SEAGraph: Unveiling the Whole Story Behind Paper Review Comments

Because every review comment tells a deeper story...

---

🎯 What is SEAGraph?

Ever wondered what reviewers really mean when they write those cryptic comments on your paper? 🤔

SEAGraph is an intelligent framework that acts as your personal "review comment interpreter"! It doesn't just read review comments—it unveils the underlying intentions, context, and research landscape behind them.

The SEAGraph Framework: Connecting the dots between papers, reviews, and research context

🧠 How Does It Work?

SEAGraph constructs two powerful knowledge structures:

1. 📚 Semantic Mind Graph (SMG): Captures the author's thought process and the logical flow of the paper
2. 🌐 Hierarchical Background Graph (HBG): Maps out the research domains, related work, and academic context

By combining these graphs with intelligent retrieval, SEAGraph generates coherent, context-aware explanations that help you truly understand what reviewers are asking for!

See SEAGraph in action: Transforming vague comments into actionable insights

---

🚀 Quick Start

Prerequisites

SEAGraph uses two separate environments due to dependency conflicts. Don't worry—we've got you covered!

🔬 Environment 1: For Nougat (PDF Parsing)

cuda==12.4
python==3.9.20
torch==2.5.1
transformers==4.38.2
nougat-ocr==0.1.17
numpy==2.0.2

🤖 Environment 2: For Mistral & Sentence-BERT (Graph Construction & Retrieval)

cuda==12.4
python==3.9.20
torch==2.5.1
vllm==0.6.4.post1
transformers==4.46.3
sentence-transformers==3.3.1
torch_cluster==1.6.3
torch_scatter==2.1.2
torch_sparse==0.6.18
torch_spline_conv==1.2.2

⚠️ Pro Tip: The Nougat environment may conflict with the Mistral environment. Consider using separate conda/virtual environments!

---

📂 Project Structure

SEAGraph/
├── 📄 data/
│   ├── paper_pdf/      # 🎓 Place your academic papers here (PDF format)
│   └── raw_review/     # 📝 Place your review comments here (TXT format)
├── 💻 code/            # 🛠️ All the magic happens here
├── 📊 result/          # ✨ Your explanations will appear here (JSON format)
└── 🎨 asset/           # 🖼️ Figures and visualizations

---

🎬 The SEAGraph Pipeline

Transform your paper and review comments into insightful explanations in 10 easy steps:

Step	Script	Description
1️⃣	pdf_parse.py	📖 Parse your PDF into machine-readable format (MMD)
2️⃣	smg.py	🧠 Construct the Semantic Mind Graph
3️⃣	review_process.py	🔍 Extract and process review comments
4️⃣	hbg_related_paper_search.py	📚 Search for related papers based on citations
5️⃣	hbg_themes_infer.py	🎯 Infer research themes and topics
6️⃣	hbg_hot_paper_search.py	🔥 Find trending papers in your field
7️⃣	hbg_process_paper.py	⚙️ Process all background papers
8️⃣	retrieve_smg.py	🎣 Retrieve relevant content from SMG
9️⃣	retrieve_hbg.py	🎣 Retrieve relevant content from HBG
🔟	rag_seagraph.py	🎉 Generate comprehensive explanations!

---

💡 Usage

It's as simple as running a single command!

python <script_name>.py --filename <your_paper_id>

Example: Process a Paper End-to-End

# Step 1: Parse the PDF
python pdf_parse.py --filename 5t44vPlv9x

# Step 2: Build the Semantic Mind Graph
python smg.py --filename 5t44vPlv9x

# Step 3: Process review comments
python review_process.py --filename 5t44vPlv9x

# ... continue with remaining steps ...

# Final step: Generate explanations
python rag_seagraph.py --filename 5t44vPlv9x

💡 Tip: Replace 5t44vPlv9x with your paper's unique identifier!

---

📊 Performance Insights

SEAGraph performance across different metrics

---

🤝 Contributing

We welcome contributions! Whether it's:

• 🐛 Bug reports
• 💡 Feature suggestions
• 📖 Documentation improvements
• 🔧 Code contributions

Feel free to open an issue or submit a pull request!

---

📜 Citation

If you find SEAGraph helpful in your research, please consider citing our work:

@article{yu2024seagraph,
  title={SEAGraph: Unveiling the Whole Story of Paper Review Comments},
  author={Yu, Jianxiang and Tan, Jiaqi and Ding, Zichen and Zhu, Jiapeng and Li, Jiahao and Cheng, Yao and Cui, Qier and Lan, Yunshi and Li, Xiang},
  journal={arXiv preprint arXiv:2412.11939},
  year={2024}
}

---

📧 Contact

Have questions? Reach out to us!

• 📮 Email: sea.ecnu@gmail.com