Analyzing Finnish Inflectional Classes through Discriminative Lexicon Models

University of Eastern Finland DRDHum 2024 Conference Submission58 Authors

Keywords: Discriminative Lexicon Model, Finnish, FastText, Word embeddings, Inflectional morphology

Abstract:

Descriptions of complex nominal systems make use of inflectional classes. Inflectional classes bring together nouns which have similar stem changes and use similar exponents in their paradigms. Establishing what the inflectional classes of a language are is far from trivial, and across grammatical descriptions, the number of classes distinguished can vary considerably. Although inflectional classes can be very useful for language teaching as well as for setting up finite state morphological systems, it is unclear whether inflectional classes are cognitively real, in the sense that native speakers would need to discover these classes in order to learn how to properly inflect the nouns of their language.

This study investigates whether the Discriminative Lexicon Model can understand Finnish inflected words without setting up inflectional classes, using a dataset with 55271 inflected nouns of 2000 high-frequency Finnish nouns of 49 inflectional classes. Two DLM models were constructed, one using Endstate of Learning (EOL), and the other using Frequency-Informed Learning (FIL). Both models were given the task of predicting Finnish FastText embeddings from words' forms, represented by trigram-based vectors. The models were trained on
40694 word tokens, and evaluated on 14577 held-out low-frequency tokens. Overall accuracy on test data was higher for EOL (78%) than for FIL (50%), which is to be expected since EOL represents optimal learning with infinite exposure. Importantly, for both models, accuracies increased for inflectional classes with more types, more lower-frequency words, and more hapax legomena. Importantly, the accuracy of the DLM models mirrors the productivity of the inflectional classes. The model struggles more with novel forms of unproductive and less productive classes, and performs far better for unseen forms belonging to productive classes.

These results demonstrate that the inflectional system of Finnish nouns can be learned without hand-crafting of inflectional classes. Crucially, the extent to which generalization is possible matches the productivity of the inflectional classes distinguished by linguistic analysis. We are currently investigating whether deep discriminative learning can provide even more accurate mappings, but we anticipate that with these even more powerful mappings, models will outperform what can be expected for human native speakers.

Submission Number: 58