Neural Machine Translation Based Translation Quality Estimation

The sentence-level machine translation quality estimation(Sentence-level QE)takes the source language sentence and the corresponding machine translation as input to estimate the quality of the translation.With the development of machine translation in recent years,machine translation quality estimation has gradually b ecome a hot topic in the field of natural language processing.With the development of deep learning in recent years,the neural machine translation(NMT)model has attracted extensive attention.Moreover,the machine translation quality estimation task and machine translation are closely related,thus this paper hopes to use the neural machine translation model to improve the performance of the machine translation quality estimation model.Firstly,this paper carries out some basic experiments,including the QE experiments using the average of the word embedding as features.Through these experiments,it is concluded that the use of the appropriate features has a greater impact then the machine learning model.One of the methods of extracting QE features by means of the neural network is to directly average the word embedding of sentence words.This method completely ignores the useful information such as word order and interaction between words.Therefore,this paper proposes a method for extracting the features that co ntain translation knowledge by means of a neural machine translation model.The results of experiments have proved that this feature is more predictive than the previous word embedding features.On the basis of this,the paper tries to improve the process of extracting the above features,but the experimental results show that the attempt does not improve the prediction power.Finally,the paper explores different combinations of this feature and other features.After adding other features,the pr edictive effect gets improved,demonstrating that this feature and other features are complementary to a certain extent.The extraction process of the above features proposed in the paper is essentially modeling the source sentence and the target sentence respectively.The Sentence-level QE task requires the model to predict the translation quality(HTER),which requires more detailed modeling of word-level(phrase-level)connections between source sentence and the corresponding machine translation.The word-level QE model is required to detect errors for each token in MT output,which is more detailed than the above method in terms of modeling the interactions of words between source sentence and the translation.Therefore,the paper also explores the use of word-level QE models to predict Sentence-level QE target(HTER)and achieves good research results.