Turkish lyrics mining for music meta-data estimation

Music retrieval has become an important problem with the widespread use of internet and related technologies for entertainment purposes. Music retrieval systems were developed for users to find songs they are looking for and similar ones in an easier manner, and list songs they might want to listen. Music objects should be analyzed and interpreted according to those analyses independent of the method that is going to be implemented. These studies on music analysis are mainly focused on two data types; content signal that is based on melodic and musical arrangement properties for music retrieval systems and meta-data information, such as name, genre, composer of the song. The use of lyrics text is very few. This study provides a basis for the prediction of meta-data of music from lyrics text in music retrieval applications. Features were chosen on the song lyrics data sets prepared according to the Turkish text and grammar structure. A system that can predict the writer, genre and relaese date of the song using the chosen features and a machine learning algorithm was presented and its performence on a large song data set generated from song writers with different styles was evaluated. Results show that this kind of an approach might be useful for music data mining and information retrieval studies.