Decoding Neural Processing of Linguistic Features From Large-Scale Intracranial Recordings and Naturalistic Language Stimuli

Abstract

Previous research has discovered a set of areas in the brain that appears to represent information about linguistic meaning. However, no study has yet produced a comprehensive survey of how semantic information is represented in the brain, utilizing high-resolution neural recording. Using data from the Brain TreeBank, a large-scale multimodal dataset of recorded brain activity, we fit Generalized Linear Models (GLM) to map language and vision stimuli to induced brain activity. This framework allows us to localize processing areas in the brain per feature, as well as explore the temporal dynamics of this processing. Findings include maps relating neural activity across the brain throughout time and different linguistics, auditory and visual tasks, and indications of neural activity per word pre-onset for different regressors, such as part-of-speech, surprisal, and delta RMS.