Near-Negative Distinction: Giving a Second Life to Human Evaluation Datasets