilitate client behavior change.