Comparisons
Document citation vs sentence citation
A document citation identifies an artifact; a sentence citation locates the specific statement supporting a claim.
The distinction
A document citation names the source artifact, often with a title, date, and stable locator. It is sufficient when the claim concerns the document as a whole, such as its existence or publication status, and it lets a reader assess authorship and broader context. It becomes less useful when a long document contains several conflicting statements, versions, or qualifications. A reader may spend substantial time searching for the exact passage that supposedly supports a narrow extracted value.
A sentence citation locates the particular source sentence used to support a claim, ideally while preserving the surrounding document and its version. It narrows the verification task and makes it easier to see whether an asserted value is quoted, inferred, or contradicted nearby. Sentence boundaries are not always the right semantic unit: a table cell, paragraph, or sentence pair may be needed. The principle is a precise evidence span, with enough context retained to avoid misleading extraction.
By dimension
| Dimension | Document citation | Sentence citation |
|---|---|---|
| Locator | The artifact or a broad section. | The specific supporting statement or span. |
| Review effort | The reader may need to search the artifact. | The supporting text is directly inspectable. |
| Context | The whole document is readily available. | Nearby context must be retained deliberately. |
| Claim granularity | Good for claims about the artifact as a whole. | Good for a specific factual assertion. |
| Versioning | A document version should be identified. | Both the span and its document version should be identified. |
| Ambiguity | Multiple passages may appear to support different readings. | The selected passage can still require interpretation. |
| Maintenance | A stable artifact link can suffice. | Span locators must survive reformatting or be tied to a retained copy. |
Choose Document citation when
- The claim concerns an entire publication or filing.
- Readers need the full source for broad analysis.
- A finer locator is unavailable but the source is still useful.
Choose Sentence citation when
- A single value or assertion needs verification.
- The source is long or contains competing statements.
- Review time and reproducibility matter.
Using both
Cite the document and its version, then identify the precise sentence or other evidence span within it and preserve enough context for interpretation.
A worked example
A fictional report states that a supplier "received an inspection finding in the last year" and cites a 90-page inspection report. A reader opens the report and finds three findings, two of them closed, one dated two years earlier, and a separate section about a different site. The citation is accurate about the document and unhelpful about the claim.
With a sentence citation, the report cites the page and the sentence that states the finding, quotes it, and records the date it was read and the version of the report. The reader sees at once that the finding is real, its date, and its site. The reader also sees what the sentence does not say, such as whether it has been closed, and can look for that in the surrounding text, which is kept with the citation.
Where each one fails
Document citation fails on long or compound sources and on anything where the claim is narrow. It also fails when the document changes in place, because the title and the URL stay the same. It tolerates sloppy summarising, since nobody can check quickly.
Sentence citation fails when the sentence is cut from its context. A quoted sentence can omit a qualification that follows it, or sit in a section about something else. It also fails when the evidence is a table, a figure or an inference across paragraphs. Benchmarks for generated text with citations show that checking whether a passage supports a claim is itself a research problem, and that even strong models often fail to give complete citation support. The citation should therefore identify the smallest span that supports the claim and keep enough neighbouring text to interpret it.
Questions to settle before choosing
Is the claim about the document as a whole, or about something in it? How long is the source? Does it contain competing statements? Is the source stable, or does it revise in place? Who will check the citation, and how much time do they have? For anything a regulator or counterparty may challenge, narrow claims need narrow citations. For general background, a document-level pointer is honest and cheaper. Whatever is chosen, say which kind of citation it is.
How the two meet in a review
The strongest form is layered: the document with its version and a retained copy, then the locator inside it, then the quoted span. The W3C web annotation model has a standard way to select a passage by quoting it with a little text on either side, which survives some changes to a page that a page position would not. A reviewer can then answer three questions in order: which document, which passage, and does the passage say it.
See also
- Golden record vs evidence-carrying record
- Citing at the sentence
- Retaining source copies
- Auditing an LLM answer against its sources
- Evidence span
- Attribution
Further reading in the articles
References
- W3C (2013). PROV-O: The PROV Ontology.
- Gao et al. (2023). Enabling Large Language Models to Generate Text with Citations. arXiv:2305.14627.
- Rashkin et al. (2021). Measuring Attribution in Natural Language Generation Models. arXiv:2112.12870.
- Bohnet et al. (2022). Attributed Question Answering: Evaluation and Modeling for Attributed Large Language Models. arXiv:2212.08037.
- W3C (2017). Web Annotation Data Model: text quote selectors.
Found an error on this page? See corrections for what counts as one, how to report it and the dated log.
Enquiry
Which do you run today: Document citation or Sentence citation?
Describe how your organization handles this now and one entity where it falls short. We reply with where a first record would sit alongside it and what it would need from you.
Or write to [email protected] in your own words.