In this thesis, I address the issue of learning form-meaning correspondences of inflectional affixes in the presence of homonymy. Homonymy is ubiquitous in all languages despite the fact that it presents a notorious problem for learning and processing. It is a common assumption that patterns of homonymy are restricted in some way and that these restrictions reflect something about the way people learn languages. In this work, I attempt to flesh out this intuition using tools from formal learning modeling. I show some quantitative evidence that inflectional paradigms have statistical preferences for certain types of non-arbitrary mappings between form and meaning. Namely, one-to-one and elsewhere mappings that can be described with defaults are preferred while all other mappings are avoided. Interestingly, the preferred types of mappings also have a nice learning property: more specifically, there are simple generalization methods that can be used for learning them. The learning model I propose takes advantage of this fact, although it is still capable of learning 'arbitrary' form-meaning mapping which are empirically attested. Overall, my learner provides a strong bias (rather than a categorical restriction) on the types of patterns it can learn; a bias motivated by the empirical data mentioned above. The model of learning I propose also predicts intermediate overgeneralization errors and subsequent corrections in the process of language acquisition. It is unique in that, unlike most formal learning models, it relies on a non-monotonic generalization strategy inspired by the blocking proposals in the realm of generative morphological theories.

back