Journal of Frontiers of Computer Science and Technology ›› 2023, Vol. 17 ›› Issue (2): 467-477.DOI: 10.3778/j.issn.1673-9418.2205085

• Artificial Intelligence·Pattern Recognition • Previous Articles     Next Articles

Knowledge Graph Completion Method Based on Multi-semantic Relation Embedding

YIN Hua, XIAO Shiran, CHEN Zhiquan, HU Zhensheng, LONG Yongchao   

  1. 1. School of Information, Guangdong University of Finance & Economics, Guangzhou 510320, China
    2. Guangdong Intelligent Commerce Engineering Technology Research Center, Guangzhou 510320, China
    3. School of Computer Science and Engineering, Sun Yat-Sen University, Guangzhou 510006, China
  • Online:2023-02-01 Published:2023-02-01



  1. 1. 广东财经大学 信息学院,广州 510320
    2. 广东省智能商务工程技术研究中心,广州 510320
    3. 中山大学 计算机科学与工程学院,广州 510006

Abstract: The knowledge graph completion method based on knowledge representation transforms entities and relationships into distributed representation, and completes missing relation by vector calculation. The existing knowledge graph representation model regards the relationship as a single vector and loses part of the relationship semantics. The traditional relational multi-semantic segmentation model is difficult to be applied in large-scale knowledge completion because of many parameters and high time consumption. In this paper, a relationship completion method of knowledge graph based on multi-semantic relationship embedding (MSRE) is proposed. The relationship angle vector is inversely calculated in complex domain space. And the semantic component cluster of each relationship is constructed based on Mean-Shift algorithm. The RotatE score function is optimized to obtain the most appropriate score of relational semantic component in the cluster. This method not only expands the relational representation, but also ensures the uniqueness in triplet operation. Experimental results of link prediction and triple classification on public datasets FB15K-237 and WN18RR show that this method can mine the potential semantics of relationships, maintain low time complexity, and improve the performance compared with mainstream models in most indicators.

Key words: knowledge graph, relation multiple semantics, relation embedding, clustering

摘要: 基于知识表示的知识图谱补全方法将实体与关系转化为分布式向量,通过向量计算补全缺失关系。现有的知识表示模型将关系看作单一向量,损失了部分关系语义。而传统关系多语义细分模型由于参数较多,时耗较大难以在大规模知识图谱上应用。提出了一种多语义关系嵌入的知识图谱补全方法(MSRE),在复数域空间中反向计算关系角度向量,基于Mean-Shift构建各关系的语义分量簇,优化RotatE得分函数为语义分量簇中最恰当的关系语义分量得分。该方法在扩充关系表示的同时,保证了三元组运算中的唯一性。在公开数据集FB15K-237、WN18RR上的链路预测和三元组分类的实验结果表明,该方法可以挖掘关系的潜在语义,保持较低的时间复杂度,且在多数指标上相较于主流模型有一定的性能提升。

关键词: 知识图谱, 关系多语义, 关系嵌入, 聚类