We propose a new approach to perform semi-supervised training of Semantic Role Labeling models with very few amount of initial labeled data. The proposed approach combines in a novel way supervised and unsupervised training, by forcing the supervised classifier to over-generate potential semantic candidates, and then letting unsupervised inference choose the best ones. Hence, the supervised classifier can be trained on a very small corpus and with coarse-grain features, because its precision does not need to be high: its role is mainly to constrain Bayesian inference to explore only a limited part of the full search space. This approach is evaluated on French and English. In both cases, it achieves very good performance and outperforms a strong supervised baseline when only a small number of annotated sentences is available and even without using any previously trained syntactic parser.