Generative Adversarial Training Can Improve Neural Language Models (2211.09728v1)

Abstract: While deep learning in the form of recurrent neural networks (RNNs) has caused a significant improvement in neural LLMing, the fact that they are extremely prone to overfitting is still a mainly unresolved issue. In this paper we propose a regularization method based on generative adversarial networks (GANs) and adversarial training (AT), that can prevent overfitting in neural LLMs. Unlike common adversarial training methods such as the fast gradient sign method (FGSM) that require a second back-propagation through time, and therefore effectively require at least twice the amount of time for regular training, the overhead of our method does not exceed more than 20% of the training of the baselines.