🎯 Sentiment Classifier — NLP Showcase

An end-to-end NLP project that compares four classical machine-learning models on real customer reviews, wrapped in an interactive web demo.

This is a refreshed version of a project I originally built for CS4563 at NYU Tandon. It takes any short piece of text — a product review, a tweet, a comment — and predicts whether the sentiment is positive or negative, using four different classical ML models trained on 3,000 labelled reviews from Amazon, IMDB and Yelp.

---

✨ Highlights

• 🔮 Live interactive demo — type any sentence and watch all four models vote in real time, with confidence bars and a side-by-side comparison.
• 📊 Model showdown — accuracy, precision, recall and F1 for Logistic Regression, Linear SVM, K-Nearest Neighbors and Multinomial Naive Bayes, plus confusion matrices for each.
• 🔍 Explainability — see the top positive and negative words each linear model latched onto.
• 🗂 Dataset explorer — browse the corpus, filter by source and sentiment, inspect class balance.
• 🧱 Clean, modular codebase — refactored from the original Colab notebook into a proper Python package with reproducible training.

🚀 Run it

# 1. install dependencies (uses pyproject.toml)
pip install -e .

# 2. train the models (writes pipelines to ./models/)
python train.py

# 3. launch the demo
streamlit run app.py

The app starts on http://localhost:8501 by default (or port 5000 when run inside this repo's bundled .streamlit/config.toml).

📈 Results

Trained on an 80/20 stratified split of the UCI Sentiment Labelled Sentences corpus (3,000 reviews):

Model	Accuracy	Precision	Recall	F1
Logistic Regression	0.84	0.83	0.85	0.84
Linear SVM	0.83	0.82	0.84	0.83
Multinomial Naive Bayes	0.82	0.82	0.82	0.82
K-Nearest Neighbors	0.80	0.80	0.79	0.79

🏗 How it works

 raw reviews ──► clean_text() ──► TF-IDF (1–2 grams) ──► [LogReg │ SVM │ KNN │ NB]
                                                              │
                                                              ▼
                                                  joblib artefacts in ./models/
                                                              │
                                                              ▼
                                                Streamlit live demo (app.py)

1. Data — UCI Sentiment Labelled Sentences, 1,000 reviews each from Amazon, IMDB and Yelp.
2. Preprocessing — lowercase, strip non-alphanumerics, collapse whitespace. Shared between training and inference so there's no train/serve skew.
3. Features — TF-IDF over unigrams + bigrams with min_df=2 and sublinear term frequency. Replaces the original word2vec setup with something tiny enough to ship inside a web app.
4. Models — four classifiers, each wrapped in its own sklearn.Pipeline so vectoriser and classifier are saved as one artefact.
5. Evaluation — accuracy, precision, recall, F1 and confusion matrices on the held-out test set; metrics persisted as JSON for the demo.

📁 Project layout

.
├── app.py                       # Streamlit demo (live UI)
├── train.py                     # Train + persist all models
├── src/sentiment/
│   ├── data.py                  # Dataset loading & text cleaning
│   └── models.py                # Pipelines, training, inference helpers
├── data/raw/                    # UCI sentiment dataset (3 .txt files)
├── models/                      # Trained pipelines + metrics.json
├── MLProject_trainedw2v.ipynb   # Original NYU notebook (kept for reference)
└── pyproject.toml

🧠 What changed from the original notebook

The original notebook was built for Google Colab and mixed exploration, training and evaluation into one long file with hard-coded Drive paths. This refresh:

• Splits everything into a small, importable Python package.
• Replaces word2vec averaging with TF-IDF n-grams — same idea (turn text into vectors) but a fraction of the size and dependency footprint, with comparable or better accuracy on this corpus.
• Wraps every model in a Pipeline so the saved artefact handles preprocessing and prediction together.
• Adds a polished web demo so the project is something you can try, not just read.

👤 Author

Mohammad Asfour — CS student at NYU Tandon School of Engineering. Built as a course project for CS4563 (Introduction to Machine Learning), later refactored into this standalone showcase.

📜 License

MIT — feel free to fork, learn from it, or build on top.