Learning the identity effect as an artificial language: Bias and generalisation

The results of two artificial grammar experiments show that individuals learn a distinction between identical and non-identical consonant pairs better than an arbitrary distinction, and that they generalise the distinction to novel segmental pairs. These results have implications for inductive models of learning, because they necessitate an explicit representation of identity. While identity has previously been represented as root-node sharing in autosegmental representations (Goldsmith 1976, McCarthy 1986), or implicitly assumed to be a property that constraints can reference (MacEachern 1999, Coetzee & Pater 2008), the model of inductive learning proposed by Hayes & Wilson (2008) assumes strictly feature-based representations, and is unable to reference identity directly. This paper explores the predictions of the Hayes & Wilson model and compares it to a modification of the model where identity is represented (Colavin et al. 2010). The results of both experiments support a model incorporating direct reference to identity.