Research On Machine Translation Metrics Based On Pre-trained Model

With the rapid development of deep learning technology,neural machine translation technology has made great process and has been widely promoted and applied in the industry.Strictly speaking,however,the current target translation of machine translation is still not comparable to the level of human translation experts.The machine translation metric is a criterion for measuring the quality of machine translation based on the reference translation.A good machine translation metric can enable researchers to find the problems existing in the machine translation system in time,help the machine system to screen out low-quality machine translations,and better promote the development process of machine translation.Therefore,this study has important research significance and application value.For the question of how to evaluate the quality of machine translation,manual evaluation is the most primitive and effective evaluation method,but this method is expensive and time-consuming.Therefore,in recent years,researchers mostly use automatic machine translation metrics to complete the task of translation evaluation.With the rise of neural network,machine translation metrics based on various methods com in succession,but at present,the most widely used machine translation metrics are still based on n-gram matching such as BLEU,The method based on n-gram matching is only sensitive to the changes of the vocabulary surface,and can not recognize the changes of the sentence semantics and grammar.At the same time,in view of the particularity of the machine translation metrics task,that is,the reference translation data with standard scores is required,so another problem faced by the machine translation metrics is the lack of large-scale training data to support the training of the model.Aiming at these two problems,this paper mainly studies the machine translation metrics based on the pre-trained model.The method based on the pre-trained model can recognize the deep semantic relationship of the sentence,and on this basis,the method based on contrastive learning is introduced to further improve the performance of the pre-trained model and solve the dilemma of the scarcity of training data faced by machine translation metrics.The main work and innovations of this paper are as follows:(1)A machine translation metric based on cross language pre-trained model multi-BERT and XLM is introduced,and on this basis,the latest cross language pre-trained model XLM-R is introduced to carry out machine translation metrics task,which is used to improve the evaluation effect of the machine translation metrics.The experimental results show that the XLM-R cross language pre-trained model achieves the best effect on this task.(2)A machine translation metric method based on contrastive learning is proposed.By using pre-trained model BART to reconstruct the corrupted sentences to generate negative samples for contrastive learning,and the cross language pre-trained models are used as the underlying models to build a contrastive learning architecture in order to improve the evaluation performance of the machine translation metric.The experimental results show that the evaluation performance of the model is further improved after the introduction of contrastive learning.(3)An unsupervised learning method is proposed,that is,the pre-trained model will not be fine-tuned after the contrastive learning,but directly used for testing to predict the evaluation score of the translation.The experimental results show that this method based on unsupervised learning can make the model no longer limited by the small amount of the training data with standard scores,and can effectively solve the problem of low resource language data scarcity to a certain extent,which further reflects the feasibility of the model.