Improving Conditional Sequence Generative Adversarial Networks by Stepwise Evaluation

Yi-Lin Tuan, Hung-yi Lee

Abstract: Conditional sequence generation is a widely researched topic. One of the most important tasks is dialogue generation, which is composed of input-output pairs with the one-to-many property. Given the recent success of generative adversarial networks (GANs), GANs have been used for sequence generation. However, there is still limited work of its application on conditional sequence generation. We investigate the influence of GAN on conditional sequence generation with three artificial grammars and dialogue generation. Moreover, we propose stepwise GAN (StepGAN) for conditional sequence generation, which predicts the reward at each time-step. StepGAN can be seen as the general version of SeqGAN. It estimates the expected returns predicted by Monte-Carlo Search in SeqGAN, but it has a lower computational cost than Monte-Carlo Search. Experimental results show that stepwise GAN can outperform other state-of-the-art algorithms in most tasks.

Keywords: conditional sequence generation, generative adversarial network, REINFORCE, dialogue generation

Withdrawal: Confirmed