摘要
DeepSeek凭借大幅下降的算力成本和强大的思维链推理能力,为各行各业应用人工智能大模型带来了新的突破和可能。本文分析了审计机关获取数据情况及其价值,剖析直接应用大模型处理结构化数据存在的输出篡改、效果波动和人工冗余三大技术瓶颈,提出“规则引擎+智能引擎”双层协同架构,即通过规则确定性逻辑与大模型语义推理的互补性协同,实现自然语言驱动的审计模式,为审计数字化转型提供了可扩展、高可控的技术路径,进一步挖掘释放结构化数据价值。通过该架构,在财政支出智能分类场景中,系统实现三级分类动态构建与摘要语义解析,破解了科目颗粒度不足与主观错配难题;在违规支出挖掘场景中,依托大模型的多维度协同研判与跨记录关联分析,有效识别科目嫁接、拆分报销等隐蔽违规行为。
DeepSeek has enabled new breakthroughs and possibilities for applying AI large language models(LLM)across industries by significantly reducing computational costs and leveraging powerful chain-of-thought reasoning.This paper analyzes data accessibility in audit institutions and its value,identifying three technical bottlenecks in directly applying LLM to structured data:output tampering,performance instability,and excessive manual intervention.To address these challenges,we propose a dual-layer collaborative architecture of"Rules Engine+Intelligence Engine".This framework synergizes the deterministic logic of rule engines with the semantic reasoning of large models,establishing a natural language-driven auditing paradigm.It provides a scalable,highly controllable technical pathway for audit digital transformation while further unlocking the value of structured data.In public expenditure intelligent classification,the architecture dynamically constructs three-tier categorization and performs semantic summarization,resolving issues of insufficient account granularity and subjective mismatches.In fraudulent expenditure detection,the architecture leverages large models'multi-dimensional collaborative analysis and cross-record correlation,effectively identifying concealed violations such as account grafting and split reimbursement.
出处
《审计研究》
北大核心
2025年第5期27-36,共10页
Auditing Research
关键词
大模型
双引擎架构
思维链推理
提示词工程
财政支出审计
large language models
dual-engine architecture
chain-of-thought reasoning
prompt engineering
fiscal expenditure audit
