Abstract: At present, State Grid Jilin Electric Power Co. , Ltd. relies on confidential personnel to manually mark the confidentiality level of documents, and its accuracy depends on the professional quality of relevant personnel, which is easy to cause the problem of inaccurate labeling. Therefore, we establish an enterprise document security classification system based on the transformer model, which can automatically extract the feature expression of text security information and intelligently assist the decision-making of enterprise secret documents. The proposed model is trained and tested on the data set constructed by the internal core commercial secret files, ordinary commercial secret files and non secret files of State Grid Jilin Electric Power Company Limited. The accuracy rate is 97. 37% and the recall rate is 98. 67% . The results show that the model achieves high recognition effect and can effectively prevent the disclosure of secret files.

Key words: security classification,  , deep learning,  , self-attention network,  , word embedding,  , enterprise secrets