An Accurate Unsupervised Method for Joint Entity Alignment and Dangling Entity Detection

Findings of the Association for Computational Linguistics: ACL 2022, pages 2330 - 2339 May 22-27, 2022 c⃝2022 Association for Computational Linguistics An Accurate Unsupervised Method for Joint Entity Alignment and Dangling Entity Detection Shengxuan Luo1,2 Sheng Yu1,2 1Center for Statistical Science, Tsinghua University 2Department of Industrial Engineering, Tsinghua University luosx18@mails.tsinghua.edu.cn syu@tsinghua.edu.cn Abstract Knowledge graph integration typically suf- fers from the widely existing dangling enti- ties that cannot find alignment cross knowl- edge graphs (KGs). The dangling entity set is unavailable in most real-world scenarios, and manually mining the entity pairs that consist of entities with the same meaning is labor- consuming. In this paper, we propose a novel accurate Unsupervised method for joint Entity alignment (EA) and Dangling entity detection (DED), called UED. The UED mines the literal semantic information to generate pseudo entity pairs and globally guided alignment informa- tion for EA and then utilizes the EA results to assist the DED. We construct a medical cross- lingual knowledge graph dataset, MedED, pro- viding data for both the EA and DED tasks. Extensive experiments demonstrate that in the EA task, UED achieves EA results compara- ble to those of state-of-the-art supervised EA baselines and outperforms the current state-of- the-art EA methods by combining supervised EA data. For the DED task, UED obtains high- quality results without supervision. 1 Introduction Entity alignment (EA) that aligns the equivalent entities in different knowledge graphs (KGs) is a fundamental technique for knowledge graph inte- gration. A typical application of EA is constructing a large-scale KG by integrating different KGs to facilitate various downstream tasks such as ques- tion answering (Savenkov and Agichtein, 2016; Yu et al., 2017; Jin et al., 2022), recommendation (Cao et al., 2019), and search engines (Xiong et al., 2017). The existing embedding-based EA methods align each entity to its closest counterpart cross KGs according to entity embeddings. In recent years, they have emerged as the dominant EA so- lutions due to their effectiveness and strong ability to utilize information such as entity name strings, entity description, attributes, and graph structure. These EA methods (Chen et al., 2017; Sun et al., 2018; Wang et al., 2018; Zhu et al., 2021a; Liu et al., 2021; Lin et al., 2021) are built upon the assumption that there exists a counterpart in the target KG for any source entity (Sun et al., 2021). Therefore, ideally, their performances are assessed by only considering the entities in the set of testing entity pairs. In the real-world scenario, four facts should be considered when aligning KGs: (1) The entities that do not have counterparts in another KG are ubiquitous. These entities are referred to as dan- gling entities, following Sun et al. (2021). There- fore, it is necessary to identify the dangling entities and then align the remaining matchable entities to their counterparts. The widely used approach of in- tegrating KGs according to the cross KG similarity between entities loses sight of identifying dangling entities. (2) Dangling entity sets are not labeled in most cases, while …

Content truncated to 500 words