摘要
针对通过Extractive方式实现自动文摘而存在文本连贯性欠缺和出现未登录词问题,提出一种基于改进Encoder-Decoder模型的新闻摘要生成方法。首先,在数据预处理的过程中融入额外的语言特征,如词语的词性和TF-IDF,使词语具有多维度的含义;其次,采用Decoder/Pointer机制在摘要中指向原文本中的位置对低频词进行处理;最后,采用注意力机制来协助模型记忆输入数据并确定其注意程度。在News2016zh数据集上进行实验,结果表明基于改进Encoder-Decoder模型与基线Encoder-Decoder相比,ROUGE-1、ROUGE-2和ROUGE-L值分别提高了32.1%、30.5%和32.5%,在摘要连贯性方面也得到了较好提升。
Aiming at the problem of text coherence deficiency and unknown words in automatic summarization by Extractive method,a news summary generation method based on improved Encoder-Decoder model was proposed.Firstly,additional linguistic features such as parts-of speech tags and TF-IDF were integrated into the process of data preprocessing to make words have multi-dimensional meanings.Secondly,low-frequency words were processed by using Decoder/Pointer mechanism to point to the position of the original text in the summary.Finally,attention mechanism was used to assist the model to memorize key input data and determine its attention level.The open Chinese news dataset called News2016zh was used as the data source.The experimental results show that the proposed method has the ROUGE-1,ROUGE-2 and ROUGE-L improved by 32.1%,30.5%and 32.5%respectively than the traditional method;and the coherence of summary has also been improved.
作者
李晨斌
詹国华
李志华
LI Chenbin;ZHAN Guohua;LI Zhihua(School of Information Science and Engineering,Hangzhou Normal University,Hangzhou Zhejiang 311121,China)
出处
《计算机应用》
CSCD
北大核心
2019年第S02期20-23,共4页
journal of Computer Applications
基金
浙江省自然科学基金资助项目(LY17D060005)
