破產預測研究主要關注企業是否可能在未來的一段時間內經歷破產。企業破產不僅對企業本身造成嚴重的損失，還會對社會經濟造成重大影響，如失業率上升、負債擴大等。儘管近年已有一些研究使用人工神經網路模型提取文本特徵以預測企業破產，但人工神經網路模型俗稱黑盒子，其得出的結果缺乏可解釋性。本研究著重於研究美國上市公司破產預測之模型，我們採用分層注意力網路 (Hierarchical Attention Network , HAN) 從10-K年度報告的管理討論與分析 (Management′s Discussion and Analysis , MD&A) 部分提取文本特徵。分層的網路可以反映文本的結構，在單詞和句子級別加入注意力機制使構建文本表示的過程中，依據重要性來區分內容以提供模型的可解釋性。除了使用分層注意力網路建構文本表示，亦運用Bi-LSTM 學習MD&A的年變化量，再結合自S&P Capital IQ 資料庫的公司財務資料做為預測模型的輸入。本研究在實驗中運用貝氏神經網路和人工神經網路對企業破產進行預測，並比較兩種分類模型在不同輸入下的預測性能，但實驗結果並不如預期，推測可能的原因是因為資料噪聲大與模型參數調配不佳。;Bankruptcy prediction research focuses on whether an enterprise is likely to experience bankruptcy in the future. Enterprise bankruptcy not only causes serious losses to the enterprise itself, but also has a major impact on the social economy, such as rising unemployment and expanding debts. Although there have been some studies in recent years using Artificial Neural Network models to extract text features to predict enterprise bankruptcy, Artificial Neural Network models are commonly known as black boxes, and the results obtained lack interpretability. This study focuses on the model of bankruptcy prediction of listed companies in the United States. We use Hierarchical Attention Network (HAN) to extract text feature from the Management′s Discussion and Analysis (MD&A) part of the 10-K annual report. The layered network can reflect the structure of the text, and the Attention Mechanism is added at the word and sentence levels to make the process of constructing the text representation, distinguishing the content according to the importance to provide the interpretability of the model. In addition to using Hierarchical Attention Networks to construct text representations, Bi-LSTM is also used to learn the annual change in MD&A, and then combined with company financial data from the S&P Capital IQ database as the input of the predicting model. In this study, Bayesian Neural Network and Artificial Neural Network are used to predict the bankruptcy of enterprises in the experiment, and the prediction performance of the two classification models is compared under different inputs, but the experimental results are not as expected. The possible reason is that the large data noise and poor adjustment of model parameters.
