Lyrics Genre Classification: Turkish and Multilingual BERT Models
Two fine-tuned BERT models that predict a song's genre from its lyrics, one for Turkish and one multilingual. Both published open source on Hugging Face.
Type
Machine learning models
Period
July 2023 - August 2023
Role
Sole author
Stack
Python, PyTorch, Hugging Face, BERT, NLP
Two multilingual BERT models fine-tuned to predict a song's genre from its lyrics alone: 93% accuracy (F1 0.90) on the Turkish model and 94% (F1 0.92) on the multilingual one.
Lyrics datasets lean heavily towards a few popular genres, so minority classes were being ignored. Tokenization, resampling and data augmentation raised minority-class recall by 15%. Both models are open source on Hugging Face.
Highlights
Fine-tuned multilingual BERT models on lyrics datasets, reaching 93% accuracy (F1 0.90) on the Turkish model and 94% (F1 0.92) on the multilingual model.
Addressed class imbalance with tokenization, resampling and data augmentation, raising minority-class recall by 15%.
Published both models open source on Hugging Face.
