PaperlayerView exact revision on arXiv

Source-linked AI summary

A model of rational interlocutors: Unification of comprehension and production

Hanlin Wu, Zhenguang G. Cai

arXiv:2609.32216v1cs.CL

TL;DR

Research on speaker modeling in comprehension and audience design in production has largely developed separately, despite both involving adjustment to a conversational partner. The paper proposes the rational interlocutor model, a unified account based on identity, fidelity, and knowledge. It treats comprehension and production as mirror-image choices and argues that perceived linguistic competence involves both fidelity and knowledge.

  • Problem

    Comprehension and production adjustments to conversational partners have largely been studied in separate research traditions.

    Source: §1 · p. 2

  • Method

    The RI model represents a partner through identity, fidelity, and knowledge, and treats comprehension and production as mirror-image choices over messages and utterances.

    Source: §1 · p. 4, §4.4 · p. 24, §4.5 · p. 27

  • Results

    Comprehenders discount less competent speakers’ utterances, while producers invest more effort for less competent partners.

    Source: §1 · p. 3

  • Takeaways & Limitations

    The model proposes that perceived linguistic competence comprises fidelity and knowledge, with contrasting profiles across L2 adults, children, and artificial partners yielding testable predictions.

    Source: §7 · p. 49

    The artificial-partner profile may become dated as beliefs about language models’ competence change.

    Source: §4.7 · p. 37

Abstract

from arXiv · show

Who we communicate with influences both our interpretation of their utterances and the design of our own. Such adjustment to the conversational partner is studied as speaker modeling in comprehension and as audience design in production, with the two literatures having developed largely separately. We argue that both adjustments express one rational computation and propose the rational interlocutor (RI) model, a computational account unifying comprehension and production. An interlocutor maintains a model of their partner, defined by three parameters: an identity parameter {\Pi} sets the messages and forms expected from the partner; a fidelity parameter {\Phi} sets how reliably messages and utterances map onto each other for them; a knowledge parameter {\Lambda} sets how knowledgeable the partner is believed to be. Comprehension and production are thus mirror-image modes of one computation over the partner model. Comprehension chooses the message the partner most likely intends to convey, weighing how well each candidate fits the utterance against how likely this partner is to mean it. Production chooses the utterance from which the partner will best recover the message, weighed against the effort of saying it. This explains why comprehenders appear to rely less on the forms produced by a linguistically less competent speaker, while producers tend to invest more effort in designing forms for them. We conjecture that perceived linguistic competence decomposes into two of these quantities: fidelity and knowledge. Their contrasting profiles across second-language (L2) adults, children, and artificial partners produce distinct and testable predictions.

1. Introduction

Listeners adjust how they interpret a speaker, while speakers adjust how they communicate with an addressee believed to be less linguistically competent. The paper proposes the rational interlocutor model to treat these adjustments as one computation over a partner model.

1. Introduction

  • Listeners discount less competent speakers’ utterances, tolerating errors and relying more on expectations about what was meant.Examples include recovering non-literal interpretations of implausible sentences and treating under-informative statements as more sensible when attributed to non-native speakers.Source: §1 · p. 3
  • Speakers invest more effort for less competent partners through greater alignment, fuller references, clearer articulation, and slower speech.The reported adjustments include children, non-native adults, and computer or voice-assistant partners.Source: §1 · p. 3
  • The RI model unifies comprehension and production as rational communication through an imperfect information channel, using identity, fidelity, and knowledge parameters.Identity sets expected meanings and forms; fidelity sets how reliably messages and utterances map onto one another; knowledge represents how knowledgeable the partner is believed to be.Source: §1 · p. 4
  • The paper conjectures that perceived linguistic competence decomposes into fidelity and knowledge, yielding distinct predictions for L2 adults, children, and artificial partners.The paper presents these contrasts as predictions to test, not as established results.Source: §1 · p. 4

2. Comprehension

Comprehenders use both speaker identity and their beliefs about a speaker’s forms and expected messages to interpret utterances. Across sentence repair, grammar, lexical access, content expectations, and pragmatic inference, the evidence shows that interpretation depends on the speaker model.

2.1 Sentence repair

  • Listeners more often recovered an implausible sentence’s non-literal meaning from an accented speaker, and their corrections reflected the source’s expected error type.The non-literal repair effect was found for alternatives one edit away, but not for active/passive alternatives requiring two edits.Source: §2.1 · p. 5

2.2 Morphosyntax

  • Listeners showed no P600 repair response to a typical non-native gender-agreement error, while speaker expectations selectively shaped responses to other agreement errors.The semantic-violation N400 did not differ reliably between native and non-native speakers in Hanulíková et al. (2012).Source: §2.2 · pp. 5–6

2.3 Lexical-semantic access

  • A speaker’s dialect biased which meaning listeners accessed for ambiguous words, and a child’s voice increased the cost of switching from an established label to a synonym.Listeners also shifted dialect-linked meaning access toward whichever interlocutor was currently speaking.Source: §2.3 · pp. 6–7

2.4 Content expectations

  • Speaker identity and mismatch type shaped responses to anomalous content, including whether listeners showed a late positivity or an N400.Stereotype violations elicited an N400, whereas biological violations elicited a late positivity in Wu and Cai (2026b).Source: §2.4 · p. 8

2.5 Pragmatic inference

  • Listeners withheld pragmatic inferences when a speaker’s behavior or identity made informative word choice less expected, including for under-informative statements and disfluencies.They also generalized selectively: exposure to a speaker’s under-informativeness shaped later judgments, but exposure to over-modification did not produce the same generalization.Source: §2.5 · pp. 9–10

2. Comprehension

  • Across comprehension, speaker identity sets expected meanings and forms, while lower trust in a speaker’s forms shifts listeners toward their own expectations.The paper distinguishes this reliance pattern from cases where identity establishes which form a speaker is expected to use.Source: §2.5 · pp. 10–11

3. Production

Speakers tailor what they say and how they say it to their addressee. The production evidence spans alignment, referential choices, and acoustic adjustments, with the paper interpreting these as investment for partners believed less able to understand.

3.1 Alignment

  • Speakers aligned more with child and non-native partners than with native adults, and adopted a partner’s perspective more when communication required matching.In the naming task, increased alignment was not attributable to memorability of recorded names or speech rate; some learner-to-learner typed dialogue showed the opposite pattern.Source: §3.1 · pp. 12–13

3.2 Referential choice

  • Speakers used more explicit references for younger children and adult L2 learners, and lexical choices tracked a partner’s dialect even without accent cues in the word itself.In one computer-partner experiment, framing affected how much detail speakers added, but not their alignment with the partner’s redundancy level.Source: §3.2 · pp. 13–14

3.3 Acoustic register

  • Speakers hyperarticulated vowels for infants and foreigners, while voice-assistant-directed speech also differed in duration and pitch from speech to a human.After a staged recognition failure, children raised pitch further; the paper interprets the adjustments as tracking perceived linguistic needs rather than affect alone.Source: §3.3 · p. 15

3. Production

  • Taken together, production adjustments include more explicit reference, greater alignment, and clearer articulation for partners believed less able to recover the message.The paper characterizes these manipulations mainly as operating on the decoding side of communication.Source: §3.3 · p. 15

4. A model of rational interlocutors

This section presents the rational interlocutor model as a unified account of how beliefs about a partner shape comprehension and production through opposite directions of one computation. It distinguishes partner identity, perceived fidelity, and knowledge, and predicts that these attributes affect how utterances are interpreted and designed.

4.1 Comprehension as inference over a noisy channel

  • Comprehension infers the intended message by weighing how well it could have produced the utterance against how likely this partner was to mean it.The channel and message prior depend on partner identity and, respectively, fidelity and knowledge; implausible messages can be resisted even when they fit the utterance.Source: §4.1 · pp. 18–19

4.2 Fidelity and the channel

  • Lower fidelity flattens the channel, making the utterance less discriminative and causing comprehension to rely more on the message prior.Identity sets each message’s canonical form, while fidelity sets how reliably utterances keep to it.Source: §4.2 · p. 20

4.3 Knowledge and the message prior

  • Knowledge (Λ) sets how strongly the message prior favors messages typical for the partner’s identity: a nonsensical remark surprises more when the partner is believed to be knowledgeable.The identity determines what is typical; knowledge determines how tightly the partner’s intended messages are expected to follow it.Source: §4.3 · pp. 23–24

4.4 The partner model: identity, fidelity, and knowledge

  • The model represents a partner through identity (Π), which sets typical messages and canonical forms; fidelity (Φ), which sets mapping reliability; and knowledge (Λ), which sets how closely intended messages fit what is typical.Identity also shapes which forms are appropriate for the partner. The same three beliefs inform comprehension and production.Source: §4.4 · p. 24
  • Fidelity and knowledge contribute distinct, additive terms to comprehension: fidelity controls utterance-based discrimination, while knowledge controls the prior’s pull toward typical messages.The partner’s identity shapes both terms. Knowledge gates identity effects in the message prior, whereas identity effects through the channel can remain when knowledge is zero; fidelity and knowledge manipulations are predicted to be additive.Source: §4.4 · pp. 25–27

4.5 Production as the mirror image of comprehension

  • Production reverses comprehension: holding the message fixed, the agent chooses an utterance that helps the partner recover it while accounting for the utterance’s cost.Recovery depends on the channel and message prior, while production cost combines the agent’s standing preference for fluent forms with forms typical for the partner’s identity.Source: §4.5 · pp. 27–28
  • The production prior represents identity-based conventions rather than channel reliability, keeping utterance cost distinct from recoverability and preserving the design-for-recovery trade-off.The model does not represent the earlier choice of what to mean, though the authors propose that the same partner model could govern that stage.Source: §4.5 · pp. 28, 29, 31
  • Production chooses utterances by balancing a noisy-channel partner’s probability of recovering the intended message against the utterance’s cost, so low fidelity favors more explicit forms.The imagined noisy-channel listener is the same distribution used in comprehension, with partner-specific rather than physical-channel noise.Source: §4.5 · pp. 29–31

4.6 One model, two goals: symmetry and its limits

  • The model unifies comprehension and production as one computation run in opposite directions, with agents alternating between modes while using the same partner model.The modes differ in what is held fixed, the channel’s direction, and the range of the prior term; dialogue should update the partner model from both what is heard and whether one’s own messages were understood.Source: §4.6 · p. 32
  • As perceived fidelity falls, comprehension discounting grows steadily, but production investment rises only until added descriptors no longer justify their effort.The reversal depends on descriptor cost and alternative confusability; sufficiently high costs can eliminate investment altogether.Source: Figure 3 · p. 33, §4.6 · pp. 33–34

4.7 Profiles of three partner types: L2 adults, children, and artificial partners

  • Perceived linguistic competence is conjectured to bundle fidelity and knowledge, which can vary independently, while identity separately sets expected messages and forms.The profiles assign L2 adults lower fidelity with high knowledge, children lower fidelity and knowledge, and artificial partners high fidelity with low knowledge.Source: §4.7 · pp. 34–35
  • Lower fidelity weakens pragmatic inferences from a partner’s forms, while lower knowledge lets incongruous content be explained without attributing an intention.Together these quantities determine how far a partner’s form can be trusted to have been chosen deliberately.Source: §4.7 · pp. 37–38
  • The model’s schematic profiles predict reduced form response but preserved content response for L2 adults, reduced responses to both for children, and enhanced form but reduced content response for artificial partners.The artificial profile’s reduced content response is predicted to diminish as estimates of artificial partners’ knowledge rise; the profiles are schematic rather than empirical measurements.Source: §4.7 · pp. 35, 37, 38, Figure 4 · p. 38

5. Acquiring and using the partner model

The section describes how partner models are acquired and updated, when they are used, and how strongly agents rely on them. It proposes that partner weight ω varies by agent and shapes adjustments in both comprehension and production.

5.1 Acquiring and updating a partner model

  • A partner model begins with category-based expectations, is refined by individuating speech evidence, and can incorporate a partner’s record of accuracy.The RI model itself holds its three parameters fixed; updating is inherited from other accounts.Source: §5.1 · pp. 39–40
  • Children detect grammatical errors from native and foreign-accented speakers, but give those errors less weight when judging a foreign-accented speaker’s reliability.Four-year-olds also simplify explanations for younger children, suggesting adjustment tracks listener capacity rather than a simple adult-child division.Source: §5.1 · pp. 40–41

5.2 When the partner model is used

  • Partner models are more likely to be consulted when beliefs are clear, candidate readings are close, evidence departs from ordinary behavior, and processing resources are available.The partner belief can change the winning interpretation when competing readings have comparable costs.Source: §5.2 · p. 41

5.3 To what extent the partner model is used

  • Partner weight ω is an agent-side property that scales reliance on the partner model, unlike identity, fidelity, and knowledge, which are beliefs about the partner.Across partners, ω scales effects from any of the three parameters; within a partner, it can be indistinguishable from a fidelity shift under the stated implementation.Source: §5.3 · p. 42
  • As ω rises, comprehension for the low-fidelity partner changes smoothly, while production adds descriptors only after a threshold; high-fidelity curves remain flat.At ω=0, the two partners produce identical behavior.Source: §5.3 · p. 42, Figure 5 · p. 43
  • Individual differences in working memory, executive control, empathy, and openness are associated with variation in partner-sensitive behavior; the authors interpret autism findings as potentially reflecting lower ω.The cited autism findings indicate partner mismatches were detected, while that information was weighted less when not explicitly attended to.Source: §5.3 · pp. 43–44

6. Situating and testing the RI model

The paper situates the RI model among accounts of partner-sensitive comprehension and production, then derives four directional predictions. These predictions concern dissociations between partner parameters, limits on adjustment, and transfer between modes.

6.1 Relation to previous accounts

  • The RI model turns the proposal that one partner model serves comprehension and production into one computation with identity, fidelity, knowledge, and partner-weight parameters.It decomposes perceived linguistic competence into fidelity and knowledge, unlike earlier accounts that treat it as one attribute.Source: §6.1 · p. 45
  • Unlike noisy-channel comprehension and rational-speech-act production considered separately, the RI model uses one partner-indexed fidelity in both modes.The account specifies fidelity and knowledge as distinct graded parameters and treats memory-based mechanisms as compatible implementation hypotheses.Source: §6.1 · pp. 45–46

6.2 Predictions

  • Lower perceived fidelity is predicted to reduce the cost and repair of form violations, while leaving content-anomaly costs comparatively unchanged after general accent costs are set aside.Identity manipulations should shift expected forms, with weaker effects as perceived fidelity falls.Source: §6.2 · pp. 47–48
  • Perceived fidelity and knowledge are predicted to be independently manipulable: expertise descriptions should mainly affect content anomalies, while accent descriptions should mainly affect form violations.Partner types’ fidelity and knowledge profiles could be rated independently of the responses they predict.Source: §6.2 · p. 48
  • Comprehension adjustment is predicted to approach a plausibility-determined ceiling as fidelity falls, while production explicitness is predicted to rise and then decline as added effort ceases to pay.The predicted production pattern is an inverted U, not a monotonic increase.Source: §6.2 · p. 48
  • A change in perceived fidelity is predicted to transfer across comprehension and production for the same partner; absent transfer would favor separate encoding and decoding fidelities.Partner weight estimated in one mode should also tend to predict adjustment in the other.Source: §6.2 · p. 48

7. Conclusion

The conclusion presents the RI model as one rational policy for partner adjustments in comprehension and production. It separates partner identity, fidelity, and knowledge, and predicts that L2 adults, children, and artificial partners can yield distinct adjustment patterns.

7. Conclusion

  • The RI model treats comprehension and production as one rational policy over a partner model, with identity setting expectations, fidelity governing message-utterance mapping, and knowledge shaping expected messages.Comprehension balances utterance fit against message expectations; production balances partner recovery against utterance effort.Source: §7 · pp. 48–49
  • The paper argues that linguistic competence decomposes into fidelity and knowledge, so L2 adults, children, and artificial partners need not form one ordered scale.The model predicts that these partner types can sometimes produce opposite response patterns.Source: §7 · p. 49

S1. Notation conventions

The notation distinguishes messages, utterances, probabilities, and functions, while Greek capitals mark beliefs about a partner and Greek lowercase letters mark agent or implementation properties.

S1. Notation conventions

  • Lowercase roman letters denote values such as messages, utterances, descriptor counts, and probabilities or densities, with specified exceptions for surprisal and the canonical-form function.Subscripted forms identify particular candidate messages and utterances.Source: S1. Notation conventions · p. 61
  • Uppercase roman letters name functions, except L0 and subscript S, which label the literal listener and speaker in the RSA model.Source: S1. Notation conventions · p. 61
  • Uppercase Greek letters Π, Φ, and Λ denote the agent’s beliefs about the partner; lowercase Greek letters α, ω, κ, and σ denote agent or implementation properties.Source: S1. Notation conventions · p. 61
  • Table S1 lists equations by kind and their implemented values; conceptual equations state the account, while implementation equations fix the forms and values used for figures.All quantities plotted in Figures 1–5 derive from the implementation equations.Source: Table S1 · p. 61, S1. Notation conventions · p. 62
Loading 2609.32216v1…