PANC: Prototype Augmented Neighbor Constraint instance completion in knowledge graphs

Abstract

Great attention has been devoted to knowledge graph completion techniques with the wide application of knowledge graph. Previous works on knowledge graph completion mainly consider predicting missing one element for given two elements in a triplet, such as (h, r, ?). However, this task lacks of reasonable consideration between entities and relations, such as (Italy, acted in, ?), which may cause the meaningless predictions. Recent attempt solves this problem by redefining a new task-instance completion, which generates and evaluates reasonable relation-tail pairs for given head entity, such as (h, ?, ?). In this work, we propose a novel Prototype Augmented Neighbor Constraint instance completion model called PANC, which consists of two modules-prototype filter and neighbor aggregation grader. A kind of coarse-grained information prototype is utilized in filters to generate more candidate relation-tail pairs and neighbor aggregation is introduced into grader so as to enhance entity embedding and constrain the combination between head entity and candidate pairs. The experiments show that our PANC outperforms the state-of-the-art instance completion techniques on two real-world datasets FB15k and JF17k. And the ablation results verify the effectiveness of the modules in our proposed PANC.