MIT neuroscientists have used an artificial language network to determine the type of sentences that stimulate the brain’s key language-processing centers most. The research revealed that complex sentences, whether due to unusual grammar or unexpected meaning, generated stronger responses in these centers. Straightforward sentences barely engaged these regions and nonsensical sequences of words yielded little effect.
For instance, the brain network was most active when reading unusual sentences like “Buy sell signals remains a particular,” from a publicly available dataset called C4. However, it went quiet when reading something straightforward, such as “We were sitting on the couch.” Primarily, the study targeted language-processing regions found in the left hemisphere of the brain, including Broca’s area, and other parts of the left frontal and temporal lobes of the brain.
The scientists began by compiling a set of 1,000 sentences from various sources — fiction, spoken word transcriptions, web text, scientific articles, etc. A group of five volunteers read these sentences as their brain activity was measured using functional magnetic resonance imaging (fMRI). The same sentences were then fed into a large language model, and its activation patterns were recorded.
An encoding model was then trained which associated the activation patterns seen in the human brain with those seen in the artificial language model. It was then able to predict how the human language network would respond to any new sentence based on how the artificial language network responded to the original sentences. Using this encoding model, 500 new sentences were identified that would generate maximum and minimum activity in the human brain network.
Three new participants were then tested with these sentences, it was found that these sentences did indeed stimulate and suppress brain activity as predicted. Further analysis of the sentences then enumerated some patterns, sentences with higher surprisal, meaning uncommon or unexpected, and those with higher linguistic complexity generate more responses in the brain.
The research has several implications, including the possibility of extending these findings to speakers of languages other than English, and exploring which stimuli may activate language processing regions in the brain’s right hemisphere. The study was financed by an Amazon Fellowship from 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.
