Research On Machine Translation Quality Estimation Based On Knowledge Transfer

With the rapid development of economy,culture and technology in nowadays so-ciety,machine translation technology is of great importance as a bridge connecting different languages.For the translation results produced by the machine translation system,the translation quality needs to be evaluated by quantitative evaluation criteria,which is machine translation evaluation.Machine translation evaluation has become a widely studied and discussed problem in the field of machine translation.Machine translation quality estimation is a machine translation evaluation method that evaluates the quality of translation results without relying on references given the source sentences and translation results of a machine translation system.Traditional machine translation quality estimation methods reflect the quality of translation results through the linguistic features of artificial design,but they depend on the language itself,as well as specific linguistic resources and tools.The machine translation quality estimation method with deep learning technology can use the neural network model to learn the features of the word embedding and the language model,but it does not fully model the interaction between the source sentences and the translation results.The machine translation quality estimation based on pure neural network uses bilingual parallel corpus and machine translation model to automatically learn high-dimensional feature representation,but this method is not suitable for low-resource language pairs that are not easily available in parallel corpus,and the automatically learned feature representations are not necessarily suitable for the quality estimation task itself.This paper focuses on the application of deep learning technology in the qual-ity estimation of machine translation,especially the research on the quality estimation method based on knowledge transfer through neural network model.The main contri-bution of this paper is as follows:1.To strengthen the interactive relationship between source sentences and translation results and obtain the feature representations that are more suitable for quality esti-mation task,this paper proposes a bi-directional quality estimation model that can integrate bilingual knowledge.By combining the bi-directional machine translation model with the quality estimation task,the mutual translation relationship between the source sentences and the translation results can be fully modeled,and the more informative and effective feature representation can be learned.By transferring the bi-directional bilingual knowledge to the quality estimation task,the representation ability of the whole machine translation quality estimation model is enhanced.At the same time,the training method based on joint learning is applied to the ma-chine translation quality estimation model,so that the automatically learned feature representations are more suitable for the quality estimation task itself.2.To solve the problem that existing methods cannot effectively model the low-resource language pairs that are not easily available in parallel corpus,this paper proposes a machine translation quality estimation model that can integrate monolingual knowl-edge,which is easier to be acquired than bilingual knowledge.This paper also ex-plores the application of the best distributed representation learning model in the quality estimation task,which can model the interaction between source sentences and translation results at a finer-grained level.At the same time,large-scale mono-lingual knowledge can be transferred to the quality estimation task,and help quality estimation model to achieve better model performance.This paper conducts experiments on the data set of the machine translation quality estimation task of WMT 2017 English-German translation,and the experimental results prove the effectiveness of the machine translation quality estimation model proposed in this paper,and it can reach and outperform the current best quality estimation model.