Traditional researches usually define named entity recognition (NER) as a sequence labeling task. But when it comes to specific domain, the entities are usually longer and contain prefixes representing domain knowledge, making their boundaries often more ambiguous compared to general entities. We draw inspiration from the word-word relation classification approach in the W2NER model and combined it with Conditional Layer Normalization (CLN), multigranularity dilated convolution, and a joint predictor that integrates Biaffine and MLP. On this basis, we enhance the model by adding Rotary Position Embedding (RoPE) to both the start and end sequences of the Biaffine classifier. Furthermore, we build a typical domain specific entity dataset in the defense domain. And we compare our model with the advanced baseline on our dataset and CLUENER2020 dataset, the experimental results verify our model outperforming the current baseline model.