Abstract: Text generative models trained via maximum likelihood estimation (MLE) suffer from the notorious exposure bias problem, and generative adversarial networks (GANs) are shown to have potential ...