An Anti-Interference Chinese Text Classification Model and Its Application in Power Systems

Abstract Automated classification of power grid work order texts is crucial for operational efficiency and system reliability, yet existing models often lack robustness in complex Chinese language scenarios involving writing errors and adversarial perturbations. To address this, we propose CCBERT, an anti-interference Chinese text classification model. CCBERT builds upon ChineseBERT and innovatively incorporates a semantic embedding generation module. This module leverages a sophisticated character selection mechanism to identify key information, employs multi-weight adversarial attacks to generate challenging samples, and utilizes a contrastive learning strategy to enforce semantic consistency between original texts and their perturbed variants in the embedding space. Extensive experiments on a real-world power grid work order dataset and the public THUCNews dataset demonstrate that CCBERT significantly outperforms baseline models, including BERT, ERNIE, RoBERTa, MacBERT, and ChineseBERT, across key classification metrics (Accuracy, F1-Score, Micro/Macro-AUC). Notably, CCBERT exhibits superior robustness and the smallest performance degradation on specially constructed adversarial (Adv) and typo test sets. These findings validate that CCBERT, through its unique semantic embedding optimization, effectively enhances the accuracy and resilience of Chinese text classification in complex, noisy environments, offering robust technological support for intelligent text processing in specialized domains like power systems.