Constraint-Guided LLM-GNN Framework for Anomalous Journal Entry Detection in Auditing

Journal Entry Testing (JET) is a fundamental audit procedure to identify any potential misstatements, fraud, and management override of controls. Traditional rule-based JET methods suffer from high false positive rates and limited ability to detect complex anomaly patterns. Recent work has shown that large language models (LLMs) can serve as effective anomaly detectors for bookkeeping data, but LLM-only approaches lack explicit modeling of accounting constraints and structural relationships among entries. We propose the Constraint-Guided LLM-GNN (CG-LGN) framework, which integrates three complementary modules: (1) a heterogeneous graph neural network (GNN) that models structural relationships among journal entries, accounts, users, and temporal attributes; (2) an accounting constraint module encoding domain-specific rules such as debit-credit integrity, unusual account combinations, and period-end concentration; and (3) an LLM-based explanation generator that produces auditor-readable interpretations for flagged entries. Experiments on synthetic journal entry data with six injected anomaly types show that CG-LGN achieves a PR-AUC of 0.49 (a 0.20 absolute gain over the strongest single-module baseline) and reduces false positives per 1,000 entries by 47%. Ablation studies confirm that the GNN and constraint modules improve detection performance, while the LLM module improves explanation quality.