Fact Verification

The initial stage where the system queries a trusted corpus—such as Wikipedia, scientific abstracts, or a proprietary knowledge graph—to find documents that support or refute the input claim. Unlike open-domain QA, this step prioritizes high recall to ensure no contradictory evidence is missed. Techniques often combine sparse retrieval (BM25) for exact term matching with dense retrieval (DPR) for semantic similarity to capture paraphrased facts.

The core reasoning engine that classifies the logical relationship between the retrieved evidence and the claim. This is typically framed as a three-way classification task:

• Entailment: The evidence confirms the claim.
• Contradiction: The evidence refutes the claim.
• Neutral: The evidence is insufficient or unrelated. Modern systems fine-tune large transformer models like RoBERTa on specialized NLI datasets (e.g., SNLI, MultiNLI) to detect subtle logical inconsistencies.

The ability to combine multiple pieces of evidence across different documents to verify a complex claim. For example, verifying 'The author of 'The Hobbit' was born in South Africa' requires joining the fact that J.R.R. Tolkien wrote The Hobbit with the fact that he was born in Bloemfontein. This often uses graph neural networks or iterative retrieval loops where an initial query spawns sub-queries to fill logical gaps.

A preprocessing step that rewrites context-dependent claims into standalone, unambiguous statements. A claim like 'He is the current president' is unverifiable without resolving the pronoun 'He' and the temporal anchor 'current'. This process uses coreference resolution and entity linking to ground the claim to specific Wikipedia IDs and timestamps, transforming it into 'Joe Biden is the president of the United States in 2024'.

The standard evaluation protocol derived from the Fact Extraction and VERification (FEVER) shared task. The metric strictly requires the system to provide the correct label (Supports/Refutes/Not Enough Info) AND the correct set of evidence sentences. A prediction is scored as correct only if both the verdict and the evidence set match the ground truth exactly, penalizing lucky guesses and ensuring genuine textual grounding.

The robustness of a verification model against semantically equivalent adversarial examples. A fact-checking system must not flip its verdict when a claim is paraphrased syntactically. Testing involves claim mutation—such as negating verbs, swapping entities, or introducing distracting clauses—to ensure the model relies on the underlying semantic logic rather than shallow lexical cues or dataset biases.

The initial stage of fact verification that sources supporting or refuting documents from a corpus. Dense Passage Retrieval (DPR) encodes claims and documents into a shared vector space for semantic matching, while sparse retrieval using BM25 anchors on lexical overlap. Modern systems often employ a hybrid search fusion approach, combining both signals via reciprocal rank fusion to maximize recall before downstream verification.

The core reasoning engine that classifies the logical relationship between a claim and retrieved evidence. An NLI model assigns labels of entailment, contradiction, or neutral. Architectures like cross-encoder re-rankers process the claim-evidence pair jointly through full self-attention, enabling deep semantic comparison that a simple cosine similarity of embeddings cannot capture.

A preprocessing step that breaks complex, multi-hop claims into atomic sub-claims. A claim like 'Tesla, founded in 2003, acquired SolarCity in 2016' is split into two verifiable units. This relies on semantic role labeling and dependency parsing to identify predicate-argument structures. Decomposition enables precise evidence retrieval for each atomic fact and prevents a single false clause from invalidating an entire compound statement.

The process of anchoring claims to structured, deterministic facts in an Enterprise Knowledge Graph. Unlike free-text evidence, a knowledge graph provides a canonical source of truth via semantic triples (subject-predicate-object). Fact verification against a graph involves link prediction and triple classification to check if a claimed relationship exists, offering high-precision verification for domain-specific assertions.

The final stage that generates a human-readable rationale for the veracity label. This goes beyond a binary true/false output by highlighting the specific evidence spans that informed the decision. Techniques leverage extractive summarization over the retrieved documents or use a generative model conditioned on the claim, evidence, and predicted NLI label to produce a coherent explanation, critical for algorithmic explainability and auditability.