Decomposing Parts of Speech: Neural Encoding of Semantic and Grammatical Dimensions and Their Combinational Structure

Conventional part-of-speech (PoS) labels such as nouns, verbs and adjectives conflate lexical-semantic roots, grammatical functions, and their combinations, making it unclear whether neural sensitivity to PoS reflects either dimension or the way they are combined. We adopted Croft's typological framework to decompose major PoS classes into semantic root type and grammatical function and define prototypical nouns, verbs, and adjectives through the prototypical mappings between semantic root and grammatical function. We fitted time-resolved encoding models to magnetoencephalography (MEG) data from three participants who each listened to approximately ten hours of narrative speech. We examined the unique predictive contributions of semantic root, grammatical function, their interaction, and a binary prototypicality predictor, and whether the three prototypes exhibited distinct neural response dynamics. Semantic root and grammatical function contributed independently to neural responses during naturalistic language comprehension, with additional contributions from their interaction beyond the main effects. Root effects were predominantly left-lateralized, function effects were more bilateral, and interaction effects were concentrated in the left temporal and temporoparietal cortex, with overlap in left temporal regions. Further analyses revealed significant neural sensitivity to the distinction between prototypical and nonprototypical combinations in two of the three datasets. Prototypical nouns and verbs showed distinct coefficient time courses in all three participants, most consistently in left superior temporal cortex, whereas adjective-related contrasts were less consistent. These within-participant findings support a decompositional account of neural word-class sensitivity involving semantic, grammatical, and the mappings between them within distributed, partially overlapping cortical networks.