Music mood classification based on lyrical analysis of Hindi songs using Latent Dirichlet Allocation

For over a decade now, due to the introduction of UTF-8 encoding, the digitization of Hindi content has increased rapidly because of which Hindi-music has accomplished popularity on the web. The focus is to identify the emotion, a person is experiencing while listening to a song track. The aim of this research work is to analyze the lyrics of Hindi-language based songs, in order to detect the mood of the listener. We used unigram and term-frequency as the main features. The songs were reduced to a level where only relevant words will be used for mood-detection. We employ unsupervised machine learning namely topic-modeling (Latent Dirichlet Allocation model) for mining the mood out of every song in the corpus. We created our own dataset of 1900 songs consisting of Bollywood tracks, bhajans (spiritual prayers) and ghazals. A mood taxonomy is used to distinguish songs into Happy or Sad. Data is applied to LDA model to discover the hidden emotions within each song. At the end of experimentation, we compare the results with manually pre-annotated dataset for validation purpose and observe good results.