Research On Interactive Machine Translation

Machine Translation (MT) technology has achieved great development in the past fewdecades. Especially, the proposition of statistical machine translation technology becomes agreat driving force in the transformation of machine translation technology to practice.However, the state-of-the-art machine translation system is still far from practical, and there isstill a considerable distance between reality and people’s expectation. If we can say that insome special areas of some special language pairs, the translation can be generally understood,then, it is hard to say that there exist any language pair between which the translation cantransfer the original meaning of the source sentence correctly.As an acceptable replacement of automatic machine translation, Computer AidedTranslation (CAT) has achieved great development. An important pattern of CAT is the socalled Interactive Machine Translation (IMT) which is one of the efforts which aim at throughthe interaction between people and machine reaching the combination of the translator’squality and the machine’s efficiency.The works of this thesis mainly include: the research and analysis of current interactivemachine translation technology, especially the state-of-art statistical phrase-based interactivemachine translation, and also pay relative more attention to the searching strategy and thegeneration of translation hypotheses. On the basis of the above work, we have implemented astate-of-art statistical phrase-based interactive machine translation system, and for problemsfound at the searching strategy and the generation of translation hypotheses, we also proposedthree refinements for the state-of-art model. We utilize dynamic distortion limitation tobalance the requirements of long distance reordering and decoding speed. And we introducethe difference function to the translation hypothesis extension as a heuristic function, to makethe final translation candidates as diverse as possible. We also use the user validated prefix tohelp the word selection of suffix based on a word co-occurrence model. All these methodsaim at optimizing the first N-best candidate translations and look forward to reducing thecognitive burden of the users. The experiential results at Chinese to English translation show the effectiveness of our methods.