Learning from Context or Names? An Empirical Study on Neural Relation Extraction