Relevance Theory (Sperber & Wilson)

The problem Grice left open

Grice's cooperative principle gave us four maxims and a calculus for deriving implicatures. But the maxims look ad hoc — why exactly four? Why those four? And how does a hearer pick which maxim a speaker is observing or flouting in real time, without obvious access to alternatives? Linguists in the 1970s and 80s pressed these questions. The answer Dan Sperber and Deirdre Wilson published in 1986 — Relevance: Communication and Cognition — is that there is no need for multiple maxims. One cognitive principle does the work.

That principle: humans process information by automatically searching for interpretations that maximise relevance, defined as a tradeoff between two quantities — cognitive effects (worthwhile changes in mental state) and processing effort. The more effects an input produces and the less effort it costs, the more relevant it is.

The two principles

Relevance theory is built on two principles, often confused. They cover distinct things:

• The cognitive principle of relevance. Human cognition tends to be geared to the maximisation of relevance. We evolved to sift inputs for those most likely to pay off in cognitive effects given the effort.
• The communicative principle of relevance. Every ostensive act of communication (one in which the communicator visibly intends to inform the audience of something) conveys a presumption of its own optimal relevance. The audience can assume: (a) the act is relevant enough to be worth processing, and (b) it is the most relevant the communicator could and would have produced, given their abilities and preferences.

The cognitive principle is a fact about psychology. The communicative principle is the contract a speaker enters by speaking ostensively — they would not say something if they did not believe it would clear the optimality bar.

Cognitive effects, formalised

Sperber and Wilson identify three kinds of cognitive effect a new input can have when combined with the hearer's existing context:

1. New contextual implication. The input combined with old assumptions yields a conclusion that neither piece would yield alone. "It's raining" + the assumption "If it's raining, the parade is cancelled" → "The parade is cancelled."
2. Strengthening an existing assumption. The input provides additional evidence for something already believed weakly. "The forecast says rain" increases your existing low-confidence belief that it will rain.
3. Contradicting and eliminating an existing assumption. The input flatly conflicts with a held belief and replaces it. "Actually, the forecast was wrong; it's sunny" kills the rain assumption.

An interpretation is relevant in proportion to how many such effects it produces and inversely to the effort needed to produce them. The cognitive system is heuristic: it pursues a path of interpretation, accumulates effects, and stops when its search rate or effect rate falls — i.e. when it has the first interpretation that satisfies the optimality presumption.

Processing effort, formalised

Effort is whatever cognitive resource an interpretation costs. The principal sources are:

• Decoding. Phonological, syntactic, and lexical decoding of the form.
• Reference resolution. Identifying what indexicals, pronouns, and definites pick out.
• Disambiguation. Choosing among lexical or structural alternatives.
• Inference. Combining the input with contextual assumptions to derive the explicature and implicatures.

An utterance with marked syntax, rare vocabulary, or difficult-to-resolve indexicals incurs more effort. By the optimality presumption, the speaker would not impose that effort unless it were rewarded with extra effects. So a marked form licenses extra inference: "That woman over there — yes, the one in the orange dress — is a poker champion" incurs effort to identify the referent, and the hearer infers there is a reason this referent in particular is being singled out.

Explicature: a missing layer

Grice treated the boundary between sentence-meaning and speaker-meaning as crisp: literal semantics, then implicatures. Sperber and Wilson — and especially Robyn Carston in Thoughts and Utterances (2002) — argue the literal sentence is almost never enough to fix what's said. Several pragmatic moves are needed before implicatures kick in:

• Disambiguation. "The bank is closed" — financial vs river.
• Reference assignment. "She is here" needs she and here assigned.
• Free enrichment. "I haven't eaten" usually means "today", supplied by context.
• Saturation. "Mary is ready" needs the hearer to fill in ready for what.
• Loosening / narrowing. "France is hexagonal" is loose; "I have a temperature" is narrowed.

The output is the explicature: the pragmatically developed proposition that captures what was explicitly said. Implicatures are then derived from the explicature plus context. This dissolves Grice's awkward "what is said" — RT treats it as already pragmatic.

Relevance theory vs Gricean pragmatics

Dimension	Gricean pragmatics	Relevance theory
Number of guiding principles	Cooperative principle + four maxims	Single cognitive principle of relevance
Hearer's reasoning	"Why did the speaker observe / flout this maxim?"	"What interpretation maximises effects for minimum effort?"
Layers of meaning	What is said vs what is implicated	Explicature vs implicature
Status	Normative — speakers should cooperate	Cognitive — humans do seek relevance
Treats metaphor as	Flouting Quality	Lexical narrowing / loosening (ad hoc concepts)
Stopping rule	Implicit; first plausible interpretation	Explicit: stop when optimal-relevance presumption is met

A worked example: ad hoc concepts

Suppose Anna says to Ben:

"My lawyer is a shark."

Grice would derive this by floating Quality: the speaker says something patently false and so must mean something else — namely, that the lawyer is aggressive and predatory. Relevance theory rejects the false-step. Instead:

1. Ben decodes the sentence and constructs an initial explicature: Anna's lawyer is a member of the category SHARK.
2. The encoded concept SHARK is too narrow to be worth processing literally — there is little effect from believing Anna's lawyer is an aquatic predator.
3. Ben adjusts the concept on the fly. He builds an ad hoc concept SHARK* that drops the literal aquatic features and retains the inferentially active ones: aggressive, predatory, dangerous to opponents.
4. The explicature becomes: Anna's lawyer is SHARK*, and Ben gets ample cognitive effects (predictions about courtroom behaviour).

The same machinery handles loose use ("France is hexagonal" → ad hoc concept HEXAGONAL* that ignores fine borders), narrowing ("I have a temperature" → TEMPERATURE* meaning specifically a high one), and approximation ("It's three o'clock" → THREE-O'CLOCK* meaning "around three"). Metaphor and literal language sit on a continuum, not a dichotomy.

Ostensive-inferential communication

Sperber and Wilson distinguish two intentions behind any ostensive act: the informative intention (to make a set of assumptions manifest to the audience) and the communicative intention (to make the informative intention itself manifest). You can have the first without the second — if you accidentally leave a love letter on the counter, you make assumptions manifest without communicating. Ostensive communication requires both: the speaker is showing the hearer they are showing them something.

Cross-linguistic and developmental data

• Procedural vs conceptual encoding. RT treats some forms (discourse markers like English well, French donc; pronouns; tense morphology) as procedural — they encode constraints on how to derive the explicature, not descriptions of the world. Blakemore's 1987 Semantic Constraints on Relevance argues but procedurally signals that the second clause rules out an inference from the first.
• Children's pragmatic development. Noveck and others show children compute explicatures earlier than scalar implicatures — explained under RT by the extra effort of generating alternatives.
• Translation. Gutt's 1991 Translation and Relevance applies the framework: a faithful translation reproduces the original's effort/effect balance, not its surface form.
• Autism. Sperber and Wilson argue pragmatic deficits in autism reflect difficulty with second-order metarepresentation — the mind-reading ostensive communication requires.

Critiques

• Operationalisation. Effort and effect are theoretical constructs, not directly measurable. Critics (Levinson 2000) argue "maximising effects for minimum effort" is too elastic for sharp predictions.
• Generalised conversational implicatures. Levinson's defaults (some ↝ not all) look more rule-bound than RT's case-by-case effort calculus comfortably allows.
• Comprehension vs production. RT focuses on the hearer; modern game-theoretic models (Frank & Goodman 2012; rational speech acts) fill in the production side.

Why relevance theory matters

• Unifies metaphor, irony, and literal language as outputs of the same machinery — no special "figurative" rules.
• Fits cognitive psychology: effort and effect map onto bounded-rationality accounts of mind.
• Predicts processing data. Motivates much of the experimental pragmatics literature on reading times for scalar inferences, indirect requests, and metaphor.
• Enriches semantics. The explicature/implicature distinction has been adopted by formal semanticists worried about underdetermination.

Common pitfalls

• "Relevance" is the everyday word. RT uses it as a technical term — a ratio of cognitive effects to processing effort. Don't read everyday salience or topic-relevance into it.
• Treating effort as bad and effects as good. A speaker can legitimately impose more effort if it is repaid with proportionally more effects (this is how poetic effects emerge in RT).
• Confusing the two principles. The cognitive principle is descriptive; the communicative principle is the presumption a hearer derives from an act being ostensive.
• Forgetting explicatures. Many things people call "implicatures" in informal pragmatics are actually explicatures in RT terms — pragmatic enrichment of what is said, not extra implicit propositions.
• Reading metaphor as Quality flouting. RT treats it as continuous with literal language via ad hoc concepts; the Gricean two-step is unnecessary.