arXiv 2408.155332025-06-27cs.CLcs.AI

LRP4RAG：通过逐层相关性传播检测检索增强生成中的幻觉

LRP4RAG: Detecting Hallucinations in Retrieval-Augmented Generation via Layer-wise Relevance Propagation

Alibaba Cloud（阿里云）；School of Computer Sciences, Universiti Sains Malaysia（计算机科学学院，马来西亚国立大学）；Department of Service Science and Operations Management, School of Management, Zhejiang University（服务科学与运营管理系，浙江大学）；State Key Laboratory for Novel Software Technology, Nanjing University（新型软件技术国家重点实验室，南京大学）

Haichuan Hu, Congqing He, Xiaochen Xie, Quanjun Zhang

阅读 PDF

arXiv 页面 ↗HTML 原文 ↗

AI总结：

本文提出LRP4RAG方法，利用逐层相关性传播算法计算RAG生成器输入输出的相关性，经提取和重采样后输入分类器检测幻觉，首次将LRP用于RAG幻觉检测且优于现有基线。

摘要：

检索增强生成（RAG）已成为缓解大型语言模型（LLMs）幻觉的主要技术。然而，不完整的知识提取和不足的理解仍可能误导LLMs生成无关甚至相互矛盾的响应，这意味着幻觉在RAG中依然存在。本文提出LRP4RAG，一种基于逐层相关性传播（LRP）算法检测RAG中幻觉的方法。具体而言，我们首先利用LRP计算RAG生成器输入与输出之间的相关性。随后对该相关性矩阵进行进一步提取和重采样。处理后的相关性数据被输入多个分类器，以判定输出是否包含幻觉。据我们所知，这是首次将LRP用于检测RAG幻觉，大量实验表明LRP4RAG优于现有基线。

英文摘要：

Retrieval-Augmented Generation (RAG) has become a primary technique for mitigating hallucinations in large language models (LLMs). However, incomplete knowledge extraction and insufficient understanding can still mislead LLMs to produce irrelevant or even contradictory responses, which means hallucinations persist in RAG. In this paper, we propose LRP4RAG, a method based on the Layer-wise Relevance Propagation (LRP) algorithm for detecting hallucinations in RAG. Specifically, we first utilize LRP to compute the relevance between the input and output of the RAG generator. We then apply further extraction and resampling to the relevance matrix. The processed relevance data are input into multiple classifiers to determine whether the output contains hallucinations. To the best of our knowledge, this is the first time that LRP has been used for detecting RAG hallucinations, and extensive experiments demonstrate that LRP4RAG outperforms existing baselines.

机构、总结、中文摘要均由AI分析而来，可能存在错误，请以论文正文为准。