Skip to content

README.md

Latest commit

 

History

History
413 lines (315 loc) · 11.2 KB

README.md

File metadata and controls

413 lines (315 loc) · 11.2 KB

Idea2Paper - Knowledge Graph-Based Academic Paper Automatic Generation System

Project Overview: Automatically transform user's research Idea into paper Story that meets top conference (ICLR) standards

Core Technologies:

  • Knowledge Graph: 16,791 nodes, 444,872 edges
  • Vector Retrieval: Two-stage recall (Jaccard+Embedding), 13x speedup
  • Large Language Models: Qwen3-14B, Qwen2.5-7B-Instruct
  • Data Source: ICLR 2025 paper dataset (8,285 papers)

🚀 Quick Start

1. Environment Setup

# Enter project directory
cd /Users/gaoge/code/mycode/Idea2Paper/Paper-KG-Pipeline

# Install dependencies
pip install -r requirements.txt

# Set API key
export SILICONFLOW_API_KEY="your_api_key_here"

2. Build Knowledge Graph (Only run once)

# Build nodes (approx. 15 minutes)
python scripts/build_entity_v3.py

# Build edges (approx. 3 minutes)
python scripts/build_edges.py

Output:

output/
├── nodes_idea.json           # 8,284 Idea nodes
├── nodes_pattern.json        # 124 Pattern nodes
├── nodes_domain.json         # 98 Domain nodes
├── nodes_paper.json          # 8,285 Paper nodes
├── edges.json                # Edge data
└── knowledge_graph_v2.gpickle # NetworkX graph

3. Generate Paper Story

# Use custom Idea
python scripts/idea2story_pipeline.py "Your research Idea description"

# Example
python scripts/idea2story_pipeline.py "Using reinforcement learning to optimize large model inference efficiency"

Output:

output/
├── final_story.json          # Final generated paper Story
├── pipeline_result.json      # Complete process results
└── log.json                  # Detailed logs

4. View Results

# View final Story
cat output/final_story.json

# View complete process
cat output/pipeline_result.json | jq '.'

📚 Core Documentation

Essential Documents (Read in order)

No. Document Content Target Audience
0 Project Overview Overall architecture, core modules, parameter configuration, execution flow Everyone
1 Knowledge Graph Construction Data sources, nodes, edge definitions, LLM enhancement, execution methods Developers
2 Recall System Three-path recall strategy, similarity calculation, performance optimization Developers
3 Idea2Story Pipeline Pattern selection, Idea Fusion, Story Reflection, Critic review Developers

Documentation Features

✅ Complete Coverage: From data construction to generation full process
✅ Execution Guide: Every document includes detailed execution methods and parameter configuration
✅ Flow Diagrams: Clear architecture and process visualization using Mermaid charts
✅ Troubleshooting: Includes common issues and solutions

🎯 Core Features

1. Knowledge Graph (16,791 nodes)

Node Types:

  • Idea (8,284): Core innovation points of papers
  • Pattern (124): Writing patterns/method templates (124 LLM-enhanced)
  • Domain (98): Research domains
  • Paper (8,285): Specific papers

Edge Types:

  • Basic connection edges: Paper→Idea, Paper→Pattern, Paper→Domain
  • Recall assistance edges: Idea→Domain, Pattern→Domain (effectiveness scoring)

2. Three-Path Recall System (13x speedup)

Path Matching Target Captured Dimension Weight
Path 1 Idea Description Core concept similarity 0.4
Path 2 Domain & Sub-domains Domain generalization ability 0.2
Path 3 Paper Title Research topic similarity 0.4

Performance: Full Embedding ~7 minutes → Two-stage recall ~27 seconds (13x speedup)

3. Idea2Story Pipeline

Core Mechanisms:

  • ✅ Pattern Multi-dimensional Classification: Stability/Novelty/Cross-Domain
  • ✅ Idea Fusion: Organic fusion at conceptual level, not technical stacking
  • ✅ Story Reflection: Evaluate fusion quality, ensure conceptual unity
  • ✅ Critic Multi-role Review: Methodology/Novelty/Storyteller
  • ✅ Intelligent Refinement: Novelty mode + score degradation rollback + fallback strategy
  • ✅ RAG Deduplication: Avoid collision with existing work

🔧 System Architecture

graph TB
    A[User Input Idea] --> B[Three-path Recall Top-10 Patterns]
    B --> C[Pattern Multi-dimensional Classification]
    C --> D[Story Generation]
    D --> E[Critic Review]
    E --> F{Pass?}
    F -->|No| G[Idea Fusion]
    G --> H[Story Reflection]
    H --> I[Story Regeneration]
    I --> E
    F -->|Yes| J[RAG Deduplication]
    J --> K{Collision?}
    K -->|Yes| L[Pivot Avoidance]
    L --> D
    K -->|No| M[Output Final Story]

    style B fill:#9ff,stroke:#333,stroke-width:2px
    style G fill:#ff9,stroke:#333,stroke-width:2px
    style M fill:#9f9,stroke:#333,stroke-width:2px
Loading

📊 Key Metrics

Data Scale

Knowledge Graph:
  - Nodes: 16,791 (Idea 8,284 + Pattern 124 + Domain 98 + Paper 8,285)
  - Edges: 444,872
  - Idea Coverage: 100% (8,284/8,285)
  - Pattern LLM Enhancement: 124/124 (100%)

Performance Metrics

Recall Speed:
  - Full Embedding: ~7 minutes
  - Two-stage Recall: ~27 seconds (13x speedup)

Pipeline Execution Time:
  - Fastest: 3 minutes (first pass)
  - Typical: 5-7 minutes (2-3 refinement rounds)
  - Slowest: 10 minutes (novelty mode)

Quality Metrics

Critic Review:
  - Pass Standard: Average score >= 7.0/10
  - Dimensions: Methodology, Novelty, Storyteller
  - Novelty Mode Improvement: 0.5-1.5 points

Fusion Quality:
  - Threshold: >= 0.65
  - Typical Value: 0.68-0.75

💡 Core Innovations

Knowledge Graph Level

✅ LLM-enhanced Patterns: Generate inductive summaries for each Pattern cluster ✅ Dual-layer Description: Specific examples + global summary, both learnable and understandable ✅ Quality-oriented Edge Weights: Calculated based on paper quality and Pattern effectiveness

Recall Level

✅ Three-path Complementary Recall: Capture relevance from idea, domain, and paper dimensions ✅ Two-stage Optimization: Jaccard coarse ranking + Embedding fine ranking, 13x speedup ✅ Real-time Calculation Path 3: Avoid pre-building redundant edges, ensure complementarity

Generation Level

✅ Idea Fusion: Organic fusion at conceptual level ✅ Story Reflection: Reflect on fusion quality ✅ Novelty Priority Mode: Automatically upgrade when stagnating ✅ Intelligent Rollback: Avoid ineffective refinements ✅ Fallback Strategy: Ensure output quality

🛠️ Parameter Configuration

Recall System Configuration

# scripts/recall_system.py

class RecallConfig:
    # Path weights
    PATH1_WEIGHT = 0.4  # Similar Idea (important)
    PATH2_WEIGHT = 0.2  # Domain-related (auxiliary)
    PATH3_WEIGHT = 0.4  # Similar Paper (important)

    # Recall quantity
    FINAL_TOP_K = 10

    # Two-stage recall
    TWO_STAGE_RECALL = True
    COARSE_RECALL_SIZE = 100

Pipeline Configuration

# scripts/pipeline/config.py

class PipelineConfig:
    # Critic threshold
    PASS_SCORE = 7.0
    MAX_REFINE_ITERATIONS = 3

    # Novelty mode
    NOVELTY_MODE_MAX_PATTERNS = 10
    NOVELTY_SCORE_THRESHOLD = 6.0

    # Reflection
    FUSION_QUALITY_THRESHOLD = 0.65

    # Rollback
    SCORE_DEGRADATION_THRESHOLD = 0.1

    # RAG deduplication
    COLLISION_THRESHOLD = 0.75

🐛 Troubleshooting

API Key Issues

# Check environment variable
echo $SILICONFLOW_API_KEY

# Set environment variable
export SILICONFLOW_API_KEY="your_key_here"

Missing Data Files

# Rebuild knowledge graph
python scripts/build_entity_v3.py
python scripts/build_edges.py

Empty Recall Results

# Check graph files
ls -lh output/nodes_*.json
ls -lh output/knowledge_graph_v2.gpickle

More Issues

Refer to the "Troubleshooting" section in each core document.

📈 Performance Optimization Recommendations

Improve Novelty

# Increase novelty mode attempts
PipelineConfig.NOVELTY_MODE_MAX_PATTERNS = 15  # Default 10

# Increase Path 1 weight
RecallConfig.PATH1_WEIGHT = 0.5  # Default 0.4

Improve Stability

# Lower fusion quality threshold
PipelineConfig.FUSION_QUALITY_THRESHOLD = 0.60  # Default 0.65

# Increase Path 3 weight
RecallConfig.PATH3_WEIGHT = 0.5  # Default 0.4

Accelerate Recall

# Reduce recall quantity
RecallConfig.PATH1_TOP_K_IDEAS = 5   # Default 10
RecallConfig.PATH3_TOP_K_PAPERS = 10 # Default 20

🔍 Monitor Key Events

# Novelty mode activation
grep "Activated【Novelty Mode】" output/log.json

# Fusion quality scores
grep "Fusion quality score" output/log.json

# Rollback events
grep "【ROLLBACK TRIGGERED】" output/log.json

# Final pass
grep "🎉 Critic Review Passed" output/log.json

📦 Code Structure

Paper-KG-Pipeline/
├── data/ICLR_25/               # Data sources
├── output/                     # Output files
├── scripts/
│   ├── idea2story_pipeline.py  # Pipeline main entry (path unchanged)
│   ├── simple_recall_demo.py   # Compatibility wrapper (preferred: demos/)
│   ├── recall_system.py        # Recall system (shim)
│   ├── tools/                  # Build/ETL tools
│   │   ├── build_entity_v3.py  # Build nodes
│   │   ├── build_edges.py      # Build edges
│   │   └── ...
│   ├── demos/                  # Demo/experiment scripts
│   │   ├── simple_recall_demo.py
│   │   ├── run_pipeline.py
│   │   └── demo_pipeline.py
│   ├── legacy/                 # Archived legacy scripts
│   │   └── generate_patterns_old.py
│   └── pipeline/               # Pipeline modules (shim)
│       ├── config.py
│       ├── manager.py
│       ├── pattern_selector.py
│       ├── planner.py          # Idea Fusion
│       ├── story_generator.py
│       ├── story_reflector.py  # Story Reflection
│       ├── critic.py
│       ├── refinement.py
│       └── verifier.py
└── docs/                       # Core documentation (4 documents)
    ├── 00_PROJECT_OVERVIEW.md
    ├── 01_KG_CONSTRUCTION.md
    ├── 02_RECALL_SYSTEM.md
    └── 03_IDEA2STORY_PIPELINE.md

📝 Changelog

V3.1 (2026-01-25)

  • ✅ Integrated documentation system, retained 4 core documents
  • ✅ All documents include execution methods, parameter configuration, and flow diagrams
  • ✅ Historical documents archived to docs/archive/

V3.0 (2026-01-22)

  • ✅ Switched to ICLR 2025 data source
  • ✅ Achieved 100% Idea coverage
  • ✅ LLM-enhanced Pattern nodes
  • ✅ Two-stage recall optimization (13x speedup)

V2.0

  • ✅ Idea Fusion mechanism
  • ✅ Story Reflection mechanism
  • ✅ Novelty priority mode
  • ✅ Intelligent rollback mechanism

🎓 Acknowledgments

  • Data Source: ICLR 2025 paper dataset
  • API Service: SiliconFlow
  • Authors: Idea2Paper Team

Version: V3.1 Last Updated: 2026-01-25 Contact: Refer to core documentation for technical support