🧠 Graph-Based Text Summarization Framework

A modular and interpretable graph-based extractive summarization system that combines semantic embeddings (SBERT) with graph ranking to identify the most informative sentences.
It supports multiple embedders (SBERT, TF-IDF, BoW), integrates entity-level relationships, and includes built-in evaluation via ROUGE metrics.

---

🚀 Quick Start

# Clone the repository
git clone https://github.com/El-Bahnasawi/Graph-Based-Text-Summarization.git
cd Graph-Based-Text-Summarization

# Install dependencies
pip install -r requirements.txt

# Run summarization
python src/summarizer.py

# Run evaluation
python experiments/ablation_study.py

---

📂 Directory Overview

Folder / File	Description
src/	Core modules (text processing, embedding, graph construction, ranking)
experiments/	Experimental scripts (ablation studies, evaluation runs)
datasets/	Input datasets for summarization and evaluation
Results/	Generated metrics, plots, and logs
diagrams/	UML and sequence diagrams for pipeline visualization
streamlit_app.py	Interactive summarization interface
LICENSE	MIT License
README.md	Project documentation

---

🧩 System Architecture

The summarization pipeline follows a five-step process:

1️⃣ Text Processing

• Sentences are split and named entities extracted using TextProcessor.
• Each sentence and its entities are mapped for embedding.

2️⃣ Sentence Embedding

• EmbeddingService computes embeddings via SBERT, TF-IDF, or BoW.
• Similarity matrices are generated for graph construction.

3️⃣ Graph Construction

• Nodes represent sentences and entities.
• Edges capture sentence-sentence, sentence-entity, and entity-entity relationships.
• The GraphBuilder module handles connectivity.

4️⃣ Graph Ranking

• The Rank.py module implements PageRank with damping and power iteration.
• Sentences are ranked based on graph centrality.

5️⃣ Result Extraction

• Top-ranked sentences form the final summary.
• Results include top indices, sentences, and similarity graphs.

---

📈 Sequence Diagram — Summarization Flow

End-to-end summarization pipeline — from preprocessing to sentence extraction.

📉 Sequence Diagram — Evaluation Flow

Evaluation process showing interaction between evaluator, dataset, and ROUGE scorer.

🧱 Class Diagrams

Summarization class structure.

Evaluation and result storage class structure.

---

🧪 Experimental Results

1️⃣ Hyperparameter Tuning — Similarity Threshold vs ROUGE L2-Norm

To identify the optimal similarity threshold, we performed a systematic sweep (0.1–0.9) for each embedder: SBERT, TF-IDF, and BoW.

Embedder	Optimal Threshold	ROUGE L2-Norm	Trend
SBERT	0.3	0.384	Peaks early; gradual decline beyond 0.5
TF-IDF	0.1	0.353	Steep decay beyond 0.2
BoW	0.1	0.370	Moderate stability; lower ceiling

Observations

• SBERT leads in stability and performance, benefiting from semantic embeddings.
• TF-IDF and BoW degrade faster with higher thresholds, showing limited adaptability.
• The effective threshold range is 0.1–0.3, balancing noise filtering and coverage.

---

2️⃣ Ablation Study — Component Contribution Analysis

We measured the contribution of each graph component by removing specific edge types (sentence–sentence, sentence–name, name–name).

ROUGE Scores Across Configurations

Configuration	ROUGE-1	ROUGE-2	ROUGE-L	ROUGE L2-Norm	Avg Time (s)
Full Model (SBERT)	0.312	0.105	0.187	0.384	0.09
No Sentence–Sentence Edges	0.262	0.074	0.160	0.320	0.04
No Sentence–Name Edges	0.333	0.116	0.203	0.410	0.03
No Name–Name Edges	0.325	0.113	0.199	0.398	0.02
Only Sentence–Sentence	0.330	0.116	0.203	0.410	0.03
TF-IDF Baseline	0.286	0.094	0.177	0.353	0.02
BoW Baseline	0.301	0.100	0.182	0.370	0.02
Only Name Relationships	0.248	0.070	0.157	0.319	0.04

---

Component Importance (Δ ROUGE L2-Norm)

Ablated Component	Δ Performance	Significance
Sentence–Sentence Edges	−0.067	High
Sentence–Name Edges	+0.022	Low
Name–Name Edges	+0.014	Low

Interpretation

• Sentence–Sentence edges form the semantic backbone of the graph.
• Entity edges provide contextual enrichment with marginal quantitative gain.
• The “Only Sentence–Sentence” setup slightly outperforms the full model, implying redundancy when entity-based edges coexist with dense sentence connectivity.

---

Processing Time Trade-offs

Configuration	Avg Time (s)	Relative to Full Model
Full Model (SBERT)	0.09	1.00×
Ablated Variants	0.02–0.04	~0.4× speed-up

Lightweight graphs significantly reduce runtime with minimal loss in summarization quality — an encouraging sign for scalability.

---

3️⃣ Summary of Findings

Metric	Finding
Best Configuration	Only Sentence–Sentence (ROUGE L2-Norm ≈ 0.41)
Worst Configuration	No Sentence–Sentence (ROUGE L2-Norm ≈ 0.32)
Most Critical Component	Sentence–Sentence edges
Average Processing Time	0.042 s per article
Overall Success Rate	100% (1,000–6,000 samples per setting)

---

4️⃣ Practical Insights

• Semantic connectivity dominates — removing it collapses coherence and recall.
• Entity linking is expendable for small-scale summarization but valuable in domain-specific contexts (e.g., biomedical or news summarization).
• SBERT embeddings offer the best compromise between quality and interpretability.
• The graph-based architecture remains robust, with ablated variants maintaining competitive ROUGE scores.

---

🙏 Acknowledgments

• Collaboration: Amr Ashraf
• Supervision: Dr. Doaa Shawky
• Institution: Zewail City of Science and Technology
• Community: Open-source contributors and tool developers

---

🧩 Citation

If you use this work in your research or project, please cite it as:

@software{bahnasawi2025graphsum,
  author = {Mahmoud El-Bahnasawi and Amr Ashraf},
  title = {Graph-Based Text Summarization Framework},
  year = {2025},
  institution = {Zewail City of Science and Technology},
  license = {MIT}
}

---

📜 License

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