BERT Sentiment Analysis

This project implements a fine-tuned BERT (Bidirectional Encoder Representations from Transformers) model to classify text into six distinct emotion categories: Sadness, Joy, Love, Anger, Fear, and Surprise. It provides an intuitive web interface powered by Flask, enabling users to input text and obtain the predicted emotion in real-time.

The model achieved an accuracy of 95% on the test set, with an average latency of ~5000 ms for inference, regardless of batch size, demonstrating consistent performance for both single and large batch predictions.

---

Table of Contents

1. Overview
2. Features
3. Project Architecture
4. Requirements
5. Installation and Setup
6. Usage
7. Model Training and Customization
8. Web Application
9. Logging and Debugging
10. Contributing
11. License

---

Overview

This project combines state-of-the-art NLP techniques with a fine-tuned BERT model to classify emotions in text. It offers:

• Real-time Predictions: Through a Flask-based web app.
• Custom Pipelines: For ingestion, transformation, and training.
• Pretrained BERT Backbone: Leveraging bert_base_en_uncased for contextualized text embeddings.

---

Features

• Sentiment Classification:
  • Classifies text into six emotion categories.
  • Uses a pre-trained BERT backbone for superior accuracy.
• Interactive Web Interface:
  • A Flask-powered interface for user interaction.
• Pipeline Structure:
  • Modular pipeline for easy customization and scalability.
• Error Handling and Logging:
  • Centralized logging and custom exception handling.

---

Project Architecture

The project is structured as follows:

BERTMODEL/
├── artifact/                 # Stores model artifacts (e.g., weights, metadata)
├── logs/                     # Logs generated during model training and runtime
├── notebooks/                # Jupyter notebooks for exploration and experimentation
├── src/                      # Core source code
│   ├── components/           # Pipeline components: data ingestion, transformation, model trainer
│   ├── pipeline/             # Train and predict pipelines
│   ├── exception.py          # Custom exception class for debugging
│   ├── logger.py             # Logging setup and utilities
│   ├── utils.py              # Helper functions
├── templates/                # HTML templates for Flask web app
│   ├── index.html            # Landing page template
│   ├── home.html             # Prediction results page
├── app.py                    # Flask application entry point
├── README.md                 # Project documentation
├── requirements.txt          # Python dependencies
├── setup.py                  # Setup script for project packaging

---

Requirements

The following dependencies are required to run the project:

• Python: Version >= 3.7
• TensorFlow: Version >= 2.10.1
• TensorFlow_text: Version >= 2.10
• Keras NLP: Version >= 0.6.0
• Additional libraries:
  • Pandas
  • NumPy
  • Scikit-learn

All dependencies are listed in the requirements.txt file with the working.

---

Installation and Setup

Follow these steps to set up the project:

Step 1: Clone the Repository

git clone https://github.com/mohdamaj/Sentiment-Analysis-using-BERT.git
cd BERT-Sentiment-Analysis

Step 2: Create and Activate a Virtual Environment

python -m venv venv

• On Unix or MacOS:

source venv/bin/activate

• On Windows:

venv\Scripts\activate

Step 3: Install Dependencies

pip install -r requirements.txt

Step 4: Train the Model

Refer to Model Training and Customization

Step 5: Run the Flask Application

python app.py

The app will run locally, and you can access it at http://127.0.0.1:5000.

---

Usage

Once the Flask app is running:

1. Open the web application in your browser.
2. Enter a text snippet in the input field.
3. Click the Submit button to see the predicted emotion.

---

Model Training and Customization

Model Architecture

The model uses the following components:

• BERT Backbone: Pre-trained bert_base_en_uncased for contextual embeddings.
• Preprocessing: Tokenization and input preparation using BertPreprocessor.
• Classification Head:
  • Dropout layer for regularization.
  • Dense layer with softmax activation for emotion classification.

Training the Model

The trigger for training takes place in data_ingestion.py.

Train the model by running:

python src/pipeline/data_ingestion.py

The trained model will be saved in the artifact/ directory.

---

Web Application

Features

• Input Form: Allows users to enter text for classification.
• Prediction Display: Shows the predicted emotion on the results page.
• Error Handling: Provides user-friendly error messages if input is missing or invalid.

Templates

• index.html: Landing page of the web app.
• home.html: Displays prediction results.

---

Logging and Debugging

• Logs are stored in the logs/ directory.
• Centralized logging is implemented using Python’s logging module.
• Errors are handled via the CustomException class in exception.py.

---

License

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