Fact Verification

The initial stage identifies check-worthy claims within unstructured text. This involves separating factual assertions from opinions, questions, or subjective statements. Systems use named entity recognition and syntactic dependency parsing to isolate the subject, predicate, and object of a claim. For example, 'The GDP grew by 3%' is extracted as a discrete, verifiable assertion with a specific numerical predicate. Advanced systems also resolve coreference to link pronouns back to their antecedents before extraction.

Once a claim is isolated, the system queries a trusted knowledge base or document corpus. This retrieval often uses dense passage retrieval to find semantically relevant evidence, not just keyword matches. The system may query structured sources like Wikidata via SPARQL or search unstructured text. The goal is to surface documents or triples that directly support or refute the claim's predicate. Document ranking algorithms prioritize authoritative sources to ground the verification in high-quality evidence.

This is the core reasoning engine. A Natural Language Inference model classifies the relationship between the claim and the retrieved evidence. The output is typically a three-way label:

• Entailment: The evidence confirms the claim.
• Contradiction: The evidence refutes the claim.
• Neutral: The evidence is insufficient to judge. Transformer-based models fine-tuned on datasets like FEVER or MultiNLI perform this logical calculus, assessing if the premise supports the hypothesis.

To communicate results to search engines, systems serialize findings using the Schema.org ClaimReview type. This structured data explicitly tags a claim with a textual review and a numeric rating. Key properties include:

• claimReviewed: The exact text of the claim.
• reviewRating: A structured rating like 'True', 'False', or 'Mostly True'.
• url: A link to the full fact-check article. This markup enables platforms like Google to display fact-check labels directly in search results.

A single claim may have multiple evidence sources with conflicting signals. The system must aggregate these into a final verdict with a calibrated confidence score. This involves weighting evidence by source authority and recency. A logistic regression layer or a soft voting ensemble combines NLI probabilities. The final output is not just a binary label but a probability distribution over verdicts, allowing downstream systems to filter results based on a required confidence threshold.

For enterprise trust, the system must show its work. Provenance tracking links every verdict back to the specific evidence sentences and source documents used. This creates an auditable lineage from claim to conclusion. Techniques like attention visualization highlight which tokens in the evidence most influenced the NLI decision. This transparency is critical for compliance and for allowing human analysts to override automated assessments when the reasoning chain is flawed.

A Schema.org structured data type that tags a specific claim with a fact-check review and rating. It enables search engines to display fact-check summaries directly in search results. Key properties include:

• claimReviewed: The exact text being fact-checked
• reviewRating: A numeric or textual rating (e.g., True, False, Mostly False)
• itemReviewed: The creative work containing the claim
• author: The organization performing the fact-check

A unique, persistent identifier assigned to an item in the Wikidata knowledge graph. Examples include Q42 for Douglas Adams and Q7186 for Marie Curie. These serve as canonical URIs for entity linking and semantic web applications. In fact verification workflows, Q-Nodes provide an unambiguous anchor point for:

• Corroborating claims against a structured knowledge base
• Resolving entity identity across disparate sources
• Establishing a single source of truth for entity attributes

A large-scale, automated knowledge base construction system developed by Google Research that fuses extracted facts from web text with existing structured data. Each probabilistic assertion receives a confidence score. This architecture underpins modern fact verification by:

• Aggregating claims from billions of web pages
• Cross-referencing extracted facts against existing knowledge graphs
• Assigning machine-learned confidence levels to each assertion
• Continuously updating as new evidence emerges

Metadata that tracks the origin, source, and transformation history of a specific fact or entity within a knowledge graph. Essential for establishing trust and auditability in automated fact verification. Key provenance attributes include:

• Source URI: The original document or dataset
• Extraction timestamp: When the fact was harvested
• Confidence score: The system's certainty in the extraction
• Attribution chain: All intermediate processing steps