Scholarlink

<p>Multi-modal entity alignment seeks to match equivalent entities across different multi-modal knowledge graphs, which integrate heterogeneous multi-modal data such as images and text to enrich entity semantics. However, variations in multi-modal data quality and their inherent unreliability present significant challenges that can negatively impact alignment results. Consequently, we propose NoCo, a noise-augmented multi-modal entity alignment method with confidence-based dynamic fusion. NoCo incorporates a modality-aware noise enhancement mechanism that adaptively injects Gaussian noise into each modality, thereby improving robustness and preventing overfitting to unreliable features. Simultaneously, a confidence-based dynamic fusion framework is designed to automatically calibrate modality contributions according to data quality, effectively down-weighting noisy inputs while amplifying reliable signals. Experimental evaluations demonstrate that NoCo effectively overcomes these challenges and achieves strong performance. Compared with the state-of-the-art method, NoCo achieves a 2.8% maximum improvement on the Multi-OpenEA datasets, 4.1% maximum improvement on FB15K-DB15K, and 5.0% on FB15K-YG15K. The code of the proposed model is stored at <a href="https://github.com/HubuKG/NoCo">https://github.com/HubuKG/NoCo</a>.</p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

Noise-augmented multi-modal entity alignment with confidence-based dynamic fusion

  • 期刊论文

  • 发表日期：

    2026年6月16日

查看全文
  • Xiangyu Luo,
  • Yan Zhang,
  • Miao Zhang,
  • Kui Xiao,
  • Wenxing Huang,
  • Zhifei Li

摘要

Multi-modal entity alignment seeks to match equivalent entities across different multi-modal knowledge graphs, which integrate heterogeneous multi-modal data such as images and text to enrich entity semantics. However, variations in multi-modal data quality and their inherent unreliability present significant challenges that can negatively impact alignment results. Consequently, we propose NoCo, a noise-augmented multi-modal entity alignment method with confidence-based dynamic fusion. NoCo incorporates a modality-aware noise enhancement mechanism that adaptively injects Gaussian noise into each modality, thereby improving robustness and preventing overfitting to unreliable features. Simultaneously, a confidence-based dynamic fusion framework is designed to automatically calibrate modality contributions according to data quality, effectively down-weighting noisy inputs while amplifying reliable signals. Experimental evaluations demonstrate that NoCo effectively overcomes these challenges and achieves strong performance. Compared with the state-of-the-art method, NoCo achieves a 2.8% maximum improvement on the Multi-OpenEA datasets, 4.1% maximum improvement on FB15K-DB15K, and 5.0% on FB15K-YG15K. The code of the proposed model is stored at https://github.com/HubuKG/NoCo.