Data & ML / Systems & interfaces · 2025
A strong baseline for movie-review sentiment
A text-classification pipeline using TF-IDF and logistic regression, with a public interface for exploring predictions.
The Streamlit demo may need to be woken after inactivity. Source code is available above.
Context
Independent modeling project
My contribution
Text preprocessing, model tuning, held-out evaluation, and Streamlit interface.
The takeaway
Approximately 90.8% held-out accuracy with a comparatively simple text-classification pipeline.
The question
How far can a straightforward statistical text model go on sentiment classification? I built a TF-IDF and logistic-regression pipeline on the 50,000-review IMDb dataset.
Method
The project uses an 80/20 training/test split. Grid search tunes the vectorizer’s n-gram range and document-frequency settings alongside the classifier’s regularization. A Streamlit interface accepts review text and displays the resulting class and probability.
Findings and limits
The project reports approximately 90.8% accuracy on the held-out test set. That supports the baseline’s usefulness on this benchmark; it does not establish accuracy on news, social media, or other domains.
Sarcasm, mixed sentiment, and domain shifts remain limitations. The displayed probability is a model output, not a guarantee of correctness or a separately calibrated confidence estimate.