WRGAN: Improvement of RelGAN with Wasserstein Loss for Text Generation

Abstract

Generative adversarial networks (GANs) were first proposed in 2014, and have been widely used in computer vision, such as for image generation and other tasks. However, the GANs used for text generation have made slow progress. One of the reasons is that the discriminator’s guidance for the generator is too weak, which means that the generator can only get a “true or false” probability in return. Compared with the current loss function, the Wasserstein distance can provide more information to the generator, but RelGAN does not work well with Wasserstein distance in experiments. In this paper, we propose an improved neural network based on RelGAN and Wasserstein loss named WRGAN. Differently from RelGAN, we modified the discriminator network structure with 1D convolution of multiple different kernel sizes. Correspondingly, we also changed the loss function of the network with a gradient penalty Wasserstein loss. Our experiments on multiple public datasets show that WRGAN outperforms most of the existing state-of-the-art methods, and the Bilingual Evaluation Understudy(BLEU) scores are improved with our novel method.