Learning Knowledge Representation with Entity Concept Information

Yuanbo Xu, Lin Yue, Hangtong Xu, Yongjian Yang

Abstract: The goal of Knowledge Representation Learning (KRL) is to learn an accurate knowledge representation that conforms to human understanding. Currently, many works have used entity multi-source information to improve the entity representation semantic precision, such as entity description, attribute, and visual information. However, few methods consider the timeliness in KRL, which significantly affects representation learning performance. In this paper, we attempt to utilize concept information with human understanding to learn an accurate, time-stable knowledge representation. Specifically, we first build a novel Knowledge Graph (KG) - Structure Concept Graph (SCG), which can provide entity structure and concept information jointly. Based on the SCG, we devise a novel KRL model that can embed entity concept information to ensure accuracy and timeliness of improving the KRL’s effect with entity structure information. We evaluate our method on two downstream tasks: the knowledge graph completion task and the zero-shot task. Experimental results on real-world datasets show that our method outperforms other baselines by building effective entities’ representations from their concept information. The source code of this paper can be obtained from https://anonymous.4open.science/r/CKRL-adma2023 .