A semantic quality generation and class-aware contrastive learning network for corporate credit risk prediction

Abstract Corporate credit risk prediction is an important task in financial risk management and intelligent decision-making, aiming to identify high-risk enterprises defined by bankruptcy outcomes based on corporate financial and operational information. However, existing methods mostly rely on structured financial indicators and still suffer from insufficient semantic representation, uncontrollable quality of generated textual descriptions, and limited capability in identifying minority high-risk samples. To address these issues, this paper proposes SMCCL-Net, a framework that couples semantic-view reliability with category-aware cross-modal representation learning. Multiple prompts are used to generate candidate financial descriptions, and semantic quality evaluation determines which candidate is used as the positive textual view of each financial record. The selected tabular-text pairs are then optimized in a shared latent space, where class frequency, subclass relationships, and cross-modal difficulty jointly determine their contributions to representation learning. Therefore, the main contribution lies in connecting quality-controlled semantic-view construction with minority-class-oriented cross-modal optimization, rather than in the isolated use of the individual techniques. Experimental results on two public corporate bankruptcy prediction datasets show that SMCCL-Net achieves the highest PR-AUC and G-mean among the compared methods, while maintaining ROC-AUC comparable to strong tabular baselines. The gains are therefore concentrated in minority-class recognition and balanced classification rather than uniform improvement across all metrics. Ablation experiments and representation space analysis further demonstrate that semantic quality filtering, cross-modal alignment, and category-aware mechanisms contribute positively to model performance. Computational analysis shows that online semantic-enhanced inference requires six LLM API calls per sample, whereas cached and tabular-only modes require no online LLM calls, highlighting the trade-off between predictive gains and deployment overhead. These results indicate that the proposed framework provides a semantic-enhanced representation learning approach for corporate bankruptcy prediction.