Towards Mitigating Hallucinations in Large Vision-Language Models by Refining Textual Embeddings