Recent research published in journals cell This provides evidence that the human brain uses a shared organizational map, or geometry, to represent word meanings across different languages. By recording individual brain cells in bilingual individuals, scientists discovered that while each language relies on different patterns of cellular activity, the overall structural relationship between word meanings is consistent. This suggests that the brain maintains an internal model of universal meaning that is language independent.
Humans have a unique ability to understand and express the same ideas in multiple languages without confusion. Previous brain imaging studies have shown that bilingual speakers rely on overlapping brain regions when processing different languages. Regions traditionally associated with language, such as the inferior frontal gyrus and posterior temporal cortex, show similar activation patterns whether a person speaks English or Spanish.
However, these extensive brain scans do not capture how the brain matches equivalent concepts across languages, keeping them functionally separate. A joint research team from Baylor College of Medicine, Rice University, and Sungkyunkwan University sought to understand this phenomenon at the level of individual brain cells.
The research team proposed that bilingual brains may use a shared neurogeometry to organize meaning. In this context, neural geometry refers to the mathematical distances and relationships between words that are represented in high-dimensional space in the brain.
“Our findings suggest that the brain may store meaning in a language-independent manner,” said Sameer Sheth, Ph.D., professor of neurosurgery at Baylor College of Medicine, McNair Scholar, Cullen Foundation Endowed Chair, and co-senior author of the study. “Different languages seem to access a shared conceptual map rather than creating completely separate representations of the world.”
The research team focused specifically on the hippocampus, a brain region known to play a central role in linking memories and concepts. Because the hippocampus is embedded deep in the brain, it typically has difficulty learning while actively processing language.
To observe these deep brain structures, the scientists recruited four fully balanced bilingual patients who spoke both English and Spanish. These patients had already undergone surgery for treatment-resistant epilepsy. This medical situation provided a rare opportunity to implant high-density microelectrodes directly into the hippocampus. Three patients were given standard microelectrodes and one patient was given the advanced Neuropixels probe, which allowed the scientists to record the electrical spikes of individual neurons.
The scientists designed three different tasks for participants to perform. In the first task, all four patients spent approximately 120 minutes passively listening to corresponding stories and podcasts in both English and Spanish. The audio content included an educational science podcast by a creator named Kurzgesagt and an excerpt from the audiobook “Eat, Pray, Love.” This provided the researchers with thousands of spoken words that could be analyzed over multiple sessions.
In the second task, two of the patients read 99 matching short phrases on a computer screen and spoke them aloud. Finally, these same two patients engaged in a natural, unstructured conversation. They conversed with native speakers of each language for 32 to 99 minutes. The scientists then matched the spoken audio with recorded neural activity to track how the brain responded to specific words.
The scientists closely analyzed the firing rates of neurons recorded in response to equivalent words across languages. They first looked for interlingual neurons, individual brain cells that respond in the same way to translated pairs like “earth” and “tierra.” They identified a small subset of these neurons and provided evidence that a small number of isolated cells carry out direct translation.
But because these particular cells are rare, researchers could not fully explain how the brain processes two languages seamlessly. This finding suggests that translation is not primarily driven by specialized lexicon neurons, but rather arises from coordinated activity across large neural populations.
“Our results show that bilingual meaning is a novel property of neural populations,” said Xinyuan Yang, a postdoctoral fellow at Baylor College of Medicine and lead author of the study. “The brain does not seem to rely on one-to-one translation cells; instead, it stores patterns of relationships between concepts across languages.”
To understand the full picture, the authors compared the patients’ neural activity with an artificial intelligence tool called Multilingual BERT. BERT is a large-scale language model that learns cross-linguistic representations by processing large amounts of text. Scientists extracted contextual word embeddings from an artificial intelligence model.
Contextual word embeddings are mathematical representations of words that capture their meaning based on the surrounding text. The scientists then mapped these artificial representations alongside the actual firing rates of human hippocampal neurons. They found remarkable similarities between the geometry of semantic representations in the hippocampus and the internal organization of modern artificial intelligence systems trained in multiple languages.
“Large language models and the human brain may be converging on similar computational solutions for expressing meaning,” said Benjamin Hayden, adjunct professor of electrical and computer engineering and linguistics at Rice University, professor of neurosurgery and McNair Scholar at Baylor University, and co-senior author of the study. “That doesn’t mean AI will work exactly like the brain, but it does suggest that there may be universal principles for organizing knowledge.”
The results revealed that the specific semantic tuning curves of most individual neurons differed significantly between English and Spanish. A semantic tuning curve is essentially a profile of how a particular neuron responds to different word meanings across different topics. These profiles did not match, indicating that the brain uses language-specific recipes to process words. An individual neuron might fire strongly for the English word “dog” but remain completely silent for the Spanish word “perro.”
Despite this difference at the cellular level, the overall population of neurons maintained a conserved geometric organization across both languages. The researchers measured the mathematical distance between neural responses to different words. They found that comprehensive maps of English word meanings tended to perfectly mirror Spanish maps.
The brain accomplishes this by using the same population of neurons but reading neuron activity from different angles or axes depending on the language being spoken. This phenomenon is similar to viewing a three-dimensional object from two different perspectives. The shape of the object remains the same, but the visible profile changes depending on the viewpoint.
“This helps explain how bilingual people can switch between languages so smoothly,” Hayden says. “The brain seems to maintain a common internal structure for meaning while at the same time allowing languages to be differentiated enough to avoid interference.”
The shared semantic structure is so robust that researchers could use it to predict neural responses. By looking at clusters of related words in English, the scientists were able to mathematically rotate the data to accurately predict how the brain would respond to the pushed out Spanish words. This form of zero-shot learning shows that the overall geometry provides sufficient information for translation, even without a direct word-to-word mapping at the individual neuron level.
This discovery could be of wide interest in the humanities and social sciences, beyond the field of science. The concept of stable, shared geometric neural maps provides evidence for a structuralist view of language, an intellectual current often traced back to the Swiss linguist Ferdinand de Saussure. Structuralism holds that meaning transcends individual cultural expressions and instead depends on underlying universal structures and systems.
The authors note that in addition to advances in basic neuroscience, these findings could have implications for the development of brain-computer interfaces. It could also inform speech rehabilitation therapy and future artificial intelligence systems designed to communicate more naturally with humans.
Although these findings provide deep insight into bilingualism, readers should be aware of several limitations. One potential misconception is that these results apply to all language learners. All patients in this study were highly skilled early-acquiring bilinguals. This means that he learned both languages around the age of 4 or 5. It remains unclear whether people who learn a second language later in life share this same neurogeometry.
Furthermore, the researchers cautioned that the study featured a very small sample size of only four participants. This is because it is extremely rare to find balanced bilingual patients who require deep brain electrode implants for medical reasons. The authors note that this study only examined English and Spanish. These two languages share many linguistic roots and structural similarities.
Another limitation involves the health status of the participants. One of the patients was recorded under general anesthesia after resection of part of the temporal lobe. Although this patient’s data were broadly consistent with those of other patients, anesthesia and surgery may have altered normal brain activity patterns.
Future studies could extend these findings by testing unrelated language pairs, such as English and Mandarin, or by observing larger populations. Scientists also want to study how participants actively learn a new language. Tracking the brain during the learning process could reveal exactly how this shared semantic geometry is formed and adjusted over time.
The study, “Shared neural geometry for bilingual semantic representation in human hippocampal neurons,” was authored by Xinyuan Yan, Ana G. Chavez, Melissa Franch, Kalman A. Katlowitz, Ivy Gautam, Brian Kim, Aaditya Krishna, Aadit Shrivastava, Katie Van Arsdel, James Belanger, Assia Chericoni, Taha Ismail, and Elizabeth. A. Mickiewicz, Danica Paulo, Hanlin Zhu, Arika M. Goldman, Vaishnav Krishnan, Atul Maheshwari, Eleonora Bartoli, Nicole R. Provenza, Sen Bom Michael Yu, Benjamin Y. Hayden, Sameer A. Sheth.
