nucleation

THEORY-172: In Frank and Goodman's one-shot reference games, listeners' mean bets matched a parameter-free listener inverting a speaker who chooses among true utterances by informativeness, with a separately measured salience prior, and not salience alone; and in Bergen, Goodman and Levy's signalling games with novel symbols, listeners read a signal as excluding what the available alternative would have conveyed

4 min read

Properties8
StatusActive
Version4
Tagspragmatics, linguistics, cognition, probabilistic-modeling
SourceLIT-821 — Predicting Pragmatic Reasoning in Language Games, LIT-809 — That's what she (could have) said: How alternative utterances affect language use, LIT-816 — Pragmatic Language Interpretation as Probabilistic Inference
ExtendsTHEORY-472 — A speaker who chooses words to minimise the information lost about a single intended referent, with words' literal meanings uniform over their extensions, picks each true word with probability inversely proportional to a power of its extension, so the size principle follows from informative communication; in one betting experiment, learners who knew the referent bet on a novel word's meaning in line with the inversion of that speaker (r = .93 on condition means), and in one free-description study speakers' word use followed it only weakly (r = .19)
Extended byTHEORY-785 — Listeners interpret an utterance by inverting a model of a speaker who chooses among alternatives by informativeness, so what a fixed form conveys depends on what else the speaker could have said, as the rational speech act account of Frank, Goodman and coauthors proposes
SupportsCLAIM-111 — A text supports a distribution over possible communicative situations, and a translation can be faithful by preserving that distribution and its response to further evidence, CLAIM-221 — The relational-constitution thesis's defeat condition describes a standard rival the record holds no counter for, intention-based accounts of illocution, which plausibly meet it for every act in the four-speakers case except the institutional one, CLAIM-236 — Under the rational-speech-act model the distribution-over-situations thesis grounds on, a form's interpretation depends on what else the speaker could have said, so matching source and target posteriors needs a correspondence of alternative sets that translation changes and no entry names, CLAIM-214 — The distribution-over-situations criterion compares each language's own listener posterior, so it needs no correspondence of alternative sets; that a literal rendering can fail it is the criterion detecting a lost implicature, and a target language may admit no rendering that passes, CLAIM-242 — The use-level counterpart of Saussure's value is an utterance's contrast with the alternatives its speaker could have used in context; values do not cross between languages, which is why fidelity is graded transport through an anchored correspondence rather than identity of value
Cited as “grounds” byARG-040 — In use, an utterance is interpreted by its contrast with the alternatives its speaker could have used, so what a translation must let a receiver reconstruct is those contrasts, not the wording or the proposition

Graph View

Source

  • Frank & Goodman (2012), LIT-821, read in NOTE-634: Eqs. 1–2, S1–S4, the three-group design and its correlations.
  • Bergen, Goodman & Levy (2012), LIT-809, read in NOTE-613: Eqs. 1–8, Experiments 1–2.
  • Goodman & Frank (2016), LIT-816, read in NOTE-606 from the authors’ manuscript: the general form and its extensions.

The claim

In one-shot reference and signalling games with a known set of alternative utterances, listeners’ interpretations match P_L(m|u) ∝ P_S(u|m)P(m), where P_S chooses among the utterances true of m in proportion to exp(α[log P_Lit(m|u) − c(u)]) and P_Lit conditions the prior on literal truth. Two consequences follow and are observed: a general utterance is read as excluding what a more specific available utterance would have conveyed (specificity implicature), and, when the agents are uncertain what the signals mean, a costlier signal is read as the less likely meaning (Horn’s principle). The measured prior alone does not predict the interpretation; its combination with the speaker model does. A literal-meaning-only listener was not fitted.

What was actually shown

Frank and Goodman measured the prior and the listener in separate groups of participants. Salience and listener bets were uncorrelated (r = .19); the model combining measured salience with the parameter-free speaker matched listener means at r = .99 (.87 without 0/100 points), and the speaker model matched speaker bets at r = .98. This could have failed: listeners could have followed salience, or literal meaning alone, or a speaker who is not informative. Bergen, Goodman and Levy showed that interpretations of signals with no prior meaning follow the contrast with the available alternative (ceiling, Experiment 1) and the relative costs of alternatives (Experiment 2); and they showed by derivation that the basic recursion cannot produce the cost effect, so the observed Horn mapping is evidence for something beyond it, which lexical uncertainty supplies.

What this does not say

  • Not that comprehension in natural conversation computes this recursion. The games give the alternatives, use single features or novel symbols, and make costs explicit in money. How alternatives are fixed in real language is listed by the framework’s authors as open.
  • Not that speakers are informative. RSA models the listener’s assumption about speakers; production is overspecified and egocentric in ways the authors acknowledge it does not capture.
  • Not that one depth of recursion is universal. The evidence the review cites has most people at minimal depth and a minority deeper.
  • Not established beyond aggregated fits. The 2012 correlations are over condition means, several fixed at 0 or 100; a 2021 reanalysis (Sikos et al., not held, not read) addresses how robust they are.
  • Not that a form’s meaning is nothing but its contrasts. The alternatives’ literal meanings, iconic or conventional, are inputs; the contrast modulates them.

Connections

  • The word-learning precursor (THEORY-472; LIT-824, NOTE-627). Frank, Goodman, Lai and Tenenbaum (2009) derive the size-principle speaker this account’s listener inverts from informative communication, and test the inversion in word learning: given the referent, learners bet on a novel word’s meaning as the inverted speaker predicts. This account widens the same mechanism to inferring the referent. NOTE-627 finds that Experiment 1 adds support here in that neighbouring form, and that its evidence (r = .93 on condition means, one model, no competitor) is of the kind promote_when says cannot settle it.

Note of 2026-10-20: what the promotion keeps

The account is held to what the two studies measured. Frank and Goodman’s listener condition is a match of condition means to one model (r = .99) against a measured prior that does not predict the bets (r = .19); a likelihood-only listener is not reported, so the title does not claim that literal meaning alone fails. Bergen, Goodman and Levy’s Experiment 1 shows the dependence of a fixed form’s reading on its alternative at ceiling. Their Experiment 2, the Horn mapping, supports lexical uncertainty, which goes beyond the basic inversion, and stays in the body as such. Whether listeners compute the inversion, rather than follow the alternatives by a simpler route, is QUESTION-209.

Note of 2026-10-20: the wider claim filed

The promotion audit narrowed this entry to what was shown. Under ADR-047 v4 the claim the narrowing dropped, that listeners interpret utterances by inverting an informative speaker, as the mechanism and not only a fit to mean bets, is now filed as THEORY-785, Proposed, which extends this entry and takes it as evidence. This entry’s title and status are unchanged.

Backlinks