Search Abstracts | Symposia | Slide Sessions | Poster Sessions
Probing the Syntax-Semantics Interface of the Language Network in Sentence Processing
Poster Session D, Thursday, October 1, 4:30 - 6:30 pm, Wangari Maathai
This poster is part of the Sandbox Series.
Peisong Yan1,2,3, Luyao Chen4, Junfeng Lu1,2,3; 1Department of Neurosurgery, Huashan Hospital, Shanghai Medical College, Fudan University, Shanghai, China, 2Shanghai Key Laboratory of Brain Function Restoration and Neural Regeneration, Shanghai, China, 3National Center for Neurological Disorders, Huashan Hospital, Shanghai Medical College, Fudan University, Shanghai, China, 4Max Partner Group at School of International Chinese Language Education, Beijing Normal University, Beijing, China
Introduction Sentence comprehension requires the brain to integrate syntactic structure with semantic content. Because syntactic and semantic information are normally tightly coupled in natural language, it remains challenging to determine whether neural responses reflect structure building, meaning integration, or their interaction. A useful approach is therefore to create sentences in which grammatical structure is preserved while rich real-world semantic content is reduced. Here, we used model-based fMRI analyses to investigate how the human language network represents syntactic and semantic dimensions during Chinese sentence comprehension. Objectives To localize brain regions involved in syntactic and semantic processing and to characterize their encoding mechanisms using a natural-unnatural comprehension paradigm and representation alignment with rule-based syntactic models and large language models. Methods Six participants completed a sentence-reading Go/No-go task during fMRI scanning across five sessions. The experiment included 720 valid Chinese sentences that varied along two dimensions: syntactic complexity and semantic naturalness. Natural sentences contained meaningful real-world content, whereas unnatural sentences used abstract shape terms to reduce rich lexical-semantic associations while preserving grammatical structure. General Linear Models were used to identify activation patterns related to syntactic complexity and semantic naturalness, and generalized Psychophysiological Interaction analyses examined task-dependent connectivity. To probe representational mechanisms, we constructed syntactic model features using HanLP-derived dependency distance and tree depth, and semantic model features using word2vec-based word-level representations and SentBERT-based sentence-level representations. Representation Similarity Analysis was then used to test whether multivoxel activation patterns matched these syntactic and semantic model spaces. Results Natural semantic processing preferentially engaged temporal and parietal components of the language network, including the anterior and posterior temporal lobes, angular gyrus, posterior superior temporal gyrus, and middle temporal gyrus. In contrast, unnatural sentences and increased syntactic complexity recruited broader frontal-temporal regions, including posterior inferior frontal gyrus and posterior middle frontal gyrus. RSA further revealed dissociable representational profiles for syntax and semantics. Syntactic model fits were distributed across the language network, but frontal regions showed a selective sensitivity to specific syntactic dimensions: BA44/45 patterns were more strongly associated with dependency distance than with tree depth, suggesting a preferential role in linear or non-adjacent dependency processing. Semantic RSA showed convergent effects for both word-level and sentence-level models in posterior temporal cortex and posterior middle frontal gyrus, indicating shared hubs for lexical and combinatorial semantic representations. Connectivity analyses further suggested that natural semantic processing involved temporal-parietal pathways, whereas unnatural sentence processing recruited stronger frontal interactions. Conclusions These findings suggest that sentence comprehension is supported by partially dissociable but interacting representational systems within the left frontal-temporal language network. By using grammatically structured but semantically impoverished unnatural sentences, this study provides a controlled way to examine how the brain maintains structure-building operations when rich semantic content is reduced. Our results indicate that distinct dimensions of syntactic structure and semantic content are encoded as separable multivoxel representational geometries. Keywords: Syntactic-semantic interface, Frontal-temporal network, fMRI, Language models.
Topic Areas: Syntax and Combinatorial Semantics,