Skip to main content

Olist Reviews NLP

NLP
Deep Learning
Predicting review scores (1–5★) from Portuguese review text: TF-IDF + classical models vs. fine-tuned BERTimbau.

Problem

Predict the review score (ordinal target, 1 to 5 stars) from text written by customers of a Brazilian e-commerce, and understand what drives the most severe complaints.

Approach

Classical approach (TF-IDF with Logistic Regression, SVM and LightGBM) compared with fine-tuning a Portuguese transformer (BERTimbau), plus topic extraction from complaints.

Technologies

Python scikit-learn LightGBM Transformers BERTimbau PyTorch Streamlit

Results

Fine-tuned BERTimbau reached QWK 0.8640, versus 0.8169 for TF-IDF + SVM. But the SVM is about 147x faster at inference and needs no GPU, and the project records that trade-off, not just the winner.

Back to top