The Machine Translation System That Combined Internet Engine

MT(Machine Translation)has made remarkable results after decades of development from the proposed to the present.During this period various methods have been raised.Now,the most popular MT methods are statistical machine translation and the rising neural machine translation.Each machine translation methods have their own unique advantages,therefore the approach combination of Multi-Engine MT was proposed in order to "lean from each other",hoping to optimize the translation results through the system combination.At present,the application of machine translation in the industry has been very mature,Baidu,Youdao and Google etc.have launched their own online MT system,this study used these online MT systems and the system trained by Moses to accomplish the combination of Multi-Engine.According to the basic operation unit,the system combination can be divided into three types:sentence-level combination,phrase-level combination and word-level combination.This study used the sentence-level combination,word-level combination and the combination based on MEMT in Chinese-English translation task.The sentence-level combination use MBR(Minimum Bayes Risk)decoding method,and different loss functions were used when decoding.When using the TER as the loss function achieves the best results,raised 0.24 BLEU point than the best results before combination.The word-level combination needs to build CN(Confusion Network)to get results.In the experiment,we compared different word alignment methods when building CN and use different features when decoding.The results show that using TER(Translation Edit Rate)for word alignment and adding multi-feature when decoding can improve the effect of system.The experiment also achieves the best results of this study,which raised 0.78 and 3.01 BLEU points than the best and worst result before combination.MEMT-based system combination performance in general,raised 0.48 BLEU point than the best results before combination.Experimental results show that the machine translation system that combined internet engine can improve the quality of translation.In the end,the study implements a web system based on B/S model,using word-level system combination approach.