For many sequence prediction tasks in Natural Language Processing, modeling dependencies between individual predictions can be used to improve prediction accuracy of the sequence as a whole. Supertagging, involves assigning lexical entries to words based on lexicalized grammatical theory such as Combinatory Categorial Grammar (CCG). Previous work has used Bayesian HMMs to learn taggers for both POS tagging and supertagging separately. Modeling them jointly has the potential to produce more robust and accurate supertaggers trained with less supervision and thereby potentially help in the creation of useful models for new languages and domains. Factorial Hidden Markov Models (FHMM) support joint inference for multiple sequence prediction tasks. Here, I use them to jointly predict part-of-speech tag and supertag sequences with varying levels of supervision. I show that supervised training of FHMM models improves performance compared to standard HMMs, especially when labeled training material is scarce. Secondly, FHMMs trained from tag dictionaries rather than labeled examples also perform better than a standard HMM. Finally, I show that an FHMM and a maximum entropy Markov model can complement each other in a single step co-training setup that improves the performance of both models when there is limited labeled training material available. / text

Parsing is the task of assigning syntactic or semantic structure to a natural language sentence. This thesis focuses on syntactic parsing with Combinatory Categorial Grammar (CCG; Steedman 2000). CCG allows incremental processing, which is essential for speech recognition and some machine translation models, and it can build semantic structure in tandem with syntactic parsing. Supertagging solves a subset of the parsing task by assigning lexical types to words in a sentence using a sequence model. It has emerged as a way to improve the efficiency of full CCG parsing (Clark and Curran, 2007) by reducing the parser’s search space. This has been very successful and it is the central theme of this thesis. We begin by an analysis of how efficiency is being traded for accuracy in supertagging. Pruning the search space by supertagging is inherently approximate and to contrast this we include A* in our analysis, a classic exact search technique. Interestingly, we find that combining the two methods improves efficiency but we also demonstrate that excessive pruning by a supertagger significantly lowers the upper bound on accuracy of a CCG parser. Inspired by this analysis, we design a single integrated model with both supertagging and parsing features, rather than separating them into distinct models chained together in a pipeline. To overcome the resulting complexity, we experiment with both loopy belief propagation and dual decomposition approaches to inference, the first empirical comparison of these algorithms that we are aware of on a structured natural language processing problem. Finally, we address training the integrated model. We adopt the idea of optimising directly for a task-specific metric such as is common in other areas like statistical machine translation. We demonstrate how a novel dynamic programming algorithm enables us to optimise for F-measure, our task-specific evaluation metric, and experiment with approximations, which prove to be excellent substitutions. Each of the presented methods improves over the state-of-the-art in CCG parsing. Moreover, the improvements are additive, achieving a labelled/unlabelled dependency F-measure on CCGbank of 89.3%/94.0% with gold part-of-speech tags, and 87.2%/92.8% with automatic part-of-speech tags, the best reported results for this task to date. Our techniques are general and we expect them to apply to other parsing problems, including lexicalised tree adjoining grammar and context-free grammar parsing.

This thesis investigates the application of structured sequence classification models to multilingual natural language processing (NLP). Many tasks tackled by NLP can be framed as classification, where we seek to assign a label to a particular piece of text, be it a word, sentence or document. Yet often the labels which we’d like to assign exhibit complex internal structure, such as labelling a sentence with its parse tree, and there may be an exponential number of them to choose from. Structured classification seeks to exploit the structure of the labels in order to allow both generalisation across labels which differ by only a small amount, and tractable searches over all possible labels. In this thesis we focus on the application of conditional random field (CRF) models (Lafferty et al., 2001). These models assign an undirected graphical structure to the labels of the classification task and leverage dynamic programming algorithms to efficiently identify the optimal label for a given input. We develop a range of models for two multilingual NLP applications: word-alignment for statistical machine translation (SMT), and multilingual super tagging for highly lexicalised grammars.