Factual Consistency

The foundational mechanism for measuring factual consistency is Natural Language Inference (NLI). A generated statement is factually consistent only if the source text entails it.

• Entailment: The source text provides sufficient evidence for the claim to be true.
• Contradiction: The claim directly opposes information in the source.
• Neutral: The claim introduces new information not discussed in the source, which is a form of extrinsic hallucination.

This strict logical grounding prevents models from adding plausible-sounding but unsupported details.

Factual consistency distinguishes between two primary failure modes:

• Intrinsic Hallucination: The output contradicts the provided source text. This is a direct violation of consistency and is easier to detect with NLI models.
• Extrinsic Hallucination: The output adds information that cannot be verified by the source text. This is a failure of faithfulness, as the model introduces ungrounded knowledge from its pre-training data.

A factually consistent summary must avoid both, remaining strictly faithful to the input context.

Modern evaluation frameworks like FActScore break generated text into a series of discrete, self-contained atomic facts. Each atomic fact is a single, verifiable claim.

• Process: A generator decomposes the text into a list of short assertions.
• Verification: Each assertion is individually checked against a knowledge source or the provided context.
• Scoring: The factual consistency score is the percentage of atomic facts that are fully supported.

This granular approach moves beyond surface-level n-gram overlap to assess deep semantic accuracy.

Factual consistency is distinct from, but often correlated with, coherence.

• Factual Consistency: Measures the truthfulness of content relative to a source. A summary can be perfectly consistent but poorly written.
• Coherence: Measures the logical flow and readability of the text itself. A summary can be highly coherent but entirely fabricated.

A high-quality output requires both properties. A coherent hallucination is often more dangerous than an incoherent one because it appears more convincing to a human reader.

While NLI-based metrics like ANLI and SummaC are standard for measuring factual consistency, they have known failure modes:

• Lexical Bias: Models may rely on word overlap rather than true logical reasoning.
• Domain Shift: An NLI model trained on generic text may fail on specialized domains like medicine or law.
• Granularity Mismatch: A single sentence may contain multiple claims, some true and some false, complicating a binary entailment decision.

Human evaluation remains the gold standard for final verification, with automated metrics serving as a scalable, high-recall filter.

Chain-of-Verification is a prompting strategy designed to improve factual consistency by making the model self-correct.

1. Draft: The model generates an initial response.
2. Question Generation: The model creates a list of fact-checking questions based on its draft.
3. Independent Verification: The model answers these questions independently, without seeing the original draft.
4. Correction: The final output is revised to align with the verified answers.

This process reduces hallucination by forcing the model to explicitly reason about and validate its own claims.