財務報表欺詐是一種白領犯罪會造成嚴重後果，包括投資者和債權人的財務損失、公司聲譽的損害以及個人和公司法律和監管後果。檢測欺詐的傳統方法非常耗時，並且需要大量的人工操作。本研究提出一種財務報表欺詐檢測系統之架構，其中包含四個方向的考量分別是採用分層注意力網路模型(HAN)從年度報告中管理討論與分析(MD&A)提取文本特徵以獲取管理階層對公司營運看法、利用Bi-LSTM及相似度分析提取MD&A的時間變化量以及利用該公司的財務報表理解該公司的營運狀況。與過去典型黑盒子的類神經網路不同，本研究利用深度模型的預測能力及注意力模型結合以得到擁有可解釋的模型，以及利用貝氏類神經網路(BNN)量化該模型的不確定性，還有藉由HAN提高提取文本特徵的準確度及效率。本研究通過提高預測分析中的準確度、效率以及將可解釋性和不確定性加入模型中為文獻做出貢獻，並為監管機構提供一種藉由檢查大量公開文本及資料以監控並預測財務報表欺詐的方法。;Financial statement fraud is a type of white-collar crime that can have serious consequences, including financial losses for investors and creditors, damage to company reputation, and legal and regulatory consequences for individuals and companies. Traditional methods for detecting fraud are time-consuming and require extensive manual operations. This study proposes an architecture for a financial statement fraud detection system that incorporates four directions of consideration. These include using a Hierarchical Attention Network (HAN) model to extract textual features from Management′s Discussion and Analysis (MD&A) in annual reports to obtain management′s perspectives on company operations, utilizing Bi-LSTM and similarity analysis to extract temporal changes in MD&A, and using the company′s financial statements to understand its operational status. Unlike typical black-box neural networks used in the past, this research utilizes the predictive ability of deep models and combines attention models to obtain an interpretable model. It also quantifies the model′s uncertainty using Bayesian neural networks (BNN), and enhances the accuracy and efficiency of extracting textual features by leveraging HAN. This study contributes to the literature by improving accuracy, efficiency, interpretability, and incorporating uncertainty in predictive analytics, and provides a method for regulatory agencies to monitor and predict financial statement fraud by examining a large volume of public text and data.
