Anchors Regularized: Adding Robustness and Extensibility to Scalable Topic-Modeling Algorithms