Multi-view Classification Model for Knowledge Graph Completion

Proceedings of the 1st Conference of the Asia-Pacific Chapter of the Association for Computational Linguistics and the 10th International Joint Conference on Natural Language Processing, pages 726–734 December 4 - 7, 2020. c⃝2020 Association for Computational Linguistics 726 Multi-view Classification Model for Knowledge Graph Completion Wenbin Jiang 1, Mengfei Guo 2∗, Yufeng Chen 2, Ying Li 1, Jinan Xu 2, Yajuan Lyu 1, Yong Zhu 1 1Baidu Inc., Beijing, China 2School of Computer and Information Technology, Beijing Jiaotong University, Beijing, China {jiangwenbin, nicole, lvyajuan, zhuyong}@baidu.com {guomengfei, jaxu, chenyf}@bjtu.edu.cn Abstract Most previous work on knowledge graph completion conducted single-view prediction or calculation for candidate triple evaluation, based only on the content information of the candidate triples. This paper describes a novel multi-view classification model for knowledge graph completion, where multiple classifica- tion views are performed based on both con- tent and context information for candidate triple evaluation. Each classification view eval- uates the validity of a candidate triple from a specific viewpoint, based on the content in- formation inside the candidate triple and the context information nearby the triple. These classification views are implemented by a uni- fied neural network and the classification pre- dictions are weightedly integrated to obtain the final evaluation. Experiments show that, the multi-view model brings very significant improvements over previous methods, and achieves the new state-of-the-art on two rep- resentative datasets. We believe that, the flexi- bility and the scalability of the multi-view clas- sification model facilitates the introduction of additional information and resources for better performance. 1 Introduction Knowledge graph (KG) is a typical kind of graph- structured knowledge base (KB). Nowdays, there exist many famous KGs such as YAGO (Suchanek et al., 2007), Freebase (Bollacker et al., 2008) and DBpedia (Lehmann et al., 2015). Large-scale KGs are widely used in many applications such as semantic searching (Kasneci et al., 2008; Schuh- macher and Ponzetto, 2014; Xiong et al., 2017), question answering (Zhang et al., 2016; Hao et al., 2017) and machine reading (Yang and Mitchell, ∗Joint first author. Guo participated in the optimization of this work during the internship in Baidu. 2017). A KG contains a set of triples indicating facts, each of which is composed of a head entity, a tail entity, and a relation indicating the relationship between the two entities. It is nearly impossible to collect a complete set of facts or triples for a KG, especially in open domains. In fact, many valuable valid triples are missing even for the existing well- built large-scale KGs such as Freebase (Socher et al., 2013; West et al., 2014). Many researchers devote their efforts to the problem of knowledge graph completion (KGC), the core operation of which is to evaluate the validity of candidate triples. Previous work on KGC mainly include two groups, embedding-based methods and classification-based methods. Embedding-based models learn embeddings for entities and relations, and evaluate candidate triples based on the embed- dings and specific distance metrics. Representative models include TransE (Bordes et al., 2013) and its extensions (Wang et al., …

Content truncated to 500 words