Identifying Semantically Difficult Samples to Improve Text Classification