Atomic Fact Decomposition

Each decomposed fact must be a standalone assertion that can be verified as true or false without requiring context from other sentences. This prevents dependency chains where one error cascades.

• Single Predicate Rule: One subject-predicate-object relationship per atom
• No Anaphora: Pronouns and implicit references are resolved to explicit named entities
• Example: 'The court ruled on March 3, 2024' is atomic; 'It was decided then' is not

Every atomic fact is individually checked against the source document using a Natural Language Inference (NLI) model. The result is a strict binary classification with no partial credit.

• Entailment: The source text logically supports the fact
• Contradiction: The source text directly refutes the fact
• Neutral: The source text does not mention or confirm the fact (treated as a hallucination)
• Hallucination Rate Formula: Contradictions + Neutrals / Total Atoms

The level of decomposition directly impacts evaluation precision. Over-decomposition creates noise; under-decomposition masks errors within complex statements.

• Coarse-Grained: Sentence-level decomposition risks accepting compound falsehoods
• Fine-Grained: Clause-level decomposition catches subtle misattributions
• Optimal Balance: Decompose to the smallest unit that still expresses a complete proposition
• Legal Example: 'The plaintiff, who filed on Tuesday, sought damages' becomes two atoms: (1) plaintiff filed Tuesday, (2) plaintiff sought damages

All entities referenced in atomic facts must be resolved to their canonical forms in the source text. This prevents evaluation failures caused by surface-form variation.

• Coreference Resolution: 'The company,' 'it,' and 'Acme Corp' are normalized to a single entity ID
• Date Normalization: 'Last Tuesday' becomes an absolute date based on document metadata
• Named Entity Linking: Parties, judges, and statutes are linked to a knowledge base for cross-document consistency

Each atomic fact retains a pointer to its precise origin in the generated summary and the source document. This creates a full audit trail for every claim.

• Summary Span: Character-level index of where the fact was expressed in the summary
• Source Span: Paragraph or line citation in the original document used for verification
• Confidence Score: The NLI model's probability score for the entailment classification
• Utility: Enables targeted human review of only the facts flagged as unsupported

Atomic decomposition evaluates faithfulness, not coverage. A summary can be perfectly faithful (all atoms entailed) yet incomplete (missing key source facts).

• Faithfulness: All stated facts are true (measured by atomic decomposition)
• Completeness: All important source facts are included (measured by recall-oriented metrics)
• Trade-off: Increasing completeness often introduces hallucination risk
• Combined Metric: F1 score balancing ROUGE recall against atomic fact precision

The degree to which a generated summary accurately reflects the stated facts of the source document without contradiction or fabrication. In the context of Atomic Fact Decomposition, consistency is measured by verifying each atomic claim against the source.

• Binary Verification: Each atom is either supported or not
• Contradiction Detection: Identifies atoms that directly oppose source statements
• Hallucination Flagging: Catches fabricated details not present in the original

A task where a model determines if a hypothesis is entailed by, contradicts, or is neutral to a given premise. NLI serves as the computational engine behind atomic fact verification.

• Entailment: The source text logically supports the atomic claim
• Contradiction: The source text directly refutes the atomic claim
• Neutral: The source text provides insufficient information to confirm or deny

Used in benchmarks like SummaC and ANLI to automate faithfulness scoring.

The technique of explicitly linking each factual statement in a generated summary back to its precise location in the source document. Atomic Fact Decomposition enables granular attribution by isolating individual claims.

• Span-Level Grounding: Maps atoms to specific sentence ranges
• Multi-Source Fusion: Tracks atoms synthesized from multiple passages
• Unattributed Detection: Flags atoms lacking any source support for human review

Critical for legal applications where citation integrity is non-negotiable.

A metric quantifying the frequency at which a language model generates factually incorrect or unverifiable information. Atomic Fact Decomposition provides a fine-grained measurement by counting unsupported atoms.

• Atom-Level Precision: Measures hallucination at the claim level, not sentence level
• Extrinsic vs. Intrinsic: Distinguishes source-contradicting from world-knowledge fabrications
• Benchmarking: Used in FActScore and similar evaluation frameworks

A summary may be fluent but have a high hallucination rate if many atoms fail verification.

The task of identifying and linking semantically related passages that discuss the same event or fact across a collection of distinct documents. When decomposing a multi-document summary, atoms must be verified against potentially multiple sources.

• Redundancy Detection: Identifies atoms repeated across sources
• Conflict Resolution: Surfaces atoms where sources disagree on facts
• Temporal Alignment: Ensures atoms reflect the correct chronological version of events

Essential for Multi-Document Fusion tasks where summaries synthesize information from case law collections.

The NLP task of identifying all linguistic expressions that refer to the same real-world entity. This preprocessing step is critical before atomic fact decomposition to ensure atoms are self-contained.

• Entity Normalization: Replaces pronouns with named entities in atoms
• Cross-Sentence Linking: Resolves references spanning sentence boundaries
• Legal Specificity: Handles complex party designations common in contracts and rulings

Without resolution, an atom like "He breached the contract" becomes unverifiable without knowing who "He" refers to.