Abstract: Unsupervised neural machine translation can train models using only a large amount of monolingual data without the need of parallel data, but it is difficult to establish the connection between two linguistically distant languages. To address this problem, this paper proposes a new neural machine translation training method without parallel sentence pairs. A bilingual dictionary is used to replace words in monolingual data, so as to establish the connection between the two languages. Meanwhile, word embedding fusion initialization and dual-encoder fusion training are used to enhance the alignment of the two languages in the same semantic space, in order to improve the performance of the machine translation system. Experiments show that, compared with other unsupervised models, our method can improve the BLEU values by 2.39 and 1.29 over the baseline system on the Chinese-English and English-Chinese translation tasks, and also achieve good results on the English-Russian and English- Arabic translation tasks with monolingual data.

Key words: neural network, neural machine translation, dictionary, unsupervised