Improving the Faithfulness of Attention-based Explanations with Task-specific Information for Text Classification