MESN: A multimodal knowledge graph embedding framework with expert fusion and relational attention

Abstract

Knowledge graph embedding is essential for knowledge graph completion and downstream applications. However, in multimodal knowledge graphs, this task is particularly challenging due to incomplete and noisy multimodal data, which often fails to capture semantic relationships between entities. While existing methods attempt to integrate multimodal features, they frequently overlook relational semantics and cross-modal dependencies, leading to suboptimal entity representations. To address these limitations, we propose MESN, a novel multimodal embedding framework that integrates relational and multimodal signals through semantic aggregation and neighbor-aware attention mechanisms. MESN selectively extracts informative multimodal features via adaptive attention and expert-driven learning, ensuring more expressive entity embeddings. Additionally, we introduce an enhanced ComplEx-based scoring function, which effectively combines structured graph interactions with multimodal information, capturing both relational and feature diversity. Extensive experiments on standard multimodal datasets confirm that MESN significantly outperforms baselines across multiple evaluation metrics. Our findings highlight the importance of relational guidance in multimodal embedding tasks, paving the way for more robust and semantically-aware knowledge representations.