Contribution
Gricean, relevance-theoretic and common-ground accounts explained communicative inference informally. This paper gives a model with no fitted parameters that predicts graded human judgements in a controlled reference game: an informative speaker, defined from surprisal under a literal listener, and a Bayesian listener who inverts that speaker with a measured prior. It separates the two ingredients experimentally, by collecting speaker, salience and listener judgements in different groups.
Key insight
Interpretation is inference about why the speaker chose this word rather than another word that was also true of the referent. If the speaker prefers words that pick out fewer objects, a word true of several objects is evidence for the one that has no more specific word available.
Assumptions
- Shared context and vocabulary. C = {o₁, …, o_n}; each word is a Boolean function on objects, shared by speaker and listener.
- Rational speaker. P(w|r_S, C) ∝ exp(α U(w; r_S, C)) with U = I(w; r_S, C) − D(w); α = 1 (Luce choice); cost D(w) constant, because the words are randomized and roughly matched.
- Informativeness as negative surprisal under the literal listener w̃_C, uniform over objects the word is true of: I = log w̃_C(r_S) = −log|w|.
- Prior as salience, measured empirically by asking a separate group which object a speaker using an unknown word meant.
- One step of recursion: the listener reasons about a speaker who reasons about a literal listener.
Key results
- Eq. 1. P(r_S|w, C) = P(w|r_S, C)P(r_S) / Σ_{r′∈C} P(w|r′, C)P(r′).
- Eq. 2 / S4. P(w|r_S, C) = |w|⁻¹ / Σ_{w′ true of r_S} |w′|⁻¹, the size principle of Xu and Tenenbaum. Checked: exp(−(−log|w|⁻¹)) = |w|⁻¹.
- Design. Three objects, each with a colour, shape and texture; two dimensions vary, one is constant; the number of objects sharing each of the target’s two values (1, 2 or 3) gives seven context types (two distinct 2/2 types). 50 random trials per numerical condition per group; one trial per participant after a manipulation check.
- Fits. Speaker r = .98 (p < .001); listener r = .99 (p < .0001), r = .87 without the 0 and 100 predictions; salience vs listener r = .19 (p = .40).
Claims
| id | claim | strength | support |
|---|---|---|---|
| C1 | Speakers’ word choices in these games follow choice in proportion to specificity | strong (for these games) | speaker condition, r = .98 over context-type means |
| C2 | Listeners’ interpretations follow the Bayesian inversion of that speaker with a measured salience prior | strong (for these games) | listener condition, r = .99; r = .87 excluding 0/100 |
| C3 | Neither prior nor likelihood alone explains listeners; their combination does | moderate | salience–listener r = .19; a likelihood-only comparison is not reported |
| C4 | Information-theoretic informativeness plus measured common knowledge can capture pragmatic inference in context generally | weak | the authors’ concluding suggestion; one game type |
Method
Model derivation in the Supplementary Text: speaker choice by softmax of utility (S1), utility as informativeness minus cost (S2), literal listener (S3), giving S4, which reduces to Eq. 2. Listener predictions multiply measured salience by S4 and normalize. Participants bet $100 across options; mean bets per term or object per context type are correlated with model predictions.
Concepts
- contextual salience: the prior probability that an object would be referred to, standing for perceptual, social and conversational common ground; measured, not computed.
- informativeness: negative surprisal of the intended referent under the literal interpretation of the word.
- size principle: a word’s likelihood is inversely proportional to the number of objects it applies to.
- language game: Wittgenstein’s term, used here for one-shot reference games.
Connections
The authors place it after Rosenberg and Cohen’s disambiguation models, game-theoretic signalling models (Benz, Jäger and van Rooij), and Dale and Reiter’s generation of referring expressions; the size principle is Xu and Tenenbaum’s. The general framework it starts is surveyed in Goodman and Frank (LIT-816) as the rational speech act model, and Bergen, Goodman and Levy (LIT-809) extend the same recursion with costs and lexical uncertainty.
Bearing on the record
- Produces, with LIT-809 and LIT-816, THEORY-172: that interpretation in reference games is predicted by inverting an informative speaker, so it depends on the alternatives the speaker had.
- The listener’s output is a posterior over referents given an utterance and a context: a distribution over what was meant, not a single meaning. That is the formal object the record would need for any account of a text as supporting a distribution over situations.
- No instruction for machine-learning practice; nothing for the anthology.
Limitations
- Correlations are over a small number of aggregated points (terms and objects by context type), several at 0 or 100 by construction; the r = .87 with those removed is the more telling number.
- Speaker, salience and listener judgements come from different people, and no pair communicates; the listener is asked to imagine a speaker.
- The words are single features of artificial objects, chosen so that cost can be taken as constant; no natural-language alternatives.
- α = 1 is set, not tested; the absence of free parameters depends on it and on the measured prior.
- No comparison model (likelihood only, literal listener with prior) is reported.
Open questions
- Does the fit hold for listeners who must infer the speaker’s alternatives rather than being shown them?
- Does it hold with deeper recursion or a fitted α? Goodman and Frank (LIT-816) report that deeper recursion is seen in only some participants.
- The 2021 reanalysis by Sikos et al. (not held, not read) bears on how robust the fit is.