基于双重多路注意力匹配的观点型阅读理解  

Dual co-matching network with multiway attention for opinion reading comprehension

作　　者：鲍亮 陈志豪 陈文章 叶锴 廖祥文[1,2,3] BAO Liang;CHEN Zhi-hao;CHEN Wen-zhang;YE Kai;LIAO Xiang-wen(School of Mathematics and Computer Science,Fuzhou University,Fuzhou 350116,Fujian,China;Fujian Provincial Key Laboratory of Networking Computing and Intelligent Information Processing,Fuzhou University,Fuzhou 350116,Fujian,China;Digital Fujian Institute of Financial Big Data,Fuzhou 350116,Fujian,China)

机构地区：[1]福州大学数学与计算机科学学院,福建福州350116 [2]福州大学福建省网络计算与智能信息处理重点实验室,福建福州350116 [3]数字福建金融大数据研究所,福建福州350116

出　　处：《山东大学学报（理学版）》2021年第3期44-53,共10页Journal of Shandong University(Natural Science)

基　　金：国家自然科学基金资助项目(61976054,61772135,U1605251);国家自然科学基金青年科学基金资助项目(41801324);福建省自然科学基金面上项目(2017J01755);模式识别国家重点实验室开放课题基金项目(201900041)。

摘　　要：观点型阅读理解旨在对于给定的问题和答案段落摘要,判断答案段落摘要相对于问题的观点倾向。现有基于神经网络的模型主要依靠精心设计的匹配网络来捕获各部分文本之间的关系,往往只考虑单向的匹配关系或是采用单一的交互机制对各个文本对之间的关系进行建模,难以有效捕捉观点型问题中问题和答案段落摘要二者之间的潜在关系。为此,提出一种基于双重多路注意力的匹配方法。该方法对<问题,答案段落摘要>二元组从2个方向同时进行匹配,并采用多种注意力机制学习二者的协同注意力表示,通过双向多视角的交互为最后的观点推断提供更丰富的推理证据。在观点型阅读理解数据集DureaderOpinion上的实验表明,该方法相对于基准模型取得了更好的效果。Opinion-based reading comprehension aims to judge the opinion polarity of the answer span to the given question. Previous approaches usually rely on a matching network to capture the relationship between texts, however either only modeling the relationship in a uni-direction, or only calculating the interactive representations with a single attention mechanism, which cannot effectively capture the correlation between the opinion-based question and the answer span. In this work, we propose dual co-matching with multiway attention(DCMA) matching method, which models the relationship between question and answer bidirectionally, and three attention mechanisms are employed to calculated the co-attention representations of question and answer. Experimental result on the opinion-based reading comprehension dataset DureaderOpinion demonstrates our model obtains state-of-the-art performance.

关 键 词：机器阅读理解 观点挖掘 注意力机制 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]