基于路径与层次注意力的关系预测方法  

Relation prediction method based on path and hierarchical attention

作　　者：石津妮 安敬民 李冠宇[1] SHI Jin-ni;AN Jing-min;LI Guan-yu(Information Science and Technology College,Dalian Maritime University,Dalian 116026,China)

机构地区：[1]大连海事大学信息科学技术学院,辽宁大连116026

出　　处：《计算机工程与设计》2023年第11期3433-3439,共7页Computer Engineering and Design

基　　金：国家自然科学基金项目(61976032)。

摘　　要：目前,大多数基于路径的关系预测方法忽略了路径和关系之间的语义相关性,针对此问题提出一种层次注意力机制,捕获不同粒度的信息,包括关系级特征和路径级特征,在多跳路径和关系之间提供一种补充的语义关联。关系级注意力是以目标关系为导向,更多关注与目标关系相匹配的关系。路径级注意力是通过捕获给定实体对的邻域子图结构实现,为结果提供可解释性。该模型可以有效缓解噪声路径和冗余路径对关系预测任务的影响。实验结果表明,在相关数据集上,该模型的关系预测性能优于现有模型。Currently,most path based relation prediction methods ignore the semantic correlation between paths and relations.To solve this problem,a hierarchical attention mechanism was proposed to capture information with different granularities,including relation-level feature and path-level features,which provided a supplementary semantic association between multi-hop paths and relation.Relation-level attention was target relation oriented,which focused more on the relations that matched the target relation.Path-level attention was achieved by capturing the neighborhood subgraph structure of a given entity pair,which was able to provide explainability for results.The model can effectively alleviate the impact of noise path and redundant path on relationship prediction task.Experimental results show that the model outperforms existing models in relationship prediction on related datasets.

关 键 词：知识图谱 层次注意力 多粒度 关系路径 目标关系 实体邻域 关系预测 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]