基于多重得分函数的知识嵌入模型  

Knowledge Embedding Model with Multiple Scoring Functions

作　　者：陈新元 谢晟祎[2] 陈庆强 刘羽 CHEN Xinyuan;XIE Shengyi;CHEN Qingqiang;LIU Yu(Faculty of Engineering, Fuzhou Technology and Business University, Fuzhou, Fujian 350715, China;Experimental Training Center, Fujian Vocational College of Agriculture, Fuzhou, Fujian 350181, China;Information Science and Engineering College, Fujian University of Technology, Fuzhou, Fujian 350118, China;Modern Education Technical Center, Fuzhou Melbourne Polytechnic, Fuzhou, Fujian 350121, China)

机构地区：[1]福州工商学院工学院,福建福州350715 [2]福建农业职业技术学院实验实训中心,福建福州350181 [3]福建工程学院信息科学与工程学院,福建福州350118 [4]福州墨尔本理工职业学院现代教育技术中心,福建福州350121

出　　处：《闽江学院学报》2022年第2期22-32,共11页Journal of Minjiang University

基　　金：福建省中青年教师教育科研项目(JAT210619)。

摘　　要：单一得分函数难以学习多种关系模式;增强表达能力往往意味着较高的模型复杂度和计算开销。在基于平移/旋转的嵌入模型基础上,设计结合多重得分函数的神经网络模型,分析其对多种关系模式的学习和表示能力,在统一框架下为同一实体/关系训练多组独立的嵌入表示以匹配得分函数,同时使用基于阈值限制的损失函数改善模型的泛化能力。在4个主流数据集上进行链路预测,比较方案与其他主流模型的性能表现,验证在嵌入表示中集成不同关系模式信息的可行性,衡量不同得分函数对整体模型的影响。It is difficult for a single scoring function to deal with various relational patterns and to enhance expressivity often means higher model complexity and computational cost.On the basis of translational/rotational embedding models,a neural network model with multiple scoring functions is proposed with discussions on model representation capabilities.In a unified framework,sets of independent embedding representation for the same entity/relation with different functions are trained.At the same time,a loss function based on threshold limit is employed to improve the generalization capacity of the model.Link prediction is carried out on four benchmark datasets to compare our model with other mainstream models.Experimental results show that our model could handle various relational modes;the influence of each scoring function on the integrated model is also demonstrated.

关 键 词：知识表示 平移/旋转嵌入 多重得分函数 模式信息集成 阈值限制 链路预测 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]