融入互注意力的风险领域实体关系抽取研究  

Research on Entity Relationship Extraction in Risk Domain Integrating Mutual Attention

作　　者：杨美芳[1] 杨波[2] YANG Mei-fang;YANG Bo(School of Information Management,Jiangxi University of Finance and Economics,Nanchang 330013,China;Postgraduate of Information Resource Management,Jiangxi University of Finance and Economics,Nanchang 330013,China)

机构地区：[1]江西财经大学信息管理学院,南昌330013 [2]江西财经大学信息资源管理研究所,南昌330013

出　　处：《小型微型计算机系统》2023年第5期991-1001,共11页Journal of Chinese Computer Systems

基　　金：国家自然科学基金项目(72064015)资助;江西省社会科学“十三五”规划项目(19TQ01)资助;江西省教育厅科技项目(GJJ200536)资助。

摘　　要：风险领域实体关系抽取是扩充现有知识图谱与泛化知识工程应用的关键问题.当前特定领域实体关系抽取面临人工标注语料的严重依赖、实体间关系的交叉互联以及远程监督标注存在噪声数据等核心难题,简单的解决方案是运用风险领域已有的知识图谱作为指导.然而,相比通用领域知识图谱,风险领域知识图谱的规模往往较小,难以满足当前领域实体关系抽取的知识需求.因此,本文既要利用已有的风险领域知识图谱,又要充分挖掘蕴含于领域文本数据中规律性的风险知识.本文提出基于知识图谱与文本互注意力的风险领域实体关系抽取方案.首先,根据已有的知识图谱抽象出风险领域实体关系及其约束条件;其次,运用少量高质的实体关系与大规模风险领域语料训练知识图谱与文本的互注意力机制模型,并融合文本表示学习与深度神经网络的方法进行风险领域实体关系的抽取.最后,针对给定的领域文本数据,综合关系约束与关系抽取结果得出风险领域实体关系类型.本文以风险领域数据为例,仅用少量的领域知识,即可获取较好的实体关系抽取效果.The extraction of entity relationships in the risk domain is a key issue for expanding the existing knowledge graph and generalizing the application of knowledge engineering.Currently,entity relationship extraction in a specific field faces serious reliance on manual annotation corpus,cross-interconnection of entity relationships,and noisy data in remote supervision and annotation.The simple solution is to use the existing knowledge map in the risk field as a guide.However,compared with general domain knowledge graphs,the scale of risk domain knowledge graphs is often smaller,and it is difficult to meet the knowledge requirements of entity relationship extraction in the current domain.Therefore,we must not only use the existing risk domain knowledge graph,but also fully mine the regular risk knowledge contained in the domain text data.This paper proposes a scheme of entity relationship extraction in risk domain based on knowledge graph and text mutual attention.First.abstract the entity relationships and constraints in the risk domain based on the existing knowledge graph;secondly,use a small amount of high-quality entity relationships and large-scale risk domain corpus to train the mutual attention mechanism model of the knowledge graph and text;then.merge the text representation The method of learning and deep neural network constructs the entity relationship extraction model in the risk domain based on the mutual attention mechanism.Finally,for a given domain text data,comprehensive relation constraints and relation extraction results are used to obtain entity relation types in the risk domain.This article takes the risk domain data as an example,only a small amount of domain knowledge can be used to obtain a better entity relationship extraction effect.

关 键 词：领域实体关系 互注意力机制 关系抽取 知识图谱 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]