Leveraging Diverse Sources in Statistical Machine Translation

Statistical machine translation is often faced with the problem of having insufficient training data for many language pairs. In this thesis, several methods have been proposed to leverage other sources to enhance the quality of machine translation systems. Particularly, we propose approaches suitable in these four scenarios: 1. when an additional parallel corpus between the source and the target language is available (ensemble decoding); 2. when parallel corpora between the source language and a third language and between that language and the target language are available (ensemble triangulation); 3. when an abundant source language monolingual corpus is available (graph propagation for paraphrasing out-of-vocabulary words); 4. when no additional resource is available (stacking). In the heart of these solutions lie two novel approaches: ensemble decoding and a graph propagation approach for paraphrasing out-of-vocabulary (oov) words. Ensemble decoding combines a number of translation systems dynamically at the decoding step. We evaluate its performance on a domain adaptation setting where a model trained on a large parliamentary domain is adapted to the medical domain, we then translate sentences from the medical domain. Our experimental results show that ensemble decoding outperforms various strong baselines including mixture models, the current state-of-the-art for domain adaptation in machine translation. We extend ensemble decoding to do triangulation on-the-fly when there exist parallel corpora between the source language and one or multiple pivot languages and between those and the target language. These triangulated systems are dynamically combined together