Project idea · ungraded

Review Sentiment Analyser

Build a sentiment analyser for product or movie reviews that predicts positive or negative, explains its decision with the words that mattered, and reports honestly where it fails (negation, sarcasm, mixed reviews).

You will use

Milestones

Build the smallest working version first. Ticks are saved in this browser only.

01Load a labelled review dataset and look at the class balance and typical review length.
02Train the TF-IDF + logistic regression pipeline and report cross-validated accuracy and F1.
03List the top positive and negative phrases and check that they make sense.
04Write an explain(review) function that shows each word's contribution to the score.
05Collect 30 misclassified reviews and group them by cause (negation, sarcasm, wrong label, ...).

Stretch goals

Compare with the embedding-bag model in PyTorch.
Add a third class for neutral or mixed reviews.
Serve the model with FastAPI as in the 'Serve your model' project.

Data

IMDB movie reviews (50,000, balanced): on Hugging Face or Kaggle.
Or write 100 short reviews of things you own and label them yourself.

Run it yourself

pip install scikit-learn pandas datasets
python sentiment.py