基于小样本的知识图谱补全模型

Knowledge Graph Completion Model Based on Few-Shot

针对知识图谱中存在实体稀疏及实体对数量有限导致知识图谱不完备的问题,提出一种基于小样本的元学习知识图谱补全模型。通过关系元传递重要信息,梯度元提升学习效率,从而快速获取三元组。在链接预测任务上进行验证该方法的有效性。实验结果表明,基于元学习的小样本知识图谱补全算法在数据集FB15K上,MeanRank相较于TransE提高41.4%,Hits@10相较于FSRL提高7%;在数据集Wordnet18上,MeanRank相较于DistMult提高37.9%,Hits@10相较于ComplEx提高18.8%;在数据集NELL-995上,MeanRank相较于TransE提高41.4%,Hits@10相较于FSRL提高18.8%。所提出的方法不仅能更好的进行知识学习,并且显著提升实体和关系的预测效率。

Aiming at the problem of incomplete knowledge map caused by sparse entities and limited number of entity pairs in knowledge map,a meta learning knowledge map completion model based on small samples is proposed. The important information is transmitted through the relation element,and the gradient element improves the learning efficiency,so as to quickly obtain the triples. The effectiveness of this method is verified on the link prediction task. The results show that the small sample knowledge map completion algorithm based on meta learning improves 41.4% compared with TransE and 7% compared with FSRL on the data set FB15K;On the data set Wordnet18,MeanRank is37.9% higher than DistMult,and hits@10 is 18.8% higher than ComplEx;On the data set NELL-995,meanrank increased by 41.4% compared with TransE,and hits@10 increased by 18.8% compared with FSRL. The proposed method can not only better learn knowledge,but also significantly improve the prediction efficiency of entities and relationships.