Rapport-building dialogue strategy enhanced with proactivity, personalization, and aizuchi backchannels, and its automatic evaluation in HRI

Rapport is a critical factor in interactions, strongly influencing user experience, engagement, and interaction outcomes. Proactive behavior, dialogue personalization, and backchannel responses play key roles in fostering rapport, preventing conversation stalls, and maintaining natural conversational flow. This study integrates these behaviors into a rapport-building dialogue strategy by prompting large language models using the CO-STAR and few-shot frameworks in a human–robot interaction setting. To evaluate dialogue quality and rapport, we conduct behavioral and subjective human evaluations and propose 2S-FLARE, a Two-Stage Feature-Led LLM-Driven Automated Rapport Evaluation framework. 2S-FLARE first extracts interaction-level features grounded in psychological theories of rapport and then performs LLM-based judgment to assess rapport. Experimental results demonstrate that the proposed dialogue strategy significantly enhances engagement and rapport, and that 2S-FLARE shows strong alignment with human judgments from both first-person and third-person perspectives.