面向自动文摘的文本结构划分  

Text structure partition for automatic summarization

作　　者：朱荷香[1] 曲维光[1] 卢俊之[2] 李素建[3] 邵艳秋[3] 

机构地区：[1]南京师范大学计算机科学系,南京210097 [2]南京师范大学文学院,南京210097 [3]北京大学计算语言研究所,北京100871

出　　处：《南京大学学报（自然科学版）》2008年第2期204-211,共8页Journal of Nanjing University（Natural Science）

基　　金：国家自然科学基金(60773173,60603093);国家社会科学基金(07BYY050);国家973项目(2004CB318102);江苏省社会科学基金(06JSBYY001);国家博士后基金(20060400027)

摘　　要：文本结构划分是自动文摘研究中的一个关键阶段,也是自然语言处理领域的重要课题.本文通过构建段落向量空间模型,提出一种综合考察相邻段落相似度和连续段落平均相似度的意义段划分方法,使文摘内容更加全面,结构更加平衡.实验结果表明,该方法能够较有效地反映文章的内容结构,对有子标题组织和无子标题组织的文章均适用;由于考虑了总起段,使得文本结构划分更加合理,为自动文摘系统的后续工作打下坚实的基础.Text structure partition is a significant stage in the automatic text summarization as well as an important issue in nature language processing. The topic partition is based on vector space model （VSM） in this paper. Different from the existing approaches that make use of the similarity of adjacent paragraphs, we put forward an algorithm for topic partition based on a comprehensive investigation of both adjacent paragraphic similarity and consecutive average paragraphic similarity. This makes the summarization more comprehensive in content and more balanced in structure. At the same time, the topic number of article is determined automatically. We also find that the importance of recapitulative paragraph is neglected by the previous investigations and propose a method to recognize the recapitulative paragraph. This makes the topic partition more reasonable. We designate three experiments as the basis of topic partition： the adjacent paragraphic similarity, the consecutive paragraphic similarity, and the comprehensive investigation of both adiacent paragraphic similarity and consecutive average paragraphic similarity. The result shows that the method we put forward is superior to the previous methods. The result also shows that the headlines contribute to topic partition, whereas our approach is also suitable for the topicpartition of the articles without headlines. This lays a good foundation for the research of the automatic text summarization system.

关 键 词：文本结构 自动文摘 向量空间模型 段落相似度 意义段划分 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]