Our annotators are very simple like preprocessors, I think they would also benefit from becoming scikit-learn transformers, particularly allowing batch processing.