Contributed Talk Sessions | Poster Sessions | All Posters | Search Papers

Poster Session A: Tuesday, August 12, 1:30 – 4:30 pm, de Brug & E‑Hall

Can LLMs inform us about predictive processing during natural listening in MEG?

Sahel Azizpour¹, Britta U. Westner, Jakub Szewczyk, Umut Güçlü², Linda Geerligs¹; ¹Donders Institute for Brain, Cognition and Behaviour, ²Radboud University Nijmegen

Presenter: Sahel Azizpour

The brain uses contextual information and prior knowledge to anticipate upcoming content during language comprehension. Recent research has shown that predictive signals can be revealed in pre-onset electrocorticography (ECoG) activity during naturalistic narrative listening, by building encoding models based on word embeddings from large language models (LLMs). Similarly, evidence for long-range predictive encoding has been observed in functional magnetic resonance imaging (fMRI) data, where incorporating embeddings for multiple upcoming words in a narrative improves alignment with brain activity. This study examines whether similar predictive information can be detected in MEG, a technique with higher temporal resolution than fMRI but a lower signal-to-noise ratio than ECoG. Our findings indicate that MEG captures pre-onset representations up to 1 second before word onset, consistent with ECoG results. However, unlike fMRI findings, incorporating future word embeddings did not enhance encoding in MEG, not even for one word into the future, which suggests that the pre-onset encoding may not reflect predictive processing. This work demonstrates that MEG combined with LLMs is a valuable approach for studying language processing in naturalistic narratives and highlights the need to study further what constitutes evidence for prediction during natural listening.

Topic Area: Language & Communication

Extended Abstract: Full Text PDF