Continual Model Refinement for Extractive Question Answering