TREE-LIKE NEURAL CODES FOR SYNTAX IN THE HUMAN BRAIN

A theory of neural computations should provide a compact and interpretable description of the neural code and the operations underlying these computations. Sentence processing has consequently been theorized to involve the parsing of words into a tree-like structure: the syntactic tree. Yet, neural responses to linguistic inputs are currently best predicted by emergent black-box representations of artificial neural networks (ANN), leading to quantitative but little qualitative progress in our understanding of neural codes. Crucially, it remains unknown if these predictions can be explained by the superior ability of these models to instantiate theoretical linguistic structures compared to hand-made linguistic features. To elucidate what and why artificial neural networks predict from brain activity, we extended interpretability works to disentangle single-electrode encoding models of intracranial electroencephalographic recordings. Remarkably, the prediction of an artificial neural network was mostly explained by a small number of interpretable components, grouped into two subspaces. Early neural responses predictions (0 to 0.5 s after word offset) were explained by wordform and syntactic univariate features. Late neural responses predictions (0.8 to 1.2s) were explained by components instantiating a tree-like neural code of each sentence's syntax. These results provide partial evidence that the brain, like artificial neural networks, instantiates syntactic trees. We then applied our framework to two hypotheses of syntactic tree structure and showed that the effect was present for a dependency-tree model but not for a constituent-tree model. Together, these results pave the way to a theory of linguistic processes based on high-dimensional spaces, interpreted through simple laws.