In recent years, semi-supervised learning with generative adversarial networks (GANs) has singled out for its performance in domains with little labeled data. In this paper, we propose a new approach called PaGAN which is a combination of a document classifier and a sentence classifier in a GAN for patent understanding. PaGAN is applied and evaluated on a realworld dataset. Experiments show outperforming results of PaGAN comparatively to baseline approaches.