A Self-Improving, RAG-Enhanced Framework for Automatic Knowledge Graph Construction from Climate Event News

Abstract

Knowledge Graphs (KGs) have emerged as important tools in information management and data analysis, with applications spanning recommendation systems, medical diagnostics, and complex event understanding. However, constructing KGs automatically from climate event news remains challenging due to the presence of redundant entities and intricate relationships that arise from semantic complexity of climate events. Traditional methods (e.g., those relying on fine-tuned BERT models) often require extensive annotated datasets and struggle to merge semantically similar entities. Recent advancements in Large Language Model (LLM)-based algorithms show promise but are prone to hallucination, leading to inaccurate KGs. To address these limitations, we propose a novel LLM-based framework that combines Retrieval-Augmented Generation (RAG) with an iterative self improving mechanism to improve the quality and accuracy of KGs extracted from climate event news. We compared the proposed method with three other commonly used prompting techniques. The experimental results showed that our method can effectively trace the driving factors of complex climate events and construct precise and reliable KGs. This framework offers a scalable, cost-effective solution for constructing domain-specific KGs, contributing to informed decision-making in managing complex crises.