Enhanced Entity Interaction Modeling for Multi-Modal Entity Alignment

Jinxu Li, Qian Zhou, Wei Chen, Lei Zhao

Abstract: Multi-modal Entity Alignment (MMEA) aims to find equivalent entities across different multi-modal knowledge graphs (MMKGs). Most existing methods focus on how to encode or fuse information from different modalities effectively, without considering the critical interactions between entities, especially those between an entity and its neighbors within each modality. To fill the gap, we propose a novel model namely Enhanced Entity Interaction Modeling for Multi-modal Entity Alignment (EIEA). Specifically, we first utilize multiple separate pre-trained models to acquire single-modal data based entities’ embeddings. Then, the module Enhanced Entity Representation (EER) is designed to mine interactions between entities and their neighborhoods, and facilitate effective multi-modal embedding fusion using a weighting mechanism. Finally, through contrastive learning, we ensure that the aligned entities have higher similarity than non-aligned ones within each modality. The extensive experiments demonstrate that EIEA outperforms the state-of-the-art baselines on three benchmark datasets.