The Dual Nature of Syntactic Node Count: Facilitating and Inhibiting Sentence Comprehension

Neurolinguistic research on syntactic processing has utilized Node Count, the number of structure-building steps executed at each word, to investigate how the effort invested in structure building affects brain activity. However, several recent studies cast doubt on the presumed link between Node Count and processing effort; they observe negative correlations between Node Count and reading times. Addressing this disagreement, we analyze information-theoretic measures and reading times using three large-scale naturalistic datasets in English across different grammar formalisms: phrase-structure grammar, dependency grammar, and combinatory categorial grammar. This study establishes a dual nature of Node Count: we conclude that it indexes two opposing forces—processing facilitation due to predictability in immediate regions and processing inhibition due to contextual semantic integration in spillover regions. We also demonstrate that Node Count has a unique psycholinguistic value as an index of human prediction, since its facilitatory effect explains reading time variance beyond that captured by a Transformer-based language model. These results highlight the role of structural representations in human language processing and underscore the need for a careful interpretation of the observed correlations between Node Count and brain activity in neurolinguistic studies.