Predicting to learn? The role of prediction (error) in children’s language processing and acquisition

Chiara Gambi

Cardiff University, School of Psychology

Starting from around their second birthday, children can quickly recognize the words they hear; what’s more, they can even predict upcoming words ahead of hearing them. Recent years have seen a surge in interest in the hypothesis that children’s ability to predict upcoming language may be key to the process of language acquisition, a hypothesis that is incorporated in influential computation models of language learning. Central to these models, in particular, is the idea that predictions drive learning when they are incorrect, because incorrect predictions, when compared to the input, generate informative learning signals (prediction error). But what is the evidence that children predict to learn? In this talk, I will present data from two large scale studies (N = 215 and N = 160, respectively) that appear to cast doubts on the role that prediction plays in language learning.
In the first study (2-to-5 year olds), we modelled individual variation in fine-grained eye-tracking measures of recognition and prediction skill using mixed effects models, and related these processing measures to individual variation in vocabulary size. Importantly, we also followed up a subset of the children (N = 55) to measure increases in vocabulary size over time. In younger children, larger increases in vocabulary size over time were positively associated with both prediction and recognition skills, but recognition skills fully mediated the relationship between prediction skills and vocabulary growth; no associations were found in older children. These findings suggest that the ability to predict upcoming words benefits word learning only indirectly, because more predictable words are recognized more quickly, thus freeing up resources for processing yet-to-be-learned words.
In the second study (2-to-4 year olds), we devised a novel tablet-based task to investigate whether children are more likely to remember novel words that violate stronger expectations based on their pre-existing linguistic knowledge. While adults formed stronger memory traces for novel words that violated a stronger prior expectation, children’s memory was unaffected. These findings suggest that prediction error may drive memory formation in adults but not in pre-schoolers. I will discuss these data in the context of the broader literature on the role of prediction (error) in language learning.