Coarse-to-Fine Highlighting: Reducing Knowledge Hallucination in Large Language Models