Semi-supervised Entity Alignment via Noisy Student-Based Self Training

Abstract

Knowledge graph is a structured data model that captures the relationships between entities in the real world. Entity alignment (EA) has drawn significant attention in recent years as a potential means of identifying corresponding entities across different knowledge graphs. Although knowledge graph embedding-based entity alignment methods have recently obtained significant progress, the shortage of training data remains a severe challenge. Conventional approaches have attempted to solve this issue through semi-supervised learning but still suffer from the negative impacts of entity alignment. To resolute the above issues, we propose a semi-supervised framework with Noisy Student-based self training Entity Alignment named NSEA. Our framework proposes a new noisy student self-training strategy for obtaining diverse entity alignment pairs, and we also design an adaptive alignment selector to infer reliable entity pairs. Through extensive experiments on benchmark datasets, we demonstrate that our method outperforms most existing models in terms of accuracy and efficiency, highlighting its usefulness for large-scale and diverse knowledge graphs with insufficient annotated data.

This work is in part supported by the National Natural Science Foundation of China under Grant 62202221, and the Natural Science Foundation of Jiangsu Province under Grant BK20220331.