A Simple but Tough-to-Beat Data Augmentation Approach for Natural Language Understanding and Generation