Simplifying emotion classification through emotion distillation

Many state-of-the-art emotion classification systems are computationally complex. In this paper we present an emotion distillation framework that decreases the need for computational complex algorithms while maintaining rich, and interpretable, emotional descriptors. These representations are important for emotionally-aware interfaces, which we will increasingly see in technologies such as mobile devices with personalized interaction paradigms and in behavioral informatics. In both cases these technologies require the rapid distillation of vast amounts of data to identify emotionally salient portions. We demonstrate that emotion distillation can produce rich emotional descriptors that serve as an input to simple classification techniques. This system obtains results that match state-of-the-art classification results on the USC IEMOCAP data.