Research On Translation Quality Estimation Technology Based On Deep Learning

Translation quality estimation aims at evaluating the quality of machine translation without references,which is of great significance in computer-assisted translation.Traditional translation quality estimation methods are based on feature engineering.Through the design and selection of bilingual features,the mapping function between features and translation quality scores is further studied.With the development of deep learning technology,state-of-the-art translation quality estimation methods use neural network to automatically learn bilingual features,which overcomes the limitations of traditional methods that require manual design and selection of features.However,through detailed analysis of specific models,it is found that there are still some problems in the current quality estimation methods based on deep learning.This paper summarize these problems into two aspects.First,the two-stage end-to-end quality estimation method which widely used at present,can learn a certain amount of bilingual structural information.But,it still fails to capture the deep structural syntactic details of the sentences.Focusing on this problem,this paper proposed a quality estimation method that incorporates syntactic knowledge.By incorporating syntactic knowledge into the current model,the accuracy of quality estimation is effectively improved.Second,the pre-trained word embedding can not only introduce prior knowledge to the translation quality estimation model,but also overcome the over-fitting problem of the downstream model.However,the pre-trained word embedding used in the current research fails to carry a comprehensive contextual information.Focusing on this problem,this paper introduced contextualized word embedding features and dependency syntax features in sentence-level subtasks,and proposed a translation quality estimation method based on multi-feature fusion.Experimental results verified the effectiveness of the method.Finally,integrating the two methods proposed in this paper,a deep learning-based translation quality estimation system is implemented.It can estimate the translation quality from different granularities and further improve the performance of each individual method.