A review of background, methods, limitations and opportunities of knowledge graph completion

Knowledge graph completion (KGC), as a pivotal technology for extracting hidden knowledge from large-scale data, has evolved into a systematic research framework through the development of knowledge graph (KG) technology in recent years. To address the many challenges faced by current research, this study systematically reviews the fundamental theories and methodological systems in the field of KGC, grouping them into four categories: embedding-based, path-based, neural network-based, and large language model (LLM)-based approaches. Current research indicates that traditional closed-domain KGC relies on standard KG embedding or relational path models, which remain effective for completing structured data. However, there are notable limitations in handling unseen entities and relations in open scenarios. With breakthroughs in neural networks and LLMs, open-domain KGC has begun to emerge, although a lack of systematic analysis and classification of model architectures still persists. To address this gap, this review conducts a multi-dimensional academic investigation, clarifying the foundational research landscape and core methodological distinctions, establishing a model classification framework that spans both closed and open domains and integrating mainstream dataset resources within the field. Furthermore, the review explores the challenges and future directions of technological development, including critical issues such as complex knowledge reasoning, improvements in domain adaptability improvement, and the deep integration of LLMs with KGs, providing theoretical foundations and practical references to guide subsequent research efforts.