KADR-LLM：基于深度检索推理的大语言模型辅助档案开放审核方法

打开文本图片集

中图分类号：TB9 文献标志码：A文章编号：1674-5124（2025）07-0009-10

Abstract：This paper addresses limitations in traditional archive declassification review systems， including low efficiency， excessive subjectivity，and inadequate semantic analysis.A KADR-LLM-based intelligent auditing framework was developed，integrating Dense Passage Retrieval （DPR） capabilities with Knowledge Augmented Reasoning Process （KARP） mechanisms to establish a three-stage "retrieval-reasoningverification" paradigm. Key innovations include： A dual-channel text preprocessng method optimizing semantic representation through paragraph truncation based on document spatial structures; A rule-driven dynamic reasoning system combining sensitive term matching with retrieval-augmented generation; A keywordguided progressve auditing strategy enabling interpretable decision-making from surface feature extraction to logical chain validation. Evaluations on OParchives datasets showed KADR-LLM achieved 79.98% accuracy in zero-shot conditions and 82.34% in few-shot scenarios， surpassing baseline models by 4.31% while demonstrating superior semantic generalization capability.

Keywords： archival opening review; large language model; dense retrieval; reasoning prompt

0 引言

在数字时代，档案开放共享对社会信息化和知识创新意义重大。（剩余15350字）