README.md

metadata

license: mit

Wikidata5M-KG

Wikidata5M-KG is an open-domain knowledge graph constructed from Wikipedia and Wikidata. It contains approximately 4.6 million entities and 21 million triples. Wikidata5M-KG is built based on the Wikidata5M dataset.

📦 Contents

wikidata5m_kg.tar.gz

This is the processed knowledge graph used in our experiments. It contains:

• 4,665,331 entities
• 810 relations
• 20,987,217 triples

After extraction, it yields a single file: wikidata5m_kg.jsonl, each line is a JSON object representing an entity with its metadata and one-hop neighborhood. Example:

{
  "entity_id": "Q10417481",                          // Unique entity ID
  "entity_description": "Arethaea polingi, or Poling's thread-legged katydid, is a species of phaneropterine katydid in the family Tettigoniidae. It is found in North America.",       // Description of an entity
  "entity_alias": ["Arethaea polingi", "arethaea polingi"],  // Aliases of an entity
  "all_one_hop_triples_str": [                       // One-hop outgoing triples (relation, tail entity)
    ["parent taxon", "Arethaea"],
    ["instance of", "Taxxon"],
    ["taxon rank", "cohesion species"]
  ]
}

wikidata5m_raw_data.tar.gz

This archive contains the original raw files used to construct wikidata5m_kg.jsonl.
The data is derived from the Wikidata5M dataset. For detailed descriptions of the files and their meanings, please refer to the original project documentation.

🔗 References

• 📘 Paper: KG-Infused RAG
• 💻 Code: GitHub Repository

📄 Citation

If you find this knowledge graph useful, please cite:

@article{wu2025kg,
  title={KG-Infused RAG: Augmenting Corpus-Based RAG with External Knowledge Graphs},
  author={Wu, Dingjun and Yan, Yukun and Liu, Zhenghao and Liu, Zhiyuan and Sun, Maosong},
  journal={arXiv preprint arXiv:2506.09542},
  year={2025}
}