A Flexible Simplicity Enhancement Model for Knowledge Graph Completion Task

Knowledge graph (KG) has gradually become the cornerstone of many Artificial Intelligence (AI) tasks, as one of the most effective ways to represent world knowledge, while these KGs still hardly cover the massive emerging knowledge in the real world. Knowledge Graph Completion (KGC) tries to reason over known facts and infer the missing links, to improve KG’s coverage. The conventional KGC algorithms usually map each entity to a unique embedding vectors, which incurs a linear growth in memory consumption for saving embedded matrices, and results in high computational costs when modeling real-world KG. Hence, this paper aims to investigate how to strengthen the simplicity (i.e., reduce complexity) of KGC model and strike a reasonable balance between accuracy and complexity. Especially, this paper proposes a novel concept-enhanced anchor-based entity representation method to learn a fixed-size vocabulary in condition of the collapsed KG, which is built by the selected anchor entities, concept semantics respect to these anchors and relation types. This work can be viewed as a flexible plug-in unit to serve many current KGC models. Experiments show that our model performs competitively in KGC task while retaining less than 10% of explicit entities in a given KG and less than 10% of parameters.