In this paper, we study the use of a probabilistic fusion approach inspired from the probFuse algorithm [14] . Our fusing technique is based on queries that are classified using some automatically extracted linguistic features [11] . In this approach, performances which are estimated during a training phase are used as an indicator of the systems to fuse. The rank list of relevant documents provided by the systems are divided into segments and used in a training/testing process to detect the

more »

... tems to fuse. Our fusion technique shows better performances than other fusion techniques like CombMNZ [10], ProbFuse [14] and Best_sys [11] .