Knowledge Graph based Mutual Attention for Machine Reading Comprehension over Anti-Terrorism Corpus

作     者：Feng Gao Jin Hou Jinguang Gu Lihua Zhang 

作者机构：School of Computer Science and TechnologyWuhan University of Science and TechnologyWuhan 430065Hubei The Key Laboratory of Rich-Media Knowledge Organization and Service of Digital Publishing ContentInsitute of Scientic and Technical Information of ChinaBeijing 100038China Wuhan University of Science and Technology Big Data Science and Engineering Research InstituteWuhan 430065Hubei Eastchina Jiaotong UniversityNanchang 330013Jiangxi 

出 版 物：《Data Intelligence》 (数据智能（英文）)

年 卷 期：2023年第5卷第3期

页      面：685-706页

核心收录：

学科分类：0303[法学-社会学] 03[法学] 081203[工学-计算机应用技术] 08[工学] 0835[工学-软件工程] 0812[工学-计算机科学与技术（可授工学、理学学位）] 

基　　金：National key research and development program(2020AAA0108500) National Natural Science Foundation of China Project(No.U1836118) Key Laboratory of Rich Media Digital Publishing,Content Organization and Knowledge Service(No.:ZD2022-10/05) 

主　　题：Machine reading comprehension Anti-terrorism domain Knowledge embedding Knowledge attention Mutual attention 

摘      要：Machine reading comprehension has been a research focus in natural language processing and intelligence ***,there is a lack of models and datasets for the MRC tasks in the anti-terrorism ***,current research lacks the ability to embed accurate background knowledge and provide precise *** address these two problems,this paper first builds a text corpus and testbed that focuses on the anti-terrorism domain in a semi-automatic ***,it proposes a knowledge-based machine reading comprehension model that fuses domain-related triples from a large-scale encyclopedic knowledge base to enhance the semantics of the *** eliminate knowledge noise that could lead to semantic deviation,this paper uses a mixed mutual ttention mechanism among questions,passages,and knowledge triples to select the most relevant triples before embedding their semantics into the *** results indicate that the proposed approach can achieve a 70.70%EM value and an 87.91%F1 score,with a 4.23%and 3.35%improvement over existing methods,respectively.