Semantic Cognition for and from Language Models

Abstract

The success of neural network language models (LMs) at showing semantically sophisticated behavior raises questions about the extent to which they are compatible with human conceptual knowledge, and if they can perhaps also advance our understanding of semantic cognition. This brief opinion explores this potentially bidirectional connection between LMs and semantic cognition. Specifically, I review a range of methodology used to establish the structure and the function of conceptual knowledge in LMs, primarily inspired by the semantic cognition literature. I then summarize three possible ways that future work could take to make contributions in the opposite direction, by using insights from language models to make conclusions about the language-based origin and mechanisms of human semantic cognition.