Atomic Fact

An atomic fact expresses exactly one claim about a subject. It cannot be further decomposed into smaller factual statements without losing its complete meaning.

• Single Predicate Rule: Contains one subject and one predicate relationship
• Decomposition Test: If a sentence can be split by 'and' into two true/false statements, it is not atomic
• Example: 'Tesla was founded in 2003' is atomic; 'Tesla was founded in 2003 and is headquartered in Austin' is not

The fact must be fully understandable in isolation without requiring external sentences to resolve pronouns, temporal references, or ambiguous entities.

• Entity Resolution: All named entities are explicit and disambiguated (e.g., 'Elon Musk' not 'he')
• Temporal Grounding: Dates and times are absolute, not relative ('on July 15, 2024' not 'last Tuesday')
• No Anaphora: Pronouns and implicit references are replaced with their canonical referents

Every atomic fact must be objectively classifiable as true, false, or unverifiable against a trusted knowledge source. This property enables automated fact-checking at scale.

• Falsifiability Criterion: There must exist evidence that could prove the statement false
• Verification Target: Designed for direct lookup against a knowledge base like Wikipedia or Wikidata
• NLI Compatibility: Structured to serve as a hypothesis in a Natural Language Inference task against a premise passage

The fact is expressed in a canonical, normalized form to reduce surface-form variation that complicates matching. This ensures the same underlying claim is always represented identically.

• Canonicalization: Standardized date formats, unit conversions, and entity naming conventions
• De-duplication: Enables efficient clustering of semantically equivalent statements
• Embedding Stability: Normalized text produces more consistent vector representations for similarity search

Each atomic fact carries a direct pointer to its source passage, enabling fine-grained provenance tracking. This is the foundation of attribution fidelity scoring.

• Passage-Level Linking: The fact is tied to a specific sentence or paragraph, not an entire document
• Provenance Chain: Maintains a verifiable trail from the fact back to the original source
• Citation Integrity: Essential for FActScore evaluation, where each atomic fact is independently verified against a reference corpus

An atomic fact is anchored to a specific time point or interval to prevent staleness and enable temporal consistency checks across a dynamic knowledge base.

• Time-Bound Assertions: 'The population of Tokyo was 13.96 million in 2020' not 'Tokyo has 13.96 million people'
• Drift Detection: Explicit timestamps allow systems to flag facts that require re-verification
• Versioning Support: Enables knowledge base maintenance where outdated facts are deprecated rather than deleted

A fine-grained evaluation metric that decomposes a generated biography into atomic facts and verifies each one independently against a trusted knowledge source like Wikipedia. FActScore calculates a factual precision score by dividing the number of supported atomic facts by the total number of atomic facts in the output, providing a direct quantitative measure of hallucination severity at the most granular level.

A data structure consisting of a subject, predicate, and object that represents a single factual assertion about an entity. Each triple—such as <Albert_Einstein> <born_in> <Ulm>—is functionally equivalent to an atomic fact expressed in machine-readable form. Semantic triples form the foundational unit of knowledge graphs and enable automated reasoning, consistency checking, and contradiction detection across large-scale factual datasets.

A prompting technique where a language model first drafts a response, then generates a series of independent fact-checking questions to systematically verify and correct its own initial output. CoVe decomposes complex statements into atomic claims, executes verification queries against external or internal knowledge, and revises the final response based on detected inconsistencies. This self-correcting loop directly operationalizes atomic fact decomposition for hallucination reduction.

An automated evaluation process that measures the degree to which a generated summary or statement aligns with the facts presented in a source document, penalizing contradictions and hallucinations. Modern implementations decompose both the source and generated text into atomic fact units, then perform pairwise entailment checks using Natural Language Inference models. A low consistency score signals that generated atomic facts lack grounding in the provided evidence.

A task in natural language processing that determines whether a hypothesis sentence can be logically inferred (entailment), contradicted, or is neutral with respect to a given premise sentence. NLI serves as the computational engine for atomic fact verification: each extracted atomic fact becomes a hypothesis tested against a source premise. High-confidence entailment predictions provide the logical signal that an atomic fact is grounded.

A W3C standard for validating RDF graphs against a set of conditions, or 'shapes,' ensuring that knowledge graph data conforms to a defined ontology and is free of logical inconsistencies. SHACL constraints can encode rules like 'a person must have exactly one birth date' or 'a product price must be a positive number,' catching atomic-level factual contradictions before they propagate into AI-generated outputs grounded on that knowledge graph.