ISSN 1000-1239 CN 11-1777/TP

计算机研究与发展 ›› 2020, Vol. 57 ›› Issue (6): 1239-1251.doi: 10.7544/issn1000-1239.2020.20190253

• 人工智能 • 上一篇    下一篇



  1. 1(南京航空航天大学计算机科学与技术学院 南京 211106);2(国网辽宁省电力有限公司 沈阳 110004) (
  • 出版日期: 2020-06-01
  • 基金资助: 

Duration-HyTE: A Time-Aware Knowledge Representation Learning Method Based on Duration Modeling

Cui Yuanning1, Li Jing1, Shen Li2, Shen Yang2, Qiao Lin2, Bo Jue2   

  1. 1(School of Computer Science and Technology, Nanjing University of Aeronautics and Astronautics, Nanjing 211106);2(State Grid Liaoning Electric Power Supply Co, LTD, Shenyang 110004)
  • Online: 2020-06-01
  • Supported by: 
    This work was supported by the State Grid Corporation Headquarters Science and Technology Project (SGLNXT00YJJS1800110).

摘要: 知识表示学习是知识获取与应用的基础,是贯穿知识图谱构建与应用全过程的重要问题,伴随含有时间标签的大型知识图谱的发展,近几年时间感知的知识表示学习成为该领域研究热点之一.针对传统方法不能有效学习知识持续时长分布规律的问题,融合超平面和有效持续时间建模,提出一种时间感知知识表示学习方法Duration-HyTE.首先,将元事实按照有效持续时间分类,对知识有效持续时间进行建模,提出知识有效可信度的计算方法,将其作用于训练过程评价函数和损失函数的计算,最后在含有时间标签的数据集Wikidata12K、YAGO11K和新建立的持续型关系数据集上进行对比实验,结果表明与其他同类方法相比,Duration-HyTE方法在实体和关系的链接预测和时间预测上性能得到有效提升,尤其在Wikidata12K数据集上,经Duration-HyTE训练得到的知识表示模型对于头尾实体的预测效果比当前最优的表示方法分别提升了25.7%和35.8%,有效提高了链接预测准确率.

关键词: 知识图谱, 持续时间建模, 时间感知, 知识表示学习, 链接预测, 有效可信度

Abstract: Knowledge representation learning is the foundation of knowledge acquisition and reasoning. It is widely used in entity extraction, entity alignment, recommendation system and other fields. It has become an important issue throughout the whole process of knowledge graph construction and application. With the development of large-scale knowledge graphs containing time labels, time-aware knowledge representation learning has become one of the research hotspots in this field in recent years. Traditional time-aware knowledge representation learning methods can not effectively use the distribution of knowledge valid duration. In this paper, we propose an improved time-aware knowledge representation learning method combined hyperplane model and duration modeling to solve this problem. Firstly, we divide meta facts into persistent facts and instantaneous facts according to their valid duration. Then we model the valid duration of knowledge, so we get the calculation method of valid reliability. Finally, we propose a new knowledge representation learning method by improving score function with valid reliability. Wikidata12K and YAGO11K are two knowledge graph data sets containing time labels. We extract two new persistent facts datasets from these two datasets. We do a series of comparative experiments on these four data sets. The results show that Duration-HyTE method of link prediction and time prediction performance has been effectively promoted. Especially on Wikidata12K dataset, the accuracy of link prediction of the head entity and tail entity of the Duration-HyTE method is improved by 25.7% and 35.8% respectively compared with other traditional and advanced knowledge representation methods.

Key words: knowledge graph, duration models, time aware, knowledge representation, link prediction, valid reliability