日本語大規模言語モデルにおける発話意図の習得と共感対話システムの構築

抄録

This paper proposes a novel framework for empathetic dialogue generation in Japanese large language models (LLMs) by employing LoRA-based adapter tuning. The proposed system incorporates both cognitive and emotional empathy through two key modules: the Next Intent Predictor and the Intent Combiner. These modules dynamically predict, select, and integrate multiple intent representations, enabling empathetic and user-centered responses tailored to the context. Evaluation results highlight significant improvements, including a notable increase in user-friendliness and empathetic quality compared to the base model, achieving a win rate exceeding 70%. Furthermore, the system demonstrates human-level competitiveness, with over 40% of its responses being rated as comparable to human-generated ones. This study demonstrates the effectiveness of combining reinforcement learning with intent-specific fine-tuning to enhance empathetic capabilities in dialogue systems. By focusing on intent-driven responses, this approach paves the way for more user-centric applications, particularly in privacy-sensitive and culturally nuanced contexts.