[4Xin2] Poster session 2

[4Xin2-105] Intention Inference and Persona Matching Method based on Psychological Models with LLM

〇Yudai Yamazaki¹, Daisuke Kamisaka¹, Kurokawa Mori¹, Asakawa Yoshinori², Yamaguchi Motomu ² (1.KDDI Research, Inc., 2.KDDI CORPORATION)

Keywords:AI, LLM

The development of Language Models (LLM), exemplified by ChatGPT, has been remarkable. The Chain of Thought (CoT) technique, introduced to enhance inference accuracy, incorporates examples or instructions in the input prompt for LLM. Challenges persist in accurately inferring user intentions, especially in complex tasks like discerning intentions from user-specific data. This arises from nuanced differences in the inference process between LLM and humans. Our study aims to address this by proposing CoT, integrating a psychological model. This technique gauges intention degrees and factors from persona data using the Likert scale, encompassing attributes, personality, and behavioral information. Evaluating the method's accuracy involved inferring volunteer participation intent from 273 participants' data, resulting in a 5.5% improvement in Mean Absolute Error compared to traditional zero-shot CoT. Additionally, we leverage this intention inference to propose a framework matching personas with similar intentions from diverse persona data elements. By infusing insights from various psychological models into LLM's inference, our method enhances intention inference accuracy and facilitates data augmentation through intention inference.