Research On Unsupervised Neural Machine Translation Technique

Machine translation is one of the most successful and progressive fields of natural language processing in recent years.After using the technology of deep learning,the performance on some language pairs has reached the human level,but it relies too much on large-scale parallel corpora.The technology of unsupervised neural machine translation only uses monolingual corpora for training,which solves the above problem.Unsupervised neural machine translation combines denoising auto-encoder and reverse translation to achieve the machine translation task.Although unsupervised neural machine translation has achieved good performance at the present stage,through the analysis of the structure and principle of the specific model,it is found that the state-of-the-art unsupervised neural machine translation model still has some problems to be solved,which are summarized in the following two aspects in this paper.First,back-translation is a widely used method in unsupervised neural machine translation,which can transform routine unsupervised learning tasks into supervised learning tasks.However,the quality of the training data cannot be controlled during the back-translation process.To solve this problem,this paper proposes a quality estimation method to filter the pseudo-parallel data generated during back translation,which can effectively improve the performance of unsupervised neural machine translation.Second,with the widespread use of the Transformer model in neural machine translation,the performance of neural machine translation has improved significantly.However,although Transformer can learn the underlying syntactic knowledge from sentences,it is far from being sufficient.Therefore,focusing on this problem,this paper integrates syntactic knowledge in unsupervised neural machine translation.By linearizing the results of syntactic parsing,syntactic knowledge is explicitly integrated into unsupervised neural machine translation,and the effectiveness of this method is verified by experiments.Finally,we integrate the two methods proposed in this paper,and designed and implemented the unsupervised neural machine translation system.This system uses the unsupervised method to complete the task of machine translation,and further improves the quality of unsupervised neural machine translation compared with the baseline method.