Original title: FSPER-Net: Fine-Grained Selective Prototype-Expert Routing for Long-Tailed Chinese Telecom Fraud Text Classification
Guosheng Tan, Yuhao Feng, Yiming Geng, Yuxuan Zhou +2 · September 2026
Sorting Chinese telecom fraud messages is hard because fraud categories overlap, some scams appear rarely, and scammers use many different scripts within one category. Traditional models often use one prototype, which is a single average example for an entire class. This average can hide important variations across different scripts. To fix this, researchers built FSPER-Net. It uses a bank of multiple script centroids alongside written descriptions of each scam. The system then applies gates to selectively adjust a primary RoBERTa-WWM classifier for each text. Multiple centroids cut distortion inside three diverse categories by 11.17–22.39%. Overall, FSPER-Net reached an 84.1236% mean Macro-F1 score on one dataset and 97.5329% on another, showing that tracking specific scripts improves fraud detection.
You're a teacher grading essays, and you keep one example of an A paper. A student writes a great essay in a different style, and you mark it down because it doesn't match your single example. So you keep several A papers, each in a different style, and compare each new essay to the style it actually resembles. In this paper, the system keeps several example texts per fraud category and lets each new text choose which examples to compare itself against.
Telecom fraud text classification must separate overlapping labels under class imbalance, and each category contains multiple writing styles. This model improves accuracy on those hard cases by learning several scripts per category, which helps fraud detection systems avoid confusing similar scams and missing rare ones.
The abstract reports that the main Macro-F1 gains were not statistically significant in paired tests, and the diagnostic gains came from a separate three-seed experiment on validation-defined difficult samples. The secondary evaluation used a different dataset. Generalization beyond Chinese telecom fraud text is not established.
Ask anything about this paper. Answers come from its abstract, and say so when the abstract doesn't cover your question.
FSPER-Net: Fine-Grained Selective Prototype-Expert Routing for Long-Tailed Chinese Telecom Fraud Text Classification — Guosheng Tan, Yuhao Feng, Yiming Geng et al. — Electronics — doi:10.3390/electronics15184337
Abstract licence: CC BY
Explanation generated by Synthesis from the abstract — it is not the original paper.