The study applies DBNs to a natural language understanding problem. Specifically, the authors compare a DBN-initialized neural network to three widely used text classification algorithms (SVM, Boosting and MaxEnt). Whereas the call–routing classification accuracy given by plain DBN-based model equals the best of the other models, recourse to additional unlabeled data for DBN pre–training and combining DBN–based learned features with the original features results into significant gains over SVMs (which performs performed better than MaxEnt and Boosting).