Toward learning better metrics for sequence generation training with policy gradient

Designing a metric manually for unsupervised sequence generation tasks, such as text generation, is essentially difficult. In a such situation, learning a metric of a sequence from data is one possible solution. The previous study, SeqGAN, proposed the framework for unsupervised sequence generation, in which a metric is learned from data, and a generator is optimized with regard to the learned metric with policy gradient, inspired by generative adversarial nets (GANs) and reinforcement learning. In this paper, we make two proposals to learn better metric than SeqGAN's: partial reward function and expert-based reward function training. The partial reward function is a reward function for a partial sequence of a certain length. SeqGAN employs a reward function for completed sequence only. By combining long-scale and short-scale partial reward functions, we expect a learned metric to be able to evaluate a partial correctness as well as a coherence of a sequence, as a whole. In expert-based reward function training, a reward function is trained to discriminate between an expert (or true) sequence and a fake sequence that is produced by editing an expert sequence. Expert-based reward function training is not a kind of GAN frameworks. This makes the optimization of the generator easier. We examine the effect of the partial reward function and expert-based reward function training on synthetic data and real text data, and show improvements over SeqGAN and the model trained with MLE. Specifically, whereas SeqGAN gains 0.42 improvement of NLL over MLE on synthetic data, our best model gains 3.02 improvement, and whereas SeqGAN gains 0.029 improvement of BLEU over MLE, our best model gains 0.250 improvement.