Language Models for XML Element Retrieval Rongmei Li 1 and Theo van der Weide 2 1 University of Twente, Enschede, The Netherlands 2 Radboud University, Nijmegen, The Netherlands Abstract. In this paper we describe our participation in the INEX 2009 ad-hoc track. We participated in all four retrieval tasks (thorough, fo- cused, relevant-in-context, best-in-context) and report initial findings based on a single set of measure for all tasks. In this first participa- tion, we test two ideas: (1) evaluate the performance of standard IR engines used in full document retrieval and XML element retrieval; (2) investigate if document structure can lead to more accurate and focused retrieval result. We find: 1) the full document retrieval outperforms the XML element retrieval using language model based on Dirichlet priors; 2) the element relevance score itself can be used to remove overlapping element results effectively. 1 Introduction INEX offers a framework for cross comparison among content-oriented XML retrieval approaches given the same test collections and evaluation measures. The INEX ad-hoc track is to evaluate system performance in retrieving relevant document components (e.g. XML elements or passages) for a given topic of request. The relevant results should discuss the topic exhaustively and have as little non-relevant information as possible (specific for the topic). The ad-hoc track includes four retrieval tasks: the Thorough task, the Focused task, the Relevant in Context task, and the Best in Context task. The 2009 collection is the English Wikipedia with XML format. The ad-hoc topics are created by the participants to represent real life information need. Each topic consists of five fields. The <title> field (CO query) is the same as the standard keyword query. The <castitle> field (CAS query) adds structural constraints to the CO query by explicitly specifying where to look and what to return. The <phrasetitle> field (Phrase query) presents explicitly a marked up query phrase. The <description> and <narrative> fields provide more information about topical context. Especially the <narrative> field is used for relevance assessment. The paper documents our first participation in the INEX 2009 ad-hoc track. Our aims are to: 1) evaluate the performance of standard IR engines (Indri search engine) used in full document retrieval and XML element retrieval; 2) investigate if document structure can lead to a more accurate and focused retrieval result. We adopt the language modeling approach [2] and tailor the estimate of query term generation from a document to an XML element according to the user