IMDb Sentiment Classifier using Hugging Face Transformers

A modular NLP pipeline for binary sentiment classification on IMDb movie reviews using Hugging Face Transformers. The project demonstrates scalable model evaluation, multilingual compatibility (tabularisai), and clean engineering practices.

Project Overview

• Goal: Predict whether a movie review expresses positive or negative sentiment
• Dataset: IMDb Dataset via Hugging Face datasets library
• Models:
  • distilbert-base-uncased-finetuned-sst-2-english
  • tabularisai/multilingual-sentiment-analysis
• Evaluation:
  • Multiclass accuracy using model-defined label IDs
  • Custom bucketed evaluation to map diverse outputs into binary classes (positive/negative)

Directory Structure

imdb-sentiment-classifier/ 
├── scripts/ 
│ └── run_pipeline.py # Entry point to run the full workflow 
├── src/ 
│ ├── load_data.py # Loads the dataset 
│ ├── preprocess.py # Tokenizes and decodes review texts │ ├── inference.py # Performs inference with chosen model
│ └── evaluate.py # Includes multiclass and binary evaluation 
├── outputs/ # Evaluation results and logs 
└── metrics_log.md # pasted results of metrics from demonstrated models 
├── requirements.txt # Project dependencies 
└── README.md # You are here

How to Run

Make sure you're in the project root directory and have Python 3.8+.

1. Install dependencies

   pip install -r requirements.txt

2. Run the pipeline

   python -m scripts/run_pipeline.py

Results

DistilBERT

Metric	Positive	Negative	Accuracy
Precision	0.89	0.95	0.92
Recall	0.93	0.90
F1-Score	0.91	0.92

The DistilBERT model achieved strong performance, with balanced precision and recall across both sentiment classes. Evaluation was based on a stratified sample of 500 reviews.

Tabularisai

When evaluated using label bucketing (e.g., grouping "Very Positive" and "Positive" together), the multilingual model achieved 100% accuracy on a reduced subset after removing neutral entries:

Classes Accuracy Positive vs Negative 1.00

Neutral predictions (not present in the original dataset) were excluded dynamically to enable fair binary classification.

Skills Demonstrated

Hugging Face Transformers & Pipelines

Tokenization, decoding, and review preprocessing

Label mapping and evaluation customization

Modular pipeline design for structured inference

Label bucketing

Metrics visualization: classification reports