TRAMM: Triple-Relational Attentive Message Passing with Dynamic Multimodal Fusion for Knowledge Graphs
Minghe Yu , Zhi Dong , Chenxuan Shan , Songran Liu , Tiancheng Zhang , Ge Yu
Multimodal knowledge graph completion (MMKGC) seeks to enhance link prediction by leveraging visual and textual information, yet naively fusing these signals often introduces noise and undermines structural consistency. Existing methods typically adopt static and relation-agnostic fusion strategies, failing to account for the relational semantics that govern howmultimodal features should be utilized. To address this limitation, we propose TRAMM (Triple-Relational Attentive Message Passing with dynamic Multimodal fusion), a structure-dominant framework built on the observation that the same entity requires different modal emphasis under different relations. At the core of TRAMM is a multimodal adaptive fusion module, which dynamically filters and enhances visual and textual features conditioned on specific relations. TRAMM further incorporates a triple-channel message passing module to align multimodal signals with structural semantics, along with a completion scoring module with logical consistency regularizers that captures canonical relational patterns during prediction. Together, these components enable adaptive multimodal fusion while preserving the geometric properties of the knowledge graph. Extensive experiments show that TRAMM achieves new state-of-the-art performance on relation-rich datasets DB15K and MKG-W, and delivers highly competitive results on the simpler MKG-Y dataset. Ablation studies confirm the contribution of each module, and qualitative case studies illustrate how our relation-guided modulation realizes the adaptive, relation-specific emphasis on different modalities.
knowledge graph completion / multimodal knowledge graph / dynamic fusion / relational reasoning / graph neural networks / link prediction
Higher Education Press 2026
/
| 〈 |
|
〉 |
PDF
(12779KB)
