Mining Structures from Massive Texts by Exploring the Power of Pre-trained Language Models