Development of a Dialogue System That Selectively Shares Information about Others Based on Interest Level

In recent years, there has been growing interest in dialogue systems designed to support the formation ofinterpersonal relationships. Previous studies have proposed dialogue systems that engage in one-on-one interactionwith a user and indirectly support relationship formation by sharing information about a third person . However, itremains unclear what types of information about such a person should be shared to effectively support relationshipformation. To address this issue, we aim to develop a dialogue system that dynamically selects and shares informationestimated to have a high interest score. As a first step, we construct an interest estimation model by fine-tuning alarge language model, resulting in an improvement of estimation accuracy by 45.4% compared to a baseline model.Next, we propose Interest-RAG, a response generation framework that integrates retrieval-augmented generation(RAG) with the interest estimation model, enabling the system to share interesting information while maintainingconversational naturalness. To evaluate our approach, we implemented it in a LINE chatbot and conducted a dialogueexperiment. The results show that our proposed system increases users’ interest the other person and enhances theirwillingness to communicate with that person, demonstrating its effectiveness in supporting relationship formation.Future work will focus on long-term deployment of the proposed system in real-world settings and on evaluating itssustained effectiveness through extended dialogue experiments.