Analyzing and Mitigating Object Hallucination in Large Vision-Language Models