How the brain builds sentences, neuron by neuron

In the fraction of a second before a person speaks, their brain weaves together complex grammar, precise vocabulary and the underlying meaning of the language. Now, researchers have tracked the electrical crackle of individual brain cells in real time during unscripted conversations, capturing how sentences are built before a single word is spoken.

By observing these neurons in a region of the human brain called the frontotemporal cortex, scientists have discovered that individual brain cells act as specialized linguistic building blocks. “We used to think language was this diffuse, whole-network phenomenon,” says Ziv Williams, a neurosurgeon at Massachusetts General Hospital (MGH) in Boston and co-author of the study. “But it turns out you have specific neurons that only care if a word is a noun, or only care if a phrase is ending.”

The work was published today in Nature1.

On the map

To capture this activity, Williams and his colleagues used electrodes that were temporarily implanted in people with epilepsy to monitor their seizures. Because these participants were awake and speaking freely, the team could observe how the brain operated as they spoke. Neuroscientist Jing Cai, also at MGH, says that this set-up provided a rare opportunity to eavesdrop on the cellular processes that underlie speech, capturing details that standard brain-imaging devices cannot obtain.

Brain implant translates thoughts to speech in an instant

Access to such data provides a “rare” glimpse into the biological machinery that governs speech, says Angela Friederici, a neuropsychologist at the Max Planck Institute for Human Cognitive and Brain Sciences in Leipzig, Germany.

The researchers found that specific neurons increased their electrical firing rate just before certain components of speech, such as a noun, were spoken. This signalled that the brain constructs sentences by firing particular neurons that can handle distinct properties of language, such as parts of speech, the researchers noticed.

To make sense of this electrical symphony, the team turned to large-language models, the same type of artificial intelligence that powers chatbots. By comparing the models’ text-processing patterns to the participants’ brain activity, the researchers discovered that both the neurons and the models tracked the sentence’s broader context, maintaining a mental record of up to five preceding words to shape the meaning of the next one.

They also found a division of labour between two core concepts: semantics, or the meaning of the words, and syntax, which dictates grammar and structure. Most cells preferentially encode either semantic or syntactic information alone, they found. Although these language-specialized neurons are scattered across the frontotemporal cortex, the researchers discovered that the left hemisphere of the brain was more active in encoding this linguistic information than was the right.