German Semantic Role Labeling is suffering from sparse data. In 
particular, only a small corpus for VerbNet-style SRL for German is 
available: the Semantic Role Triple Corpus for German (SR3de). This 
corpus contains parallel data with VerbNet, FrameNet and ProbBank 
annotations. We extend the VerbNet-annotations using data from the other 
SRL frameworks, and show that a simple Multi-Task Learning architecture 
with similar auxiliary tasks (peer task) can improve the quality of the 
automatic SRL.