Rich lexicons and restrictive grammars: Maximum likelihood learning in optimality theory

This dissertation develops a theory of phonological learning within a proposed probabilistic formulation of Optimality Theory (Prince and Smolensky 1993/2004), Maximum Likelihood Learning of Lexicons and Grammars (MLG). MLG undertakes the problem of concurrently learning grammars and lexicons, the two main subproblems of phonological learning. A major challenge is the learning of grammars that are simultaneously restrictive and have generalizing capacity, two contradictory requirements. The proposed solution combines the principle of richness of the base with maximum likelihood learning of Optimality Theoretic grammars. The generalizing capacity of grammars is attributed to formal linguistic theory, due to implicational markedness universals. The identification of restrictive grammars is the consequence of maximum likelihood learning in conjunction with explicit reliance on richness of the base. A possible learning algorithm EMGL, a two-stage Expectation Maximization procedure, is proposed. The algorithm is shown to successfully learn restrictive grammars and lexicons that together comprise explanatorily adequate language models with generalizing capacity. The success of the teaming theory is dependent on the success of the formal linguistic system it incorporates. Since the ultimate goal of MLG is to build a foundation for modeling of child phonological acquisition, the formal linguistic system must be capable of capturing the implicational markedness universals observed not only in adult languages but also during acquisition. A portion of the thesis is devoted to extending the formal linguistic system to account for apparent inconsistencies of markedness universals in the area of syllable structure. The proposed learning theory builds a foundation for the computational modeling of child phonological acquisition, accounting for the end states of two stages of acquisition: phonotactic learning and morphophonemic learning. Additionally, predictions are made concerning the roles of frequency in learning. The algorithm is applied to the proposed OT theory of syllable structure to initiate modeling of the acquisition of syllable structure in Polish.