Training and models for statistical machine translation

Machine translation is the application of machines to translate text or speech from one natural language to another. Nowadays machines typically refer to computers and computer software.One common algorithmic approach to machine translation is the statistical phrase-based translation model. With statistical models, the two key aspects are the model itself and the training thereof. This work describes and analyzes both these aspects.We will present an algorithm and procedure for consistent training of phrase-based statistical translation models. The training involves a forced alignment method, where a phrase alignment is found between the source training instances and the given translation, using all models and components of the translation process. We will show that this leads to smaller models with higher translation quality.We will investigate different smoothing techniques that improve the generalization of our translation models when applied to previously unseen, out-of-domain test data. We introduce phrase count features and consistently trained word-lexicon models. Both attempts result in improvements in translation quality.We will show models that use machine learning components to improve lexical selection. In order to train these proposed extended lexicon models, we will use efficient algorithms and data structures for training large-scale sparse maximum entropy models. These methods reduce the time and memory required to train the extended lexicon models by several orders of magnitude.By combining this work on consistent training of phrase-based and efficient machine-learning we show the feasibility of a full discriminative model for phrase-based statistical machine translation.The resulting machine translation systems improve translation quality on a range of tasks and language pairs, tested in public evaluation campaigns