Sentiment Analysis Using BERT, LSTM, GRU, and RNN

A comparative analysis of modern and traditional deep learning architectures for sentiment classification on the Sentiment140 dataset.

---

📋 Table of Contents

• Project Overview
• Dataset Details
• Automatic Dataset Setup
• Workflow
• Models Architectures
• Installation and Setup
• Project Structure
• Usage
• Results & Comparison
• License

---

🔍 Project Overview

This project focuses on identifying the sentiment (Positive or Negative) of Twitter messages. We evaluate multiple deep learning models to compare their performance, training time, and accuracy:

1. BERT (Bidirectional Encoder Representations from Transformers) - State-of-the-art transformer based approach.
2. LSTM (Long Short-Term Memory) - Captures long-term dependencies in text.
3. GRU (Gated Recurrent Unit) - A faster alternative to LSTM with similar gating mechanisms.
4. RNN (Recurrent Neural Network) - The foundational sequence model.

---

📊 Dataset Details

The Sentiment140 dataset contains 1,600,000 tweets extracted using the Twitter API.

• Source: Kaggle Sentiment140
• Labels:
  • 0 -> Negative
  • 4 -> Positive
• Features: Tweet ID, Date, User, and the actual Text.

---

📥 Automatic Dataset Setup

Since the dataset is large (~145MB), this repository includes an automated download and organization system:

• Automatic Retrieval: Uses kagglehub to download the latest version directly from Kaggle.
• Local Persistence: The notebook and scripts automatically copy the downloaded CSV to a root-level Dataset/ folder for easy access and consistency.

---

⚙️ Workflow

The project follows a modular pipeline for data processing and model evaluation:

graph TD
    A[Data Acquisition: Sentiment140 Dataset] --> B[Data Cleaning & Mapping]
    B --> B1[Mapping Labels: 0-Neg, 4-Pos]
    B --> B2[Text Preprocessing: Regex, Stopwords]
    
    B1 --> C[Exploratory Data Analysis - EDA]
    C --> D[Data Splitting: Train/Val/Test]
    
    D --> E1[BERT Tokenization]
    D --> E2[Sequence Padding & Tokenization]
    
    E1 --> F1[Model 1: BERT Fine-Tuning]
    E2 --> F2[Model 2: RNN]
    E2 --> F3[Model 3: LSTM]
    E2 --> F4[Model 4: GRU]
    
    F1 --> G[Comparative Evaluation]
    F2 --> G
    F3 --> G
    F4 --> G
    
    G --> H[Final Analysis & Visualization]

Loading

(Workflow defined in Flow/sentiment_flow.mmd)

---

📂 Project Structure

Sentiment_Analysis_Using_BERT-LSTM-GRU-RNN/
├── Dataset/                   # Dataset folder (auto-created)
├── Flow/
│   └── sentiment_flow.mmd     # Mermaid workflow diagram
├── Sentiment_Analysis_Comparative_Study.ipynb  # Main research notebook
├── train_models.py           # Modular training script
├── requirements.txt           # Python dependencies
├── .gitignore                 # Files to ignore (Data, Models, etc.)
├── LICENSE                    # MIT License
└── README.md                  # Project documentation

---

🚀 Installation and Setup

1. Clone the repository:

   git clone https://github.com/SANJAI-s0/Sentiment_Analysis_Using_BERT-LSTM-GRU-RNN.git
   cd Sentiment_Analysis_Using_BERT-LSTM-GRU-RNN

2. Create a virtual environment:

   python -m venv venv
   source venv/bin/activate  # On Windows: venv\Scripts\activate

3. Install dependencies:

   pip install -r requirements.txt

4. Run the Analysis:

   • Open the .ipynb file in Jupyter or VS Code.
   • Or run the script: python train_models.py

---

🧠 Models Architectures

1. BERT

• Framework: HuggingFace Transformers (PyTorch).
• Pre-trained: bert-base-uncased.
• Fine-tuned with a linear classifier layer.

2. Recurrent Architectures (RNN, LSTM, GRU)

• Framework: TensorFlow/Keras.
• Hidden Layers: Bidirectional configurations.
• Dropout for regularization to avoid overfitting.

---

📈 Results & Comparison

Model	Accuracy	F1-Score	Training Time
BERT	~91%	TBD	High
LSTM	~84%	TBD	Medium
GRU	~83%	TBD	Medium
RNN	~76%	TBD	Low

---

📄 License

This project is licensed under the MIT License - see the LICENSE file for details.

---

📧 Contact

Sanjai - GitHub Profile

Project Link: https://github.com/SANJAI-s0/Sentiment_Analysis_Using_BERT-LSTM-GRU-RNN