Semantic predictability selectively modulates delta-band syntax tracking: evidence for good-enough parsing

Abstract

Neural populations exhibit low-frequency phase synchrony to inferred phrase and constituent boundaries during speech comprehension, widely interpreted as reflecting the endogenous construction of syntactic structure. Good-enough processing (GEP) accounts suggest this structure-building is not obligatory for every utterance but graded: listeners may scale back syntactic computation when meaning can be recovered from local lexical-semantic context. Whether this graded reliance on syntax is reflected uniformly in delta-band neural activity remains unresolved, partly because syntactic processing comprises several dissociable demands that prior work has not modeled concurrently. We reanalyzed an open-source EEG dataset recorded during audiobook listening and used delta-band temporal response functions to model four syntactic constructs concurrently: syntactic surprisal, closing-node count, and integration and storage cost from Dependency Locality Theory. For each construct, we tested whether its contribution to the delta-band response was modulated by local semantic dissimilarity, computed from contextualized word embeddings. Semantic dissimilarity significantly modulated the response to syntactic surprisal and closing-node count, showed a smaller but reliable modulation of storage cost, and showed no modulation of integration cost; a repeated-measures comparison confirmed this graded hierarchy differed reliably across constructs. These results indicate that semantic predictability does not discount syntactic processing across the board but selectively affects the anticipation and construction of structure and its ongoing maintenance, while leaving retrieval of an already-resolved dependency largely unaffected. We interpret this pattern as sharpening, rather than merely reiterating, the claim that listeners prioritize efficiency during real-time comprehension.