Treebanks are a valuable resource for the training of parsers that perform automatic annotation of unseen data. It has been shown that changes in the representation of linguistic annotation have an impact on the performance of a certain annotation task. We focus on the task of Topological Field Parsing for German using Probabilistic Context-Free Grammars in the present research. We investigate an iterative algorithm for tuning the label set of a given treebank to this task and show that the number of parses proposed by a context-free grammar is reduced considerably in addition to an increase in labeled precision and recall for the annotation of node labels. We also show that the optimal refinement can be achieved with a relatively small number of changes to the treebank. 1