Skip to main content


A little knowledge can be a dangerous thing: Children's statistical learning leads to declines in learning performance


Recent research at the University of Maryland has found that infants use complex statistical knowledge when they are processing language. New findings from Aaron Steven White and colleagues suggest that children use knowledge about the sorts of sentence structures a word shows up in to guide their expectations about how sentences will unfold over time. These results support the view that children have abstract knowledge about sentence structure at least by 16 months of age. Not only do these infants have such knowledge, but they can deploy it quickly in learning new words.

White is a trainee in Maryland’s “Biological and Computational Foundations of Language Diversity” program, which is supported by NSF’s Integrative Graduate Education and Research Traineeship (IGERT) program. By integrating methodological and analytical techniques that span the fields of linguistics, psychology, and computer science, his research is some of the first to show that very young children associate statistical knowledge about sentence structure with specific words. Prepositions can clue listeners in to the role of different objects in an event. For example, in the sentence “She’s tapping with the blicket”, we know “the blicket” refers to an instrument of tapping. This contrasts with a sentence like “She’s tapping the blicket”, where we know “the blicket” refers to a thing being tapped. Using text analysis techniques, White and his colleagues found that the latter sentence-type is far more frequent than the former in speech to children. Using data from children’s eye-movements while they hear such sentences, White and his colleagues have shown that children understand the difference between the two sentences, and use this difference in learning novel words, by 16 months but fail to show it as 19 months. Interestingly, the 19-month-olds act as though they heard “She’s tapping the blicket” regardless of what sentence type they actually did hear. This suggests that, once children have gained more experience with processing language, they base their predictions on what is most likely to come up next in a sentence. By exploring the time course of children’s eye movements, they have demonstrated that providing 19-month-olds with novel verbs—-like “She’s remming with the blicket”—-allows them to show the same sensitivity found in the younger children.

This finding supports the presence of statistical knowledge about sentence structure linked to specific verbs. This work has several important implications. First, it demonstrates that children have abstract syntactic knowledge from a very young age. Second, it helps to establish that infants know quite a bit about the statistical properties of words and sentences in their language by the second year of life. Finally, because this work depends on understanding statistical properties of the input and online processing, it is possible to link it up with extensive literatures on corpus statistics and adult sentence processing. Ultimately, this work paves the way for understanding the developing interplay between language processing and the statistics of experience in first language acquisition.

Address Goals

White’s research addresses the Learning strategic goal through the training of young researchers in skills that combine linguistic analysis, experimentation with infants, and computational modeling. These skills are rarely found in the same individual. White’s research represents an important discovery in its own right, because it shows that children rely on different sources of linguistic information in their environment at different developmental stages, sometimes leading to an unexpected decline in learning performance.