A high-accuracy unsupervised statistical learning method for joint dangling entity detection and entity alignment

Abstract

Dangling entities are common in knowledge graphs but there is a lack of research on entity alignment involving them. Most existing studies leverage neural network methods through supervised learning. However, these data-driven methods suffer from poor interpretability and high computation overhead. In this paper, we propose a Simple Unsupervised Dangling entity detection and entity Alignment method (SUDA)¹ without employing neural networks. Our method consists of three modules: entity embedding, dangling entity detection, and entity alignment. While the state-of-the-art Simple but Effective Unsupervised entity alignment method (SEU)² is incapable of dealing with dangling entities, SUDA further extends it and addresses the bilateral dangling entities problem. Theoretical proof of our method is given. We also design a new adjacent matrix for incorporating richer entity relations. Then we construct entity similarity outlier intervals to detect dangling entities and align entities through assignment problem after removing them. Extensive experiments demonstrate that our method outperforms those supervised and unsupervised methods. Additionally, in the entity alignment tasks, SUDA consumes less runtime compared to neural network methods, while maintaining high efficiency, interpretability, and stability. Code is available at https://github.com/skyccong/SUDA.git.