Music genre classification with word and document vectors

In these days, music genre classification (MGC) is a quite popular research field. The main goal of the MGC studies is automatically detecting music genre (eg., rap, rock). In literature, features are generally extracted from the music's melodic content or lyrics for this task. In this study, we have performed lyrics based MGC on a Turkish dataset. We have just used lyrics as feature source and considered the MGC as a classical text classification problem. However, we represented the features using word (word2vec) and document (doc2vec) vector methods which are quite popular recently. Also, we have compared these methods with traditional Bag of Words (BoW) feature model. In addition, we have investigated the impact of preprocessing steps and vector dimension on both word and document vectors. We have conducted experiments on Support Vector Machine algorithm. Our experimental results show that word vector can be employed for feature representation.