Jointly Learning Entity and Relation Representations for Entity Alignment

Proceedings of the 2019 Conference on Empirical Methods in Natural Language Processing and the 9th International Joint Conference on Natural Language Processing, pages 240–249, Hong Kong, China, November 3–7, 2019. c⃝2019 Association for Computational Linguistics 240 Jointly Learning Entity and Relation Representations for Entity Alignment Yuting Wu1, Xiao Liu1, Yansong Feng1,2∗, Zheng Wang3 and Dongyan Zhao1,2 1Wangxuan Institute of Computer Technology, Peking University, China 2The MOE Key Laboratory of Computational Linguistics, Peking University, China 3School of Computing, University of Leeds, U.K. {wyting,lxlisa,fengyansong,zhaodongyan}@pku.edu.cn z.wang5@leeds.ac.uk Abstract Entity alignment is a viable means for inte- grating heterogeneous knowledge among dif- ferent knowledge graphs (KGs). Recent devel- opments in the field often take an embedding- based approach to model the structural infor- mation of KGs so that entity alignment can be easily performed in the embedding space. However, most existing works do not explic- itly utilize useful relation representations to assist in entity alignment, which, as we will show in the paper, is a simple yet effective way for improving entity alignment. This paper presents a novel joint learning frame- work for entity alignment. At the core of our approach is a Graph Convolutional Network (GCN) based framework for learning both en- tity and relation representations. Rather than relying on pre-aligned relation seeds to learn relation representations, we first approximate them using entity embeddings learned by the GCN. We then incorporate the relation ap- proximation into entities to iteratively learn better representations for both. Experiments performed on three real-world cross-lingual datasets show that our approach substantially outperforms state-of-the-art entity alignment methods. 1 Introduction Knowledge graphs (KGs) transform unstructured knowledge into simple and clear triples of <head entity, relation, tail entity> for rapid response and reasoning of knowledge. They are an effec- tive way for supporting various NLP-enabled tasks like machine reading (Yang and Mitchell, 2017), information extraction (Wang et al., 2018a), and question-answering (Zhang et al., 2018b). Even though many KGs originate from the same resource, e.g., Wikipedia, they are usually created independently. Therefore, different KGs often use ∗Corresponding author. different expressions and surface forms to indicate equivalent entities and relations – let alone those built from different resources or languages. This common problem of heterogeneity makes it diffi- cult to integrate knowledge among different KGs. A powerful technique to address this issue is En- tity Alignment, the task of linking entities with the same real-world identity from different KGs. Classical methods for entity alignment typically involve a labor-intensive and time-consuming pro- cess of feature construction (Mahdisoltani et al., 2013) or rely on external information constructed by others (Suchanek et al., 2011). Recently, efforts have been devoted to the so-called embedding- based approaches. Representative works of this direction include JE (Hao et al., 2016), MTransE (Chen et al., 2017), JAPE (Sun et al., 2017), IP- TransE (Zhu et al., 2017), and BootEA (Sun et al., 2018). More recent work (Wang et al., 2018b) uses the Graph Convolutional Network (GCN) (Kipf and Welling, 2017) to jointly embed multiple KGs. Most of the recent works (e.g., …

Content truncated to 500 words