基于混合过滤编码的神经中文生成式摘要  

Neural Chinese abstractive summarization based on hybrid filter encoding

作　　者：蓝雯飞[1] 周伟枭 许智明 朱容波[1] 罗一凡 LAN Wenfei;ZHOU Weixiao;XU Zhiming;ZHU Rongbo;LUO Yifan(College of Computer Science, South-Central University for Nationalities, Wuhan 430074, China;College of Mechanical Engineering and Automation, Fuzhou University, Fuzhou 350108, China)

机构地区：[1]中南民族大学计算机科学学院,武汉430074 [2]福州大学机械工程及自动化学院,福州350108

出　　处：《中南民族大学学报（自然科学版）》2021年第3期305-311,共7页Journal of South-Central University for Nationalities：Natural Science Edition

基　　金：国家自然科学基金资助项目(61772562)。

摘　　要：在神经生成式摘要任务中,由于输入文档与参考摘要之间不存在显式的对齐关系,通常会导致重复生成相同单词的问题以及生成的摘要与输入文档语义不相关、准确性低的问题.为更好解决该问题,提出了混合过滤编码网络(HFEN)并在HFEN中集成混合过滤编码机制(HFEM)、注意力机制、指针生成器.HFEM分为管道过滤编码机制(PFEM)、特征融合过滤编码机制(FFFEM).其中,FFFEM具体通过添加特征融合层实现.在中文摘要领域基准数据集(LCSTS)上的实验结果表明:HFEN相较于基线模型生成了准确性更高、重复单词更少的摘要,ROUGE指标有较大提升.ive summarization based on hybrid filter encoding[J].Journal of South-Central University for Nationalities(Natural Science Edition),2021,40(3):305-311.Neural Chinese abstractive summarization based on hybrid filter encodingLAN Wenfei1,ZHOU Weixiao1*,XU Zhiming2,ZHU Rongbo1,LUO Yifan1(1 College of Computer Science,South-Central University for Nationalities,Wuhan 430074,China;2 College of Mechanical Engineering and Automation,Fuzhou University,Fuzhou 350108,China)Abstract For the neural abstractive summarization task,because there is no explicit alignment relationship between the input document and the reference summary,this often leads to problems includes repeatedly generating the same words and generating a summary that is not semantically relevant to the input document and is not accurate enough.To solve these problems,this paper proposes a hybrid filter encoding network(HFEN)and integrates the hybrid filter encoding mechanism(HFEM),attention mechanism and pointer generator into HFEN.HFEM is divided into pipeline filter encoding mechanism(PFEM)and feature fusion filter encoding mechanism(FFFEM),FFFEM is realized by adding feature fusion layer.Experimental results on the LCSTS showed that HFEN produced a more accurate summary with fewer repeated words than the baseline model,with significant improvement in ROUGE.

关 键 词：神经中文生成式摘要 混合过滤编码网络 混合过滤编码机制 管道过滤编码机制 特征融合过滤编码机制 指针生成器 

分 类 号：TP391.4[自动化与计算机技术—计算机应用技术]