Proceedings of the Annual Meeting of the Cognitive Science Society

A fundamental challenge in cognitive science is whether the human brain employs a structured, relational representational scheme for semantics that is shared across individuals. While existing electroencephalography (EEG) studies have primarily focused on decoding or correlational analyses, it remains unclear whether sentence-level EEG activity encodes semantic structure at the level of representational geometry, beyond stimulus-specific decoding. This study investigates whether EEG signals recorded during natural reading can be aligned with representations capturing shared semantic structure and systematic correspondence with text-based distributional semantic models. Using the ZuCo 2.0 dataset, we learn an alignment between sentence-level EEG activity and a shared semantic embedding space and evaluate cross-subject generalization under a strict leave-one-subject-out protocol. Representational alignment is quantified using cross-subject semantic retrieval and Representational Similarity Analysis (RSA), linking neural activity to abstract semantic relations. Our results suggest that human EEG activity during language processing is consistent with a stable, relational organization of semantic information, consistent with structural properties of semantic organization captured in computational models, without implying representational equivalence.