Sentence Polarity Asymmetries #
@cite{he-kaiser-iskarous-2025}
Empirical data and domain types for modeling sentence polarity asymmetries with fuzzy interpretations in a possibly wonky world.
Key Phenomena #
Two asymmetries between positive and negative polarity:
- Cost asymmetry: Negation elicits higher production cost than positive polarity
- Presupposition asymmetry: Negation presupposes prominence of its positive counterpart in common ground, but not vice versa
Domain #
Part-whole relations (e.g., house-bathroom, classroom-stove):
- States: s_pos (A has B), s_neg (A doesn't have B)
- Utterances: u_pos, u_neg, u_null
- Costs: Cost(u_null)=0, Cost(u_pos)=1, Cost(u_neg)=2
Models #
| Model | Description |
|---|---|
| Standard RSA | Baseline with Boolean semantics |
| fuzzyRSA | Soft semantics with polarity-specific interpretation |
| wonkyRSA | Complex prior for common ground update |
| funkyRSA | Combination of fuzzy + wonky |
Equations
Instances For
Equations
- One or more equations did not get rendered due to their size.
Instances For
Equations
- One or more equations did not get rendered due to their size.
Equations
- One or more equations did not get rendered due to their size.
All states
Equations
- One or more equations did not get rendered due to their size.
Instances For
Three utterances for part-whole communication.
uPos: "A has B" (positive polarity, cost 1)uNeg: "A doesn't have B" (negative polarity, cost 2)uNull: Say nothing (cost 0)
- uPos : HKIUtterance
- uNeg : HKIUtterance
- uNull : HKIUtterance
Instances For
Equations
Instances For
Equations
- One or more equations did not get rendered due to their size.
Instances For
Equations
- One or more equations did not get rendered due to their size.
Equations
- One or more equations did not get rendered due to their size.
All utterances
Equations
- One or more equations did not get rendered due to their size.
Instances For
Equations
Instances For
Equations
- One or more equations did not get rendered due to their size.
Instances For
Get polarity of an utterance
Equations
- Phenomena.Presupposition.Studies.HeKaiserIskarous2025.HKIUtterance.uPos.polarity = Phenomena.Presupposition.Studies.HeKaiserIskarous2025.Polarity.positive
- Phenomena.Presupposition.Studies.HeKaiserIskarous2025.HKIUtterance.uNeg.polarity = Phenomena.Presupposition.Studies.HeKaiserIskarous2025.Polarity.negative
- Phenomena.Presupposition.Studies.HeKaiserIskarous2025.HKIUtterance.uNull.polarity = Phenomena.Presupposition.Studies.HeKaiserIskarous2025.Polarity.null
Instances For
Utterance costs from the paper.
Cost(u_null) = 0 Cost(u_pos) = 1 Cost(u_neg) = 2
Negation has higher cost due to marked form / longer linguistic realization.
Equations
- Phenomena.Presupposition.Studies.HeKaiserIskarous2025.utteranceCost Phenomena.Presupposition.Studies.HeKaiserIskarous2025.HKIUtterance.uNull = 0
- Phenomena.Presupposition.Studies.HeKaiserIskarous2025.utteranceCost Phenomena.Presupposition.Studies.HeKaiserIskarous2025.HKIUtterance.uPos = 1
- Phenomena.Presupposition.Studies.HeKaiserIskarous2025.utteranceCost Phenomena.Presupposition.Studies.HeKaiserIskarous2025.HKIUtterance.uNeg = 2
Instances For
Negative utterances cost more than positive (Asymmetry Hypothesis 1)
Boolean literal semantics: which utterance is true in which state.
- u_pos is true only in s_pos
- u_neg is true only in s_neg
- u_null is true in both (vacuously)
Equations
- One or more equations did not get rendered due to their size.
Instances For
Equations
Instances For
Equations
- One or more equations did not get rendered due to their size.
Instances For
All world types
Equations
- One or more equations did not get rendered due to their size.
Instances For
Prior probability over states.
In the He et al. study, priors were normed empirically for 81 part-whole pairs (e.g., house-bathroom has high prior, classroom-stove has low prior).
- p_pos : ℚ
P(s_pos) - probability of positive state
Non-negative
At most 1
Instances For
Get prior probability of a state
Equations
Instances For
Uniform prior: P(s_pos) = P(s_neg) = 0.5
Equations
- One or more equations did not get rendered due to their size.
Instances For
High prior: P(s_pos) = 0.9 (typical part, e.g., house-bathroom)
Equations
- One or more equations did not get rendered due to their size.
Instances For
Low prior: P(s_pos) = 0.1 (atypical part, e.g., classroom-stove)
Equations
- One or more equations did not get rendered due to their size.
Instances For
Parameters for fuzzy interpretation functions.
For negative utterances: constant probability n For positive utterances: sigmoid function with parameters (L, k, x0, c)
- n : ℚ
Negative interpretation: [u_neg] = n
- L : ℚ
Sigmoid maximum value
- k : ℚ
Sigmoid steepness
- x0 : ℚ
Sigmoid midpoint
- c : ℚ
Sigmoid vertical shift
n in [0,1]
Instances For
Best-fit parameters from the paper (Section 4.2).
n = 0.8, α = 1, θ = {L=0.7, k=6, x0=0.35, c=0.3}
Equations
- One or more equations did not get rendered due to their size.
Instances For
Configuration for RSA model instances
- prior : HKIPrior
Prior over states
- alpha : ℕ
Rationality parameter
- p_wonky : ℚ
Wonkiness prior P(wonky) for wonkyRSA
- fuzzyParams : FuzzyParams
Fuzzy parameters for fuzzyRSA
Instances For
Default configuration with uniform prior
Equations
Instances For
High-prior configuration (typical parts like house-bathroom)
Equations
Instances For
Low-prior configuration (atypical parts like classroom-stove)
Equations
Instances For
Asymmetry Hypothesis 1: Negation has higher production cost.
Marked forms like negation use more complex structures and longer linguistic forms, eliciting higher production cost.
Equations
- Phenomena.Presupposition.Studies.HeKaiserIskarous2025.asymmetryHypothesis1 = "Marked forms like negation elicit higher production cost than " ++ "their unmarked positive-polarity counterparts."
Instances For
Asymmetry Hypothesis 2: Negation presupposes positive prominence.
Negation presupposes that its positive-polarity counterpart is relevant or prominent in common ground, but not vice versa.
Equations
- One or more equations did not get rendered due to their size.
Instances For
Simple sigmoid approximation using rational arithmetic.
Equations
- One or more equations did not get rendered due to their size.
Instances For
Fuzzy interpretation for negative utterances. [u_neg] = n (constant) [u_neg] = 1 - n
Equations
- Phenomena.Presupposition.Studies.HeKaiserIskarous2025.fuzzyNegInterpretation n Phenomena.Presupposition.Studies.HeKaiserIskarous2025.HKIState.neg = n
- Phenomena.Presupposition.Studies.HeKaiserIskarous2025.fuzzyNegInterpretation n Phenomena.Presupposition.Studies.HeKaiserIskarous2025.HKIState.pos = 1 - n
Instances For
Fuzzy interpretation for positive utterances.
Equations
- One or more equations did not get rendered due to their size.
Instances For
Fuzzy interpretation for null utterance (no information).
Instances For
Combined fuzzy meaning function for fuzzyRSA.
Equations
- One or more equations did not get rendered due to their size.
Instances For
World-conditioned prior for wonkyRSA.
Equations
- Phenomena.Presupposition.Studies.HeKaiserIskarous2025.worldConditionedPrior cfg Phenomena.Presupposition.Studies.HeKaiserIskarous2025.WorldType.wonky x✝ = 1 / 2
- Phenomena.Presupposition.Studies.HeKaiserIskarous2025.worldConditionedPrior cfg Phenomena.Presupposition.Studies.HeKaiserIskarous2025.WorldType.normal x✝ = cfg.prior.prob x✝
Instances For
Goal projection for wonkyRSA.
Equations
- Phenomena.Presupposition.Studies.HeKaiserIskarous2025.wonkyGoalProject x✝² x✝¹ x✝ = (x✝¹ == x✝)
Instances For
Standard scenario has correct dimensions.
wonkyRSA has 2 goals (normal, wonky).
Negative utterances have higher cost in our model.
fuzzyRSA with low prior: positive utterance becomes less reliable.
Negative interpretation is constant regardless of prior.
Map He et al.'s sentence polarity to compositional context polarity.
Equations
- Phenomena.Presupposition.Studies.HeKaiserIskarous2025.toContextPolarity Phenomena.Presupposition.Studies.HeKaiserIskarous2025.Polarity.positive = Core.NaturalLogic.ContextPolarity.upward
- Phenomena.Presupposition.Studies.HeKaiserIskarous2025.toContextPolarity Phenomena.Presupposition.Studies.HeKaiserIskarous2025.Polarity.negative = Core.NaturalLogic.ContextPolarity.downward
- Phenomena.Presupposition.Studies.HeKaiserIskarous2025.toContextPolarity Phenomena.Presupposition.Studies.HeKaiserIskarous2025.Polarity.null = Core.NaturalLogic.ContextPolarity.upward
Instances For
Cost aligns with UE/DE distinction: DE costs more.
Equations
Instances For
Equations
- One or more equations did not get rendered due to their size.
Instances For
Part-whole model.
Equations
Instances For
The "has" relation: which containers have which parts.
Equations
- One or more equations did not get rendered due to their size.
- Phenomena.Presupposition.Studies.HeKaiserIskarous2025.has_sem part container = false
Instances For
Positive sentence meaning: "A has B".
Equations
- Phenomena.Presupposition.Studies.HeKaiserIskarous2025.posMeaning container part = Phenomena.Presupposition.Studies.HeKaiserIskarous2025.has_sem part container
Instances For
Negative sentence meaning: "A doesn't have B" = neg(has(A, B)).
Equations
Instances For
Key theorem: negative meaning is compositionally derived via neg.
Lift He et al. sentences to world-indexed propositions.
Equations
Instances For
Negative sentence meaning as world-indexed proposition. ⟦"A doesn't have B"⟧ = pnot(⟦"A has B"⟧)
Equations
- One or more equations did not get rendered due to their size.
Instances For
The grounded polarity from Entailment.Polarity (uses pnot).
Equations
Instances For
Negative sentences have DE polarity (from Montague's proof).
Positive sentences have UE polarity (identity = no negation).
Structural complexity: count of functional heads in the derivation.
Equations
- Phenomena.Presupposition.Studies.HeKaiserIskarous2025.structuralComplexity Phenomena.Presupposition.Studies.HeKaiserIskarous2025.HKIUtterance.uNull = 0
- Phenomena.Presupposition.Studies.HeKaiserIskarous2025.structuralComplexity Phenomena.Presupposition.Studies.HeKaiserIskarous2025.HKIUtterance.uPos = 1
- Phenomena.Presupposition.Studies.HeKaiserIskarous2025.structuralComplexity Phenomena.Presupposition.Studies.HeKaiserIskarous2025.HKIUtterance.uNeg = 2
Instances For
Utterance cost equals structural complexity.
Negation adds exactly one unit of complexity (the Neg functional head).