Docs / Studies / Studies.ZaslavskyMaldonadoCulbertson2021

Let’s Talk (Efficiently) about Us: Person Systems Achieve Near-Optimal Compression

Zaslavsky, N., Maldonado, M. & Culbertson, J. (2021). In Proceedings of the 43rd Annual Meeting of the Cognitive Science Society (pp. 938–944)

Zaslavsky, Maldonado and Culbertson apply the Information Bottleneck to personal pronouns. The person space has eleven referents, built from the speaker, the addressee and others, one, exactly two or more of them. A speaker's meaning for a referent spreads over the space in proportion to $\exp(-\gamma d)$, $d$ a weighted Hamming distance between the referents' features, and a pronoun system is an encoder of meanings into forms. Zwicky's generalization, that no language gives the inclusive the form of the second person to the exclusion of the exclusive, is where the paper's models part: weighting the speaker and the addressee alike rates the unattested second-inclusive system as efficient as attested ones, while weighting the speaker sixteen times the addressee, the egocentric model, separates them. With the weights alike, exchanging the speaker and the addressee is a symmetry of the meanings, and under a need distribution that the exchange also preserves an encoder and its exchange have the same complexity and the same accuracy. The second-inclusive system is the exchange of a first-inclusive one, so the unbiased model values the two exactly alike unless the need distribution tells the speaker from the addressee: the separation needs the egocentric weights or an asymmetric need. Harbour's two person features, by contrast, generate the partition of the first system's groups and never that of the second's, Zwicky's tripartition: a hard constraint where the paper finds a soft bias.

Implementation notes

The paper encodes number in two bits that its unbiased model counts as one feature; here the number term of the distance is whether the numbers differ. The precision $\gamma$, which the paper fixes from the speaker's memory capacity, and the need distribution, which it infers from corpora, are parameters.

References

Symmetries of the Information Bottleneck

An encoder relabelled by a symmetry of the source has the encoder's complexity.

theorem complexity_comapL91–96
{M : Type u_1} {W : Type u_3} [MeasurableSpace M] [MeasurableSpace W] [Fintype M] [Fintype W] [DiscreteMeasurableSpace M] [MeasurableSingletonClass M] [MeasurableSingletonClass W] {source : MeasureTheory.Measure M} [MeasureTheory.IsProbabilityMeasure source] (encoder : ProbabilityTheory.Kernel M W) [ProbabilityTheory.IsMarkovKernel encoder] {e : M ≃ M} (he : MeasureTheory.Measure.map (⇑e) source = source) :
L131–147

An encoder relabelled by a symmetry of the source that the meanings follow has the encoder's accuracy.

theorem accuracy_comapL149–157
{M : Type u_1} {U : Type u_2} {W : Type u_3} [MeasurableSpace M] [MeasurableSpace U] [MeasurableSpace W] [Fintype M] [Fintype U] [Fintype W] [DiscreteMeasurableSpace M] [DiscreteMeasurableSpace U] [MeasurableSingletonClass M] [MeasurableSingletonClass U] [MeasurableSingletonClass W] [Nonempty M] (meanings : ProbabilityTheory.Kernel M U) [ProbabilityTheory.IsMarkovKernel meanings] {source : MeasureTheory.Measure M} [MeasureTheory.IsProbabilityMeasure source] (encoder : ProbabilityTheory.Kernel M W) [ProbabilityTheory.IsMarkovKernel encoder] {e : M ≃ M} {f : U ≃ U} (he : MeasureTheory.Measure.map (⇑e) source = source) (hm : ∀ (m : M), meanings (e m) = MeasureTheory.Measure.map (⇑f) (meanings m)) :
ZaslavskyKempRegierTishby2018.accuracy meanings source (encoder.comap ⇑e ⋯) = ZaslavskyKempRegierTishby2018.accuracy meanings source encoder

The person space

L163–188
L193
L194
L266–269
L271–273

Inclusive systems

L288–296
L298–306

Weighting the speaker and the addressee alike, under a need distribution that their exchange preserves, the second-inclusive system and the first-inclusive system have the same value of the Information Bottleneck objective at every trade-off.

theorem objective_secondInclusiveL311–324
{w : Weights} (h : w.speaker = w.addressee) (γ : ℝ) (need : MeasureTheory.Measure Referent) [MeasureTheory.IsProbabilityMeasure need] (hneed : MeasureTheory.Measure.map (⇑Referent.swapEquiv) need = need) (β : ℝ) :
ZaslavskyKempRegierTishby2018.objective (meanings w γ) need (ProbabilityTheory.Kernel.deterministic secondInclusive ⋯) β = ZaslavskyKempRegierTishby2018.objective (meanings w γ) need (ProbabilityTheory.Kernel.deterministic firstInclusive ⋯) β

Built from , , , , ,

Against Harbour's parameters

On the referents with more than one member, the first-inclusive system groups the inclusive with the exclusive and the second-inclusive system groups it with the second person, the person partitions of the standard tripartition and of Zwicky's.

The two systems the unbiased model values alike are told apart by Harbour's parameters, which generate the partition of the first-inclusive system's groups and never that of the second-inclusive system's, Zwicky's tripartition.

theorem groupGraph_mem_image_setting_graphL352–360
:
groupGraph firstInclusive ∈ Finset.image Harbour2016.Setting.graph Finset.univ ∧ groupGraph secondInclusive ∉ Finset.image Harbour2016.Setting.graph Finset.univ

Built from , ,