A growing number of scientific articles are retracted each year. Many of these articles continue to be cited and influence research and clinical decisions, posing serious risks. In this paper, we propose Pub-Guard-LLM, a large language model-based system for assessing retraction risks in biomedical articles. The system predicts retraction risks and provides textual explanations. We introduce an open-source benchmark, PubMed Retraction, containing over 12,000 real-world biomedical articles with metadata and retraction labels. Our results show that Pub-Guard-LLM outperforms various baselines. We further show that adding metadata, such as author, institution, and journal information, improves prediction performance. However, metadata can also introduce bias. We study metadata-related bias in retraction risk predictions using a framework for transparent bias detection. To reduce bias, we explore several methods, including retrieval-augmented generation, counterfactual fine-tuning, debate, and argumentation-based approaches. Our results reveal a trade-off between prediction performance and metadata bias. Pub-Guard-LLM provides a step towards reliable and fair retraction risk assessment for biomedical research.
翻译:暂无翻译