Identifying the Emotional Polarity of Song Lyrics through Natural Language Processing

The analysis of song lyrics can provide important meta-information about a song, such as its genre, emotion, and theme; however, obtaining this information can be difficult. Classifying a song based solely on its lyrics can be challenging for a variety of reasons, and as yet no single algorithm results in highly accurate classifications. In this paper, we examine why the task of classification based solely on lyrics is so challenging by predicting the emotional polarity of a song based on the song’s lyrics. We use natural language processing to classify a song as either positive or negative according to four sets of frequency-based or machine learning algorithms which we develop and test on a dataset of 420 songs. We determine which factors complicate classification, such as generic subjectivity lexicons and non-subjective lyrics, and suggest strategies for overcoming these difficulties.