Mitigating Large Language Model Hallucinations via Autonomous Knowledge Graph-based Retrofitting