Making Pre-trained Language Models Better Few-shot Learners