[3Xin2] Poster session 1

[3Xin2-06] LLM Sentence Generation with Personality Cloning by Ten Utterance Example Prompting

〇Yuka Tsubota¹, Yoshinobu Kano¹ (1.Shizuoka University)

Keywords:generative AI, prompting, personality

For dialogue systems to engage in seamless and natural communication with humans, it is necessary for these systems to behave more human-like. It is believed that replicating the personalities of real individuals can be effective in achieving this. This study proposes a method for replicating personalities without training, based on large language models, using profile information and examples of utterances. Specifically, it examines and experiments with prompt techniques for replicating the personalities of specific individuals in generative AI, including ChatGPT. Collecting a large amount of dialogue data for a specific individual is not easy and requires a significant amount of time for learning. Therefore, this research aims to replicate the personalities of specific individuals without using dialogue data. The term "personality" in this paper refers to the character of a person that can be judged solely based on textual dialogue input and output. The results of the experiments revealed that by combining profile information with 10 examples of utterances into prompts, it is possible to improve the replicability of a person's personality without using actual utterances of the person in question, but rather a few generated examples by GPT-4.