[4F5-OS-29c] OS-29

4:00 PM - 4:15 PM JST(7:00 AM - 7:15 AM UTC)

[4F5-OS-29c-03] Emotion-Based Memory Management for Long-Term Dialogue by Large Language Models

〇Yuri Minamitani¹, Masatoshi Nagano^1,2, Tadahiro Taniguchi^1,3 (1. Kyoto University , 2. The University of Electro-Communications, 3. Ritsumeikan University)

Keywords:

Large Language Model,Emotion Estimation,Long-Term Dialogue System

The importance of communication in healthcare and home environments is widely recognized. In particular, appropriately understanding conversational content is known to foster trust and a sense of security in others. In recent years, large language models (LLMs) have demonstrated human-like conversational abilities, raising expectations that they can take on the social roles to help address labor shortages. However, conventional LLMs are insufficient to be in charge of such social roles because it is difficult for them to have long-term memory, and they generate hallucinations or contextually inconsistent responses. Therefore, this paper proposes a system that enables LLMs to infer, remember, and forget important information from long-term dialogues based on psychological insights. Specifically, we construct an algorithm that estimates emotion-related sentences from user dialogues and manages memory accordingly. To evaluate the effectiveness of the proposed method, we conducted question-answering tasks on various long-term dialogue datasets. The results demonstrated that our approach outperforms previous methods in accuracy and clear advantages in managing important information from long-term dialogues.