BSIP/Universal Image Group via Getty Images
“Language is a vast field, and we are newbies in this field. We know a lot about how different areas of the brain are involved in linguistic tasks, but the details are unclear,” says Mohsen Jamali, a computational neuroscientist at Harvard Medical School who led a recent study on the mechanisms of human language understanding.
“What was unique about our study was that we looked at a single neuron. There are a lot of such studies in animals, electrophysiology studies, but very few in humans. We had a unique opportunity to access human neurons,” Jamali adds.
Examining the brain
In Jamali’s experiment, a series of recorded words was played to patients who, for clinical reasons, had implants that monitored the activity of neurons in the left prefrontal cortex, which is primarily responsible for language processing. “We had data from two types of electrodes: old-fashioned tungsten microarrays that can pick up the activity of a small number of neurons, and NeuroPixel probes, which are the latest development in electrophysiology,” Jamali said. NeuroPixels were inserted into a human patient for the first time in 2022 and were able to record the activity of more than 100 neurons.
“So we asked patients to participate in the operating room. We mixed sentences with words, including nonsense sounds that weren’t real words but sounded like words. We also had short stories about Elvis,” Jamali explains. He said the aim was to see if there was any structure to the neuronal responses to language. The nonsense words were used as a control to see if the neurons responded differently to them.
“The electrodes used in the study recorded voltages; this was a continuous signal with a sampling rate of 30 kHz. The key part was to distinguish how many neurons were in each recording channel. We used statistical analysis to isolate individual neurons in the signal,” says Jamali. His team then synchronized the neuronal activity signals to the recordings played back to the patient down to the millisecond, and began analysing the collected data.
Put away the words
“First, we converted the words in our word set into vectors,” Jamali says. Specifically, his team used Word2Vec, a technique used in computer science to find relationships between words in large amounts of text. What Word2Vec can do is tell us whether certain words have something in common, for example, if they’re synonyms. “Each word is represented by a vector in a 300-dimensional space. We then looked at the distance between those vectors and concluded that if the distance was close, the words belonged to the same category,” Jamali explains.
The team then used these vectors to identify words that clustered together, suggesting that they had something in common (which they later confirmed by looking at which words were clustered together). The team then determined whether certain neurons responded differently to different clusters of words, and they found that they did.
“We ended up with nine clusters, and we looked at which words were in those clusters and labeled them,” Jamali says. They found that each cluster corresponded to a distinct semantic domain: certain neurons responded to words related to animals, while other groups responded to words related to emotions, activities, names, weather, and so on. “The majority of the neurons we registered had one preferred domain, but some had multiple preferred domains, like two or three,” Jamali explains.
Understanding Mechanisms
The team also tested whether neurons were stimulated by the mere sound of a word, or by the meaning of the word. “Apart from meaningless words, another control we used in this study was homonyms,” Jamali says. The aim was to test whether neurons responded differently to the word “sun” and the word “son,” for example.
They found that the response depended on the context: when it was clear from the sentence that the word referred to a star, the sound stimulated neurons associated with meteorological phenomena. When it was clear that the same sound referred to a person, it stimulated neurons associated with relatives. “Presenting the same words randomly, without context, did not evoke as strong a response as when the context was present,” Jamali claims.
But language processing in our brains involves more than different semantic categories processed by different groups of neurons.
“There are a lot of unanswered questions in language processing. One of them is: to what extent does structure, i.e. syntax, matter? Is structure represented in a distributed network, or can we find a subset of neurons that encode structure but not meaning?” Jamali asked. Another thing his team wants to study is what the neural processing during speech is like, in addition to comprehension. “How are these two processes related in terms of brain regions and how information is processed?” Jamali adds.
Finally, and according to Jamali, the most challenging part will be to use the Neuropixels probe to see how information is processed in the different layers of the brain. “The Neuropixels probe will go deep into the cortex, and we can look at the neurons along the electrodes and determine: ‘information from this layer, which is responsible for meaning, goes to this layer, which is responsible for something else.’ We want to find out how much information is processed in each layer. This will be difficult, but it will be interesting to see how different areas of the brain are engaged simultaneously when language stimuli are presented,” concludes Jamali.
Nature, 2024. DOI: 10.1038/s41586-024-07643-2
