MIT neuroscientists, aided by an artificial language network, have published a study revealing that complex sentences, both in terms of unusual grammar and unexpected meaning, produce stronger responses in the brain’s primary language processing areas. The centers have a less marked response to straightforward and nonsensical sentences.
For instance, when reading unusual sentences like “Buy sell signals remains a particular,” chosen from a public language dataset, this brain network was most active. Nonetheless, it showed less activity when reading straightforward sentences such as “We were sitting on the couch.”
Researchers, including Evelina Fedorenko, Associate Professor of Neuroscience at MIT and member of MIT’s McGovern Institute for Brain Research, focused on language-processing regions located in the brain’s left hemisphere. This region includes Broca’s area and other parts of the left frontal and temporal lobes of the brain.
To understand what kind of linguistic input drives the left hemisphere language network, the researchers compiled a set of 1,000 sentences from varying sources such as fiction, spoken word transcriptions, web text, scientific articles, and more. They measured the language network activity of five human participants as they read each sentence, and then fed these same sentences into a large language model like ChatGPT. An “encoding model” was applied to compare the activation patterns seen in the human brain and those observed in the artificial language model, allowing the researchers to predict how the human language network would respond to any new sentence based on the artificial language model’s responses.
The model was used to identify 500 new sentences that were either “drive” sentences (which would generate maximal brain activity) or “suppress” sentences (which would lead to minimal brain activity). Experiments with three new human participants confirmed that these selected sentences did indeed drive or suppress brain activity as predicted.
To understand why certain sentences drove activity more than others, the researchers evaluated sentences based on 11 different linguistic properties. The study revealed that sentences with higher surprisal and linguistic complexity generated higher responses in the brain.
The researchers found that sentences that require work to understand, but still make some sense, generated the largest responses in the language network. They now plan to see if these findings can be extended in speakers of other languages.
The research was funded by several prestigious organizations, including the Science Hub, the American Association of University Women, the MIT-IBM Watson AI Lab, the National Institutes of Health, the McGovern Institute, the Simons Center for the Social Brain, and MIT’s Department of Brain and Cognitive Sciences.
