startthat.work home

CK-17 Source-span and quotation check by code (quote-and-verify; extractive grounding; citation verification)

Require each extracted statement to point to the exact passage it came from, confirm by string match that the passage exists, and produce a pass or fail list per statement.

Check extracted resultsEvidence: DocumentedVerification technique, runs in step 6Strongest statement: describe

What it does

Requires each extracted statement to point to the exact passage it came from, then confirms by string match that the passage exists. A reviewer or entailment check then confirms the passage supports the claim. Artefact: a pass or fail list per statement.

When to use it

Use when a model has produced statements with quotations or citations and you need to know whether the quoted passages really exist in the source.

When it does not fit

Needs the source as text. Misfit: the quoted span is a paraphrase, or exists but does not support the claim.

Result format

The result is a pass or fail list with one row per extracted statement, giving the statement, the exact passage it points to, whether that passage exists by string match and whether the passage supports the claim. It ends with the list of statements that failed.

Requirements on the material

PropertyRequirement typeRequirement
Content typeAlways appliesrequires=text
Unit of analysisAlways appliesrequires=claim
Appraises source credibilityApplies in some casesno

Not restricted on: Structure, Variable types, Time structure, How the sample was chosen, Level of detail available, Number of sources, Whether sources can be joined, Amount of material, Known data problems, Producer interest, Source dependence, Origin, Document genre, Language, Personal information, Security classification, Licence.

Minimum materialthe source in machine-readable text
Additional sources needednone
Not for (rules)none
Not for (text)Wrong choice when the source is available only as images or scans. A quoted span that exists may still not support the claim, so a reviewer or entailment check must follow.

Choosing and ordering

Kind of techniqueVerification technique, runs in step 6
Phase tagvalidate
Other phase tagsreport
Kind of findingcontradiction; gap
Question answeredwhat
Strongest statement it supportsdescribe
Needs as input
Makes as output
Must exist firstNone
Loop targetreturn to verify and prepare

Library checks by stakes

Each stakes level includes the checks of the level below it. Checks that need a person apply at high stakes only, so the process leaves them out at lower stakes (decision 38). Repeatability: repeatable with fixed settings.

StakesCheckPerson involvedCounts as a different method
exploratoryinput validationnoneno
decision-informinginput validationnoneno
decision-informingsource-span citationnoneno
decision-informingspot checkoptionalno
high-stakesinput validationnoneno
high-stakessource-span citationnoneno
high-stakesspot checkoptionalno
high-stakeshuman review sampleneededno
high-stakesindependent reviewneededno

Failure modes and limits

Model suitabilitymodel plus code
Typical model failureFabricates quotations that look verbatim, or attaches real quotes to claims they do not support.
Error sourcesinput data; data transformation; language model
Certainty limitsCertainty drops when passages are paraphrases and not exact matches (indirectness) and when matched passages are not read for support (risk of bias).

Evidence and tag basis

Documented. One approach extracts both the supporting source span and the inconsistent span, and human evaluation found the span extraction helpful (abstract). NIST requires review and verification of sources and citations in outputs. (Kryscinski et al., arXiv 1910.12840; NIST AI 600-1)

Tag basis: input profile inference; selection inference; sequencing inference; robustness inference. All tag values are inference unless a source is named.