MIT researchers have used an artificial language network to decipher which types of sentences activate the brain’s language processing centers most effectively. Their study reveals that sentences of higher complexity with unusual grammar or unexpected meanings engage these centers to a greater degree than straightforward sentences or nonsensical series of words. Their findings are based on examinations using functional magnetic resonance imaging (fMRI) of five human volunteers reading 1,000 sentences compiled from diverse sources.
Using a large language model similar to ChatGPT, the team noted the activation patterns of both the model and readers when presented with each sentence. Alongside this, the research team trained an encoding model that connected activation patterns shown in the human brain with those observed in the artificial language model, enabling the prediction of how the human language network will react to new sentences.
The study consequently used the encoding model to pinpoint 500 ‘drive sentences’, anticipated to create the most activity in the human brain, and ‘suppress sentences’, expected to provoke the least activity. In a separate group of three individuals, it was confirmed that these new sentences drove and suppressed brain activity as theorized.
Moreover, in terms of linguistic complexity, the brain’s language network was most activated by sentences of moderate complexity adhering to English grammar rules but challenging to comprehend, such as “Jiffy Lube of — of therapies, yes.” Either overly simplistic sentences or those too confusing to understand incited minimal activation. By gathering ratings from participants on crowd-sourcing platforms concerning the sentences’ grammaticality, plausibility, emotional valence, and ease of visualization, the researchers discovered that surprise and linguistic complexity most correlated with the response of the language network. Brain responses were most heightened by sentences of higher surprisal relative to its frequency in other sentences.
Going forward, the research team plans to expand their study to test speakers of other languages and explore stimuli types that might activate the language processing centers in the right hemisphere. The International Doctoral Fellowship from the American Association of University Women, an Amazon Fellowship from the Science Hub, the MIT-IBM Watson AI Lab, the National Institutes of Health, MIT’s Department of Brain and Cognitive Sciences, the McGovern Institute, and the Simons Center for the Social Brain funded the research.
