Automatic Evaluation Method for Machine Translation Quality Based on Cross-Attention Recurrent Neural Network

This paper proposes an automatic scoring model of machine translation quality of cross-attention recurrent neural networks (MTQ-CARNNs). This method combines the advantages of recurrent neural network (RNN) in processing sequence data with the ability of the cross-attention mechanism in capturing complex semantic relationships between texts to achieve accurate automatic evaluation of machine translation (MT) quality. A quality assessment model of MT based on CARNN is proposed. The model uses RNN layer to encode the source text and the target text, and calculates the correlation score between the source text and the target text through the cross-attention mechanism, so as to achieve the translation quality score. The experimental results show that MTQ-CARNN shows high NIST value in multiple translation tasks and different training rounds, that is, the accuracy of translation quality assessment is significantly improved. Especially when dealing with complex semantic relationships and long-distance dependencies, the advantages of MTQ-CARNN are more obvious. In addition, MTQ-CARNN also shows strong generalization ability, and can give more accurate evaluation results for different types of translation tasks and different fields of text. The cross-attention mechanism of MTQ-CARNN not only improves the accuracy of translation quality assessment but also enhances the interpretability of the model. By visualizing the cross-attention score, we can intuitively understand the focus of the model during the evaluation process, thus further verifying the rationality of the model.