Mitigating Hallucinations in Medical Information Extraction

This technical report addresses the hallucination challenges faced by large language models (LLMs) in the context of Medical Information Extraction (MIE) tasks. It introduces a novel approach called ALternate Contrastive Decoding (ALCD) which redefines MIE tasks as an identify-and-classify process. The report outlines how the identification and classification functions of LLMs can be separated by selectively masking the optimization of tokens during fine-tuning. During inference, the method alternates between contrasting output distributions from sub-task models to enhance the capabilities of identification and classification while minimizing the influence of other inherent abilities in LLMs. The report also proposes an adaptive constraint strategy to adjust the scale and scope of contrastive tokens effectively. Comprehensive experiments conducted on various medical information extraction tasks demonstrate that ALCD significantly improves the resolution of hallucination issues compared to conventional decoding methods, thereby enhancing the reliability of LLMs in medical applications.