A Survey on Knowledge Graph Embedding

Abstract

Knowledge graph (KG) is used to represent the relationships between different concepts in the real world. It is a special network in which nodes represent entities and edges represent relationships. KGs can intuitively model the connections between facts, but in many applications, there are certain limitations in directly using symbolic logic to represent knowledge in KGs and perform calculations, making it difficult to achieve expected results in downstream tasks. Meanwhile, with the explosive growth of Internet capacity, the traditional KG structure faces the problems of computational inefficiency and management difficulties. To alleviate these problems, Knowledge graph embedding (KGE) is proposed to improve the computational efficiency by embedding entities and relations in the KG into a low-dimensional, dense and continuous vector space, and thus the solution of some problems in the knowledge graph is transformed into vector operations. Moreover, KGE also can be used as a pre-trained model which is more beneficial to downstream applications, such as applications based on deep learning. In this paper, we classify KGE into three categories, namely translational distance models, semantic matching models and neural network based models. The advantages and disadvantages of different embedding methods are compared, while the main applications of KGE are summarized. Some current challenges of KGE are summarized, as well as some views on the future research directions of KGE.

Topics & Concepts