Annotator Rationales for Labeling Tasks in Crowdsourcing