Deep Reinforcement Learning for Entity Alignment

Findings of the Association for Computational Linguistics: ACL 2022, pages 2754 - 2765 May 22-27, 2022 c⃝2022 Association for Computational Linguistics Deep Reinforcement Learning for Entity Alignment Lingbing Guo1,2,3 and Yuqiang Han1,2 and Qiang Zhang1,2∗and Huajun Chen1,2,3∗ 1 College of Computer Science and Technology, Zhejiang University 2 Hangzhou Innovation Center, Zhejiang University 3 Alibaba-Zhejiang University Joint Reseach Institute of Frontier Technologies {lbguo, hyq2015, qiang.zhang.cs, huajunsir}@zju.edu.cn Abstract Embedding-based methods have attracted in- creasing attention in recent entity alignment (EA) studies. Although great promise they can offer, there are still several limitations. The most notable is that they identify the aligned entities based on cosine similarity, ignor- ing the semantics underlying the embeddings themselves. Furthermore, these methods are shortsighted, heuristically selecting the closest entity as the target and allowing multiple enti- ties to match the same candidate. To address these limitations, we model entity alignment as a sequential decision-making task, in which an agent sequentially decides whether two en- tities are matched or mismatched based on their representation vectors. The proposed re- inforcement learning (RL)-based entity align- ment framework can be flexibly adapted to most embedding-based EA methods. The ex- perimental results demonstrate that it consis- tently advances the performance of several state-of-the-art methods, with a maximum im- provement of 31.1% on Hits@1. 1 Introduction Entity alignment (EA) is one of the most crucial tasks in knowledge graph (KG) studies. It aims to seek the potentially aligned entity pairs between two KGs, such that distributed knowledge can be linked for better supporting downstream applica- tions. Generally, a fact in a KG can be represented by a triplet (e1 x, r1, e1 y), where e1 x, e1 y denote the head and tail entities in the first KG G1. r1 is the relation connecting them. With a small number of known alignment pairs as anchors, embedding- based entity alignment (EEA) methods can learn the representations of entities belonging to respec- tive KGs in a unified space and exploit the underly- ing aligned pairs based on the embedding distance. For example, e2 x will be chosen as target entity for ∗Corresponding authors. Outer Space Jitters Of Cash and Hash Guns a Poppin! Q7112014 Q7078739 Q5619358 Figure 1: Different evaluation strategies. The ranking strategy (left) heuristically selects the candidate with the largest similarity. The sequential strategy (right) al- lows each candidate to be matched only once. Deeper color indicates higher similarity. Diagonals are correct matches. Cells with yellow borders are the selected en- tities, while those with dotted borders denote the ex- cluded entities. e1 e1x if its embedding is closest to the embedding of x in vector space. Although recent EEA methods (Chen et al., 2017; Sun et al., 2017, 2018; Guo et al., 2019; Wu et al., 2019; Tang et al., 2020; Wang et al., 2020; Sun et al., 2020a) have made great performance improvement, they rarely consider the evaluation process. For example, in Figure 1, all three films are directed by Jules White and have similar casts. This makes the EEA methods …

Content truncated to 500 words