Research On Neural Translation Quality Estimation Method Based On Contextual Word Embedding

With the progress of globalization and the continuous innovation of natural language processing,human beings are increasingly dependent on machine translation systems.Although machine translation technology has helped humans to ease the barrier of communication between different languages,there's still a shortcoming: the quality of translation is difficult to judge.In order to further improve the user experience of the machine translation system and help users identify the machine translation of poor quality,the researchers proposed a method of estimating the quality of the machine translation.Translation quality estimation is the automatic scoring of the quality of the machine translation in the absence of a human translation as reference.This paper proposes a translation quality estimation method based on contextword vectors."Context" in this article means the contextual information of words with rich syntactic and semantic knowledge.It can reflect the fluency of a text and is an important concept in natural language processing tasks.Based on the existing neural network-based quality estimation method,this paper combines fluency features to estimate translation quality.In order to introduce the features of translation fluency into the existing neural quality estimation model,we propose a method of introducing context-word vectors into a neural translation quality estimation model using a stacked bidirectional long-term and short-term memory network—CUNQE.CUNQE merges context-word vectors with traditional translation quality vectors through network parallel connection.The translation quality estimation method combining context-word vectors and quality vectors comprehensively applies the loyalty and fluency features of machine translation,and is more comprehensive in the evaluation of machine translation quality.On CWMT18 and WMT18 task data sets,the effect of neural translation quality estimation method combining ELMo and BERT context-word vectors was verified and analyzed.The experimental results show that the translation quality estimation method based on context-word vector can significantly improve the correlation between translation quality estimation and human evaluation,and the neural translation quality estimation method based on BERT context-word vector fusion has the best performance.The experimental analysis further reveals that the translation quality estimation method incorporating contextual word vectors can improve the effect of translation quality estimation by utilizing the fluency feature of the translation,and indicates that the pretraining mode of the language model,the use method and quality of contextual word vectors are the key factors affecting the performance of the neural quality estimation model.