[1Win4] Poster session 1

[1Win4-36] Experience-Based Character-Like Utterance Generation and Its Experimental Evaluation for Real-World Events

〇Seiya Takahashi¹, Shinji Iwata², Koya Ihara^1,2, Shohei Kato¹ (1.Nagoya Institute of Technology, 2.CyberAgent)

Keywords:character, dialogue system, LLM, experience

Generating character-like utterances based on experiences (experience utterance) is important for developing character-like dialogue systems that share real-world experiences with users. To generate character-like experience utterance, we structured real-world events using a 5W1H+impression frame in a large language model. However, experience utterances generated with an impersonal impression lacked character-like qualities. In this study, we introduced a 5W1H+character-like impression frame and conducted a user experiment to evaluate its effectiveness. In the experiment, users assessed character-likeness and experience consistency using a 7-point Likert scale from experience utterance generated under three conditions: no impression, impersonal impression, and character-like impression. The results demonstrated the effectiveness of character-like impression, which incorporate personality into impressions. However, character-like impression did not show significant improvement compared to the no impression condition. User feedback and additional analysis suggest that these results may be influenced by utterance style and experimental methodology.