Erion Cano and Maurizio Morisio, Polytechnic University of Turin, Italy

Music emotion recognition today is based on techniques that require high quality and large emotionally labeled sets of songs to train algorithms. Manual and professional annotations of songs are costly and hardly accomplished. There is a high need for datasets that are public, highly polarized, large in size and following popular emotion representation models. In this paper we present the steps we followed to create two such datasets using intelligence of last.fm community tags. In the first dataset, songs are categorized based on an emotion space of four clusters we adopted from literature observations. The second dataset discriminates between positive and negative songs only. We also observed that last.fm mood tags are biased towards positive emotions. This imbalance of tags was reflected in cluster sizes of the resulting datasets we obtained; they contain more positive songs than negative ones.

Music Sentiment Analysis, Ground-truth Dataset, User Affect Tags, Semantic Mood Spaces