Glossary
Evidence span
An evidence span is the specific passage or region of a source used to support a claim.
Provenance and citation
An evidence span is the smallest useful portion of source material that supports an interpretation. It may be a sentence, table cell, paragraph or image region. The span needs enough context to preserve negation, headings, footnotes and qualifiers. Its locator should identify the source version as well as the position, because page numbers and offsets can shift when a document changes.
In practice, reviewers compare a proposed claim directly with its span. A headline may say a product is approved while the body limits approval to one jurisdiction; the supporting span must carry that condition. Extraction outputs should retain the original words rather than only a cleaned value. Where a claim depends on multiple passages, each relevant span should be linked.
For organizations defending values, a precise span shortens the path from conclusion to evidence and exposes overinterpretation. A retained source copy keeps the underlying material available; a sentence-level citation tells the reader which claim the span supports; provenance records how that claim moved into a final view. None of these steps makes a poor source authoritative.
Example
A Halden Therapeutics AG (fictional) trial status cites a table row and its footnote, because the footnote limits the row to one reporting period.
Addressing a span so it can be found again
An evidence span must be stored in a form that lets a reader find the same passage later. A character offset into the text is fragile: it is valid only for the exact version of the text it was computed against. The W3C Web Annotation Data Model defines selectors that are more robust. A text quote selector stores the exact quoted text together with a short prefix and suffix, so that the passage can be located by searching, even if offsets have shifted. A text position selector stores offsets and can be paired with the quote selector. A fragment selector uses an identifier or a page. The model also allows a source to be identified by a specific state, such as a particular retrieved version, which links the span to the retained copy it was taken from.
The standard datasets for claim verification ask for the same thing. FEVER requires a system to retrieve the sentences from Wikipedia that serve as evidence for or against each claim, and scores the evidence as well as the verdict. A verdict without the right evidence is counted as a failure, which expresses the point that the span is part of the answer.
Choosing the right size
A span that is too small drops the qualifiers that determine meaning: a clause such as enrolment completed loses the preceding word expected. A span that is too large buries the support in a page of text and invites the reader to stop checking. A workable rule is to include the whole sentence that expresses the claim, plus any adjacent sentence needed to resolve a pronoun, a date or a table header. For tables, include the header row and the row label. For figures and scanned pages, record the page, the region and a transcription, and keep the image.
Spans should be checked mechanically. Confirm that the stored quote occurs in the retained copy, once, at the recorded position; a quote that is not found signals that the copy and the span have drifted apart. Count the share of claims with at least one verified span, and treat claims without one as unsupported until shown otherwise. The article on how to evaluate LLM citations at the claim level uses spans in this way.
A worked example
A claim states that a manufacturing site of Halden Therapeutics AG (fictional) received an inspection finding in the previous year. The evidence span stored with the claim is a quote selector: the exact sentence from the regulator's letter, with the thirty characters before and after, the letter's date and the digest of the retained PDF. When a reviewer opens the claim, the viewer loads the retained copy, searches for the quote, and highlights it on page 2. If the regulator reissues the letter with different wording, the old copy still contains the original passage, and the new document is stored as a separate source with its own span. Offsets alone would have broken when the document was reformatted.
Related terms and boundaries
An evidence span is the evidential half of a sentence-level citation, and the span's context is why a claim keeps its qualifiers. It depends on a retained source copy for stability, is produced during information extraction, and is the unit chosen in chunking for retrieval. It shows what a source says, not whether the source is reliable.
Common misreadings
- Selecting a span too narrow to include its qualifier.
- Recording an offset without a source version.
- Substituting a summary for the original passage.
Appears in
- Document citation vs sentence citation
A document citation identifies an artifact; a sentence citation locates the specific statement supporting a claim.
- Golden record vs evidence-carrying record
A golden record presents preferred values; an evidence-carrying record keeps the dated source sentence with each value.
- Knowledge graph vs document store
A graph represents entities and relations explicitly; a document store preserves documents as units of content.
- LLM extraction vs rule-based extraction
Language models interpret varied prose; explicit rules suit stable, mechanically defined fields and formats.
- 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.
- Evaluating extraction against a gold set
Compare extracted claims with reviewed annotations, including spans, scope and abstentions.
References
- Buneman, Khanna and Tan (2001), Why and Where: A Characterization of Data Provenance, ICDT.
- W3C (2017), Web Annotation Data Model.
- Thorne et al. (2018), FEVER: a large-scale dataset for Fact Extraction and VERification, NAACL.
- Gao et al. (2023), Enabling Large Language Models to Generate Text with Citations, EMNLP.
- Rashkin et al. (2021), Measuring Attribution in Natural Language Generation Models, Computational Linguistics.
Found an error on this page? See corrections for what counts as one, how to report it and the dated log.
Enquiry
Where does "Evidence span" 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.