Skip to content

Latest commit

 

History

History
108 lines (91 loc) · 4.03 KB

README.md

File metadata and controls

108 lines (91 loc) · 4.03 KB
Raw

Sarcasm Detection using Deep Learning

This project combines traditional machine learning and advanced deep learning techniques to accurately detect sarcasm in text. By leveraging both TF-IDF-based Random Forest models and the power of BERT, it ensures robust performance across various datasets.

Features

1. Advanced Text Preprocessing

  • Custom TextPreprocessor Class: Handles text cleaning, lemmatization, and tokenization.
  • Stopword Removal: Removes common words that don’t contribute to meaning.
  • Special Character Handling: Filters out unnecessary symbols and punctuations.

2. Dual Model Architecture

  • Traditional ML Model: Uses TF-IDF features and a Random Forest classifier for sarcasm detection.
  • Deep Learning Model: Implements a fine-tuned BERT model for contextual understanding.
  • Ensemble Prediction: Combines predictions from both models for improved accuracy.

3. BERT Integration

  • Pre-Trained Model: Uses BERT for advanced feature extraction.
  • Custom Dataset Class: Prepares input for BERT with attention masks and tokenization.
  • Context Understanding: Captures nuanced meanings for accurate sarcasm detection.

4. Better Feature Engineering

  • TF-IDF Representation: Replaces basic CountVectorizer for better feature weighting.
  • N-Gram Support: Includes uni-grams, bi-grams, and tri-grams for richer features.
  • BERT Word Embeddings: Utilizes BERT's contextual embeddings for deep learning.

5. Improved Training Process

  • Train/Validation Split: Ensures reliable performance evaluation.
  • Cross-Validation: Supports robust model tuning.
  • Learning Rate Optimization: Implements dynamic learning rates for faster convergence.
  • Dropout Regularization: Prevents overfitting in deep learning models.

6. Additional Functionality

  • Device Support: Compatible with both CPU and GPU for training and inference.
  • Comprehensive Evaluation Metrics: Includes accuracy, precision, recall, F1-score, and more.
  • Flexible Predictions: Choose between traditional ML, BERT, or ensemble methods.

Requirements

  • Python 3.12+
  • Jupyter Notebook
  • Libraries:
    • TensorFlow / PyTorch
    • NLTK
    • Transformers
    • Scikit-Learn
    • re
    • Matplotlib
    • Collections
    • Numpy
    • Pandas

Installation

  1. Clone the repository

    git clone https://github.com/ajitashwathr10/Sarcasm-Detection.git
cd Sarcasm-Detection

  2. Set up a virtual environment (Optional)

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

  3. Install Dependencies

    pip install -r requirements.txt

  4. Download Pre-Trained BERT Model

    • This project uses the Hugging Face Transformers library. Ensure you have internet access when running the model to automatically download the pre-trained BERT weights.

  5. Verify Installation

    python test_installation.py

Additional Notes

  • For GPU support, ensure you have CUDA installed and install the GPU-compatible version of TensorFlow or PyTorch 
    pip install torch torchvision torchaudio --index-url https://download.pytorch.org/whl/cu117
    pip install tensorflow-gpu

Now you're ready to start detecting sarcasm!

Usage

  1. Preprocess Data
    • Prepare your dataset using the TextPreprocessor class to clean and tokenize text.
  2. Train Models
    • Train the traditional and deep learning models using the provided training pipelines.
  3. Make Predictions
    • Use the ensemble or individual models to predict sarcasm in new textual data.
  • Example: 
    from sarcasm_detector import SarcasmDetector
detector = SarcasmDetector()
detector.train("path/to/dataset.csv")
predictions = detector.predict(["This is such a great day!", "Oh, what a surprise..."])
print(predictions)

Contributing

We welcome contributions to enhance this project. Feel free to submit pull requests or raise issues!

License

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