It is important to avoid harms that may result from classifiers.
Representationalharms:harmscausedbyasystemthatdemeansasocialgroup.Forexample,AfricanAmericannamesaremorelikelytobeassignednegativeemotioninthesentimentanalysis.
Censorship harms: like in toxicity detection, false-positive errors could lead to the censoring of discourse about certain groups, like gay people and blind people.
AmodelcardforNLPcanhelptoclearandfindtheharms:Amodelcardincludesthefollowinginformation:•trainingalgorithmsandparameters•trainingdatasources,motivation,andpreprocessing•evaluationdatasources,motivation,andpreprocessing•intendeduseandusers•modelperformanceacrossdifferentdemographicorothergroupsandenvironmentalsituations
