Inflectional defaults and principal parts: an empirical investigation

We describe an empirical method to explore and contrast the roles of default and principal part information in the differentiation of inflectional classes. We use an unsupervised machine learning method to classify Russian nouns into inflectional classes, first with full paradigm information, and then with particular types of information removed. When we remove default information, shared across classes, we expect there to be little effect on the classification. In contrast when we remove principal part information we expect there to be a more detrimental effect on classification performance. Our data set consists of paradigm listings of the 80 most frequent Russian nouns, generated from a formal theory which allows us to distinguish default and principal part information. Our results show that removal of forms classified as principal parts has a more detrimental effect on the classification than removal of default information. However, we also find that there are differences within the defaults and principal parts, and we suggest that these may in part be attributable to stress patterns.