Many recent experiments in the automatic classification of discourse relations have limited themselves to a small set of coarse categories. While there are eminent reasons to do so – annotation for a small sets of categories can be created more reliably, and possibly also be defined in a more clear – cut way than finer distinctions – it is an interesting question whether the finer-grained distinctions present in some annotated corpora can be reconstructed reliably. The present paper investigates the feasibility of such fine-grained tagging of discourse relations using data from the Penn Discourse Treebank. 1