Can NLP Models Correctly Reason Over Contexts that Break the Common Assumptions?