[Paper Review] Disentangling Syntax and Semantics in the Brain with Deep Networks

Charlotte Caucheteux, Alexandre Gramfort+1 more|arXiv (Cornell University)|Mar 2, 2021

Neurobiology of Language and Bilingualism53 references33 citations

TL;DR

The paper proposes a taxonomy to factorize GPT-2 activations into lexical, compositional, syntactic, and semantic representations, then maps these factorized activations to fMRI brain data from 345 subjects listening to narratives, revealing distributed, non-modular syntactic and semantic substrates.

ABSTRACT

The activations of language transformers like GPT-2 have been shown to linearly map onto brain activity during speech comprehension. However, the nature of these activations remains largely unknown and presumably conflate distinct linguistic classes. Here, we propose a taxonomy to factorize the high-dimensional activations of language models into four combinatorial classes: lexical, compositional, syntactic, and semantic representations. We then introduce a statistical method to decompose, through the lens of GPT-2's activations, the brain activity of 345 subjects recorded with functional magnetic resonance imaging (fMRI) during the listening of ~4.6 hours of narrated text. The results highlight two findings. First, compositional representations recruit a more widespread cortical network than lexical ones, and encompass the bilateral temporal, parietal and prefrontal cortices. Second, contrary to previous claims, syntax and semantics are not associated with separated modules, but, instead, appear to share a common and distributed neural substrate. Overall, this study introduces a versatile framework to isolate, in the brain activity, the distributed representations of linguistic constructs.

Motivation & Objective

• Motivate a clear, operable taxonomy to separate lexical, compositional, syntactic, and semantic representations in both artificial and biological neural networks.
• Develop a method to extract syntactic representations from a deep language model and decompose brain representations accordingly.
• Assess whether shared brain-model representations are localized or distributed across syntax and semantics.
• Evaluate the mapping between GPT-2 activations (factorized by linguistic class) and fMRI signals from a large cohort during naturalistic listening.

Proposed method

• Define a five-point taxonomy distinguishing lexical, compositional, syntactic, and semantic representations in distributed activations.
• Isolate syntactic representations by synthesizing sentences with the same syntax and averaging GPT-2 activations (overline{X}).
• Map model activations X (and syntactic extractions overline{X}) to brain signals Y using a spatio-temporal encoding model with ridge regression and FIR delays.
• Decompose brain scores into lexical, compositional, syntactic, and semantic components across GPT-2 layers and compare against syntactic-synthesized baselines.
• Apply the method to the Narratives fMRI dataset (345 subjects, ~4 hours of stories) and test across layers and architectures for generalization.

Experimental results

Research questions

• RQ1 Can a robust taxonomy disentangle lexical, compositional, syntactic, and semantic representations in both GPT-2 and brain activity?
• RQ2 Do syntactic and semantic representations map to distinct, localized brain modules or to distributed substrates?
• RQ3 How do compositional representations compare to lexical representations in their brain mapping, and where are they distributed?
• RQ4 Is the brain-model mapping stable across layers and across different transformer architectures?

Key findings

• Compositional representations recruit a more widespread cortical network than lexical ones, involving bilateral temporal, parietal, and prefrontal cortices.
• Syntax and semantics are not confined to separated modules; they share a common and distributed neural substrate.
• Contextual (deep) layers (e.g., GPT-2 layer 9) yield higher brain scores than lexical embeddings, reflecting stronger encoding of linguistic structure.
• Syntactic representations can be extracted from GPT-2 by averaging activations over syntactically matched synthesized sentences, and remain informative for brain mapping.
• Compositional syntax and compositional semantics show distributed brain involvement, peaking in regions including temporal and prefrontal cortices, cingulate, supramarginal, and middle-frontal areas.
• Findings generalize across layers and architectures, with middle GPT-2 layers often yielding strongest brain scores and similar trends observed in other transformer models.