Proceedings of the Annual Meeting of the Cognitive Science Society

Recent advances in large language models (LLMs) have renewed interest in understanding the cognitive mechanisms underlying human language use. In this study, we focus on the Japanese word game ``Shiritori'' as a simple task related to language use, and aim to clarify the cognitive factors involved in its characteristics. To this aim, we model the execution of Shiritori based on the basic memory mechanisms of a cognitive architecture and investigate the characteristics that appear in Shiritori word sequences through simulations with different parameter settings. The results show that word sequences with different properties were obtained depending on the levels of lexical activation, inhibition, and semantic association. To complement the simulation findings, we also conducted a preliminary human evaluation in which participants rated the robot-generated word sequences. By constructing a model that can flexibly control Shiritori behavior, we explore its potential applications as a stimulus for research on human–robot interaction and language acquisition support.