Wonnacott, E. 1 , Brown, H. 2 & Nation, K. 3
1 University College London
2 University of Warwick
3 University of Oxford
Language acquisition requires generalization, but learners must also acquire lexically based restrictions. A growing body of research suggests this is achieved, at least in part, by tracking distributional statistics at and above the level of particular lexical items. We used artificial language learning to investigate sensitivity to distributional statistics in 6-year-olds and adults. The language comprised nouns that were followed by one of two meaningless novel particles; noun-particle relationships were manipulated across input sets. At test, learners produced their own noun-particle combinations. Lexically based learning was shown when a learner restricted usage of a noun to an attested particle, generalization was shown when they produced unattested combinations. We examined the effect of two variables: overall lexicality of a language (proportion of words in the input restricted to occur with a single particle) and input skew (whether a majority particle occurred with most nouns in the language). Lexicality aided learning for adults but not children; both groups were aided by input skew. These results demonstrate that learning is affected by distributional statistics above the level of words or bigrams. Findings are discussed within the framework offered by models that capture generalization as rational inference such as hierarchical Bayesian and simplicity-based models.