Research on Named Entity Recognition Models for Cybersecurity

In response to the challenges faced by existing Chinese cybersecurity named entity recognition models regarding the handling of Chinese language complexity and ambiguity during training, which utilizes a mixed corpus of Chinese and English in the network security vertical domain. The model is named BERT-BiGRU-Self-Attention-CRF. This model addresses the issue of polysemy by utilizing BERT to process cybersecurity texts and obtain contextual representations for each word. Then, the obtained word vectors are processed forward and backward in a Bidirectional Gated Recursive Unit (BiGRU) to better capture the contextual dependencies of the required network security related data. Inputting the output of the BiGRU layer into an attention mechanism allows the attention mechanism to assign different weights to each word in the sentence. Finally, a conditional random field is employed for data labeling, accurately annotating the entities. Through relevant comparative experiments, under the same experimental environment, F1 value of the model has enhanced by 3.46% contrasted to the LSTM-CRF model. In addition, compared model based on pre trained model BERT, F1 score increased by 1.84%. Compared to the BiGRU model with BERT, the F1 score is 0.80% higher. These findings fully demonstrate the significant role and advantages of the model in the vertical domain of Chinese English mixed cybersecurity.