Modality Relational Augmentation for Knowledge Graph Completion

作者：Li Du, Yongli Wang · 发表于：2025 5th International Conference on Digital Society and Intelligent Systems (DSInS) · 年份：2025 · DOI：10.1109/dsins68311.2025.11330070

A bstract-Multimodal Knowledge Graph Completion (MMKGC) integrates structural triples with multimodal entity representations for inferring missing facts. However, modal diversity and imbalance are prevalent in real-world MMKGs, where the inconsistency in representation learning is hampered by unevenly distributed modalities. To address these issues, we propose MIMORA-a novel framework that fuses relational and multimodal semantics via two complementary modules. The Cross-Modal Relational Attention Fusion (CMRAF) module enables selective cross-modal fusion and fine-grained semantic alignment by adaptively capturing relation-dependent cross-modal interactions. The Collaborative Augmentation of Modality Embeddings (CAME) module leverages an adversarial generation technique to synthesize and augment limited modalities, thereby enhancing representational completeness and alleviating modality imbalance. Extensive experiments on the recently constructed DivMMKGC benchmark (covering textual, visual, and numerical modalities) validate MIMORA's effectiveness and generalization in complex multimodal scenarios. These experiments demonstrate that MIMORA consistently achieves state-of-the-art performance across multiple datasets and maintains robustness under severe modality imbalance scenarios.