Annotated corpora constitute a crucial resource to acquire or induce linguistic knowledge about how languages are used. In this sense, it is widely admitted that tagged corpora appear to be a very useful resource for computational and linguistic analysis of languages. The more explicit linguistic information they contain, the more interesting and useful they are. In this paper we present the theoretical basis for semantic annotation of two treebanks, CESS-ESP and CESS-CAT, focusing specially on the verbal semantic classes that determine the mapping between syntactic functions and semantic roles.