LLM-Based Digital Twin Water Conservancy Knowledge Graph Construction

In the study, a novel method for extracting domain-specific knowledge in digital twin water conservancy construction is proposed, addressing the challenges of interdisciplinary complexity and the limitations in existing knowledge extraction models' understanding capabilities of domain-specific knowledge. Utilizing a large language model, the novel method incorporates the domain knowledge of digital twin water conservancy into a local model framework. It employs prompt-based fine-tuning and leverages the semantic understanding and generative capabilities of the model to enhance knowledge extraction precision. To optimize entity extraction, innovative heterogeneous entity alignment strategies are introduced. The efficacy of the proposed method is validated through comparative and ablation studies conducted on a water domain corpus. Results indicate a significant improvement over conventional models, with F1 scores for entity and relationship extraction at 88.63% and 84.46%, respectively, and an entity extraction precision rate of 90.11%. Ablation experiments further demonstrate that our method outperforms the baseline large language model, enhancing the F1 scores for entity and relation extraction by 5.5 and 3.2 percentage points, respectively. By adjusting the amount of domain knowledge injected, the influence of the amount of domain text on the knowledge extraction effect is explored.