基于循环交互注意力网络的问答立场分析  

Answer stance detection based on recurrent interactive attention network

作　　者：骆旺达 刘宇瀚 梁斌 徐睿峰 LUO Wangda;LIU Yuhan;LIANG Bin;XU Ruifeng(School of Computer Science and Technology,Harbin Institute of Technology(Shenzhen),Shenzhen 518055,China)

机构地区：[1]哈尔滨工业大学(深圳)计算机科学与技术学院,深圳518055

出　　处：《清华大学学报（自然科学版）》2021年第9期913-919,共7页Journal of Tsinghua University(Science and Technology)

基　　金：国家自然科学基金面上项目(61876053);国家自然科学基金重点项目(61632011);深圳市基础研究学科布局项目(JCYJ20180507183527919,JCYJ20180507183608379)。

摘　　要：针对现有问答立场分析方法未考虑问答文本间交互依赖关系的不足,该文提出一种基于循环交互注意力(recurrent interactive attention,RIA)网络的问答立场分析方法。该方法模拟人类的问答阅读理解机制,采用交互注意力机制和循环迭代策略,结合问题和回答的相互联系分析问答文本的立场信息。此外,为了处理问题文本无法明确表达自身立场的情况,该方法将问题转换为陈述句。在中文社交问答数据集上的实验结果表明,由于有效地表示了问答对依赖关系,本文方法的性能优于已有方法。Most existing answer stance detection methods ignore the interactive dependence between question and answer. This paper describes an answer stance detection method based on a recurrent interactive attention(RIA) network. This method simulates the human interactions in question-answer reading comprehension using an interactive attention mechanism and iterations to simulate the interactive dependence between question and answer for detecting the answer stance. In addition, since the question text cannot explicitly express its stance, the question text is transformed into declarative sentences. Tests on a Chinese social media question-answer dataset show that this method outperforms existing answer stance detection methods due to the effective representation of the interactive dependence between the question and the answer.

关 键 词：问答立场分析 循环交互注意力 问题表示 

分 类 号：TP391.1[自动化与计算机技术—计算机应用技术]