Contribution
A review, but also the paper that fixes the RSA framework’s canonical form and names its extension, uRSA. It argues that one utility-theoretic version of Grice’s cooperative principle, with Bayesian inference about the speaker, replaces the maxims and makes quantitative predictions, and it gathers the evidence for that across reference games, implicature and non-literal language.
Key insight
The listener does not decode a meaning and then repair it when a maxim is broken; every interpretation is inference to the best explanation of why the speaker said this, among what she could have said. Uncertainty about the speaker herself, her topic, her knowledge, even her lexicon, enters the same posterior, and that is what produces the non-literal readings.
Assumptions
- A literal semantics ⟦u⟧ assigning true or false to each world; the literal listener conditions the prior on it.
- An approximately rational speaker: softmax over a given set of alternative utterances with rationality α.
- A shared prior P(w) over worlds, and in uRSA a prior P(s) over speaker types.
- Bounded recursion: the standard form is listener → speaker → literal listener.
- Computational level only, in Marr’s sense; no algorithmic claim.
Key results
The review’s own formal content:
- RSA: P_L(w|u) ∝ P_S(u|w)P(w); P_S(u|w) ∝ exp(αU(u; w)); U(u; w) = log P_Lit(w|u); P_Lit(w|u) ∝ δ_⟦u⟧(w)P(w).
- Worked example (Fig. 1): faces with hat and glasses, glasses only, neither; utterances “glasses”, “hat”. “Glasses” is read as the glasses-only face, because a speaker meaning the other face would have said “hat”.
- Utility refinements: cost (written in the Box as U = log P_Lit(w|u) + cost(u); the sign must be negative for cost to disfavour an utterance, as in LIT-809’s Eq. 3); expected utility under speaker knowledge k, U(u; k) = E_{P(w|k)}[U(u; w)]; topic relevance by a projection t, U(u; w, t) = log Σ_{w′: t(w′)=t(w)} P_Lit(w′|u); social utilities such as kindness.
- uRSA: P_L(w, s|u) ∝ P_S(u|w, s)P(s)P(w).
The empirical results it reports, by its account: reference-game fits and a replication; RSA fits to spatial-language games with noisier fits; sensitivity to message cost in dollars and to typing difficulty; scalar implicature modulated by the speaker’s access; hyperbole, irony and metaphor accounting for “almost all of the explainable variance”; deeper-than-minimal recursion in about 15% of participants in one study.
Claims
| id | claim | strength | support |
|---|---|---|---|
| C1 | Basic RSA predicts graded judgements in simple signalling games across samples and stimuli | moderate | cited studies [24], [66], [13]; not shown here |
| C2 | uRSA, by joint inference over the speaker’s topic, accounts for hyperbole, irony and metaphor | moderate | cited studies [45], [44], [43]; not shown here |
| C3 | Uncertainty over word meanings accounts for vagueness phenomena and for embedded implicatures that basic RSA misses | moderate | cited [51], [65], [9] |
| C4 | Most comprehenders use minimal recursion; a minority go deeper | weak | mixed cited evidence, the authors’ summary |
| C5 | RSA replaces the maxims with one utility-theoretic cooperative principle | — | a framing claim, not an empirical one |
Concepts
- RSA: a class of models in which comprehension arises from recursive reasoning about what speakers would have said, given communicative goals.
- uRSA: RSA with joint inference over the speaker’s intended meaning and other aspects of the interaction (topic, context, word meanings).
- literal listener: the base case that conditions the prior on the utterance’s truth.
- topic / QUD: a function t from worlds to the part of the world the speaker is trying to convey.
- social recursion: speaker and listener modelling each other to some depth.
- implicature, scalar implicature, conversational maxims: as in the glossary; standard.
Connections
Grice (and Lewis’s signalling games) supply the idea; game-theoretic pragmatics (Benz, Jäger, van Rooij) and natural language generation the tools; Bayesian cognitive modelling the inference. It connects RSA to the question under discussion, to noisy-channel comprehension, to compositional semantics à la Montague, to grammatical theories of alternatives, and to neural speaker–listener models in NLP (Andreas and Klein). The language-change box links in-the-moment pragmatics to iterated learning (Kirby, Cornish and Smith 2008, which the record holds as LIT-771; and Kirby, Tamariz, Cornish and Smith 2015, where a communicative pressure that the review says “can be modeled via RSA” offsets learnability) and to typological efficiency (Regier, Kay and Khetarpal on colour; Kemp and Regier on kinship; Xu and Regier on numerals).
Bearing on the record
- THEORY-172 takes its formal statement from here and its evidence from LIT-821 and LIT-809, read directly.
- The uRSA posterior P_L(w, s|u) is a joint distribution over the world and the circumstances of the utterance (speaker’s topic, knowledge, lexicon) given what was said. The record has no other worked model of an utterance supporting a distribution over the situation of its saying.
- Topic-relative utility makes informativeness relative to a projection of the world, so a speaker can be maximally informative about one question and uninformative about others; this is the framework’s own version of fidelity relative to a question.
- LIT-769 and THEORY-155. The language-change box cites iterated learning and says whether typological distributions arise from iterated learning with RSA-like users is open. That is a question, not a result.
- No instruction for machine-learning practice; nothing for the anthology. It mentions neural speaker–listener models only as an application.
Limitations
- A review: its empirical claims are summaries of other papers, several of them conference papers or then in press.
- The set of alternative utterances is given, not derived; the authors list this as open.
- The fits are of listeners’ aggregate judgements; speakers’ actual production (egocentric, overspecifying) is acknowledged as not well captured.
- Computational cost grows with worlds and utterances; no algorithmic account.
- This reading is of the submitted manuscript, not the published text.
Open questions
Its own Box: depth of recursion; how alternatives are computed; how informativeness relates to social goals; dialogue over many turns; pragmatics and language change; how it is computed quickly and at scale.