Alleviating Hallucinations in Large Language Models with Scepticism Modeling