Search Abstracts | Symposia | Slide Sessions | Poster Sessions
Decomposing the neural correlates of language: lexical, syntactic, and semantic organisation across the cortex
Poster Session B, Wednesday, September 30, 4:30 - 6:30 pm, Wangari Maathai
James Fodor1, Leila Wehbe1; 1Carnegie Mellon University
Numerous studies have established that vector embeddings extracted from large language models (LLMs) are predictive of brain activity during language reading and listening tasks. However, it remains unclear which linguistic features drive these correlations, and to what extent language processing in LLMs mirrors that in the brain. In this study, we seek to shed light on these questions by analysing an existing fMRI dataset consisting of three participants each listening to 80 autobiographical narratives, totalling about twenty hours of audio. We conduct several novel analyses. First, we decompose the explained variance of brain activity based on distinct categories of linguistic features, finding that about 60% of variance is explained by lexical and word-sense features, 30% by syntactic and semantic role features, and the remaining 10% by broader story context. Second, we analyse variation of semantic features across all 80 stories, showing that representations of semantic features such as actions, body parts, social interaction, and numerical quantity are spatially segregated across the cortex. We interpret these findings in the light of ongoing disputes about the function of core language regions in the temporal and inferior frontal regions, specifically how this function relates to semantic processing in a broader range of frontal, parietal, and cingulate cortical regions. On the basis of our findings, we argue that these regions play a supportive role in representing grounded contextual semantic information, after the relevant semantic features have been identified and extracted by linguistic operations performed by the core language areas.
Topic Areas: Computational Approaches, Meaning: Lexical Semantics