Exploring Supervised and Unsupervised Rewards in Machine Translation