Fund misuse detection in financial auditing using ensemble one-class classifiers with fine-tuned BERT representations

Despite advances in auditing research, most studies remain at the level of conceptual frameworks and provide limited technical solutions for tackling important and complex auditing problems. To address this gap, we focus on fund misuse detection in financial auditing and formulate it as a multi-class classification task, where the classes of interest constitute only a small fraction of the data and require simultaneous extraction and categorization of relevant text segments. We propose the Fine-Tuned Bidirectional Encoder Representations from Transformers with One-Class Ensemble (BERT-OCEn) model, which employs a task-specific fine-tuned Bidirectional Encoder Representations from Transformers (BERT) to generate domain-adapted text representations. Built upon these representations, BERT-OCEn incorporates a collection of one-class classifiers based on a newly designed Multi-Context Attentive One-Class Classifier (MC-AOCC) to capture class-specific characteristics in an interpretable manner. A three-layer fully connected neural network further integrates the outputs of these MC-AOCC classifiers to produce unified multi-class predictions. Experiments on two real-world datasets demonstrate that BERT-OCEn effectively identifies instances of fund misuse. These findings highlight the importance of technically grounded solutions for complex auditing tasks and provide auditors with practical tools to improve efficiency and support data-driven decision-making.