FSPER-Net: Fine-Grained Selective Prototype-Expert Routing for Long-Tailed Chinese Telecom Fraud Text Classification

Fine-grained Chinese telecom fraud classification must distinguish overlapping labels under class imbalance while representing multiple scripts within each class. A single prototype can obscure script-specific modes, whereas uniform fusion cannot adapt prototype contributions to individual texts. We propose FSPER-Net, a Fine-grained Selective Prototype-Expert Routing Network, to separate script representation from control over its influence. A dual-source bank combines script-level centroids with Chinese semantic descriptions; class-source and instance-conditioned gates integrate this evidence through bounded corrections to a primary RoBERTa-WWM classifier. Multiple centroids reduced training-set within-class cosine distortion by 11.17–22.39% in three heterogeneous categories. In matched three-seed refinement, the full model exceeded single-prototype and single-source variants by 0.3510 and 0.2797–0.2844 Macro-F1 percentage points, respectively, and direct expert replacement by 0.3695 points. On ten-class FGRC-SCD, FSPER-Net attained the highest mean Macro-F1 among four core models over five seeds (84.1236%), although paired differences were not statistically significant. A separate three-seed diagnostic showed a 1.5569-point gain over PSCL on validation-defined difficult samples. The secondary Telecom_Fraud_Texts_5 evaluation yielded 97.5329% Macro-F1. These findings support script-aware representation and controlled prototype integration for fraud taxonomies combining semantic overlap, within-class diversity, and long-tailed supervision.