End-to-End Differentiable GANs for Text Generation

Despite being widely used, text generation models trained with maximum likelihood estimation (MLE) suffer from known limitations. Due to a mismatch between training and inference, they suffer from exposure bias. Generative adversarial networks (GANs), on the other hand, by leveraging a discriminator, can mitigate these limitations. However, discrete nature of text makes the model non-differentiable hindering training. To deal with this issue, the approaches proposed so far, using reinforcement learning or softmax approximations are unstable and have been shown to underperform MLE. In this work, we consider an alternative setup where we represent each word by a pretrained vector. We modify the generator to output a sequence of such word vectors and feed them directly to the discriminator making the training process differentiable. Through experiments on unconditional text generation with Wasserstein GANs, we find that while this approach, without any pretraining is more stable while training and outperforms other GAN based approaches, it still falls behind MLE. We posit that this gap is due to autoregressive nature and architectural requirements for text generation as well as a fundamental difference between the definition of Wasserstein distance in image and text domains.