Abstract
Identifying the neural correlates of specific syntactic features is challenging due to the inherent entanglement of syntax and semantics in natural language. This paper presents a novel framework using Mean Vector Null space Projection (MVNP) to surgically remove syntactic information from word embeddings. By analyzing the resulting decrease in the ability of these embeddings to predict fMRI brain activity, the study maps the distributed and hierarchical representation of syntax across the human cortex. The findings highlight a significant overlap between syntactic and semantic processing networks and identify previously overlooked regions involved in grammar processing.
Introduction
While neuroscientists have identified general brain regions involved in language processing, pinpointing the neural correlates of fine-grained syntactic features—such as Part-of-Speech (POS) or Dependency (DEP) relations—remains difficult. The primary obstacle is "entanglement": in natural speech, syntax and semantics are deeply intertwined. Traditional fMRI studies often use artificial, manually designed stimuli to isolate these variables, which may not reflect the complexity of natural language processing. This study proposes a data-driven approach using Large Language Models (LLMs) to "probe" the brain by isolating and removing specific syntactic threads from word representations.
Methods
The researchers developed a three-stage framework to identify brain regions sensitive to specific syntactic features:
- Feature Elimination (MVNP): The core of the methodology is Mean Vector Null space Projection (MVNP). This technique calculates the mean vector for each category within a feature (e.g., the average vector for all nouns). It then projects the word embeddings into a "null space" where these mean vectors are neutralized, effectively blinding the embedding to that specific feature while preserving other information.
- Brain Encoding: Original and "impoverished" embeddings (from BERT or ELMo) are used to predict fMRI signals from 19 subjects listening to naturalistic stories using Ridge Regression.
- Significance Testing: A block-bootstrap test identifies "feature-correlated" voxels—areas where prediction accuracy significantly drops when a specific feature is removed.
Datasets and Models
Syntactic labels were derived from Ontonotes 5.0 and the English Web Treebank. Brain imaging data consisted of fMRI recordings from subjects listening to The Moth Radio Hour. Pre-trained BERT-base and ELMo models provided the initial word embeddings for the study.
Results
The application of MVNP successfully isolated syntactic features, significantly reducing classification accuracy for target features while maintaining semantic integrity. The subsequent brain encoding analysis yielded several key insights:
- Distributed Representation: Syntactic features are not localized to a single "grammar center" but are distributed across the temporal, frontal, and parietal lobes.
- Syntax-Semantic Overlap: There is a massive overlap between syntactic and semantic brain networks, suggesting the language network processes these elements holistically.
- Hierarchical Organization: Certain regions, such as the Superior Temporal Gyrus (STG) and Angular Gyrus (AG), act as hubs for multiple features, while others, like the precuneus, show more specific sensitivity.
- Novel Regions: The study identified the precuneus and cingulate gyrus as being sensitive to multiple syntactic features, areas often overlooked in classical models.
Discussion and Conclusion
The findings confirm that syntactic features are encoded across a broad cortical network. However, the study notes limitations, including the current method's restriction to categorical features and the assumption of a linear relationship between embeddings and neural signals. Future research could explore non-linear transforms and multi-voxel pattern analysis (MVPA) to further decode the brain's complex linguistic representations.
Summary updated