Keynotes | K&Ts | GACs | Talks | Posters | Search

Poster C in Poster Session C: Wednesday, August 5, 9:30 – 11:15 am, Kimmel Center, Shorin & Rosenthal Rooms

Word Meaning, not Co-occurrence Statistics, is Essential for Predictive Speech Processing

Andrey Zyryanov¹, Victoria Pierz¹, Yulia Oganian¹; ¹Eberhard-Karls-Universität Tübingen

Presenter: Andrey Zyryanov

Speech processing is widely viewed as predictive: the brain computes a prediction error, or how word predictions differ from the actual perceived word, to infer sentence meaning. However, whether prediction error is computed from representations of word meaning or co-occurrence statistics is unknown. To address this, we examined how well surprisal from a large language model (LLM), a co-occurrence-based proxy of prediction error, accounts for how humans process words when their meaning is ambiguous. In one self-paced reading and one MEG experiment, ambiguity abolished the ability of LLM surprisal to predict word processing demands: slower reading and stronger neural responses. This suggests that the language network uses word meaning, rather than co-occurrence statistics alone, to compute prediction error. These findings highlight the limits of LLMs as models of the human language faculty.

Topic Area: Auditory, Speech & Language Processing