Songs Recommendation using Context-Based Semantic Similarity between Lyrics

With the rapid growth of the internet, many songs and music are readily available for users on various platforms. The number, however, gets so huge that the user might get overwhelmed when it comes to selecting a follow-up song. A recommender system comes in handy in such situations, where users can choose a recommended piece based on their likes and dislikes. There can be various metrics in developing a song recommender system, lyrics being one of them. In this paper, a song recommendation system is proposed on English songs, which uses the contextual embeddings to extract the context out of the song lyrics, identifies semantic similarity between these lyrics, and gives the most similar songs to the user based upon his choice. The dataset is taken from musixmatch.com, containing around 3300 songs. Following the pre-processing of the data, the context is extracted from the lyrics using Google's Universal Sentence Encoder algorithm. The proposed methodology achieves an F1 score of 0.7700, which shows that the accuracy of the proposed model is better than the available models in the literature for the song recommendation system using the lyrics of the songs.