
摘要
多模态实体对齐(MMEA)旨在发现不同知识图谱(KGs)中相同的实体,这些实体与相关的图像相关联。然而,当前的MMEA算法依赖于知识图谱级别的模态融合策略来表示多模态实体,这忽略了不同实体在模态偏好上的差异,从而降低了对模糊图像和关系等模态噪声的鲁棒性。本文介绍了一种名为MEAformer的多模态实体对齐变换器方法,用于元模态混合,该方法动态预测各模态之间的相互关联系数,实现更细粒度的实体级别模态融合与对齐。实验结果表明,我们的模型不仅在多种训练场景下(包括监督、无监督、迭代和低资源设置)实现了最先进的性能,而且参数数量有限、运行效率高且具有可解释性。我们的代码可在https://github.com/zjukg/MEAformer 获取。
代码仓库
zjukg/MEAformer
官方
pytorch
GitHub 中提及
基准测试
| 基准 | 方法 | 指标 |
|---|---|---|
| entity-alignment-on-dbp15k-fr-en | MEAformer (w/o surf) | Hits@1: 0.845 |
| entity-alignment-on-dbp15k-fr-en | MEAformer | Hits@1: 0.996 |
| entity-alignment-on-dbp15k-fr-en | MEAformer (w/o surf & iter ) | Hits@1: 0.770 |
| entity-alignment-on-dbp15k-fr-en | Unsup. MEAformer (w/o iter ) | Hits@1: 0.973 |
| entity-alignment-on-dbp15k-fr-en | MEAformer (w/o iter ) | Hits@1: 0.991 |
| entity-alignment-on-dbp15k-fr-en | Unsup. MEAformer | Hits@1: 0.993 |
| entity-alignment-on-dbp15k-ja-en | MEAformer (w/o surf) | Hits@1: 0.842 |
| entity-alignment-on-dbp15k-ja-en | MEAformer | Hits@1: 0.991 |
| entity-alignment-on-dbp15k-ja-en | Unsup. MEAformer (w/o iter ) | Hits@1: 0.958 |
| entity-alignment-on-dbp15k-ja-en | MEAformer (w/o iter ) | Hits@1: 0.977 |
| entity-alignment-on-dbp15k-ja-en | Unsup. MEAformer | Hits@1: 0.987 |
| entity-alignment-on-dbp15k-ja-en | MEAformer (w/o surf & iter ) | Hits@1: 0.764 |
| entity-alignment-on-dbp15k-zh-en | Unsup. MEAformer (w/o iter ) | Hits@1: 0.917 |
| entity-alignment-on-dbp15k-zh-en | MEAformer (w/o surf & iter ) | Hits@1: 0.771 |
| entity-alignment-on-dbp15k-zh-en | Unsup. MEAformer | Hits@1: 0.962 |
| entity-alignment-on-dbp15k-zh-en | MEAformer (w/o surf) | Hits@1: 0.847 |
| entity-alignment-on-dbp15k-zh-en | MEAformer | Hits@1: 0.973 |
| entity-alignment-on-dbp15k-zh-en | MEAformer (w/o iter ) | Hits@1: 0.948 |
| entity-alignment-on-fbdb15k | MEAformer (seed 20%) | Hits@1: 0.578 |
| entity-alignment-on-fbdb15k | MEAformer (seed 40% w/o iter) | Hits@1: 0.619 |
| entity-alignment-on-fbdb15k | MEAformer (seed 60% w/o iter) | Hits@1: 0.765 |
| entity-alignment-on-fbdb15k | MEAformer (seed 40%) | Hits@1: 0.690 |
| entity-alignment-on-fbdb15k | MEAformer (seed 20% w/o iter) | Hits@1: 0.417 |
| entity-alignment-on-fbdb15k | MEAformer (seed 60%) | Hits@1: 0.784 |
| entity-alignment-on-fbyg15k | MEAformer (seed 60% w/o iter) | Hits@1: 0.703 |
| entity-alignment-on-fbyg15k | MEAformer (seed 40%) | Hits@1: 0.612 |
| entity-alignment-on-fbyg15k | MEAformer (seed 20% w/o iter) | Hits@1: 0.327 |
| entity-alignment-on-fbyg15k | MEAformer (seed 60%) | Hits@1: 0.724 |
| entity-alignment-on-fbyg15k | MEAformer (seed 20%) | Hits@1: 0.444 |
| entity-alignment-on-fbyg15k | MEAformer (seed 40% w/o iter) | Hits@1: 0.560 |
| multi-modal-entity-alignment-on-umvm-dbp-fr | MEAformer (w/o surf) | Hits@1: 0.845 |
| multi-modal-entity-alignment-on-umvm-dbp-fr | MEAformer (w/o surf & w/o iter) | Hits@1: 0.771 |
| multi-modal-entity-alignment-on-umvm-dbp-ja | MEAformer (w/o surf) | Hits@1: 0.842 |
| multi-modal-entity-alignment-on-umvm-dbp-ja | MEAformer (w/o surf & w/o iter) | Hits@1: 0.764 |
| multi-modal-entity-alignment-on-umvm-dbp-zh | MEAformer (w/o surf & w/o iter) | Hits@1: 0.772 |
| multi-modal-entity-alignment-on-umvm-dbp-zh | MEAformer (w/o surf) | Hits@1: 0.847 |