Training language GANs from Scratch

By Cyprien de Masson d'Autume, Mihaela Rosca, Jack W. Rae, Shakir Mohamed. 2019

91 citations recorded by the source corpus.

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.

Subfields: Artificial Intelligence, Computer Vision and Pattern Recognition, Natural Language Processing, Generative Models

How researchers see this paper

The researchers whose trusted networks rank it highest, strongest advocate first. Every researcher gets a different order.

In Artificial Intelligence

In Computer Vision and Pattern Recognition

In Natural Language Processing

In Generative Models

Citations in the corpus

ColdGANs: Taming Language GANs with Cautious Sampling Strategies (2020, methodology): “To our knowledge, this is the ﬁrst time that MLE falls short [5, 23] w.r.t. GAN-based approaches for this task.”
Red Teaming Language Models with Language Models (2022, background): “Another promising direction is to jointly train the red LM and target LM, similar to Generative Adversarial Networks (Goodfellow et al., 2014; d’Autume et al., 2019).”
Language GANs Falling Short (2018, methodology): “Our results are further validated in independent follow-up works (d’Autume et al., 2019; Alihosseini et al., 2019)where temperature sweeps show MLE outperforming GANs.”
Residual Energy-Based Models for Text (2021, background): “Generative Adversarial Networks (Goodfellow et al., 2014; Yu et al., 2017; de Masson d’Autume et al., 2019; Scialom et al., 2020; Caccia et al., 2020) also relate to EBMs, except that in GANs the discriminator is only used at training time, and the goal is to use it to help improve the generator.”
Step-unrolled Denoising Autoencoders for Text Generation (2021, methodology): “We also compare to several other baselines reported in (d’Autume et al., 2019; Caccia et al., 2020): Se-qGAN (Yu et al., 2017), LeakGAN (Guo et al., 2017), RankGAN (Lin et al., 2017), MaliGAN (Che et al., 2017) and autoregressive (AR) language models.”
Adding A Filter Based on The Discriminator to Improve Unconditional Text Generation (2020, methodology): “Unconditional text generation, which generates novel, reasonable and meaningful sentences, is a stepping stone for the above tasks, thus becoming a hot topic recently [30, 10, 7].”
A Review on Generative Adversarial Networks: Algorithms, Theory, and Applications (2020, background): “ScratchGAN [389] is proposed to train a language GAN from scratch – without maximum likelihood pretraining.”
A Discriminator Improves Unconditional Text Generation without Updating the Generator (2020, unknown): “Although (d’Autume et al., 2019) trains a language GAN from scratch, its performance is only comparable with LM.”