Song Genre and Artist Classification via Supervised Learning from Lyrics

1
Song Genre and Artist Classification via
Supervised Learning from Lyrics

• Adam Sadovsky
• Xing Chen
• CS 224N Final Project

2
Introduction

• Goal develop a classifier that classifies songs
  by their genre and/or artist using only their
  lyrics
• Use EvilLyrics 1 program to build corpus of
  approximately fifteen popular albums per genre
  (rock, rap/hip-hop, and country)
• Use Maxent and SVM classifiers with cross
  validation to classify lyrics
• For Part-of-Speech (POS) features, we use the
  Stanford Log-linear Part-Of-Speech Tagger 2 to
  label all words in our corpus with a POS

3
Feature Selection
Look at differences between
Bag-of-Words artist diction and content Word
Endings artist style Line Length song pattern
and rhythm

• Country

Brooks Dunn - Again ain't it funny, the turns
life puts you through. don't know what's round
the bend, man, you don't know where it's leadin'
you. close your eyes, say a prayer, take it on
the chin. it's a dawn sun, comes back
again. baby, i thought that love was over and
gone forever... never gonna come back to
me. never gonna hold me again.
Pearl Jam - Come Back If I keep holding
out Will the light shine through? Under this
broken roof It's only rain that I feel I've been
wishin' out the days Oh oh oh Come back Know
that I still remain true I've been wishin' out
the days Please say that if you hadn't have gone
now I wouldn't have lost you another way From
wherever you are Oh oh oh oh Come back
PRP VBP VBG
Rock
Number of Lines song length Repetition style
and rhythm Punctuation writing
style Part-of-Speech statistics writing style
4
Genre Classification

• Attempt to distinguish between rap, rock and
  country

Performance
SVM Confusion Matrix country, rap, rock a
b c lt-- classified as 157 4 32 a
country 6 152 3 b rap 46 8 119
c rock
Classifier Accuracy ()
Maxent 76.45
SVM 81.21
Feature Performance
Best Alone (Maxent/SVM) Bag-of-words (75 /
72) Word endings (73 / 72) POS tags (61 /
61)
Most Significant Ablations Bag-of-Words (3 /
4) Word endings (3 / 3)
5
Classifying Artists

• Classifier might perform better when each group
  of lyrics is by the same artist
• Two new datasets beatles, u2, blink_182 (all
  rock) and snoop_dogg, beatles, garth_brooks
  (rap, rock, country)
• Results

Dataset Classifier Accuracy
beatles, u2, blink Maxent 0.70464
beatles, u2, blink SVM 0.72050
snoop, beatles, garth Maxent 0.83866
snoop, beatles, garth SVM 0.84115
6
References

• 1 http//www.evillabs.sk/evillyrics/
• 2 http//nlp.stanford.edu/software/tagger.shtml