TAC 2009 Update Summarization of ICL

For the update summarization task of TAC 2009, we submitted two runs using two different methods. The first one is manifold ranking method, which models all sentences as a graph. The topic description is deemed as the only labeled node and assigned with an initial score, then the scores of all the sentences in the documents are learned by spreading the initial score on the graph. The second method is called clustering method, which divides all sentences in one topic into several clusters. A good summary should care for all these clusters. The more significant a cluster appears, the more likely sentences will be chosen from this cluster. With these two methods, we conduct some experiments and discuss what has an effect on the performance of the update summarization task.