Music Mood Classification via Deep Belief Network

In this paper we present a study on music mood classification by using only lyrics information. Specially considering the Chinese songs, the Chinese word-segmentation has caused intolerable errors and inadequate use of lyrics information. Our work proposes to use bag-of-character features instead of bag-of-word features to avoid the word segmentation error, which makes the classification more inaccurate. Further more, the use of DBN which trains useful features automatically can make more use of lyrics information. With the experiments on 1200 songs collected, we demonstrate that our high level features trained by DBN and joint bag-of-character perform much better than the traditional features in music mood classification.