The paper describes a new tagset for the morphological disambiguation of Sanskrit, and compares the accuracy of two machine learning methods (CRF, deep recurrent neural networks) for this task, with a special focus on how to model the lexicographic information. It reports a significant improvement over previously published results.