Abstract: Aimed at the problems of poor translation quality and low hiding capacity of text steganography methods based on machine translation, a text steganography method based on Transformer neural machine translation is proposed. Through the use of advanced neural machine translation model and Beam Search decoder with hidden module, the corresponding BPE code could be output according to the secret information at each time step, and the secret information could be embedded in the process of translation. Experimental results show that this method can generate steganographic translation with high quality and large hidden capacity. Compared with previous methods, BLEU value and hiding capacity are improved by 4.84 and 1.63 percentage points, respectively.