Larger corpus, deeper model, better scoring function on models, more models