Typhoon: Towards an Effective Task-Specific Masking Strategy for Pre-trained Language Models