Emotion classification plays a key role in understanding human communications in the modern era of social media. While previous
research have focused mainly on eight or fewer emotions, in this work we introduce AffectVec, an embedding incorporating 68 emotions. Our approach improves previous systems in a number of ways; first, the fine-grained vectors better represent the richness of human motions, second, unlike common vector representation techniques the scores on each dimension are interpretable. We compare our embedding to human-annotated corpora and vector representations in a series of statistical and machine learning experiments.