节点文献

基于指针网络生成抽象式新闻摘要

GENERATING ABSTRACTIVE NEWS SUMMARIES BASED ON POINTER NETWORK

  • 推荐 CAJ下载
  • PDF下载
  • 不支持迅雷等下载工具,请取消加速工具后下载。

【作者】 郭倩黄继风宋俊典陈海光

【Author】 Guo Qian;Huang Jifeng;Song Jundian;Chen Haiguang;Shanghai Soft China Information Technology Co., Ltd.;College of Information, Mechanical and Electrical Engineering, Shanghai Normal University;

【机构】 上海软中信息技术有限公司上海师范大学信息与机电工程学院

【摘要】 基于指针网络与引入注意力机制的编码器解码器神经网络模型,构建混合网络模型生成抽象式新闻摘要。实验采用搜狐新闻语料库作为数据集,先对数据集进行数据清洗,包括去除空格、特殊字符、停用词等,利用中文分词系统ICTCLAS[1]对清洗过的语料库分词,再把数据集划分为训练集、验证集、测试集。模型的搭建是在加入注意力机制的编码器解码器模型的基础上引入了指针网络,该网络会生成一个权衡概率,用来分配从词典中生成新词与从原文中复制词语的权重。生成的抽象新闻摘要采用ROUGE评分机制进行评测,评测结果比单独采用加入注意力机制的编码器解码器模型平均高出2分。

【Abstract】 Based on the pointer network and the encoder-decoder neural network with attention mechanism, this paper constructs a hybrid network model to generate abstractive news summary. The experiment used Sohu news corpus as the data set. Firstly, the data set was cleaned, including removing spaces, special characters and stop words, etc. The cleaned corpus was segmented by using Chinese word segmentation system ICTCLAS[1]. Then, the data set was divided into a training set, a verification set, and a test set. The model was constructed by introducing a pointer network based on the encoder-decoder model with attention mechanism. The network generated a trade-off probability, which was used to allocate the weight of new words generated from the dictionary and words copied from the original. The generated abstractive news summary is evaluated by the ROUGE scoring mechanism, and the evaluation results are on average 2 points higher than the encoder-decoder model with attention mechanism alone.

【基金】 上海市科技人才计划项目(18PJ1431600);上海市科研计划项目(18ZC2425002);中小企业发展专项资金项目(XQ-ZXQY-02-18-5857)
  • 【文献出处】 计算机应用与软件 ,Computer Applications and Software , 编辑部邮箱 ,2020年06期
  • 【分类号】TP391.1;G210.7
  • 【被引频次】7
  • 【下载频次】262
节点文献中: 

本文链接的文献网络图示:

本文的引文网络