Knowledge Graph Construction: Extraction, Learning, and Evaluation

A Knowledge Graph (KG), which structurally represents entities (nodes) and relationships (edges), offers a powerful and flexible approach to knowledge representation in the field of Artificial Intelligence (AI). KGs have been increasingly applied in various domains—such as natural language processing (NLP), recommendation systems, knowledge search, and medical diagnostics—spurring continuous research on effective methods for their construction and maintenance. Recently, efforts to combine large language models (LLMs), particularly those aimed at managing hallucination symptoms, with KGs have gained attention. Consequently, new approaches have emerged in each phase of KG development, including Extraction, Learning Paradigm, and Evaluation Methodology. In this paper, we focus on major publications released after 2022 to systematically examine the process of KG construction along three core dimensions: Extraction, Learning Paradigm, and Evaluation Methodology. Specifically, we investigate (1) large-scale data preprocessing and multimodal extraction techniques in the KG Extraction domain, (2) the refinement of traditional embedding methods and the application of cutting-edge techniques—such as Graph Neural Networks, Transformers, and LLMs—in the KG Learning domain, and (3) both intrinsic and extrinsic metrics in the KG Evaluation domain, as well as various approaches to ensure interpretability and reliability.