Structural bias in phonology

Chomsky and Halle (1968) propose an Evaluation Procedure that prefers featurally simple rules and grammars, and Bach and Harms (1972) propose that this bias for structural simplicity can explain instances of historical change in terms of rule simplication in learning. In this talk, I discuss an alternative model of structural bias (Pater and Moreton 2012) cast in a Maximum Entropy framework (Goldwater and Johnson 2003, Hayes and Wilson 2008), which does not invoke an explicit simplicity metric. This MaxEnt model closely resembles the Configural Cue Model (Gluck and Bower 1988), a domain-general model of concept learning, which was abandoned twenty years ago because it does not match the classic results in visual concept learning (Shepard et al. 1961 et seq.). I present the results of an experiment that supports the predictions of the MaxEnt model against those of a model preferring featurally simple rules (Moreton et al. 2013). This phonotactic learning study differs from the classic visual concept learning studies in a number of ways. Based on a review of further studies (see esp. also Kurtz et al. 2013), it appears that the key difference is that the classic visual concept methodology encourages explicit hypothesis formation.