In this project I built a model do classify the genre of a music, based on its lyric. The model was trained with some popular brazilian music styles, such as Axé, Funk, MPB, Pagode, Samba and Sertanejo. I collected music lyrics in LETRAS.MUS.BR web site, I developed web scrappers, one for each music genre, using Python package BeaultifulSoup. Almost 1k lyrics was collected for each genre. The web crawlers codes are available here.

Genre Distribution:

First I followed these steps for Data Preparation:

• Text cleaning and normalization, removing html tags and special characters and turning all words in lower case.
• Data partition in train set (66%) and test set (33%).
• Targel label encoding, using sklearn LabelEncoder.

Then I built 4 different models:

• Bag-of-Words (TF-IDF): use of TF-IDF vectorizer, feature selection and Logistic Regression with penalty, using RandomizedSearch to tune trainable tasks.
• Continuous Skipgram (Pre-Trained - Brazilian Portuguese): pre-trained word embedding Continuous Skupgram with a Logistic Regression with penalty, using RandomizedSearch to tune trainable tasks.
• Continuous Skipgram (Task-Specific): training a word embedding Continuous Skupgram, using Gensim package, with a Logistic Regression with penalty, using RandomizedSearch to tune trainable tasks.
• BERT (Pre-Trained - Brazilian Portuguese): use of the BERT model pre-trained in Portuguese, fine-tuning with HuggingFace package class BertForSentenceClassifier.

Results:

• Confusion Matrix (BERT)

The Colab Notebook of this study is available here.

The data sets are available here.

• NLP
• Text Classification
• Continuous Skipgram
• BERT
• PyTorch
• RandomizedSearch
• BeautifulSoup