What is it about?

The ability to predict events has been proposed as a cornerstone of human cognition, with some calling brains “prediction machines”. Within the domain of language, the ability to predict upcoming words has been widely studied. One idea is that predicting upcoming words speeds our understanding of what’s being said. Given the variability within language use, however, people are unlikely to be able to always successfully predict. We therefore assessed the costs of failed predictions, during both language production and comprehension. Our findings suggest that accurate predictions speed the production and understanding of words, but inaccurate predictions slowed the production or understanding of words, even when the predictions were almost right. We interpret these results to suggest that predicting is not so much useful for speeding our production or understanding of language since we are often unable to generate entirely accurate predictions. So why might we predict? One reason might be that when we predict and are wrong, we learn from the experience and ultimately come to better represent the whole system we’re predicting. (In fact, this is basically how Large Language Models like ChatGPT work.)

Featured Image

Why is it important?

That we can produce and understand language so quickly and successfully is a biological miracle. But, the scientific community and the public at large don’t really know where our language abilities come from – at least not yet. (And before you think that because ChatGPT seems to understand and produce language, we must have it figured out – as reflected by the interpretability problem, we don’t know how ChatGPT works either!) Current evidence as well as common sense suggests that prediction is beneficial, and many prominent theories argue prediction is a fundamental component of language processing. However, exactly why is still debated. This research helps to put an important piece of the puzzle in place. Predicting is hard, especially about the future. There are many reasons we may do so anyway, for example, to test what we know of the language we speak, but with this research, we can rule out one reason: We don’t predict simply because it makes us speak or understand faster.

Read the Original

This page is a summary of: Is predicting during language processing worth it? Effects of cloze probability and semantic similarity on failed predictions., Journal of Experimental Psychology Learning Memory and Cognition, April 2024, American Psychological Association (APA),
DOI: 10.1037/xlm0001347.
You can read the full text:



The following have contributed to this page