openbmb/factnet-bench

FactNet Benchmarks

This repository contains three benchmark datasets derived from FactNet:

1. Knowledge Graph Completion (KGC)

The KGC benchmark evaluates a model's ability to complete missing links in a knowledge graph.

• Format: (subject, relation, object) triples
• Splits: Train/Dev/Test
• Task: Predict missing entity (either subject or object)
• Construction: Extracted from entity-valued synsets and projected to (S, P, O) triples with careful cross-split collision handling

2. Multilingual Knowledge Graph QA (MKQA)

The MKQA benchmark evaluates knowledge graph question answering across multiple languages.

• Languages: Multiple (en, zh, de, fr, etc.)
• Format: Natural language questions with structured answers
• Task: Answer factoid questions using knowledge graph information
• Construction: Generated from FactSynsets with canonical mentions across languages

3. Multilingual Fact Checking (MFC)

The MFC benchmark evaluates fact verification capabilities across languages.

• Languages: Multiple (en, zh, de, fr, etc.)
• Labels: SUPPORTED, REFUTED, NOT_ENOUGH_INFO
• Format: Claims with associated evidence units
• Construction:
  • SUPPORTED claims generated from synsets with FactSenses
  • REFUTED claims generated by value replacement
  • NOT_ENOUGH_INFO claims generated with no matching synsets
  • Each claim associated with gold evidence units with character spans

Usage

from datasets import load_dataset

# Load the KGC benchmark
kgc_dataset = load_dataset("factnet/kgc_bench")

# Load the MKQA benchmark for English
mkqa_en_dataset = load_dataset("factnet/mkqa_bench", "en")

# Load the MFC benchmark for English
mfc_en_dataset = load_dataset("factnet/mfc_bench", "en")

# Example of working with the MFC dataset
for item in mfc_en_dataset["test"]:
    claim = item["claim"]
    label = item["label"]
    evidence = item["evidence"]
    print(f"Claim: {claim}")
    print(f"Label: {label}")
    print(f"Evidence: {evidence}")

Construction Process

FactNet and its benchmarks were constructed through a multi-phase pipeline:

1. Data Extraction:

   • Parsing Wikidata to extract FactStatements and labels
   • Extracting Wikipedia pages using WikiExtractor
   • Parsing pagelinks and redirects from SQL dumps

2. Elasticsearch Indexing:

   • Indexing Wikipedia pages, FactStatements, and entity labels
   • Creating optimized indices for retrieval

3. FactNet Construction:

   • Building FactSense instances by linking statements to text
   • Aggregating FactStatements into FactSynsets
   • Building inter-synset relation edges

4. Benchmark Generation:

   • Constructing KGC, MKQA, and MFC benchmarks from the FactNet structure

Citation

If you use FactNet benchmarks in your research, please cite:

@article{shen2026factnet,
  title={FactNet: A Billion-Scale Knowledge Graph for Multilingual Factual Grounding},
  author={Shen, Yingli and Lai, Wen and Zhou, Jie and Zhang, Xueren and Wang, Yudong and Luo, Kangyang and Wang, Shuo and Gao, Ge and Fraser, Alexander and Sun, Maosong},
  journal={arXiv preprint arXiv:2602.03417},
  year={2026}
}

Acknowledgements

FactNet was built using Wikidata and Wikipedia data. We thank the communities behind these resources for their invaluable contributions to open knowledge.