Publication day! Hand-code fewer documents and get better model performance with, basically, stratified sampling of training observations: doi.org/10.1017/pan.20…