In order to face different requirements at TALP Research Center we have built a highly parameterized environment allowing to instantiate specific summarizers for different summarization tasks in different languages. This paper describes and analyzes how our system deals with the DUC 2006 task of providing summary-length answers to complex questions. The given query is used to detect relevant passages. After that, semantic similarities between these relevant sentences are detected and then used as input of an iterative graph-based algorithm to avoid redundancy and obtain a cohesioned text. NIST human evaluations are used to analyze several aspects of our system and a specific analysis for each of the three different kinds of submitted summaries is reported.

Peer Reviewed