With the assistance of an artificial language network, MIT neuroscientists have discovered what types of sentences serve to stimulate the brain’s primary language processing regions. In a study published in Nature Human Behavior, they revealed that these areas respond more robustly to sentences that display complexity, either due to unconventional grammar or unexpected meaning.
Evelina Fedorenko, Associate Professor of Neuroscience at MIT and a member of MIT’s McGovern Institute for Brain Research, said that straightforward or simplistic sentences prompt less of a response, while an unusual sentence structure or unfamiliar words trigger the brain to work harder.
The researchers centered their study on the language-processing regions located in the brain’s left hemisphere, such as Broca’s area and other areas of the left frontal and temporal lobes. To determine what type of sentence or linguistic input most stimulated these areas, they compiled 1,000 sentences from a variety of sources, including fiction, web text, spoken word transcriptions, and scientific articles.
Five human participants read through these considered sentences as their language network activity was tracked through functional magnetic resonance imaging (fMRI). Afterwards, the researchers input these 1,000 sentences into a large language model — similar to the structure of ChatGPT — and recorded the resulting activation patterns of the model in response.
From here, the team built an ‘encoding model’ to draw parallels between the activation patterns seen in the human brain and those observed in the artificial language model. This enabled them to project how the human language network is likely to respond to any subsequent sentences based on the artificial language network response to the initial 1,000 sentences.
The encoding model identified another 500 sentences projected to produce maximum brain activity (“drive” sentences) and those most likely to result in minimal activity (“suppress” sentences). A second group of human participants was then used to verify that these newly identified sentences did indeed either spark or suppress brain activity as anticipated.
The researchers found that certain linguistic properties would influence the intensity of the language network’s responses. For example, sentences with higher ‘surprisal’ (uncommonness compared to other sentences) tended to elicit stronger brain responses, and sentences that were either too simple or overly complex yielded the least activation. The largest responses arose from sentences that would require additional effort to comprehend due to slightly unusual, complex phrasing, or odd word selection.
The research team aims to evaluate whether these findings transfer to speakers of languages other than English. They also plan to investigate what types of stimuli might engage the language processing regions in the right hemisphere of the brain. The study was funded by Amazon Fellowship from the Science Hub, an International Doctoral Fellowship from the American Association of University Women, the MIT-IBM Watson AI Lab, and the National Institutes of Health, among others.