| 基于MD&A文本隐式情感的上市公司财务困境预测 |
| Research on financial distress prediction of listed companies based on implicit sentiment of MD&A |
| 摘要点击 80 全文点击 0 投稿时间：2024-11-18 修订日期：2025-08-09 |
| 查看/发表评论 下载PDF阅读器 |
| 中文关键词 财务困境预测, 隐式情感分析, 深度学习, 图神经网络 |
| 英文关键词 Financial distress prediction, Implicit sentiment analysis, Deep learning, Graph Neural Network |
| 基金项目 |
| 投稿方向 |
| 作者 | 单位 | 邮编 | | 闫志华 | 山西财经大学 | 030006 | | 唐锡晋* | 中国科学院数学与系统科学研究院 | 100190 | | 黄晓辉 | 中国科学院数学与系统科学研究院 | | | 闫绪娴 | 山西财经大学 | |
|
| 中文摘要 |
| 管理层讨论与分析(Management discussion and analysis, MD&A)文本中包含大量的隐式情感信息, 但是无法使用情感词直接判断情感类别. 本文提出融合句子结构特征、上下文语义特征和外部金融情感信息的金融文本隐式情感识别模型: 使用依存句法分析将MD&A文本中的语句表示为依存句法树, 基于图注意力机制获得关键语句的结构特征; 使用Bi-LSTM和注意力机制获取关键上下文语义信息; 使用FinBERT金融预训练模型引入金融领域知识和情感信息. 在此基础上, 本文将MD&A文本的隐式情感特征应用于企业财务预测任务, 提升机器学习算法的预测能力. 实验表明, 该模型可以提升MD&A文本的隐式情感特征识别能力, 隐式情感特征可以大幅度提升上市公司财务困境预测的效果. |
| 英文摘要 |
| Management discussion and analysis (MD&A) text contains a large amount of implicit sentiment information, but it is unable to use sentiment words to directly determine sentiment polarity. In this paper, we propose a financial text implicit sentiment detection model (SC-ISA) that fuses sentence structural features, contextual semantic features, and external financial sentiment information: use dependent syntactic analysis to represent the sentence as a tree structure, and obtain key sentence structural features based on the graph attention mechanism; employ Bi-LSTM and the attention mechanism to obtain key contextual semantic information; utilize FinBERT financial pretraining model to introduce financial domain knowledge and sentiment information. On this basis, the implicit sentiment features are applied to the corporate finance prediction task to improve the prediction ability of the machine learning algorithms. Experiments show that the SC-ISA model can improve the performance of implicit sentiment detection based on MD&A and the implicit sentiment features can substantially improve the accuracy of the prediction of financial distress of listed companies. |
| 关闭 |
|
|
|
|
|
