Research on the Automated Identification of Sensitive Words in Electric Power Data Based on Bert-att-Kan

Power-sensitive data identification is of great significance to the power industry, but the data types are complex and their distribution is seriously unbalanced, making it challenging to address using traditional methods. Therefore, this paper proposes a recognition method based on dual-label marking and dual-module collaboration. In this method, "sensitive" or "insensitive" labels are assigned based on the data’s sensitivity, and "few-class" or "multi-class" labels are defined based on the sample size of each class, to reflect the data’s distribution characteristics and introduce label correlation. During model training, a two-branch network is employed to identify sensitivity and category separately, utilizing focal loss, cross-entropy loss, and label correlation loss to ensure consistency between the predictions of the two branches and their corresponding relationships. Experimental results demonstrate that the proposed method significantly improves recognition performance when the data is limited and unevenly distributed.