Despite the excellent performance of black box approaches to modeling\nsentiment and emotion, lexica (sets of informative words and associated\nweights) that characterize different emotions are indispensable to the NLP\ncommunity because they allow for interpretable and robust predictions. Emotion\nanalysis of text is increasing in popularity in NLP; however, manually creating\nlexica for psychological constructs such as empathy has proven difficult. This\npaper automatically creates empathy word ratings from document-level ratings.\nThe underlying problem of learning word ratings from higher-level supervision\nhas to date only been addressed in an ad hoc fashion and has not used deep\nlearning methods. We systematically compare a number of approaches to learning\nword ratings from higher-level supervision against a Mixed-Level Feed Forward\nNetwork (MLFFN), which we find performs best, and use the MLFFN to create the\nfirst-ever empathy lexicon. We then use Signed Spectral Clustering to gain\ninsights into the resulting words. The empathy and distress lexica are publicly\navailable at: http://www.wwbp.org/lexica.html.\n