Documentation

Linglib.Studies.Alsop2024

[Als24] — The pragmatics of free choice any #

[Als24] (Glossa 9(1)) argues that You may read any book does not entail that each book may be read on its own ([menendez-benito-2010], [dayal-2013]'s Viability Constraint) — it carries a particularly robust exclusiveness implicature, derived pragmatically from [szabolcsi-2019]'s weaker semantics. The derivation combines [CAG19]'s ambiguity-driven free choice with [FB20b]'s Global Intentions architecture: each utterance comes with its own set of licit exhaustified parses, the speaker chooses an utterance–parse pair jointly, and the listener infers state and intended parse together.

The model (the paper's §5, eqs. (36)–(41)): seven permission states over a two-class domain (their Table 1); four utterances with 12 utterance–parse pairs in total (3 for may S, 3 for may P, 2 for may any, 4 for may every; truth conditions in their Table 2); L0(s|u,p) ∝ P(s)·⟦u⟧ᵖ(s); S1(u,p|s) ∝ exp(α·log L0) over all 12 pairs; L1(s,p|u) ∝ P(s)·S1(u,p|s) with parse-marginal L1(s|u). Speaker optimality α = 100, equal costs.

Instantiated on the canonical pipeline: L0 is RSA.Canonical.L0OfBool over the Table-2 matrix, the joint speaker is RSA.Canonical.S1 with powUtility, the parse-marginal speaker is its PMF.map along Parse.utt, and the pragmatic listener is RSA.Canonical.L1.

Main statements #

Context manipulation (verified prose) #

The paper's Tables 5–9 manipulate the state prior, which enters L0 (eq. (36)); independently recomputed, every reported cell reproduces exactly. The not every implicature is absent at a uniform prior (the 50/50 split above), derived under a 70%-Only-1 prior (L1(Only 1) ≈ 1, robust for all scanned α ≥ 1), and the Any-#-biased prior shifts L1 to 0.93 while S1 stays 50/50 — a prior-driven shift, not an implicature (the paper's eq. (1)). Robustness of exclusiveness to a 70%-S-or-2 prior (Table 5: 0.49/0.49/0.02) genuinely needs α ≥ 45, so the paper's α = 100 does real work there. These prior-in-L0 results are recorded as prose rather than theorems per the parameter-dependence policy.

Implementation notes #

Theorems are at the paper's α = 100 (the parse preference at every α ≥ 1) with a uniform prior, by -certificate dominance bounds — no numeric reflection. The previous version of this file modelled two global interpretation functions with an interpretation prior at α = 2 — the [CAG19] architecture that the paper explicitly replaces with [FB20b]'s — and included a negation finding with no counterpart in the paper's model (its utterance set has no negation; NPI any is set aside in the paper's §2.1).

States, utterances, parses (the paper's Tables 1–2) #

The seven permission states (their Table 1), each a set of accessible worlds over {take nothing, take S, take P, take both}; every state makes taking nothing accessible.

Instances For
    @[implicit_reducible]
    Equations
    def Alsop2024.instReprFCIState.repr :
    FCIStateStd.Format
    Equations
    Instances For
      @[implicit_reducible]
      Equations

      The four utterances (their (31)).

      Instances For
        @[implicit_reducible]
        Equations
        def Alsop2024.instReprUtterance.repr :
        UtteranceStd.Format
        Equations
        Instances For

          The 12 utterance–parse pairs (their (32)–(35)): per-utterance licit exhaustified parses, a the weakest. May any has exactly two — the weak parse (34a, Szabolcsi: every class may be taken, possibly only together) and the strong parse (34b, Dayal: every class may be taken on its own).

          Instances For
            @[implicit_reducible]
            Equations
            def Alsop2024.instReprParse.repr :
            ParseStd.Format
            Equations
            Instances For
              @[implicit_reducible]
              Equations
              @[implicit_reducible]
              Equations
              @[implicit_reducible]
              Equations
              def Alsop2024.meaning :
              ParseFCIStateBool

              Truth conditions for each utterance–parse pair (their Table 2).

              Equations
              Instances For

                The canonical GI pipeline #

                @[reducible, inline]
                noncomputable abbrev Alsop2024.l0 :
                UnitParsePMF FCIState

                Per-parse literal listener (eq. (36) at a uniform state prior): uniform on the parse's extension.

                Equations
                Instances For
                  noncomputable def Alsop2024.speaker (α : ) :
                  FCIState × UnitPMF Parse

                  The joint speaker over utterance–parse pairs (eqs. (37)–(38)): S1(u,p|s) ∝ L0(s|u,p)^α, equal costs.

                  Equations
                  Instances For
                    noncomputable def Alsop2024.speakerU (α : ) (s : FCIState × Unit) :

                    The parse-marginal speaker (eq. (39)): S1(u|s) = Σ_p S1(u,p|s).

                    Equations
                    Instances For
                      @[reducible, inline]
                      noncomputable abbrev Alsop2024.prior :
                      PMF (FCIState × Unit)

                      Uniform joint state prior.

                      Equations
                      Instances For
                        noncomputable def Alsop2024.listener (α : ) (u : Utterance) :
                        PMF (FCIState × Unit)

                        The pragmatic listener over states (eqs. (40)–(41), parse-marginal): the canonical posterior of the parse-marginal speaker.

                        Equations
                        Instances For

                          Extension sizes (their Table 2 row sums) #

                          B1 zeros: may any and may every exclude single-class states #

                          theorem Alsop2024.speakerU_onlyS_mayAny (α : ) ( : α 0) :

                          Both parses of may any are false at Only-S, so the speaker never produces it there.

                          theorem Alsop2024.mayAny_rules_out_onlyS (α : ) ( : α 0) :

                          Hearing may any, the listener assigns zero posterior to Only-S.

                          theorem Alsop2024.speakerU_onlyS_mayEvery (α : ) ( : α 0) :

                          All four parses of may every are false at Only-S.

                          theorem Alsop2024.mayEvery_rules_out_onlyS (α : ) ( : α 0) :

                          Hearing may every, the listener assigns zero posterior to Only-S.

                          The mechanism: the strong parse wins #

                          theorem Alsop2024.s1_prefers_strong_parse {α : } ( : α 0) :

                          At the Only-1 state the speaker prefers the strong parse (34b) of may any to the weak parse (34a), for every exponent α ≥ 1: the weak parse is true in five states (L0 = 1/5), the strong in two (L0 = 1/2), and within one state the softmax partition cancels. At α = 100 the ratio is (5/2)^100, the paper's "almost 100% of the time".

                          Exclusiveness (their Table 3, uniform prior, α = 100) #

                          Per-state speaker bounds: at the exclusiveness states the strong parse alone gives may any more than a third of the speaker's mass; at the non-exclusive states only the weak parse survives and is dominated.

                          theorem Alsop2024.exclusiveness_derived :
                          (listener 100 Utterance.mayAny).toOuterMeasure {(FCIState.only2, ()), (FCIState.sOr2, ()), (FCIState.pOr2, ())} < (listener 100 Utterance.mayAny).toOuterMeasure {(FCIState.only1, ()), (FCIState.anyNum, ())}

                          The exclusiveness implicature (their Table 3): hearing may any at a uniform prior, the listener puts more posterior mass on the exclusiveness states {Only 1, Any #} (where each class may be taken on its own; the paper's 0.50 + 0.50) than on the non-exclusive states {Only 2, S or 2, P or 2} (each ≈ 0).

                          May S communicates Only-S (their Table 3: 0.67 vs 0.33) #

                          theorem Alsop2024.literal_s_communicates_onlyS :
                          (listener 100 Utterance.mayS).toOuterMeasure {(FCIState.sOr2, ())} < (listener 100 Utterance.mayS).toOuterMeasure {(FCIState.onlyS, ())}

                          Hearing may S, the listener prefers Only-S to S-or-2 (their Table 3: 0.67 vs 0.33): the dedicated exhaustified parse (32c) is only available at Only-S.

                          The strict Only-1 / Any-# asymmetry (refining their Table 3) #

                          theorem Alsop2024.exclusiveness_strict_asymmetry :
                          (listener 100 Utterance.mayAny).toOuterMeasure {(FCIState.anyNum, ())} < (listener 100 Utterance.mayAny).toOuterMeasure {(FCIState.only1, ())}

                          The asymmetry the paper's Table 3 rounds away: at a uniform prior, L1(Only 1 | may any) strictly exceeds L1(Any # | may any) — the paper reports 0.50/0.50 at α = 100 (the difference is ≈ 2·10⁻³¹). Both parses of may any carry the same weight at the two states, but may every's parse (35b) is true at Any # and not at Only 1, strictly inflating the speaker's partition there, so the speaker is less likely to choose may any at Any #. A formaliser's refinement, not a claim of the paper's.