Abstract: In the construction of a knowledge graph for radar maintenance, the strong specialization and scarcity of annotated datasets pose significant challenges in training named entity recognition (NER) models, with traditional model training failing to meet application requirements. Building upon the BiGRU-CRF model, this paper introduces a pre-trained model and proposes the ERNIE-BiGRU-CRF model. First, taking a specific radar model as an example, maintenance data were collected and preprocessed. The doccano platform was used for manual annotation, resulting in over 1,100 labeled NER data entries in the radar maintenance domain. Next, dynamic word embeddings for the radar maintenance training data were obtained using the ERNIE pre-trained model, while BiGRU captured bidirectional semantic information. Finally, the most reasonable sequence labeling results were derived through CRF constraints. Experimental results show that, with limited training data, the proposed model achieves strong recognition performance. Compared to BiGRU-CRF and BiLSTM-CRF models, it demonstrates an improvement in F1-score, effectively addressing the issues of insufficient training data and suboptimal training performance in the radar maintenance domain. This model holds practical value for the automated construction of knowledge graphs in radar maintenance.

Key words: radar maintenance, named entity recognition, enhanced representation through knowledge integration(ERNIE) model, large language model