Graph Pattern Entity Ranking Model for Knowledge Graph Completion

Proceedings of NAACL-HLT 2019, pages 988–997 Minneapolis, Minnesota, June 2 - June 7, 2019. c⃝2019 Association for Computational Linguistics 988 Graph Pattern Entity Ranking Model for Knowledge Graph Completion Takuma Ebisu1,2 and Ryutaro Ichise2,1,3 1SOKENDAI, Chiyoda-ku, Tokyo, Japan 2National Institute of Informatics, Chiyoda-ku, Tokyo, Japan 3National Institute of Advamced Industrial Science and Technology, Koto-ku, Tokyo, Japan {takuma,ichise}@nii.ac.jp Abstract Knowledge graphs have evolved rapidly in recent years and their usefulness has been demonstrated in many artificial intelligence tasks. However, knowledge graphs often have lots of missing facts. To solve this prob- lem, many knowledge graph embedding mod- els have been developed to populate knowl- edge graphs and these have shown outstand- ing performance. However, knowledge graph embedding models are so-called black boxes, and the user does not know how the informa- tion in a knowledge graph is processed and the models can be difficult to interpret. In this pa- per, we utilize graph patterns in a knowledge graph to overcome such problems. Our pro- posed model, the graph pattern entity ranking model (GRank), constructs an entity ranking system for each graph pattern and evaluates them using a ranking measure. By doing so, we can find graph patterns which are useful for predicting facts. Then, we perform link prediction tasks on standard datasets to eval- uate our GRank method. We show that our ap- proach outperforms other state-of-the-art ap- proaches such as ComplEx and TorusE for standard metrics such as HITS@n and MRR. Moreover, our model is easily interpretable be- cause the output facts are described by graph patterns. 1 Introduction Knowledge graphs can be used to describe real- world relations as facts in a form that a computer can easily process and has been used for many ar- tificial intelligence tasks (Hakimov et al., 2012; Daiber et al., 2013; Bordes et al., 2014). In a knowledge graph, a fact is represented by a la- beled and directed edge, called a triple (h, r, t), where h and t are entity nodes and r is a rela- tion label of an edge from h to t. Knowledge graphs such as YAGO (Suchanek et al., 2007), DBpedia (Auer et al., 2007), and Freebase (Bol- lacker et al., 2008) have developed rapidly in re- cent years and are used for many artificial intel- ligence tasks such as question answering, content tagging, fact-checking, and knowledge inference. Although some knowledge graphs already contain millions of entities and billions of facts, they might still be incomplete and some facts may be missing. Hence, we need to develop a system that can pre- dict missing facts to complete knowledge graphs automatically. Many kinds of models for link prediction have been developed to estimate unknown facts. Knowledge graph embedding models, which are the most widely used approach in this field, map entities and relations in a knowledge graph onto a vector space and obtain the latent underlying fea- tures. However, these models are generally diffi- cult to interpret, as we do not know how informa- tion is processed in …

Content truncated to 500 words