🤖 Sarcasm Detection using Deep Learning (Bi-LSTM)

A complete Natural Language Processing (NLP) project that detects sarcasm in text using a Bidirectional LSTM (Bi-LSTM) model. The system is built with a clean, modular architecture and evaluated on multiple datasets to ensure real-world generalization.

---

🚀 Project Overview

Sarcasm detection is a challenging NLP task because sarcastic sentences often appear positive on the surface but convey negative intent.

This project builds an end-to-end pipeline that:

• Processes raw text data
• Converts it into numerical representations
• Trains a deep learning model
• Evaluates performance using multiple metrics
• Tests generalization on unseen datasets

---

🧠 Problem Statement

Given a sentence:

"Oh great, another assignment"

The model predicts:

Sarcastic 😏

---

🏗️ Project Structure

sarcasm-detector/
│
├── data/
│   ├── dataset.json
│   └── new_dataset.json
│
├── model/
│   ├── sarcasm_model.keras
│   └── history.pkl
│
├── tokenizer/
│   └── tokenizer.pkl
│
├── plots/
│   ├── accuracy.png
│   └── loss.png
│
├── src/
│   ├── data_loader.py
│   ├── preprocess.py
│   ├── model.py
│   ├── train.py
│   ├── predict.py
│   ├── evaluate.py
│   ├── visualize.py
│   └── test_new_data.py
│
├── requirements.txt
└── README.md

---

⚙️ Installation

1. Clone the repository

git clone <your-repo-url>
cd sarcasm-detector

2. Create virtual environment

python3 -m venv venv
source venv/bin/activate

3. Install dependencies

pip install -r requirements.txt

---

📊 Dataset

This project uses the News Headlines Sarcasm Dataset from Kaggle.

• Format: JSON (line-by-line)

• Fields:

  • headline → text
  • is_sarcastic → label (0 or 1)

Download it from: [https://www.kaggle.com/search?q=News+Headlines+Sarcasm+Dataset]

Note: Dataset is subject to its own license.

---

🔄 Pipeline

1. Data Loading

• Reads JSON dataset
• Extracts text and labels

2. Preprocessing

• Tokenization (word → integer mapping)
• Padding (fixed sequence length)

3. Model Architecture

• Embedding Layer
• Bidirectional LSTM (64 units)
• Dropout (regularization)
• Bidirectional LSTM (32 units)
• Dense layers
• Sigmoid output (binary classification)

---

🧠 Model Summary

• Total Parameters: ~748K
• Type: Deep Learning (Bi-LSTM)
• Loss: Binary Crossentropy
• Optimizer: Adam

---

🏋️ Training

Run:

cd src
python train.py

Features:

• Early stopping (prevents overfitting)
• Model saving (.keras format)
• Tokenizer persistence
• Training history saved for visualization

---

🔮 Prediction

Run:

python predict.py

Example:

Enter a sentence: Oh great, another exam
Sarcastic 😏

---

📈 Evaluation

Run:

python evaluate.py

Metrics:

• Accuracy
• Precision
• Recall
• F1-score
• Confusion Matrix

---

📊 Visualization

Run:

python visualize.py

Generates:

• Accuracy vs Epoch graph
• Loss vs Epoch graph

Saved in:

plots/

---

🔍 Cross-Dataset Testing

Run:

python test_new_data.py

This evaluates the model on a completely different dataset to test generalization.

---

📊 Results

✅ Primary Dataset

• Accuracy: ~85–87%
• F1 Score: ~0.86

✅ Cross-Dataset Performance

• Accuracy: ~91%
• F1 Score: ~0.90

---

🧠 Key Insights

• The model generalizes well across datasets
• Slight difficulty in detecting subtle sarcasm
• Performance depends on dataset style and distribution

---

⚠️ Challenges

• Sarcasm is context-dependent
• Limited understanding of real-world tone
• Domain mismatch (news vs conversational text)

---

🔥 Future Improvements

• Replace LSTM with Transformer models (BERT)
• Add Streamlit UI for interactive predictions
• Deploy as a web application
• Fine-tune on conversational sarcasm datasets

---

🧑‍💻 Technologies Used

• Python
• TensorFlow / Keras
• NumPy / Pandas
• Scikit-learn
• Matplotlib

---

🏁 Conclusion

This project demonstrates:

• End-to-end NLP pipeline design
• Deep learning model implementation
• Proper evaluation and validation
• Real-world testing using multiple datasets

---

⭐ If you like this project

Give it a ⭐ on GitHub!

---