Topic Model Approach to Song Classification Based on Lyrics

Abstract

The development of technology has allowed man to record enormous amount of information in digital format. But these unstructured and uncategorized overwhelming information provides very little use. So these information has to be categorized, in other words, classified. But it is impossible for human to go through all the information manually. So the machines have been given the task of processing information. One such popular method for text document processing is topic models. The applicability of Latent Dirichlet Allocation(LDA) in the domain of music information retrieval has been evaluated in this research. As the mood based song classification is one significant and difficult problem in music information retrieval, a topic model approach has been proposed by this research. The document collection used for the research was the song lyrics of artists in 60s to 90s. The methodology of the research in a short is as follows. The latent topics in the lyrics collection are uncovered using LDA algorithm and their respective topic distributions are derived. Once the topics are identified, they are automatically labeled according to their meaning by a novel algorithm which uses hierarchical clustering. Then each latent topic is assigned with a mood according to Russell s Circumflex of Affects. The topic distributions resulted in lyrics analysis are then used to classify a newly given lyric to its relevant topic. Once a lyric is assigned to a topic, the mood category that topic belongs, becomes the mood category of that particular lyric. Finally this methodology has been tested with three different experiments. Two of them were evaluated against user feedbacks and the other was evaluated against a ground truth data set build upon the user review tags of songs in the last.fm music website. Results of the experiments were compared with the previous works in the field of mood base song classification