Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

FactNet: A Billion-Scale Knowledge Graph for Multilingual Factual Grounding

Yingli Shen, Wen Lai, Jie Zhou, Xueren Zhang, Yudong Wang, Kangyang Luo, Shuo Wang, Ge Gao, Alexander Fraser, Maosong Sun

TL;DR

FactNet addresses factual grounding and provenance for multilingual AI by constructing a billion-scale knowledge graph that deterministically aligns Wikidata statements with native Wikipedia evidence across 316 languages. Its three-layer data model (FactStatement, FactSense, FactSynset) and a deterministic canonicalization policy pi yield auditable, byte-level traceability, complemented by a rich set of relation signals. The authors introduce FactNet-Bench, a reproducible evaluation suite for knowledge graph completion, multilingual question answering, and closed-context fact checking, with strict leakage controls and stratified, snapshot-based splits. Empirical results demonstrate high grounding precision (0.921 design-weighted) and robust provenance integrity, while also revealing trade-offs in recall for long-tail languages and the impact of language biases. Together, FactNet and FactNet-Bench provide a publicly available, auditable foundation for training and evaluating trustworthy multilingual grounding systems.

Abstract

While LLMs exhibit remarkable fluency, their utility is often compromised by factual hallucinations and a lack of traceable provenance. Existing resources for grounding mitigate this but typically enforce a dichotomy: they offer either structured knowledge without textual context (e.g., knowledge bases) or grounded text with limited scale and linguistic coverage. To bridge this gap, we introduce FactNet, a massive, open-source resource designed to unify 1.7 billion atomic assertions with 3.01 billion auditable evidence pointers derived exclusively from 316 Wikipedia editions. Unlike recent synthetic approaches, FactNet employs a strictly deterministic construction pipeline, ensuring that every evidence unit is recoverable with byte-level precision. Extensive auditing confirms a high grounding precision of 92.1%, even in long-tail languages. Furthermore, we establish FactNet-Bench, a comprehensive evaluation suite for Knowledge Graph Completion, Question Answering, and Fact Checking. FactNet provides the community with a foundational, reproducible resource for training and evaluating trustworthy, verifiable multilingual systems.

FactNet: A Billion-Scale Knowledge Graph for Multilingual Factual Grounding

TL;DR

FactNet addresses factual grounding and provenance for multilingual AI by constructing a billion-scale knowledge graph that deterministically aligns Wikidata statements with native Wikipedia evidence across 316 languages. Its three-layer data model (FactStatement, FactSense, FactSynset) and a deterministic canonicalization policy pi yield auditable, byte-level traceability, complemented by a rich set of relation signals. The authors introduce FactNet-Bench, a reproducible evaluation suite for knowledge graph completion, multilingual question answering, and closed-context fact checking, with strict leakage controls and stratified, snapshot-based splits. Empirical results demonstrate high grounding precision (0.921 design-weighted) and robust provenance integrity, while also revealing trade-offs in recall for long-tail languages and the impact of language biases. Together, FactNet and FactNet-Bench provide a publicly available, auditable foundation for training and evaluating trustworthy multilingual grounding systems.

Abstract

While LLMs exhibit remarkable fluency, their utility is often compromised by factual hallucinations and a lack of traceable provenance. Existing resources for grounding mitigate this but typically enforce a dichotomy: they offer either structured knowledge without textual context (e.g., knowledge bases) or grounded text with limited scale and linguistic coverage. To bridge this gap, we introduce FactNet, a massive, open-source resource designed to unify 1.7 billion atomic assertions with 3.01 billion auditable evidence pointers derived exclusively from 316 Wikipedia editions. Unlike recent synthetic approaches, FactNet employs a strictly deterministic construction pipeline, ensuring that every evidence unit is recoverable with byte-level precision. Extensive auditing confirms a high grounding precision of 92.1%, even in long-tail languages. Furthermore, we establish FactNet-Bench, a comprehensive evaluation suite for Knowledge Graph Completion, Question Answering, and Fact Checking. FactNet provides the community with a foundational, reproducible resource for training and evaluating trustworthy, verifiable multilingual systems.
Paper Structure (116 sections, 10 equations, 5 figures, 22 tables)

This paper contains 116 sections, 10 equations, 5 figures, 22 tables.

Table of Contents

  1. Introduction
  2. FactNet: Design and Construction
  3. Data Model
  4. Auditable Canonicalization into FactSynsets
  5. Grounding: FactSense Extraction
  6. Relational Structure
  7. Reproducibility, Format, and Licensing
  8. Resource Statistics and Quality Assessment
  9. Scale, Definitions, and Long-Tail Structure
  10. Content Distribution and Representation
  11. Audit Protocol and Evaluation Estimands
  12. Grounding Quality of FactSenses
  13. Canonicalization and Structural Quality
  14. Reproducibility and Integrity
  15. FactNet-Bench: Tasks and Experiments
  16. ...and 101 more sections

Figures (5)

  • Figure 1: FactNet Architecture. The graph couples Wikidata claims with native evidence (from Wikipedia) via three layers: FactStatement (atomic unit), FactSense (grounded span with byte-offsets), and FactSynset (cross-lingual normalization). RelationEdges facilitate structural reasoning.
  • Figure 2: FactNet construction workflow. The pipeline processes dumps through three deterministic stages: (1) view extraction, (2) statement canonicalization, and (3) evidence matching. By avoiding stochastic models, we ensure every generated FactSynset and FactSense retains a stable, auditable trace back to the source snapshot.
  • Figure 3: Results on FactNet-Bench: (a) KGC under leakage control, (b) MKQA semantic parsing (18 langs), (c) MFC verification and evidence quality. Error bars show std. over 3 seeds.
  • Figure 4: Language-rank distribution diagnostics. This figure complements Table \ref{['tab:top_languages']} and justifies stratified sampling and tier-wise reporting.
  • Figure 5: Evidence-gap funnel visualization. We recommend plotting both macro- and micro-averaged funnels, and stacking the dominant loss reasons per tier. This figure operationalizes the "evidence gap" and helps users choose between improving extraction coverage versus restricting to the strong-evidence subset.