The paper presents an unlexicalized probabilistic parsing model for German trained on the Negra treebank. Evaluation is performed with respect to constituency and dependency measures. It is observed that existing models based on Parent Encoding and Markovization optimize for constituency measures at the expense of dependency performance (at least in German). Several linguistically inspired transformation and annotation schemes are proposed which do help with dependency measures. Finally, it is shown that performance compares well with published results for German.