| 摘 要: 针对医疗知识图谱中因邻域异质性致使实体对齐准确率较低的问题,提出了一种多特征融合的实体对齐方法。该方法运用图注意力网络对不同知识图谱进行建模,通过采样筛选出可体现待对齐实体丰富信息的相邻节点,并组成子图。在此基础上,通过计算子图结构中的交叉图注意力进行实体匹配,并融合关系和实体名称,以丰富长尾实体的特征信息,提升了实体对齐的准确度。实验结果表明,该方法在 DBpedia15KZH-EN (DBpedia15K Chinese-English Cross-lingual Knowledge Graph Dataset)数据集上的 Hits@1指标分别达到了87.25%和85.79%,
优于 MTransE(Multilingual Knowledge Graph Embedding with TransE)、GCN-Align(Graph Convolutional Network for Entity Alignment)、RNM(Relational Neighborhood Matching)等对比模型,充分证明了该方法的有效性。 |
| 关键词: 医疗知识图谱;实体对齐;多特征融合;图注意力网络 |
|
中图分类号: TP391
文献标识码: A
|
|
| Research on Medical Entity Alignment Method Based on Multi-feature Fusion |
|
GUO Zhenhua, SONG Bo
|
(School of Information Science and Technology, Qingdao University of Science and Technology, Qingdao 266061, China)
1603823510@qq.com; f022110013@mails.qust.edu.cn
|
| Abstract: To address the low accuracy of entity alignment caused by neighborhood heterogeneity in medical knowledge graphs, this paper proposes a mult-i feature fusion entity alignment method. The method employs a graph attention network to model different knowledge graphs, sampling and filtering neighboring nodes that reflect rich information of the entities to be aligned, thereby constructing subgraphs. On this basis, cross-graph attention within the subgraph structure is calculated for entity matching, while integrating relational and entity name features to enhance the characterization of long-tail entities, thereby improving alignment accuracy. Experimental results demonstrate that the proposed method achieves Hits@ 1 scores of 87.25% and 85.79% on the DBpedia15KZH-EN (DBpedia 15K ChineseEnglish Cross-lingual Knowledge Graph Dataset), outperforming baseline models such as MTransE (Multilingual Knowledge Graph Embedding with TransE), GCN-Align (Graph Convolutional Network for Entity Alignment), and RNM (Relational Neighborhood Matching). These results fully validate the effectiveness of the proposed approach. |
| Keywords: medical knowledge graph; entity alignment; mult-i feature fusion; graph attention network |
