In this paper, we are presenting a learning model for XML document classification based on Bayesian networks. Then, we are proposing a model which simplifies the arborescent representation of the XML document that we have, named coupled model and we will see that this approach improves the response time and keeps the same performances of the classification. Then, we will study an extension of this generative model to the discriminating model thanks to the formalism of the Fisher’s kernel. At last, we have applied a ponderation of the structure components of the Fisher’s vector. We finish by presenting the obtained results on the XML collection by using the CBS and SVM methods