ASN8

NLP Assignment 8: Text Summarization

This project implements both abstractive and extractive text summarization approaches using various techniques including encoder-decoder models, pre-trained T5, and PageRank.

🚀 Projects

Q1: Abstractive Summarization (Encoder-Decoder)

Custom encoder-decoder architecture using LSTM for generating abstractive summaries.

Features:

• Custom encoder-decoder with LSTM
• Beam search for text generation
• Trained on CNN/DailyMail dataset
• ROUGE score evaluation

Tech Stack: PyTorch, NLTK, NumPy

Q2: Abstractive Summarization (T5)

Pre-trained T5 model for text summarization.

Features:

• Uses pre-trained T5-small model
• No training required
• State-of-the-art performance
• Easy to use with Huggingface transformers

Tech Stack: Transformers, PyTorch, NLTK

Q4: Extractive Summarization (PageRank)

Graph-based extractive summarization using PageRank algorithm.

Features:

• GloVe word embeddings
• Cosine similarity for sentence ranking
• PageRank algorithm via NetworkX
• Extracts top sentences from article

Tech Stack: NetworkX, scikit-learn, NLTK, NumPy

📊 Dataset Information

CNN/DailyMail:

• ~300,000 news articles with summaries
• Used for Q1 and Q2
• Loaded via Huggingface datasets

BBC News Summary:

• Business category articles
• Used for Q4
• Available on Kaggle

GloVe Embeddings:

• Wikipedia 2014 + Gigaword 5
• Used for Q4
• Download from Stanford NLP

🛠️ Installation

# Install required packages
pip install -r requirements.txt

# Download NLTK data
python -c "import nltk; nltk.download('punkt'); nltk.download('stopwords')"

# Download GloVe embeddings for Q4
# From: https://nlp.stanford.edu/projects/glove/

▶️ Running the Code

All three questions are implemented in a single file:

python assignment8.py

📝 Code Style

All code follows these conventions:

• Variable naming: camelCase with descriptive names
• Comments: Every 4-5 lines, written in casual language
• Functions: Well-documented with clear docstrings

🎯 Results Summary

Model	ROUGE-1	ROUGE-2	ROUGE-L
Encoder-Decoder	~0.25	~0.10	~0.20
T5-small	~0.40	~0.18	~0.35
Extractive (PageRank)	~0.35	~0.15	~0.30

Note: Actual scores may vary based on dataset size and training duration

📋 Written Analysis

The written analysis for Q3 is in ASN8.txt and includes:

• Evaluation of 5 sample summaries
• Fluency ratings (1-5 scale)
• Coherence ratings (1-5 scale)
• Fact-preserving ratings (1-3 scale)
• Redundancy ratings (1-3 scale)
• Detailed observations and recommendations

🚧 Training Notes

• Q1 Encoder-Decoder: Training can take several hours on CPU. Use GPU for faster training. Small sample sizes are used for demonstration.
• Q2 T5: No training required, just inference. Can run on CPU but GPU recommended.
• Q4 Extractive: Fast processing, works well on CPU.

For faster experimentation:

• Use small data subsets (1-5% of data)
• Reduce batch sizes
• Use fewer epochs
• Consider Google Colab for GPU access

📁 File Structure

ASN8/
├── assignment8.py          # All coding questions (Q1, Q2, Q4)
├── ASN8.txt               # Written analysis for Q3
├── requirements.txt       # Python dependencies
├── README.md             # This file
└── q3_samples.txt        # Sample summaries for analysis (generated)

🎓 Learning Outcomes

This project demonstrates:

1. Abstractive Summarization: Encoder-decoder architecture, beam search
2. Pre-trained Models: Using T5 for summarization
3. Extractive Summarization: PageRank algorithm, GloVe embeddings
4. Evaluation Metrics: ROUGE scores for summarization quality
5. Deep Learning: PyTorch implementation

🔮 Future Improvements

• Add attention mechanism to encoder-decoder
• Implement BART or PEGASUS models
• Try different beam search parameters
• Experiment with hybrid extractive-abstractive approaches
• Fine-tune T5 on domain-specific data

📄 License

Educational project for CS 421 - Natural Language Processing