Cultural Evolution of Topic Preference of Chatting Agents Based on Generative Models

Abstract

This study proposes an agent-based model for the cultural evolution of chatting agents based on a generative model. Agents are in a two-dimensional space reflecting their social proximity and possess invariant genetic traits as well as multiple cultural traits acquired from others. Each individual utters a sentence generated by GPT-2 using these traits as its topic preference. It moves closer or further away from others when the commonality of words in the utterance with them is greater or less than a threshold. Each agent also receives topic words as its cultural traits from closely neighboring individuals, corresponding to cultural evolution. Experiments without cultural evolution, based on the agents having either positive or negative words as their genetic trait, showed that such polarity of topic preference affected their basic social behavior, and positive agents tended to form clusters, while negative agents tended to wander between clusters. Cultural evolution weakened this tendency, leading to the emergence of major and novel cultural traits shared by many individuals regardless of their polarity. These findings may reflect group dynamics and cultural evolution through conversational interactions inface-to-face situations or SNS.