LDA based emotion recognition from lyrics

Abstract

Music is one way to express emotion. Music can be felt/heard either using an instrument or as a song which is a combination of instrument and lyrics. Emotion Recognition in a song can be done either using musical features or lyrical features. But at times musical features may be misinterpreting, when the music dominates the lyrics. So a system is proposed to recognize emotion of the song using Latent Dirichlet Allocation (LDA) modelling technique. LDA is a probabilistic, statistical approach to document modelling that discovers latent semantic topics in large corpus. Since there is a chance of more than one emotion occurring in a song, LDA is used to determine the probability of each emotion in a given song. The sequences of N-gram words along with their probabilities are used as features to construct the LDA. The system is evaluated by conducting a manual survey and found to be 72% accurate. © Springer International Publishing Switzerland 2014.