Genuine phrase-based statistical machine translation with supervision

This thesis addresses mainly two issues that have not been addressed in Statis-tical Machine Translation. One issue is that even though research has been evolving from word-based approaches to phrase-based ones, because words were consistently found to be inappropriate translation units, the fact is that words are still considered in the composition of phrases, either to determine translation equivalents or to check language fluency. Such consideration might result in the attempt of establishing relations between words within a phrase translation equivalent even when sometimes its phrases should be considered as a whole. Attempts to further partition such phrases would produce incorrect translation units that would introduce unwanted noise in the translation pro-cess. Besides, the internal fluency of an identified multi-word phrase should not require checking. As such, phrases should indeed be considered units, avoiding incorrect translation equivalents that might be identified from their partition, as well as only considering the fluency of a phrase with other phrases and not within the phrase itself. The other issue is that supervision, in the form of trans-lation lexica, is generally overlooked, with SMT research focusing mainly on the identification of translation units without any human intervention and without considering already known translation units. As such, no importance has been attributed to the inclusion of verified lexica, with only some rarely used dic-tionaries to score translation candidates and not really as a source of translation units. Indeed, translation equivalents should be memorized, checked and used as a source of translation units, avoiding the need to keep identifying the same translation units, in particular if those are frequently used. This Thesis presents a truly Phrase-Based approach to SMT, using contiguous and non-contiguous phrases, along with Supervision, in which phrases are not divided and verified lexica is built, kept and used to propose translations of complete sentences