Automatic recognition of topic transition for newspaper articles and application to document summary

自动识别报纸文章的主题转换并应用于文档摘要

基本信息

批准号：
15500086
负责人：
SUZUKI Yoshimi
金额：
$ 2.5万
依托单位：
University of Yamanashi
依托单位国家：
日本
项目类别：
Grant-in-Aid for Scientific Research (C)
财政年份：
2003
资助国家：
日本
起止时间：
2003 至 2004
项目状态：
已结题

项目摘要

In this study, we paid attention to the automatic summary for the newspaper articles.The following was researched as a first step to summarize multi-documents precisely.(1)A subject template is made from the large-scale corpus, and we extract the subsequent articles of a target article using that subject template correctly. The extracted subsequent articles are classified in the subject cluster.(2)Every subject cluster is summarized, and the whole of the subsequent articles is summarized in consideration of a connection between the clusters.Regarding (1), we proposed a method for topic tracking using subject templates and machine learning (support vector machines). And also, we showed that our methods can extract subsequent articles with high accuracy using large corpus (the corpus by Topic detection and Tracking and articles of Mainichi Shimbun newspaper) (research paper 4,5).Regarding (2), we found that we have to extract synonyms of each word for multi-document summarization. We proposed a method to identify synonym pairs from Japanese newspaper (3,4). For identifying synonyms, we compared Lin's method with Hindle's method and we found Lin's method is better than Hindle's method for Japanese documents.We proposed a method which is based on Lin's method for Japanese documents. Moreover, we performed some experiments of sentence extraction using automatically extracted synonym pairs and title of newspaper article (research paper 1,2). The method is as the following, firstly from newspaper articles we extracted synonyms of words in titles of newspaper articles using the proposed method which is based on Lin's method, then we performed sentence extraction using the results. The results show that identifying synonyms is useful for sentence extraction.

在本研究中，我们关注的是报纸文章的自动摘要。作为多文档精确摘要的第一步，我们进行了以下研究。(1)从大规模语料库中制作主题模板，并提取主题模板。目标文章的后续文章正确使用该主题模板。将提取的后续文章分类到主题簇中。（2）对每个主题簇进行总结，并考虑到簇之间的连接来总结整个后续文章。关于（1），我们提出了一种主题跟踪方法使用主题模板和机器学习（支持向量机）。此外，我们还表明，我们的方法可以使用大型语料库（主题检测和跟踪的语料库以及每日新闻报纸的文章）高精度地提取后续文章（研究论文 4,5）。关于（2），我们发现我们必须提取每个单词的同义词以进行多文档摘要。我们提出了一种从日本报纸中识别同义词对的方法（3,4）。在同义词识别方面，我们将Lin's方法与Hindle's方法进行了比较，发现Lin's方法对于日语文档比Hindle's方法要好。我们提出了一种基于Lin's方法的日语文档方法。此外，我们使用自动提取的同义词对和报纸文章标题进行了一些句子提取实验（研究论文1,2）。该方法如下，首先使用基于Lin方法的所提出的方法从报纸文章中提取报纸文章标题中的单词的同义词，然后使用结果进行句子提取。结果表明，识别同义词对于句子提取很有用。