Corporate credit risk prediction models often encounter persistent barriers when attempting to identify high-risk commercial entities from numerical accounting records alone. Standard predictive architectures typically process structured financial statements, operational indicators, and solvency metrics in isolation, leaving the complex economic semantics embedded across interconnected variables largely implicit. To capture these latent relationships without destabilizing model training, a research team developed a cross-modal learning architecture designated as SMCCL-Net. The framework introduces a coordinated mechanism that translates raw accounting records into evaluated semantic views, linking the reliability of generated textual descriptions directly to downstream minority-class contrastive optimization. Across comprehensive evaluations conducted on two historical corporate bankruptcy benchmarks, the method achieved higher precision-recall area under the curve and geometric mean scores compared to competitive tabular and deep learning baselines, demonstrating that structured financial signals can be effectively augmented when textual generation is coupled with rigorous consistency filtering.
The study specifically targets the prediction of corporate bankruptcy outcomes, which represent severe credit events that are inherently rare within market populations. In corporate financial datasets, bankrupt enterprises constitute a small fraction of observed entities, causing traditional machine learning and standard deep neural networks to tilt their decision boundaries toward the normal majority class. While classical techniques such as logistic regression, random forests, and gradient boosting algorithms provide baseline interpretability, they rely heavily on manual feature engineering and struggle to articulate high-level economic context from nonlinear variable combinations. Conversely, deep tabular models and graph neural networks frequently optimize representations strictly within numerical feature spaces or depend heavily on the completeness of enterprise relational networks. Introducing external textual descriptions through large language models offers a potential mechanism to supply explicit economic semantics, yet unconstrained generative text introduces noise, hallucination, and representational divergence that can undermine cross-modal alignment.
To overcome these limitations, SMCCL-Net implements a structured two-stage pipeline consisting of a semantic quality generation optimization module and a class-aware cross-modal contrastive learning module. Instead of feeding unverified language descriptions directly into a classification network, the system generates candidate textual analyses using three distinct prompt templates grounded solely in reported financial indicators: key-indicator prompts, financial expert perspectives, and comprehensive interpretations. Crucially, these prompts restrict candidate generation to current-period accounting variables such as leverage, solvency, profitability, and cash flow, strictly prohibiting future outcome information, true bankruptcy labels, and explicit categorical risk assertions like default or bankruptcy. The framework deterministically transforms the structured financial record into a standardized indicator profile following a fixed template, which then serves as a ground-truth reference for semantic validation.
Candidate language outputs undergo multi-metric semantic quality filtering before any textual representation is paired with tabular features. The evaluation calculates an integrated quality score by normalizing and weighting four complementary consistency measures: cosine similarity between tabular and textual embeddings, Pearson correlation of feature-text vectors, an evaluator-based GPTScore assessing financial consistency and completeness, and lexical overlap measured via BLEU. In the established experimental configuration, the weights are set to 0.30 for cosine similarity, 0.20 for Pearson correlation, 0.35 for GPTScore, and 0.15 for BLEU. This weighting intentionally limits the influence of surface word matching while prioritizing relational agreement and embedding coherence. The candidate description that attains the highest cumulative quality score is designated as the positive textual view for that specific enterprise record, ensuring that noisy, redundant, or divergent textual generations are filtered out prior to cross-modal projection.
Once the validated textual description is selected, the tabular features and text are projected into a shared 128-dimensional latent space using a three-layer multilayer perceptron and a Sentence-BERT encoder, respectively. A bidirectional InfoNCE loss objective aligns the representations of corresponding tabular-text pairs from identical enterprises while treating pairs from distinct enterprises as negative samples. To resolve class imbalance at the representation level rather than merely adjusting classification decision thresholds, the second module incorporates an inverse-frequency category weighting scheme, a two-level hierarchy-aware contrastive loss, and an adaptive hard sample mining schedule. Category weights inversely scale with class sample frequencies, amplifying the optimization contribution of the sparse bankruptcy cases during representation training.
The hierarchy-aware learning structure captures fine-grained corporate risk profiles by partitioning the coarse normal and high-risk classes into finer subclasses. Using standardized training features, K-means clustering is conducted independently within each primary class, with silhouette analysis identifying three stable subclasses per group, yielding six subclasses in total. In the category-aware contrastive objective, pairs belonging to the same coarse class and subclass are weighted as strong positives with a factor of 1.0, while pairs sharing the coarse class but residing in different subclasses are weighted as weak positives with a factor of 0.5. Meanwhile, sample difficulty is measured dynamically by calculating the cross-modal discrepancy between normalized tabular and textual embeddings. The proportion of hard samples mined per batch rises linearly from 10 percent to 30 percent across the first 60 training epochs before stabilizing, preventing anomalous or difficult samples from destabilizing early latent space geometry.
Following cross-modal alignment, the network fuses the tabular vector, the textual vector, and their absolute difference vector into a unified representation, which is processed by a linear classification head optimized via a composite joint loss function. The total loss combines standard cross-entropy classification, bidirectional cross-modal alignment weighted at 0.5, and category-aware contrastive loss weighted at 0.8. Empirically, the entire training pipeline was executed over 100 epochs using the AdamW optimizer with a learning rate of 1e-4 and a batch size of 128 across five independent random seeds. By conditioning the construction and optimization weighting of cross-modal pairs directly on semantic selection scores and subclass assignments, the framework ensures that text quality filtering directly regulates downstream representation geometry.
The empirical efficacy of SMCCL-Net was demonstrated on two public corporate bankruptcy benchmarks: the Polish Companies Bankruptcy Dataset from the UCI Machine Learning Repository, analyzing 10,503 firms under its third-year forecasting subtask with 65 numerical indicators, and the American Companies Bankruptcy Prediction Dataset, covering 8,262 publicly listed firms across 78,682 firm-year observations from 1999 to 2018. Across both benchmarks, SMCCL-Net achieved the highest PR-AUC and G-mean scores among twelve baseline methods, including classic classifiers like XGBoost and LightGBM, modern tabular models such as TabNet, FT-Transformer, ExcelFormer, TabPFN, and TabM, as well as multimodal controls like Dual-Encoder Fusion and Template-Text Fusion.
Statistical significance testing via paired t-tests across five random seeds confirmed that the observed improvements in minority-class recognition are statistically robust. After applying Holm-Bonferroni corrections to control family-wise error across baseline comparisons, PR-AUC improvements remained statistically significant, exhibiting adjusted p-values between 0.030 and 0.048 and positive 95 percent confidence intervals. G-mean improvements likewise remained significant, with corrected p-values spanning 0.024 to 0.036. In contrast, overall ROC-AUC scores for SMCCL-Net remained comparable to, but did not significantly exceed, top-tier tabular baselines such as TabM, scoring 0.928 versus 0.930 on the Polish data and 0.922 versus 0.924 on the American data. This performance divergence emphasizes that the architectural advantages of SMCCL-Net are specifically concentrated in precision-recall balance and minority-class discrimination rather than uniform metric inflation across majority classes.
Ablation experiments and representation analyses verified the operational mechanisms driving these gains. Visualizations using t-SNE and feature distribution plots revealed that SMCCL-Net produces more compact high-risk cluster structures and reduces overlap between healthy and bankrupt enterprises compared to base tabular encoders and nearest-neighbor retrieval models. Mechanism ablations removing the semantic generation module caused the steepest declines in PR-AUC and G-mean, while shuffling textual descriptions across non-matching enterprises degraded performance, confirming that predictive gains require genuine semantic correspondence rather than generic textual noise. Furthermore, full coupling of quality filtering with class-aware contrastive optimization outperformed configurations implementing either mechanism in isolation, confirming that downstream performance depends on directly linking semantic view selection to minority-class contrastive weighting.
Notwithstanding its empirical advantages, the researchers detailed concrete operational trade-offs and deployment constraints associated with the architecture. Under online semantic-enhanced inference, each evaluated firm requires three generative API requests and three automated quality evaluations, totaling six large language model calls per prediction. While cached inference eliminates online generative calls by storing precomputed texts for batch processing, latency-critical environments may necessitate reverting to the tabular-only fallback model, which operates without semantic enhancements. Additionally, the study noted that the empirical findings are currently bounded by corporate bankruptcy labels derived from numerical accounting ratios on two datasets, excluding real-time macroeconomic shocks, supply-chain disruptions, and news sentiment. The authors concluded that future investigations must explore multi-model verification and determine whether the coupled semantic contrastive approach generalizes effectively to broader credit domains like loan default and bond rating classifications.
Subject of Research: Corporate credit risk prediction using cross-modal contrastive learning and semantic generation
Article Title: A semantic quality generation and class-aware contrastive learning network for corporate credit risk prediction
Article References: Liu, T. (2026). A semantic quality generation and class-aware contrastive learning network for corporate credit risk prediction. Discover Computing, 29(1), Article 874. https://doi.org/10.1007/s10791-026-10767-w
Image Credits: AI Generated
DOI: 10.1007/s10791-026-10767-w
Keywords: corporate credit risk, bankruptcy prediction, contrastive learning, cross-modal representation, SMCCL-Net, class imbalance, tabular data, large language models, precision recall AUC, G-mean, ROC-AUC, Polish bankruptcy dataset, American bankruptcy dataset, hard sample mining, subclass clustering, financial indicators, representation learning
News Source: Denise Maddox. (November 12, 2026). Cross-Modal Contrastive Framework Improves Tabular Corporate Bankruptcy Prediction for Rare Defaults. Scienmag.



