Glossary
Claim
A claim is an attributable assertion that a subject has a specified property or relationship, usually with a time qualification.
Truth discovery
A claim is a statement that can be evaluated, such as an entity having an identifier or a trial having a status on a date. It is not the same thing as the value alone. Its subject, predicate, object, scope and time determine what would count as agreement or contradiction. Attribution identifies who made the assertion, while evidence identifies the material that can be inspected.
In practice, claims are extracted from documents or received as structured assertions. The original wording matters because a paraphrase can broaden a qualified statement: planned, registered and completed refer to different states. A claim should keep its source location, observation date and any uncertainty expressed in the source. These details let reviewers test the interpretation without trusting a summary.
Organizations that act on values need to know whether a displayed fact is directly stated, inferred or contested. A claim is the unit from which such a record can be built. Data conflicts arise between incompatible claims; provenance explains their origins; truth discovery evaluates competing claims. None of those concepts should erase the original assertion or its qualification.
Example
A Halden Therapeutics AG (fictional) notice states that a trial was planned as of 4 March; the claim does not say the trial had started.
The parts of a claim
A claim has a subject, a predicate and an object, so that it can be represented as a statement in the form used by RDF. Around that statement sit the qualifications that decide what it means: the time or period to which it applies, the scope, such as a jurisdiction or a population, any hedge in the source, such as planned, estimated or reportedly, and the unit or scale. Wikidata's data model is a working example of the structure: a statement has a main value, qualifiers that narrow it, references that support it, and a rank that records whether it is preferred or deprecated. The model lets contradictory statements on the same item coexist, each with its own references.
Distinguishing the claim from the fact matters because the same fact can be claimed many times and the same claim can be about more than one fact. When a record shows a single value with no claim behind it, it has already settled every question that the claim structure would have kept open.
Keeping the claim attached to its evidence
A claim needs two attachments to be useful for a decision. The first is attribution: which agent made the assertion, and when. The PROV data model expresses this as attribution of an entity to an agent and derivation from earlier entities. The second is the span of source text that expresses the claim, which the W3C Web Annotation model can address precisely using a quoted passage with its surrounding text; see evidence span. With both attachments a reviewer can check that the claim was read correctly and decide how much weight its author deserves.
The article on how to evaluate LLM citations at the claim level argues that the claim, not the document, is the right unit for checking whether a cited source supports a statement, because a document can contain a supporting sentence and a contradicting one. The same reasoning applies when a generated answer contains several claims behind a single citation.
A worked example
A press notice from Halden Therapeutics AG (fictional) states that it expects to begin enrolling patients in the third quarter. An extractor stores the claim as: Halden Therapeutics AG, enrolment start, third quarter. That loses the verb expects, and the loss changes the meaning from an announced intention to an event. The corrected claim keeps the modality, the quarter, the publication date of the notice and the sentence that expresses it. When a registry later shows enrolment opened in the fourth quarter, the two claims are compatible: one was a forecast, the other an observation, and neither is wrong. A record that stored only the value would present them as a conflict and might pick the wrong one.
Related terms and boundaries
A claim is the unit on which most neighbouring ideas operate. A data conflict is a relation between claims; truth discovery is a method for choosing among them; provenance and attribution record where each came from; an evidence span pins it to the words of the source; and information extraction is the process that usually produces it from text. Keeping the claim as the stored unit, with its qualifiers, is what allows each of these to be examined without reconstructing the source from memory.
Checks a reviewer can run
Open a stored claim and check for its subject, property, value, qualifiers, period, source and passage. Compare ten claims with their source sentences to see whether hedges, negations and scope survived. Ask whether the record can hold two incompatible claims about one subject. Ask who is attributed with a claim that was quoted from a third party.
Common misreadings
- Reducing a qualified sentence to an unqualified value.
- Confusing an extractor's interpretation with source wording.
- Omitting the date or subject of the assertion.
Appears in
- Reporting source conflicts
Preserve competing values and their grounds until a scoped decision can be defended.
Found an error on this page? See corrections for what counts as one, how to report it and the dated log.
Enquiry
Where does "Claim" bear on a decision of yours?
This entry sits under truth discovery. 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.