Reformulations participate in structuring of discourse, especially in dialogues, and also contribute to the dynamics of the discourse. Reformulation is a signicant act which has to satisfy precise objectives. The purpose of our work is to automatically predict the reason for which a speaker performs a reformulation. We use a classication with eleven pragmatic functions inspired by the existing work and by the data analyzed. The reference data are built through manual and consensual annotations of spontaneous reformulations introduced by three markers (c'est-à-dire, je veux dire, disons) in French. The data are provided by spoken corpora and a corpus with forum discussions on health issues. We exploit supervised categorization algorithms and a set with several features (syntactic, formal, semantic and discursive) for the prediction of the reformulation categories. The distribution of sentences is not homogeneous across categories. The experiments are positioned at two levels: general and specic. Our results indicate that it is easier to predict the types of functions at the general level (the average F-measure is around 0.80), than at the level of individual categories (the average F-measure is around 0.40). We study the inuence of various parameters.