The invention discloses an automatic text generation method. The method comprises the following steps of: data obtaining; data preprocessing; model construction; loss definition; model pre-training; and model adversarial training model verification. According to the method and device, a concurrent neural network and a variational auto encoder are combined, and an adversarial network training modeis added, so that the quality of generated texts is prevented from becoming bad along with elongation of the lengths of the texts, and the problem of generating single texts is avoided. According to the method, the concurrent neural network commonly used in the field of natural language processing is applied as basis and the variational auto encoder is fused, so that text distribution can be better learnt. For later effect enhancement, a thought of a generative adversarial network is used, a convolutional neural network is used as a discriminator, reinforcement learning is used for carrying out training, and models are globally trained, so that better effect is achieved. The invention furthermore discloses an automatic text generation device.