Attribution Span Annotation

The core principle is identifying the shortest contiguous text segment that provides sufficient evidence for a claim. This avoids citing entire paragraphs or documents, forcing the system to pinpoint the exact supporting phrase.

• Goal: Maximize precision, minimize extraneous text.
• Example: For the claim 'Revenue grew 15%', the annotation is the sentence 'Q3 revenue increased by 15% year-over-year,' not the entire earnings report.
• Contrast: Unlike document-level citation, span annotation provides a direct pointer to the evidence.

Creating training data requires a strict protocol for human annotators. They must map every factual clause in a model-generated summary back to a verifiable source span.

• Task: Annotators read the summary and source, then highlight the minimal supporting text.
• Edge Case: If a claim is unsupported, it is flagged as a hallucination rather than assigned a span.
• Inter-Annotator Agreement: High agreement on span boundaries is a critical quality metric, often measured by exact match or token-level F1.

Models are trained to jointly generate a summary and predict the corresponding source spans. This is often framed as an extractive-then-abstractive or end-to-end sequence tagging task.

• Architecture: A model may output summary tokens alongside pointer indices marking the start and end of evidence spans.
• Loss Function: Combines standard cross-entropy loss for text generation with a span prediction loss (e.g., for start/end token probabilities).
• Output: The final system produces a summary with inline, clickable citations linked to specific highlighted text in the source document.

The definition of a 'minimal span' involves a critical design trade-off between forensic precision and user comprehension.

• Token-Level: Citing a single word or number is maximally precise but can be jarring and lack context for a human reader.
• Sentence-Level: Citing a full sentence is more readable but may include redundant information.
• Best Practice: A sub-sentence clause is often the optimal balance, providing a complete, self-contained proposition that is both precise and human-readable.

Attribution span quality is evaluated using Attributable to Identified Sources (AIS) metrics. A claim is considered attributable only if a human judge agrees the annotated span fully supports it.

• AIS Precision: The percentage of generated claims that have a correctly attributed span.
• AIS Recall: The percentage of source facts that are correctly cited in the summary.
• Automated Proxies: Natural Language Inference (NLI) models can act as automated evaluators, checking if the annotated span entails the generated claim.

Span annotation is a direct mechanism for enforcing and verifying factual consistency. It creates an explicit, auditable link between output and input.

• Hallucination Prevention: By requiring a source span for every claim, the model is structurally discouraged from generating unsupported content.
• Debugging: When a factual error is found, the faulty span annotation immediately identifies the source of the grounding failure.
• Contrast with Faithfulness Metrics: While a faithfulness score quantifies consistency, span annotation provides the mechanistic proof and provenance.

The overarching mechanism of anchoring every factual claim in a generated response to a specific, verifiable location within a source document. While citation grounding provides the link between claim and source, attribution span annotation defines the exact minimal text segment that constitutes the evidence. This distinction is critical: grounding says which document, span annotation says exactly which words within that document support the claim. Fine-grained citation systems rely on span-level precision to enable user verification without requiring them to read entire source documents.

An automated metric that quantifies the degree to which a generated summary's factual assertions align with source documents. Attribution span annotation directly enables more rigorous consistency scoring by providing precise evidence boundaries. Key aspects include:

• Span-level NLI: Running natural language inference on the exact attributed span rather than the full document
• Precision/Recall of evidence: Measuring whether the annotated span fully supports the claim without including irrelevant text
• Atomic fact decomposition: Breaking claims into individual verifiable facts, each requiring its own attribution span

A verification process using Natural Language Inference (NLI) to determine if a generated statement is logically supported by the provided source text. Attribution span annotation narrows the entailment check to the most relevant text segment, dramatically reducing false positives where a claim appears supported by a document that contains the right entities but in a different context. The entailment decision is made on the minimal span, not the full document, increasing the precision of hallucination detection.

The systematic logging of every piece of information's origin and modification history throughout the synthesis pipeline. Attribution span annotation adds a critical layer to provenance by recording:

• The exact character offsets of supporting text
• The document version and chunk identifier
• The model's confidence in the span-to-claim alignment This granular audit trail is essential for enterprise compliance and allows downstream systems to trace any generated statement back to its precise evidentiary basis.

The task of determining whether a hypothesis is supported by a corpus of multiple documents, requiring evidence synthesis across sources. Attribution span annotation becomes significantly more complex in this setting, as a single claim may require multiple spans from different documents to be fully supported. This introduces challenges around:

• Span fusion: combining partial evidence from multiple sources
• Conflict resolution: identifying when spans from different documents contradict each other
• Redundancy elimination: ensuring multiple supporting spans aren't simply restating the same information

A quantitative evaluation measure designed specifically to assess whether a generated summary is factually consistent with and fully supported by the input source text. Attribution span annotation serves as the ground truth for training and evaluating faithfulness metrics. Human annotators mark precise spans, creating gold-standard datasets that teach models to distinguish between:

• Fully supported claims with tight, relevant spans
• Partially supported claims where the span covers some but not all facts
• Unsupported claims where no valid span exists in the source material