Context-dependent type-level models for unsupervised morpho-syntactic induction

Abstract

This thesis improves unsupervised methods for part-of-speech (POS) induction and morphological word segmentation by modeling linguistic phenomena previously not used. For both tasks, we realize these linguistic intuitions with Bayesian generative models that first create a latent lexicon before generating unannotated tokens in the input corpus. Our POS induction model explicitly incorporates properties of POS tags at the type-level which is not parameterized by existing token-based approaches. This enables our model to outperform previous approaches on a range of languages that exhibit substantial syntactic variation. In our morphological segmentation model, we exploit the fact that axes are correlated within a word and between adjacent words. We surpass previous unsupervised segmentation systems on the Modern Standard Arabic Treebank data set. Finally, we showcase the utility of our unsupervised segmentation model for machine translation of the Levantine dialectal Arabic for which there is no known segmenter. We demonstrate that our segmenter outperforms supervised and knowledge-based alternatives.