|Title:||Restrictiveness and Phonological Grammar and Lexicon Learning|
|Comment:||in Proceedings of CLS 43|
|Abstract:||A major challenge for models of phonological learning is the learning of restrictive grammars that prohibit unobserved phonological patterns. Within Optimality Theory, a well-known solution to this learning problem is a ranking bias. Researchers have proposed a number of such biases, including Markedness >> Faithfulness, Specific Faithfulness >> General Faithfulness, and General Markedness >> Specific Markedness (Smolensky 1996; Prince and Tesar 2004; Hayes 2004, Tessier 2006). As Alderete and Tesar (2002) point out, however, ranking biases do not provide a general solution. The inadequacy of ranking biases is particularly apparent when grammar learning is viewed in the context of the larger phonological learning problem, the problem of learning both grammars and lexicons of underlying forms.
This paper demonstrates via simulation that the general solution to the restrictiveness problem developed in Jarosz (2006) subsumes the effects of the three ranking biases above. The same solution also naturally extends to the full phonological learning problem, identifying restrictive grammar and lexicon combinations in a case where ranking biases are insufficient. The model formalizes the learning of a probabilistic OT grammar and lexicon as the gradual optimization of an objective function, likelihood (the probability of the observed data under a hypothesized grammar and lexicon combination). The solution to the restrictiveness problem exploits a fundamental property of this well-known optimization-based approach to learning: the grammar-lexicon combinations favored by likelihood are those that best fit the observed data distribution, generating only observed forms and as few others as possible. Reliance on this basic property of likelihood maximization to favor restrictive grammars in the domain of phonological learning depends crucially on the way in which the generative model of the grammar and lexicon is formalized. The paper discusses the properties of the model that make this solution possible and presents simulations illustrating the model's capacity to identify restrictive grammars.