TaylorGAN: Neighbor-Augmented Policy Update for Sample-Efficient Natural Language Generation

TL;DR

TaylorGAN introduces a neighbor-augmented policy update method using Taylor expansion to improve sample efficiency and stability in natural language generation, enabling training from scratch without pre-training and outperforming existing GAN-based models.

Contribution

The paper proposes TaylorGAN, a novel gradient estimation technique that enhances sample efficiency in GAN-based NLG by incorporating off-policy updates and Taylor expansion.

Findings

Outperforms existing GAN-based NLG methods on multiple metrics.

Can be trained from scratch without maximum likelihood pre-training.

Requires smaller batch sizes for effective training.

Abstract

Score function-based natural language generation (NLG) approaches such as REINFORCE, in general, suffer from low sample efficiency and training instability problems. This is mainly due to the non-differentiable nature of the discrete space sampling and thus these methods have to treat the discriminator as a black box and ignore the gradient information. To improve the sample efficiency and reduce the variance of REINFORCE, we propose a novel approach, TaylorGAN, which augments the gradient estimation by off-policy update and the first-order Taylor expansion. This approach enables us to train NLG models from scratch with smaller batch size -- without maximum likelihood pre-training, and outperforms existing GAN-based methods on multiple metrics of quality and diversity. The source code and data are available at https://github.com/MiuLab/TaylorGAN