基于协同注意力和递归随机游走的实体链接方法  

ENTITY LINKING METHOD BASED ON CO-ATTENTION AND RECURRENT RANDOM WALK

作　　者：李圣杰 周新 史一民[1] Li Shengjie;Zhou Xin;Shi Yimin(Faculty of Information Science and Technology,Dalian Maritime University,Dalian 116026,Liaoning,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：《计算机应用与软件》2022年第8期163-170,186,共9页Computer Applications and Software

基　　金：国家自然科学基金项目(61976032,61371090,61602076,61702072)。

摘　　要：针对当前大多数实体链接方法,在获取指称和候选实体之间的语义信息时,没有揭示注意力的焦点,实体链接准确性有待提高的问题,提出一种基于协同注意力和递归随机游走的实体链接方法——Att-RRW。采用协同注意力机制识别出指称上下文和候选实体描述中最具有区别性的单词,提高指称和候选实体之间的局部相关性的准确度;采用递归随机游走实现了将局部兼容性和实体之间的一致性结合起来的集成实体链接。在四个数据集上的实验验证表明,Att-RRW的整体性能优于当前主流的实体链接方法。Aiming at the problem that most of the current entity linking methods do not reveal the focus of attention when acquiring the semantic information between the mention and candidate entities,the accuracy of entity linking needs to be improved,this paper proposes an entity linking method based on co-attention and recurrent walk—Att-RRW.It used co-attention mechanism to identify the difference between the mention context and the candidate entity description words,improving the accuracy of the local correlation between the mention and the candidate entity.Recurrent random walk was used to realize an integrated entity link that combines local compatibility and consistency between entities.Experiments on four data sets verify that the overall performance of Att-RRW is better than the current mainstream entity linking methods.

关 键 词：实体链接 递归随机游走 协同注意力 神经网络 集成实体链接 

分 类 号：TP3[自动化与计算机技术—计算机科学与技术]