A Convolution Quaternion-Based Inverse Relational Rotation Embedding Knowledge Representation Method

Abstract

Knowledge graphs have been widely used in various fields, but knowledge graphs are usually incomplete, so it is necessary to perform reasoning based on multiple relational patterns to complete graph completion. The RotatE model realizes the modeling and reasoning of multiple relational patterns in the complex space. However, the RotatE model has the problems of insufficient connection between different entity vectors and the low flexibility of rotation between entities and relations. Therefore, we propose a convolution quaternion-based inverse relational rotation embedding knowledge representation method (cQuaIE). On the basis of RotatE, the correlation between the head entity and the tail entity is enhanced by introducing an inverse relation vector. And utilize a more expressive quaternion representation to model entities and relations. Finally, a convolutional neural network is used to process the input quaternion embedding vector to improve the accuracy of model link prediction. Experiments show that the model outperforms existing majority knowledge representation models in the link prediction task.