A Transformer-based Unified Neural Network For Quality Estimation Of Machine Translation

In the context of economic globalization,international communication is becoming more and more frequent,and the use of computer can greatly reduce the burden of communication between different native speakers.Although there are many advantages in machine translation,but this method still has many limitations,including the difficult way to evaluate the quality of machine translation results.To improve the quality of machine translation systems and thus to serve more customers,the key is to evaluate the quality of machine translation results.Therefore,there is a quality estimation of machine translation task.The goal of quality estimation is to evaluates the quality of the machine translation results without reference translation.Quality estimation can not only improve the performance of the machine translation system,but also plays an important role in automatic post-editing and computer aided translation.This paper proposes a Transformer-based unified neural network for quality estimation of machine translation—TUNQE.In order to use the bilingual related knowledge learned from the bilingual parallel corpus by the Transformer model,we combine the Transformer bottleneck layer and the Bi-LSTM network layer to form a unified neural network for translation quality estimation.The TUNQE model consists of two modules: one is the feature extraction module implemented using the Transformer encoder-decoder;the other is quality estimation module implemented using a recurrent neural network to calculate the quality score.Quality estimation module is essentially a supervised regression module.We combine these two modules to obtain an end-to-end neural network model.During the training process of model,the feature extraction module is pre-trained using a large scale bilingual corpus,then the feature extraction module and quality estimation module joint training using the training datasets provided by the quality estimation of machine translation task.This paper verified the performance of the TUNQE model using the quality estimation datasets provided by the WMT2018,WMT2019 and the CWMT2018sentence-level quality estimation task.The experimental results show that the TUNQE method can significantly improve the correlation between the quality estimation score and the artificial evaluation score,and further improve the performance of the integrated BERT word vector.The analysis of experimental shows the influence of the mask method of the Transformer decoder on the model performance.