A Shared Componential Semantic Geometry Recovered from Brain Activity Transfers across Subjects and Languages

Is the representational geometry of word meaning a property of a brain or language, or a shared structure that surfaces differently across them? Using a contrastive multilayer perceptron, we learn a semantic embedding from Chinese word-level fMRI, supervised only by human componential semantic ratings. The learned geometry is strongly subject-invariant (leave-one-subject-out 2-vs-2 $\approx 0.97$; subject identity decodes near chance while semantic category decodes substantially above chance) and its dissimilarity structure transfers zero-shot across a language boundary: on the $42$-concept Chinese--English overlap it predicts the representational geometry of independent English concept fMRI (partial RSA $= 0.146$, $91\%$ of the lower human-rating noise ceiling, significant in five of five seeds), whereas a raw-brain control is near zero and a GloVe embedding estimated directly from English text reaches a comparable $0.177$. Human componential features thus index a semantic geometry that is invariant across individuals and a language boundary, recoverable from neural activity, and alignable across systems with a compact learned code.