Docs / Pragmatics / Probabilistic / Latent
Latent variables
import Linglib.Pragmatics.Probabilistic.Latent · namespace RSA
A latent the speaker knows and the listener does not is a coordinate of the listener's state, and her beliefs about the world and about the latent are the two marginals of her posterior. In the lexical-uncertainty model of Bergen, Levy and Goodman the latent is a lexicon with a prior independent of the state's. The paper states the listener's belief about the state as the listener of the speaker averaged over lexica, and the two presentations agree; her belief about the lexicon is the listener of the speaker averaged over states. A state is possible after an utterance exactly when it has positive prior and some lexicon of positive prior has its speaker produce the utterance there (the union of the lexica's supports), and two states the speakers of every such lexicon treat alike keep their prior odds. A threshold, a comparison class, a scope reading or a question under discussion enters the same way. When the listener hears only the form of the speaker's choice, the latent is the choice, and its prior depends on the state: the joint of the world prior and the speaker.
References
- Bergen et al. 2016, Pragmatic Reasoning through Semantic Inference formalized
- M. Franke and L. Bergen, Theory-Driven Statistical Modeling for Semantics and Pragmatics: A Case Study on Grammatically Generated Implicature Readings (2020)
Marginals of a listener over a product
A latent the speaker also knows, a lexicon or a threshold or a question, is a coordinate of the
state: the speaker at a pair is fun p ↦ speaker α C (L p.2) Measure.dirac p.1, addressing that
latent's listener, and her listener holds a prior over pairs. Its marginals pool the
prior-weighted production probabilities over the other coordinate. When the prior over pairs is a
product, this is lexical uncertainty (Linglib.Pragmatics.Probabilistic.Latent).
At a prior giving every pair the same positive mass, the first marginal of the listener prefers the point with the greater production probability summed over the second coordinate.
At a prior giving every pair the same positive mass, the second marginal of the listener prefers the point with the greater production probability summed over the first coordinate.
The heard form
When the listener hears only the form obs u of the speaker's choice u (eqs. 18b/21b), her
state is the pair of the world and the choice, her prior the joint of the world prior and the
speaker, and her channel the point mass at the form. Her first marginal is her belief about the
world and her second her belief about the choice.
The listener of the form prefers the state with the greater prior-weighted production probability pooled over the choices with the form heard.
Between two choices with the form heard, the listener of the form prefers the one with the greater prior-weighted production probability pooled over the states.
At a uniform prior the listener of the form prefers the state likelier to produce a choice with the form heard.
At a uniform prior the listener of the form prefers, of two choices with the form heard, the one the states together are likelier to produce.
A latent with an independent prior
Lexical uncertainty. Against a product prior, the first marginal of the listener is the listener of the speaker averaged over the second coordinate by its prior.
Against a product prior, the second marginal of the listener, her belief about the latent, is the listener of the speaker averaged over the first coordinate by its prior.
Against a product prior, the listener keeps the prior odds of two states at which the speaker gives the utterance the same mass at every value of positive prior of the second coordinate.
Against a product prior, a state is possible after an utterance exactly when it has positive prior and the speaker produces the utterance there at some value of positive prior of the second coordinate.