Exploiting Semantic Reconstruction to Mitigate Hallucinations in Vision-Language Models