DGSAN: Discrete generative self-adversarial network

Although GAN-based methods have received many achievements in the last few years, they not been entirely successful generating discrete data. The most crucial challenge of these is difficulty passing gradient from discriminator to generator when outputs are discrete. Despite fact that several attempts made alleviate this problem, none existing improved performance text generation compared with maximum likelihood approach terms both quality and diversity. In paper, we proposed a new framework for data by an adversarial which there no need pass generator. method has iterative manner each defined based on discriminator. It leverages discreteness model real distribution implicitly. Moreover, supported theoretical guarantees, experimental results generally show superiority DGSAN other popular or recent sequential