Reference Anchoring

The core mechanism of reference anchoring is linking a generated text segment to a specific character offset range in the source. This requires the model to output not just an answer but a structured object containing {answer: "...", evidence_span: [start_char, end_char], source_document_id: "..."}. This contrasts with document-level retrieval where the entire document is cited as a monolithic block. Span-level grounding enables automated verification pipelines to programmatically extract and compare the exact supporting text without human parsing.

Models capable of reference anchoring are often fine-tuned using contrastive learning objectives. The training data consists of triplets: (query, positive_span, negative_span). The model learns to maximize the similarity between the generated claim and the correct evidence span while minimizing similarity with distractor spans from the same or different documents. This teaches the model to discriminate between superficially relevant and genuinely supporting passages, reducing hallucinated citations.

A critical distinction in reference anchoring is between extractive grounding and abstractive grounding:

• Extractive: The answer text is a verbatim substring of the source span. Verification is trivial via string matching.
• Abstractive: The answer paraphrases or synthesizes the source span. Verification requires natural language inference (NLI) models to assess entailment. Anchoring systems must declare which mode they operate in, as abstractive grounding carries higher risk of attribution errors where the cited span does not fully entail the claim.

Complex claims often require support from multiple non-contiguous spans across one or more documents. Advanced anchoring systems output a citation graph rather than a single pointer. For example, a claim like 'Company X acquired Company Y for $Z billion in Q3 2024' might anchor the acquisition fact to a press release span, the financial figure to an SEC filing span, and the date to a news article span. Each anchor carries its own confidence score and attribution type (supporting, contradicting, background).

Reference anchoring is not a post-hoc feature but an integrated component of the generation pipeline. After generating a claim and its anchors, a verification module performs a closed-loop check: it retrieves the cited span, runs an NLI model to score entailment, and either accepts the claim-anchor pair or triggers regeneration. This self-correcting architecture is essential for high-stakes domains like legal document analysis and medical summarization where citation integrity is non-negotiable.

Reference anchoring requires standardized output schemas for interoperability. Common formats include:

• W3C Web Annotation Data Model: Uses target and body with precise TextPositionSelector or TextQuoteSelector objects.
• Provenance JSON-LD: Extends schema.org with hasEvidence properties linking claims to source spans.
• Custom span-pointer JSON: Lightweight format with {claim, source_url, text_span, start_offset, end_offset, confidence}. These schemas enable downstream tools to render clickable citations that navigate directly to the highlighted source text.