The Structural and Statistical Basis of Morphological Generalization in Arabic

This thesis examines learnability and generalization of the morphology of Modern Standard Arabic, focusing on two subsystems: the noun plural and the masdar of form I verbs. Using psycholinguistic experiments and computational analyses, I assess two aspects of generalization. First, I address the learnability of a morphological system based on the predictability of the morphological variant of an unseen form based on analogy to existing forms. Second, I assess how speakers generalize existing morphological patterns to previously unseen forms using nonce-form tasks.;More generally, this thesis investigates how speakers learn and generalize morphological patterns in systems with two characteristics: coarse-grained representations, as both systems contain non-concatenative patterns that require a high level of abstraction to represent; and high uncertainty, in that there are 30+ patterns for both systems under investigation, and the pattern that an existing words takes is somewhat unpredictable. By studying systems with these characteristics, I examine the key questions of: 1) what is the basis of analogy in morphological generalization in Arabic?; and 2) how do speakers decide among the possible outcomes when there are a large number of possibilities with varying likelihoods?;First, I demonstrate that speakers generalize existing noun plurals primarily on the basis of the coarse-grained CV template, and select among the possible morphological variants in a probabilistic manner, indicating that they track lexical statistics on this coarse-grained level. Second, I demonstrate that the masdar is quite predictable on the basis of type statistics on the coarse-grained representation of the verb pattern, which disproves previous claims that the masdar of form I verbs is unpredictable. Finally, in a nonce-form task, I show that speakers also generalize existing masdar patterns in a probabilistic manner, but do so not on type statistics on the verb pattern, but potentially on the CV template. For both of these systems, speakers utilize the full range of possibilities in generalization, indicating that they generalize low-probability patterns even when there are 30+ possibilities. The implications of these findings for theories of Arabic morphology as well as theories of learnability and generalization of complex morphological systems are discussed.