🧠 RNN Sentiment Analysis

Teaching a Recurrent Neural Network to read between the lines of a movie review

Raw Text → NLP Cleaning → TF-IDF Vector → PyTorch RNN → 😊 / 😞

📖 Table of Contents

• Overview
• Architecture
• The Model
• Pipeline
• Results
• Tech Stack
• Project Structure
• Getting Started
• What This Demonstrates
• Author

⚡ Overview

A Recurrent Neural Network, built from scratch in PyTorch, that reads an IMDb movie review and decides — did the viewer love it, or hate it?

This project walks through the complete deep-learning workflow for text classification, end to end:

  Preprocess  →  Vectorize  →  Tensorize  →  Build RNN  →  Train  →  Evaluate

No shortcuts, no black-box libraries doing the thinking — the RNN, the loss, the training loop, all handwritten and transparent.

🧩 Architecture

flowchart LR
    A["📝 Raw Review"] --> B["🧹 NLP Cleaning"]
    B --> C["🔢 TF-IDF\n5,000 Features"]
    C --> D["🔥 PyTorch RNN\nHidden Size 128"]
    D --> E["📐 Linear Layer"]
    E --> F["🎯 Positive / Negative"]

    style A fill:#2b2f38,stroke:#8A2BE2,color:#fff
    style B fill:#2b2f38,stroke:#8A2BE2,color:#fff
    style C fill:#2b2f38,stroke:#EE4C2C,color:#fff
    style D fill:#2b2f38,stroke:#EE4C2C,color:#fff
    style E fill:#2b2f38,stroke:#2EB872,color:#fff
    style F fill:#2b2f38,stroke:#2EB872,color:#fff

Loading

🧠 The Model

The heart of this project is a single nn.RNN layer doing the heavy lifting.

⚙️ Parameter	🔧 Value
Architecture	nn.RNN
Hidden Size	128
Input Features	5,000
Loss Function	BCELoss
Optimizer	Adam
Epochs	10
Output	Binary Classification
Test Accuracy	85.16%

🔄 Pipeline

flowchart TD
    A["📦 IMDb Dataset"] --> B["🔡 Lowercase"]
    B --> C["🧽 Strip URLs / HTML / Punctuation"]
    C --> D["🚫 Remove Stopwords"]
    D --> E["🌱 Porter Stemming"]
    E --> F["🔢 TF-IDF Vectorization"]
    F --> G["📊 TensorDataset"]
    G --> H["🔁 DataLoader"]
    H --> I["🔥 PyTorch RNN"]
    I --> J["🎯 Binary Output"]

    style A fill:#1e2127,stroke:#4d90fe,color:#fff
    style J fill:#1e2127,stroke:#2EB872,color:#fff

Loading

📊 Results

🎯 85.16% Test Accuracy

RNN  ████████████████████████████████████████████░░░░░  85.16%

Trained for 10 epochs with Adam + BCELoss, evaluated on a held-out IMDb test split.

🛠️ Tech Stack

Icon	Technology	Role
🐍	Python	Core programming language
🔥	PyTorch	RNN architecture & training
🔤	NLTK	NLP preprocessing & stemming
📊	Scikit-learn	TF-IDF vectorization & data splitting
🐼	Pandas	Data loading & handling
📓	Jupyter / Colab	Notebook development

🗂️ Project Structure

RNN-for-Sentiment-Analysis/
│
├── 📓 imdb-sentiment-analysis-rnn.ipynb    # Full training & evaluation notebook
├── 📄 IMDB Dataset.csv                     # Raw movie review dataset
└── 📘 README.md                            # You are here

🚀 Getting Started

1. Clone the repository

git clone https://github.com/sumitjhadev/RNN-for-Sentiment-Analysis.git
cd RNN-for-Sentiment-Analysis

2. Install dependencies

pip install pandas nltk scikit-learn torch

3. Launch the notebook

jupyter notebook

Then open imdb-sentiment-analysis-rnn.ipynb and run the cells top to bottom. 🎬

✨ What This Demonstrates

• 🧠 Building a Recurrent Neural Network from scratch in PyTorch
• 🔢 Converting raw text into numerical features with TF-IDF
• 🧹 Practical, real-world NLP preprocessing
• ⚙️ Training with Adam optimizer + BCELoss
• 📦 Using PyTorch TensorDataset and DataLoader
• 📊 Evaluating a binary text classification model
• 🎯 Hitting 85.16% test accuracy

🏁 Final Takeaway

The dataset gives the task — but the real story here is the RNN: architecture, training loop, and evaluation, built and understood from the ground up.

🧠 BUILD · TRAIN · EVALUATE · LEARN

---

👤 Author

Sumit Jha 🔗 @sumitjha.ai  •  💻 github.com/sumitjhadev

⭐ If this project helped you understand RNNs better, consider giving it a star! ⭐