Article Details

Abstract

In the financial market, the timely and accurate identification of negative news is crucial for risk early warning and investment decision-making. Aiming at the limitations of traditional text classification models, such as the difficulty in capturing deep semantic features and weak generalization ability, this paper proposes an improved BERT model based on adversarial training (AT-BERT). The proposed model utilizes the pre-trained BERT model to extract dynamic word embeddings, introduces an attention mechanism to assign weights to key semantic features, and employs the Fast Gradient Method (FGM) adversarial training strategy to enhance model robustness. Experimental results show that the model achieves an average accuracy of 97.53%, an AUC value of 0.990 4, and an F1 score of 0.983 0 under 5-fold cross-validation. Compared with traditional machine learning models, classic Deep Learning models, and the baseline BERT model, the proposed model demonstrates improvements across all evaluation metrics, showing great application potential on public financial datasets.

Keywords

financial negative news; BERT; attention mechanism; adversarial training; financial text classification

References

1. [1] Cao Huanhuan, Xie Xingsheng Research and Application of Negative News Judgment Algorithm [J]. Small Microcomputer System, 2015, 36 (5): 1047-1051
2. [2] Qin Miao, Hu Erqin Identification and prediction of positive and negative emotions in public opinion of automobile companies [J]. Journal of Hubei University of Technology, 2022, 37 (2): 116-120
3. [3] Yu Fei, Li Jing, Liu Zhiguang, etc User demand identification method based on scenario analysis and negative comments [J]. Mechanical Design, 2025, 42 (8): 205-211
4. [4] Qin Xiao, Zhang Jinyong, Zheng Hongchun Text Level BERTGCN: A Text Classification Model Based on BERT Optimization [J]. Guangxi Science, 2025, 32 (6): 1212-1221
5. [5] Geng Qing, Sun Xujin, Yang Li, etc A Vector Data Semantic Classification Method Based on Improved BERT Model [J]. Journal of Spatiotemporal Information, 2025, 32 (5): 538-545
6. [6] Li Jun, Shao Jiarong Research on Predicting Emotional Characteristics of Financial News Based on Interpretable Large Language Model [J/OL]. Intelligent Computers and Applications, 2025:1-10 https://doi.org/10.20169/j.issn.2095-2163.25062602.
7. [7] Wan Yongzhi, Hu Huijun, Zhang Yaofeng Financial Text Sentiment Analysis Method Based on Multi level Knowledge Enhanced Large Model [J/OL]. Journal of Wuhan University (Science Edition), 2026 (4): 531-542
8. [8] Yang Xiaofeng Research on "Title Party" Recognition Method Based on NLP Technology [J]. Modern Information Technology, 2020, 4 (20): 124-127
9. [9] Yin Pengbo, Pan Weimin, Zhang Haijun, etc Research on Identifying Headline Party News Based on BERT BiGA Model [J]. Data Analysis and Knowledge Discovery, 2021, 5 (6): 126-134
10. [10] Zhang Yangshen, Zhou Weixiang, Zhang Yuyao, etc A negative news recognition method based on sentiment computing and hierarchical multi head attention mechanism [J]. Acta Sinica, 2020, 48 (9): 1720-1728
11. [11] Wu Tianlong, Shen Wenhui Stock market text classification model based on improved BERT Attention [J]. Journal of Yunnan University for Nationalities (Natural Science Edition), 2026, 35 (4): 531-537560
12. [12] Yang Zhuofeng, Li Yang, Li Deyu A semi supervised multi label classification method for financial events [J]. Data Collection and Processing, 2024, 39 (2): 385-394
13. [13] Long Jie, Lu Chong, Han Liying, etc Hidden Emotion Analysis Model Driven by Multi Head Self Attention under Adversarial Training [J]. Computer Engineering and Design, 2026, 47 (3): 808-816
14. [14] Ji Xunsheng, Cai Zhiwan, Hu Kai Research on Text Classification Method Integrating Adversarial Training and Improved BERT [J]. Computer and Digital Engineering, 2025, 53 (8): 2200-22042245
15. [15] Sun Hao, Pu Yifei Knowledge enhanced BERT text classification model based on dynamic hierarchical adversarial collaborative optimization [J]. Journal of Sichuan University (Natural Science Edition), 2026, 63 (3): 586-596

About the authors

Huang Jianghe (2005-), male, Han ethnicity, from Nanyang, Henan, undergraduate student, research interests include digital finance; Corresponding author: Wang Jingsai (1997-), male, Han ethnicity, from Nanyang, Henan, teaching assistant, master's student, research interests include data science and statistical machine learning and financial mathematics.