摘要
为了改进生成式文本的摘要模型,本文提出了一种基于主题的生成对抗与指针网络结合的文本摘要模型.首先通过LDA主题建模方法获取主题词,在获取单词的主题向量后,将结合主题的词向量与传统的序列注意力相结合,形成新的复合注意力共同影响单词的生成,然后通过加入生成对抗网络以在指针生成网络上取得了更好的效果.实验采用gigaword数据集进行训练,采用ROUGE评分机制进行评分,结果证明由于融入主题因素,相比单独采用指针网络我们的模型提升了摘要结果的可读性及准确性,具有更好的表现.
In order to improve the summarization model of generative text, this paper proposes a text summarization model based on topic Generative Adversarial and Pointer Network.First, the topic words are obtained through the LDA topic modeling method.After obtainingthe topic wordsvectors, the word vectors that combine the topics and the traditional sequence attention are combined to form a new compound attention that affects the generation of the words, and then generated by addingGenerative Adversarial network to achieve better results in the pointernetwork.The experiment uses the gigaword data set for training, and uses the ROUGE scoring mechanism for scoring.The results prove that due to the integration of Topicfactors, our model improves the readability and accuracy of the summary results and has better performance than using the pointer network alone.
作者
郭继峰
费禹潇
孙文博
谢培浇
张健
GUO Ji-feng;FEI Yu-xiao;SUN Wen-bo;XIE Pei-jiao;ZHANG Jian(College of Information and Computer Engineering,Northeast Forestry University,Harbin 150000,China;School of Artificial Intelligence,Wuxi Vocational College of Science and Technology,Wuxi 214029,China)
出处
《小型微型计算机系统》
CSCD
北大核心
2023年第1期199-203,共5页
Journal of Chinese Computer Systems
基金
国家自然科学基金项目(613000980)资助
黑龙江省自然科学基金项目(LH2019C003)资助。
关键词
指针网络
生成对抗
主题模型
文本摘要
pointer network
generative adversarial
topic model
text summarization
