Search Abstracts | Symposia | Slide Sessions | Poster Sessions
Local Syntactic Structure Emerges in Predictive Neural Decoding
Poster Session D, Thursday, October 1, 4:30 - 6:30 pm, Wangari Maathai
Alessandro Tavano1,2, Cosimo Iaia1,2; 1Goethe University Frankfurt, Germany, 2Cooperative Brain Imaging Center, Frankfurt am Main, Germany
Human language models (LLMs) and the human brain both exploit context to support prediction, but whether they rely on comparable structural information remains unclear. We addressed this question by combining predictive neural decoding with theory-driven measures of syntactic structure. Building on expectation-based accounts of comprehension and dependency-locality theory, we asked whether prior knowledge modulates decoding as a function of local versus non-local syntactic context (Hale, 2001; Levy, 2008; Gibson, 2000). We analyzed data from 25 participants in an openly accessible MEG dataset (Gwilliams et al., 2025). First, we extracted seven syntactic measures commonly used in neuroscience and computational linguistics, and grouped them into memory-related features, such as tree depth and open nodes – which govern the moment-to-moment evolution of the underlying syntactic tree - , and integration-related features, such as closing nodes, which complete the syntactic tree or branch. Second, to capture the facilitatory effects of prediction in humans, we modeled prior knowledge using the Brown corpus and extracted marginal transition probabilities for memory and integration features at five context points. Third, we fit Ridge Regression decoding models to the MEG data and compared the decoding scores across features. Finally, we modelled decoding dynamics using Generalized Additive Mixed Models with an AR(1) autocorrelation structure. Prior knowledge selectively enhanced decoding for memory-related, but not integration-related, syntactic features. Crucially, these effects were not monotonic: decoding followed an inverted-U profile, with strongest enhancement at nearby contexts and weaker or absent effects at longer distances. This pattern is consistent with a local sharpening of syntactic expectations rather than a broad increase in predictive gain. The effect was robust across syntactic measures and was observed regardless of whether decoding was aligned to onset or offset, suggesting that it reflects a stable property of how syntactic context is represented over time. By contrast, integration measures were best decoded from brain activity using raw features, that is the current syntactic state. These findings provide neural evidence for local relashionship dynamically governing syntactic information during comprehension in humans. While both humans and LLMs may share a next-word prediction operational principle, they radically differ in what counts as context in syntactic predictions, local for humans and non-local for LLMs.
Topic Areas: Computational Approaches, Syntax and Combinatorial Semantics