Phonological generalization from distributional evidence

We propose a model of L2 phonological learning in which the acquisition of novel phonological category inventories proceeds not by mapping L2 inputs onto existing category inventories available in L1 and other already known languages, but rather through general categorization processes in which L1 and other language knowledge serves as an inductive bias. This approach views linguistic knowledge as hierarchically organized such that the outcome of acquisition of a language—L1 or otherwise—includes not only knowledge of the specific language in question, but also beliefs about how any language is likely to be structured. In this paper we test a set of predictions regarding how two key types of information can come together to drive L2 learning: distributional information within a single phonetic dimension and generalization bias derived from existing knowledge of language. We tested these predictions by training adult monolingual English speakers in a distributional learning paradigm (Maye & Gerken, 2000; Maye, Werker, & Gerken, 2002) on a novel contrast, segmental length, and testing them on categorization of short and long segments for both trained and untrained items. Results show both learning and generalization from one class of segments (sonorants) to another class (obstruents), broadening the empirical range of phonetic contrasts for which distributional learning has been shown to be effective and providing evidence for our approach to L2 learning as one of inductive inference and generalization rather than of mapping.