图书情报工作 ›› 2019, Vol. 63 ›› Issue (11): 108-117.DOI: 10.13266/j.issn.0252-3116.2019.11.012

• 知识组织 • 上一篇    下一篇

基于序列到序列模型的抽象式中文文本摘要研究

余传明1, 朱星宇1, 龚雨田1, 安璐2   

  1. 1. 中南财经政法大学信息与安全工程学院 武汉 430073;
    2. 武汉大学信息管理学院 武汉 430072
  • 收稿日期:2018-06-15 修回日期:2018-12-16 出版日期:2019-06-05 发布日期:2019-06-05
  • 通讯作者: 安璐(ORCID:0000-0002-5408-7135),教授,博士生导师,通讯作者,E-mail:anlu97@163.com。
  • 作者简介:余传明(ORCID:0000-0001-7099-0853),教授;朱星宇(ORCID:0000-0001-8122-3000),硕士研究生;龚雨田(ORCID:0000-0002-0434-2492),硕士研究生。
  • 基金资助:
    本文系国家自然科学基金面上项目"大数据环境下基于领域知识获取与对齐的观点检索研究"(项目编号:71373286)和教育部哲学社会科学研究重大课题攻关项目"提高反恐怖主义情报信息工作能力对策研究"(项目编号:17JZD034)研究成果之一。

Research of Abstractive Chinese Text Summarization Based on Seq2seq Model

Yu Chuanming1, Zhu Xingyu1, Gong Yutian1, An Lu2   

  1. 1. School of Information and Safety Engineering, Zhongnan University of Economics and Law, Wuhan 430073;
    2. School of Information Management, Wuhan University, Wuhan 430072
  • Received:2018-06-15 Revised:2018-12-16 Online:2019-06-05 Published:2019-06-05

摘要: [目的/意义]为更好地处理文本摘要任务中的未登录词(out of vocabulary,OOV),同时避免摘要重复,提高文本摘要的质量,本文以解决OOV问题和摘要自我重复问题为研究任务,进行抽象式中文文本摘要研究。[方法/过程]在序列到序列(sequence to sequence,seq2seq)模型的基础上增加指向生成机制和覆盖处理机制,通过指向生成将未登录词拷贝到摘要中以解决未登录词问题,通过覆盖处理避免注意力机制(attention mechanism)反复关注同一位置,以解决重复问题。将本文方法应用到LCSTS中文摘要数据集上进行实验,检验模型效果。[结果/结论]实验结果显示,该模型生成摘要的ROUGE (recall-oriented understudy for gisting evaluation)分数高于传统的seq2seq模型以及抽取式文本摘要模型,表明指向生成和覆盖机制能够有效解决未登录词问题和摘要重复问题,从而显著提升文本摘要质量。

关键词: 抽象式文本摘要, 序列到序列模型, 注意力机制, 覆盖机制, 指向生成机制

Abstract: [Purpose/significance] To deal with the Out Of Vocabulary (OOV) in text summarization while avoiding duplication of summaries, this article focuses on solving the OOV problem and the self-duplication and carries out a profiling study.[Method/process] Bases on the sequence-to-sequence model, a pointer generator module and a coverage processing module are added. An attempt is made to copy the OOV into abstractive summary to solve the problem of OOV by means of the pointer generator module. The coverage processing module tries to avoid the Attention Mechanism paying attention to the same position repeatedly to solve the duplicate problem. The model is applied to the Chinese summarization dataset LCSTS to conduct experiments to test the effectiveness.[Result/conclusion] Experiment results show that the ROUGE of the generated summary is much higher than that of seq2seq model and extractive model, indicating that in the abstractive Chinese text summary, the pointer generator module and the coverage mechanism module can effectively solve the problem of OOV and the repetition of the summary, thereby significantly improving text summary quality.

Key words: abstractive text summarization, sequence-to-sequence model, attention mechanism, coverage mechanism, pointer generator mechanism

中图分类号: