Abstract
Generative adversarial networks (GANs) have great successes on synthesizingdata. However, the existing GANs restrict the discriminator to be a binaryclassifier, and thus limit their learning capacity for tasks that need tosynthesize output with rich structures such as natural language descriptions.In this paper, we propose a novel generative adversarial network, RankGAN, forgenerating high-quality language descriptions. Rather than training thediscriminator to learn and assign absolute binary predicate for individual datasample, the proposed RankGAN is able to analyze and rank a collection ofhuman-written and machine-written sentences by giving a reference group. Byviewing a set of data samples collectively and evaluating their quality throughrelative ranking scores, the discriminator is able to make better assessmentwhich in turn helps to learn a better generator. The proposed RankGAN isoptimized through the policy gradient technique. Experimental results onmultiple public datasets clearly demonstrate the effectiveness of the proposedapproach.