Abstract: Abstract: Various multimodal knowledge graphs with distinct ontology designs have been developed in relevant research fields, and integrating these heterogeneous knowledge graphs remains a critical challenge. The entity alignment task for multimodal knowledge graphs enables the identification of equivalent entities across different graphs, eliminates redundant data, and facilitates the construction of an integrated multimodal knowledge graph with comprehensive contents and a unified architecture. Nevertheless, most existing entity alignment algorithms perform well for multimodal entity fusion under normal conditions,whereas they struggle to detect and filter out noise, as well as extract discriminative features from low-quality data. To address this issue, a multi-head latent attention-driven entity alignment model is proposed for multimodal knowledge graphs under low-quality data scenarios. By embedding the multi-head latent attention mechanism and quality-aware multimodal fusion strategy into the model framework, the robustness against poor data is substantially improved. Comparative experiments are conducted against four state-of-the-art classical multimodal knowledge graph entity alignment approaches. Experimental results demonstrate that the proposed model maintains stable alignment performance on datasets containing dynamic noise, with overall performance superior to the comparative schemes corresponding to two types of public benchmark datasets.

Key words: multimodal knowledge graphs, entity alignment, multi-head latent attention, knowledge graphs, Keywords: quality-aware fusion

摘要： 摘 要: 在专业领域中，存在众多本体设计各异的多模态知识图谱，如何融合本体架构不同的知识图谱是一大难题。依托多模态知识图谱实体对齐任务能够识别图谱间的相似实体，剔除冗余信息，进而搭建内容更全面、架构统一的融合型多模态知识图谱。然而现有多数实体对齐算法虽能完成常规场景下的多模态实体融合，在低质量数据场景中却难以精准甄别、滤除噪声，有效提取关键特征。为此构建一种面向低质量数据、由多头潜在注意力驱动的多模态知识图谱实体对齐模型，将多头潜在注意力机制与质量感知多模态融合策略嵌入模型架构，以此提升模型面对劣质数据时的鲁棒性能。选取四类经典多模态知识图谱实体对齐算法开展对比实验，实验结果显示，该模型在含动态噪声的数据集上仍可维持稳定的对齐性能，综合性能优于两类公开基准数据集对应的对比方案。

关键词: 多模态知识图谱, 实体对齐, 多模态实体融合, 多头潜在注意力, 质量感知融合