Human language comprehension is a complex process involving the integration of contextual information across words to build meaning. This process requires resolving local dependencies between words and integrating information from the broader narrative context. Traditionally, neuroimaging studies have used controlled experiments to isolate specific linguistic computations and map them onto brain activity. However, these approaches have limitations in generalizability to real-world language complexity. Recent advances in natural language processing (NLP) with transformer-based models, particularly the Transformer architecture, offer a new way to analyze language processing. Transformers use a layered "attention head" mechanism to process sequences of words, incorporating context-sensitive information through structured computations. Previous research has focused on the internal representations ("embeddings") generated by these models. This study shifts the focus to the circuit computations themselves – the "transformations" – which integrate contextual information. The researchers hypothesize that these functionally specialized transformations, particularly those implemented by individual attention heads, can provide a complementary window into linguistic processing in the brain. They aim to determine if these transformations, extracted from the BERT model, can effectively predict brain activity during natural language comprehension, and if the functional specialization observed in attention heads corresponds to functional specialization in the brain.

The paper reviews the existing literature on neuroimaging studies of language processing, highlighting the limitations of traditional controlled experiments in capturing the complexity of natural language comprehension. It discusses the advantages of using naturalistic comprehension paradigms. The authors mention previous work utilizing word embeddings to model brain activity but emphasize the novelty of focusing on the "transformations" within the Transformer architecture. Prior work on BERT and other Transformer models has demonstrated emergent functional specialization in attention heads, with specific heads showing selectivity for certain linguistic operations (e.g., resolving direct objects, tracking nominal modifiers). The review establishes the context for the current study, which seeks to bridge the gap between the insights from BERTology (the study of BERT's internal workings) and neurobiological understanding of language processing.

The researchers employed a model-based encoding framework to map Transformer features onto fMRI data acquired while participants listened to naturalistic spoken stories. The main features analyzed were transformations from the BERT-base model. These transformations capture the contextual information added to word embeddings at each layer. They also analyzed embeddings (contextualized semantic content) and transformation magnitudes (overall activity of attention heads). Other linguistic features, such as classical linguistic annotations (parts of speech, syntactic dependencies), and non-contextual word embeddings (GloVe) were included as comparisons. The brain data were spatially downsampled to a 1000-parcel cortical parcellation, grouped into ten regions of interest (ROIs) across the cortical language network. Encoding models were estimated using banded ridge regression with three-fold cross-validation for each subject and story. Confound variables (phonemes, phoneme rate, word rate, silence indicator) were included. Model performance was evaluated using the correlation between predicted and actual time series, normalized by a noise ceiling estimated using intersubject correlation (ISC). To further investigate functional specialization, they analyzed the individual contributions of each attention head within the transformations, examining their correspondence to syntactic dependencies and brain activity using PCA.

The study found that BERT transformations, considered in aggregate, effectively predict brain activity, performing comparably to embeddings and outperforming classical linguistic features and GloVe embeddings in most language ROIs. Transformations from earlier layers showed more unique variance in brain activity than embeddings. The layerwise analysis revealed that transformations exhibit more layer-specific performance fluctuations than embeddings, peaking at earlier layers. Analyzing individual attention heads revealed a structured mapping between their properties (layer, look-back distance) and brain activity in a low-dimensional cortical space. Heads with longer look-back distances were associated with prefrontal and anterior temporal areas. Finally, a strong correspondence was found between attention heads predicting specific syntactic dependencies and the brain regions associated with processing those dependencies. This correspondence is not trivial, requiring both the functional organization of heads and the model’s learning of linguistic structures.

The findings demonstrate that the transformations within the BERT model offer a robust basis for understanding human brain activity during language comprehension. The superior performance of transformations over other features highlights the importance of contextually rich information in neural processing. The layer-specific nature of transformations suggests a more granular mapping of computational steps onto the cortical hierarchy than embeddings alone. The structured relationship between attention heads and brain regions supports the hypothesis of shared functional specialization. The study's findings challenge the strict separation of syntax and semantics in neural language processing, suggesting that these aspects are intertwined in real-world comprehension. The results suggest that future language models with more biologically-inspired architecture could provide further insights into the functional organization of the brain during language processing.

This research provides strong evidence for shared computational principles between transformer-based language models and the human brain. The transformations within these models offer a powerful tool to understand the neural basis of language, surpassing traditional linguistic features in predicting brain activity. Future studies should explore modifications to Transformer architecture for enhanced model-brain correspondence, incorporate acoustic and prosodic features of speech, and investigate the role of long-range fiber tracts in brain language networks. These investigations will provide further understanding of the complex relationship between large language models and the human brain.

The study relies on pre-trained language models, potentially limiting the generalizability to other models or training procedures. The use of a fine-grained cortical parcellation and averaging across subjects could obscure some finer-grained aspects of individual differences in brain organization. The focus on naturalistic spoken stories limits the control available in experimental manipulations. The study did not directly incorporate acoustic and prosodic features into the model. Finally, despite the insights offered into the computational properties of the attention mechanisms, the authors do not suggest that the model provides a mechanistic model of the brain.