SentiGAN: Generating Sentimental Texts via Mixture Adversarial Networks

TLDR

Generating texts with different sentiment labels is increasingly studied; GANs have shown promise but suffer from poor quality, lack of diversity, and mode collapse. The paper proposes SentiGAN, a framework with multiple generators and a multi‑class discriminator, to overcome these GAN limitations in sentiment text generation. Multiple generators are trained simultaneously to produce texts of different sentiment labels without supervision, and a penalty‑based objective forces each generator to generate diversified examples of a specific sentiment label. Experiments on four datasets show that SentiGAN consistently outperforms state‑of‑the‑art methods in sentiment accuracy and text quality, with each generator accurately focusing on its assigned sentiment label.

Abstract

Generating texts of different sentiment labels is getting more and more attention in the area of natural language generation. Recently, Generative Adversarial Net (GAN) has shown promising results in text generation. However, the texts generated by GAN usually suffer from the problems of poor quality, lack of diversity and mode collapse. In this paper, we propose a novel framework - SentiGAN, which has multiple generators and one multi-class discriminator, to address the above problems. In our framework, multiple generators are trained simultaneously, aiming at generating texts of different sentiment labels without supervision. We propose a penalty based objective in the generators to force each of them to generate diversified examples of a specific sentiment label. Moreover, the use of multiple generators and one multi-class discriminator can make each generator focus on generating its own examples of a specific sentiment label accurately. Experimental results on four datasets demonstrate that our model consistently outperforms several state-of-the-art text generation methods in the sentiment accuracy and quality of generated texts.