Angry Message Detector

A simple neural network classifier that detects angry/urgent messages vs neutral ones using PyTorch and TF-IDF features.

Features

• Binary text classification (angry vs neutral)
• TF-IDF vectorization with uni/bi-grams
• Small feedforward neural network
• Training with backpropagation
• Easy to adapt to your own data

Installation

# Clone the repository
git clone https://github.com/yourusername/angry-message-detector.git
cd angry-message-detector

# Install dependencies
pip install -r requirements.txt

Quick Start

Using the example data

python angry_message_detector.py

This will train the model on the included example data and show predictions on test messages.

Using your own data

1. Prepare a CSV file with columns text and label:

   • text: your message content
   • label: 1 for angry/urgent, 0 for neutral

2. Place your CSV in the data/ directory (e.g., data/my_messages.csv)

3. Run the training script:

python train_with_csv.py --data data/my_messages.csv --epochs 50

Project Structure

angry-message-detector/
├── README.md
├── requirements.txt
├── angry_message_detector.py    # Main script with example data
├── train_with_csv.py            # Script for training with CSV data
├── data/
│   └── example_messages.csv     # Example training data
└── models/
    └── .gitkeep

Example Data Format

text,label
"This is unacceptable, I've asked three times already!",1
"Hello, could you please check my order status?",0
"I am extremely disappointed with your service.",1
"Thanks for your help, appreciated!",0

How It Works

1. Text Vectorization: Converts messages to TF-IDF numerical features
2. Neural Network: 2-layer feedforward network (input → 64 hidden units → output)
3. Training: Uses binary cross-entropy loss and backpropagation
4. Prediction: Outputs probability that a message is angry (threshold: 0.5)

Model Architecture

Input (TF-IDF features) → Linear(hidden=64) → ReLU → Dropout(0.1) → Linear(1) → Sigmoid

Customization

Edit hyperparameters in the script:

# In angry_message_detector.py or train_with_csv.py
EPOCHS = 25           # Number of training passes
HIDDEN = 64           # Hidden layer size
LEARNING_RATE = 1e-3  # Optimizer learning rate
BATCH_SIZE = 4        # Mini-batch size
MAX_FEATURES = 300    # TF-IDF vocabulary size

Performance

On the small example dataset (10 messages), the model quickly learns to distinguish angry from neutral messages. For real-world use, you'll need:

• At least 100-1000 labeled examples
• Balanced classes (similar number of angry/neutral)
• Larger max_features (e.g., 5000-20000)

Saving and Loading Models

See train_with_csv.py for examples of:

• Saving trained models to models/
• Loading models for inference
• Saving the TF-IDF vectorizer

License

MIT

Contributing

Pull requests are welcome! For major changes, please open an issue first.

Acknowledgments

Built with PyTorch and scikit-learn.