Prasertsom, Smith & Culbertson (2026)

@cite{prasertsom-smith-culbertson-2026}

Domain-general categorisation explains constrained cross-linguistic variation in noun classification. Cognition 271, 106411.

Key Claims

@cite{prasertsom-smith-culbertson-2026} ask why animacy — but not colour — appears in noun classification systems, despite colour being perceptually salient. They argue the answer lies in a domain-general categorisation principle: features with higher predictive power (the ability to predict other features of category members) yield more compact, more distinct categories, which are easier to learn. Animacy is more predictive than colour, explaining its cross-linguistic prevalence via cultural transmission amplification (@cite{griffiths-kalish-2007}, @cite{kirby-et-al-2007}, @cite{smith-2011}).

Argument Chain

1. Typological observation (§1): Animacy is universal in noun categorization (@cite{aikhenvald-2000}); colour is absent from every attested system — despite high perceptual salience.
2. Learning bias (Exp 1a–b): Participants learn animacy-based artificial noun classes better than colour-based ones.
3. Domain generality (Exp 2a–c): Non-linguistic sorting also shows animacy preference, ruling out grammar-specific mechanisms.
4. Mechanism (§4.1, Table 6): Predictive power — animacy predicts other features (shape, horn, appendages) better than colour does, yielding more compact categories (Table 2) and higher classifier accuracy (Table 3).
5. Causal evidence (Exp 3a–b): Manipulating predictive power modulates sorting preferences, though the effect on noun class learning is indirect — participants who notice colour's predictive structure and sort by colour learn animacy-based noun classes worse.

Formalization

The typological observation (§1) connects to existing NounCategorization infrastructure. Predictive power is formalized via Table 6's conditional entropy data from the experimental stimuli. The connection to Core.Prominence.AnimacyLevel — the Silverstein hierarchy — makes explicit that animacy's grammatical privilege spans both case marking and noun classification.

def Phenomena.Agreement.Studies.PrasertsonSmithCulbertson2026.isAttestedInTypology (p : Core.NounCategorization.SemanticParameter) : Bool

Whether a semantic parameter is attested in any system in our typology. Derived from allSystems data rather than stipulated.

Equations

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

theorem Phenomena.Agreement.Studies.PrasertsonSmithCulbertson2026.animacy_attested : isAttestedInTypology Core.NounCategorization.SemanticParameter.animacy = true

Animacy is attested in the typological data.

theorem Phenomena.Agreement.Studies.PrasertsonSmithCulbertson2026.humanness_attested : isAttestedInTypology Core.NounCategorization.SemanticParameter.humanness = true

Humanness is attested in the typological data.

theorem Phenomena.Agreement.Studies.PrasertsonSmithCulbertson2026.shape_attested : isAttestedInTypology Core.NounCategorization.SemanticParameter.shape = true

Shape is attested in the typological data.

theorem Phenomena.Agreement.Studies.PrasertsonSmithCulbertson2026.colour_not_attested : isAttestedInTypology Core.NounCategorization.SemanticParameter.colour = false

Colour is NOT attested in any system in our typology.

theorem Phenomena.Agreement.Studies.PrasertsonSmithCulbertson2026.animacy_colour_asymmetry : isAttestedInTypology Core.NounCategorization.SemanticParameter.animacy = true ∧ isAttestedInTypology Core.NounCategorization.SemanticParameter.colour = false

The central asymmetry: animacy is attested, colour is not.

inductive Phenomena.Agreement.Studies.PrasertsonSmithCulbertson2026.StimulusFeature : Type

Features in the experimental stimuli (Fig. 10, Tables 4–5).

• animacy : StimulusFeature
• colour : StimulusFeature
• horn : StimulusFeature
• shape_ : StimulusFeature
• appendages : StimulusFeature

instance Phenomena.Agreement.Studies.PrasertsonSmithCulbertson2026.instDecidableEqStimulusFeature : DecidableEq StimulusFeature

Equations

• Phenomena.Agreement.Studies.PrasertsonSmithCulbertson2026.instDecidableEqStimulusFeature x✝ y✝ = if h : x✝.ctorIdx = y✝.ctorIdx then isTrue ⋯ else isFalse ⋯

def Phenomena.Agreement.Studies.PrasertsonSmithCulbertson2026.instBEqStimulusFeature.beq : StimulusFeature → StimulusFeature → Bool

Equations

• Phenomena.Agreement.Studies.PrasertsonSmithCulbertson2026.instBEqStimulusFeature.beq x✝ y✝ = (x✝.ctorIdx == y✝.ctorIdx)

instance Phenomena.Agreement.Studies.PrasertsonSmithCulbertson2026.instBEqStimulusFeature : BEq StimulusFeature

Equations

• Phenomena.Agreement.Studies.PrasertsonSmithCulbertson2026.instBEqStimulusFeature = { beq := Phenomena.Agreement.Studies.PrasertsonSmithCulbertson2026.instBEqStimulusFeature.beq }

def Phenomena.Agreement.Studies.PrasertsonSmithCulbertson2026.instReprStimulusFeature.repr : StimulusFeature → ℕ → Std.Format

Equations

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

instance Phenomena.Agreement.Studies.PrasertsonSmithCulbertson2026.instReprStimulusFeature : Repr StimulusFeature

Equations

• Phenomena.Agreement.Studies.PrasertsonSmithCulbertson2026.instReprStimulusFeature = { reprPrec := Phenomena.Agreement.Studies.PrasertsonSmithCulbertson2026.instReprStimulusFeature.repr }

structure Phenomena.Agreement.Studies.PrasertsonSmithCulbertson2026.ConditionalEntropy : Type

Conditional entropy H(row|col) from Table 6. Lower values mean the column feature is more predictive of the row feature. Values are in bits; the maximum for a binary feature is 1.0, for a 4-valued feature is 2.0.

• row : StimulusFeature
• col : StimulusFeature
• hBits_x1000 : ℕ

  H(row|col) in bits, scaled ×1000 for exact rational arithmetic

def Phenomena.Agreement.Studies.PrasertsonSmithCulbertson2026.instReprConditionalEntropy.repr : ConditionalEntropy → ℕ → Std.Format

Equations

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

instance Phenomena.Agreement.Studies.PrasertsonSmithCulbertson2026.instReprConditionalEntropy : Repr ConditionalEntropy

Equations

• Phenomena.Agreement.Studies.PrasertsonSmithCulbertson2026.instReprConditionalEntropy = { reprPrec := Phenomena.Agreement.Studies.PrasertsonSmithCulbertson2026.instReprConditionalEntropy.repr }

def Phenomena.Agreement.Studies.PrasertsonSmithCulbertson2026.table6_horn_given_animacy : ConditionalEntropy

Table 6 data, predictive-animacy stimulus set. In this set, animacy is the predictive feature: knowing animacy reduces uncertainty about horn/shape/appendages more than knowing colour does. H(Horn|Animacy) = 0.406 < H(Horn|Colour) = 0.906.

Equations

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

def Phenomena.Agreement.Studies.PrasertsonSmithCulbertson2026.table6_horn_given_colour : ConditionalEntropy

Equations

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

def Phenomena.Agreement.Studies.PrasertsonSmithCulbertson2026.table6_shape_given_animacy : ConditionalEntropy

Equations

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

def Phenomena.Agreement.Studies.PrasertsonSmithCulbertson2026.table6_shape_given_colour : ConditionalEntropy

Equations

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

def Phenomena.Agreement.Studies.PrasertsonSmithCulbertson2026.table6_appendages_given_animacy : ConditionalEntropy

Equations

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

def Phenomena.Agreement.Studies.PrasertsonSmithCulbertson2026.table6_appendages_given_colour : ConditionalEntropy

Equations

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

theorem Phenomena.Agreement.Studies.PrasertsonSmithCulbertson2026.animacy_predicts_horn_better : table6_horn_given_animacy.hBits_x1000 < table6_horn_given_colour.hBits_x1000

Table 6: animacy is more predictive of horn than colour is. H(horn|animacy) = 0.406 < H(horn|colour) = 0.906.

theorem Phenomena.Agreement.Studies.PrasertsonSmithCulbertson2026.animacy_predicts_shape_better : table6_shape_given_animacy.hBits_x1000 < table6_shape_given_colour.hBits_x1000

Table 6: animacy is more predictive of shape than colour is.

theorem Phenomena.Agreement.Studies.PrasertsonSmithCulbertson2026.animacy_predicts_appendages_better : table6_appendages_given_animacy.hBits_x1000 < table6_appendages_given_colour.hBits_x1000

Table 6: animacy is more predictive of appendages than colour is.

def Phenomena.Agreement.Studies.PrasertsonSmithCulbertson2026.table6_animacy_given_horn : ConditionalEntropy

Table 6, "Both stimulus sets" column (reverse direction): Horn/shape/appendages predict animacy better than they predict colour. H(Animacy|Horn) = 1.451 < H(Colour|Horn) = 1.951. This is consistent: predictive power is symmetric — if animacy predicts horn well, horn also predicts animacy well.

Equations

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

def Phenomena.Agreement.Studies.PrasertsonSmithCulbertson2026.table6_colour_given_horn : ConditionalEntropy

Equations

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

theorem Phenomena.Agreement.Studies.PrasertsonSmithCulbertson2026.horn_predicts_animacy_better_than_colour : table6_animacy_given_horn.hBits_x1000 < table6_colour_given_horn.hBits_x1000

inductive Phenomena.Agreement.Studies.PrasertsonSmithCulbertson2026.SemanticBasis : Type

Semantic basis for categorization (both experimental and typological).

• animacy : SemanticBasis
• colour : SemanticBasis

instance Phenomena.Agreement.Studies.PrasertsonSmithCulbertson2026.instDecidableEqSemanticBasis : DecidableEq SemanticBasis

Equations

• Phenomena.Agreement.Studies.PrasertsonSmithCulbertson2026.instDecidableEqSemanticBasis x✝ y✝ = if h : x✝.ctorIdx = y✝.ctorIdx then isTrue ⋯ else isFalse ⋯

def Phenomena.Agreement.Studies.PrasertsonSmithCulbertson2026.instBEqSemanticBasis.beq : SemanticBasis → SemanticBasis → Bool

Equations

• Phenomena.Agreement.Studies.PrasertsonSmithCulbertson2026.instBEqSemanticBasis.beq x✝ y✝ = (x✝.ctorIdx == y✝.ctorIdx)

instance Phenomena.Agreement.Studies.PrasertsonSmithCulbertson2026.instBEqSemanticBasis : BEq SemanticBasis

Equations

• Phenomena.Agreement.Studies.PrasertsonSmithCulbertson2026.instBEqSemanticBasis = { beq := Phenomena.Agreement.Studies.PrasertsonSmithCulbertson2026.instBEqSemanticBasis.beq }

instance Phenomena.Agreement.Studies.PrasertsonSmithCulbertson2026.instReprSemanticBasis : Repr SemanticBasis

Equations

• Phenomena.Agreement.Studies.PrasertsonSmithCulbertson2026.instReprSemanticBasis = { reprPrec := Phenomena.Agreement.Studies.PrasertsonSmithCulbertson2026.instReprSemanticBasis.repr }

def Phenomena.Agreement.Studies.PrasertsonSmithCulbertson2026.instReprSemanticBasis.repr : SemanticBasis → ℕ → Std.Format

Equations

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

structure Phenomena.Agreement.Studies.PrasertsonSmithCulbertson2026.CategoryCompactness : Type

Intra-category similarity from Word2Vec embeddings of 472 frequent physical nouns from CHILDES (§4.1.1). Higher InSim means category members are more similar to each other in distributional semantic space.

• basis : SemanticBasis
• category : String
• inSim_x1000 : ℕ

  Mean cosine similarity × 1000 (for exact arithmetic)

instance Phenomena.Agreement.Studies.PrasertsonSmithCulbertson2026.instReprCategoryCompactness : Repr CategoryCompactness

Equations

• Phenomena.Agreement.Studies.PrasertsonSmithCulbertson2026.instReprCategoryCompactness = { reprPrec := Phenomena.Agreement.Studies.PrasertsonSmithCulbertson2026.instReprCategoryCompactness.repr }

def Phenomena.Agreement.Studies.PrasertsonSmithCulbertson2026.instReprCategoryCompactness.repr : CategoryCompactness → ℕ → Std.Format

Equations

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

def Phenomena.Agreement.Studies.PrasertsonSmithCulbertson2026.compactness_animate : CategoryCompactness

Equations

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

def Phenomena.Agreement.Studies.PrasertsonSmithCulbertson2026.compactness_inanimate : CategoryCompactness

Equations

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

def Phenomena.Agreement.Studies.PrasertsonSmithCulbertson2026.compactness_warm : CategoryCompactness

Equations

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

def Phenomena.Agreement.Studies.PrasertsonSmithCulbertson2026.compactness_cool : CategoryCompactness

Equations

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

def Phenomena.Agreement.Studies.PrasertsonSmithCulbertson2026.inSimAvg_animacy_x1000 : ℕ

InSim(Avg) × 1000: weighted average across categories. Animacy: 0.149, Colour: 0.147.

Equations

• Phenomena.Agreement.Studies.PrasertsonSmithCulbertson2026.inSimAvg_animacy_x1000 = 149

def Phenomena.Agreement.Studies.PrasertsonSmithCulbertson2026.inSimAvg_colour_x1000 : ℕ

Equations

• Phenomena.Agreement.Studies.PrasertsonSmithCulbertson2026.inSimAvg_colour_x1000 = 147

theorem Phenomena.Agreement.Studies.PrasertsonSmithCulbertson2026.animacy_categories_more_compact : inSimAvg_animacy_x1000 > inSimAvg_colour_x1000

Animacy-based categories are more compact on average (Table 2). InSim(Avg) animacy (0.149) > InSim(Avg) colour (0.147).

theorem Phenomena.Agreement.Studies.PrasertsonSmithCulbertson2026.animate_most_compact : compactness_animate.inSim_x1000 > compactness_inanimate.inSim_x1000 ∧ compactness_animate.inSim_x1000 > compactness_warm.inSim_x1000 ∧ compactness_animate.inSim_x1000 > compactness_cool.inSim_x1000

The animate category is the most compact single category.

structure Phenomena.Agreement.Studies.PrasertsonSmithCulbertson2026.ClassifierAccuracy : Type

5-fold cross-validated logistic classifier accuracy from normalised Word2Vec embeddings (§4.1.2, Table 3).

• basis : SemanticBasis
• accuracy_x1000 : ℕ

  Mean accuracy × 1000

def Phenomena.Agreement.Studies.PrasertsonSmithCulbertson2026.instReprClassifierAccuracy.repr : ClassifierAccuracy → ℕ → Std.Format

Equations

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

instance Phenomena.Agreement.Studies.PrasertsonSmithCulbertson2026.instReprClassifierAccuracy : Repr ClassifierAccuracy

Equations

• Phenomena.Agreement.Studies.PrasertsonSmithCulbertson2026.instReprClassifierAccuracy = { reprPrec := Phenomena.Agreement.Studies.PrasertsonSmithCulbertson2026.instReprClassifierAccuracy.repr }

def Phenomena.Agreement.Studies.PrasertsonSmithCulbertson2026.logistic_animacy : ClassifierAccuracy

Equations

• Phenomena.Agreement.Studies.PrasertsonSmithCulbertson2026.logistic_animacy = { basis := Phenomena.Agreement.Studies.PrasertsonSmithCulbertson2026.SemanticBasis.animacy, accuracy_x1000 := 858 }

def Phenomena.Agreement.Studies.PrasertsonSmithCulbertson2026.logistic_colour : ClassifierAccuracy

Equations

• Phenomena.Agreement.Studies.PrasertsonSmithCulbertson2026.logistic_colour = { basis := Phenomena.Agreement.Studies.PrasertsonSmithCulbertson2026.SemanticBasis.colour, accuracy_x1000 := 742 }

theorem Phenomena.Agreement.Studies.PrasertsonSmithCulbertson2026.animacy_classifier_more_accurate : logistic_animacy.accuracy_x1000 > logistic_colour.accuracy_x1000

Animacy categories are easier for a classifier to learn (Table 3).

structure Phenomena.Agreement.Studies.PrasertsonSmithCulbertson2026.LearningResult : Type

Result of a noun class learning experiment.

• basis : SemanticBasis
• accuracy_x1000 : ℕ

  Mean proportion correct × 1000

instance Phenomena.Agreement.Studies.PrasertsonSmithCulbertson2026.instReprLearningResult : Repr LearningResult

Equations

• Phenomena.Agreement.Studies.PrasertsonSmithCulbertson2026.instReprLearningResult = { reprPrec := Phenomena.Agreement.Studies.PrasertsonSmithCulbertson2026.instReprLearningResult.repr }

def Phenomena.Agreement.Studies.PrasertsonSmithCulbertson2026.instReprLearningResult.repr : LearningResult → ℕ → Std.Format

Equations

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

def Phenomena.Agreement.Studies.PrasertsonSmithCulbertson2026.exp1a_animacy : LearningResult

Experiment 1a: Ease of learning (§2.1). Animacy condition: 0.915 mean accuracy. Colour condition: 0.841 mean accuracy. β_SemanticBasis = 1.859, p = 0.023.

Equations

• Phenomena.Agreement.Studies.PrasertsonSmithCulbertson2026.exp1a_animacy = { basis := Phenomena.Agreement.Studies.PrasertsonSmithCulbertson2026.SemanticBasis.animacy, accuracy_x1000 := 915 }

def Phenomena.Agreement.Studies.PrasertsonSmithCulbertson2026.exp1a_colour : LearningResult

Equations

• Phenomena.Agreement.Studies.PrasertsonSmithCulbertson2026.exp1a_colour = { basis := Phenomena.Agreement.Studies.PrasertsonSmithCulbertson2026.SemanticBasis.colour, accuracy_x1000 := 841 }

theorem Phenomena.Agreement.Studies.PrasertsonSmithCulbertson2026.exp1a_animacy_learned_better : exp1a_animacy.accuracy_x1000 > exp1a_colour.accuracy_x1000

def Phenomena.Agreement.Studies.PrasertsonSmithCulbertson2026.exp1b_animacy_generalizers : ℕ

Experiment 1b: Extrapolation from ambiguous data (§2.2). When animacy and colour are confounded in training, 77.5% of participants (62/80) generalise to animacy-based classification at test.

Equations

• Phenomena.Agreement.Studies.PrasertsonSmithCulbertson2026.exp1b_animacy_generalizers = 62

def Phenomena.Agreement.Studies.PrasertsonSmithCulbertson2026.exp1b_total_participants : ℕ

Equations

• Phenomena.Agreement.Studies.PrasertsonSmithCulbertson2026.exp1b_total_participants = 80

theorem Phenomena.Agreement.Studies.PrasertsonSmithCulbertson2026.exp1b_above_chance : exp1b_animacy_generalizers * 2 > exp1b_total_participants

structure Phenomena.Agreement.Studies.PrasertsonSmithCulbertson2026.SortingResult : Type

Result of an image sorting experiment (§3). Participants sort 16 images into two groups; sorting strategy is inferred from AMI with reference sorts. Proportions are expressed as parts per thousand.

• animacy_x1000 : ℕ

  Proportion ×1000 who sorted by animacy

• colour_x1000 : ℕ

  Proportion ×1000 who sorted by colour

• other_x1000 : ℕ

  Proportion ×1000 who sorted by another strategy

instance Phenomena.Agreement.Studies.PrasertsonSmithCulbertson2026.instReprSortingResult : Repr SortingResult

Equations

• Phenomena.Agreement.Studies.PrasertsonSmithCulbertson2026.instReprSortingResult = { reprPrec := Phenomena.Agreement.Studies.PrasertsonSmithCulbertson2026.instReprSortingResult.repr }

def Phenomena.Agreement.Studies.PrasertsonSmithCulbertson2026.instReprSortingResult.repr : SortingResult → ℕ → Std.Format

Equations

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

def Phenomena.Agreement.Studies.PrasertsonSmithCulbertson2026.exp2a : SortingResult

Exp 2a: original stimuli (same as Exp 1).

Equations

• Phenomena.Agreement.Studies.PrasertsonSmithCulbertson2026.exp2a = { animacy_x1000 := 875, colour_x1000 := 25, other_x1000 := 100 }

def Phenomena.Agreement.Studies.PrasertsonSmithCulbertson2026.exp2b : SortingResult

Exp 2b: increased within-category colour similarity.

Equations

• Phenomena.Agreement.Studies.PrasertsonSmithCulbertson2026.exp2b = { animacy_x1000 := 683, colour_x1000 := 150, other_x1000 := 167 }

def Phenomena.Agreement.Studies.PrasertsonSmithCulbertson2026.exp2c : SortingResult

Exp 2c: further decreased within-category animacy similarity.

Equations

• Phenomena.Agreement.Studies.PrasertsonSmithCulbertson2026.exp2c = { animacy_x1000 := 617, colour_x1000 := 183, other_x1000 := 200 }

theorem Phenomena.Agreement.Studies.PrasertsonSmithCulbertson2026.exp2a_animacy_preferred : exp2a.animacy_x1000 > exp2a.colour_x1000

theorem Phenomena.Agreement.Studies.PrasertsonSmithCulbertson2026.exp2b_animacy_preferred : exp2b.animacy_x1000 > exp2b.colour_x1000

theorem Phenomena.Agreement.Studies.PrasertsonSmithCulbertson2026.exp2c_animacy_preferred : exp2c.animacy_x1000 > exp2c.colour_x1000

theorem Phenomena.Agreement.Studies.PrasertsonSmithCulbertson2026.sorting_bias_robust : exp2a.animacy_x1000 > exp2a.colour_x1000 ∧ exp2b.animacy_x1000 > exp2b.colour_x1000 ∧ exp2c.animacy_x1000 > exp2c.colour_x1000

The animacy bias persists even when visual similarity is manipulated to favour colour-based sorting (Exp 2b, 2c).

theorem Phenomena.Agreement.Studies.PrasertsonSmithCulbertson2026.bias_weakens_with_manipulation : exp2a.animacy_x1000 > exp2b.animacy_x1000 ∧ exp2b.animacy_x1000 > exp2c.animacy_x1000

The bias weakens as colour similarity increases — evidence that it is not absolute but can be modulated.

structure Phenomena.Agreement.Studies.PrasertsonSmithCulbertson2026.Exp3Result : Type

Experiment 3a results (§4.2): predictive power manipulation had a main effect of semantic basis (β = 1.251, p = 0.003) but the interaction with predictive feature was NOT significant (β = 0.276, p = 0.519). The prediction was "not straightforwardly confirmed."

However, the sorting strategy × semantic basis interaction WAS significant (β = 1.058, p < 0.001): participants who sorted by animacy learned animacy-based noun classes better.

• animacyBasis_x1000 : ℕ

  Accuracy ×1000 for animacy-basis condition

• colourBasis_x1000 : ℕ

  Accuracy ×1000 for colour-basis condition

instance Phenomena.Agreement.Studies.PrasertsonSmithCulbertson2026.instReprExp3Result : Repr Exp3Result

Equations

• Phenomena.Agreement.Studies.PrasertsonSmithCulbertson2026.instReprExp3Result = { reprPrec := Phenomena.Agreement.Studies.PrasertsonSmithCulbertson2026.instReprExp3Result.repr }

def Phenomena.Agreement.Studies.PrasertsonSmithCulbertson2026.instReprExp3Result.repr : Exp3Result → ℕ → Std.Format

Equations

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

def Phenomena.Agreement.Studies.PrasertsonSmithCulbertson2026.exp3b_animacyBasis_x1000 : ℕ

Exp 3b (§4.3): among participants who sorted by colour (successful manipulation), there was NO significant difference between animacy-based and colour-based noun class learning (0.800 vs 0.803). β = −0.510, p = 0.314.

Equations

• Phenomena.Agreement.Studies.PrasertsonSmithCulbertson2026.exp3b_animacyBasis_x1000 = 800

def Phenomena.Agreement.Studies.PrasertsonSmithCulbertson2026.exp3b_colourBasis_x1000 : ℕ

Equations

• Phenomena.Agreement.Studies.PrasertsonSmithCulbertson2026.exp3b_colourBasis_x1000 = 803

theorem Phenomena.Agreement.Studies.PrasertsonSmithCulbertson2026.exp3b_no_animacy_advantage : exp3b_colourBasis_x1000 ≥ exp3b_animacyBasis_x1000

Exp 3b: when the animacy bias is neutralised by selecting colour-sorters, the learning advantage for animacy disappears. This is consistent with the domain-general account — the bias is in categorisation, not in grammar.

theorem Phenomena.Agreement.Studies.PrasertsonSmithCulbertson2026.prominence_encodes_categorization_parameters : Core.Prominence.AnimacyLevel.human.rank > Core.Prominence.AnimacyLevel.animate.rank ∧ Core.Prominence.AnimacyLevel.animate.rank > Core.Prominence.AnimacyLevel.inanimate.rank ∧ isAttestedInTypology Core.NounCategorization.SemanticParameter.animacy = true ∧ isAttestedInTypology Core.NounCategorization.SemanticParameter.humanness = true

The animacy hierarchy that governs noun classification (@cite{aikhenvald-2000}, @cite{silverstein-1976}) is the same hierarchy that governs differential argument marking (Core.Prominence).

@cite{prasertsom-smith-culbertson-2026} provide a domain-general explanation for WHY animacy is grammatically privileged in both domains: animacy is highly predictive of other referent features, making it a good basis for categorization in general.

This theorem connects the prominence hierarchy to the noun categorization typology: the levels of AnimacyLevel correspond to the SemanticParameter.animacy / .humanness distinction.

theorem Phenomena.Agreement.Studies.PrasertsonSmithCulbertson2026.argument_chain : isAttestedInTypology Core.NounCategorization.SemanticParameter.colour = false ∧ isAttestedInTypology Core.NounCategorization.SemanticParameter.animacy = true ∧ table6_horn_given_animacy.hBits_x1000 < table6_horn_given_colour.hBits_x1000 ∧ inSimAvg_animacy_x1000 > inSimAvg_colour_x1000 ∧ logistic_animacy.accuracy_x1000 > logistic_colour.accuracy_x1000 ∧ exp1a_animacy.accuracy_x1000 > exp1a_colour.accuracy_x1000 ∧ exp2a.animacy_x1000 > exp2a.colour_x1000

The paper's core argument, assembled from the evidence above:

1. Colour is absent from all attested noun categorization systems
2. Animacy is present in all attested systems
3. Animacy is more predictive of other features than colour (Table 6)
4. Animacy-based categories are more compact in distributional space (Table 2)
5. Animacy-based categories are easier for classifiers to learn (Table 3)
6. Humans learn animacy-based noun classes better (Exp 1a)
7. Humans prefer animacy-based sorting even non-linguistically (Exp 2a–c)

The connection is: predictive power → compactness → learnability → typological prevalence (via cultural transmission).