Paciic Association for Computational Linguistics toward Pragmatic Use of Evaluation Resources

Natural language processing requires more and more voluminous textual corpora, and lots of evaluations are led in this eld. So, from the viewpoint of improving heterogeneous corpora (unrestricted texts) morpho-syntactic tagging, an adaptive approach to the type of text using resources provided by an evaluation will be presented. To justify this approach, basic statistics are applied on French MULTITAG data. This exploration reveals a few striking facts. A given text could have positive variations for a tagger negative for another. A priori typology does not explain some performance variations. So, it is necessary to build typology correlated to performance. Characterization function obtained by character set gives some clues to explain variations in performances.