Factual Consistency

Factual consistency is the property of a generated summary containing only statements that can be directly supported by the source document. It ensures the output does not contradict, embellish, or hallucinate facts not present in the original text. This is distinct from fluency or coherence; a summary can be perfectly well-written yet factually inconsistent. The core principle is strict textual entailment: every atomic fact in the summary must be entailed by the source.

Evaluating factual consistency relies on several automated and human-driven approaches:

• Natural Language Inference (NLI): Treats each summary sentence as a hypothesis and the source as a premise, classifying pairs as entailed, neutral, or contradictory.
• Question Answering (QA): Generates questions from the summary and attempts to answer them from the source; a mismatch indicates inconsistency.
• FactCC: A BERT-based model fine-tuned specifically to detect factual errors in generated summaries.
• Human Evaluation: Expert annotators manually label factual errors at the atomic fact level, often using frameworks like the Fabbri et al. typology.

Factual errors in summaries fall into distinct categories:

• Intrinsic Hallucination: The summary fabricates information not present in the source.
• Extrinsic Hallucination: The summary introduces real-world knowledge not contained in the source document.
• Contradiction: The summary states the opposite of what the source asserts.
• Misattribution: A correct fact is linked to the wrong entity or context.
• Overgeneralization: A specific claim is broadened into an unsupported universal statement.

Factual consistency is the direct inverse of hallucination in summarization. While hallucination is a broad term for any generated content ungrounded from reality, factual consistency specifically measures grounding against a provided source document. A summary can be factually consistent with its source yet still contain factual errors if the source itself is wrong. This distinction is critical: factual consistency evaluates faithfulness to the input, not absolute truthfulness.

Several strategies improve factual consistency in generation:

• Source Grounding: Constraining the decoder to attend exclusively to source tokens.
• Contrastive Decoding: Amplifying the probability of tokens that an expert model prefers over an amateur model, reducing hallucination.
• Post-hoc Correction: Using a separate verification model to detect and rewrite inconsistent spans.
• Chain-of-Density Prompting: Iteratively refining summaries to increase information density without introducing unsupported facts.
• Retrieval-Augmented Generation (RAG): Explicitly retrieving source passages before generation to anchor the output.

Key benchmarks for measuring factual consistency include:

• XSum-Factuality: A dataset of CNN/DailyMail and XSum summaries annotated for factual errors.
• SummaC: A benchmark providing both NLI-based and QA-based consistency metrics with a standardized evaluation protocol.
• FRANK: A human-annotated dataset with fine-grained error typology across multiple summarization models.
• TRUE: A comprehensive collection of human-annotated factual consistency evaluations unifying several prior datasets.

A measurement of the uncertainty or randomness in a model's token predictions that correlates with the generation of non-factual or unsupported content. High entropy states often precede hallucinations, making this metric a valuable internal risk indicator for real-time monitoring.

• Serves as an early warning system for factual drift
• Calculated from the probability distribution over the model's vocabulary
• Can trigger fallback mechanisms when thresholds are exceeded

The technique of anchoring a language model's generated output to specific, provided source documents to minimize factual errors. Unlike general knowledge generation, grounded outputs must be verifiable against the input data, creating a direct chain of accountability.

• Requires explicit document retrieval before generation
• Forms the backbone of RAG architectures
• Enables citation and provenance tracking

The accuracy with which a generated summary correctly links claims and facts back to their precise origin points within the source material. High attribution fidelity ensures proper provenance and allows human reviewers to verify every assertion against its original context.

• Measures citation precision at the sentence level
• Critical for legal, medical, and financial applications
• Evaluated using entailment models and span overlap metrics

A generation technique that improves text quality by searching for tokens that maximize the probability difference between a strong expert model and a weaker amateur model. This approach amplifies desirable behaviors like factual accuracy while suppressing common failure modes.

• Requires no external knowledge base
• Amplifies knowledge from larger models
• Reduces hallucination without fine-tuning

A decoding strategy that contrasts the logit outputs from a later, mature transformer layer against an earlier, premature layer to surface factual knowledge. By exploiting the fact that factual knowledge localizes to specific layers, DoLa reduces hallucinations without requiring an external model or retraining.

• Operates within a single model's internal layers
• No additional compute overhead for retrieval
• Effective for long-form generation tasks

A methodology that uses Natural Language Inference models to automatically evaluate whether a generated summary is entailed by, contradicts, or is neutral to the source document. This provides a scalable, automated proxy for human factuality judgments.

• Uses models fine-tuned on MNLI, ANLI, and similar benchmarks
• Produces entailment, contradiction, and neutral scores per sentence
• Enables continuous monitoring in production pipelines