Reward Augmented Maximum Likelihood for Neural Structured Prediction