全文下载： 
202102015.pdf
文章编号： 1672-6987（2021）02-0104-08； DOI： 10.16351/j.1672-6987.2021.02.015
孙驰， 杜军威*， 李浩杰(青岛科技大学 信息科学技术学院，山东 青岛 266061)
摘要： 面对阅读理解数据集存在多文档、多段落候选，而参考答案只出现在一个段落里面的情景,提出基于段落选择的分层融合阅读理解模型。段落选择器用于在候选文档中选出最佳段落，基于双向注意力的分层信息融合模型用于在最佳段落中定位参考答案的起始和结束位置，降低传统直接选择机器阅读理解模型的复杂性和计算量。在DuReader中文数据集上进行实验效果对比，结果显示所提出的方法较基准方法在Bleu4和RougeL指标分别提高了约12%和11%，进一步验证了所提出方法的有效性。
关键词： 阅读理解； 分层融合； 多段落； 段落选择器
中图分类号： TP 391文献标志码： A
引用格式： 孙驰， 杜军威，李浩杰. 基于段落选择的分层融合阅读理解模型［J］. 青岛科技大学学报(自然科学版), 2021, 42(2): 104111.
SUN Chi，DU Junwei，LI Haojie. A hierarchical and integrated reading comprehension model based on paragraph selection［J］. Journal of Qingdao University of Science and Technology（Natural Science Edition), 2021，42(2)： 104111.
A Hierarchical and Integrated Reading Comprehension Model
Based on Paragraph Selection
SUN Chi, DU Junwei, LI Haojie
(College of Information Science and Technology, Qingdao University of Science and Technology， Qingdao 266061, China)
Abstract: This paper designs a hierarchical integrated reading comprehension model based on paragraph selection and provides ideas for processing multidocument multiparagraph reading comprehension data sets. The paragraph selector selects the best paragraph from the candidate documents. The hierarchical reading comprehension model uses twoway attention and hierarchical information fusion to locate the starting and ending positions of the reference answers from the best paragraphs. This model reduces the complexity and computation of the traditional direct selection machine reading comprehension model. A comparison of experimental results on DuReader Chinese data sets shows that the method proposed in this paper improves the indexes of BLEU4 and rougel by about 12% and 11% respectively compared with the benchmark method, which further verifies the effectiveness of the method proposed in this paper.
Key words: reading comprehension; layer fusion; multiparagraph; paragraph selector
收稿日期： 20200426
基金项目： 国家自然科学基金项目（61402246,61973180）；山东省重点研发计划项目（2018GGX101052）；山东省自然科学基金项目（ZR2018MF007,ZR2019MF014,ZR2019MF033）.
作者简介： 孙驰(1994—),男,硕士研究生.*通信联系人.
