Music Genre Classification by Lyrics using a Hierarchical Attention Network

We adapt the hierarchical attention network for the task of genre classification using lyrics. Utilising a large dataset of intact song lyrics we apply a recurrent neural network model which tries to learn importance of words, lines and sections in the genre classification task. This hierarchical structure attempts to replicate the structure of lyrics and enable learning of which sections, lines or words play importance in different genres. We explore the differences between taking layers at the line level or at the section level. We test the model over a 117 genre dataset and a smaller 20 genre dataset, classifying over a much higher number of genres than previous research. Experiments show that the hierarchical attention network outperforms basic non-neural models and in certain situations simple neural models, although high classification accuracies still present a tough challenge. We visualise the performance of the model by extracting from songs the lines and words it applies most weight to in the classification task.