Music subject classification based on lyrics and user interpretations

That music seekers consider song subject metadata to be helpful in their searching/browsing experience has been noted in prior published research. In an effort to develop a subject‐based tagging system, we explored the creation of automatically generated song subject classifications. Our classifications were derived from two different sources of song‐related text: 1) lyrics; and 2) user interpretations of lyrics collected from songmeanings.com. While both sources contain subject‐related information, we found that user‐generated interpretations always outperformed lyrics in terms of classification accuracy. This suggests that user interpretations are more useful in the subject classification task than lyrics because the semantically ambiguous poetic nature of lyrics tends to confuse classifiers. An examination of top‐ranked terms and confusion matrices supported our contention that users' interpretations work better for detecting the meaning of songs than what is conveyed through lyrics.