startthat.work home

TC-04 Claim and citation verification against source text (fact verification; claim checking; citation support check; quotation accuracy check)

Labels each claim as supported, refuted or not enough information against retrieved evidence, or rates whether a cited source supports its claim, and produces a label with evidence sentences.

Trace cause and test claimsEvidence: TestedVerification technique, runs in step 6Strongest statement: describe

What it does

Labels a claim as supported, refuted or not enough information against retrieved evidence, or checks that a cited source supports the claim it is attached to. Artefact: a label with evidence sentences, or a supported, minor-error or major-error rating per citation. Read the evidence sentences, and read repeated claims for a shared origin.

When to use it

Use when you need to check whether claims in a text, or the sources cited for them, are actually backed by the source text.

When it does not fit

Fits when the corpus or cited source contains the relevant sentence. Misfit: evidence spread across documents or needing inference, paywalled sources, and opinions or predictions extracted as claims. The minor-error class is subjective (kappa of 0.26 in one study).

Result format

The result is a table with one row per claim or citation, giving the claim, the source, the label (supported, refuted or not enough information, or a supported, minor-error or major-error rating) and the evidence sentences behind it. It names the repeated claims that share an origin.

Requirements on the material

PropertyRequirement typeRequirement
Content typeAlways appliesrequires=text
Unit of analysisAlways appliesrequires=claim
Number of sourcesAlways appliesrequires=two or more
Source dependenceChanges the orderprefers=independent
Appraises source credibilityApplies in some casesno

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

Minimum materialClaims in separable form and the cited source or an evidence corpus that contains the relevant passage.
Additional sources needednone
Not for (rules)none
Not for (text)Wrong choice when evidence is spread across documents or needs inference, the source is paywalled, or the claims extracted are opinions or predictions. The minor-error class is subjective (kappa of 0.26 in one study).

Choosing and ordering

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

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: judgement dependent.

StakesCheckPerson involvedCounts as a different method
exploratorysource-span citationnoneno
decision-informingsource-span citationnoneno
decision-informinghuman review sampleneeded; left out below high stakesno
high-stakessource-span citationnoneno
high-stakeshuman review sampleneededno
high-stakesgold standard agreementconditionalno
high-stakesindependent reviewneededno

Failure modes and limits

Model suitabilityhuman judgement required
Typical model failureLabels a claim as supported because a similar sentence exists, and counts a claim repeated across sources as independently confirmed.
Error sourcesinput data; language model; source selection; coder or rater judgement
Certainty limitsCertainty drops when evidence is spread across documents, needs inference or is behind a paywall, when labels are judged subjectively, and when repeated claims share one origin (inconsistency and indirectness).

Evidence and tag basis

Tested. A 2018 baseline pipeline reached 31.87 per cent accuracy even with correct evidence supplied, and annotators agreed at Fleiss kappa 0.6841 (not a statement about current models). A meta-analysis of 28 medical studies gave a pooled citation error rate of 25.4 per cent, with 11.9 per cent major errors. (Thorne et al., FEVER, arXiv 1803.05355; Jergas and Baethge, PeerJ 2015)

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