Joint pairwise learning and masked language models for neural machine translation of English

Abstract

The translation activity of language is a link and bridge for the integration of politics, economy, and culture in various countries. However, manual translation requires high quality of professional translators and takes a long time. The study attempts to introduce dual learning on the basis of traditional neural machine translation models. The improved neural machine translation model includes decoding of the source language and target language. With the help of the source language encoder, forward translation, backward backtranslation, and parallel decoding can be achieved; At the same time, adversarial training is carried out using a corpus containing noise to enhance the robustness of the model, enriching the technical and theoretical knowledge of existing neural machine translation models. The test results show that compared with the training speed of the baseline model, the training speed of the constructed model is 115 K words/s and the decoding speed is 2647 K words/s, which is 7.65 times faster than the decoding speed, and the translation quality loss is within the acceptable range. The mean bilingual evaluation score for the “two-step” training method was 16.51, an increase of 3.64 points from the lowest score, and the K-nearest-neighbor algorithm and the changing-character attack ensured the semantic integrity of noisy source language utterances to a greater extent. The translation quality of the changing character method outperformed that of the unrestricted noise attack method, with the highest bilingual evaluation study score value improving by 3.34 points and improving the robustness of the model. The translation model constructed by the study has been improved in terms of training speed and robustness performance, and is of practical use in many translation domains.