Fact Verification

The initial stage where a natural language processing system identifies and isolates discrete, check-worthy factual assertions from unstructured text. This process often involves breaking complex, multi-clause sentences into atomic claims that can be individually verified.

• Named Entity Recognition (NER) identifies key entities like people, dates, and locations
• Semantic Role Labeling (SRL) determines the predicate-argument structure of each assertion
• Output is a structured set of subject-predicate-object triples ready for evidence retrieval

The system queries a trusted document corpus—often a vector database or inverted index—to find passages semantically relevant to the extracted claim. This stage prioritizes recall to ensure no exculpatory evidence is missed.

• Dense retrieval using bi-encoder models maps claims and documents into a shared embedding space
• Sparse retrieval via BM25 provides exact keyword matching as a complementary signal
• Re-ranking models then score candidate passages for relevance before passing them to the veracity prediction stage

A classification model, often a fine-tuned transformer architecture, evaluates the relationship between the claim and the retrieved evidence. The core task is Recognizing Textual Entailment (RTE).

• Entailment: The evidence logically supports the claim
• Contradiction: The evidence refutes the claim
• Neutral: The evidence is insufficient to determine truth
• The model outputs a citation confidence score reflecting the probability of entailment

For complex claims that cannot be verified by a single document, the system performs multi-hop reasoning. It chains together multiple pieces of evidence across a citation graph or knowledge base to construct a logical proof.

• Iteratively retrieves new evidence based on intermediate conclusions
• Uses graph neural networks to traverse relationships in a provenance graph
• Essential for verifying claims requiring aggregation, comparison, or temporal reasoning across sources

Not all evidence is equal. Verification systems assign a source authority score to each document based on its historical reliability, domain expertise, and citation patterns. This score modulates the influence of evidence on the final verdict.

• Factors include PageRank-like algorithms on the citation graph
• Domain-specific trust lists (e.g., peer-reviewed journals vs. self-published blogs)
• Temporal freshness to penalize outdated information
• Prevents adversarial manipulation by low-credibility sources

The final stage produces a human-readable explanation for the verdict, grounding each determination in specific evidence passages. This is the foundation of attribution protocols and reference anchoring.

• Generates a concise rationale summarizing the supporting or refuting evidence
• Links each claim to precise text spans in source documents via reference anchoring
• Enables end-user citation transparency and independent auditability of the system's reasoning