Self-training with Few-shot Rationalization