基于多角度交叉注意力机制的知识库问答方法  

QUESTION ANSWERING OVER KNOWLEDGE BASE BASED ON MULTI-VIEW CROSS-ATTENTION MECHANISM

作　　者：代发扬 符海东[1,2] 高峰 顾进广 Dai Fayang;Fu Haidong;Gao Feng;Gu Jinguang(School of Computer Science and Technology,Wuhan University of Science and Technology,Wuhan 430065,Hubei,China;Hubei Province Key Laboratory of Intelligent Information Processing and Real-time Industrial,Wuhan 430065,Hubei,China;Big Data Science and Engineering Research Institute,Wuhan University of Science and Technology,Wuhan 430065,Hubei,China;Key Laboratory of Rich Media Digital Publishing Content Organization and Knowledge Service of Press and Publication Administration,Beijing 100083,China)

机构地区：[1]武汉科技大学计算机科学与技术学院,湖北武汉430065 [2]湖北省智能信息处理与实时工业系统重点实验室,湖北武汉430065 [3]武汉科技大学大数据科学与工程研究院,湖北武汉430065 [4]新闻出版署富媒体数字出版内容组织与知识服务重点实验室,北京100083

出　　处：《计算机应用与软件》2023年第12期33-40,共8页Computer Applications and Software

基　　金：国家自然科学基金项目(U1836118,61673304);国家社科基金重大计划项目(11&ZD189);湖北省自然科学基金项目(2018CFB194)。

摘　　要：近些年知识库问答的方法通常利用多视角信息来表示候选答案,忽略了这些信息间的相互影响,将问题的单词与候选答案的多视角信息计算相关性,忽略了二者在整体与细节上的信息。基于上述问题,提出一个多角度交叉注意力模型,通过多视角交叉注意力机制获取候选答案多视角信息间的交叉影响;将问题与候选答案信息进行整体表示,运用双向交叉注意力机制来计算其二者在整体级别上的关联性,最终提高获取答案的正确率。利用FreeBase知识库与WebQuestions数据集进行实验,F1值达到55.84%,优于最近表现较好的方法。In recent years,question answering over knowledge base methods usually use multi-view information to represent candidate answers,ignoring the mutual influence between these information,and they calculate the correlation between the word of the question and the multi-view information of the answer,ignoring the overall and detailed information of the two.Based on the above problems,a multi-view cross-attention model is proposed.Multi-view cross-attention mechanism was used to obtain the cross-influence between multi-view information of the candidate answer.The question and the candidate answer information were represented as a whole,and the two-way cross-attention mechanism was used to calculate the correlation between the two at the overall level.The correct rate of obtaining the answer was improved.Experiments were conducted on the FreeBase knowledge base and WebQuestions data sets,and the F1 value reached 55.84%,which was better than the recent methods that performed well.

关 键 词：知识库问答 多视角信息 多视角交叉注意力机制 双向交叉注意力机制 

分 类 号：TP3[自动化与计算机技术—计算机科学与技术]