Recent approaches for building syntactic language models include the combination of Probabilistic Tree Substitution Grammars (PTSGs) and Bayesian learning methods. While PTSGs have appealing features for syntax modeling, Bayesian methods provide a framework for inducing compact grammars that do not overfit the training corpus. In this paper, we apply these approaches to learn syntactic language models from a Brazilian Portuguese treebank. © 2012 Springer-Verlag.