Where does relevance collide?

Evidence is relevant if it has any tendency to make a fact of consequence more or less probable than it would be without it. Federal Rule of Evidence 401 fixed that wording in 1975, Rule 402 makes relevant evidence admissible unless something else excludes it, and Rule 403 lets a judge exclude relevant evidence whose probative value is substantially outweighed by unfair prejudice or waste of time. James argued in 1941 that relevancy is a matter of logic and experience rather than a rule of law, and the drafters took that view: the threshold is low, a brick need not be a wall. Relevance here is a relation between an item of proof and a legally material proposition, decided by a judge before the fact-finder ever sees it.[1][2]

Unit

An item of evidence offered in a proceeding

Recovery means

Not applicable, the question is admission, not restoration

Measured by

A judge's ruling on the 401 threshold and the 403 balance

An input is relevant to a person to the extent that it yields positive cognitive effects (new conclusions, strengthened or revised beliefs) at low processing effort. Grice listed "be relevant" in 1975 as one maxim among several and left it undefined. Sperber and Wilson made it the whole theory in 1986: human cognition is geared to maximise relevance, every utterance carries a presumption of its own relevance, and a hearer stops interpreting at the first reading that satisfies that presumption. Relevance here is a property of a stimulus for a mind at a moment, graded by a ratio of effect to effort, and it is claimed to explain comprehension rather than to evaluate it.[3][4]

Unit

An utterance or stimulus for an individual at a time

Recovery means

Not applicable

Measured by

Not measured directly; inferred from interpretations hearers arrive at and from reading-time and eye-movement studies

The relation between a retrieved object and a user's need, judged either by the system, as topical match between query and document, or by the user, as usefulness for the task at hand. Cleverdon's Cranfield tests fixed the evaluation method in the 1960s: a fixed collection, fixed queries, and relevance judgments made in advance by assessors, from which precision and recall are computed. Saracevic reviewed the notion in 1975 and again in 2007, and insisted that system relevance and user relevance are different manifestations that cannot be reduced to one another. TREC carries the Cranfield design forward at scale. Relevance here is a judgment about a document, binary or graded, pooled across assessors and treated as ground truth.[5][6][7]

Unit

A query–document pair

Recovery means

Not applicable, the criterion is retrieval effectiveness, not a return to a state

Measured by

Assessor relevance judgments; precision, recall, and graded measures such as nDCG

A feature is relevant if removing it changes what can be predicted about the target. Kohavi and John gave the standard definitions in 1997: a feature is strongly relevant if the Bayes-optimal classifier degrades without it, weakly relevant if it helps only in some subset of features, and irrelevant otherwise; they also showed that relevance defined this way does not settle whether a learner should use the feature, and proposed wrappers that select features by the learner's own accuracy. Blum and Langley's review in the same issue set out filter, wrapper, and embedded approaches. Relevance here is a statistical dependence between a variable and a target, defined relative to the other variables available and to the learner.[8][9]

Unit

A feature in relation to a target and a feature set

Recovery means

Not applicable

Measured by

Held-out accuracy with and without the feature; mutual information and other filter statistics

A learner's perception that instruction connects to their goals, interests, and prior experience. Keller's ARCS model of 1987 lists relevance as the second of four conditions for motivation, after attention and before confidence and satisfaction, and prescribes strategies for producing it: goal orientation, motive matching, familiarity. It is measured by self-report on the Instructional Materials Motivation Survey, where relevance is a subscale of the learner's rating of the materials. Relevance here is a judgment the learner makes about a lesson, and its criterion is effort and persistence, not truth.[10]

Unit

A learner's perception of a piece of instruction

Recovery means

Not applicable

Measured by

IMMS relevance subscale; persistence and effort as outcomes

The properties of an event that make it likely to be selected and reported as news. Galtung and Ruge proposed twelve factors in 1965 from a study of foreign news in Norwegian papers (frequency, threshold, unambiguity, meaningfulness, consonance, unexpectedness, continuity, composition, and reference to elite nations, elite people, persons, and the negative) and hypothesised that factors are additive and that selection amplifies distortion. Harcup and O'Neill tested the list against British newspapers in 2001 and revised it again in 2017 for the social-media era, adding exclusivity, shareability, and audio-visuals. Relevance here is one named value among many, meaning proximity or bearing on the audience, and the wider construct is a property of events as journalists rank them.[11][12][13]

Unit

An event as a candidate for publication

Recovery means

Not applicable

Measured by

Content analysis of published stories coded for the presence of each value