​The Neural Basis of Hierarchical Structure Building during Phrase Processing

Abstract

Human language comprehension requires transforming linear word sequences into hierarchical structure. Using Mandarin as a case study, we investigated how the brain builds phrase-level structure from two sequentially presented words. Participants processed the same phrases in two tasks emphasizing either the semantic meaning of the phrase or the composition relation between the two words that form the phrase. We focused on three components of phrase building–the syntactic category of individual words, the syntactic category of the resulting constituent, and the composition relation between the two words–and examined them using representational similarity analysis, Bayesian hierarchical modeling, and variance decomposition with semantic controls. Word-category representations emerged during second-word processing and were observed mainly in left middle temporal and inferior frontal regions. Phrasal constituent category representations were found in left posterior temporal, left supramarginal, and right inferior frontal regions. Composition-relation representations were detected in left angular and left frontal regions and remained after controlling for constituent category, indicating that word-to-word relations and resulting constituent categories capture distinct aspects of composition. Variance decomposition showed that structural effects in several regions could not be fully reduced to embedding-derived semantic similarity. Task demands further modulated these representations differently: word category effects were largely task-independent, whereas constituent category effects showed frontal modulation and composition relation effects were strongest when the task explicitly required relation judgments. Together, these findings show that different aspects of phrase structure are represented in different brain regions and help explain how the brain combines words into phrases and builds more complex linguistic structures.