A Textual Approach Based on Passages Using IR-n in WikipediaMM Task 2008

In this paper we have focused our efforts on comparing the behaviour of two relevance feedback methods in this task LCA and PRF and in checking if our passage based information rerieval (IR) system is useful in a competition with small sized documents. Furthermore we have added an adaptation to this domain based on decompound in single terms those file names which use a Camel Case notation. We base our decision on the belief that the most meaningful information of an image file appointed by a human is on the file name itself. Thus, it is important to make visible this terms when they are hidden in a compounded file name. Finally we have added a geographical query expansion and a visual concept expansion. We have obtained a 29th place within a total of 77 runs with our baseline run which only used the passage IR system -, and a 3rd place obtained with our best run which used the passage IR system with Camel Case decompounding -. It shows us on one hand the usefulness of our passage based IR system in this domain, and on the other hand it confirms our belief in the existence of specially meaningful information within the file names. In the the relevance feedback respect, we have obtained contradictory results about the suitability of LCA or PRF to the task, but we have found that LCA has a more robust behavior than PRF.