《计算机应用研究》|Application Research of Computers

一种改进的自适应知识图谱嵌入式表示方法

Improved adaptive embedding method for knowledge graph representation

免费全文下载 (已被下载 次)  
获取PDF全文
作者 孟小艳,蒋同海,周喜,韩云飞,马博
机构 1.中国科学院新疆理化技术研究所,乌鲁木齐 830011;2.中国科学院大学,北京 100049;3.中国科学院新疆理化技术研究所 新疆民族语音语言信息处理实验室,乌鲁木齐 830011;4.新疆农业大学 计算机与信息工程学院,乌鲁木齐 830052
统计 摘要被查看 次,已被下载
文章编号 1001-3695(2021)01-007-0039-05
DOI 10.19734/j.issn.1001-3695.2019.11.0605
摘要 知识图谱的嵌入式表示方法以基于翻译的TransE最为经典,但在处理复杂关系时存在局限;使用欧氏距离作为得分函数中的度量,每个特征维度以相同的权重参与计算,准确性会受到无关维度的影响,灵活性不高。因此,针对这两个缺陷,提出一种自适应的知识图谱嵌入式表示方法TransAD。利用自适应度量方法更换度量函数,在得分函数中引入对角权重矩阵,为每一个特征维分别赋予权重,增加模型的表示能力。同时受TransD方法的启发,将实体与关系通过动态映射矩阵建立空间投影模型,来增强模型对复杂关系的处理能力,最后将两种优化集成在一个模型中。实验结果表明,新方法TransAD优于Trans(<i>E</i>,<i>H</i>,<i>R</i>,<i>D</i>),在链路预测和三元组分类任务的各项指标上均有提升,有一定的先进性。
关键词 知识图谱; 知识表示学习; 嵌入式表示; 自适应
基金项目 中国科学院STS计划项目(KFJ-STS-QYZD-102)
中国科学院青年创新促进会项目(Y9290802)
中科院西部之光—西部青年学者A类资助项目(2018-XBQNXZ-A-003)
自治区天山青年计划项目(2018Q032)
本文URL http://www.arocmag.com/article/01-2021-01-007.html
英文标题 Improved adaptive embedding method for knowledge graph representation
作者英文名 Meng Xiaoyan, Jiang Tonghai, Zhou Xi, Han Yunfei, Ma Bo
机构英文名 1.The Xinjiang Technical Institute of Physics & Chemistry,Chinese Academy of Sciences,Urumqi 830011,China;2.University of Chinese Academy of Sciences,Beijing 100049,China;3.Xinjiang Laboratory of Minority Speech & Language Information Processing,Chinese Academy of Sciences,Urumqi 830011,China;4.College of Computer & Information Engineering,Xinjiang Agricultural University,Urumqi 830052,China
英文摘要 TransE, the embedding representation method, is the most classic translation-based method. But it also has two defects. One is the limitation in dealing with complex relations. The other one is Euclidean distance is used as a measure in the scoring function and each feature dimension is calculated with the same weight, so the accuracy will be affected by irrelevant dimensions and the flexibility is lower. Therefore, in view of these two defects, this paper proposed an adaptive KG embedding representation method, namely, TransAD. It replaced the measure function and then introduced a diagonal weight matrix into the score function to assign weights to each feature dimension respectively to increase the representation ability of the model. At the same time, inspired by TransD, it built a spatial projection model of entity and established relationship through dynamic mapping matrix to enhance the processing ability of the model for complex relations. Finally, it integrated the two optimizations into the TransAD model. The experimental results show that TransAD is superior to Trans(<i>E</i>, <i>H</i>, <i>R</i>, <i>D</i>) , and it is advanced in various indexes of link prediction and triad classification tasks and has certain advantages.
英文关键词 knowledge graph; knowledge represents learning; embedding presentation; adaptive
参考文献 查看稿件参考文献
 
收稿日期 2019/11/1
修回日期 2020/1/8
页码 39-43
中图分类号 TP391;TP182
文献标志码 A