Towards Improving Selective Prediction Ability of NLP Systems