NLP - Text category classifier

Marcos V. O. Assis (mvoassis@gmail.com)

---

Objective:

Generate an algorithm capable of classifying articles based on their titles.

Language:

For development and testing, words in Portuguese-BR were used.

Approach:

1. Download Word2Vec pre-trained CBOW model from NILC (cbow_s300)
2. Load the model using the Gensin library.
3. Vectorize the article's titles using the NLTK library and CBOW model.
4. Training a Logistic Regression classification model using the Scikit Learn library.

Categories considered:

• Mundo (World)
• Cotidiano (Daily life)
• Mercado (Market)
• Esporte (Sports)
• Ilustrada (Illustrated)
• Colunas (Column)

Files

• Train dataset -> Link - 90000 entries
• Test dataset -> Link - 20513 entries

Results

• 0.8 Accuracy score using CBOW (against 0.3 Accuracy from Dummy Classifier)
• Regarding individual categories, the proposed model achieved F1-Score:
  1. colunas - 0.78
  2. cotidiano - 0.69
  3. esporte - 0.90
  4. ilustrada - 0.23
  5. mercado - 0.81
  6. mundo - 0.79
• CBOW is slightly better than Skipgram for this problem.