Towards Mitigating Hallucination in Large Language Models via Self-Reflection