Error Driven Learning Based Machine Translation Post-edit Modeling

The existing Machine Translation engine cannot meet the needs of users,and users often need to correct a large number of similar errors existing inmachine translations. Corrections of the repeatable error will cost too much timeand hurt the user feedback. Therefore, there are many researches on machinetranslation automatic post editing. However, the mainstream of the post editingmethod is based on statistical machine translation (SMT), but opacity of SMTbrings problems to researches. In this context, this paper introduces error drivenlearning mechanisms, respectively, for translation errors and word order errorswith extracting post-editing rule, so as to improve the quality of machinetranslation. The main research contents in this paper include the followingaspects:(1) Error driven learning modeling for the translation word error. Themachine translation engine may appear missing words, extra words, and thewrong word. Therefore, with error driven learning framework, we learn the postediting rules for translation word errors. And through the changes of machinetranslation quality, we can know performance of the post-editing rules.(2) Error driven learning modeling for the wrong order words. The machinetranslation engine may appear wrong order words. Therefore, we defined wordorder errors through aligned crossover. With the error driven learning framework,we learn the post editing rules for wrong order words. And through the changesof machine translation quality, we can know performance of the post-editingrules.(3) Mixed strategy based error driven learning modeling for the translationword errors and wrong order words. Machine translation errors contains thetranslation word errors and wrong order errors, in order to simultaneously solvethe translation word errors and wrong order errors, we fusion the above twomethods in the level of model. At first, through the cascade error driven learningwe fusion post editing rules, then the iteratively. At last, through the changes ofmachine translation quality, we can know performance of the post-editing rules.