Statistical Models for Unsupervised Learning of Morphology and POS Tagging

This thesis concentrates on two fields in natural language processing. The main contribution of the thesis is in the field of morphology learning. Morphology is the study of how words are formed combining different language constituents (called morphemes) and morphology learning is the process of analysing words, by splitting into these constituents. In the scope of this thesis, morphology is learned mainly by paradigmatic approaches, in which words are analysed in groups, called paradigms. Paradigms are morphological structures having the capability of generating various word forms. We propose approaches for capturing paradigms to perform morphological segmentation. One of the approaches proposed captures paradigms within a hierarchical tree structure. Using a hierarchical structure covers a wide range of paradigms by spotting morphological similarities.

The second scope of the thesis is part-of-speech (POS) tagging. Parts-of-speech are linguistic categories, which group words having similar syntactic features, i.e. noun, adjective, verb etc. In the thesis, we investigate how to exploit POS tags to learn morphology. We propose a model to capture paradigms through syntactic categories. When syntactic categories are provided, the proposed system can capture paradigms well. Following this approach, we extend it for the case of having no syntactic categories provided. To this end, we propose a joint model, in which POS tags and morphology are learned simultaneously. Our results show that a joint model is possible for learning morphology and POS tagging.

We also study morpheme labelling, for which we propose a clustering algorithm that groups morphemes showing similar features. The algorithm can capture morphemes having similar meanings.