ARENA Deliverable [P5_WP2_02] Report 2: estimate shared and unique info between LM and brain w.r.t. linguistic property
This work investigates which linguistic properties contribute to the alignment between language model representations and human brain activity during language comprehension. Using fMRI recordings and representations from pretrained language models, the framework selectively removes information related to specific linguistic properties (surface, syntactic, and semantic) from model representations and measures the resulting change in brain predictivity. The analyses quantify the shared contribution of different linguistic properties to both systems and identify which properties are most responsible for the observed brain–language model alignment, highlighting a particularly strong role for syntactic information
Publication Date: 2026-06-18