Apprentissage de conversion de documents semi-structurés à partir d'exemples

We propose here the method called ISM -Incremental Structure Mappingwhich allows one to convert XML documents from heterogeneous sources to a mediated schema. Unlike existing methods, ISM is document centric and takes into account both the structural information and the content information. It does not need to specify manually correspondences between schema and is based on Machine Learning methods in order to transform documents to the mediated schema. The method learns the transformation using a set of documents expressed in both the input schema and the output schema. At last, ISM has a low complexity and can be used with very large XML collections. This methods is experienced here on a set of corpus and different tasks. The results show that the algorithm is able to learn complex transformations and to transform large corpora, particularly for the conversion of HTML documents to semantically