Diversity-Promoting GAN: A Cross-Entropy Based Generative Adversarial Network for Diversified Text Generation

TLDR

Existing text generation methods tend to produce repeated and “boring” expressions. To tackle this problem, we propose a new text generation model, called Diversity‑Promoting Generative Adversarial Network (DP‑GAN). DP‑GAN assigns low reward to repeated text and high reward to novel, fluent text, and employs a language‑model discriminator that better distinguishes novelty without saturation. Experiments on review and dialogue generation show DP‑GAN generates substantially more diverse and informative text than existing baselines.

Abstract

Existing text generation methods tend to produce repeated and "boring" expressions. To tackle this problem, we propose a new text generation model, called Diversity-Promoting Generative Adversarial Network (DP-GAN). The proposed model assigns low reward for repeatedly generated text and high reward for "novel" and fluent text, encouraging the generator to produce diverse and informative text. Moreover, we propose a novel language-model based discriminator, which can better distinguish novel text from repeated text without the saturation problem compared with existing classifier-based discriminators. The experimental results on review generation and dialogue generation tasks demonstrate that our model can generate substantially more diverse and informative text than existing baselines.