Glossary
Sentence-level citation
A sentence-level citation ties a particular sentence or claim to the exact source material offered as support.
Provenance and citation
A sentence-level citation makes the support for a statement inspectable at the level where the statement is made. It identifies source material relevant to that sentence, rather than attaching one general bibliography to a page of mixed claims. Its purpose is traceability, not decoration. The cited passage must actually entail the statement, including its qualifiers, subject and time.
In practice, a citation may point to a document section, page, paragraph or retained excerpt. A sentence containing several independent assertions may need several citations or be split. A citation to a source that merely mentions the subject is insufficient. Reviewers should be able to compare the cited wording with the sentence and detect when a paraphrase overstates the evidence.
For a defensible record, this granularity makes errors local and correctable. It lets a reader challenge one assertion without rejecting an entire report. An evidence span is the precise portion of source content; attribution names the source or creator; a sentence-level citation connects the rendered assertion to that supporting material.
Example
A Halden Therapeutics AG (fictional) summary cites the exact notice sentence for a planned study date, while a separate sentence about status has separate support.
Why the sentence and not the document
A document-level citation tells a reader that a source exists somewhere in the vicinity of the claim. A sentence-level citation tells a reader which statement is supported and by which passage. The distinction decides whether a cited answer can be checked in seconds or only by rereading the source. Research on generated text has made the same distinction: Rashkin and colleagues propose attribution to identified sources as a measurable property of generated statements, Gao and colleagues evaluate citation quality in language-model answers, and Liu and colleagues audited generative search engines and found that many statements were not fully supported by the sources cited for them.
Systems that quote a supporting passage are easier to verify. DeepMind's GopherCite trained a model to support its answers with verbatim quotes, and Bohnet and colleagues formalised attributed question answering as a task with its own evaluation. The common thread is that a citation is a testable claim: this passage supports this statement.
Testing a citation
A citation can fail in four ways. The source may not contain the statement at all. It may contain a related but different statement, for example a forecast where the answer states an outcome. It may contain the statement but contradict it elsewhere. Or it may support only part of a compound sentence. A check should therefore read the passage and decide whether it entails the statement, partly supports it, is unrelated, or contradicts it. Automated judges can help at scale, and their reliability should be calibrated against human labels, as described in the article on an LLM judge groundedness calibration protocol.
Two requirements are easy to overlook. The cited passage must be retrievable in the form the author saw, which is why a retained source copy matters, since pages change. And a multi-step answer needs citations for every step, a requirement tested in the article on multi-hop RAG citation completeness. The article on evaluating LLM citations at the claim level combines these checks.
A worked example
An answer states: Halden Therapeutics AG (fictional) completed enrolment in its Phase 2 study in May. It cites a registry page. The page reads: enrolment is expected to complete in May. The citation points to a real page on the right company and the right study, and a document-level check would pass it. A sentence-level check finds that the passage supports an expectation and not a completion, so the statement is unsupported as written. The correction is either to reword the answer to state the expectation or to find a later source that reports completion. The reviewer's note records the verdict, the passage and the reason. The same pattern, a right source cited for a stronger statement, is the commonest failure found in audits of cited answers.
Related terms and boundaries
A sentence-level citation is the pairing of a claim with an evidence span. It is one form of grounding and an input to measuring faithfulness. It relies on attribution to name the source and on provenance to explain how the source was obtained.
Common misreadings
- Citing a relevant document that does not support the sentence.
- Using one citation for several unrelated assertions.
- Dropping a source's uncertainty in paraphrase.
Related terms
Appears in
- Document citation vs sentence citation
A document citation identifies an artifact; a sentence citation locates the specific statement supporting a claim.
- Auditing an LLM answer against its sources
Split an answer into claims and inspect cited passages for support, inference and missing qualifications.
- Citing at the sentence
Bind each assertion to the exact source passage, its locator, reading date and retained copy.
- Detecting source drift
Compare current source content with the retained cited version and reassess changed support.
References
- Buneman, Khanna and Tan (2001), Why and Where: A Characterization of Data Provenance, ICDT.
- Rashkin et al. (2021), Measuring Attribution in Natural Language Generation Models, Computational Linguistics.
- Gao et al. (2023), Enabling Large Language Models to Generate Text with Citations, EMNLP.
- Liu, Zhang and Liang (2023), Evaluating Verifiability in Generative Search Engines.
- Menick et al. (2022), Teaching language models to support answers with verified quotes.
- Bohnet et al. (2022), Attributed Question Answering: Evaluation and Modeling for Attributed Large Language Models.
Found an error on this page? See corrections for what counts as one, how to report it and the dated log.
Enquiry
Where does "Sentence-level citation" bear on a decision of yours?
This entry sits under provenance and citation. If a decision in your organization turns on it, send the entity concerned, the sources you fuse to describe it and who may read which of them. We reply with what a first record would cover.
Or write to [email protected] in your own words.