The LIA summarization system at DUC-2007

This paper presents the LIA summarization systems participating to DUC 2007. This is the second participation of the LIA at DUC and we will discuss our systems in both main and update tasks. The system proposed for the main task is the combination of seven different sentence selection systems. The fusion of the system outputs is made with a weighted graph where the cost functions integrate the votes of each system. The final summary corresponds to the best path in this graph. Our experiments corroborate the results we obtained at DUC 2006, the fusion of the multiple systems always outperforms the best system alone. The update task introduces a new kind of summarization, the over the time update summarization. We propose a cosine maximization-minimization approach. Our system relies on two main concepts. The first one is the cross summary redundancy removal which tempt to limit the redundancy between the update summary and the previous ones. The second concept is the novelty detection in a cluster of documents. In the DUC 2007 main and update evaluations, our systems obtained very good results in both automatic and human evaluations.