Title of Thesis: Learning Structured Classifiers for Statistical Dependency Parsing Learning Structured Classifiers for Statistical Dependency Parsing

In this thesis, I present three supervised and one semi-supervised machine learning approach for improving statistical natural language dependency parsing. I first introduce a generative approach that uses a strictly lexicalised parsing model where all the parameters are based on words, without using any part-of-speech (POS) tags or grammatical categories. Then I present an improved large margin approach for learning dependency parsers from treebank data that allows a more general set of linguistic features to be used. Specifically, I incorporate local constraints that enforce the correctness of each individual link, rather than just scoring the whole parse tree. For dealing with sparse data, I smooth the lexical parameters according to their underlying word similarities using Laplacian regularization. Third, I present a simpler and more efficient approach to training dependency parsers by applying a boosting-like procedure to standard supervised training methods. By using logistic regression as an efficient base classifier (for predicting dependency links between word pairs), I am able to efficiently train a dependency parsing model, via structured boosting, that achieves state-of-the-art results in English, and surpasses state-of-the-art in Chinese. Finally, I propose a novel semi-supervised training algorithm for learning dependency parsers. By combining a supervised large margin loss with an unsupervised least squares loss, I obtain a discriminative, convex, semi-supervised training algorithm for dependency parsing.