Research On Improving Tense Translation In Chinese-English Neural Machine Translation

In recent years,with the application of deep learning in the field of natural language processing,rapid advances have been made in many areas,especially in the task of machine translation,neural machine translation(NMT)has achieved many breakthroughs in comparison with traditional statistical machine translation(SMT).It also brought the translation effect to an unprecedented height.However,when a language like Chinese lacks morphological changes is translated into English,because the Chinese verb tense is not clearly identified,the English tense is directly indicated by the verb changes,which makes it very difficult to maintain tense consistency before and after translation.In the era of statistical machine translation,many scholars have studied the issue of maintaining the tense consistency between the source side and the target side of translation.Since entering the age of neural machine translation,few people have been involved.However,through our investigation,we find that the current neural machine translation system still has a serious problem of tense consistency.Therefore,we propose two methods to try to solve the tense consistency issues for the current state of the Chinese-English neural machine translation system through different ideas.The first method to solve the tense consistency is to adopt the method of transmitting the Chinese temporal information from the source side to the target side.Firstly,we use the neural network to construct the Chinese tense tagging model.The verb tense of Chinese sentences is obtained through this model before translation.In the process of translation,use the alignment matrix of the Attention to pass the word tense of the source side to the target side.Reduce the probability of English words that do not match the tense of the candidate word produced by the translation and the source side tense.In this way,the tense consistency in word level from Chinese to English can be basically achieved.For the tense annotation model and the NMT system with tense annotation,we conducted detailed experiments.The experimental results also show that our model is effective.The second method of this paper is to look at the problem of tense consistency from another angle.According to the deep-learning "end-to-end" idea,we try to generate the target-side tense directly through the neural network,instead of acquiring it from the source side.By constructing the tense Attention module,we use it to pay attention to the content related to the temporal expression at the source side,and then generate the tense corresponding to each time step of the Decoder,and then use this tense prediction to compare with the tense of the target candidate words.Reduce the probability of candidate words that are inconsistent in tense.The experimental results also demonstrate the effectiveness of this method for solving tense consistency problems.