基于四元数胶囊网络的知识图谱补全模型

知识图谱采用RDF三元组的形式描述现实世界中的关系和头、尾实体，即（头实体，关系，尾实体）或（主语，谓语，宾语）。为补全知识图谱中缺失的事实三元组，将四元数融入胶囊神经网络模型预测缺失的知识，并构建一种新的知识图谱补全模型。采用超复数嵌入取代传统的实值嵌入来编码三元组结构信息，以尽可能全面捕获三元组全局特性，将实体、关系的四元数嵌入作为胶囊网络的输入，四元数结合优化的胶囊网络模型可以有效补全知识图谱中丢失的三元组，提高预测精度。链接预测实验结果表明，与CapsE模型相比，在数据集WN18RR中，该知识图谱补全模型的Hit@10与正确实体的倒数平均排名分别提高3.2个百分点和5.5%，在数据集FB15K-237中，Hit@10与正确实体的倒数平均排名分别提高2.5个百分点和4.4%，能够有效预测知识图谱中缺失的事实三元组。

Knowledge Graph Completion Model Based on Quaternion Capsule Network

Knowledge Graph(KG) uses RDF triplets to describe relationships as well as head and tail entities in the real world, namely (head entity, relationship, tail entity) or (subject, predicate, object).In order to complete the missing fact triplets in knowledge graphs, quaternion is integrated into the capsule neural network model to predict missing knowledge, and on this basis a new knowledge graph completion model is proposed.In order to capture the global features of triplets as much as possible, hypercomplex embedding is used to encode triplet structure information instead of traditional real-valued embedding.In this paper, the quaternion embedding of entities and relationships is used as the input of the capsule network.Quaternion combined with the optimized capsule network model can effectively complete the missing triplets in the knowledge map and improve the prediction accuracy.The proposed model is tested in a link prediction experiment and compared with the CapsE model.Results show that on the WN18RR dataset, the proportion of entities correctly completed by the proposed model is 3.2 percentage points higher than CapsE in Hit@10 and 5.5% higher in the reciprocal average ranking.On the FB15K-237 dataset, the proportion of entities correctly completed by the proposed model is 2.5 percentage points higher in Hit@10 and 4.4% higher in the reciprocal average ranking.The model can effectively predict the missing fact triplets in a knowledge graph.