面向检索增强式私有问答大模型的幻觉检测方法OA

The existence of Large Language Models'(LLMs)hallucination phenomenon seriously restricts its implementation in practical sce-narios,existing hallucination detection methods were singular,and there was a lack of hallucination detection work under the paradigm of Re-trieval-Augmented Generation(RAG).To address these issues,a hallucination detection method was proposed for private Question-Answer-ing(Q&A)LLMs under the paradigm of RAG.This method integrated two types of hallucination detection methods,one based on uncertainty measure indexes and the other on LLMs'automatic evaluation,which effectively harnessed the hallucination features in the LLMs'generation process and the LLMs'automatic evaluation capability.To enhance the automatic assessment of open-source LLMs,a method of instruction fine-tuning was applied to distill the automatic evaluation ability from closed-source LLMs to open-source ones.Furthermore,a hallucination assessment dataset under the RAG's paradigm was constructed to verify the effectiveness of the proposed method.Experimental results on the above dataset indicate that the proposed integrated method achieved the highest AUC-ROC value,significantly increasing by 11.1％and 4.3％compared to the two baseline methods of uncertainty measure index and LLMs'automatic evaluation,respectively,proving the complementary nature of the two baseline methods.Moreover,the instruction fine-tuned open-source LLM exhibits an 18.6％increase in the AUC-ROC val-ue,with a significant improvement in automatic evaluation capability,validating the effectiveness of the instruction fine-tuning method in en-hancing the automatic evaluation ability of open-source LLMs.