CRL/NYU System at DUC-2004

We participated in two multi-document sum- marization tasks (Task 2 and Task 5) at the DUC-2004 formal run and evaluated the per- formance of our summarization system. Our system based on sentence extraction also uses a module to estimate similarity between sen- tences. The similarity information was used for either selecting the representative sentence among similar sentences or gathering key sen- tences that have similar structures but different contents. We also incorporated a module which categorized document sets into two groups cor- responding to the distribution of key sentences.