Educational resources can be easily found on the Web. Most search engines base their algorithms on a resource's text or popularity, requiring teachers to navigate the results until they find an appropriate resource. This makes searching for resources a tedious and cumbersome task. Specialized repositories contain resources that are annotated with metadata that aim to facilitate the discovery of quality resources. Nevertheless, the abundance and variety of content make searching a complex task. Recommender systems can assist teachers in finding the proper content by determining clusters of similar users and inferring users interest in a resource, assembling clusters of similar resources, or a mix of both. Probabilistic model-based (PMB) techniques, on the other hand, make it possible to classify resources into more than one cluster with various degrees of probability. In this paper, we used recommender systems and PMB techniques to analyze a dataset produced by primary and secondary level teachers over the course of four years and under natural conditions. We found that a hybrid recommendation, the Collaborative Topic Regression (CTR) technique, performs better than other approaches despite the high sparseness of the dataset. In addition, learning resources annotated with curriculum metadata had a positive impact on recommender's accuracy whereas free-text or other metadata negatively impacted the results.