[4K1-GS-9] Human interface:

[4K1-GS-9-03] Emotional Analysis of Persona-designated Character with LLM

〇Kazuma Murakami¹, Naoki Mori², Makoto Okada² (1. Osaka Prefecture University, 2. Osaka Metropolitan University)

Keywords:Human Agent Interaction, LLM, Kansei Engineering, Entertainment Applications

Research on Large Language Models (LLMs) like ChatGPT has gained momentum in recent years. These advanced LLMs can produce high-quality outputs, leading to significant achievements in more complex tasks. However, ChatGPT, which currently leads in performance, has yet to disclose internal specifications, making the construction of an LLM independently a costly endeavor. As a result, there is a growing trend in research focusing on the behavior of these models rather than improving the models themselves. This study forcuses on dialogues with characters whose interactions have been substantially enhanced by LLMs, aiming to achieve more relevant and interactive conversations with the real world. In this process, a character persona was assigned, and the decision whether to speak was based on assumed visual information and the character's internal state. Moreover, LLMs were utilized to numerically assess the character's emotions based on these contextual factors. Using the emotion vectors evaluated by the LLM and the author's assessments, the character's propensity to speak was framed as a binary classification problem, inputting the emotion vectors. Numerical results indicated that the emotional assessments successfully reflected the designated personas, and the determination to speak or not showed significant results compared to baseline models.