One of the main research challenges in semantic role labeling (SRL) is the development of applications for languages other than English. For Brazilian Portuguese, recent projects in lexical semantics are about to provide the necessary computational resources for research in this area. However, the amount of annotated data provided is not significant enough for successful supervised learning. Hence, we propose to use a semi-supervised approach capable of taking advantage of both annotated and unannotated data available. In this paper, we outline the methodology for the development of this SRL system, the same as the benchmark to be used to test its performance. © 2012 Springer-Verlag.