A hallucination detection and mitigation framework for faithful text summarization using LLMs

Recent advancements in large language models (LLMs) have significantly propelled the field of automatic text summarization. Nevertheless, the domain continues to face substantial challenges, particularly the issue of hallucination where summaries contain information not present in the source text. Such problems not only compromise the factual accuracy of summaries but also lead to diminished user satisfaction. Existing methods exhibit limitations in effectively detecting and mitigating hallucinations, often lacking transparency in their underlying mechanisms. This paper introduces a hallucination detection and mitigation framework that employs a Question-Answer Generation, Sorting, and Evaluation (Q-S-E) methodology to enable the quantitative detection of hallucinations in summaries. Leveraging LLMs, the framework incorporates an iterative hallucination resolution mechanism, which enhances the transparency of the modification process and improves the faithfulness of text summarization. Experimental results on three benchmark datasets CNN/Daily Mail, PubMed, and ArXiv demonstrate that our approach markedly improves the factual consistency of summaries while preserving their informational completeness.