Bilingual speakers seem to switch effortlessly between languages, even reviving a long-unused one with just a little exposure. Scientists have long puzzled over how the brain manages this, and most existing evidence comes from brain imaging studies that only offer a broad view of neural activity, missing the finer details of how individual languages are actually processed.
To dig deeper, researchers from Baylor College of Medicine recorded activity from single neurons in four bilingual (English-Spanish) volunteers who had epilepsy-related electrodes implanted in their hippocampus.
As the participants listened, read, and spoke in both languages, the team found that the brain seems to operate on two levels: individual neurons often fired for just one language, but networks of neurons formed a shared “concept map” that was largely language-independent, grouping related ideas, like “dog” and “wolf”, together regardless of which language was being used.
Remarkably, this conceptual map appeared to be universal across languages. Using data from English alone, researchers could accurately predict how related Spanish words would cluster together. This pattern echoes findings from other studies, including fMRI research showing that speakers of very different languages, and even fans of fictional languages like Klingon or Na’vi, seem to tap into a similar underlying neural scaffold for meaning.
To make sense of the neural data, researchers compared it to Google’s mBERT, a multilingual AI language model. The comparison showed a striking parallel: much like in the brain, mBERT organizes words based on meaning and context, creating a similar “semantic geometry” where related concepts sit close together. This suggests that both biological brains and AI language models may converge on similar strategies for representing meaning across languages, even without directly translating words one-to-one.
The study focused specifically on semantics, or meaning, leaving open questions about how the brain handles syntax and grammar, which may rely on separate neural networks entirely. Looking ahead, researchers hope to track people as they learn a brand-new language to see how these conceptual maps form in real time. Beyond deepening our understanding of language processing, the findings could also help inspire the next generation of more efficient and capable AI language models.