A Study On Mongolian-Chinese Machine Translation Based On Neural Network

With the increasing demand for information exchange and the need of stable national ethnic areas,the research and application of Mongolian-Chinese machine translation is imminent.The study of Mongolian-Chinese machine translation have great significance to promote the informatization of Inner Mongolia Autonomous Region and facilitate the social progress and economic development of the autonomous region,prosper and develop the cultural and educational cause of ethnic minorities and progress of science and technology.In recent years,the rapid development of neural machine translation is expected to replace statistical machine translation as the new mainstream technology.According to the characteristics of Mongolian language and the latest method of machine translation,we study the Mongolian-Chinese machine translation based on neural network.This article mainly does the following research work.(1)Using TensorFlow to build the Mongolian-Chinese end-to-end neural machine translation system.The framework of this system is the Encoder-Decoder framework,The encoder and Decoder both use recurrent neural network.Then the word-level Mongolian-Chinese machine translation experiment was done on this system.The best NIST and BLEU values of the results were 8.9223 and 0.6067.(2)Conventional neural machine translation models use fixed modest-size vocabulary,Mongolian is an Agglutinative language,it has many different types of morphological changes,there are many variants of Mongolian nouns and verbs with the same root,which represent similar concepts.Different deformations of the same stem are trained as different units,which increases the size of vocabulary and increases the difficulty of neural machine translation.In order to weigh the characters rich in Mongolian morphology and the vocabulary limitation of neural machine translation,this paper uses phrase-based statistical machine translation and CRF model to segment Mongolian word.Firstly,use the phrase-based statistical machine translation segmentation model to segment the Mongolian words,and then use model based on the conditional random field to segment the unknown words.(3)After segment the Mongolian word,Mongolian stem and affixes were incorporated into the machine translation of the Mongolian-Chinese neural machine,and a morpheme-based neural machine translation experiment was conducted.Experiments show that morpheme-based neural machine translation is better than word-based translation,the NIST and BLEU of the optimal model reached 9.4216 and 0.6320.