Machine translation is a challenging task that its difficulties arise from several characteristics of natural language. The main focus of this work is on reordering as one of the major problems in MT and statistical MT, which is the method investigated in this research. The reordering problem in SMT originates from the fact that not all the words in a sentence can be consecutively translated. This means words must be skipped and be translated out of their order in the source sentence to produce a fluent and grammatically correct sentence in the target language. The main reason that reordering is needed is the fundamental word order differences between languages. Therefore, reordering becomes a more dominant issue, the more source and target languages are structurally different. The aim of this thesis is to study the reordering phenomenon by proposing new methods of dealing with reordering in SMT decoders and evaluating the effectiveness of the methods and the importance of reordering in the context of natural language processing tasks. In other words, we propose novel ways of performing the decoding to improve the reordering capabilities of the SMT decoder and in addition we explore the effect of improving the reordering on the quality of specific NLP tasks, namely named entity recognition and cross-lingual text association. Meanwhile, we go beyond reordering in text association and present a method to perform cross-lingual text fragment alignment, based on models of divergence from randomness. The main contribution of this thesis is a novel method named dynamic distortion, which is designed to improve the ability of the phrase-based decoder in performing reordering by adjusting the distortion parameter based on the translation context. The model employs a discriminative reordering model, which is combining several fea- 2 tures including lexical and syntactic, to predict the necessary distortion limit for each sentence and each hypothesis expansion. The discriminative reordering model is also integrated into the decoder as an extra feature. The method achieves substantial improvements over the baseline without increase in the decoding time by avoiding reordering in unnecessary positions. Another novel method is also presented to extend the phrase-based decoder to dynamically chunk, reorder, and apply phrase translations in tandem. Words inside the chunks are moved together to enable the decoder to make long-distance reorderings to capture the word order differences between languages with different sentence structures. Another aspect of this work is the task-based evaluation of the reordering methods and other translation algorithms used in the phrase-based SMT systems. With more successful SMT systems, performing multi-lingual and cross-lingual tasks through translating becomes more feasible. We have devised a method to evaluate the performance of state-of-the art named entity recognisers on the text translated by a SMT decoder. Specifically, we investigated the effect of word reordering and incorporating reordering models in improving the quality of named entity extraction. In addition to empirically investigating the effect of translation in the context of crosslingual document association, we have described a text fragment alignment algorithm to find sections of the two documents in different languages, that are content-wise related. The algorithm uses similarity measures based on divergence from randomness and word-based translation models to perform text fragment alignment on a collection of documents in two different languages. All the methods proposed in this thesis are extensively empirically examined. We have tested all the algorithms on common translation collections used in different evaluation campaigns. Well known automatic evaluation metrics are used to compare the suggested methods to a state-of-the art baseline and results are analysed and discussed.

In Statistical Machine Translation (SMT), inference needs to be performed over a high-complexity discrete distribution de ned by the intersection between a translation hypergraph and a target language model. This distribution is too complex to be represented exactly and one typically resorts to approximation techniques either to perform optimisation { the task of searching for the optimum translation { or sampling { the task of nding a subset of translations that is statistically representative of the goal distribution. Beam-search is an example of an approximate optimisation technique, where maximisation is performed over a heuristically pruned representation of the goal distribution. For inference tasks other than optimisation, rather than nding a single optimum, one is really interested in obtaining a set of probabilistic samples from the distribution. This is the case in training where one wishes to obtain unbiased estimates of expectations in order to t the parameters of a model. Samples are also necessary in consensus decoding where one chooses from a sample of likely translations the one that minimises a loss function. Due to the additional computational challenges posed by sampling, n-best lists, a by-product of optimisation, are typically used as a biased approximation to true probabilistic samples. A more direct procedure is to attempt to directly draw samples from the underlying distribution rather than rely on n-best list approximations. Markov Chain Monte Carlo (MCMC) methods, such as Gibbs sampling, o er a way to overcome the tractability issues in sampling, however their convergence properties are hard to assess. That is, it is di cult to know when, if ever, an MCMC sampler is producing samples that are compatible iii with the goal distribution. Rejection sampling, a Monte Carlo (MC) method, is more fundamental and natural, it o ers strong guarantees, such as unbiased samples, but is typically hard to design for distributions of the kind addressed in SMT, rendering an intractable method. A recent technique that stresses a uni ed view between the two types of inference tasks discussed here | optimisation and sampling | is the OS approach. OS can be seen as a cross between Adaptive Rejection Sampling (an MC method) and A optimisation. In this view the intractable goal distribution is upperbounded by a simpler (thus tractable) proxy distribution, which is then incrementally re ned to be closer to the goal until the maximum is found, or until the sampling performance exceeds a certain level. This thesis introduces an approach to exact optimisation and exact sampling in SMT by addressing the tractability issues associated with the intersection between the translation hypergraph and the language model. The two forms of inference are handled in a uni ed framework based on the OS approach. In short, an intractable goal distribution, over which one wishes to perform inference, is upperbounded by tractable proposal distributions. A proposal represents a relaxed version of the complete space of weighted translation derivations, where relaxation happens with respect to the incorporation of the language model. These proposals give an optimistic view on the true model and allow for easier and faster search using standard dynamic programming techniques. In the OS approach, such proposals are used to perform a form of adaptive rejection sampling. In rejection sampling, samples are drawn from a proposal distribution and accepted or rejected as a function of the mismatch between the proposal and the goal. The technique is adaptive in that rejected samples are used to motivate a re nement of the upperbound proposal that brings it closer to the goal, improving the rate of acceptance. Optimisation can be connected to an extreme form of sampling, thus the framework introduced here suits both exact optimisation and exact iv sampling. Exact optimisation means that the global maximum is found with a certi cate of optimality. Exact sampling means that unbiased samples are independently drawn from the goal distribution. We show that by using this approach exact inference is feasible using only a fraction of the time and space that would be required by a full intersection, without recourse to pruning techniques that only provide approximate solutions. We also show that the vast majority of the entries (n-grams) in a language model can be summarised by shorter and optimistic entries. This means that the computational complexity of our approach is less sensitive to the order of the language model distribution than a full intersection would be. Particularly in the case of sampling, we show that it is possible to draw exact samples compatible with distributions which incorporate a high-order language model component from proxy distributions that are much simpler. In this thesis, exact inference is performed in the context of both hierarchical and phrase-based models of translation, the latter characterising a problem that is NP-complete in nature.