ISSN 1000-1239 CN 11-1777/TP

计算机研究与发展 ›› 2020, Vol. 57 ›› Issue (7): 1460-1471.doi: 10.7544/issn1000-1239.2020.20190643

• 人工智能 • 上一篇    下一篇

基于重排序的迭代式实体对齐

曾维新1,赵翔1,2,唐九阳1,2,谭真1,王炜3   

  1. 1(国防科技大学信息系统工程重点实验室 长沙 410073);2(地球空间信息技术协同创新中心(武汉大学) 武汉 430079);3(新南威尔士大学计算机科学与工程学院 澳大利亚悉尼 2052) (zengweixin13@nudt.edu.cn)
  • 出版日期: 2020-07-01
  • 基金资助: 
    国家自然科学基金项目(61872446,61902417,71690233,71971212);湖南省自然科学基金项目(2019JJ20024);湖南省研究生科研创新项目(CX20190033)

Iterative Entity Alignment via Re-Ranking

Zeng Weixin1, Zhao Xiang1,2, Tang Jiuyang1,2, Tan Zhen1, Wang Wei3   

  1. 1(Science and Technology on Information Systems Engineering Laboratory, National University of Defense Technology, Changsha 410073);2(Collaborative Innovation Center of Geospatial Technology (Wuhan University), Wuhan 430079);3(School of Computer Science and Engineering, The University of New South Wales, Sydney, Australia, 2052)
  • Online: 2020-07-01
  • Supported by: 
    This work was supported by the National Natural Science Foundation of China (61872446, 61902417, 71690233, 71971212), the Natural Science Foundation of Hunan Province of China (2019JJ20024), and the Postgraduate Scientific Research Innovation Project of Hunan Province (CX20190033).

摘要: 现有的知识图谱无法避免地存在不完整这一问题.缓解此问题的可行方法是引入外部知识图谱中的知识.在此过程中,实体对齐是最关键的步骤.当前最先进的实体对齐解决方案主要依靠知识图谱的结构信息来判断实体的等价性,但在真实世界知识图谱上,大部分实体只具有较低的节点度数以及微少的结构信息.此外,标注数据的缺乏也大大限制了实体对齐模型的效果.为解决上述问题,提出将不受节点度数影响的实体名信息与结构信息相结合,从更全面的角度实现实体对齐.在此基本框架上,利用基于课程学习的迭代训练方法从易至难地选择高置信度结果加入到训练数据中,扩增标注数据的规模.最后使用词移距离模型进一步改进实体名信息的利用方式,并对前序对齐结果重排序,提升实体对齐准确率.在跨语言以及单语言实体对齐任务上的实验结果表明,提出的实体对齐方法性能远好于当前最好的方法.

关键词: 实体对齐, 课程学习, 迭代训练, 重排序, 知识图谱对齐

Abstract: Existing knowledge graphs (KGs) inevitably suffer from the problem of incompleteness. One feasible approach to tackle this issue is by introducing knowledge from other KGs. During the process of knowledge integration, entity alignment (EA), which aims to find equivalent entities in different KGs, is the most crucial step, as entities are the pivots that connect heterogeneous KGs. State-of-the-art EA solutions mainly rely on KG structure information for judging the equivalence of entities, whereas most entities in real-life KGs are in low degrees and contain limited structural information. Additionally, the lack of supervision signals also constrains the effectiveness of EA models. In order to tackle aforementioned issues, we propose to combine entity name information, which is not affected by entity degree, with structural information, to convey more comprehensive signals for aligning entities. Upon this basic EA framework, we further devise a curriculum learning based iterative training strategy to increase the scale of labelled data with confident EA pairs selected from the results of each round. Moreover, we exploit word mover’s distance model to optimize the utilization of entity name information and re-rank alignment results, which in turn boosts the accuracy of EA. We evaluate our proposal on both cross-lingual and mono-lingual EA tasks against strong existing methods, and the experimental results reveal that our solution outperforms the state-of-the-arts by a large margin.

Key words: entity alignment, curriculum learning, iterative training, re-ranking, knowledge graph alignment

中图分类号: