基于复数空间内矩阵映射的知识表示方法

Knowledge representation using matrix mapping in complex space

为同时建模知识图谱中存在的多对一、多对多等复杂关系以及对称、逆、组合等关系模式,提出一种基于复数空间内矩阵映射的知识表示方法(MMCS方法)。将实体以及实体间关系表示为离散的复数矢量,利用矩阵映射将离散的复数矢量聚簇为相近的矢量,用于表示实体间的多对一、多对多等复杂关系。利用复数空间的旋转策略,表示离散复数矢量之间的对称、逆、组合等关系模式。在公开的数据集上进行相关的链接预测实验,实验结果表明,MMCS在预测三元组中缺失的实体时比其它方法有显著提升。

To simultaneously model the many-to-one,many-to-many and other complex relationships and relationship modes such as symmetry,inverse and composition in the knowledge graph,a knowledge representation method based on matrix mapping in the complex space(MMCS)was proposed.The entities and the inter-entity relationships were represented as discrete complex vectors.The discrete complex vectors were clustered into similar vectors,matrix mapping was used to represent complex relationships between entities,such as many-to-one and many-to-many.The relation patterns including symmetry,inverse and composition between discrete complex vectors were represented through the use of rotation strategy of the complex space.The public datasets were used for the related link prediction experiment.Experimental results show that using MMCS has a significant improvement in predicting missing entities in triples compared to other methods.