In this paper we present some findings from an evaluation of dependency-based features for argument identification and classification in a pipelined semantic role labeling system. In total over 100 feature types were evaluated. Results indicate that most features can be discarded with sustained or even improved performance. We further find that arguments of nominal and verbal predicates seem to rely on different feature types - while arguments of verbal predicates rely more on structural features, arguments of nominal predicates rely more on lexical features.