计算机科学
人工智能
理论计算机科学
任务(项目管理)
基于知识的系统
数据挖掘
算法
特征(语言学)
知识图
知识表示与推理
匹配(统计)
钥匙(锁)
度量(数据仓库)
出处
期刊:2020 International Conference on Computer Engineering and Application (ICCEA)
[Institute of Electrical and Electronics Engineers]
日期:2026-05-01
卷期号:: 699-703
标识
DOI:10.1109/iccea69962.2026.11743596
摘要
Retrieval-Augmented Generation (RAG) improves large language models by grounding responses in external evidence, but the retrieved context itself may still contain incorrect, outdated, or mutually contradictory information. As a result, retrieval relevance alone is insufficient to guarantee reliable generation. To address this issue, we propose KG-Verified RAG, a post-retrieval fact verification framework that uses a knowledge graph as an external factual anchor. Given a user query and its retrieved supporting documents, the framework extracts candidate factual triplets from the evidence, performs entity linking and relation linking to align textual claims with the graph, and assigns each claim one of three statuses: Verified, Conflicted, or Not Found. Because ambiguous entities and flexible relation expressions may lead to false conflicts, we further introduce a lightweight arbitration module for semantically difficult cases. The resulting verification report is then injected into the generation stage, enabling the language model to prioritize supported facts, down-weight conflicting statements, and handle uncertain evidence more explicitly. Experiments on open-domain question answering tasks and controlled factual-conflict settings show that the proposed method improves factual consistency, answer reliability, and robustness against misleading retrieved evidence.
科研通智能强力驱动
Strongly Powered by AbleSci AI