nucleation

LIT-821: Predicting Pragmatic Reasoning in Language Games

2 min read

Properties4
StatusActive — read 2026-10-09 (NOTE-634); worth reading as the one-page origin of the rational speech act model's quantitative test: a listener who inverts, by Bayes' rule, a speaker who chooses words in proportion to their informativeness (exp of minus surprisal under a literal listener, which reduces to the size principle |w|⁻¹), combined with an empirically measured salience prior, predicts mean listener bets in three-object reference games at r = .99 with no fitted parameters (α set to 1). Read it knowing that the fit is to means over seven context types, that the prior is measured in a separate group, and that the speaker and listener groups never interact. Source of THEORY-172.
Tagspragmatics, linguistics, cognition, probabilistic-modeling
Cited as “paper” byNOTE-634 — Predicting pragmatic reasoning in language games
Cited as “source” byTHEORY-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, THEORY-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

Graph View

Michael C. Frank and Noah D. Goodman (2012), Science 336(6084):998 — DOI-10.1126/science.1218633

Key takeaways

  • The listener. P(r_S|w, C) ∝ P(w|r_S, C) P(r_S): the prior is contextual salience, measured in a separate group who bet on the referent of an unknown word; the likelihood is a model of the speaker.
  • The speaker. P(w|r_S, C) ∝ exp(α U), U = log w̃_C(r_S) − D(w), with w̃_C the literal listener (uniform over the objects the word is true of). With α = 1 and constant cost this is |w|⁻¹ normalized over the true words: the size principle. So the model has no free parameters.
  • The data. 745 participants (206 speaker, 276 salience, 263 listener), one trial each, three objects varying on two of colour, shape and texture, seven context types.
  • The result. Speaker bets against the model r = .98; listener bets against prior × likelihood r = .99, and r = .87 with predictions of 0 and 100 removed; salience and listener bets themselves r = .19.
  • The framing. Grice, Sperber and Wilson, and Clark give informal accounts; the contribution is a quantitative prediction from an information-theoretic informativeness plus measured common ground.

Standing in the record

Filed on 2026-10-09 at the owner’s request, from the manuscript bibliography of 2026-10-09 (work what-survives-translation): one of the works the manuscript considered and dropped from its final reference list.

Read on 2026-10-09 (NOTE-634); with LIT-809 and LIT-816 it is the source of THEORY-172. A later reanalysis exists (Sikos, Venhuizen, Drenhaus and Crocker 2021, “Reevaluating pragmatic reasoning in language games”, PLOS ONE, DOI-10.1371/journal.pone.0248388); the record does not hold it and it was not read.

Backlinks