Data-oriented models of parsing and translation [thesis]

The merits of combining the positive elements of the rule-based and data-driven approaches to MT are clear: a combined model has the potential to be highly accurate, robust, cost-effective to build and adaptable. While the merits are clear, however, how best to combine these techniques into a model which retains the positive characteristics of each approach, while inheriting as few of the disadvantages as possible, remains an unsolved problem. One possible solution to this challenge is the Data-Oriented Translation (DOT) model originally proposed by Poutsma (1998, 2000, 2003), which is based on Data-Oriented Parsing (DOP) (e.g. (Bod, 1992; Bod et al., 2003)) and combines examples, linguistic information and a statistical translation model. In this thesis, we seek to establish how the DOT model of translation relates to the other main MT methodologies currently in use. We find that this model differs from other hybrid models of MT in that it inextricably interweaves the philosophies of the rule-based, example-based and statistical approaches in an integrated framework. Although DOT embodies many positive characteristics on a theoretical level, it also inherits the computational complexity associated with DOP. Previous experiments assessing the performance of the DOT model of translation were small in scale and the training data used was not ideally suited to the task (Poutsma, 2000, 2003). However, the algorithmic limitations of the DOT implementation used to perform these experiments prevented a more informative assessment from being carried out. In this thesis, we look to the innovative solutions developed to meet the challenges of implementing the DOP model, and investigate their application to DOT. This investigation culminates in the development of a DOT system; this system allows us to perform translation experiments which are on a larger scale and incorporate greater translational complexity than heretofore. Our evaluation indicates that the positive characteristics of the model identified on a theoretical level are also in evidence when it is subjected to empirical assessment. For example, in terms of exact match accuracy, the DOT model outperforms an SMT model trained and tested on the same data by up to 89.73%. The DOP and DOT models for which we provide empirical evaluations assume contextfree phrase-structure tree representations. However, such models can also be developed for more sophisticated linguistic formalisms. In this thesis, we also focus on the efforts which have been made to integrate the representations of Lexical-Functional Grammar (LFG) with DOP and DOT. We investigate the usefulness of the algorithms developed for DOP (and adapted here to Tree-DOT) when implementing the (more complex) LFG-DOP and LFG-DOT models. We examine how constraints are employed in these models for more accurate disambiguation and seek an alternative methodology for improved constraint specification. We also hypothesise as to how the constraints used to predict both good parses and good translations might be pruned in a motivated fashion. Finally, we explore the relationship between translational equivalence and limited generalisation reusability for both the tree-based and LFG-based DOT models, focussing on how this relationship differs depending on which formalism is assumed.