Docs / Pragmatics / Probabilistic / QUD

Speakers who address a question

import Linglib.Pragmatics.Probabilistic.QUD · namespace RSA

A question under discussion is a function q : W → X from states to answers. In the models of Kao and colleagues the speaker knows the answer q w at her state and is informative about the answer alone: she addresses the pushforward L.map q of the listener's belief about the state. At the identity question she is the speaker who knows the state, at a constant one she only weighs costs, and a coarser question lets her say more. Under a literal listener she produces an utterance exactly when it is true at a state of positive prior with her answer, whether or not it is true at hers. A listener unsure of the question holds it as a latent coordinate of the state, as in lexical uncertainty, and such an utterance leads her to states where it is false, which is how the models derive hyperbole, metaphor and irony. The question is sufficient for the listener: she learns the speaker's answer, and given the answer, or the question and the answer when she is unsure of the question, she believes what the prior does. In these models an utterance names one coordinate of the state, and at a prior that factors through that coordinate the literal listener does not depend on its marginal.

References

Declarations

L46–52

The speaker

The speaker who addresses the question q. At the state w she knows the answer q w and soft-maximizes the log of the listener's probability of that answer, less the cost.

qudSpeakerL58–62
noncomputable def RSA.qudSpeaker {W : Type u_1} {U : Type u_2} {X : Type u_3} [MeasurableSpace W] [MeasurableSpace U] [MeasurableSpace X] [MeasurableSingletonClass X] [Fintype U] [Fintype X] (α : ℝ) (C : U → ℝ) (L : ProbabilityTheory.Kernel U W) (q : W → X) (w : W) :
MeasureTheory.Measure U
  • qudSpeaker α C L q w = (speaker α C (⇑(L.map q)) MeasureTheory.Measure.dirac) (q w)

Used by instIsZeroOrProbabilityMeasureQudSpeaker, qudSpeaker_eq_of_eq, qudSpeaker_id, qudSpeaker_const, qudSpeaker_apply_singleton_ne_zero_of_coarser, qudSpeaker_literalListener_apply_singleton_ne_zero_iff, listener_prod_fst_qudSpeaker_literalListener_apply_singleton_ne_zero_iff, cond_listener_qudSpeaker_preimage, cond_listener_prod_qudSpeaker_preimage

Used in KaoBergenEtAl2014, KaoGoodman2015, KaoWuEtAl2014, ScontrasPearl2021, SpinosoDiPianoEtAl2025

L68–69
L73–75

At the identity question she is the speaker who knows the state.

theorem qudSpeaker_idL78–81
{U : Type u_2} {X : Type u_3} [MeasurableSpace U] [MeasurableSpace X] [MeasurableSingletonClass X] [Fintype U] [Fintype X] {α : ℝ} (C : U → ℝ) (L : ProbabilityTheory.Kernel U X) (x : X) :
qudSpeaker α C L id x = (speaker α C (⇑L) MeasureTheory.Measure.dirac) x

Built from

At a constant question the answer tells the listener nothing and she only weighs costs.

theorem qudSpeaker_constL83–90
{W : Type u_1} {U : Type u_2} {X : Type u_3} [MeasurableSpace W] [MeasurableSpace U] [MeasurableSingletonClass U] [MeasurableSpace X] [MeasurableSingletonClass X] [Fintype U] [Fintype X] {α : ℝ} (C : U → ℝ) (L : ProbabilityTheory.Kernel U W) [Countable W] [MeasurableSingletonClass W] [ProbabilityTheory.IsMarkovKernel L] (x : X) (w : W) (u : U) :
(qudSpeaker α C L (fun (x_1 : W) => x) w) {u} = ENNReal.ofReal (Real.exp (-(α * C u))) / ∑ u' : U, ENNReal.ofReal (Real.exp (-(α * C u')))

Built from

A coarser question lets her say more: an utterance she can produce addressing a question she can produce addressing any question that identifies at least the states it does.

theorem qudSpeaker_apply_singleton_ne_zero_of_coarserL93–101
{W : Type u_1} {U : Type u_2} {X : Type u_3} [MeasurableSpace W] [MeasurableSpace U] [MeasurableSingletonClass U] [MeasurableSpace X] [MeasurableSingletonClass X] [Fintype U] [Fintype X] {α : ℝ} {C : U → ℝ} {L : ProbabilityTheory.Kernel U W} {q : W → X} {X' : Type u_4} [MeasurableSpace X'] [MeasurableSingletonClass X'] [Fintype X'] [Countable W] [MeasurableSingletonClass W] [ProbabilityTheory.IsFiniteKernel L] (hα : 0 < α) {q' : W → X'} {w : W} (h : ∀ (w' : W), q w' = q w → q' w' = q' w) {u : U} (hu : (qudSpeaker α C L q w) {u} ≠ 0) :
(qudSpeaker α C L q' w) {u} ≠ 0

Built from

Nonliteral production. The speaker who addresses a question to a literal listener produces an utterance exactly when it is true at a state of positive prior with the same answer as hers, whether or not it is true at her own state.

theorem qudSpeaker_literalListener_apply_singleton_ne_zero_iffL105–114
{W : Type u_1} {U : Type u_2} {X : Type u_3} [MeasurableSpace W] [MeasurableSpace U] [MeasurableSingletonClass U] [MeasurableSpace X] [MeasurableSingletonClass X] [Fintype U] [Fintype X] {α : ℝ} [Fintype W] [DiscreteMeasurableSpace W] (hα : 0 < α) (C : U → ℝ) (μ : MeasureTheory.Measure W) [MeasureTheory.IsFiniteMeasure μ] (sem : U → Set W) (q : W → X) (w : W) (u : U) :
(qudSpeaker α C (literalListener μ sem) q w) {u} ≠ 0 ↔ ∃ (w' : W), q w' = q w ∧ w' ∈ sem u ∧ μ {w'} ≠ 0

Built from

The listener

L118–129

A question is sufficient for the listener. Against a speaker who addresses a question, the listener who learns the speaker's answer, at a state where the speaker can produce the utterance, holds the prior conditioned on that answer.

theorem cond_listener_qudSpeaker_preimageL131–142
{W : Type u_1} {U : Type u_2} {X : Type u_3} [MeasurableSpace W] [MeasurableSpace U] [MeasurableSingletonClass U] [MeasurableSpace X] [MeasurableSingletonClass X] [Fintype U] [Fintype X] {α : ℝ} [DiscreteMeasurableSpace W] (C : U → ℝ) (L : ProbabilityTheory.Kernel U W) (μ : MeasureTheory.Measure W) [MeasureTheory.IsFiniteMeasure μ] (q : W → X) {u : U} {w : W} (h : (qudSpeaker α C L q w) {u} ≠ 0) :
((listener μ (qudSpeaker α C L q)) u)[|q ⁻¹' {q w}] = μ[|q ⁻¹' {q w}]

Built from

A listener unsure of the question who learns the question and the speaker's answer to it, at a state where the speaker can produce the utterance, holds the prior conditioned on them.

theorem cond_listener_prod_qudSpeaker_preimageL144–159
{W : Type u_1} {U : Type u_2} {X : Type u_3} [MeasurableSpace W] [MeasurableSpace U] [MeasurableSingletonClass U] [MeasurableSpace X] [MeasurableSingletonClass X] [Fintype U] [Fintype X] {α : ℝ} {B : Type u_4} [MeasurableSpace B] [MeasurableSingletonClass B] [DiscreteMeasurableSpace W] (C : U → ℝ) (L : ProbabilityTheory.Kernel U W) (μ : MeasureTheory.Measure W) [MeasureTheory.IsFiniteMeasure μ] (ν : MeasureTheory.Measure B) [MeasureTheory.IsFiniteMeasure ν] (q : B → W → X) {u : U} {b : B} {w : W} (h : (qudSpeaker α C L (q b) w) {u} ≠ 0) :
((listener (μ.prod ν) fun (p : W × B) => qudSpeaker α C L (q p.2) p.1) u)[|(q b ⁻¹' {q b w}) ×ˢ {b}] = (μ.prod ν)[|(q b ⁻¹' {q b w}) ×ˢ {b}]

Built from