面向主题的无监督文本摘要生成方法  

Topic-oriented Unsupervised Text Summarization Generation Method

作　　者：蒋杰 JIANG Jie(Guangdong University of Technology,Guangzhou 510006,China)

机构地区：[1]广东工业大学计算机学院,广东广州510006

出　　处：《电脑知识与技术》2021年第28期127-130,148,共5页Computer Knowledge and Technology

摘　　要：随着人工智能的迅速发展,传统的文本摘要技术也从抽取式摘要向生成式摘要发展。当源文档中存在着多个主题的内容时,现有的大多数生成式文本摘要方法会尽全面地对多主题内容进行概括和总结,而对其中包含某个主题的内容可能无法进行细致描述。针对这一问题,本文提出了一个面向主题信息的无监督文本摘要生成方法。在给定目标主题信息的情况下,在解码器中利用注意力机制将目标主题信息与文本本身的主题信息进行信息融合,从而使得包含了目标主题的文本得到较大的注意力权重,以此生成与目标主题相关的文本摘要。在大规模英文文本数据集Yelp上的实验结果验证了所提方法的有效性和先进性。With the rapid development of artificial intelligence,traditional text summarization techniques have also evolved from ex⁃tractive summaries to generative summaries.When there are multiple topics in the source document,most of the existing generative text summaries will summarize and summarize the content of multiple topics in a comprehensive manner,and the content of a cer⁃tain topic may not be able to be described in detail.In response to this problem,this paper proposes a topic-oriented unsupervised text summary generation method.Given the target topic information,in the decoder,the attention machanism is used to fuse the tar⁃get topic information with the topic information of the text itself,so that the text which contains the target topic information gets a larger attention weight,so as to generate a text summary related to the target topic.The experimental results on the large-scale Eng⁃lish text data set Yelp verify the effectiveness and advancement of the proposed method.

关 键 词：自然语言处理 序列到序列模型 无监督学习 生成式摘要 主题信息 注意力机制 

分 类 号：TP18[自动化与计算机技术—控制理论与控制工程]