Conclusion and Future work:

In our project, we spend lots of time on data preparation. We are glad that with 400 data for each genre, we got pretty good results on classifying two different genres. For classifying three different genres, we can classify ~80% of song correctly with DNN. It is also quite interesting to extract features for each genre and got a brief idea of how lyrics of different genre type in common or different to each other. There are more works that we want to continue to do:

• Develop a classifier that can distinguish more genres. Now we have developed a genre classification system that can classify lyrics into Country, Rap and Jazz. In the future, we would like to add more genres such as Pop, Rock, Punk…etc. This could be quite challenging since some genres are not easily distinguishable if we only use TFIDF to get features. We might need to find additional features to complete the task.

• Since now we use TFIDF to extract the feature of the lyrics, we don’t consider the meaning of the sentence, or the meaning of sequence of the words. This could be one of the reason that we can’t achieve very high accuracy when classifying lyrics into three genres. In the future, we would like to try LSTM or RNN, to fit our model and improve our accuracy.