Using Probabilistic Language Models for Tracking Modulations in MEG Spectral Power During Auditory Narrative Comprehension

Abstract

This study set to establish a direct link between formally-modelled predictive language comprehension processes and scalp-recorded electrophysiological signals. We recorded MEG while participants listened to 4–8 minutes long auditory stories (narratives) with no secondary linguistic task. Predictive language comprehension was modeled with probabilistic language models. On the basis of language-model output, two information-theoretic complexity metrics, word surprisal and word entropy, were computed word-by-word for all stories and correlated with modulations of MEG power envelopes in the theta (4-8 Hz) and beta (12-18 Hz) frequency bands. We used the framework of mutual information analysis to quantify the strength of statistical relationship between linguistic and MEG signals. In this preliminary analysis, we were not able to confirm any significant statistical relationship between either word entropy or word surprisal and theta- or beta-band MEG signals. To confirm that mutual information as implemented in our analysis could otherwise reveal meaningful statistical relationships in our signals, we show that there was stronger audio-MEG phase alignment in the theta than in the beta frequency band. We conclude by evaluating the current approach and outline possible avenues for follow-up research.