Proceedings of the Annual Meeting of the Cognitive Science Society

We implemented two instance-based learners, the K-Nearest Neighbors model and the Generalized Context Model, and a rule-based learner, the Minimal Generalization Learner, adapted for linguistic data. We fit these on three distinct, variable patterns of word variation in Hungarian: paradig- matic leveling and vowel deletion in verbs and vowel har- mony in nouns. We tested their predictions using a Wug task. The best learners were combined into an ensemble model for each pattern. All three learners explain variation in the test data. The best ensemble models of inflectional variation in the data combine instance-based and rule-based learners. This result suggests that the best psychologically plausible learn- ing model of morphological variation combines instance-based and rule-based approaches and might vary from case to case.