THREE LANGUAGES, ONE MAP OF MEANING IN THE BRAIN

A Chinese reader, an English listener and a French one can follow the same story, even though their languages share almost no words and arrange sentences differently. Somewhere, the brain must keep what the story means. Yi Guo, Shuqiang Wang and colleagues, at the Shenzhen Institutes of Advanced Technology of the Chinese Academy of Sciences and two universities in Hong Kong, looked for that common ground in brain scans.

Not the same activity, the same distances

Their key idea is to compare geometry rather than activity. Two people, or two languages, may produce different patterns of brain activity for the same sentence. But one can still ask: which passages does the brain treat as close to each other, and which as far apart? If those relations hold across languages, the brain shares a common map of meaning even when the details differ.

227 brains, three languages

The team reanalysed four public sets of functional MRI recordings, 227 participants in all:

• adults listening to The Little Prince in English (49), Chinese (35) or French (28) — three separate groups of people;
• six Chinese–English bilinguals reading stories in both languages;
• twelve people listening to news in Mandarin, used to pre-train the system;
• 97 native French speakers listening to passages of Alice’s Adventures in Wonderland, as an independent test; 79 of them also heard the same passages in English.

They built SemDecoder, a brain decoder. A multilingual language model proposes possible next words; the brain signals decide which ones to keep. It recovered meaningful content in all three languages and beat an earlier reference decoder on every measure and dataset. One example from the paper: the line “For me, it was like a spring of fresh water in the desert” came back from the brain as “Anymore, it was just like a rain of pure water in the desert.” In French, a sentence about a picture in “a book about the virgin forest” came back as “a book about the black forest”. The brain regions that mattered most were in the temporal lobes on both sides, in all three languages.

The shared map

Then the authors compared the geometry. Comparing two brains first means putting them in a common frame: that is alignment, a computed rotation that looks for the closest correspondence between the two groups’ activity patterns, without assuming the same brain locations light up. Once aligned, the arrangement of story events largely coincided across Chinese, English and French. The authors then scored every passage on three properties of meaning — how concrete rather than abstract it is, how much it is about people and relationships, and how emotionally arousing it is — and looked for the direction on the brain map along which each score rises: those three directions pointed the same way in all three languages. Passages far apart in one language tended to be far apart in another. The match held for individual listeners and throughout the story.

It was not uniform across the brain. The cross-language match was weakest in areas close to the senses and movement, and strongest in high-level association regions — the ones that do not receive signals straight from the ears or eyes but combine them. The bottom of that gradient is no surprise: a Chinese sentence and a French one share none of their sounds, and the auditory areas hear that difference before any meaning, so agreement can only appear further along. This gradient closely followed the brain’s main known axis of organisation, which runs from sensory and motor areas to the most integrative regions (correlation of 0.80 across 400 brain regions). The authors suggest that these high-level regions hold relations between meanings that depend less on the form of any one language.

Reading a language it never learned

The practical test was generalisation. A decoder trained on one language worked above chance on another it had never seen, in all six directions between the three languages, and also in bilinguals. A decoder trained on Chinese listeners picked out, better than an untrained network, which of twelve passages French speakers were hearing — in French or in English. The more it relied on the high-level regions, the better it generalised.

Finally, when only a small amount of brain data was available in the target language, pre-training on the other languages improved decoding at every data size tested. The authors point to a question of fairness: brain recordings are not equally available for every language, and borrowing from well-studied languages could help the others — though they did not test that directly. Their study covers three languages; whether the shared map extends to the full diversity of human languages, and how it changes with a person’s own experience of several languages, remains open.