Relation Prototype Driven Multimodal Knowledge Graph Completion

Multimodal knowledge graph completion (MMKGC) aims to infer missing knowledge by integrating structural information with textual and visual modalities. Although relations provide essential semantic signals for accurate inference, most existing methods primarily focus on enhancing entity representations while paying limited attention to explicit multimodal modeling of relations. This limitation leads to insufficient exploitation of relational semantics and weakens the ability to perform relation-guided reasoning. To address the mentioned issue, we propose RelPro, a novel method which introduces shared relation prototype contrastive learning to enhance fine-grained multimodal representations of the relations and improve MMKGC performance. RelPro selects informative and representative multimodal features to construct relation prototypes and further leverages contrastive learning to strengthen the capability of the relation prototypes to capture multimodal semantics, allowing the model to incorporate these prototypes more effectively into the prediction process. Extensive experiments on benchmark MMKGC datasets demonstrate that RelPro achieves substantial improvements over competitive baselines, highlighting the effectiveness of the enhanced relation modeling in MMKGC.