ChatGPT时代下基于深度自编码网络的企业内部审计智能预警——以往来款项审计为例

An intelligent risk warning model of enterprise internal audit based on deep autoencoder network in the ChatGPT era:Audit case of current accounts

以ChatGPT为代表的生成式人工智能技术被誉为第二次信息革命,其强大的数据深入分析能力为企业智能化内部审计提供了新思路.针对现有审计风险预警中传统机器学习泛化能力提升有限以及特征分析维度不足等问题,提出一种基于ChatGPT技术内核的深度自编码网络方法,来对往来款项这一重要会计活动中的风险做出事前判别.首先,根据影响因素从业务匹配、期限结构、减值损失、关联交易、单体统计和文本信息多个角度筛选提取审计特征;随后,考虑到风险样本的不平衡性以及业财指标在经营周期下的前后时变特性,基于无监督和深度学习思想,构建了以添加注意力机制双向长短期记忆(Bi-LSTM)作为神经网络的深度自编码器(DAE)预训练模型,并借鉴多任务学习思想,利用融合模型迁移的集成框架量化审计风险概率以保证预警的稳定性;最后,通过大数据技术采集企业往来业务和财务的真实数据对上述方法进行了多方面对比验证.试验结果表明,该方法有助于不同预警时间窗口下审计特征的高效精准挖掘,相较常用的监督学习和迭代聚类法能显著提升审计风险预警的精度和鲁棒性,同时能识别出导致风险产生的关键因素以快速定位审计疑点,为企业改善内部审计的质量和效率提供智能化决策支持.

The generative AI technology represented by ChatGPT,considered as the second information revolution,have transformed the depth of data analysis,offering new perspectives for intelligent internal audits in enterprises.In response to the limitations in the existing audit risk warnings,such as the limited improvement in the generalization capability of traditional machine learning and the insufficient feature analysis dimensions,we propose a method based on the core technology of ChatGPT—A deep autoencoder network.This method aims to pre-determine risks in the critical accounting activity of incoming funds.First,based on influencing factors,audit features are selected and extracted from various perspectives including business matching,term structure,impairment loss,related transactions,individual statistics,and text information.Subsequently,considering the imbalance of risk samples and the temporal charac-teristics of financial indicators over the operating cycle,an unsupervised and deep learning-based approach is employed.This involves constructing a deep autoencoder(DAE)pre-training model with the addition of an attention mechanism and employing bidirectional long short-term mem-ory(Bi-LSTM)as the network.Additionally,drawing from the concept of multi-task learning,an integrated framework with model transfer is utilized to quantify audit risk probabilities,en-suring the stability of warnings.Finally,real data from enterprise transactions and finances are collected by using big data technology for comprehensive comparative validation of the proposed method.Experimental results indicate that this method effectively and accurately extracts audit features under different warning time windows.In comparison to common practices like super-vised learning and iterative clustering,it significantly enhances the precision and robustness of audit risk warnings.Moreover,it also identifies key factors leading to risk,enabling quickly swift localization of audit suspicions.Our study can provide intelligent decision support for enterprises to improve the quality and efficiency of internal audit.