Rees, Reksnes & Rohde (2026) @cite{rees-reksnes-rohde-2026} #
Why are you telling me this? The availability and timing of relevance inferences. Journal of Memory and Language 148, 104741.
@cite{sperber-wilson-1986} @cite{bergen-grodner-2012} @cite{kravtchenko-demberg-2022} @cite{bott-noveck-2004}
Key contribution #
Identifies relevance inferences: pragmatic inferences from trivial utterances where listeners reason about WHY the speaker chose to speak. Unlike scalar implicatures, these are not triggered by a weaker scalar term — there is no scale to search, and no stronger alternative the speaker chose not to use. Instead, the speaker's decision to produce a trivial utterance like "the library walls are blue" violates expectations of relevance, prompting the listener to infer enriched content (e.g., the walls used to be a different colour).
This distinction from scalar implicatures is the paper's central theoretical contribution: the phenomenon is about why did the speaker choose to speak at all?, not why did the speaker use a weaker term?.
Formalization strategy #
We model relevance inferences using RT's comprehension procedure (@cite{sperber-wilson-1986}). A trivial utterance's literal interpretation fails the relevance threshold when the speaker is knowledgeable (its content is already expected), forcing the comprehension procedure to search further. The enriched interpretation ("something has changed") passes threshold when warranted by speaker knowledge.
Speaker factors as effect modulation #
Both manipulated factors — Speaker Knowledge and Speaker Style — modulate the EFFECTS of the enriched interpretation, not the threshold:
Speaker Knowledge (Exps 1–4): A knowledgeable speaker's trivial utterance produces stronger contextual implications for the "changed" interpretation, because the inference is warranted (the speaker would notice changes to a familiar location). Cf. @cite{bergen-grodner-2012}: speaker knowledge modulates SI computation in the same direction.
Speaker Style (Exp 2): A quiet speaker's decision to speak amplifies the effects of the enriched interpretation — if they bothered to mention it, there must be a strong reason. This connects to @cite{burnett-2019}'s persona framework.
Both are instances of effect strengthening, the monotonicity property
of RT's comprehension procedure (RTScenario.selects_of_strengthened_effects).
Emphasis Cue null effect (Exp 1) #
"Hey, guess what" does not significantly increase inference rates. This follows from RT's communicative principle: every act of ostensive communication already presumes its own optimal relevance (@cite{sperber-wilson-1986}, 2nd ed). An explicit emphasis cue is redundant with the ostensive act of speaking itself.
Processing cost (Exps 3–4) #
Relevance inferences are costly to compute: cross-experiment analysis of Exps 3–4 shows inference-endorsing "Yes" responses (Exp 4, where "Yes" = inference) are slower than non-inference "Yes" responses (Exp 3, where "Yes" = no inference). Cf. @cite{bott-noveck-2004}: scalar implicature computation shows the same response-time cost pattern.
Graded predictions #
RT's comprehension procedure is qualitative (binary selection), but the monotonicity theorem is independent of threshold shape: it holds for both a step function (qualitative RT) and any sigmoid extension (graded RT). The ordinal predictions — Speaker Knowledge and Speaker Style each strengthen the enriched interpretation — produce graded proportion predictions under any noisy-threshold model.
Relation to atypicality inferences #
@cite{kravtchenko-demberg-2022} found that informationally redundant utterances (explicitly mentioning a script-typical action like "eating at a restaurant") trigger atypicality inferences — listeners infer the action was atypical for the character. Rees et al.'s relevance inferences extend this: even non-redundant but TRIVIAL utterances trigger enriched inferences about why the speaker chose to speak.
Data summary (4 experiments, N = 777) #
- Exp 1 (N=197): Speaker Knowledge significant (p<.001); Emphasis Cue n.s.
- Exp 2 (N=192): Speaker Knowledge significant (p<.001); Speaker Style significant (p=.001); Knowledge × Style interaction (p=.026) driven by increased inferencing in unfamiliar+quiet condition
- Exp 3 (N=194): Speaker Knowledge replicates; response time pattern numerically consistent with processing cost but not statistically robust
- Exp 4 (N=194): Speaker Knowledge replicates; cross-experiment analysis with Exp 3 confirms inference computation is costly
Candidate interpretations for a trivial utterance like "The library walls are blue."
A listener can: (a) accept the literal content transparently — "the walls are blue" (b) infer that the situation has changed — relevance inference (c) treat the utterance as a social gesture with no enrichment
The paper's experimental task is binary ("same" vs "different"), but the three-candidate decomposition captures the RT comprehension procedure's search space: the procedure considers candidates in accessibility order and stops at the first that passes threshold. The phatic interpretation is maximally costly to access and never selected, but its presence in the candidate set reflects that phatic communication is always available as a last resort (@cite{crystal-1991}, @cite{zegarac-clark-1999}).
- literal : TrivialInterp
Transparent acceptance: "The walls are blue" (no enrichment)
- changed : TrivialInterp
Relevance inference: "The walls USED TO BE different"
- phatic : TrivialInterp
Phatic: social gesture, no contentful enrichment
Instances For
Equations
- One or more equations did not get rendered due to their size.
Instances For
Sample size per experiment (after exclusions).
Equations
Instances For
Speaker Knowledge was significant (p < .001) in all four experiments.
Instances For
The direction of the Knowledge effect is consistent across all experiments: familiar School > unfamiliar PM.
- schoolGreaterThanPM : KnowledgeDirection
Instances For
Equations
- One or more equations did not get rendered due to their size.
Instances For
Familiar location, normal speaker (School condition).
The literal reading has low effects: a knowledgeable speaker saying something mundane about a familiar location is unsurprising (effects 1). The enriched "something has changed" reading is warranted and produces strong cognitive effects (effects 3): contextual implication that the speaker noticed a change to the familiar environment.
The enriched effect decomposes as EffectType.contextualImplication:
a new conclusion ("the walls changed") derivable only from the input
plus the contextual assumption that the speaker is knowledgeable.
Equations
- One or more equations did not get rendered due to their size.
Instances For
Unfamiliar location, normal speaker (Prime Minister's office condition).
The literal reading is genuinely informative: neither speaker nor listener knew what the PM's office looked like, so "the walls are blue" is novel information (effects 2, passes threshold). The enriched reading is not warranted: the speaker has no baseline knowledge of the location, so inferring change is unjustified (effects 1, fails threshold).
Equations
- One or more equations did not get rendered due to their size.
Instances For
Quiet speaker, familiar location.
A quiet speaker's decision to speak carries extra weight: the rarity of their speech makes each utterance more noteworthy, amplifying the cognitive effects of the enriched interpretation (effects 4 vs 3). The literal reading remains subthreshold (effects 1).
This connects to @cite{burnett-2019}'s persona framework: "quiet"
is a stable speaker quality (Core.SocialMeaning.StanceLevel.quality)
that modulates the listener's expectations of the speaker's
communicative threshold.
Equations
- One or more equations did not get rendered due to their size.
Instances For
Quiet speaker, unfamiliar location.
The quiet speaker's reticence provides a small boost to the enriched interpretation's effects even in the unfamiliar case (effects 2 vs 1 for normal+unfamiliar), but the literal reading still passes threshold (effects 2) and is more accessible, so the qualitative comprehension procedure selects literal.
This scenario captures the Experiment 2 interaction: the Knowledge × Style interaction (p = .026) was driven by increased inferencing in the unfamiliar+quiet condition compared to unfamiliar+normal. The authors speculate that the quiet speaker mentioning something about an unfamiliar location leads participants to infer the speaker is trying to convey some additional meaning, but the only available response was "different" — even if the actual inference was something other than "the situation has changed."
The qualitative RT model predicts literal selection here (both literal
and changed have effects 2, but literal is more accessible). The
elevated inference rate (~30% vs ~15% for normal+unfamiliar) reflects
graded/noisy-threshold effects: the closer .changed gets to
threshold, the more often a noisy threshold lets it through.
Equations
- One or more equations did not get rendered due to their size.
Instances For
Familiar + normal speaker: the comprehension procedure selects the "changed" interpretation.
The literal reading fails threshold (1 < 2), so the procedure searches further. The enriched reading passes (3 ≥ 2) and is accepted.
This formalizes the core finding across all four experiments: trivial utterances from knowledgeable speakers trigger relevance inferences at higher rates than from non-knowledgeable speakers.
Unfamiliar + normal speaker: the literal reading is selected.
The literal reading passes threshold (2 ≥ 2) and is the most accessible candidate, so the procedure stops there. The enriched reading is never reached — there is no need to search further.
Quiet + unfamiliar: literal is still selected.
Despite the reticence-driven boost to the enriched interpretation (effects 2, at threshold), the literal reading is more accessible and also passes threshold, so the satisficing procedure stops there.
The familiar and quiet-familiar scenarios are structural variants: same candidates, same accessibility ordering, same search path.
Speaker Style preserves selection via effect strengthening.
The quiet-familiar scenario is a structural variant of the familiar
scenario with strictly higher effects for .changed (4 > 3) and
equal effects for all other candidates. By the monotonicity theorem
(selects_of_strengthened_effects), the .changed interpretation
is selected — the quiet speaker's reticence REINFORCES the inference.
This formalizes Experiment 2's Speaker Style finding (p = .001): quiet speakers produce more relevance inferences.
Speaker Knowledge increases the enriched interpretation's effects.
This is the ordinal prediction that holds independently of whether the comprehension procedure is a step function (qualitative RT) or a sigmoid (graded extension). Under any noisy-threshold model, this ordering produces higher inference rates in the familiar condition.
Speaker Knowledge decreases the literal interpretation's effects.
Complementary to the enriched strengthening: a knowledgeable speaker saying something mundane is LESS informative literally (they could say more interesting things about a familiar location).
Speaker Style (quiet) strengthens the enriched interpretation beyond the Speaker Knowledge baseline.
Reticence also boosts the enriched interpretation in the unfamiliar condition, formalizing the source of the Knowledge × Style interaction in Experiment 2 (p = .026).
In the data, inferencing in the unfamiliar+quiet condition was elevated
relative to unfamiliar+normal, narrowing the Knowledge effect gap.
In the model, this is a strict increase in .changed effects (2 > 1)
that does not cross the selection boundary (literal still wins) but
would produce higher inference rates under a noisy threshold.
The Knowledge effect gap is smaller under quiet speaker style than under normal style: reticence narrows the difference between familiar and unfamiliar conditions for the enriched interpretation.
familiar − unfamiliar gap: Normal: 3 − 1 = 2 Quiet: 4 − 2 = 2
In this qualitative encoding the gaps are equal. The observed interaction (p = .026) reflects that the quiet speaker's boost to the unfamiliar condition proportionally closes more of the gap than the boost to the familiar condition, an effect visible only under a graded (noisy-threshold) extension where the enriched interpretation's distance from threshold matters.
Why no emphasisScenario #
Every act of ostensive communication already creates a presumption
of optimal relevance (@cite{sperber-wilson-1986}, 2nd ed). The emphasis
cue "Hey, guess what" is an ostensive stimulus, but so is the act of
speaking itself. Since the presumption of relevance is already triggered
by the base utterance, the emphasis cue is redundant — it maps to no
parameter change in the RTScenario.
This accounts for Experiment 1's null finding for Emphasis Cue (n.s.): "Hey, guess what" does not significantly increase inference rates. The decision to speak is itself a sufficient ostensive stimulus; explicit cueing adds nothing to the relevance presumption.
There is no formal theorem here because the claim is that the emphasis
cue changes NO model parameter — familiarScenario already models
both the cued and uncued conditions. A theorem familiarScenario = familiarScenario would be vacuously true (rfl), encoding no content.
The enriched interpretation has nonzero processing effort.
Computing the "something has changed" inference requires reasoning about the speaker's goals and knowledge — effort beyond decoding the literal content. This nonzero effort maps onto the response time cost observed in the cross-experiment analysis of Exps 3–4: inference-endorsing "Yes" responses (Exp 4) are slower than non-inference "Yes" responses (Exp 3).
Cf. @cite{bott-noveck-2004}: scalar implicature computation shows the same response-time cost pattern in sentence verification tasks.
The literal interpretation is effort-free: it requires only decoding the semantic content, no pragmatic reasoning.
In the current scenarios, effort does not block the enriched interpretation (effortWeight = 0). The effort manifests as processing cost (response time), not as relevance reduction.
A future extension could set effortWeight > 0 to model time-pressure contexts where pragmatic enrichment IS blocked by effort — predicting lower inference rates under cognitive load (cf. @cite{bale-etal-2025} for load effects on competence assumption cancellation).
Clause (b): the speaker's choice to speak #
The deep mechanism behind relevance inferences is clause (b) of optimal relevance (@cite{sperber-wilson-1986}): the ostensive stimulus is the most relevant one compatible with the communicator's abilities and preferences. The speaker chose to SPEAK rather than remain silent. Since silence would have been easier (zero effort, zero risk of irrelevance), the hearer infers the speaker had a communicative goal that required speaking. The hearer then searches for an interpretation that makes the utterance relevant.
Unlike the scalar implicature case ("some" vs "all"), the alternative
here is SILENCE, not a stronger utterance. The existing ClauseBArgument
type models stimulus-level alternatives (alternative utterances the
speaker could have produced), not the speaking-vs-silence choice. A
future extension could add a SpeakingDecision type to model this.
The "something has changed" inference is a contextual implication: a new conclusion derivable only from the utterance content PLUS the contextual assumption that the speaker is knowledgeable about the location.
This classifies relevance inferences within S&W's effect taxonomy. The inference is not strengthening (no prior assumption about the walls being different) and not contradiction (the literal content isn't contradicted). It is a genuinely new conclusion that requires both the input and the context.
Equations
Instances For
RT correctly predicts the Knowledge effect direction:
the familiar scenario selects .changed while the unfamiliar
selects .literal, matching the School > PM pattern replicated
across all four experiments (p < .001 in each).
Compare @cite{bergen-grodner-2012}: speaker knowledge modulates scalar implicature computation in the same direction — listeners are less likely to compute SIs when the speaker has "skimmed" rather than "read meticulously." Rees et al. extend this to relevance inferences from trivial utterances, showing that speaker knowledge effects generalize beyond scale-based inferencing.
RT correctly predicts the Style effect direction:
quiet+familiar selects .changed (via monotonicity), matching
the Quiet > Normal pattern in Experiment 2 (p = .001).
The Style effect is captured by effect strengthening, not threshold lowering. Increasing effects for the target while holding blocking candidates constant preserves (and strengthens) selection.