Multilingual Experiments of UTA at CLEF 2003: The Impact of Different Merging Strategies and Word Normalizing Tools

There are two main translation approaches in a multilingual information retrieval task: either to translate the topics or to translate the datasets. The first one is an easier and more common approach. There are two indexing approaches: either to index the languages in the same index, or to build separate indexes for different languages. If the latter approach is used, retrieved result sets must be merged. Several merging strategies have been developed, but there have been no breakthroughs. University of Tampere used the raw score approach as the baseline in the CLEF 2003 runs and tested two new merging approaches. No remarkable differences were found between the merging methods. The index words can be stored as such (inflected index), or normalized. The most common normalizing method in IR is stemming. The tests of University of Tampere show that stemming is a good normalizing method for English, which is a language with weak morphology. For Finnish stemming gives remarkably worse results, the index built with the morphological analyzer gave better results.