Cascade Sampling via Dual Uncertainty for Active Entity Alignment

Jiye Xie, Jiaxin Li, Jiawei Tan, Hongxing Wang

Abstract: Entity Alignment (EA) aims to find and unite equivalent entities across different knowledge graphs for knowledge fusion. It requires pre-aligned entity pairs as seed alignments to train an EA model. Recent effort has employed active learning (AL) to query more informative seed alignments for effective EA modeling at a lower cost. However, it still challenges existing AL methods to find and diversify seed alignments since true alignments themselves are sparse and unavailable before getting annotated. To address this issue, we manipulate seed alignment query based on entity selection on a single knowledge graph and deploy active learning on the EA task by querying entities that behave with (i) Matching Uncertainty determined by the EA model in training and (ii) Novelty-oriented Uncertainty estimated through diverse entity identification. To adapt the query set to changes in the EA model and aligned entities during AL iterations, we propose a dynamic cascade sampling strategy by trading-off between matching uncertainty and novelty-oriented uncertainty in a two-stage manner. Experiments on real-world benchmark datasets show the effectiveness of the proposed approach in comparison with state-of-the-art methods.