A Compositional Model of Semantic Fluency

The ability to recall semantically connected concepts---be it animals, summer fruits, or cities in Italy---is a remarkable capacity of the human mind. Such semantic fluency is thought to rely on traversing a mental space in which concepts are represented in terms of their meanings. However, the structure, properties, and navigability of this representational space remain enigmatic and highly debated. Existing approaches rely either on complex, uninterpretable distributional word-embeddings or on rigid, hand-crafted category norms. Here, we exploit the strengths of both, introducing Conceptome: a version of a compositional, interpretable, feature-based representation of semantic concepts, constructed by leveraging large language models. We use Conceptome to develop Conceptome-search, an auto-regressive model of how humans explore semantic spaces. We validate Conceptome and Conceptome-search using an animal fluency task, showing that they outperform state-of-the-art models in predicting human choices and capture key behavioral patterns such as interference. Our work, hence, offers new insights into the mechanisms underlying semantic fluency and memory retrieval.