Identifying the Source of Vulnerability in Explanation Discrepancy: A Case Study in Neural Text Classification