Open Hierarchical Relation Extraction

Proceedings of the 2021 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, pages 5682–5693 June 6–11, 2021. ©2021 Association for Computational Linguistics 5682 Open Hierarchical Relation Extraction Kai Zhang1∗, Yuan Yao1∗, Ruobing Xie2, Xu Han1, Zhiyuan Liu1† , Fen Lin2, Leyu Lin2, Maosong Sun1 1Department of Computer Science and Technology Institute for Artificial Intelligence, Tsinghua University, Beijing, China Beijing National Research Center for Information Science and Technology, China 2WeChat Search Application Department, Tencent, China drogozhang@gmail.com, yaoyuanthu@163.com Abstract Open relation extraction (OpenRE) aims to ex- tract novel relation types from open-domain corpora, which plays an important role in completing the relation schemes of knowl- edge bases (KBs). Most OpenRE methods cast different relation types in isolation with- out considering their hierarchical dependency. We argue that OpenRE is inherently in close connection with relation hierarchies. To ad- dress the bidirectional connections between OpenRE and relation hierarchy, we propose the task of open hierarchical relation extrac- tion and present a novel OHRE framework for the task. To effectively integrate hierarchy in- formation into relation representations for bet- ter novel relation extraction, we propose a dy- namic hierarchical triplet objective and hierar- chical curriculum training paradigm. We also present a top-down hierarchy expansion algo- rithm to add the extracted relations into exist- ing hierarchies with reasonable interpretability. Comprehensive experiments show that OHRE outperforms state-of-the-art models by a large margin on both relation clustering and hierar- chy expansion. The source code and experi- ment details of this paper can be obtained from https://github.com/thunlp/OHRE. 1 Introduction Open relation extraction (OpenRE) aims to extract novel relations types between entities from open- domain corpora, which plays an important role in completing the relation schemes of knowledge bases (KBs). OpenRE models are mainly cate- gorized into two groups, namely tagging-based and clustering-based methods. Tagging-based methods consider OpenRE as a sequence labeling ∗ indicates equal contribution † Corresponding author: Z.Liu (liuzy@tsinghua.edu.cn) significant person relative father participant winner Training instances … Representation Learning Test instances … OHRE child spouse Relation Clustering Novel relations … Hierarchy Expansion Figure 1: The workflow of OHRE framework. Trained with relation hierarchy and labeled instances, OHRE extracts novel relations from open-domain corpora and adds them into the existing hierarchy. task, which extracts relational phrases from sen- tences (Banko et al., 2007; Cui et al., 2018). In contrast, clustering-based methods aim to cluster relation instances into groups based on their se- mantic similarities, and regard each cluster as a relation (Yao et al., 2011; Wu et al., 2019). However, most OpenRE models cast different re- lation types in isolation, without considering their rich hierarchical dependencies. Hierarchical orga- nization of relations has been shown to play a cen- tral role in the abstraction and generalization ability of human (Tenenbaum et al., 2011). This hierar- chical organization of relations also constitutes the foundation of most modern KBs (Auer et al., 2007; Bollacker et al., 2008). Figure 1 illustrates an ex- ample of relation hierarchy in Wikidata (Vrandeˇci´c and Krötzsch, 2014). Such relation …

Content truncated to 500 words