Asymmetric Communication Models #
@cite{xu-etal-2024} @cite{milroy-milroy-1985} @cite{labov-2011}
A communicative interaction in which speaker and listener operate on
different lexicons / different channels. The canonical case is
variation/innovation diffusion: a speaker has acquired a novel encoding
E*, the listener has not yet, so the speaker's production policy is
conditioned on L' while the listener's interpretation is conditioned
on L (@cite{xu-etal-2024} §2.1).
This generalizes the standard symmetric Shannon setup (@cite{kemp-regier-2012}, @cite{zaslavsky-kemp-regier-tishby-2018}, where speaker and listener share one channel) and serves as substrate for any linguistic-modeling case in which the two ends of a communication act have different generative models. Concrete uses:
- Lexicalization spread (@cite{xu-etal-2024})
- Iterated learning (parent / child generation mismatch; @cite{kirby-tamariz-cornish-smith-2015})
- L1 / L2 communicative accommodation
- Native / non-native listener mismatch in adaptation studies
Two layers #
ChannelAccess is a bare-bones non-finite, non-normalized score function
C → W → ℝ — the most general view of "what a participant returns when
asked about (concept, form)". CommChannel (in InformationTheory/Channel.lean)
is the Shannon-strength version (finite alphabets + row-stochastic).
CommChannel.toAccess projects the latter onto the former. AsymmetricCommModel
holds two ChannelAccesses so it can be instantiated either by lifting two
CommChannels (ofCommChannels) or by stipulating two arbitrary score
functions over a non-finite type (e.g., String-valued lexicons).
Bare-bones channel access: a score function C → W → ℝ with no
finiteness or normalization constraints. The minimal interface
needed to compute per-pair surprisals.
Equations
- Pragmatics.Communication.ChannelAccess C W = (C → W → ℝ)
Instances For
Project a Shannon CommChannel onto its bare access function.
Lives in Pragmatics.InformationTheory so dot notation
ch.toAccess resolves.
Instances For
A two-channel communication model: speaker uses produce, listener
uses comprehend. The asymmetry is structural — produce and
comprehend are independent fields and may disagree.
@cite{xu-etal-2024}'s headline conceptual move over @cite{kemp-regier-2012}
and @cite{zaslavsky-kemp-regier-tishby-2018}: the speaker's production
policy is conditioned on the expanded lexicon L' while the listener's
interpretation is conditioned on the existing lexicon L. This
structural distinction lives in the variation-theory tradition of
@cite{labov-2011} and @cite{milroy-milroy-1985}.
- produce : ChannelAccess C W
Speaker channel
p(w | c, L'). - comprehend : ChannelAccess C W
Listener channel
p(w | c, L)(or, equivalently, the listener's score function for inferringcfromwunder their lexicon).
Instances For
Symmetric special case: speaker and listener share one channel.
Recovers the Kemp-Regier / Zaslavsky setup as
AsymmetricCommModel.symmetric ch.
Equations
- Pragmatics.Communication.AsymmetricCommModel.symmetric ch = { produce := ch, comprehend := ch }
Instances For
Construct an asymmetric model from two finite-alphabet
CommChannels — one for the speaker on L', one for the listener
on L.
Equations
- Pragmatics.Communication.AsymmetricCommModel.ofCommChannels speakerCh listenerCh = { produce := speakerCh.toAccess, comprehend := listenerCh.toAccess }
Instances For
An asymmetric model is symmetric iff its two channels coincide as functions.
Equations
- m.IsSymmetric = (m.produce = m.comprehend)