Tackling Long-Tailed Relations and Uncommon Entities in Knowledge Graph Completion

Proceedings of the 2019 Conference on Empirical Methods in Natural Language Processing and the 9th International Joint Conference on Natural Language Processing, pages 250–260, Hong Kong, China, November 3–7, 2019. c⃝2019 Association for Computational Linguistics 250 Tackling Long-Tailed Relations and Uncommon Entities in Knowledge Graph Completion ∗ Zihao Wang1, Kwun Ping Lai1, Piji Li12, Lidong Bing3, Wai Lam1, 1The Chinese University of Hong Kong 2Tencent AI Lab, Shenzhen, China 3R&D Center Singapore, Machine Intelligence Technology, Alibaba DAMO Academy zihaowangbupt@gmail.com {kplai, wlam}@se.cuhk.edu.hk pijili@tencent.com l.bing@alibaba-inc.com Abstract For large-scale knowledge graphs (KGs), re- cent research has been focusing on the large proportion of infrequent relations which have been ignored by previous studies. For exam- ple few-shot learning paradigm for relations has been investigated. In this work, we further advocate that handling uncommon entities is inevitable when dealing with infrequent rela- tions. Therefore, we propose a meta-learning framework that aims at handling infrequent re- lations with few-shot learning and uncommon entities by using textual descriptions. We de- sign a novel model to better extract key in- formation from textual descriptions. Besides, we also develop a novel generative model in our framework to enhance the performance by generating extra triplets during the train- ing stage. Experiments are conducted on two datasets from real-world KGs, and the results show that our framework 1 outperforms pre- vious methods when dealing with infrequent relations and their accompanying uncommon entities. 1 Introduction Modern knowledge graphs (KGs)(Bollacker et al., 2008; Lehmann et al., 2015; Vrandeˇci´c and Kr¨otzsch, 2014) consist of a large number of facts, where each fact is represented as a triplet consist- ing of two entities and a binary relation between them. KGs provide rich information and it has been widely adopted in different tasks, such as question answering (Yih et al., 2015), information extraction (Bing et al., 2017, 2015, 2016) and im- age classification (Marino et al., 2017). However, ∗The work described in this paper is substantially sup- ported by a grant from the Research Grant Council of the Hong Kong Special Administrative Region, China (Project Code: 14204418). 1The implementation of our framework can be found in https://github.com/ZihaoWang/ Few-shot-KGC. KGs still have the issue of incomplete facts. To deal with the problem, Knowledge Graph Com- pletion (KGC) task is introduced to automatically deduce and fill the missing facts. There exist many previous works focusing on this task and embedding-based methods (Bordes et al., 2013; Wang et al., 2014; Trouillon et al., 2016) achieves the best performance among them. Recent works such as (Xiong et al., 2018) have pointed out that relations in KGs follow a long-tailed distribution. To be more precise, a large proportion of relations have only a few facts in KGs. However, previ- ous works of KGC usually focused on small pro- portions of frequent relations and ignored the re- maining ones. One observation is that they of- ten conducted experiments on small datasets such as FB15k and WN18 (Bordes et al., 2013) where a relation typically possesses thousands of facts. Moreover, after analyzing real-world …

Content truncated to 500 words