We address the problem of learning to map automatically flat and semi-structured documents onto a mediated target XML schema. We propose a machine learning approach where the mapping between input and target documents is learned from examples. Complex transformations can be learned using only pairs of input and corresponding target documents. From a machine learning point of view, the structure mapping task raises important complexity challenges. Hence we propose an original model which scales well to real world applications. We provide learning and inference procedures with low complexity. The model sequentially builds the target XML document by processing the input document node per node. We demonstrate the efficiency of our model on two structure mapping tasks. Up to our knowledge, there are no other model yet able to solve these tasks.