Search Abstracts | Symposia | Slide Sessions | Poster Sessions
Brains and language models converge on a shared conceptual space across different languages
Poster Session C, Saturday, September 13, 11:00 am - 12:30 pm, Field House
Zaid Zada1, Samuel Nastase1, Uri Hasson1; 1Princeton University
Human languages are vastly diverse, including different sounds, alphabets, syntax, and more. Despite these substantial differences in linguistic form, we can translate between languages and retain the intended meaning. While the forms of language are different, the meaning and function of these forms must be partly shared across speakers of different languages. How do different languages converge on a shared neural substrate for conceptual meaning? What is shared and what is language-specific? Honey and colleagues (2012) measured neural activity in anglophones and russophones when they listened to their respective languages, and found that neural activity was shared across languages in high-level language and default-mode areas. Recent work on interpreting multilingual large language models (LLMs) has found multilingual conceptual features at intermediate layers of the neural network (Lindsey et al., 2025). In the current study, we used LLMs to identify neural representations of the shared conceptual meaning of the same story as heard by speakers of three different languages. We used an open fMRI dataset where three groups of participants listened to an audiobook of The Little Prince in their respective native languages: English (n = 49), Mandarin Chinese (n = 35), and French (n = 29). The story was ~100 minutes long and was divided into 9 scanning sessions. We aligned the audio and text across the three languages for each sentence, resulting in a total of 1650 sentences. Using the transcripts, we extracted contextual word embeddings from three unilingual BERT models: one trained solely on English text, one trained solely on Chinese text, and one trained solely on French text. Despite the different models and languages, we found notable similarities between each pair of language embeddings after a simple rotation of the embedding features, especially in the middle layers. Next, we trained voxelwise encoding models to predict each subject’s BOLD responses from the word embeddings for their own native language. First, we found that BERT embeddings predicted activity throughout the language network for each language. Critically, we then evaluated how well the encoding model predictions generalize across languages. For example, we evaluated predictions from the English model trained on the brain activity of an English subject against the brain activity of French and Chinese subjects. We found that models trained to predict neural activity for one language in speakers of that language generalize to different subjects listening to the same content in a different language throughout high-level language and default-mode regions. Our results so far suggest that the neural representations of meaning behind different languages are shared across speakers of different languages and that language models trained on different languages converge on this shared meaning. These findings suggest that shared meaning arises from our shared environment and how we interact with it (and each other) in similar ways. 1. Honey et al. https://doi.org/10.1523/JNEUROSCI.1800-12.2012 2. Lindsey et al. https://transformer-circuits.pub/2025/attribution-graphs/biology.html
Topic Areas: Meaning: Lexical Semantics, Computational Approaches