Construction and Dynamic Update Technology of Knowledge Graph for Domain-specific Large Language Models in Electric Power Communication

Abstract

Aiming at the problems in knowledge management of Electric Power Communication (EPC), such as insufficient use of unstructured information and lagging update of dynamic knowledge, this article studies the construction and dynamic update technology of knowledge map of EPC-specific Large Language Model (LLM). Firstly, the domain knowledge system is analyzed, and the core types of entities and relationships are defined. A knowledge extraction algorithm based on domain-enhanced prompt learning is proposed, which combines EPC terminology dictionary and prompt template library to improve the accuracy of entity and relationship recognition. An ontology alignment method integrating semantic similarity and rule reasoning is also designed to realize the unified representation of multi-source heterogeneous data. The incremental updating framework of time sequence awareness is constructed, and the inconsistency in the process of knowledge evolution is dealt with through conflict detection and confidence assessment mechanism. The results show that the F1 value of entity recognition is 92.3%, the F1 value of relation extraction is 89.7%, and the accuracy of ontology alignment is 94.6%. In the dynamic update scenario, it takes an average of 12.5 seconds to process 1000 pieces of data at a time, and the accuracy of conflict handling is 90.3%. The research shows that the proposed method can effectively construct a knowledge map adapted to EPC field and realize efficient dynamic update, which provides reliable knowledge support for intelligent operation and maintenance of EPC network.