English

Journal of Chinese Agricultural Mechanization

Journal of Chinese Agricultural Mechanization ›› 2021, Vol. 42 ›› Issue (9): 136-142.DOI: 10.13733/j.jcam.issn.2095-5553.2021.09.19

Previous Articles     Next Articles

 An improved method for vegetable Chinese knowledge graph representation learning

Du Yaru, Huang Yuan, Gao Xinna, Wu Meng, Li Haijie, Yang Yingru.    

  • Online:2021-09-15 Published:2021-09-15

一种面向蔬菜中文知识图谱表示学习的改进方法

杜亚茹;黄媛;高欣娜;武猛;李海杰;杨英茹;   

  1. 石家庄市农林科学研究院;石家庄市农业信息化工程技术创新中心;河北省都市农业技术创新中心;
  • 基金资助:
    农业部岗位科学家项目(CARS—23—C06)
    石家庄市科学技术研究与发展计划项目(201490074A)

Abstract:  Since the TransE model ignores the semantic differences of vectors of the same class when dealing with the relationship between different categories, it has poor performance in the representation learning of the complex attribute relations of onetomany, manytoone, and manytomany existing in the knowledge graph of the vegetable domain. Based on the TRANSE method, PTA (pathbased TransE for Attribute) model was proposed in this paper to improve it. Firstly, the vegetable domain entities and relationships were mapped into a lowdimensional dense vector space by Word2Vector. Secondly, the attribute relation and the upanddown relation were combined to form the relation path for training. Mean Rank and Hits@10 were used to evaluate the link prediction effect of the presentation learning model. The experimental results showed that compared with the TransE model, the PTA model was better than the TransE model without considering the relation classification, and the Mean Rank was up to 6 orders advanced. In the case that attribute relationships were classified according to complexity, the value of link prediction Hits@10was up to 13% higher than that of the TransE model.

Key words: vegetable, knowledge graph, representation learning, TransE model, PTA model, link prediction

摘要: 由于TransE模型在处理不同类别关系时,忽略了同类别向量的语义差别,在对蔬菜领域知识图谱中存在的一对多、多对一和多对多的复杂属性关系进行表示学习时效果较差。本文在TransE方法的基础上,提出PTA(Path-based TransE for Attribute)模型对其进行改进。首先,将蔬菜领域实体和关系通过word2vector映射到一个低维稠密向量空间中;其次,将复杂的属性关系和上下位关系进行结合构成关系路径进行训练;并采取Mean Rank和Hits@10两个评价指标衡量表示学习模型的链接预测效果。试验结果表明:PTA模型较TransE模型在不考虑关系分类的情况下,链接预测效果有较大的提高,Mean Rank最高提前6个次序;且在对属性关系按照复杂度进行分类的情况下,链接预测Hits@10的值较TransE模型最高提高13%。

关键词: 蔬菜, 知识图谱, 知识表示学习, TransE, PTA模型, 链接预测

CLC Number: