@cite{ronderos-etal-2024}

@cite{sedivy-etal-1999}

Perceptual, Semantic, and Pragmatic Factors Affect the Derivation of Contrastive Inferences. Open Mind: Discoveries in Cognitive Science 8, 1213–1227.

Core Argument

Cross-linguistic eye-tracking (English, Hindi, Hungarian; N = 97) using the @cite{sedivy-etal-1999} contrastive inference paradigm shows that adjective type modulates whether listeners draw contrastive inferences:

• Color adjectives elicit contrastive inferences — contra Sedivy (2003, 2004), who argued color adjectives are used descriptively and therefore do not trigger contrastive interpretations.
• Scalar adjectives elicit contrastive inferences — replicating @cite{sedivy-etal-1999}.
• Material adjectives do NOT elicit contrastive inferences — interpreted as an effect of low visual salience of material properties.

Three factors interact:

1. Perceptual: color contrast is visually salient, material is not
2. Semantic: scalar adjectives require comparison-class computation (more distributed gaze in baseline), color/material do not
3. Pragmatic: informativity expectations drive contrastive inference only when perceptual access is fast enough

Connection to Noise Theory

The contrastive inference pattern aligns with the noise discrimination ordering from RSA.Noise: color (0.98) > size (0.60) > material (0.40). High discrimination → strong contrastive signal → contrastive inference; low discrimination → weak signal → no contrastive inference. This extends @cite{kursat-degen-2021}'s production-side finding (redundant modification rate) to the comprehension side (contrastive inference).

Verified Data

All regression coefficients and cluster statistics verified against paper text (§3.1–§3.4). SE is not reported for target-advantage or baseline analyses and is omitted here.

inductive Phenomena.Reference.Studies.RonderosEtAl2024.AdjType : Type

The three adjective types tested. Maps to Core.PropertyDomain: color → .color, material → .material, scalar → .size (the scalar items are spatial dimensions).

• color : AdjType
• material : AdjType
• scalar : AdjType

instance Phenomena.Reference.Studies.RonderosEtAl2024.instDecidableEqAdjType : DecidableEq AdjType

Equations

• Phenomena.Reference.Studies.RonderosEtAl2024.instDecidableEqAdjType x✝ y✝ = if h : x✝.ctorIdx = y✝.ctorIdx then isTrue ⋯ else isFalse ⋯

instance Phenomena.Reference.Studies.RonderosEtAl2024.instBEqAdjType : BEq AdjType

Equations

• Phenomena.Reference.Studies.RonderosEtAl2024.instBEqAdjType = { beq := Phenomena.Reference.Studies.RonderosEtAl2024.instBEqAdjType.beq }

def Phenomena.Reference.Studies.RonderosEtAl2024.instBEqAdjType.beq : AdjType → AdjType → Bool

Equations

• Phenomena.Reference.Studies.RonderosEtAl2024.instBEqAdjType.beq x✝ y✝ = (x✝.ctorIdx == y✝.ctorIdx)

def Phenomena.Reference.Studies.RonderosEtAl2024.instReprAdjType.repr : AdjType → ℕ → Std.Format

Equations

• One or more equations did not get rendered due to their size.

instance Phenomena.Reference.Studies.RonderosEtAl2024.instReprAdjType : Repr AdjType

Equations

• Phenomena.Reference.Studies.RonderosEtAl2024.instReprAdjType = { reprPrec := Phenomena.Reference.Studies.RonderosEtAl2024.instReprAdjType.repr }

def Phenomena.Reference.Studies.RonderosEtAl2024.AdjType.toDomain : AdjType → Core.PropertyDomain

Map adjective type to PropertyDomain.

Equations

• Phenomena.Reference.Studies.RonderosEtAl2024.AdjType.color.toDomain = Core.PropertyDomain.color
• Phenomena.Reference.Studies.RonderosEtAl2024.AdjType.material.toDomain = Core.PropertyDomain.material
• Phenomena.Reference.Studies.RonderosEtAl2024.AdjType.scalar.toDomain = Core.PropertyDomain.size

inductive Phenomena.Reference.Studies.RonderosEtAl2024.Language : Type

Three languages tested.

• english : Language
• hindi : Language
• hungarian : Language

instance Phenomena.Reference.Studies.RonderosEtAl2024.instDecidableEqLanguage : DecidableEq Language

Equations

• Phenomena.Reference.Studies.RonderosEtAl2024.instDecidableEqLanguage x✝ y✝ = if h : x✝.ctorIdx = y✝.ctorIdx then isTrue ⋯ else isFalse ⋯

def Phenomena.Reference.Studies.RonderosEtAl2024.instBEqLanguage.beq : Language → Language → Bool

Equations

• Phenomena.Reference.Studies.RonderosEtAl2024.instBEqLanguage.beq x✝ y✝ = (x✝.ctorIdx == y✝.ctorIdx)

instance Phenomena.Reference.Studies.RonderosEtAl2024.instBEqLanguage : BEq Language

Equations

• Phenomena.Reference.Studies.RonderosEtAl2024.instBEqLanguage = { beq := Phenomena.Reference.Studies.RonderosEtAl2024.instBEqLanguage.beq }

def Phenomena.Reference.Studies.RonderosEtAl2024.instReprLanguage.repr : Language → ℕ → Std.Format

Equations

• One or more equations did not get rendered due to their size.

instance Phenomena.Reference.Studies.RonderosEtAl2024.instReprLanguage : Repr Language

Equations

• Phenomena.Reference.Studies.RonderosEtAl2024.instReprLanguage = { reprPrec := Phenomena.Reference.Studies.RonderosEtAl2024.instReprLanguage.repr }

def Phenomena.Reference.Studies.RonderosEtAl2024.Language.n : Language → ℕ

Equations

• Phenomena.Reference.Studies.RonderosEtAl2024.Language.english.n = 49
• Phenomena.Reference.Studies.RonderosEtAl2024.Language.hindi.n = 27
• Phenomena.Reference.Studies.RonderosEtAl2024.Language.hungarian.n = 21

def Phenomena.Reference.Studies.RonderosEtAl2024.totalN : ℕ

Total participants analyzed across all three languages (108 recruited, 97 after exclusions).

Equations

• One or more equations did not get rendered due to their size.

theorem Phenomena.Reference.Studies.RonderosEtAl2024.total_is_97 : totalN = 97

inductive Phenomena.Reference.Studies.RonderosEtAl2024.Condition : Type

Contrast condition: whether the visual display contains a competitor object of the same category that differs on the adjective dimension.

• contrast : Condition
• noContrast : Condition

instance Phenomena.Reference.Studies.RonderosEtAl2024.instDecidableEqCondition : DecidableEq Condition

Equations

• Phenomena.Reference.Studies.RonderosEtAl2024.instDecidableEqCondition x✝ y✝ = if h : x✝.ctorIdx = y✝.ctorIdx then isTrue ⋯ else isFalse ⋯

def Phenomena.Reference.Studies.RonderosEtAl2024.instBEqCondition.beq : Condition → Condition → Bool

Equations

• Phenomena.Reference.Studies.RonderosEtAl2024.instBEqCondition.beq x✝ y✝ = (x✝.ctorIdx == y✝.ctorIdx)

instance Phenomena.Reference.Studies.RonderosEtAl2024.instBEqCondition : BEq Condition

Equations

• Phenomena.Reference.Studies.RonderosEtAl2024.instBEqCondition = { beq := Phenomena.Reference.Studies.RonderosEtAl2024.instBEqCondition.beq }

instance Phenomena.Reference.Studies.RonderosEtAl2024.instReprCondition : Repr Condition

Equations

• Phenomena.Reference.Studies.RonderosEtAl2024.instReprCondition = { reprPrec := Phenomena.Reference.Studies.RonderosEtAl2024.instReprCondition.repr }

def Phenomena.Reference.Studies.RonderosEtAl2024.instReprCondition.repr : Condition → ℕ → Std.Format

Equations

• One or more equations did not get rendered due to their size.

def Phenomena.Reference.Studies.RonderosEtAl2024.itemsPerType : ℕ

Number of items per adjective type (8 adjectives × 3 nouns).

Equations

• Phenomena.Reference.Studies.RonderosEtAl2024.itemsPerType = 24

def Phenomena.Reference.Studies.RonderosEtAl2024.totalTrials : ℕ

Total experimental trials (+ 72 filler trials).

Equations

• Phenomena.Reference.Studies.RonderosEtAl2024.totalTrials = Phenomena.Reference.Studies.RonderosEtAl2024.itemsPerType * 3

theorem Phenomena.Reference.Studies.RonderosEtAl2024.trials_is_72 : totalTrials = 72

structure Phenomena.Reference.Studies.RonderosEtAl2024.ClusterResult : Type

Result from cluster-based permutation test on target-advantage difference curves (Contrast − No-Contrast).

• clusterStart : ℕ

  Start of significant cluster (ms post-adjective onset)

• clusterEnd : ℕ

  End of significant cluster (ms)

• sumT : Float

  Sum of t-values across the cluster

• significant : Bool

  Whether the cluster reached significance (p < 0.01)

def Phenomena.Reference.Studies.RonderosEtAl2024.instReprClusterResult.repr : ClusterResult → ℕ → Std.Format

Equations

• One or more equations did not get rendered due to their size.

instance Phenomena.Reference.Studies.RonderosEtAl2024.instReprClusterResult : Repr ClusterResult

Equations

• Phenomena.Reference.Studies.RonderosEtAl2024.instReprClusterResult = { reprPrec := Phenomena.Reference.Studies.RonderosEtAl2024.instReprClusterResult.repr }

def Phenomena.Reference.Studies.RonderosEtAl2024.cluster_color : ClusterResult

Color: significant cluster 240–600ms (§3.1).

Equations

• Phenomena.Reference.Studies.RonderosEtAl2024.cluster_color = { clusterStart := 240, clusterEnd := 600, sumT := 39.61, significant := true }

def Phenomena.Reference.Studies.RonderosEtAl2024.cluster_scalar : ClusterResult

Scalar: significant cluster 260–500ms (§3.1).

Equations

• Phenomena.Reference.Studies.RonderosEtAl2024.cluster_scalar = { clusterStart := 260, clusterEnd := 500, sumT := 33.07, significant := true }

def Phenomena.Reference.Studies.RonderosEtAl2024.cluster_material : ClusterResult

Material: no significant cluster (§3.1).

Equations

• Phenomena.Reference.Studies.RonderosEtAl2024.cluster_material = { clusterStart := 0, clusterEnd := 0, sumT := 0.0, significant := false }

def Phenomena.Reference.Studies.RonderosEtAl2024.cluster_interaction : ClusterResult

Adjective Type × Condition interaction: significant cluster 280–600ms (§3.1).

Equations

• Phenomena.Reference.Studies.RonderosEtAl2024.cluster_interaction = { clusterStart := 280, clusterEnd := 600, sumT := 37.96, significant := true }

structure Phenomena.Reference.Studies.RonderosEtAl2024.TargetAdvantageResult : Type

Mixed-effects regression result for target-advantage score (mean proportion of looks to target, 200–800ms window). Note: paper reports β, t, and p but not SE.

• beta : Float

  Fixed-effect coefficient (Contrast − No-Contrast)

• tStat : Float

  t-statistic

• pValue : Option Float

  p-value (exact when reported, else threshold)

• significant : Bool

  Whether effect is significant (p < 0.05)

def Phenomena.Reference.Studies.RonderosEtAl2024.instReprTargetAdvantageResult.repr : TargetAdvantageResult → ℕ → Std.Format

Equations

• One or more equations did not get rendered due to their size.

instance Phenomena.Reference.Studies.RonderosEtAl2024.instReprTargetAdvantageResult : Repr TargetAdvantageResult

Equations

• Phenomena.Reference.Studies.RonderosEtAl2024.instReprTargetAdvantageResult = { reprPrec := Phenomena.Reference.Studies.RonderosEtAl2024.instReprTargetAdvantageResult.repr }

def Phenomena.Reference.Studies.RonderosEtAl2024.targetAdv_color : TargetAdvantageResult

Color: β = 0.24, t = 2.41, p < 0.05 (§3.2).

Equations

• Phenomena.Reference.Studies.RonderosEtAl2024.targetAdv_color = { beta := 0.24, tStat := 2.41, significant := true }

def Phenomena.Reference.Studies.RonderosEtAl2024.targetAdv_scalar : TargetAdvantageResult

Scalar: β = 0.19, t = 2.02, p < 0.05 (§3.2).

Equations

• Phenomena.Reference.Studies.RonderosEtAl2024.targetAdv_scalar = { beta := 0.19, tStat := 2.02, significant := true }

def Phenomena.Reference.Studies.RonderosEtAl2024.targetAdv_material : TargetAdvantageResult

Material: β = 0.10, t = 1.08, p = 0.28 (§3.2).

Equations

• Phenomena.Reference.Studies.RonderosEtAl2024.targetAdv_material = { beta := 0.10, tStat := 1.08, pValue := some 0.28, significant := false }

structure Phenomena.Reference.Studies.RonderosEtAl2024.BaselineResult : Type

Baseline comparison of overall looks to both target and competitor in the No-Contrast condition (200–800ms). Tests whether adjective types differ in how efficiently participants fixate on the critical images, independent of contrastive inference. Positive β means more looks to both critical images for the first adjective type. Note: paper reports β, z, and p but not SE.

• beta : Float

  Fixed-effect coefficient

• zStat : Float

  z-statistic

• significant : Bool

  Whether effect is significant

instance Phenomena.Reference.Studies.RonderosEtAl2024.instReprBaselineResult : Repr BaselineResult

Equations

• Phenomena.Reference.Studies.RonderosEtAl2024.instReprBaselineResult = { reprPrec := Phenomena.Reference.Studies.RonderosEtAl2024.instReprBaselineResult.repr }

def Phenomena.Reference.Studies.RonderosEtAl2024.instReprBaselineResult.repr : BaselineResult → ℕ → Std.Format

Equations

• One or more equations did not get rendered due to their size.

def Phenomena.Reference.Studies.RonderosEtAl2024.baseline_colorVsScalar : BaselineResult

Color vs Scalar in No-Contrast: β = 0.25, z = 2.80, p < 0.01 (§3.3). Participants fixated more on both critical images in color trials than in scalar trials.

Equations

• Phenomena.Reference.Studies.RonderosEtAl2024.baseline_colorVsScalar = { beta := 0.25, zStat := 2.80, significant := true }

def Phenomena.Reference.Studies.RonderosEtAl2024.baseline_materialVsScalar : BaselineResult

Material vs Scalar in No-Contrast: β = 0.24, z = 2.40, p < 0.05 (§3.3). Participants fixated more on both critical images in material trials than in scalar trials. Interpretation: scalar adjectives require comparison-class computation, leading to more distributed gaze across all four display images including distractors.

Equations

• Phenomena.Reference.Studies.RonderosEtAl2024.baseline_materialVsScalar = { beta := 0.24, zStat := 2.40, significant := true }

def Phenomena.Reference.Studies.RonderosEtAl2024.preNounBaselineClean : Bool

No significant effects in pre-noun window (trial onset to noun onset): no condition differences, no adjective-type differences, no interactions. Confirms effects in critical window are not due to anticipatory looking (§3.4).

Equations

• Phenomena.Reference.Studies.RonderosEtAl2024.preNounBaselineClean = true

theorem Phenomena.Reference.Studies.RonderosEtAl2024.contrastive_inference_pattern : cluster_color.significant = true ∧ cluster_scalar.significant = true ∧ ¬cluster_material.significant = true ∧ targetAdv_color.significant = true ∧ targetAdv_scalar.significant = true ∧ ¬targetAdv_material.significant = true

Color and scalar adjectives elicit contrastive inferences; material adjectives do not. Both cluster and target-advantage analyses converge.

theorem Phenomena.Reference.Studies.RonderosEtAl2024.interaction_significant : cluster_interaction.significant = true

The interaction between adjective type and condition is significant: the contrastive effect is not uniform across adjective types.

theorem Phenomena.Reference.Studies.RonderosEtAl2024.all_betas_positive : targetAdv_color.beta > 0 ∧ targetAdv_scalar.beta > 0 ∧ targetAdv_material.beta > 0

All target-advantage betas are positive: the Contrast condition always shows numerically more looks to target than No-Contrast, even for material (just not significantly).

theorem Phenomena.Reference.Studies.RonderosEtAl2024.effect_size_ordering : targetAdv_color.beta > targetAdv_scalar.beta ∧ targetAdv_scalar.beta > targetAdv_material.beta

Color has the largest contrastive effect, then scalar, then material: β_color > β_scalar > β_material.

theorem Phenomena.Reference.Studies.RonderosEtAl2024.scalar_baseline_disadvantage : baseline_colorVsScalar.significant = true ∧ baseline_materialVsScalar.significant = true

In the No-Contrast baseline, both color and material attract more fixations to the critical images (target + competitor) than scalar adjectives. Scalar adjectives require comparison-class computation, leading to more distributed gaze across all four display images.

theorem Phenomena.Reference.Studies.RonderosEtAl2024.discrimination_predicts_contrastive : AdjType.color.toDomain = Core.PropertyDomain.color ∧ Core.PropertyDomain.color.noiseDiscrimination = some RSA.Noise.colorDiscrimination ∧ targetAdv_color.significant = true ∧ AdjType.material.toDomain = Core.PropertyDomain.material ∧ Core.PropertyDomain.material.noiseDiscrimination = some RSA.Noise.materialDiscrimination ∧ ¬targetAdv_material.significant = true ∧ RSA.Noise.colorDiscrimination > RSA.Noise.materialDiscrimination

The adjective types that elicit contrastive inferences (color, scalar) both have noise discrimination > 0.40 (material's value). The adjective type that does NOT elicit contrastive inferences (material) has the lowest discrimination.

theorem Phenomena.Reference.Studies.RonderosEtAl2024.noise_ordering_matches_effect_ordering : RSA.Noise.colorDiscrimination > RSA.Noise.sizeDiscrimination ∧ RSA.Noise.sizeDiscrimination > RSA.Noise.materialDiscrimination ∧ targetAdv_color.beta > targetAdv_scalar.beta ∧ targetAdv_scalar.beta > targetAdv_material.beta

The full discrimination ordering (color > size > material) matches the contrastive effect ordering (β_color > β_scalar > β_material). Both orderings agree on which adjective types produce the strongest pragmatic effects.

theorem Phenomena.Reference.Studies.RonderosEtAl2024.converging_evidence_with_kursat_degen : ¬targetAdv_material.significant = true ∧ RSA.Noise.materialDiscrimination < RSA.Noise.sizeDiscrimination ∧ RSA.Noise.materialDiscrimination < RSA.Noise.colorDiscrimination

Connection to @cite{kursat-degen-2021}: both studies find that material adjectives produce the weakest pragmatic effects. @cite{kursat-degen-2021} shows material is used redundantly less (production); this study shows material fails to elicit contrastive inferences (comprehension). Both are predicted by low noise discrimination for material properties.

theorem Phenomena.Reference.Studies.RonderosEtAl2024.agrees_with_sedivy_on_scalar : targetAdv_scalar.significant = true ∧ SedivyEtAl1999.exp1_competitor_contrast.significant = true ∧ SedivyEtAl1999.exp2_competitor_contrast.significant = true ∧ SedivyEtAl1999.exp3_competitor_contrast.significant = true ∧ AdjType.scalar.toDomain = Core.PropertyDomain.size ∧ SedivyEtAl1999.adjDomain = Core.PropertyDomain.size ∧ Core.PropertyDomain.size.requiresComparisonClass = true

Agreement with @cite{sedivy-etal-1999}: both studies find that scalar adjectives (size domain) trigger contrastive inferences. This study replicates the core Sedivy finding cross-linguistically.

theorem Phenomena.Reference.Studies.RonderosEtAl2024.disagrees_with_sedivy_on_color : targetAdv_color.significant = true ∧ Core.PropertyDomain.color.requiresComparisonClass = false ∧ Core.PropertyDomain.color.noiseDiscrimination = some RSA.Noise.colorDiscrimination ∧ RSA.Noise.colorDiscrimination > RSA.Noise.sizeDiscrimination

Disagreement with Sedivy (2003, 2004): this study finds that color adjectives DO trigger contrastive inferences (β = 0.24, p < 0.05), while Sedivy's later work argued color adjectives are used descriptively and do not trigger contrastive interpretations.

The two-route model resolves this: @cite{sedivy-etal-1999}'s comparison-class mechanism predicts color should NOT trigger (it doesn't require comparison class), but the perceptual discrimination mechanism predicts it SHOULD trigger (color has the highest discrimination at 0.98). The disagreement suggests that perceptual salience can override the semantic-restrictiveness prediction.

theorem Phenomena.Reference.Studies.RonderosEtAl2024.two_routes : SedivyEtAl1999.adjDomain = Core.PropertyDomain.size ∧ Core.PropertyDomain.size.requiresComparisonClass = true ∧ SedivyEtAl1999.exp1_competitor_contrast.significant = true ∧ Core.PropertyDomain.color.requiresComparisonClass = false ∧ Core.PropertyDomain.color.noiseDiscrimination = some RSA.Noise.colorDiscrimination ∧ targetAdv_color.significant = true ∧ Core.PropertyDomain.material.requiresComparisonClass = false ∧ Core.PropertyDomain.material.noiseDiscrimination = some RSA.Noise.materialDiscrimination ∧ ¬targetAdv_material.significant = true ∧ RSA.Noise.colorDiscrimination > RSA.Noise.materialDiscrimination

Two independent mechanisms can drive contrastive inference:

1. Semantic restrictiveness (@cite{sedivy-etal-1999}): adjectives requiring comparison-class computation are pragmatically marked, triggering inference. Predicts: size YES, color NO, material NO. Confirmed by @cite{sedivy-etal-1999} across 3 experiments.

2. Perceptual discrimination (@cite{ronderos-etal-2024}): high discrimination provides a strong pragmatic signal, enabling inference even for non-restrictive adjectives. Predicts: color YES (high discrimination despite no comparison class), material NO (low discrimination AND no comparison class).

Together these explain the full pattern: size triggers inference via route 1, color triggers inference via route 2, material fails both.