Machine translation has achieved impressive performance with the advances in deep learning and rely on large scale parallel corpora.There have been a large number of attempts to extend these successes to low-resource lan-guage,yet requiring large parallel sentences.In this study,we build the Mongo-lian-Chinese neural machine translation model based on unsupervised methods.Cross-lingual word embedding training plays a crucial role in unsupervised ma-chine translation which generative adversarial networks(GANs)training meth-ods only perform well between two closely-related languages,yet the self-learn-ing method can learn high-quality bilingual embedding mappings without any parallel corpora in low-source language.In this work,apply the self-learning method is better than using GANs to improve the BLEU score of 1.0.On this basis,we analyze the Mongolian word lexical features and use stem-affixes seg-mentation in Mongolian to replace the Bytes-Pair-Encoding(BPE)operation,so that the cross-lingual word embedding training is more accurate,and obtain higher quality bilingual words embedding to enhance translation performance.We reporting BLEU score of 15.2 on the CWMT2017 Mongolian-Chinese da-taset,without using any parallel corpora during training.