Enhanced Cybersecurity Entity Recognition Using DeBERTa, Transformer-CNN Hybrids, and BiLSTM-Softmax

In recent years, cybersecurity threats have been increasing, making data-driven security intelligence analysis a key research focus. Artificial intelligence techniques, particularly knowledge graph-based methods, offer support for detecting complex and unknown network attacks in multi-source heterogeneous threat intelligence data. Cybersecurity entity recognition serves as the foundation for constructing threat intelligence knowledge graphs. However, the complexity of security entities in open network text data makes traditional deep learning methods less effective in accurate identification. To address this challenge, we propose a cybersecurity entity recognition model based on DeBERTa, Transformer-CNN hybrids, and BiLSTM-Softmax. The DeBERTa model is leveraged to generate character-level feature representations, enhancing contextual understanding. A Transformer-CNN hybrid is employed to effectively extract crucial security entity features by combining convolutional feature extraction with self-attention mechanisms. Finally, BiLSTM-Softmax is used to generate BIO labels for each character in the sequence. Experimental results on a large-scale annotated cybersecurity entity dataset demonstrate that the proposed approach achieves superior performance compared to LSTM-CRF, BiLSTM-CRF, and traditional entity recognition models.