Construction of Knowledge Graph in the power Field Based on Language model

With the continuous expansion of the scale and the increasing complexity of the structure of the smart distribution network, knowledge in the power domain is still stored in unstructured text form at present. The content structure is complex and the query efficiency is low, making it difficult to support comprehensive analysis and decision-making during operation and maintenance. To address this issue, this paper constructs a domain-specific knowledge graph based on a language model using real-world data from the power sector. First, the original text is preprocessed through stop word filtering and Begin-Inside-Outside (BIO) annotation. Then, a Bidirectional Encoder Representations from Transformers - Bidirectional Long Short-Term Memory - Conditional Random Field (BERT- BiLSTM-CRF) model is applied for accurate entity recognition. Next, a Bidirectional Encoder Representations from Transformers - Attention (BERT-Attention) model is used to extract relationships between entities, identifying various semantic relations such as "cause," "action," "contain," and "exist." Finally, the extracted knowledge triples are visually stored and managed using the Neo4j graph database, enabling intuitive visualization of the knowledge graph. Experimental results show that the proposed model outperforms baseline models in both entity recognition and relation extraction tasks, with improvements in precision, recall, and F1 score for named entity recognition by 12.26%, 11.03%, and 11.63%, respectively, and for relation extraction by 5.15%, 7.76%, and 6.48%, respectively.