Large Language Model Framework for Multi-Document Financial Anomaly Detection in Intelligent Auditing via Semantic Mapping and Risk Reasoning

Abstract

This study addresses the requirements of multi-document financial anomaly detection in intelligent auditing by developing a deep language model framework that integrates a semantic mapping mechanism, cross voucher consistency modeling, and a risk reasoning function, allowing the system to handle structural heterogeneity, dispersed information, and cross document chain distribution in financial texts. The method first encodes amount fields, business elements, and voucher structures through a unified semantic representation layer to achieve standardized modeling of multi source data. It then applies a cross document consistency measurement strategy to align the semantic associations of business chains and identify potential conflicts and abnormal logic. The risk reasoning module combines local evidence with global chain information to generate anomaly judgments consistent with audit logic. A comprehensive sensitivity evaluation is conducted across multiple dimensions, including risk threshold variation, text noise, context window expansion, computing environment differences, and changes in transaction density and business complexity. The results show that the framework maintains stable semantic integration across diverse document types, effectively captures key anomalies along cross structural chains, and demonstrates strong interpretability and adaptability under varied audit conditions, providing a scalable technical paradigm for automated risk identification in complex financial texts.