Document representation models can convert unstructured text data into structured data, which is the basis of many natural language processing tasks. Currently, wordbased models cannot deal with unregistered words and documents in the document representation tasks. The generative adversarial network (GAN) can use two neural networks to deal with confrontation so as to learn the distribution of the original data well. We propose a Wasserstein adversarial document model (WADM), which uses denoising autoencoder as its discriminant network and obtains document representation directly by its hidden layer. Experiments show that the WADM can extract document features accurately and has stronger document representation capability than word-based models.