Inductive Relation Prediction with Logical Reasoning Using Contrastive Representations

Proceedings of the 2022 Conference on Empirical Methods in Natural Language Processing, pages 4261 - 4274 December 7-11, 2022 ©2022 Association for Computational Linguistics Inductive Relation Prediction with Logical Reasoning Using Contrastive Representations Yudai Pan1,2 Jun Liu1,2∗ Lingling Zhang1,3 Tianzhe Zhao1 Qika Lin1 Xin Hu1 and Qianying Wang4 1School of Computer Science and Technology, Xi’an Jiaotong University 2National Engineering Lab for Big Data Analytics 3Shaanxi Provincial Key Laboratory of Big Data Knowledge Engineering 4Lenovo Research, Beijing, China {pyd418,qikalin,dr.huxin}@foxmail.com, {liukeen,zhanglling}@xjtu.edu.cn ztz9712265887@stu.xjtu.edu.cn, wangqya@lenovo.com Abstract Relation prediction in knowledge graphs (KGs) aims at predicting missing relations in incomplete triples, whereas the dominant em- bedding paradigm has a restriction on handling unseen entities during testing. In the real- world scenario, the inductive setting is more common because entities in the training pro- cess are finite. Previous methods capture an in- ductive ability by implicit logic in KGs. How- ever, it would be challenging to preciously ac- quire entity-independent relational semantics of compositional logic rules and to deal with the deficient supervision of logic caused by the scarcity of relational semantics. To this end, we propose a novel graph convolutional network (GCN)-based model LogCo with log- ical reasoning by contrastive representations. LogCo firstly extracts enclosing subgraphs and relational paths between two entities to supply the entity-independence. Then a contrastive strategy for relational path instances and the subgraph is proposed for the issue of deficient supervision. The contrastive representations are learned for a joint training regime. Finally, prediction results and logic rules for reason- ing are attained. Comprehensive experiments on twelve inductive datasets show that LogCo achieves outstanding performance comparing with SOTA inductive baselines. 1 Introduction Knowledge graphs (KGs) store plenty of facts by triples consisting of entities and relations. They have been widely used in different application sce- narios, such as relation extraction (Hu et al., 2021), question answering (Abdelaziz et al., 2021) and information retrieval (Verlinden et al., 2021). Ex- tracting triples from contexts consumes plenty of re- sources. Therefore, some dominant methods aim at the KG completion or relation prediction by learn- ing representations of relations and entities, such ∗ Corresponding author as TransE(Bordes et al., 2013), RESCAL (Nickel et al., 2011), R-GCN (Schlichtkrull et al., 2018) and CompGCN (Vashishth et al., 2020). However, the above methods predict relations assuming for a transductive setting, which means the entities are fixed during training and testing. In the application scenario, there will be new enti- ties during testing. For the scenario in Figure 1(a), training and testing entities have no intersection, so the previous transductive methods will not accu- rately predict the relation liveIn between entities Bill Gates and W.A. in the test set without retraining the whole model. Thus, some studies like GraiL (Teru et al., 2020) focus more on models owning inductive ability, which can handle unseen entities by implicit first-order logic rules (Horn, 1951) fit- ting the cognition of human beings. For example, by the following logic rule r1: workIn(X, Z) ∧ locatedIn(Z, Y ) → liveIn(X, Y ), the relation liveIn …

Content truncated to 500 words