TSCAN

Abstract

A topic is defined as a seminal event or activity along with all directly related events and activities. It is represented as a chronological sequence of documents by different authors published on the Internet. In this paper, we define a task called topic anatomy, which summarizes and associates core parts of a topic graphically so that readers can understand the content easily. The proposed topic anatomy model, called TSCAN, derives the major themes of a topic from the eigenvectors of a temporal block association matrix. Then, the significant events of the themes and their summaries are extracted by examining the constitution of the eigenvectors. Finally, the extracted events are associated through their temporal closeness and context similarity to form the evolution graph of the topic. Experiments based on the official TDT4 corpus demonstrate that the generated evolution graphs comprehensibly describe the storylines of topics. Moreover, in terms of content coverage and consistency, the produced summaries are superior to those of other summarization methods based on human composed reference summaries.