基于多跳路径的知识库补全  

Knowledge Bases Completion Based on Multi-hop Paths

作　　者：王引苗 韩志敏 WANG Yinmiao;HAN Zhimin(Artificial Intelligence Institute,School of Automation,Hangzhou Dianzi University,Hangzhou 310018)

机构地区：[1]杭州电子科技大学自动化学院人工智能研究所,杭州310018

出　　处：《北京大学学报（自然科学版）》2022年第1期7-12,共6页Acta Scientiarum Naturalium Universitatis Pekinensis

基　　金：国家重点研发计划(2018AAA0101601)资助。

摘　　要：为了对知识库(KBs)进行补全,提出一种新的基于路径的推理方法,使用注意力机制,将实体与其类型相结合,共同对路径中的实体进行表示,并使用注意力机制对每条路径预测的关系向量与给定关系的表示向量之差的绝对值进行汇总来计算模型的置信度。在基准数据集WN18RR和FB15k-237上的实验结果表明,与现有的基于路径的关系推理方法相比,所提方法具有更好的性能。In order to complement knowledge bases(KBs),the authors propose a new path-based reasoning method,which uses the attention mechanism to combine entities and their types to represent the entities in the path and use the attention mechanism to summarize the absolute value of the difference between the relationship vector predicted by each path and the representation vector of the given relationship to calculate the confidence of the model.The results of experiment on benchmark data sets WN18 RR and FB15 k-237 show that the proposed model has better performance than the existing path-based relational reasoning methods.

关 键 词：知识库(KBs)补全 基于路径的推理 注意力机制 

分 类 号：TP391.1[自动化与计算机技术—计算机应用技术] TP311.13[自动化与计算机技术—计算机科学与技术]