A Computational Model of General Rule Learning with Unnatural Classes

This paper presents the results of a computational model of generalized phonological rule learning (Calamaro and Jarosz, 2012), which is used to model experimental studies on the learning of phonotactic patterns governed by natural and unnatural classes. I focus on two papers with conflicting results on the learnability of natural and unnatural rules. Saffran and Thiessen (2003) find that a phonotactic pattern of positional voicing restrictions governed by a natural class of segments is learned by infants, but a similar pattern governed by an unnatural class is not learned. In contrast, Chambers, Onishi, and Fisher (2003) find that infants can learn a phonotactic pattern governed by an unnatural class of segments. The computational model presented in this paper is able to account for these seemingly conflicting results, explaining both the learnability and unlearnability of rules governed by unnatural classes.