Docs / Studies / Studies.FoxKatzir2021

Notes on Iterated Rationality Models of Scalar Implicatures

Fox, D. & Katzir, R. (2021). Journal of Semantics, 38(4), 571–600

An iterated rationality model pairs the messages with the cells of the partition they induce by peeling: round after round, each message that identifies a cell is paired with it and both are set aside, until a round identifies nothing. A message identifies a cell when it is true there and at no other remaining cell; it mirror-identifies the cell when it is the only remaining message true there, which is identification under the inverse truth relation; and it does so probabilistically when the cell is the strict mode of the Bayesian listener's posterior. With the naive speaker, who chooses uniformly among the messages true at her cell and is the rational speaker of the Rational Speech Acts model at vanishing rationality, probabilistic peeling at flat priors reads the disjunction of two atoms as their conjunction. The paper's case against the model is that this does not extend: with three or more atoms peeling stops after the atoms, while the grammatical theory's exhaustification strengthens the disjunction of any set of atoms to the cell where exactly those atoms hold, whatever their number; and a prior that makes one disjunct's cell even slightly likelier than the other's strengthens the disjunction to that cell.

Implementation notes

  • Which messages are true at which cells is a relation SetRel M T. Under a downward-entailing operator the numerals' truth relation is the inverse of the upward one, so the two notions of identification trade places.
  • The disjunctions take any finite set of logically independent atoms: a cell is the set of atoms that hold, a message the disjunction or the conjunction of a set of atoms, and an atom the disjunction of itself alone. Two atoms are the paper's §4 and three its §5.
  • A criterion takes the remaining messages and cells as sets, so that the infinite scales of §3 fit. Peeling iterates a round on finite sets as many times as there are messages, after which no round changes anything.
  • The naive speaker chooses among the messages of the original set throughout peeling, as the paper's tables after a round keep the probabilities of the full table.
  • Probabilistic identification compares the posterior of RSA.listener, equivalently the prior times the speaker's probability of the message. At flat priors with the naive speaker it is the comparison of how many messages each cell makes true; among the disjunctions of n atoms a cell of k atoms makes 2 ^ n - 2 ^ (n - k) of them true, so it is the comparison of sizes.
  • The grammatical theory's operator is Exhaustification.exhIEII, Bar-Lev and Fox's innocent exclusion and inclusion, read off the cells of a single atom.
  • Franke's iterated best response is the light system of Studies/Franke2011.lean, iterated from the literal receiver, a naive hearer, or from the literal sender, a naive speaker.

Errata

  • §4.1 says that with the alternatives A or B, A and B "mirror-image cell identification in (10) will fail as well". (10) is the probabilistic notion, which succeeds there; the mirror-image notion that fails is (4) (Disjunction.peel_mirrorIdentifies_disjunctions).
  • §5.2 names the third state to be peeled, three times, as ¬A ∧ B ∧ ¬C, the cell where only B holds. The state meant is ¬A ∧ B ∧ C.
  • (10), unlike (2) and (4), does not require the message to be true at the cell, so every message identifies the last remaining cell (probMirrorIdentifies_of_subset_singleton). Footnote 31 says that adding the three-way conjunction to van Rooij's alternatives keeps A or B or C from identifying any cell; under (10) as stated it is paired with the cell where no atom holds, at which it is false (Disjunction.peel_probMirrorIdentifies_vanRooij_insert).
  • Footnote 33 says that in Franke's system, starting from a naive speaker or from a naive hearer, A or B among the alternatives of a three-way disjunction comes to mean that two of the three disjuncts hold but not all three. In the light system that holds from a naive speaker (Disjunction.franke_naiveSpeaker); from a naive hearer the cell where all three hold keeps its tie with the cells where two do, and A or B comes to mean that at least two hold (Disjunction.franke_naiveHearer).

TODO

  • The second repair of §5.2, on which a message may identify several cells and messages with fewer disjuncts take precedence, is not formalized.
  • The embedded configurations that §4.2 calls analogous, free choice under a possibility modal and disjunctive antecedents, are not formalized here; Studies/ChampollionAlsopGrosu2019.lean has the modal model.

References

Cell identification

Which messages are true at which cells is a relation R : SetRel M T.

L96–98
L104–105
L107–109

Peeling

L120–122
L133–137
L139–142
L147–149
L151–153
L163–182
L184–193
L195–202

Decidability over finite sets

L214–216
L218–220

Probabilistic identification and the naive speaker

L231–236
L239–242
L245–249
L253–256
L258–259
L262–266
L268–272
L274–278
L281–296

The naive speaker as a rational speaker

The naive speaker is the RSA speaker at vanishing rationality. At a flat prior and no cost, the rational speaker of the Rational Speech Acts model weighs each true message by the inverse size of its extension to the power of her rationality, and as the rationality goes to zero she chooses uniformly among the true messages, as the naive speaker (12) does.

theorem tendsto_speaker_naiveSpeakerL311–322
{M : Type u_1} {T : Type u_2} [Fintype M] [Fintype T] [DecidableEq M] [DecidableEq T] [MeasurableSpace M] [DiscreteMeasurableSpace M] [MeasurableSpace T] [DiscreteMeasurableSpace T] (sem : M → Finset T) (t : T) (m : M) :
Filter.Tendsto (fun (α : ℝ) => ((RSA.speaker α 0 (⇑(RSA.listener (ProbabilityTheory.uniformOn Set.univ) (RSA.truthful fun (u : M) => ↑(sem u)))) MeasureTheory.Measure.dirac) t).real {m}) (nhdsWithin 0 (Set.Ioi 0)) (nhds (((naiveSpeaker {p : M × T | p.2 ∈ sem p.1} Finset.univ) t).real {m}))

Built from

Disjunctions (§§2–5)

A message is the disjunction or the conjunction of a set of logically independent atoms, and a cell is the set of atoms that hold.

L340
L349–351
L353–354
L359–360
L362–364
L381–383

Number independence on the grammatical theory ((16), (20)). Against the disjunctions of the nonempty sets of atoms, exhaustification by innocent exclusion and inclusion strengthens the disjunction of any nonempty set of atoms to the cell where exactly those atoms hold: each of its disjuncts holds and every other atom fails. The worlds where a single disjunct holds are the minimal ones; no alternative is excludable but the disjunctions of other atoms, and all the others are includable.

theorem Disjunction.exhIEII_orL390–454
{ι : Type u_1} [Fintype ι] [DecidableEq ι] {S : Finset ι} (hS : S.Nonempty) :
Exhaustification.exhIEII ((fun (m : Msg ι) => ↑(extension m)) '' ↑disjunctions) ↑(extension (Connective.or, S)) = {S}

Built from , ,

L456–458
L462–484

Three or more disjuncts (§5.1). With at least three atoms, peeling the disjunctions by probabilistic identification with the naive speaker at flat priors pairs each atom with the cell where it alone holds and then stops: each remaining disjunction is true at two distinct cells of two atoms, which the naive speaker treats alike, so no disjunction is strengthened. The grammatical theory strengthens every one of them (exhIEII_or).

theorem Disjunction.peel_probMirrorIdentifies_of_three_leL487–586
{ι : Type u_1} [Fintype ι] [DecidableEq ι] (h : 3 ≤ Fintype.card ι) :
peel (ProbMirrorIdentifies (ProbabilityTheory.uniformOn Set.univ) ⇑(naiveSpeaker truth disjunctions)) disjunctions Finset.univ = Finset.image (fun (i : ι) => ((Connective.or, {i}), {i})) Finset.univ

Built from , , , ,

Two atoms (§§2–4, §6)

The conjunctive reading of disjunction (§4.2). With the naive speaker and flat priors, peeling by probabilistic identification first pairs A with the cell where only A holds and B with the cell where only B holds, and then pairs A or B with the cell where both hold. Neither identification (2) nor its mirror image (4) gets started on these alternatives.

theorem Disjunction.peel_probMirrorIdentifies_twoL617–626
:
peel (ProbMirrorIdentifies (ProbabilityTheory.uniformOn Set.univ) ⇑(naiveSpeaker truth disjunctions)) disjunctions Finset.univ = {((Connective.or, {0}), {0}), ((Connective.or, {1}), {1}), ((Connective.or, Finset.univ), Finset.univ)}

Built from , , , ,

L644–653

Sensitivity to priors (§6.1, (22)). Any prior on which the cell where only B holds is likelier than the cell where only A holds, and not too much less likely than the cell where both hold, makes A or B identify the cell where only B holds in the first round of peeling, so that the disjunction is strengthened to the wrong cell. The paper's case, where the other two cells are each 0.9999 times as likely as it, is one.

theorem Disjunction.probMirrorIdentifies_or_of_priorL655–672
{μ : MeasureTheory.Measure (Cell (Fin 2))} [MeasureTheory.IsFiniteMeasure μ] (h₁ : μ.real {{0}} < μ.real {{1}}) (h₂ : 2 * μ.real {Finset.univ} < 3 * μ.real {{1}}) (Ms : Set (Msg (Fin 2))) :
ProbMirrorIdentifies μ (⇑(naiveSpeaker truth disjunctions)) Ms Set.univ (Connective.or, Finset.univ) {1}

Built from , , , , ,

L676–681
L684–695

Three atoms (§5)

L701–707
L713–723

Franke's iterated best response on three atoms (footnotes 30 and 33)

Franke's light system (Franke 2011, Franke2011.receiverChain, Franke2011.senderChain) is iterated best response at flat priors, its message type here the disjunctions and its meanings their extensions.

L736–742
L745–752

A finite scale (§2.2, (14), (21))

The alternatives some, many and all partition the context set into four cells.

L769
L771
L780
L788–791
L806–817

Infinite scales (§3)

Read as at least, the numerals partition the context set into the cells where John has exactly k children. No numeral is true at a single cell, so identification (2) never starts, while each numeral is the only one true at its own cell once the smaller numerals are peeled, so mirror-image identification (4) peels the numerals one by one. Under a downward-entailing operator the truth relation is the inverse one, and the two notions trade places.

L859–863
L866–867
L870–871
L874–876