Training language GANs from Scratch
This addresses the challenge of unstable GAN training in natural language processing, offering a potential alternative to pre-training methods, though it appears incremental as it combines existing techniques.
The paper tackled the problem of training Generative Adversarial Networks (GANs) for natural language generation from scratch, without relying on maximum likelihood pre-training, and achieved comparable performance to traditional methods on EMNLP2017 News and WikiText-103 corpora in terms of quality and diversity metrics.
Generative Adversarial Networks (GANs) enjoy great success at image generation, but have proven difficult to train in the domain of natural language. Challenges with gradient estimation, optimization instability, and mode collapse have lead practitioners to resort to maximum likelihood pre-training, followed by small amounts of adversarial fine-tuning. The benefits of GAN fine-tuning for language generation are unclear, as the resulting models produce comparable or worse samples than traditional language models. We show it is in fact possible to train a language GAN from scratch -- without maximum likelihood pre-training. We combine existing techniques such as large batch sizes, dense rewards and discriminator regularization to stabilize and improve language GANs. The resulting model, ScratchGAN, performs comparably to maximum likelihood training on EMNLP2017 News and WikiText-103 corpora according to quality and diversity metrics.