Sarcasm Detection (NLP Project)

This repository explores several ML/DL approaches for sarcasm detection on the Sarcasm Headlines Dataset (Onion/HuffPost headlines). We implemented and compared models ranging from classical baselines to transformer-based models.

Notebooks

• LogReg+SVM.ipynb
• LSTM.ipynb (my contribution)
• BiLSTM.ipynb
• BERT_baseline.ipynb
• BERT_improved.ipynb
• GUI_demo+words_importance.ipynb

My Contribution (LSTM)

I implemented and evaluated the LSTM sarcasm detector, including:

• Data preprocessing + vocabulary construction
• Baseline LSTM (random embeddings) vs improved LSTM (GloVe-100d init, fine-tuned)
• Training loop + validation + early stopping
• Metrics + confusion matrices
• Error analysis (false positives/negatives)
• Word-importance visualization (leave-one-out masking)

Dataset

• Sarcasm_Headlines_Dataset.json (included)

Report

• Final Report_Sarcasm Detection.pdf

Running

Open any notebook and run cells top-to-bottom.

Notes for LSTM.ipynb:

• Uses PyTorch, NumPy/Pandas, scikit-learn, matplotlib/seaborn
• Downloads GloVe 6B 100d on first run (large download) and reuses it on later runs
• If you want a clean setup, create a Python venv and install dependencies from requirements.txt

Install

Create a venv, then:

pip install -r requirements.txt

PyTorch install depends on your machine (CPU vs GPU). Install it from the official selector: https://pytorch.org/get-started/locally/