An Attentional Recurrent Neural Network-Based Automatic Diagnosis Method for Machine Translation Errors

Abstract

Although machine translation has achieved great progress in recent years, reliability always remains a concern in such task. Due to linguistic complexity and uncertainty, various errors are inevitably accompanied with machine translation results. It is also important to develop automatic error diagnosis methods for machine translation results. To deal with such an issue, this work combines both recurrent neural networks (RNN) and attention mechanisms. This work proposes an attentional RNN-based automatic error diagnosis method for machine translation. The collected machine translation results and corresponding reference results are collected to establish the experimental dataset. A large model named Recurrent Neural Machine Translation (RNMT) is utilized to extract semantic features from sentences generated by machine translation. A self-supervised diagnosis algorithm is accordingly formulated by introducing Jaccard similarity testing. Finally, the experiments are conducted on the specific dataset to evaluate the performance of the proposal. The research results show that the automatic diagnosis method obtained in this study has high accuracy, and it can meet the usage requirements in various aspects such as grammar, tense, single vocabulary and multi-vocabulary.