Domain Knowledge Graph Construction Approach Based on Large Language Models

Current approaches for constructing domain knowledge graphs suffer from limitations such as reliance on manually defined rules, insufficient annotated data, and excessive dependence on domain experts, leading to inefficient processes and inconsistent quality. To address the above problems, this paper proposes a novel LLM-based framework for building domain knowledge graphs. The framework leverages the information extraction capabilities of LLMs to automatically identify core components of knowledge graphs—including entities, relational triples, and events — from unstructured text. To enhance performance on domain-specific data and mitigate knowledge hallucination, the framework integrates retrieval-augmented generation (RAG) with external knowledge bases and incorporates a reflective case library. Additionally, it features a comprehensive error detection and correction mechanism, along with a redundancy processing module, ensuring the generated knowledge graphs are concise and reliable. To validate the effectiveness of our approach, comparative and ablation experiments were conducted on CrossNER, CrossRE, and SciERC datasets. Results show: on CrossRE/SciERC, relation extraction F1-scores reach $0.575 / 0.578$ ($13.2 \% / 10.1 \%$ higher than SAC-KG); ablation studies confirm RAG cuts hallucinations by $\sim 8 \%$ and the reflective case library improves error avoidance by $\sim 5\%$.