A Framework for Entity-Relationship Extraction in the Electric Domain Based on Large Language Models

The Entity Relationship Extraction task involves accurately extracting target entities and uncovering their complex relational interdependencies within unstructured text, thereby establishing foundational structural knowledge for downstream applications including knowledge graph construction and semantic network population. While conventional approaches predominantly focus on sentence-level entity relationship extraction in general-domain texts. However, with the continuous expansion of practical needs, the task of entity relationship extraction no longer merely satisfies the extraction of relationships between entities from a single sentence. Instead, it requires accurately identifying the complex relationships between entities from multiple sentences.meanwhile, in the context of the Electric-power industry, in the task of entity relationship extraction in the electric professional field, due to the presence of a large number of professional terms and specific expressions in the texts of the power professional field, the relationships between entities are more complex and difficult to define. Using previous methods cannot accurately identify the entities in the power professional field and their relationships from the texts. To address these challenges, we propose a new framework for entity-relationship extraction in the electric domain based on large language models and modular fine-tuning, which we have named the R-E-M(Relationship-Entity Matching) framework. This framework comprises four components. First, it enhances the large language model's capability to recognize domainspecific entities in the power industry through targeted modular fine-tuning. Next, an entity-relationship extraction module identifies relevant entities and their corresponding relationships within the text. Subsequently, the entity-relationship matching module establishes predefined matching patterns between entities and relationships for each identified relation type. Finally, the prompt selection module dynamically chooses appropriate prompts based on different relationship types while specifying standardized output formats for the model. From the final experimental results, our proposed new framework has significantly improved the large language model's performance in the entity-relationship extraction task for power domain texts, elevating the Micro-F1 score from 0.185 to 0.575 -a substantial gain of 0.39, demonstrating remarkable enhancement.