Robust Question Answering Through Sub-part Alignment