.png){kind=icon}
.png){kind=icon}
Return
Semantic- and relation-based graph neural network for knowledge graph completion
.png){kind=icon}
PRE
AI
DOI：10.1007/s10489-024-05482-2.png)
.png){kind=icon}
Original request for help.png){kind=icon}
Share.png){kind=icon}
Save
Abstract
En 中文
Knowledge graph completion (KGC) refines missing entities, relationships, or attributes from a knowledge graph, which is significant for referral systems, biological informatics, and search engines. As an effective KGC approach, a graph neural network (GNN) learns to aggregate information from neighboring nodes by iteratively passing messages between them. However, the semantic and relational information contained in knowledge graphs is rarely used in the existing GNN-based approaches for KGC (i.e., only structure information is used). Hence, a semantic- and relation-based GNN (SR-GNN) model, which combines the semantic similarity information between neighboring entities and the relational features of knowledge graphs, is proposed. First, we develop an entity semantic aggregation module that learns semantic similarity information among neighboring entities connected to the same central entity via an RNN. Second, we propose a relational aggregation module that captures the different semantics among different types of relations through a GRU. This enables the model to better comprehend semantic relationships and be applied to KGC tasks requiring relationship embedding vectors. Extensive studies conducted on the FB15k-237, WN18RR, WN18 and YAGO3-10 datasets reveal that, when compared to 17 baseline models, the SR-GNN exhibits state-of-the-art performance in terms of the MRR and H@n metrics. Significantly, the MRR metric improves by 10.2%\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$10.2 \%$$\end{document} on the FB15K-237 dataset and by 4.2%\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$4.2 \%$$\end{document} on the WN18RR dataset over those of the rival models.
Keywords:
Knowledge graphs
Knowledge graph completion
Knowledge graph embedding
Recurrent neural network