Abstract: To expend the existing neuro-lingusitic programming capabilities of Chinese-English dialog robots to more languages and meet the needs of human-computer interaction scenarios in mixed languages, we analyze the preprocessing mechanism of new language characteristics and propose a multi-language robots model of deep learning.We construct a translation model through multi-task joint training, introduce discriminator antagonism training and word orientation. Innovative methods such as word vector corpus sharing, localized mining mapping vector space, and cross-language knowledge distillation technology have realized knowledge transfer and automatic iteration in different language environments. The results show that the cross-language model achieves the expected results in both monolingual and mixed-language tests, which proves the validity of the model.