Research And Implementation Of Chinese-Vietnamese Neural Machine Translation Integrating Translation Knowledge

Vietnam is located in the Indo-China Peninsula,adjacent to China's Guangxi and Yunnan provinces,and is an important partner of our country in politics and economy fields.Under the “One Belt and One Road” strategy,there is huge potential cooperation between China and Vietnam.The increasingly economic and cultural communication between the two countries has put more and more pressing demands on the machine translation technology of the two languages.However,the demand for parallel corpus and the scarcity of corpus in existing machine translation techniques have become major contradictions that hinder the development of this technology.In recent years,neural machine translation technology has achieved performance beyond statistical machine translation on multiple language pairs,and achieved unprecedented success in translation fluency.However,this technology also requires a large amount of parallel corpus as training data.At the same time,the results of statistical machine translation research in the past should not be ignored.Therefore,how to integrate the translation knowledge in statistical machine translation into the neural machine translation framework has become the current research hotspot.Based on this question,this paper applies contextual constraints and language features to word alignment modeling,studies the word alignment model based on recurrent neural networks,and further studies how to integrate translation knowledge into neural machine translation models.Main aspects of the research are as follows:(1)Chinese-Vietnamese word alignment method based on bidirectional LSTM and linguistic features.Word alignment is the basis of various machine translation models(eg,phrase based translation model,syntactic based translation model),as well as the basis for phrase alignment,and syntax tree alignment.Therefore,it is an important part of statistical machine translation methods.In this paper,we use the recurrent neural network to model the word alignment and introduce the context information of the alignment position.Besides,the Chinese-Vietnamese linguistic differences are integrated,and the model parameters are adjusted based on the supervised training to obtain better performance.(2)Chinese-Vietnamese neural machine translation method integrating lexical translation probabilities.The parameters of neural machine translation methods are poorly interpreted,and it is difficult to integrate external translation knowledge such as word alignment and dictionaries.At the same time,although the internal attention mechanism can show the correspondence between source language words and target language words,but the relationships do not have explicit restrictions.In this paper a method is proposed to integrate lexical translation probabilities into neural machine translation.Firstly,a memory neural network is used to store the lexical translation probabilities and convert the probabilities from discrete features to continuous features.Then in the decoding stage of neural machine translation,the alignment probability of the target language word and each source language word is calculated according to the decoding context information.Finally,this information is combined with the attention mechanism to help choosing target word in the decoding stage.(3)Chinese-Vietnamese neural machine translation system integrating translation knowledge.The system includes a machine translation module,a web backend service module,and a web front end service module.The implementation of the translation model is based on Google's implementation of NMT and a lot of improvements are made in this paper,mainly including the modification of the traditional attention mechanism,and the integration of lexical translation probabilities into the neural machine translation framework.The system will play an important role in the integration of translation knowledge with neural machine translation.