Making Document-Level Information Extraction Right for the Right Reasons