Decoupling Mixture-of-Graphs: Unseen Relational Learning for Knowledge Graph Completion by Fusing Ontology and Textual Experts

Proceedings of the 29th International Conference on Computational Linguistics, pages 2237–2246 October 12–17, 2022. 2237 Decoupling Mixture-of-Graphs: Unseen Relational Learning for Knowledge Graph Completion by Fusing Ontology and Textual Experts Ran Song1,2, Shizhu He3,4, Suncong Zheng6, Shengxiang Gao1,2 ∗, Kang Liu3,4,5, Zhengtao Yu1,2, and Jun Zhao3,4 1 Faculty of Information Engineering and Automation, Kunming University of Science and Technology 2 Yunnan Key Laboratory of Artificial Intelligence, Kunming, China 3 Institute of Automation, Chinese Academy of Sciences, Beijing, China 4 School of Artificial Intelligence, University of Chinese Academy of Science, Beijing, China 5Beijing Academy of Artificial Intelligence, Beijing, China 6 Tencent AI Lab {song_ransr,junzhao123}@163.com, {shizhu.he,kliu}@nlpr.ia.ac.cn, congzheng@tencent.com, {gaoshengxiang.yn,ztyu}@hotmail.com Abstract Knowledge Graph Embedding (KGE) has been proposed and successfully utilized for knowl- edge Graph Completion (KGC). But classic KGE paradigm often fail in unseen relation representations. Previous studies mainly uti- lize the textual descriptions of relations and its neighbor relations to represent unseen relations. In fact, the semantics of a relation can be ex- pressed by three kinds of graphs: factual graph, ontology graph, textual description graph, and they can complement each other. A more com- mon scenario in the real world is that seen and unseen relations appear at the same time. In this setting, the training set (only seen relations) and testing set (both seen and unseen relations) own different distributions. And the train-test inconsistency problem will make KGE meth- ods easily overfit on seen relations and under- performance on unseen relations. In this paper, we propose decoupling mixture-of-graph ex- perts (DMoG) for unseen relations learning, which could represent the unseen relations in the factual graph by fusing ontology and tex- tual graphs, and decouple fusing space and rea- soning space to alleviate overfitting for seen relations. The experiments on two unseen-only public datasets and a mixture dataset verify the effectiveness of the proposed method, which improves the state-of-the-art methods by 6.84% in Hits@10 on average. 1 Introduction Knowledge Graphs (KGs) such as Freebase (Bol- lacker et al., 2008), DBpedia (Lehmann et al., 2015) and YAGO (Mahdisoltani et al., 2014) contain large amounts of entities, relations and facts, which can be used to support many NLP tasks. Knowledge- dependent tasks rely heavily on the coverage of KGs. And the incompleteness of those KGs is an ∗ Corresponding author Scientist University of Graz Student Property coparticipatesWith Factual Graph Ontology Graph Unseen Relation Seen Relation academicAdvisor Nernst Boltzmann Gilbert Textual Graph A scientist is a person who conducts scientific research…. A student is prima rily a person enrol led in a school … An academic advis or is a professional staff …. A person is a being that has certain cap acities…. Berkeley, California Training Testing Niwica Figure 1: The semantics of a relation in a KG is ex- pressed by three kinds of graphs: factual graph, ontol- ogy graph and textual graph. And the knowledge graph completion involving unseen relations in factual graph, which struggle in previous KGE methods, could be alle- viated by utilizing their ontology and textual graphs. urgent issue …

Content truncated to 500 words