Factual Consistency Metric

The systematic quantification of a model's propensity to generate content that is nonsensical or unfaithful to the provided source material. This assessment often uses factual consistency metrics as a primary signal, combining them with confidence calibration and semantic entropy to produce a composite risk score. Key techniques include:

• Analyzing output distributions for high uncertainty
• Cross-referencing generated entities against a knowledge graph
• Measuring the density of unsupported claims per token

A directional relationship between two text fragments where a premise (the source document) logically implies a hypothesis (the generated summary). This is the foundational NLP task underpinning many factual consistency metrics. A generated sentence is factually consistent if and only if it is entailed by the source. Modern approaches use decomposable attention models and pre-trained NLI classifiers fine-tuned on datasets like ANLI to score this relationship.

The broader machine learning task of determining the inferential relationship between a premise and a hypothesis, typically classified as entailment, contradiction, or neutral. In factual consistency evaluation, NLI models serve as automated judges. A summary sentence that contradicts the source document is flagged as a hallucination. Key benchmark datasets include:

• SNLI (Stanford Natural Language Inference)
• MultiNLI (Multi-Genre NLI)
• FEVER (Fact Extraction and VERification)

The algorithmic process of ordering retrieved documents or passages by their probative value to a specific claim before a final consistency judgment is made. Effective factual consistency metrics depend on high-quality evidence ranking to ensure the model is comparing the generated text against the most relevant source segments. Techniques include cross-encoder re-ranking and monoBERT architectures that compute fine-grained relevance scores between a claim and candidate evidence.

The natural language generation step that produces a human-readable explanation for why a generated statement was deemed factually consistent or inconsistent. This moves beyond a simple numeric metric to provide auditable reasoning. A justification typically cites the specific source passage, identifies the conflicting or supporting entities, and articulates the logical connection. This is critical for explainable fact-checking systems and debugging hallucination patterns.

The practice of evaluating factual consistency metrics against inputs specifically designed to fool the evaluator. A robust metric must not be brittle to paraphrastic manipulation or distractor sentences. Testing methodologies include:

• Semantically Equivalent Adversarial Rules (SEARs): Generating text with identical meaning but different syntax
• Adversarial NLI: Constructing hypothesis-premise pairs that confuse standard entailment models
• Stress-testing against known failure modes like negation and coreference swaps