基于改进指针生成网络的文本摘要
Text Summarization Based on Improved Pointer-Generator Network
杨尚儒 1廖闻剑2
作者信息
- 1. 武汉邮电科学研究院 武汉 430074;南京烽火天地通信科技有限公司 南京 210019
- 2. 南京烽火天地通信科技有限公司 南京 210019
- 折叠
摘要
随着人们每天接收的消息越来越多,能够短时高效地找到自己想要的内容,获取想要的信息是提升自己的关键,因此文本摘要变得必不可少.人工生成文章的摘要是一项费时费力的任务,自动生成可读性高、流畅性强的摘要变得很有必要.摘要生成有很多方法,又分为抽取式摘要和生成式摘要.指针生成网络因为其能有效解决未登录词的问题,仍然是一种非常流行的文本摘要方法.在我们的工作中,仍然使用传统的指针生成网络为基本框架,引入Transformer中的编码器部分作为预处理,提升编码质量;另外,引入未登录词惩罚来提高生成摘要文本的新颖性.实验结果表明,该模型在NLPCC数据集上取得了良好的效果.
Abstract
With people receive more and more information every day,it's important to be able to get the information we want in a short period of time,so text summaries have become indispensable.Manually generating abstracts for articles is a time-consum-ing and laborious task,and it becomes necessary to automatically generate abstracts with high readability and fluency.There are many ways to generate abstracts,which are divided into extractive abstracts and generative abstracts.The pointer generation net-work is still a very popular text summarization method because it can effectively solve the problem of unregistered words.In our work,the traditional pointer generation network is still used as the basic framework,and the encoder part in Transformer is intro-duced as preprocessing to improve the coding quality.In addition,the penalty of unregistered words is introduced to improve the novelty of the generated abstract text.Experimental results show that the model has achieved good results on the NLPCC data set.
关键词
编码器-解码器/注意力机制/Transformer/指针生成网络Key words
encoder-decoder/attention mechanism/Transformer/pointer-generator network引用本文复制引用
出版年
2024
国家科技期刊平台
NSTL
万方数据