DEER: Descriptive Knowledge Graph for Explaining Entity Relationships

Proceedings of the 2022 Conference on Empirical Methods in Natural Language Processing, pages 6686 - 6698 December 7-11, 2022 ©2022 Association for Computational Linguistics DEER : Descriptive Knowledge Graph for Explaining Entity Relationships Jie Huang∗,1 Kerui Zhu∗,1 Kevin Chen-Chuan Chang1 Jinjun Xiong2 Wen-mei Hwu1,3 1University of Illinois at Urbana-Champaign, USA 2University at Buffalo, USA 3NVIDIA, USA {jeffhj, keruiz2, kcchang, w-hwu}@illinois.edu jinjun@buffalo.edu Abstract We propose DEER (Descriptive Knowledge Graph for Explaining Entity Relationships) – an open and informative form of modeling en- tity relationships. In DEER, relationships be- tween entities are represented by free-text rela- tion descriptions. For instance, the relationship between entities of machine learning and algo- rithm can be represented as “Machine learning explores the study and construction of algo- rithms that can learn from and make predic- tions on data.” To construct DEER, we pro- pose a self-supervised learning method to ex- tract relation descriptions with the analysis of dependency patterns and generate relation de- scriptions with a transformer-based relation de- scription synthesizing model, where no human labeling is required. Experiments demonstrate that our system can extract and generate high- quality relation descriptions for explaining en- tity relationships. The results suggest that we can build an open and informative knowledge graph without human annotation.1 1 Introduction Relationships exist widely between entities. For example, a person may be related to another person or an institution, and a scientific concept can be connected to another concept. At the same time, relationships between entities can be subtle or com- plex, e.g., the relationship between machine learn- ing and algorithm. To model relationships between entities, re- searchers usually construct knowledge graphs (KGs) (Ji et al., 2021; Hogan et al., 2021), where nodes are entities, e.g., machine learning, and edges are relations, e.g., subclass of (Figure 2). However, KGs usually require a pre-specified set of relation types, and the covered relation types are usually coarse-grained and simple. This indi- cates existing KGs lack two desired features. The 1Code and data are available at https://github.com/ jeffhj/DEER. ∗Asterisk indicates equal contribution. first is openness: for entities with a relationship not covered by the type set, KGs cannot handle their relationship directly. Besides, in many cases, the relationship between entities is complex or id- iosyncratic that it cannot be simply categorized to a relation type. For instance, for related entities ma- chine learning and algorithm, Wikidata (Vrandeˇci´c and Krötzsch, 2014) does not include a relation for them, and it is also not easy to come up with a relation type to describe their relationship. The second feature is about informativeness. With the relational facts in KGs, humans may still have difficulty in understanding entity relationships. For instance, from fact “(data mining, facet of, database)” in Wikidata, humans may guess data mining and database are related fields, but they cannot understand how exactly they are related, e.g, why is it a facet? and what is the facet? Although techniques like knowledge graph rea- soning (Lao et al., 2011; Xiong et al., 2017; Chen et al., 2018) or …

Content truncated to 500 words