Farmer Credit Default Prediction With Survey-Text Soft Information: A Multimodal Bidirectional Attention Fusion Deep Learning Framework

Farmer credit default identification in the context of inclusive finance is significantly hindered by sample scarcity, heterogeneous data sources, and the underutilization of “soft information” embedded in investigation reports. This study proposes BIAF-mDnet, an end-to-end multimodal deep learning framework designed to jointly model structured variables and unstructured investigation texts. Specifically, BIAF-mDnet employs a structured residual enhancement network to learn deep representations of structured data, alongside a frozen RoBERTa projection encoder to extract semantic text representations and achieve latent space alignment. Within the fusion stage, a bidirectional interactive attention fusion (BIAF) module and a gating mechanism are introduced to facilitate cross-modal evidence alignment and adaptive reweighting, thereby enhancing the quality of the joint representation. To address default sample scarcity and class imbalance, a class-weighted Focal Loss is incorporated during training to strengthen minority class learning and optimize the misclassification structure. Using real-world farmer credit data from a national commercial bank, we construct a multimodal dataset comprising numerical, categorical, and textual information derived from loan officers’ pre-loan investigation reports. Empirical results demonstrate that, under multimodal input, BIAF-mDnet achieves the best overall performance, with an Accuracy of 0.9686, an AUC of 0.9728, and an F1-score of 0.8091, while reducing T1E and T2E to 0.0154 and 0.2192, respectively. Compared with its structured-only counterpart, the multimodal version improves AUC by 0.0224 and F1-score by 0.0703, and statistical tests confirm that these gains are significant for most evaluation metrics. These results indicate that textual soft information, when effectively aligned and fused with structured features, provides stable complementary risk cues and consistently enhances predictive performance. SHAP-based interpretability analysis further reveals that the introduction of text does not result in a uniform shift in the importance of structured features; rather, it leads to a redistribution and reordering of explanatory weights, providing local explanations that align more closely with actual risk scenarios. This research provides robust methodological support and practical insights for rural financial institutions seeking to improve risk identification accuracy and credit resource allocation efficiency.