Matched by a 91.19% F1 score on the IMDb benchmark.
PROJECT 04 / NLP · Deployment
Sentiment Analyzer
A plain language tool that reads a written review and estimates whether the overall opinion is positive or negative, helping large volumes of feedback become easier to summarize.
DATASETIMDb Large Movie Review Dataset
SIZE50,000 labeled reviews · 25k train / 25k test
MEASURED EVIDENCE
What can be verified.
Metrics and outputs drawn from the project artifacts—not estimates added for presentation.
Improvement over the 89.46% TF-IDF and Logistic Regression baseline.
25,000 training and 25,000 test examples from IMDb.
Three epochs on an Apple M3 CPU after adapting to memory constraints.
VALIDATION
The public repository includes baseline comparisons, training details, failure analysis, reproducible inference code, and links to both the deployed application and model.
Why this project matters
Organizations can receive thousands of reviews, survey comments, and support messages, far more than a person can read quickly. This project began with the question of whether a language model could reliably summarize the overall tone of that feedback.
What was developed
The project fine tuned DistilBERT using 25,000 labeled IMDb movie reviews and built a Gradio interface where a user can enter text and receive a real time sentiment prediction. The training examples taught the model patterns commonly associated with positive and negative opinions.
What it means
The model reached 91.2% accuracy on the project benchmark and demonstrates how unstructured text can become a measurable signal. A business could use that signal to find broad trends, while people still review nuanced, urgent, or ambiguous comments.
04 / WHERE THIS WORK APPLIES
From project to practical use.
Sentiment analysis can support e commerce reviews, customer service monitoring, brand research, employee surveys, product feedback, hospitality, media analysis, and nonprofit or public program feedback.
05 / LIMITATION
What the project does not solve yet.
Training on IMDb reviews limits generalization to other domains, and the model may struggle with sarcasm, mixed sentiment, or language outside the training distribution.
06 / NEXT ITERATION
Future advancement
Add domain adaptive and multilingual training, confidence calibration, bias testing, and production monitoring for language drift.
TECHNICAL TOOLKIT
DistilBERTTransformersIMDbGradioPython