Research And Implementation Of Optimization And Post-Processing Technology In Chinese To Mongolian Neural Machine Translation

The rise and development of neural network technology along with deep learning research has made great progress in machine translation.However,because Chinese(chi)and Mongolian(mon)have great differences in lexical and syntactic aspects,there are many research difficulties in chi-mon machine translation research.Some key technologies need to be broken and the key models of translation need to be further optimized.This thesis mainly implements a machine translation system of chi-mon neural network,researches and implements the post-processing technology and model optimization method of Mongolian translation.The specific contents of this thesis are as follows:(1)Construct a chi-mon neural machine translation model.The RNN,CNN,Transformer network structure and the existing chi-mon aligned corpus are used to complete the construction of the chi-mon translation model.(2)Using the special characters in Mongolian to design and implement post-processing methods for Mongolian translation special characters,to reduce the translation errors caused by special characters in Mongolian corpus during training to improve Mongolian translation quality.(3)Implement on the model compression method and apply it to the neural network machine translation model task to reduce the memory space required by the model while ensuring the performance of the model.(4)Build a translation system.The B/S architecture is adopted to realize a translation system composed of a combination of a browser side,a web server side and a translation server side,and a corresponding auxiliary translation function is added while completing the basic translation function.Finally,post-processing techniques were used to increase the BLEU values of the chi-mon neural machine translation model using 60,000 corpus based on RNN,CNN,and Transformer to 23.21(+2.17),24.74(+0.71),and 25.57(+0.91),respectively.The post-processing method was also verified in the CNN model obtained using260,000 corpus training.The implementation of the model compression method makes the model compression rate reach 75% and the translation speed is also greatly improved.Based on the above research,the chi-mon translation system is completed to meet the needs of users.