FBS Colloquia No.430Perceptual and Cognitive Neuroscience Laboratory

Linguistics in the Era of LLMs: Language Models as Linking Hypotheses between Linguistic Theory and the Brain

Large language models (LLMs) produce strikingly human-like language, and recent work has shown that their internal representations predict human brain activity with surprising accuracy. Does this mean that LLMs process language the way humans do? This talk addresses this question and introduces computational psycholinguistics and neurolinguistics, a field that uses language models to study how humans process language and how the brain supports it. The approach compares model predictions with reading times and brain activity recorded while people read or listen to naturalistic stories. As a case study, we present evidence that a model sensitive to the hierarchical structure of sentences better accounts for brain activity during Japanese reading. We conclude by discussing open questions, including what the predictive success of LLMs actually tells us about human language.