MNLEA-Multi-Modal Neighbor Context Learning for Entity Alignment

Multi-modal entity alignment (MMEA) plays a crucial role in Multi-modal Knowledge Graphs (MMKGs), aiming to identify entities in different graphs that represent the same meaning in real world. MMEA has gained significant attention in various domains and have driven the development of numerous downstream applications. However, current MMEA algorithms rely on the entities themselves, which ignores the importance of neighbor contextual features, thus compromising robustness against noise and making them challenging to handle newly emerging entities in real world graphs. In this work, we introduce MNLEA, a multi-modal neighbor context learning approach for MMEA, which uses a new Neighbor Attention Network (NAN) to learn the neighbor context for each entity and employs a cross-modal fusion strategy to aggregate entity features for MMEA. MNLEA can effectively align entities that have never appeared during training strategy owing to the NAN module. We conducted extensive experimental validation on several commonly used MMEA datasets. Experimental results demonstrate that our model can indeed improve the performance of MMEA in multiple scenarios.