DOI: https://doi.org/10.21203/rs.3.rs-1249163/v1
At present, high-quality English-Chinese parallel corpus is currently in a stage of scarcity. How to use existing monolingual corpus data for unsupervised training has become an important research direction of English-Chinese machine translation. This work proposes an unsupervised English neural machine translation system and deploys it in a wireless network to perform unsupervised English translation. The specific work content of this work is as follows: 1) To alleviate the issue of low accuracy for unsupervised English-Chinese dictionary induction methods, an unsupervised English-Chinese dictionary induction method based on translation model is proposed. First, an unsupervised English-Chinese statistical machine translation model is built, and with this model, a bilingual dictionary is summarized through statistical word alignment technology. After that, the bilingual dictionary summarized by the translation model is combined with the language model, unsupervised translation model is initialized, unsupervised English-Chinese neural machine translation model is optimized through the back translation method. 2) To alleviate the problem that the unsupervised translation model built based on the denoising autoencoder is not natural and accurate, this paper implements a pre-trained English-Chinese language model based on the masked sequence-to-sequence (MASS) method. First, use monolingual corpus combined with random masking, and train the Transformer together to obtain a language model.