大语言模型在国家审计中的应用探索  

Research on the Application of Large Language Models in National Auditing

作　　者：杨麟 张宪礼 于微伟 Yang Lin;Zhang Xianli;Yu Weiwei

机构地区：[1]江苏省审计厅,210024

出　　处：《审计研究》2024年第6期22-29,共8页Auditing Research

摘　　要：大语言模型具备强大的自然语言处理能力,在官方基准测试中展示出了卓越性能,目前国家审计面临非结构化数据体量大、手段少、效率低等挑战,大语言模型为审计智能化提供了新的方法路径。本文分析了大语言模型的技术基础和发展现状,从基于通识能力创建审计知识库、基于推理能力开发审计智能工具等角度讨论了大语言模型在国家审计领域的应用前景,剖析在行业实施组织和应用落地等方面的难点,提出了技术可行性和实现路径。经论证,利用无加工标注的审计行业数据和无监督训练可以影响大语言模型权重,提升通用大语言模型在审计行业领域的智能水平,大语言模型具备推动审计智能化转型的潜力。本文还给出了构建政务语料库、微调训练和私有化部署以及国家审计应用等应用示例。Large language model has strong natural language processing capabilities and has demonstrated excellent performance in official benchmark test.At present,national audit faces the challenges of large volume of unstructured data,few means and low efficiency.Large language model provides a new method and path for the intelligent audit.This paper briefly introduces the technical basis of large language models and the development status at home and abroad,discusses in detail the application prospect of building audit knowledge base based on general ability and developing audit intelligence tool based on reasoning abilities,deeply analyzes the organizational difficulties in industry implementation and the technical difficulties in application implementation,and puts forward the technical feasibility and implementation path.It has been demonstrated that,using unmarked audit industry data,unsupervised training can affect the weight of large language models,improve the intelligence level of general large language models in the audit industry,and that large language models have the potential to promote the intelligent transformation of auditing.From the practical perspective,this paper provides examples such as constructing government corpus,fine-tuning training and privatization deployment,and national audit application.

关 键 词：大语言模型 国家审计 审计应用 

分 类 号：TP391.1[自动化与计算机技术—计算机应用技术] F239.44[自动化与计算机技术—计算机科学与技术]