Reading Contexts for Structured Documents Retrieval
Résumé
This paper focuses on the retrieval of parts of structured document called doxels. We propose a notion of reading context of a doxel and we exploit it to extend an Indexing Language Model (LM) with Dirichlet smoothing. We interpret a context of a doxel as a propagation of the content of the connected doxels via document structure links. We experiment this model on INEX corpus 2009, and test different context propagations. We measure a significant increase in results using contexts, compared to a reference approach without the use of context for 3 types of doxels. Moreover, our proposal outperforms the best result obtained for the Focused evaluation for the Ad Hoc task at INEX 2009.