Information-Restricted Neural Language Models Reveal Different Brain Regions' Sensitivity to Semantics, Syntax, and...

Area of Science:

• Neuroscience
• Psycholinguistics
• Computational Linguistics

Background:

• Understanding the neural basis of language comprehension, specifically the separation versus integration of syntactic and semantic processing, is a key neurolinguistic question.
• Previous research has explored brain regions involved in lexical and supra-lexical language processing, but the precise interplay between syntax and semantics remains debated.

Purpose of the Study:

• To investigate the extent to which syntactic and semantic information are processed in distinct or overlapping brain regions during speech comprehension.
• To develop and apply a novel approach using information-restricted neural language models to probe the neural encoding of linguistic features.

Main Methods:

• Trained lexical (GloVe) and supra-lexical (GPT-2) language models on text corpora with selectively removed syntactic or semantic information.
• Used features from these information-restricted models to predict functional magnetic resonance imaging (fMRI) time courses of participants listening to naturalistic text.
• Manipulated contextual information size for GPT-2 to analyze supra-lexical processing integration windows.

Main Results:

• Most language comprehension brain regions responded to both syntactic and semantic features, but with varying relative importance.
• Semantic and syntactic features showed more spatial dissociation in the left hemisphere compared to the right.
• The right hemisphere demonstrated sensitivity to longer contextual information than the left hemisphere.

Conclusions:

• The brain employs a complex organization for syntactic and semantic processing, with both shared and distinct neural substrates.
• Information-restricted language models offer a powerful tool to dissect the neural representation of linguistic information.
• Findings contribute to understanding the spatial organization of language processing in the human brain, particularly hemispheric differences in context integration.