No-CapsE:一种基于节点共现的四元数胶囊网络知识图谱补全模型

No-CapsE:A Quaternion Capsule Network Knowledge Graph Completion Model Based on Node Co-occurrence

知识图谱补全是通过预测知识图谱中缺失的事实进行填补,以解决知识图谱中的数据稀疏问题。CapsE、QuatE等用于知识图谱嵌入的模型在链接预测方面已经取得了较好的表现,但是,CapsE模型因其在复数空间进行链接预测与数据挖掘,会受限于其数据维度,使得数据挖掘不够深入,QuatE采用四元数构造超复数平面进行逻辑旋转,但其方法简单,无法有效地构建复杂关系。为此,提出一种改进的胶囊网络补全方法No-CapsE,在超复数平面构建胶囊网络。将数据用四元数进行表示并输入到四元数卷积网络中,输出的特征向量作为胶囊网络的输入,通过点积操作进行评分并依据评分判定三元组的正确性。此外,为了提高模型的训练速度,提出节点共现的思想,将实体和关系都视作节点。在公开数据集FB15K-237与WN18RR上进行链接预测实验,同时为了进一步探究所提模型的性能与效果,进行消融实验。实验结果均表明,相较于对比模型,No-CapsE的知识图谱补全效果更好,可以应用于大规模链接预测任务。

Knowledge graph completion addresses sparse data by predicting and filling in missing facts within the knowledge graph.Several models used for knowledge graph embeddings,such as CapsE and QuatE,have achieved satisfactory link prediction results.However,the capability of the CapsE model to perform link prediction and data mining in complex spaces may be restricted by its data dimensions,resulting in insufficient depth of data mining.QuatE uses quaternions to construct a hypercomplex plane for logical rotation.However,this method is simple and cannot effectively construct complex relationships.Therefore,this study proposes an improved capsule network completion method,No-CapsE,to construct a capsule network in a hypercomplex plane.Specifically,the data are represented by quaternions and input into a quaternion convolutional network.The output feature vectors are used as inputs to the capsule network,and the accuracy of the triplet is evaluated using dot product operations and judged based on the score.This study proposes the concept of node co-occurrence,representing entities and relationships as nodes to enhance model training speed.Finally,link prediction experiments are conducted on the publicly available datasets FB15K-237 and WN18RR.Ablation experiments are designed and conducted to explore further the performance and effectiveness of the model used in this study.The results of both experiments indicate that the No-CapsE knowledge graph completion is more effective and suitable for large-scale linkage prediction tasks.