Seems like lemmatizing the text before it's rated by humans (during the training of the model, that is) would get around a healthy portion of the grammatically-induced scoring differences.