Multiplicative composition of graded word meanings #
Graded approaches to meaning ([Erk22]'s "probabilistic turn") assign each
word a graded value rather than a Boolean; an utterance meaning is then
composed by multiplying the word values, treating words as independent
constraints: [DHG+20]'s continuous semantics scores a referring
expression as the product of per-word values, and the incremental models of
[WD21] (ℚ-valued) and [SW23] (ℚ≥0-valued)
apply the same composition to utterance prefixes. prodMeaning is this
composition for a per-word lex function lex : U → W → R, stated over any
CommMonoid R. prodMeaning_perm — the composed meaning is independent of
token order — is the listener-level sanity check of
[SW23].
The normalized counterpart on PMF is the Product of Experts
(Core/Probability/ProductOfExperts.lean), the combination rule of
[EH24]'s situation description systems.
Product meaning: the per-word lex composed multiplicatively over a token
list, prodMeaning lex us w = (us.map (lex · w)).prod.
Equations
- Semantics.Probabilistic.prodMeaning lex us w = (List.map (fun (x : U) => lex x w) us).prod
Instances For
Any permutation of the token list yields the same meaning.