Feature-aware Conditional GAN for Category Text Generation

Category text generation receives considerable attentions since it is beneficial for various natural language processing tasks. Recently, the generative adversarial network (GAN) has attained promising performance in generation, attributed to its training process. However, there are several issues GANs, including discreteness, instability, mode collapse, lack of diversity and controllability etc. To address these issues, this paper proposes a novel GAN framework, feature-aware conditional (FA-GAN), controllable category generation. In FA-GAN, generator sequence-to-sequence structure improving sentence diversity, which consists three encoders special encoder category-aware encoder, one relational-memory-core-based decoder with Gumbel SoftMax activation function. The discriminator an additional classification head. generate sentences specified categories, multi-class loss supplemented training. Comprehensive experiments have been conducted, results show that FA-GAN consistently outperforms 10 state-of-the-art approaches on 6 datasets. case study demonstrates synthetic generated by can match required categories aware features conditioned sentences, good readability, fluency, authenticity.