Semantics circle: LLMs and human linguistic cognition

Abstract

Some recent work in computational cognitive science suggests that LLMs such as GPT-3 are successful as models of human linguistic cognition (a separate matter from the success of these models as engineering tools). A particularly explicit statement of this claim was made by Steven Piantadosi, whose paper (entitled “Modern language models refute Chomsky’s approach to language”) argues that LLMs are significantly better linguistic theories than proposals emerging from within generative linguistics. Piantadosi takes this to amount to a refutation of the generative approach.

Since LLMs were designed to be useful engineering tools, discovering that they work in the same way that humans do would be startling, and this talk shows why it is also wrong. But the exercise is constructive: it sharpens the methodology of studying complex systems from the outside and serves as a reminder of some key (often quite old) discoveries about human linguistic cognition.