Statistical natural language processing can be difficult for morphologically rich languages. The observed vocabularies of such languages are very large, since each word may have been inflected for morphological properties like person, number, gender, tense, or others. This unfortunately masks important generalizations, leads to problems with data sparseness and makes it hard to generate correctly inflected text.
This thesis tackles the problem of inflectional morphology with a novel, unified statistical approach. We present a generative probability model that can be used to learn from plain text how the words of a language are inflected, given some minimal supervision. In other words, we discover the inflectional paradigms that are implicit, or hidden, in a large unannotated text corpus.
This model consists of several components: a hierarchical Dirichlet process clusters word tokens of the corpus into lexemes and their inflections, and graphical models over strings – a novel graphical-model variant – model the interactions of multiple morphologically related type spellings, using weighted finite-state transducers as potential functions.
We present the components of this model, from weighted finite-state transducers parameterized as log-linear models, to graphical models over multiple strings, to the final non-parametric model over a corpus, its lexemes, inflections, and paradigms. We show experimental results for several tasks along the way, including a lemmatization task in multiple languages and, to demonstrate that parts of our model are applicable outside of morphology as well, a transliteration task. Finally, we show that learning from large unannotated text corpora under our non-parametric model significantly improves the quality of predicted word inflections.