XFIRM at INEX 2005: ad-hoc and relevance feedback tracks. Karen Sauvagnat, Lobna Hlaoua, and Mohand Boughanem IRIT-SIG, 118 route de Narbonne, F-31 062 Toulouse Cedex 4, France Abstract. This paper describes experiments carried out with the XFIRM system in the INEX 2005 framework. The XFIRM system uses a rele- vance propagation method to answer CO and CAS queries. Runs were submitted to the ad-hoc and relevance feedback tracks. 1 Introduction The approach we used for our participation at INEX 2005 is based on the XFIRM system [6], and uses a relevance propagation method. The XFIRM system was adapted for submitting runs to the ad-hoc track (for CO, CO+S, and CAS tasks) and the relevance feedback track. 2 Experimental setup 2.1 XFIRM data model The XFIRM system is based on a relevance propagation method. We use a generic data model that allows the implementation of many IR models and the processing of heterogeneous collection. We consider that a structured document sd i is a tree, composed of leaf nodes ln ij and attributes a ij and simple nodes n ij (all nodes that are not leaf nodes or attributes). Structured document: sd i =({n ij } , {ln ij } , {a ij }) In order to easily browse the document tree and to quickly find ancestors- descendants relationships, the model uses a representation of nodes and at- tributes based on the Xpath Accelerator approach [2]. All leaf nodes are indexed, because even the smallest leaf nodes can be relevant or can give information on the relevance of their ancestors. Intuitively, title or subtitle nodes are not informative, but if a query term occurs in those nodes, such information can be useful for evaluating the relevance of the ancestor node. Such an approach has other advantages: the index process can be done automat- ically, without any human intervention and the system will be so able to handle heterogeneous collections automatically; and secondly, even the most specific query concerning the document structure will be processed, since all the docu- ment structure is stored. During query processing, relevance values are assigned to leaf nodes and rele- vance score of inner nodes are then computed dynamically, thanks to a propaga- tion of leaf nodes score through the document tree. An ordered list of subtrees is then returned to the user.