nucleation

LIT-824: Informative Communication in Word Production and Word Learning

3 min read

Properties4
StatusActive — read 2026-10-09 (NOTE-627); worth reading as the paper that first writes the informative speaker down: a speaker soft-maximises the utility −D_KL(M_S ‖ w̃), the information lost about her intended meaning when it is replaced by the uniform distribution over a word's extension, so for a known referent she picks a true word with probability ∝ |w|^(−α), the size principle derived from communicative premises. A learner who knows the referent inverts that speaker to infer what a novel word means; 700 web participants' bets track the model's predictions (r = .93). In a free-description task speakers' use of Flickr tag-words correlates only weakly with informativeness (r = .19 overall, .51 without colour terms, .02 for colour terms). It has no listener inferring a referent and no recursion: that step is Frank and Goodman 2012 (LIT-821).
Tagspragmatics, cognition, linguistics, probabilistic-modeling, information-theory
Cited as “paper” byNOTE-627 — Informative communication in word production and word learning
Cited as “source” byTHEORY-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), THEORY-816 — Learners inferring what a novel word means from its use reason about a speaker who chose the word to be informative about the referent, so the size principle in word learning is a consequence of that inference and not a separate rule of generalisation

Graph View

Michael C. Frank, Noah D. Goodman, Peter Lai and Joshua B. Tenenbaum (2009), in N. Taatgen and H. van Rijn (eds), Proceedings of the 31st Annual Conference of the Cognitive Science Society, Amsterdam, pp. 1228–1233.

Key takeaways

  • The informative speaker. Meanings are distributions over the objects in a context. A word’s literal meaning w̃_C is uniform over its extension. The speaker chooses w with P(w | M_S, C) ∝ exp(α U), where U = −D_KL(M_S ‖ w̃_C) + F, and F stands for other costs (set to 0 here). With α = 1 this is the Luce choice rule.
  • The size principle falls out. When the intended meaning is a single object, D_KL reduces to −log w̃_C(o_S), so P(w | M_S, C) ∝ |w|^(−α) over the true words (Eq. 10): a speaker prefers words that pick out smaller parts of the context. The authors offer this as a communicative derivation of Shepard’s and Tenenbaum and Griffiths’s size principle.
  • The learner inverts the speaker. A learner who knows the referent but not the lexicon updates P(L | w, M_S, C) ∝ P(w | L, M_S, C) P(L). With two features and two candidate lexicons, the probability that the word names feature f₁ is |f₁|⁻¹ / (|f₁|⁻¹ + |f₂|⁻¹) (Eq. 12): $67 on “red” in the one-red-circle display.
  • Experiment 1 (learning). 700 web participants split 50 when the two features are equally frequent, and the Figure 1 case gave 67.
  • Experiment 2 (production). 44 participants wrote descriptions of 50 of 304 Flickr “superball” photos, so that someone could pick each out. The probability of using a ball’s tag-word correlates with Eq. 10 at r = .19 overall. Without basic colour terms it is .51, and for colour terms alone .02. Adding a Google-Images frequency estimate of F barely helps (.22 and .52). The authors blame how well a tag fits the ball.
  • No recursion. The authors contrast their framework with game-theoretic pragmatics, which “rel[ies] on complex, recursive computations”. There is one speaker and one learner who inverts it, and no pragmatic listener who infers a referent.

Standing in the record

Filed on 2026-10-09 from the manuscript bibliography of 2026-10-09 (work what-survives-translation) at the owner’s request: a work the manuscript considered and dropped from its final reference list. See the curation entry of that day.

Read on 2026-10-09 (NOTE-627). It is the informative-speaker precursor of the rational-speech-act line the record already holds. The same authors’ 2009 Psychological Science paper (LIT-815) is a different work: a word learner that infers referential intentions, with no informative speaker. The listener who inverts this speaker to find a referent, with a salience prior, is Frank and Goodman 2012 (LIT-821). The lexical-uncertainty recursion is Bergen, Goodman and Levy 2012 (LIT-809), and the framework’s statement is Goodman and Frank 2016 (LIT-816). Experiment 1 adds to the evidence for THEORY-172 in its word-learning form, with the referent given and the lexicon inferred. Experiment 2 is the record’s only evidence on the speaker side, and it is weak.

Backlinks