摘要
该文基于胶囊神经网络出色的维度信息挖掘能力,加入多尺度卷积以进一步增强其特征提取和交互能力,提出了基于多尺度卷积的胶囊网络知识图谱嵌入模型.首先,通过TransE算法训练得到实体和关系的初始化嵌入向量;其次,通过多尺度卷积生成不同的特征图,将得到的特征图进行特征融合,融合后得到的特征图重组为相对应的胶囊;最后,利用动态路由指定从第一层胶囊到第二层胶囊的连接,经过路由得到的第二层胶囊利用squash函数得到最终向量长度,该向量长度决定三元组的置信度.知识图谱链接预测任务的实验结果表明,较嵌入模型CapsE,本文提出的模型在WN18RR数据集上指标Hit@10提高1.8%,MRR提高1.4%,在FB15k-237数据集上Hit@10提高2.2%,MR提高4.8%.
This paper is based on the excellent dimensional information mining ability of capsule neural networks,and incorporates multi-scale convolution to further enhance their feature extraction and interaction capabilities.A capsule network knowledge graph embedding model based on multi-scale convolution is proposed.Firstly,the initialization embedding vectors for entities and relationships are trained using the TransE algorithm.Secondly,different feature maps are generated through multi-scale convolution,and the resulting feature maps are fused to form corresponding capsules.Finally,dynamic routing is used to specify the connection from the first layer capsule to the second layer capsule.The second layer capsule obtained through routing is then used to obtain the final vector length using the squash function,which determines the confidence level of the triplet.Compared with the embedding models CapsE,the proposed model in this paper has a 1.8%improvement in Hit@10 and 1.4%improvement in MRR metrics on the WN18RR dataset,and a 2.2%improvement in Hit@10 and 4.8%improvement in MR on the FB15k-237 dataset.
作者
周淑霄
王艳娜
周子力
王妍
董兆安
ZHOU Shuxiao;WANG Yanna;ZHOU Zili;WANG Yan;DONG Zhaoan(School of Cyber Science and Engineering,273165,Qufu;School of Computer Science,Qufu Normal University,276826,Rizhao,Shandong,PRC)
出处
《曲阜师范大学学报（自然科学版）》
CAS
2024年第2期93-99,共7页
Journal of Qufu Normal University(Natural Science)
基金
山东省自然科学基金(ZR2020MF149)
山东省高校科技计划(J18KB161)
教育部产学合作协同育人项目(202102291003)。
关键词
知识图谱
多尺度卷积
胶囊网络
知识图谱嵌入
神经网络
knowledge graph
multi-scale convolution
capsule network
knowledge graph embedding
neural network
