Documentation

Linglib.Studies.QingGoodmanLassiter2016

[QGL16]: an RSA model of projective content #

The listener jointly infers the world state and the context set the speaker assumed (CogSci 2016, pp. 1110–1115). Under the right QUD this derives presupposition projection with no special semantic mechanism: L0 answers the QUD within the context set (eq. 5), S1 scores QUD-projected informativity (eq. 6, α = 6), and L1 inverts jointly over worlds and context sets (eq. 7). Domain: 13 utterances × 4 worlds × 9 context sets.

Main results #

Implementation notes #

α = 6 is a natural power, so the chain is exact ℚ≥0: the speaker is RSA.Score.s1 over the QUD-aggregated literal listener (dead cells fall back to silence — one fallback declaration covers both faces), listeners are PMF.ofScores of the prior-weighted speaker mass, and each prediction closes by the ofScores comparison family with one kernel certificate.

The paper uses 15 context sets with 5% noise; we model the 9 derivable from past/now observations — the omitted 6 carry only the noise prior (≥ 18× lower) and do not affect the qualitative predictions.

TODO #

Prove the structural equivalence with [ST25] (factives; latent = "private assumptions", their fn. 10: the difference from QGL's "common ground" is interpretive, not computational) and [War22] (genus-species) — all three compute L1(w, C | u, Q) ∝ S1(u | w, C, Q) · P(w) · P(C) over different domains. Blocked on ScontrasTonhauser2025.lean reaching the PMF face.

World states #

World state: (past, now) where past = John smoked, now = John smokes. Flat inductive for tactic enumerability.

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

      Utterances (Table 1, negations, silence) #

      Content utterances about John's smoking habits: 6 positive utterances from Table 1 and their 6 negations.

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

          Utterances with the null alternative: none is silence, some a a content assertion from Table 1.

          Equations
          Instances For

            Literal semantics #

            Literal truth conditions from Table 1. Negation of u is U - ⟦u⟧.

            The change-of-state utterances evaluate the corresponding Features.ChangeOfState.CoSType at the world's temporal projections — presupposition and assertion by construction (Features.ChangeOfState.CoSType.eval_iff), not by bridge theorem.

            Equations
            Instances For

              Context sets #

              Context sets: subsets of worlds representing common ground. These are the 9 context sets derivable from observations about past and now (eq. 8). The paper's full model uses 15 non-empty subsets with 5% noise; the omitted 6 (e.g., change = {(T,F),(F,T)}) have negligible prior.

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

                  QUD #

                  Questions under discussion.

                  Instances For
                    @[implicit_reducible]
                    Equations
                    def QingGoodmanLassiter2016.instReprQUD.repr :
                    QUDStd.Format
                    Equations
                    • One or more equations did not get rendered due to their size.
                    Instances For
                      @[implicit_reducible]
                      Equations
                      • One or more equations did not get rendered due to their size.

                      Priors #

                      Utterance prior (eq. 1): Pr(u) ∝ 2^{-#content-words(u)}. Negation and auxiliaries excluded from count.

                      • 2 content words: stopped/started/always/never smoking → 1/4
                      • 1 content word: smokes/smoked → 1/2
                      • 0 content words: silence → 1
                      Equations
                      Instances For

                        The score chain #

                        Literal listener within the context set (eq. 5; the uniform world prior cancels, and empty extensions normalize to the zero row).

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

                          Speaker (eq. 6 at α = 6): RSA.Score.s1 over the QUD-aggregated literal listener with the utterance prior as multiplicative cost; dead cells fall back to silence.

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

                            Listener world score (eq. 7, world side): context-prior-weighted speaker mass.

                            Equations
                            Instances For
                              def QingGoodmanLassiter2016.ctxScore (qud : QUD) (u : Utterance) (cs : ContextSet) :
                              ℚ≥0

                              Listener context score (eq. 7, context side).

                              Equations
                              Instances For
                                noncomputable def QingGoodmanLassiter2016.l1World (qud : QUD) (u : Utterance) :

                                World-side pragmatic listener (eq. 7).

                                Equations
                                Instances For
                                  noncomputable def QingGoodmanLassiter2016.l1Ctx (qud : QUD) (u : Utterance) :

                                  Context-side pragmatic listener (eq. 7).

                                  Equations
                                  Instances For
                                    noncomputable def QingGoodmanLassiter2016.l1WorldEvent (qud : QUD) (u : Utterance) (P : WorldStateBool) :
                                    ENNReal

                                    Event marginal of the world-side listener.

                                    Equations
                                    Instances For

                                      Listener predictions #

                                      QUD_now symmetry #

                                      Under QUD_now, wTT and wFT are indistinguishable: qAggQ .now is definitionally symmetric in the now-cell, so projection must be measured on the context-set side (l1Ctx), not the world marginal.

                                      QUD answer #

                                      L1 infers the QUD answer now=T from "didn't stop smoking".

                                      World elimination #

                                      "Didn't stop smoking" is literally false at wTF, concentrating L1 on wTT.

                                      Projection under QUD_max (Figure 1c) #

                                      The identity QUD breaks the QUD_now degeneracy — under +past, "didn't stop" narrows to exactly wTT — but projection stays incomplete: wTT = wFF, so the past marginals still coincide.

                                      Incomplete projection under QUD_max: wTT = wFF (Figure 1c). The involution (past, now) ↦ (¬now, ¬past) fixes ⟦didn't stop⟧ and permutes the context sets prior-preservingly, so world marginals cannot fully register the presupposition.

                                      Context set projection (the paper's main result, Figure 3) #

                                      Under +past, "didn't stop" narrows L0 to wTT (aggregate 1, 1⁶ = 1); under −past it spreads ((1/2)⁶ = 1/64). S1 rewards informativity, so L1 infers the speaker assumed a +past common ground — projection as context-set inference, with (wTT, +past) the joint mode (Figure 3).

                                      +past beats universe despite the prior (6 vs 9): "didn't stop" narrows to wTT under +past (1⁶ = 1) but spreads over three worlds under universe ((1/4)⁶).

                                      −now contexts are dispreferred (p. 1114): −now already entails the QUD answer, so "didn't stop smoking" adds nothing there.

                                      "Stopped smoking" #

                                      "Stopped smoking" → L1 infers now=F (the assertion).

                                      Structural properties #

                                      "didn't stop smoking" excludes exactly the stopping world.

                                      Under a +past context set, "didn't stop" is maximally informative for QUD_now: it narrows to exactly wTT.

                                      "stopped smoking" denotes exactly the stopping world (T,F).

                                      "always smoked" denotes exactly (T,T).

                                      "never smoked" denotes exactly (F,F).

                                      3 of 9 context sets entail past=T.