Text this: Reducing hallucinations of large language models via hierarchical semantic piece