Recursive question decomposition for answering complex geographic questions

Abstract

This paper describes the GIRSA-WP system and the experiments performed for GikiCLEF 2009, the geographic information retrieval task in the question answering track at CLEF 2009. Three runs were submitted. The first one contained only results from the InSicht QA system; it showed high precision, but low recall. The combination with results from the GIR system GIRSA increased recall considerably, but reduced precision. The second run used a standard IR query, while the third run combined such queries with a Boolean query with selected keywords. The evaluation showed that the third run achieved significantly higher mean average precision (MAP) than the second run. In both cases, integrating GIR methods and QA methods was successful in combining their strengths (high precision of deep QA, high recall of GIR), resulting in the third-best performance of automatic runs in GikiCLEF. The overall performance still leaves room for improvements. For example, the multilingual approach is too simple. All processing is done in only one Wikipedia (the German one); results for the nine other languages are collected by following the translation links in Wikipedia. © 2010 Springer-Verlag Berlin Heidelberg.