关键词: 多模态数据, 多模态知识图谱, 多模态知识图谱补全, 实体对齐, 关系推理

Abstract: Traditional knowledge graphs were employed to provide a unified and machine-interpretable representation of information on the web; however, their limitations in handling multimodal applications were increasingly recognized. To address these limitations, Multi-Modal Knowledge Graphs (MMKGs) were proposed as an effective solution. Nevertheless, the integration of multimodal data into knowledge graphs was often accompanied by challenges such as inadequate modality fusion and reasoning difficulties, which significantly constrained the application and further advancement of MMKGs. Multi-Modal Knowledge Graph Completion (MMKGC) techniques were introduced to facilitate the comprehensive integration of cross-modal information during the construction phase and to enable the prediction of missing links after construction, thereby mitigating issues in modality fusion and reasoning. A detailed overview of MMKGC was presented, including widely used benchmark datasets and evaluation metrics. Subsequently, existing approaches were classified into fusion tasks during the MMKG construction phase and reasoning tasks in the post-construction phase. The former focused on key techniques such as entity alignment and entity linking, while the latter encompassed relation inference, missing information completion, and multimodal expansion. Representative MMKGC methods in each category were thoroughly introduced and their characteristics systematically analyzed. Finally, the prevailing challenges faced by MMKGC were examined, and a comprehensive summary was provided.

Key words: multi-modal data, multi-modal knowledge graph, Multi-Modal Knowledge Graphs Completion (MMKGC), entity alignment, relation inference