Recently, the identification of free connective phrases as signals for discourse relations has received new attention with the introduction of statistical models for their automatic extraction. The limited amount of annotations makes it still challenging to develop well-performing models. In our work, we want to overcome this limitation with semi-supervised learning from unlabeled news texts. We implement a self-supervised sequence labeling approach and filter its predictions by a second model trained to disambiguate signal candidates. With our novel model design, we report state-of-the-art results and in addition, achieve an average improvement of about 5\% for both exactly and partially matched alternativelylexicalized discourse signals due to weak supervision.