Beyond Word Segmentation

Abstract

The term statistical learning was originally used to describe sensitivity to conditional relations between syllables in the context of word segmentation. Subsequent research has demonstrated that infants are sensitive to many other kinds of statistical information. The range of statistical learning phenomena presents a challenge to prior theories and models, which have primarily focused on a single aspect of learning. From our perspective, sensitivity to conditional information yields discrete representations (such as words). Integration across these representations yields sensitivity to distributional information. To achieve sensitivity to both kinds of statistical information, we propose a framework that combines processes that extract units from the input with processes that compare across these extracted items. We review the literature on statistical learning to show how these processes map onto prior research, and we discuss how the interaction between these processes gives rise to more complex patterns of learning than either process achieves in isolation.