Presentation Information
[4K4-GS-6b-02]Persona Assignment to Large Language Models Using Structured Psychological Scales
〇Satoko Shiga1, Izumi Nitta1, Chika Yamanami1, Junichi Suga1, Shoko Miyagawa2, Sera Muto2, Hiroyuki Ito2, Mika Omori2 (1. Fujitsu Limited, 2. Ochanomizu University)
Keywords:
large language model,persona,psychological scale,personality
Assigning personas to Large Language Models (LLMs) poses challenges. While score-based instructions are intuitive, LLMs can interpret them ambiguously. Dialogue corpora provide concrete examples but lack structure, hindering control and evaluation. To overcome these issues, we propose using psychological scales, which measure personality traits and values, as “structured concrete examples.” We define personas by incorporating Q&A examples from these scales into few-shot prompts. This framework provides a unified structure for both persona definition and evaluation, thereby enabling objective quantitative assessment. We compare our method with a conventional score-based approach using personas defined by values related to employee promotion. Results show that our method yields more consistent persona-aligned responses for both seen and unseen items, enabling objective quantitative assessment and improved controllability.
