Research on BERT-Based Named Entity Method for Network Security

In the field of cybersecurity information, there are numerous challenges such as a large amount of redundant information, ambiguous entity boundaries, and polysemy. We propose a named entity recognition (NER) model based on the BERT-BiLSTM-MVCA-CRF architecture for the cybersecurity domain. First, we use the BERT pre-trained model to obtain dynamic word embeddings. These word embeddings are then used as input to a BiLSTM network to capture contextual feature vectors of the text. Furthermore, we introduce the MVCA mechanism to fully capture multi-scale spatial information and different channel correlations. Finally, the processed results are fed into a CRF layer for labeling. To validate the effectiveness of the model, we conducted a series of experiments using a cybersecurity dataset. The results showed a significant improvement, with the F1 score reaching 89.22%, outperforming other NER models.