Two- and three-year-olds' linguistic generalizations are prudent adaptations to the language they hear

Studies of children's “statistical learning” mechanisms have established that even infants are very competent at extracting grammar-like structure from sequences of language-like sounds. We review some recent work exploring how these mechanisms might be used to extract functional grammatical knowledge from real speech. We use statistical analysis of large samples of transcribed child-directed speech to make predictions about the generalizations children will make, which we then test in the lab. We provide evidence that children's generalizations are input-driven: they are more likely to be made not only where the input gives supporting evidence, but also where the input gives no opportunity for concrete reuse and thus pushes the child to make an inductive inference. Keywords: Statistical learning; child-directed speech; rational models