NUS at DUC 2007: Using evolutionary models of text

This paper presents our new, query-based multi-document summariza-tion system used in DUC 2007. Cur-rent graph-based approaches to text summarization, such as TextRank and LexRank, assume a static graphmodel which does not model how input text emerges. A suitable evolution-ary graph model that is related to human writing/reading process may impart a better understanding of the text and improve the subsequent sum-marization process. We propose a timestamped graph (TSG) model mo-tivated by human writing and reading processes, and show how input text emerges under the construction phase of TSG. We applied TSG on both the main task and update summary task in Document Understanding Confer-ences (DUC) 2007 and achieved satis-factory results. We also suggested a modified MMR re-ranker for the up-date task.