Our brain is not a “next-word-prediction” machine, but rather makes structure-based predictions about how a sequence of words will continue. The syntax of what is being said actively limits when and to what extent the prediction is made.
How well individual words are processed in language comprehension depends heavily on how efficiently the brain can anticipate them based on the context up to that point. A common hypothesis is that the brain actively predicts upcoming words. However, since human working memory naturally has only limited processing capacity, researchers from the ESI, New York University, and Zhejiang University in Hangzhou hypothesized that the brain uses a more efficient mechanism to organize information.
One possible mechanism, therefore, could be that the brain does not process language word by word, but rather groups it into more complex, hierarchically organized units known as constituents. From this, the scientists derived another central hypothesis: namely, the assumption that predictions are more accurate within such constituents than across their boundaries. They were able to demonstrate this through experiments on brain activity and behavior.
According to their findings, the brain strikes a compromise by not optimizing language comprehension solely for prediction accuracy, but rather by balancing the “effort” required for efficient prediction quality with the available cognitive resources.
Click here for the paper
