Shrinking Embeddings for Hyper-Relational Knowledge Graphs

Proceedings of the 61st Annual Meeting of the Association for Computational Linguistics Volume 1: Long Papers, pages 13306–13320 July 9-14, 2023 ©2023 Association for Computational Linguistics Shrinking Embeddings for Hyper-Relational Knowledge Graphs Bo Xiong1, Mojtaba Nayyeri1, Shirui Pan2, Steffen Staab1,3 1University of Stuttgart, 2Griffith University, 3University of Southampton Abstract Link prediction on knowledge graphs (KGs) has been extensively studied on binary rela- tional KGs, wherein each fact is represented by a triple. A significant amount of important knowledge, however, is represented by hyper- relational facts where each fact is composed of a primal triple and a set of qualifiers compris- ing a key-value pair that allows for express- ing more complicated semantics. Although some recent works have proposed to embed hyper-relational KGs, these methods fail to capture essential inference patterns of hyper- relational facts such as qualifier monotonic- ity, qualifier implication, and qualifier mutual exclusion, limiting their generalization capa- bility. To unlock this, we present ShrinkE, a geometric hyper-relational KG embedding method aiming to explicitly model these pat- terns. ShrinkE models the primal triple as a spatial-functional transformation from the head into a relation-specific box. Each qual- ifier “shrinks” the box to narrow down the possible answer set and, thus, realizes quali- fier monotonicity. The spatial relationships be- tween the qualifier boxes allow for modeling core inference patterns of qualifiers such as im- plication and mutual exclusion. Experimental results demonstrate ShrinkE’s superiority on three benchmarks of hyper-relational KGs. 1 Introduction Link prediction on knowledge graphs (KGs) is a central problem for many KG-based applications (Zhang et al., 2016; Lukovnikov et al., 2017; Lu et al., 2023; Xiong et al., 2022b; Chen et al., 2022). Existing works (Sun et al., 2019; Bordes et al., 2013) have mostly studied link prediction on bi- nary relational KGs, wherein each fact is repre- sented by a triple, e.g., (Einstein, educated_at, Uni- versity of Zurich). In many popular KGs such as Freebase (Bollacker et al., 2007), however, a lot of important knowledge is not only expressed in triple- shaped facts, but also via facts about facts, which taken together are called hyper-relational facts. For example, ((Einstein, educated_at, University of Zurich), {(major:physics), (degree:PhD)}) is a hyper-relational fact, where the primary triple (Einstein, educated_at, University of Zurich) is contextualized by a set of key-value pairs {(ma- jor:physics),(degree:PhD)}. Like much other re- lated work, we follow the terminology established for Wikidata (Vrandeˇci´c and Krötzsch, 2014) and use the term qualifiers to refer to the key-value pairs.1 The qualifiers play crucial roles in avoiding ambiguity issues. For instance, Einstein was edu- cated_at several universities and the qualifiers for degree and major help distinguish them. In order to predict links in hyper-relational KGs, pioneering works represent each hyper- relational fact as either an n-tuple in the form of r(e1, e2, · · · , en) (Wen et al., 2016; Zhang et al., 2018; Fatemi et al., 2020; Liu et al., 2020; Abboud et al., 2020) or a set of key-value pairs in the form of {(ki : vi)}m i=1 (Guan …

Content truncated to 500 words