Boosting Explainability through Selective Rationalization in Pre-trained Language Models