Abstract:

Knowledge graph completion focuses on how to improve the missing information in knowledge graph. Knowledge graph completion task has many applications, for example, it can be applied to the knowledge graph of rail transit system to support the system design and maintenance optimization. The existing algorithm has high time com-plexity for the real knowledge graph, and it does not make good use of the data out of the knowledge graph. In view of the above two limitations, this paper proposes a knowledge graph completion framework based on active learning. Combined with the idea of active learning, this framework uses link prediction to predict the top-k pairs of entities which are most likely to generate links in the missing knowledge graph. The framework fully considers the internal and external information, and uses the combination of the internal data of the knowledge graph and the external data to realize the missing completion of the knowledge graph. Based on Freebase and DBpedia datasets, a comparative experiment is carried out for the existing work. Experiment results show that, the enhance link prediction (ELP) algorithm has better effect and active learning ability. The relationship verification combining internal data and exter-nal data of the knowledge graph can verify the triples more effectively.

Key words: active learning, knowledge graph completion, link prediction, relationship verification

摘要：

知识图谱补全任务研究如何补全知识图谱中的缺失关系。知识图谱补全任务有许多广泛的应用，例如可以应用到轨道交通运维知识库中以支撑轨道交通的系统设计、运维优化。现有的算法在用于现实的大规模知识图谱时时间开销巨大，并且无法很好地利用知识图谱外部的数据信息。针对以上两点局限性，提出了一种基于主动学习的知识图谱补全框架。该框架结合主动学习的思想，利用链接预测预先筛选缺失知识图谱中最有可能产生链接的前k对实体对，然后充分考虑知识图谱内部信息和外部信息，采用内外部数据相结合的方式实现知识图谱的缺失补全。基于Freebase和DBpedia数据集，针对已有的工作进行了对比实验，实验结果表明提出的增强链接预测算法（ELP）效果更好，并且具有主动学习能力；提出的内部数据和外部数据相结合的关系验证方法能更有效地验证三元组。

关键词: 主动学习, 知识图谱补全, 链接预测, 关系验证