Morphological complexity and unsupervised learning

This paper addresses the question of whether it is possible to use machine learning techniques on linguistic data to validate linguistic theory. We determine how readily inflectional classes recognized by linguists can be inferred by an unsupervised learning method when it is presented with the paradigms of a small number (80) of high frequency Russian noun lexemes. We interpret this as a measure of the validity of the linguistic theory. Inflectional classes are of particular interest, because they constitute a kind of autonomous morphological complexity that has no direct relationship to other levels of linguistic description, and hence there is no other objective way of assessing a theoretical characterization of them. Using the same method, we also examine the status of principal parts and defaults in inflectional classes, and the relationship between inflectional classes and stress in Russian nominal morphology. Our experiments suggest that this is an effective and interesting technique for shedding additional light on theoretical claims.