PhD candidate at Donders Centre for Cognition in Nijmegen, Netherlands, Sahel Azizpourlindy combines MEG and large language models to study the neural indices of predictive processing.

The brain uses contextual information and prior knowledge to predict future content during language comprehension. Previously, it has been demonstrated that contextual word embeddings, derived from Large Language Models, can be linearly mapped to brain data. Recently this method has been used to study neural signatures of predictive processing. One study found that in a naturalistic listening setting, predictive signatures of an upcoming word can be observed in its pre-onset signal, measured with ECoG. In the fMRI domain, another study has shown that including embeddings of multiple upcoming words improves the model’s fit to brain data. This has been interpreted as an indication that the brain encodes long-range predictions. In this study we examine whether the same predictive information can be found in MEG data, a signal with lower signal-to-noise ratio than EcOG and higher temporal resolution than fMRI. We show that: 1) The signatures of pre-onset predictions are also detectable in MEG data similarly to ECoG and 2) Contrary to what has been observed in the fMRI data, including future embeddings does not improve brain mapping in MEG signals. These findings provide a novel avenue for studying predictive processing during language comprehension with naturalistic stimuli.