PIP: Worked Phenomena #
@cite{keshet-abney-2024} @cite{karttunen-1969} @cite{partee-1972} @cite{roberts-1989} @cite{stone-1997}Concrete examples demonstrating how PIP handles the core anaphora puzzles via description-based retrieval over finite models:
- Stone's puzzle: "A wolf might come. It would eat you first."
- Bathroom sentences: "Either there's no bathroom, or it's upstairs."
- Paycheck pronouns: "John spent his paycheck. Bill saved it."
Each example uses a small finite world/entity model with native_decide
to verify PIP's predictions.
Equations
- One or more equations did not get rendered due to their size.
Instances For
Equations
- Semantics.Dynamic.PIP.Phenomena.instBEqSWorld.beq x✝ y✝ = (x✝.ctorIdx == y✝.ctorIdx)
Instances For
Equations
- One or more equations did not get rendered due to their size.
Instances For
Equations
- Semantics.Dynamic.PIP.Phenomena.instBEqSEntity.beq x✝ y✝ = (x✝.ctorIdx == y✝.ctorIdx)
Instances For
Formula label for "a wolf".
Equations
- Semantics.Dynamic.PIP.Phenomena.αWolf = { idx := 0 }
Instances For
Epistemic accessibility from actual world.
The speaker considers both wolfIn and noWolf possible.
Equations
- Semantics.Dynamic.PIP.Phenomena.sAccess Semantics.Dynamic.PIP.Phenomena.SWorld.actual Semantics.Dynamic.PIP.Phenomena.SWorld.wolfIn = true
- Semantics.Dynamic.PIP.Phenomena.sAccess Semantics.Dynamic.PIP.Phenomena.SWorld.actual Semantics.Dynamic.PIP.Phenomena.SWorld.noWolf = true
- Semantics.Dynamic.PIP.Phenomena.sAccess x✝¹ x✝ = false
Instances For
Wolf predicate: true when the wolf variable is bound to wolf.
Equations
Instances For
Come-in predicate: wolves come in only at wolfIn.
Equations
- One or more equations did not get rendered due to their size.
Instances For
Stone's sentence 1: "A wolf might come."
might(∃^αWolf x. wolf(x) ∧ comeIn(x))
The existential is inside the modal:
- x (the wolf) is local — bound by ∃ inside might
- The world variable is external — bound by might
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.
This is the crucial property: even though the wolf exists only in possible worlds, the descriptive content "wolf(x)" is available in the discourse state for subsequent anaphora.
Stone's sentence 2: "It would eat you first."
"It" = DEF_αWolf{x} — retrieves the wolf via its description. "would" = must in the same accessibility relation.
The pronoun succeeds because αWolf was registered by sentence 1 and labels survive modal operators.
Equations
- One or more equations did not get rendered due to their size.
Instances For
The full Stone's puzzle discourse: sentence 1; sentence 2.
The discourse is well-defined (non-trivially consistent) because:
- Sentence 1 registers αWolf and filters to worlds where a wolf might come in
- Sentence 2 retrieves the wolf via DEF_αWolf and continues in the modal context
Equations
Instances For
After the full discourse, the label is still available. Labels are monotonically accumulated through conjunction, retrieveDef, and would/must because all these operators preserve labels.
End-to-end test: Stone's discourse is consistent on a concrete model.
After processing "A wolf might come. It would eat you first.", the discourse state is non-empty: the assignment g_wolf (with vWolf ↦ wolf) at the actual world survives the full pipeline:
- might: g_wolf at.actual survives because (g_wolf,.wolfIn) is in
the body result (the wolf exists at.wolfIn via
existsLabeled) - retrieveDef αWolf: succeeds because the label was registered
- would/must: g_wolf at.actual survives because
modalExpandadds (g_wolf,.wolfIn) and (g_wolf,.noWolf) to the body's input, and the atom predicate (vWolf ≠ ⋆) holds for all of them
Without modalExpand, step 3 would fail: must would check accessible
worlds.wolfIn/.noWolf but find no pairs there in the body result.
Negative test: an assignment with unbound wolf variable does NOT survive
Stone's discourse. The existsLabeled in sentence 1 extends assignments
with vWolf ↦.some.wolf, and the atom predicate isWolf requires
g.indiv vWolf ==.some.wolf. An unbound assignment (vWolf = ⋆)
fails this predicate check, so it cannot appear in the body result
of might, and is correctly rejected.
This confirms the discourse genuinely filters — it's not vacuously accepting all assignments.
Equations
- One or more equations did not get rendered due to their size.
Instances For
Equations
- Semantics.Dynamic.PIP.Phenomena.instBEqBWorld.beq x✝ y✝ = (x✝.ctorIdx == y✝.ctorIdx)
Instances For
Equations
- One or more equations did not get rendered due to their size.
Instances For
Equations
- Semantics.Dynamic.PIP.Phenomena.instBEqBEntity.beq x✝ y✝ = (x✝.ctorIdx == y✝.ctorIdx)
Instances For
Variable for the bathroom.
Equations
- Semantics.Dynamic.PIP.Phenomena.vBath = { idx := 1 }
Instances For
Bathroom predicate.
Equations
Instances For
Upstairs predicate: the bathroom is upstairs in the bath world.
Equations
- One or more equations did not get rendered due to their size.
Instances For
Partee's bathroom sentence: "Either there's no bathroom, or it's upstairs."
PIP analysis: ¬∃^αBath x.bathroom(x) ∨ upstairs(DEF_αBath{x})
First disjunct: negation of labeled existential
- The label αBath is registered even under negation
- Records description "bathroom(x)"
Second disjunct: uses DEF_αBath to retrieve the bathroom
- The description "bathroom(x)" is available because labels survive negation (the fundamental PIP insight)
This contrasts with standard DPL/DRT where negation blocks accessibility of discourse referents introduced in the negated scope.
Equations
- One or more equations did not get rendered due to their size.
Instances For
The bathroom label is registered after the first disjunct, even though it's under negation.
End-to-end test: the full bathroom sentence is consistent on a concrete model.
Given a universal input (all assignment-world pairs), the bathroom sentence produces a non-empty output. The witness (g₀,.noBath) survives via the first disjunct (negation): g₀ doesn't bind vBath to any entity, so it's NOT in the existsLabeled output, meaning negation keeps it.
This verifies the full compositional pipeline: disj → negation → existsLabeled → atom (first disjunct) disj → label floating → retrieveDef → atom (second disjunct)
Negative test: a bathroom-bound assignment at the no-bathroom world is rejected by the full bathroom sentence.
At .noBath, both disjuncts fail:
- First (negation): g_bath IS in the existential's output (it matches
the
isBathroompredicate), so negation removes it - Second (upstairs):
isUpstairsrequiresw ==.bath, but w =.noBath
This tests the genuine semantic content: the sentence says either there's no bathroom OR the bathroom is upstairs. A bathroom that isn't upstairs violates both conditions.
Paycheck pronouns:
"John spent his paycheck. Bill saved it."
"it" ≠ John's paycheck (which was already spent). "it" = "his paycheck" with "his" re-evaluated for Bill.
PIP analysis:
- "his paycheck" introduces label α with description "paycheck-of(x, possessor)" where possessor is contextually resolved
- "it" = DEF_α{x} — re-evaluates the description relative to the current assignment, where the subject is now Bill
This is the cleanest argument for description-based over value-based anaphora: the pronoun carries a function (λperson. person's paycheck), not a fixed entity.
Value-based account: "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
- Semantics.Dynamic.PIP.Phenomena.instBEqPEntity.beq x✝ y✝ = (x✝.ctorIdx == y✝.ctorIdx)
Instances For
Label for the paycheck description.
Equations
- Semantics.Dynamic.PIP.Phenomena.αPaycheck = { idx := 2 }
Instances For
Variable for the paycheck.
Equations
- Semantics.Dynamic.PIP.Phenomena.vPaycheck = { idx := 2 }
Instances For
Variable for the possessor.
Equations
- Semantics.Dynamic.PIP.Phenomena.vPossessor = { idx := 3 }
Instances For
Equations
- One or more equations did not get rendered due to their size.
Instances For
Equations
- Semantics.Dynamic.PIP.Phenomena.instBEqPWorld.beq x✝ y✝ = (x✝.ctorIdx == y✝.ctorIdx)
Instances For
Paycheck-of predicate: relates a paycheck to its owner.
The key property: this predicate is relational — it depends on both the paycheck variable and the possessor variable. When the possessor changes (John → Bill), the paycheck changes too.
Equations
- One or more equations did not get rendered due to their size.
Instances For
Description-based anaphora correctly predicts the paycheck reading because the description "paycheck-of(x, possessor)" is re-evaluated when the possessor variable changes.
In a value-based system, "it" would be bound to johnsPaycheck. In PIP, "it" = DEF_α{x} with α = "paycheck-of(x, possessor)", and the possessor variable gets rebound to bill.
Summation: Σxφ = ⋃{g(x) : g ∈ G, ⟦φ⟧^{M,{g},w*} = 1}
PIP's formal construct for collecting entity values across assignments. Given a plural context G, a variable x, and a restricting formula φ, summation returns the set of all g(x) values for assignments g that satisfy φ. This is what handles summation pronouns:
"Every farmer bought a donkey. They paid a lot for them."
"them" = Σx(donkey(x)) = the set of ALL donkeys across all farmer-donkey pairs in the plural context G.
Equations
- One or more equations did not get rendered due to their size.
Instances For
Unfiltered summation: collects ALL values of v across assignments.
This is a special case of summationFiltered with a trivially true filter.
Equations
Instances For
Unfiltered summation is summation with trivial filter.
The summation set is non-empty when the context is non-empty and the variable is bound (has a non-⋆ value).
The three systems compared in @cite{keshet-abney-2024}:
Value-based: Pronouns store entity values directly. Works for simple anaphora, fails for modal/negation/paycheck cases.
Description-based (PIP): Pronouns carry formula labels. Handles all cases uniformly.
File-change / DRT: Pronouns access discourse referents. Handles quantifier donkey cases but struggles with modal contexts (requires modal subordination stipulations).
PIP subsumes the file-change approach: descriptions that happen to pick out a unique entity degenerate to value-based behavior.
- valueBased : AnaphoraStrategy
- descriptionBased : AnaphoraStrategy
- fileChange : AnaphoraStrategy
Instances For
Equations
- One or more equations did not get rendered due to their size.
Instances For
PIP uses description-based anaphora.
Equations
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
PIP covers all phenomena.
Value-based fails on modal/negation/paycheck cases.