Research On Knowledge Representation Learning Based On Rotations In Three-dimensional Space

The knowledge graph is a structured representation of the information of the objective world and has been widely used.There are many tasks related to the problems and applications of knowledge graphs.Knowledge graph completion,also known as link prediction,aims to predict missing links in the knowledge graph.Path query answering aims to give the correct answer to the query on the knowledge graph.Rule mining aims to mine easy-to-understand rules based on existing facts in the knowledge graph.Carrying out these tasks requires the model to have good reasoning ability.Knowledge representation learning,which aims to learn low-dimensional embeddings of entities and relations in the knowledge graph,is an effective way to deal with the above tasks and has drawn great attention.The reasoning ability of knowledge representation learning models heavily relies on their ability to model various relation patterns in the knowledge graph.Mainly,there are three types of relation patterns in the knowledge graph,i.e.,symmetry/antisymmetry,inversion,and composition.The composition pattern can be further divided into the commutative composition pattern and the non-commutative composition pattern.However,most existing models fail to model the non-commutative composition pattern.To address this issue,this paper proposes a model based on rotations in the three-dimensional space called Rotate3 D,which maps entities to the three-dimensional space and defines relations as rotations from head entities to tail entities.By using the non-commutative composition property of rotations in the three-dimensional space,Rotate3 D can naturally preserve the order of the composition of relations.The theoretical analysis shows that Rotate3 D can model all the main relation patterns in the knowledge graph,which provides a guarantee for the reasoning ability of Rotate3 D.This paper verifies the effectiveness of the Rotate3 D model on link prediction,path query answering,and rule mining tasks.Experiments on typical public datasets show that Rotate3 D outperforms existing state-of-the-art models for link prediction and path query answering,which demonstrates the superior reasoning ability of Rotate3 D.Experimental results on rule mining show that the Rotate3D-based rule mining algorithm is highly scalable and can mine reasonable and high-quality rules.Moreover,in order to better understand the model,based on the results of the theoretical analysis,this paperconducts case studies of inferring the three main relation patterns through visualization,intuitive examples,and quantitative analysis.Case studies demonstrate that Rotate3 D can effectively model these relation patterns with a marked improvement in modeling the composition pattern.