MAKGED is the first multi-agent framework for collaborative error detection in knowledge graphs, addressing two critical challenges in KG quality management:

1. Multi-Perspective Analysis: Overcomes single-view limitations through bidirectional subgraph analysis
2. Transparent Decision-Making: Implements explainable error detection via structured agent discussions

Key innovations:

• 🎯 4 Specialized Agents: Head/Tail × Forward/Backward agent architecture
• 🔍 Hybrid Embeddings: Combines GCN-based structural features with LLM semantic features
• 🤖 LLM-Powered Collaboration: Implements 3-round discussion protocol with tiebreaker mechanism
• 🏭 Industrial Proven: Validated with China Mobile's KGs

1. Bidirectional Subgraph Construction

   • Head_Forward/Backward Subgraphs
   • Tail_Forward/Backward Subgraphs

2. Hybrid Embedding Generator

   graph TD
     A[Subgraph Structure] --> B[3-Layer GCN]
     C[Triple Text] --> D[Llama-2 Embedding]
     B --> E[Concatenation Layer]
     D --> E
     E --> F[Unified Representation]
   

   Loading

3. Multi-Agent Discussion Protocol

   • Phase 1: Independent Analysis
   • Phase 2: 3-Round Discussion
   • Phase 3: Summarizer Voting (for ties)

If you find our paper and resource useful in your research, please consider giving a star ⭐ and citation 📝.

@article{li2025harnessing,
  title={Harnessing Diverse Perspectives: A Multi-Agent Framework for Enhanced Error Detection in Knowledge Graphs},
  author={Li, Yu and Huang, Yi and Qi, Guilin and Feng, Junlan and Hu, Nan and Zhai, Songlin and Xue, Haohan and Chen, Yongrui and Shen, Ruoyan and Wu, Tongtong},
  journal={arXiv preprint arXiv:2501.15791},
  year={2025}
}

For technical inquiries:
Yu Li - Southeast University

⬆ Back to Top