Docs / Studies / Studies.BrochhagenFrankeVanRooij2018

Coevolution of Lexical Meaning and Pragmatic Use

Brochhagen, T., Franke, M. & van Rooij, R. (2018). Cognitive Science, 42(8), 2757–2789

A population of language users, each pairing a lexicon with a level of pragmatic reasoning, evolves by the replicator mutator dynamic: types reproduce in proportion to their expected communicative success and pass to the next generation by iterated Bayesian learning. The speakers send false messages, so any type can be mistaken for any other. Over the two states of some, a lexicon with the upper bound of some communicates better than one without, and the claim that every other type beats literal use of the unbounded lexicon fails at low rationality. Learners who sample from the posterior keep the prior; learners who adopt the likeliest type favour the type others are mistaken for.

Implementation notes

  • The pragmatic speakers of eq. (6) maximize the chance that the hearer guesses the state, so they are RSA.actionSpeaker at the utility of a correct guess, not RSA.speaker, whose logarithmic utility Frank & Goodman 2012 never sends a false message (note 2).
  • A type's data (produce) draw each state from the state prior, so they have total mass one; the posterior over types is the same as under the printed likelihood.
  • The factor 1.05 of Fig. 2 is the prior's parameter, from the caption. invariant_biasedPrior shows for every factor that the stationary ratio of L_lack to L_bound users is the factor, which is the "exactly 1.05" of p. 2770.
  • Appendix B's learners at l = 10 are replaced by learners who adopt the likeliest type, the limit l → ∞ and Griffiths & Kalish 2007's mapLearner; the printed Q at l = 10 is LH to the printed precision.
  • The complexities of Table 2 are taken as printed; the lexicon prior is left unnormalized.
  • Swapping none and all reverses the scale and preserves the flat prior, so the model is symmetric under it (success_image). It takes the lexicon {none, some, all} to {all, not all, none}, which therefore communicates exactly as well with every type; only the prior separates them, by 7 : 6, since not all costs one unit more than some (success_lexMirror). The paper does not state this. Its six lexica "of the third kind" (p. 2770) are the permutations of the first; the permutations of the second leave the lower bound of not all to pragmatics instead.
  • Brochhagen et al. 2016 is the precursor, over two states and two messages, with a speaker that weighs the literal hearer by an exponent and a prior that penalizes upper-bounded messages; it is not formalized separately. Its update (p. 2083) weights the mutation into type i by the share of i in the previous generation, which loses mass: with no mutation, equal fitness and the uniform population over its twelve types in both generations, the next population has total mass 1/12. Its prose (p. 2082) says that L_a gives m₁ an upper bound, but by its matrix the bounded message, true only at the first state, is m₂.

TODO

  • The simulations of §3.3 are not formalized.
  • That the costs of Table 2 are the least derivation costs in the grammar of Table 1.
  • Convergence, not just invariance, of the chains of Fig. 2(b) and Appendix B.
  • The type-level fitness of §2.4.2, the mean of a type's success as speaker and as hearer, over the 432 types of §3.2.

References

Language users

L84–93
L95–98
L102–104
L106
L108–110
L112–113
L115–119
L121–122
L125–128
L132–138
L142–148
L151–156
L165–179
L183–189

Note 2: a speaker who maximizes the log of the literal hearer's credence in the state, as in Frank & Goodman 2012, never sends a false message, whereas every speaker of the paper, maximizing the hearer's chance of guessing the state, sends every message with positive probability.

theorem speaker_eq_zero_and_agents_fst_ne_zeroL191–199
{S : Type u_1} {M : Type u_2} [Fintype S] [MeasurableSpace S] [DiscreteMeasurableSpace S] [DecidableEq S] [Fintype M] [MeasurableSpace M] [DiscreteMeasurableSpace M] (pr : MeasureTheory.Measure S) (lam : ℝ) (L : M → Set S) {α : ℝ} (hα : 0 < α) {s : S} {m : M} (h : s ∉ L m) :
((RSA.speaker α 0 (⇑(RSA.literalListener pr L)) MeasureTheory.Measure.dirac) s) {m} = 0 ∧ ∀ (n : ℕ), ((agents pr lam L n).1 s) {m} ≠ 0

Built from

Some and all: the restricted type space of §3.1

The states are the two vertices of the I corner of the square of opposition, some but not all (Triangle.someNotAll) and all (Triangle.all), equally likely.

L222
L224
L243–244
L315–321
L324–328
L331–366

§3.1 (p. 2768): literal L_bound communicates better with its own kind than pragmatic L_bound does, which gains nothing from reasoning over a lexicon with an upper bound, and better than literal L_lack, whose some is ambiguous; and pragmatic L_bound communicates better than pragmatic L_lack. These hold at every rationality.

theorem selfSuccess_ltL438–456
{lam : ℝ} (hlam : 0 < lam) :

Built from , ,

§3.1 (p. 2768) claims that every other type communicates better than literal L_lack. With pragmatic L_bound this holds exactly when λ > log(5/3), and pragmatic L_lack does worse than literal L_lack whenever λ ≤ 2 log(5/2), as at the paper's λ = 1.

theorem selfSuccess_Llack_zero_ltL458–481
{lam : ℝ} (hlam : 0 < lam) :
(selfSuccess lam Llack 0 < selfSuccess lam Lbound 1 ↔ Real.log (5 / 3) < lam) ∧ (lam ≤ 2 * Real.log (5 / 2) → selfSuccess lam Llack 1 < selfSuccess lam Llack 0)

Built from , ,

Mutation alone on the four types of §3.1

A type pairs one of the two lexica with a reasoning level, 0 for literal and 1 for pragmatic use. A type's data are k states drawn from the prior, each with the message the type sends there.

L511
L513
L526–527
L540–545

§3.1 (p. 2770), Fig. 2(b): learners who sample from the posterior (l = 1) keep the prior Griffiths & Kalish 2007, so a population under mutation alone settles where users of L_lack outnumber users of L_bound by the factor by which the prior favours them, 1.05 in Fig. 2.

theorem invariant_biasedPriorL556–566
{lam : ℝ} (k : ℕ) (c : NNReal) :
((KirbyDowmanGriffiths2007.exponentiatedPosterior (produce lam k) (biasedPrior c) 1).comp (produce lam k)).Invariant (biasedPrior c) ∧ (biasedPrior c).real {t : Lex × Fin 2 | t.1 = Lex.lack} / (biasedPrior c).real {t : Lex × Fin 2 | t.1 = Lex.bound} = ↑c

Built from , ,

Three states: the lexica of §3.2

The states are the three vertices of the triangle of opposition, none (Triangle.none), some but not all (Triangle.someNotAll) and all (Triangle.all), equally likely, and a lexicon sends each of three messages to a set of states.

L699–717

The replicator mutator dynamic

A population is a probability measure over types. The replicator step reweights each type by its fitness, and the mutator step passes the result through the kernel Q that sends a teacher's type to the type its learners acquire.

L743–744
L746–750
L752–758
L762–763
L766–771

Appendix A: a coordination game

Two types of agent, negative and positive, earn 1 when two negatives meet, 2 when two positives meet, and 0 otherwise. A tenth of the learners of each type acquire the other.

L792
L816–821

Appendix B: mutation that favours the type the others are mistaken for

Three equally likely types produce three data. Type t₁ produces its own datum d₁ with chance 1 - 2ε; types t₂ and t₃ produce theirs with chance 1 - (ε + δ) and d₁ with chance δ.

L890–892
L912–917
L925–936
L938–943
L960–989

App. B (p. 2789): learners who sample from the posterior (l = 1) keep the flat prior, while learners who adopt the likeliest type learn LH itself, whose only stationary population is (δ, ε, ε) / (δ + 2ε). This population favours t₁ exactly when δ > ε.

theorem invariant_mapLearner_comp_lh_iffL991–1008
{ε δ : ℝ} (hε : 0 < ε) (hδ : 0 ≤ δ) (h : δ + 2 * ε < 1) (ν : MeasureTheory.Measure (Fin 3)) [MeasureTheory.IsProbabilityMeasure ν] :
((KirbyDowmanGriffiths2007.exponentiatedPosterior (lh ε δ) (ProbabilityTheory.uniformOn Set.univ) 1).comp (lh ε δ)).Invariant (ProbabilityTheory.uniformOn Set.univ) ∧ (((GriffithsKalish2007.mapLearner (lh ε δ) (ProbabilityTheory.uniformOn Set.univ)).comp (lh ε δ)).Invariant ν ↔ ν = lhStationary ε δ) ∧ ((lhStationary ε δ).real {1} < (lhStationary ε δ).real {0} ↔ ε < δ)

Built from ,