Letterboxd Movie Sentiment API

Solo project · 2025 · deployed

#nlp#machine-learning#python#deployed

GitHub →

Overview

Letterboxd reviews are not straightforward: people joke, use slang, flip meanings, and the star rating doesn’t always match the text. This project tests how far a sentiment system can get using weak labels from star ratings, a simple baseline model, a transformer fine-tuned on the same data, and a small human-labeled evaluation set — served through an actual API.

The dataset is a scraped collection of Letterboxd reviews with movie name, release year, review text, star ratings (including broken Unicode characters that needed repair), review date, and like/comment counts.

Architecture

Raw scraped reviews
  -> Fix corrupted star-rating characters, convert to numeric
  -> Build strict binary weak labels from star ratings
  -> Clean review text, parse dates, drop unusable rows
  -> TF-IDF baseline model
  -> Fine-tuned DistilBERT model
  -> FastAPI backend serving predictions

Results

After cleaning, the dataset held numeric star ratings, clean review text, a loose 3-class sentiment label, and a strict binary sentiment label, across roughly 3,500 total reviews. A 49-sample manually labeled set (independent of the star-rating weak labels) was used to check whether the weak-label assumption actually held up against human judgment — the kind of check that’s easy to skip and easy to regret skipping.

Code Highlights

Keeping the weak-label pipeline (derived from star ratings) and the manual evaluation set (derived from actual human reading) as two clearly separate data sources — rather than treating the weak labels as ground truth — is what makes the accuracy claims here honest rather than circular.