Lexical-Semantic Content, Not Syntactic Structure, Is the Main Contributor to ANN-Brain Similarity of fMRI Responses in the Language Network

Abstract

Representations from artificial neural network (ANN) language models have been shown to predict human brain activity in the language network. To understand what aspects of linguistic stimuli contribute to ANN-to-brain similarity, we used an fMRI dataset of responses to n = 627 naturalistic English sentences (Pereira et al., 2018) and systematically manipulated the stimuli for which ANN representations were extracted. In particular, we i) perturbed sentences’ word order, ii) removed different subsets of words, or iii) replaced sentences with other sentences of varying semantic similarity. We found that the lexical semantic content of the sentence (largely carried by content words) rather than the sentence’s syntactic form (conveyed via word order or function words) is primarily responsible for the ANN-to-brain similarity. In follow-up analyses, we found that perturbation manipulations that adversely affect brain predictivity also lead to more divergent representations in the ANN’s embedding space and decrease the ANN’s ability to predict upcoming tokens in those stimuli. Further, results are robust to whether the mapping model is trained on intact or perturbed stimuli, and whether the ANN sentence representations are conditioned on the same linguistic context that humans saw. The critical result—that lexical-semantic content is the main contributor to the similarity between ANN representations and neural ones—aligns with the idea that the goal of the human language system is to extract meaning from linguistic strings. Finally, this work highlights the strength of systematic experimental manipulations for evaluating how close we are to accurate and generalizable models of the human language network.

Artificial Neural Network Language Models Predict Human Brain Responses to Language Even After a Developmentally Realistic Amount of Training

“…Some recent work has begun to attempt isolating the aspects of model representations that affect model-to-brain alignment. For example, Kauf et al (2024) performed a series of experiments where model representations were obtained for different perturbations of a linguistic stimulus (e.g., scrambling the word order or dropping/replacing some of the words) and then related to neural representations of an intact stimulus in order to see which perturbations affect model representations negatively. They found that word-level and compositional semantic information appears to be more important than information related to the syntactic structure in the model-to-brain alignment.…”

Modeling dynamic social vision highlights gaps between deep learning and humans

“…To gain insight into what aspect of the captions produced high alignment between the language models and the brain/behavior, we performed selective perturbations to the captions as in [27,52], but instead of deletion, we performed masking to keep the overall syntactic structure of the sentence intact. We used the spaCy [53] package and the en_core_web_sm model in particular to identify the parts of speech for masking.…”