Doxels in context for retrieval: from structure to neighbours
Résumé
We present in this paper a new way of considering retrieval of structured documents: we intend to make the exploration of structured documents elements (doxels) easier by relying heavily on the inter-relations. These relations may be defined at creation time of the documents (like navigation or reference links), but may also be created by the information retrieval system (like kNN neighbours). Unlike Pagerank or HITS that separate features coming from links and elements coming from the content, we integrate these two aspects by defining relative specificity and exhaustivity between doxels. We use these features as well as the doxel content in a comprehensive matching process, in a way to facilitate the exploration of the result space by a user to select the relevant doxels and in addition by indicating potential good neighbours to view from one doxel. Results of experiments on the INEX2005 test collection are presented.