Emotional analysis of song lyrics using NLP produceed better music recommendation results based on song moods

Role and Responsibilities

As the part of my computer science degree, I delivered a capstone project designing a song recommendation system that used emtoional sentiment analysis techniques as a part of broader natural language processing research to make playlists for users based on their provided mood.

Process and Approach

Over the course of 8 months, there were multiple pillars required for the recommendation system to be successful. Multiple APIs and song datasets were cross-referenced and pulled from in order to populate a training dataset.

GeniusAPI and Last.FM API were used to gather millions of song IDs that were web-scraped to pull lyrics from since most datasets don't publically share song lyric data.

Once the data was scraped, cleaned and ingested, it was then trained against a set of emotional lexicon text data from NRC.

Multiprocessing technologies such as Apache Spark were used to handle multiple streams of data at once. Python was used to calls the APIs and scrape the web.

Results

• A dataset of 200,000 songs was emotionally analyzed and referenced to create recommendation systems based on a user's selected mood
• Proof of concept that streaming platform music recommendation systems can be further advanced and modulated for optimal user satisfaction