Article Preview
TopIntroduction
Knowledge graphs, such as WordNet (Miller, 1995), Freebase (Bollacker et al., 2008), and DBpedia (Lehmann et al., 2015), store a great number of knowledge triples in the form of directed graphs, where each node denotes an entity, and each edge denotes a relation. A knowledge triple (head entity, relation, tail entity) (denoted by (, , ) in this paper) stands for an edge with two end nodes in the graph, indicating that there exists a specific relationship between the head and tail entities. Recently, applications of knowledge graphs in artificial intelligence domains, including information systems (Saghafi & Wand, 2020), recommender systems (Zhang et al., 2016), information retrieval (Xiong et al., 2017), question-answering systems (Hao et al., 2017; Huang et al., 2019), and natural language processing (Yang & Mitchell, 2019), have gained more and more attention. There often exist multiple knowledge graphs in overlapped application domains or even the same domain. Moreover, some knowledge graphs are created with crowdsourcing (Chai et al., 2018), due to the uncertainty of crowdsourcing annotation (Zhang, Wu, & Sheng, 2016), they are generally incomplete, resulting in difficulties for a single knowledge graph to meet the various requirements of many AI applications. Integrating heterogeneous knowledge among different knowledge graphs through entity alignment can make knowledge graphs satisfy wider requirements in different applications. However, this goal is not easy to achieve. Because entities and relations in different knowledge graphs may be represented by different languages or symbols, it is generally ineffective for symbolic-based methods (Ngomo & Auer, 2011; Volz et al., 2009) to precisely align entities. For example, “https://www.wikidata.org/entity/ Q7769108” in Wikidata are the same entity in the real world, it is almost impossible to measure the similarity between them by symbolic-based methods. In order to solve this difficult problem, researchers have proposed quite a few embedding-based entity alignment approaches. Such approaches first embed symbolic entities and relations of a knowledge graph into a low-dimensional vector space and then find out entity alignments by measuring the similarities between the corresponding embeddings.
Generally, knowledge graph embedding models focus on representing a single knowledge graph as vectors to capture the semantics of entities and relations according to their structural information in the knowledge graph. For example, as one of the most popular knowledge graph embedding models in entity alignment methods, TransE (Bordes et al., 2013) interprets each relation as a translation from the head entity to the tail entity. However, such separately-trained embeddings cannot be directly used to compute the similarity between entities from different knowledge graphs, because embeddings of different knowledge graphs are in different semantic spaces (i.e., vector space).