Social media data are already being used to classify individuals into mutually exclusive social groups. Here we propose a model based on self-categorization theory that classifies which of two social identities is salient within the same person using text data. Based on over 500,000 online forum posts and seven prototype-based style features, a trained classifier correctly distinguishes between posts written by the same person in two different social contexts – a parenting forum and a feminist forum – significantly above chance level (AUC = .74). We then apply the trained classifier to a new dataset (N = 153) obtained from an online experiment where salience of group membership is manipulated. We show that our trained model distinguishes between salient parent and feminist identities significantly above chance level when the topic is irrelevant to either identity (AUC = .69). We discuss applications but also limitations of a text-based prediction of salient social identities.