Quantifying Human-Perceived Answer Utility in Non-factoid Question Answering