An Exploration of Mood Classification in the Million Songs Dataset

As the music consumption paradigm moves towards streamingservices, users have access to increasingly large catalogsof music. In this scenario, music classification playsan important role in music discovery. It enables, for example, search by genres or automatic playlist creation based on mood. In this work we study the classification of songmood, using features extracted from lyrics alone, basedon a vector space model representation. Previous work inthis area reached contradictory conclusions based on experimentscarried out using different datasets and evaluationmethodologies. In contrast, we use a large freelyavailabledataset to compare the performance of differentterm-weighting approaches from a classification perspective.The experiments we present show that lyrics can successfullybe used to classify music mood, achieving accuraciesof up to 70% in some cases. Moreover, contraryto other work, we show that the performance of the differentterm weighting approaches evaluated is not statisticallydifferent using the dataset considered. Finally, we discuss the limitations of the dataset used in this work, and the need for a new benchmark dataset to progress work in this area.