Beyond MLE: Convex Learning for Text Generation