IMDB Sentiment Analysis (RNN)

A recurrent neural network built with PyTorch to classify IMDB movie reviews as positive or negative.

Overview

This project implements a text preprocessing pipeline and an RNN from scratch to perform binary sentiment classification on the IMDB Dataset of 50,000 movie reviews.

Dataset

• IMDB Dataset of 50K Movie Reviews
• 49,582 reviews after removing duplicates (25k+ positive, 25k+ negative)
• Labels encoded via LabelEncoder (positive = 1, negative = 0)

Preprocessing

• Lowercasing
• URL removal
• Punctuation removal
• HTML tag removal
• Stopword removal (NLTK)
• Stemming (Porter Stemmer)
• Vectorization via TF-IDF (TfidfVectorizer, max 5000 features)

Model Architecture

• Single-layer RNN (nn.RNN), hidden size 128
• Fully connected output layer (128 → 1)
• Sigmoid activation for binary classification

Training

• Loss function: Binary Cross-Entropy Loss
• Optimizer: Adam
• Epochs: 10
• Batch size: 64

Results

• Test accuracy: 85.61%

Requirements

torch
scikit-learn
nltk
pandas

Usage

python train.py

Trains the RNN on the preprocessed reviews and evaluates accuracy on the held-out test set.

Author

Samir B K GitHub: github.com/Samir-BK