The Neural Automatic Post-Editing Based On Quality Estimation

Recent years,with the successful application and further study of deep learning in the Machine Translation(MT),the corresponding method of Automatic Post-Edition(APE)has been shifted from the traditional statistical model into deep learning model.Thus,how to apply the machine translation model effectively to the post edition has become one of the hot issues in machine translation researches.One pervasive problem existed in APE is over-correction.Traditionally,most current solutions to this problem are to add penalty parameter in sorting the candidate translation to control the MT modification degree caused by post-edition system.However,this kind of solution has failed to take account of performance of post-edition system and over-correction.Thus,aiming at the shortage of exiting method,this paper has proposed a neural automatic post edition based on the quality estimation(QE).First,the author has counted the number of edits needed in the original machine translations prone to overcorrection and made a further analysis by Zipf distribution fitting.Then,in the accordance with analyzed results,the original machine translations have been classified into several types,on which neural APE models are based.In the end,the author has created hierarchical sorting method and jointed sub-models by using QE method.In order to test the performance of the proposed method,the author has performed experimental validation in the WMT'16 and WMT'17 machine translation APE evaluation tasks.And the result shows the neural APE method based on QE will improve the quality of original machine translation and compared to the conventional method of neural APE,the proposed method can effectively reduce the frequency of over-correction in the post-editions.