Information Theoretic Retrieval with Structured Queries and Documents

Information retrieval through statistical language modeling has become popular thanks to its firm theoretical background and good retrieval performance. One goal of current research on structured information retrieval is thus to extend such models to take advantage of structure information. As a structure may be present on documents or queries or both, we are interested in supporting not only unstructured queries on structured documents, but also structured queries on unstructured documents as well as structured queries on structured documents. Most of research work has considered the first task, i.e., unstructured queries over structured docs, while some papers have addressed using structured or semistructured queries on unstructured docs. Here we take a unified approach. Our basic retrieval model is the well known Kullback-Leibler divergence, with backoff smoothing. In this paper we show how it can be extended to model and support structured/unstructured queries on structured/ unstructured documents. We make a very general assumption on the type of structure imposed on queries and/or documents, suitable for describing various structured data. We also study how the extended model can be efficiently computed. We finally report on our experiments at INEX 2006, in which we used a rough approximation of the presented model. A full implementation of the model and a more significant evaluation of its retrieval effectiveness are left for future work.