Mitigating Entity-Level Hallucination in Large Language Models