Training Machine Translation for Human Acceptability

Discriminative training, a.k.a. tuning, is an important part of Statistical Machine Translation. This step optimises weights for the several statistical models and heuristics used in a machine translation system, in order to balance their relative effect on the translation output. Different weights lead to significant changes in the quality of translation outputs, and thus selecting appropriate weights is of key importance. This thesis addresses three major problems with current discriminative training methods in order to improve translation quality. First, we design more accurate automatic machine translation evaluation metrics that have better correlation with human judgements. An automatic evaluation metric is used in the loss function in most discriminative training methods, however what the best metric is for this purpose is still an open question. In this thesis we propose two novel evaluation metrics that achieve better correlation with human judgements than the current de facto standard, the BLEU metric. We show that these metrics can improve translation quality when used in discriminative training. Second, we design an algorithm to select sentence pairs for training the discriminative learner from large pools of freely available parallel sentences. These resources tend to be noisy and include translations of varying degrees of quality and suitability for the translation task at hand, especially if obtained using crowdsourcing methods. Nevertheless, they are crucial when professionally created training data is scarce or unavailable. There is very little previous research on the data selection for discriminative training. Our novel data selection algorithm does not require knowledge of the test set nor uses decoding outputs, and is thus more generally useful and efficient. Our experiments show that with this data selection algorithm, translation quality consistently improves over strong baselines. Finally, the third component of the thesis is a novel weighted ranking-based optimisation algorithm for discriminative training. In contrast to previous approaches, this technique assigns a different weight to each training instance according to its reachability and its relationship to test sentence being decoded, a form of transductive learning. Our experimental results show improvements over a modern state-of-the-art method across different language pairs. Overall, the proposed approaches lead to better translation quality when compared strong baselines in our experiments, both in isolation and when combined, and can be easily applied to most existing statistical machine translation approaches