Documentation

Linglib.Phenomena.Modality.Studies.ChampollionAlsopGrosu2019

@cite{champollion-alsop-grosu-2019} — Free Choice Disjunction as RSA #

@cite{champollion-alsop-grosu-2019} @cite{bergen-levy-goodman-2016} @cite{fox-2007} @cite{franke-2011}"Free choice disjunction as a rational speech act" Proceedings of SALT 29: 238-257.

The Model #

Domain: "You may take an apple or a pear" with 2 items {A, B}. 5 states based on permission structure. 4 utterances. 2 interpretation functions (I₁ literal vs I₂ exhaustified), following @cite{bergen-levy-goodman-2016}.

Parameters: α = 2 (paper uses α = 100; qualitative predictions hold at α = 2, where "L1 assigns only 70% probability to the FCI states" — p. 249).

Key Innovation #

Standard RSA cannot derive free choice because disjunction is always less informative than its disjuncts. Adding semantic uncertainty — speakers and listeners reason about which interpretation function is being used — creates an avoidance pattern that drives the inference.

The two interpretation functions represent optional exhaustification:

How Free Choice Emerges #

  1. S1 wants to convey "Only One" (each item individually permitted)
  2. If S1 says "You may A", L0 might interpret via I₂ as "Only A"
  3. To avoid this misunderstanding, S1 uses disjunction
  4. L1 reasons: "S1 chose Or to prevent me thinking Only A or Only B"
  5. L1 infers: Only One or Any Number → Free choice

Qualitative Findings #

#FindingTheorem
1FCI derived (uniform prior)fci_derived
2FCI robust to biased priorfci_robust_to_prior
3EI holds under uniform priorei_uniform
4EI weakened under biased priorei_prior_sensitive

State Space (Table 2) #

State◇A◇B◇(A∧B)FCI?EI?
Only ATFFnoyes
Only BFTFnoyes
Only OneTTFyesyes
Any NumberTTTyesno
Only BothTTTnono

Note: "Only Both" means you may ONLY take both together (not either alone).

The 5 states from Table 2, based on permission structure.

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.

      The 4 utterances from (5).

      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.

          Two interpretation functions representing optional exhaustification.

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

              Free choice inference: each item is individually permitted. ◇(A∧¬B) ∧ ◇(B∧¬A). True at {onlyOne, anyNumber}.

              Equations
              Instances For
                theorem RSA.FreeChoice.I2_refines_I1 (u : Utterance) (w : FCState) :
                I2 u w = trueI1 u w = true

                I₂ refines I₁ for all utterances: I₂(u,w) = true → I₁(u,w) = true.

                I₁(Or) is true everywhere (maximally uninformative).

                I₂(A) singles out exactly onlyA.

                noncomputable def RSA.FreeChoice.cfg (worldPr : FCState) (hp : ∀ (w : FCState), 0 worldPr w) :

                @cite{champollion-alsop-grosu-2019} RSA model with semantic uncertainty. Two interpretation functions serve as latent variables. S1 score is rpow(L0, α) — standard belief-based RSA.

                Equations
                • One or more equations did not get rendered due to their size.
                Instances For
                  @[reducible, inline]

                  Uniform prior: all states equally likely.

                  Equations
                  Instances For
                    @[reducible, inline]

                    Biased prior: P(anyNumber) = 3, others = 1. Models a context where taking any combination is a priori more likely, testing prior sensitivity of FCI vs EI.

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

                      FCI is derived: L1 assigns more mass to FCI states (Only One + Any Number) than non-FCI states upon hearing "Or". This is the central result of the paper.

                      FCI is robust to prior manipulation: holds even when anyNumber is a priori 3× more likely.

                      EI is prior-sensitive: a prior biased toward anyNumber defeats EI.

                      The 4 qualitative findings from @cite{champollion-alsop-grosu-2019}.

                      Instances For
                        Equations
                        • One or more equations did not get rendered due to their size.
                        Instances For
                          noncomputable def RSA.FreeChoice.formalize :

                          Map each finding to its RSA formalization.

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

                            The RSA model accounts for all 4 findings from @cite{champollion-alsop-grosu-2019}.

                            The model derives FCI even without the conjunction alternative. This requires either removing the Only Both state or adding a null utterance. We define the null utterance version (Table 8).

                            Utterances with null option (no conjunction).

                            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.

                                RSAConfig for the model without conjunction (Table 8). Replaces "and" with a null utterance (true everywhere under both interpretations).

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

                                  FCI is derived even without the conjunction alternative (Tables 7-8). The avoidance mechanism between A/B and Or is sufficient — the conjunction alternative is not essential.

                                  Bridge content (merged from RSA_ChampollionAlsopGrosu2019Bridge.lean) #

                                  Bridge: RSA Free Choice Disjunction → Phenomena Data #

                                  @cite{champollion-alsop-grosu-2019}

                                  Connects the RSA free choice model from @cite{champollion-alsop-grosu-2019} to empirical data in Phenomena.Modality.FreeChoice.

                                  Bridge Theorems #

                                  Connection to Empirical Data #

                                  The model predicts the patterns in Phenomena.Modality.FreeChoice:

                                  1. Free Choice Permission (coffeeOrTea):

                                    • "You may have coffee or tea" → "You may have coffee AND you may have tea"
                                    • Derived: L1 assigns ~100% to FCI states
                                  2. Exclusivity Cancelability:

                                    • EI ("not both") is sensitive to world knowledge
                                    • FCI is robust across priors
                                  3. Ross's Paradox (postOrBurn):

                                    • "Post the letter" semantically entails "Post or burn"
                                    • But pragmatically, adding "or burn" triggers free choice
                                    • The asymmetry comes from the alternative structure