一种基于对比策略强化知识推理的元学习框架
A Meta-learning Framework for Enhancing Knowledge Reasoning Based on Contrastive Strategies
张希权 1党建武 1王阳萍1
作者信息
- 1. 兰州交通大学 电子与信息工程学院,兰州 730070
- 折叠
摘要
随着知识图谱中新实体不断涌现,现有嵌入式模型难以获取新实体嵌入,导致模型预测精度降低.为此,提出一种基于对比策略强化知识推理的元学习框架.首先,该框架通过对路径重要度排序,提升隐式信息的提取效率;然后,使用一种对比策略表示强关联路径,并依据负路径与目标关系之间余弦相似性剔除过多的冗余信息;最后,使用元学习来转移独立于实体的嵌入.实验结果表明:在链接预测知识图谱任务中,Hits@10 指标最高可达96.24%,说明该框架可以有效提升预测精度.
Abstract
With the continuous emergence of new entities in the knowledge graph,it is difficult for existing embedded models to obtain new entity embeddings,resulting in a decrease in model prediction accuracy.Accordingly,the study proposes a meta-learning framework based on contrastive strategies to strengthen knowledge reasoning.First,the framework improves the extraction efficiency of implicit information by sorting the importance of paths,and then u-ses a comparison strategy to represent strongly associated paths.Excessive redundant information is removed according to the cosine similarity between negative paths and target relations,and finally meta-learning is used to transfer entity-inde-pendent embeddings.The experimental results show that in the link prediction knowledge map task,the value of the Hits@10 index can reach up to 96.24%,indicating that the framework can effectively improve prediction accuracy.
关键词
归纳推理/知识补全/对比策略/元学习Key words
inductive reasoning/knowledge completion/comparison strategy/meta-learning引用本文复制引用
基金项目
国家自然科学基金(62067006)
国家自然科学基金(62367005)
甘肃省教育科技创新项目(2021jyjbgs-05)
中央引导地方科技发展专项(22ZY1QA002)
甘肃省知识产权计划(21ZSCQ013)
高校科研创新平台重大培育项目(2024CXPT-17)
出版年
2024
NETL
NSTL
维普
万方数据