Integrating textual semantics and financial indicator graphs: an interpretable deep learning approach for financial risk classification
Year: 2026
Type: article
Abstract: Financial risk classification requires the joint analysis of heterogeneous textual disclosures and structured accounting information, yet existing methods often model these modalities independently and provide limited evidence for individual predictions. To address these limitations, this study proposes Financial Graph-enhanced BERT for Risk Classification (FinGraph-BERT), a multimodal framework for classifying company-year financial reports into modified Altman Z \('\) -score-based proxy-risk categories. The model uses a domain-pretrained financial language encoder to capture contextual risk signals and constructs an intra-report financial indicator graph from reproducible accounting and ratio dependencies. A semantic–structural alignment mechanism reduces the representation gap between textual and graph features, followed by interaction-aware multimodal fusion for High-Risk, Medium-Risk, and Low-Risk classification. To reduce information leakage, reports are divided using a firm-grouped and class-stratified partition, ensuring that the same firm does not appear in different experimental subsets. Experiments on the reconstructed FinQA dataset show that FinGraph-BERT achieves an Accuracy of \(0.814\pm 0.008\) and a Macro-F1 score of \(0.796\pm 0.009\) over 30 independent runs. Compared with BERT+GAT, it improves Accuracy by 3.3 percentage points and Macro-F1 by 3.7 percentage points. Ablation and statistical analyses confirm the contributions of graph reasoning, semantic–structural alignment, and multimodal fusion. Integrated Gradients further provides token- and indicator-level explanations, while deletion and insertion tests show that highly attributed features influence predictions more strongly than randomly ranked features. These results demonstrate the effectiveness of FinGraph-BERT for interpretable report-level proxy-risk classification, rather than direct prediction of future bankruptcy or default events.
Source: Scientific Reports
Authors: Haorui Gu, Ailian Bian
Institution: Kookmin University
Language: English
Subfield: Accounting
Domain: Social Sciences
Open Access status: gold
Estimated APC paid: $2,690
OpenAlex ID: W7208834152
FWCI:
Cites:
Cited by:
Related to: