Documentation

Linglib.Phenomena.Anaphora.Studies.KeshetAbney2024

Keshet & Abney (2024): Intensional Anaphora #

@cite{keshet-abney-2024} @cite{partee-1972} @cite{roberts-1987} @cite{stone-1999} @cite{brasoveanu-2010}

Formalizes the core contributions of @cite{keshet-abney-2024}'s PIP (Plural Intensional Presuppositional predicate calculus) and connects them to the theory-neutral anaphora data in Phenomena/Anaphora/.

Paper's Core Claim #

Pronouns carry descriptive content (formula labels), not stored entity values. A pronoun presupposes that its antecedent description has a non-empty extension in every world of the context set (paper item 9).

This single hypothesis, implemented via PIP's formula labels and felicity conditions, uniformly explains:

  1. Modal subordination — labels survive modals (paper §2.6.3, items 59–60)
  2. Bathroom sentences — labels survive negation (paper §3.4, item 92b)
  3. Donkey anaphora — labels survive ∀ = ¬∃¬ (paper §2.6.2, items 53–56)
  4. Paycheck pronouns — descriptions re-evaluated (paper §2.6.4, items 66–67)
  5. Intensional anaphora — might blocks, must allows (paper §3.1–3.3)

Architecture #

This study file imports:

and proves that PIP's predictions match the empirical data via worked finite models with native_decide verification.

Stone's puzzle world model (@cite{stone-1999}, @cite{roberts-1987}).

Three possible worlds:

  • actual: the evaluation world (no wolf present)
  • wolfIn: a world where a wolf comes in
  • noWolf: a world where no wolf comes in
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.
      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.
          Instances For
            Equations
            • One or more equations did not get rendered due to their size.
            Instances For

              "A wolf might come in." (paper item 59a)

              might(∃^αWolf x. wolf(x) ∧ comeIn(x))

              Label αWolf records the description "wolf(x) that comes in".

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

                After "A wolf might come in", the label αWolf is registered.

                "It would eat you first." (paper item 59b)

                "It" = DEF_αWolf{x}; "would" = must with inherited accessibility.

                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.
                  Instances For

                    End-to-end test: Stone's discourse is consistent on a concrete model.

                    After "A wolf might come. It would eat you first.", the discourse state is non-empty: g_wolf (with vWolf ↦ wolf) at actual survives the pipeline.

                    @cite{partee-1972}'s bathroom sentence world model.

                    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.
                        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.
                            Instances For

                              "Either there's no bathroom, or it's upstairs." (paper item 92b)

                              PIP analysis: ¬∃^αBath x.bathroom(x) ∨ upstairs(DEF_αBath{x})

                              Label αBath is registered under negation and floated to the second disjunct.

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

                                "John spent his paycheck. Bill saved it." (paper items 66–67)

                                "it" carries descriptive content "paycheck-of(x, possessor)" which is re-evaluated when the possessor variable is rebound to Bill.

                                Value-based: "it" → John's paycheck → wrong referent Description-based: "it" → "the paycheck of [current subject]" → Bill's paycheck

                                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.
                                    Instances For

                                      Relational paycheck predicate: depends on both paycheck and possessor.

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

                                        Summation: Σxφ = ⋃{g(x) : g ∈ G, ⟦φ⟧^{M,{g},w*} = 1}

                                        Collects entity values across assignments. For "Every farmer bought a donkey. They paid a lot for them.", "them" = Σx(donkey(x)).

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

                                          Modal continuation type: whether a modal inherits its accessibility relation from prior discourse context.

                                          PIP predicts modal subordination is felicitous iff the second modal subordinates — i.e., it inherits the accessibility relation established by the first modal (paper §2.6.3). "Would" is analyzed as must with the inherited R; "could" as might with the inherited R.

                                          Modals that establish their own accessibility relation (epistemic "must", future "will", indicative mood) cannot access entities introduced under a prior modal's scope.

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

                                              PIP predicts modal subordination felicity iff the second modal subordinates (inherits the accessibility relation from the first).

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

                                                External/local binding modes under modals (paper §2.1).

                                                "Andrea might be eating a cheeseburger. #It is large." (paper item 79)

                                                The burger description is world-dependent: BURGER_w([b]) holds only at worlds where Andrea is eating a burger. At worlds where she isn't, Σb(BURGER_w(b)) = ∅, failing the existential presupposition SINGLE(ΣbE).

                                                Felicity condition (paper item 83): ∀w(MIGHT_w(ΣwE) → SINGLE(ΣbE)) fails because some context-set worlds have no burger.

                                                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.
                                                    Instances For

                                                      World-dependent predicate: burger exists only at burgerW.

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

                                                          Might does NOT give reflexivity at .actual — ibAccess .actual .actual = false. This is why might blocks anaphora: the evaluation world is not among the accessible worlds, so the description need not hold there.

                                                          "There must be some sort of animal in the shed. It's making quite a racket!" (paper item 88)

                                                          The animal description is world-INdependent: ANIMAL_w([x]) ∧ SINGLE(x) holds at ALL accessible worlds (realistic modal base includes actual).

                                                          Felicity condition (paper item 90): ∀w(MUST_w(ΣwX) → SINGLE(ΣxX)) succeeds because must guarantees X at every accessible world including w.

                                                          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.
                                                              Instances For

                                                                World-INdependent predicate: holds at ALL worlds.

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

                                                                    Must allows anaphora via Kratzer's realistic modal base.

                                                                    The animal accessibility relation is reflexive at .actual (the evaluation world), so must_realistic_at (derived from the T axiom) guarantees that the description predicate holds at .actual. This is the Kripke-semantic grounding of why must enables intensional anaphora.

                                                                    The paper's deeper argument about must (items 106–107): in different accessible worlds, different animals could be in the shed. The summation across assignments gives MULTIPLE animals, not a single one.

                                                                    Must still allows anaphora because:

                                                                    1. The accessibility relation is realistic (actual ∈ β_w*)
                                                                    2. The animal AT the actual world is singular (SINGLE)

                                                                    The summation Σx ANIMAL_w*([x]) — evaluated at the discourse world w* — gives the singleton {cat}. The summation across ALL worlds would give {cat, dog, raccoon}, but the world variable in ΣxX is bound by the discourse-level Σw, fixing it to w*.

                                                                    This enriched model shows why Stone/Brasoveanu's system incorrectly predicts plurality: it would sum across all accessible worlds, getting {cat, dog, raccoon} — failing SINGLE.

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

                                                                            Cross-world summation yields PLURAL — Stone/Brasoveanu would incorrectly predict plurality here since they sum across all accessible worlds. Different animals satisfy the description at different worlds: cat at actual, dog at shedW1.

                                                                            "There may already be a winner in the mayoral race. #She is a woman." (paper item 85)

                                                                            This is PIP's strongest argument against Stone/Brasoveanu's "in" predicate. The candidates (alice, bob) are real people who exist in the actual world. A Stone/Brasoveanu-style presupposition requiring only that the referent "exist in the world of evaluation" would wrongly predict felicity.

                                                                            PIP correctly blocks anaphora because the description WINNER_w([x]) is world-dependent: in worlds where the tabulation isn't complete, there is no winner, so Σx WINNER_w([x]) = ∅, failing SINGLE (paper item 87):

                                                                            ∀w(MIGHT_w(Σw WINNER_w([x])) → SINGLE(Σx WINNER_w([x])))

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

                                                                                    "There may already be a winner."

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

                                                                                      Contrast with Stone/Brasoveanu: the entities EXIST in the actual world (alice and bob are real candidates), but the description WINNER is empty there. The "in" predicate would say alice/bob exist → felicitous. PIP says the DESCRIPTION yields nothing at actual → infelicitous.

                                                                                      The label is registered (the mechanism works), but the description cannot be satisfied at the actual world.

                                                                                      The paper's core contribution (§3): might blocks anaphora, must allows it.

                                                                                      The mechanism is the same for both (label + retrieveDef). The difference:

                                                                                      • must guarantees the description holds at the evaluation world (realistic base)
                                                                                      • might only guarantees SOME accessible world

                                                                                      Since the pronoun's existential presupposition (paper item 9) requires the description to hold at the evaluation world, might fails and must succeeds.

                                                                                      Labels are registered in BOTH cases — the asymmetry is about world-dependence of the description, not label availability.

                                                                                      PIP provides a unified account via TWO mechanisms:

                                                                                      1. Label monotonicity: labels survive all operators → modal subordination, bathroom sentences, donkey anaphora

                                                                                      2. World-dependent descriptions + existential presupposition: pronouns presuppose their description holds at the evaluation world → might blocks anaphora, must allows it

                                                                                      No stipulated accommodation rules, no "in" predicate (contra @cite{stone-1999} / @cite{brasoveanu-2010}), no special accessibility conditions.

                                                                                      Evidence: all 5 phenomena are verified by the theorems above:

                                                                                      The might/must asymmetry is grounded in the T axiom of modal logic.

                                                                                      PIP's must agrees with Core.ModalLogic.kripkeEval .necessity (must_truth_agrees_kripkeEval). When the accessibility relation is reflexive at the evaluation world (realistic modal base), the T axiom (must_realistic_at) guarantees the description holds there — enabling anaphora. Non-reflexive might lacks this guarantee.

                                                                                      PIP predicts the standard cross-sentential anaphora pattern:

                                                                                      • Indefinite persistence (Karttunen 1969): ∃^α introduces a label that persists through sequential conjunction → pronoun resolves via DEF_α.
                                                                                      • Standard negation blocks (Heim 1982): negation filters the info state, and the CONJUNCTION version ("John didn't see a bird. It was singing.") fails because sequential conjunction makes the second sentence evaluate in a context where no bird-assignments survive.
                                                                                      • Double negation enables (Elliott & Sudo 2025): ¬¬∃^α x.φ registers α in the body; label monotonicity through both negations preserves it.

                                                                                      The difference between standard negation blocking and double negation enabling is exactly PIP's label monotonicity: labels survive negation (labels_survive_negation), but the info state does not survive single negation in sequential discourse.

                                                                                      PIP predicts that universals and negative quantifiers block cross-sentential anaphora: ∀x.φ = ¬∃x.¬φ does not introduce a labeled existential, so no DEF_α is available.

                                                                                      PIP vs DPL: The Architectural Difference #

                                                                                      DPL negation is a test: ⟦¬φ⟧(g, h) iff g = h ∧ ¬∃k. φ(g, k). The output assignment equals the input — no bindings are exported through negation. This is why ¬¬∃xφ ≠ ∃xφ in DPL (dpl_dne_fails_anaphora): double negation doesn't recover the binding.

                                                                                      PIP negation propagates labels from the body: (negation φ d).labels = (φ d).labels. The info state is complemented, but the label registry survives. This is exactly what enables bathroom sentences and double-negation anaphora.

                                                                                      The following theorems make this architectural difference explicit.

                                                                                      DPL negation resets the output assignment — it cannot export bindings.

                                                                                      This is the key structural property of DPL that blocks cross-negation anaphora: after ¬φ, the output assignment equals the input, so any variables bound inside φ are inaccessible.

                                                                                      PIP negation preserves labels — it CAN export descriptive content.

                                                                                      This is the fundamental advantage of PIP over DPL: even though the info state is complemented (like DPL's test), the label registry propagates outward. The pronoun resolves via DEF_α (label lookup), not via assignment binding.

                                                                                      The contrast: DPL negation blocks anaphora (test), PIP negation allows it (labels survive). This is the architectural reason bathroom sentences are infelicitous in DPL but felicitous in PIP.

                                                                                      Concretely:

                                                                                      • dpl_dne_fails_anaphora: ¬¬∃x.φ ≠ ∃x.φ in DPL (double negation doesn't recover binding)
                                                                                      • bathroom_mechanism: labels survive through negation in PIP (the bathroom sentence works because αBath is registered despite negation)
                                                                                      theorem Phenomena.Anaphora.Studies.KeshetAbney2024.pip_felicity_agrees_with_andFilter {W : Type u_1} (φ ψ : Semantics.PIP.Felicity.PIPExpr W) (w : W) :
                                                                                      (φ.conj ψ).felicitous w = ({ presup := φ.felicitous, assertion := φ.truth }.andFilter { presup := ψ.felicitous, assertion := ψ.truth }).presup w

                                                                                      PIP's F operator and Core.Presupposition.PrProp.andFilter implement the same Karttunen conjunction clause (@cite{karttunen-1973}).

                                                                                      PIP Felicity (PIPExpr.felicitous for .conj φ ψ): φ.felicitous w && ((φ.truth w).not || ψ.felicitous w)

                                                                                      PrProp.andFilter (Core.Presupposition): p.presup w && (!p.assertion w || q.presup w)

                                                                                      These are structurally identical: the first conjunct's felicity/presupposition must hold, AND the second conjunct's felicity/presupposition must hold whenever the first conjunct is true. The first conjunct's truth can "satisfy" the second conjunct's presupposition.

                                                                                      This theorem proves the correspondence by showing that PIP's F operator on conjunction produces the same Bool value as andFilter when we interpret truth as assertion and felicitous as presup.

                                                                                      PIP's F operator for negation agrees with PrProp.neg: both preserve the presupposition/felicity of the negated expression unchanged.

                                                                                      PIP's presupposition construct φ|ψ agrees with PrProp.andFilter where the presupposition is an atom (always felicitous) and the assertion is the presupposed content.

                                                                                      F(φ|ψ) = Fφ ∧ ψ andFilter(⟨Fφ, Tφ⟩, ⟨ψ, trivial⟩).presup = Fφ ∧ (!Tφ ∨ ψ)

                                                                                      These differ slightly: andFilter is the conjunction clause, while φ|ψ is a dedicated presupposition construct. The equivalence holds when the presupposition ψ is not conditioned on φ's truth — which is correct, since ψ must hold unconditionally for φ|ψ to be felicitous.