CoChat

The role of pragmatic mechanisms in referential communication and categorization: An emergent communication model

Open access
Kobrock, Kristina, Ohmer, Xenia, Bruni, Elia, Gotzner, Nicole · 2026 · PLOS Computational Biology

The DOI matches its Crossref record; 76 of the 107 references it cites were found in a registry or an index, and 22 are not checked yet; 42 of the 120 sources its claims cite were checked for support.

Abstract

We model pragmatic mechanisms of referential communication and categorization in a multi-agent framework of emergent communication. Pragmatic theories and experimental work predict that speakers consider the context in their choice of referring expressions. In addition to this context-based reasoning, utility-based pragmatic reasoning about the listener’s likely interpretation of an utterance influences the speaker’s production choices. We aim to investigate these two factors and their role in referring expression generation and categorization in a computational model of language emergence and language use. We model communication in interaction and consider an efficiency tradeoff between speaker and listener utilities. Our results show that an emerging language becomes more effective, ambiguous, and efficient when speakers and listeners communicate in a shared context. This is achieved by an efficient tradeoff between production and comprehension where languages can afford to be simpler in production when they are sufficiently informative in context, placing more burden on the listener’s side. We further demonstrate that incorporating utility-based pragmatics, as modeled with the Rational Speech Acts framework, improves the linguistic efficiency of language use only in languages that emerged with a shared context between interlocutors, but not in languages that emerged without such contextual information during training. We conclude that context-based pragmatics plays a role in referential communication and categorization by shaping an emerging language. Efficient reference in a communicative situation can benefit especially from utility-based pragmatics if the language that is being used has emerged in context. This might suggest that utility-based pragmatics hinges on mechanisms that naturally emerge when context is available during the evolution of a language. In summary, we show that human-like language and category systems emerge as an optimized tradeoff between speaker and listener needs in interaction, i.e., under efficiency considerations of simplicity and informativeness.

Verified 2026-11-19 · 22 references not checked yet

Overview

CoChat Summary

TL;DR

In an emergent communication model with concept-level reference games, a shared context between speaker and listener produced more ambiguous, shorter, and more efficient languages that generalized better. Adding Rational Speech Acts (RSA) utility-based pragmatics improved efficiency mainly for languages that had emerged with shared context, not for context-unaware languages.

Key findings

  • Context-aware agents achieved higher validation accuracy than context-unaware agents (validation accuracies ≥0.96 vs. ≥0.86; main effect of condition M = −0.03, CrI [−0.04, −0.02], pd = 100%, 0% in ROPE). Test accuracies were also higher in the context-aware condition across datasets (e.g., D(3,4): 0.93 ± 0.02 vs. 0.79 ± 0.02).
  • Context-aware languages were shorter and more ambiguous: messages in the context-unaware condition tended to be at least six symbols longer, and entropy-based scores were higher in the context-unaware condition for NMI (M = 0.14, CrI [0.13, 0.16]), effectiveness (M = 0.22, CrI [0.20, 0.25]), and consistency (M = 0.04, CrI [0.03, 0.05]), all pd = 100%, 0% in ROPE.
  • In Experiment 2, RSA shortened messages and improved lexicon-level efficiency especially for context-aware agents: context-unaware generic-concept messages averaged 10.5 symbols and specific-concept messages 17.2 symbols; with RSA these dropped to 5.5 and 10. Context-aware messages averaged 3.6 symbols for generic and 3.9 for specific concepts, with RSA reducing lengths by about 1 symbol.
  • RSA did not improve test accuracy for context-unaware agents and sometimes worsened it (e.g., D(3,16): 0.95 ± 0.04 without RSA vs. 0.72 ± 0.03 with RSA), while context-aware + RSA remained high (e.g., D(3,16): 0.96 ± 0.07 vs. 0.95 ± 0.09). RSA reduced lexicon sizes mainly for context-aware agents (interaction M = −41.42, CrI [−58.62, −24.36], pd = 100%, 2.00% in ROPE), and message distributions became more Zipf-like as more pragmatic mechanisms were available.

Methods

  • Multi-agent emergent communication model using a concept-level reference game: a speaker sends a symbolic message about a target concept, and a listener labels objects as target or distractor. Agents were implemented with feed-forward layers and GRUs (hidden size 128), trained with batch size 16, learning rate 0.001, Gumbel-Softmax, a message-length cost, and early stopping in the EGG framework.
  • Six symbolic datasets D(n,k) with 3–5 attributes taking 4–16 values; concepts ranged from generic to specific, and contexts ranged from coarse to fine. Data were split 60% train, 20% validation, 20% test; game size was 10 targets and 10 distractors, with scaling factor 10.
  • Experiment 1 compared context-unaware speakers with context-aware speakers sharing context with listeners. Experiment 2 used trained models at inference and compared context-unaware, context-unaware + RSA, context-aware, and context-aware + RSA speakers. RSA was a level-1 speaker using an internal listener model and a length cost. Results are means and standard deviations over five runs; Bayesian hierarchical models were fit in R with brms and bayestestR (2000 iterations, 1000 warm-up; all Rhat < 1.01, ESS > 1000).

Limitations

  • Stated in the paper: RSA reasoning was implemented only for the speaker at inference, not for the listener, because listener-side inference over possible world states would be computationally costly. RSA agents could only use messages that had been produced during training, so they could not generate novel shorter messages even when more efficient. The authors also state they cannot draw inferences about listener-side processing costs from these simulations.
  • Stated in the paper: the RSA framework presupposes a small fixed set of possible messages, which is cognitively implausible and intractable in full form; the model approximates this by using training utterances.
  • Evident from the text: the findings come from symbolic toy datasets and artificial neural-network agents, not human language users, and are based on five runs per condition across six controlled datasets. The context-aware condition uses a specific shared-context manipulation in which speaker and listener distractors differ from targets in the same concept-defining attributes, so generalization to richer natural communicative settings is not directly tested.

Model-written summary. Check it against the paper before relying on it.

Integrity signals

Registry

Crossref · 10.1371/journal.pcbi.1014326

Cites

76 of 107 references verified0 citation errors · 9 not checkable

Statistics

Nothing to recompute.

Figures

Fig 1 Model architecture. In this example, the speaker has to communicate the concept of a blue circle and the listener has to identify all blue circles. The loss function rewards shorter messages from the speaker (efficiency) and correct selections from the listener (informativeness). This figure was created by the first author.
Fig 1 Model architecture. In this example, the speaker has to communicate the concept of a blue circle and the listener has to identify all blue circles. The loss function rewards shorter messages from the speaker (efficiency) and correct selections from the listener (informativeness). This figure was created by the first author.CC BY · PubMed Central
Fig 2 Game scenarios in Experiment 1. A: Context-unaware - The speaker agent only observes the set of target objects, i.e., the target concept (displayed in a green frame for visualization purposes). B: Context-aware - The speaker agent observes the target concept in a context, i.e., a set of target objects (visualized in blue) and a set of distractor objects (visualized in green). This figure was created by the first author.
Fig 2 Game scenarios in Experiment 1. A: Context-unaware - The speaker agent only observes the set of target objects, i.e., the target concept (displayed in a green frame for visualization purposes). B: Context-aware - The speaker agent observes the target concept in a context, i.e., a set of target objects (visualized in blue) and a set of distractor objects (visualized in green). This figure was created by the first author.CC BY · PubMed Central
Fig 3 Game scenarios added in Experiment 2. A: Context-unaware + RSA - Speaker agents trained in the context-unaware condition reason about the listener’s likely interpretation of their message. B: Context-aware + RSA - Speaker agents trained in the context-aware condition reason about the listener’s likely interpretation of their message. This figure was created by the first author.
Fig 3 Game scenarios added in Experiment 2. A: Context-unaware + RSA - Speaker agents trained in the context-unaware condition reason about the listener’s likely interpretation of their message. B: Context-aware + RSA - Speaker agents trained in the context-aware condition reason about the listener’s likely interpretation of their message. This figure was created by the first author.CC BY · PubMed Central
Fig 4 Message lengths per concept hierarchy level. Concept specificity, or the amount of information that needs to be communicated, increases with the number of fixed attributes that is shared among target objects from more generic concepts (with fewer fixed attributes) on the left, to more specific concepts (with more fixed attributes) on the right.
Fig 4 Message lengths per concept hierarchy level. Concept specificity, or the amount of information that needs to be communicated, increases with the number of fixed attributes that is shared among target objects from more generic concepts (with fewer fixed attributes) on the left, to more specific concepts (with more fixed attributes) on the right.CC BY · PubMed Central
Fig 5 NMI, consistency and effectiveness scores for each level of the conceptual hierarchy.
Fig 5 NMI, consistency and effectiveness scores for each level of the conceptual hierarchy.CC BY · PubMed Central
Fig 6 Distribution of message lengths for different levels of the conceptual hierarchy.
Fig 6 Distribution of message lengths for different levels of the conceptual hierarchy.CC BY · PubMed Central
Fig 7 Tradeoff between lexicon size and informativeness. The lexicon size is normalized by the number of concepts in a dataset.
Fig 7 Tradeoff between lexicon size and informativeness. The lexicon size is normalized by the number of concepts in a dataset.CC BY · PubMed Central
Fig 8 Zipf’s law like distribution of message frequency plotted for dataset D(4,4). D(4,4) is a medium-sized dataset where objects consist of four attributes which can each take four different values.
Fig 8 Zipf’s law like distribution of message frequency plotted for dataset D(4,4). D(4,4) is a medium-sized dataset where objects consist of four attributes which can each take four different values.CC BY · PubMed Central
Fig 9 Zipf’s law like distribution of message length plotted for dataset D(4,4).
Fig 9 Zipf’s law like distribution of message length plotted for dataset D(4,4).CC BY · PubMed Central
Fig 10 Dataset examples. The three objects in the top row are the target objects that form a target concept together. The three objects in the bottom row are the distractor objects that form the context. A: An example for a specific concept LARGE BLUE CIRCLE in a fine context (two attributes shared). B: An example for a specific concept LARGE BLUE CIRCLE in a coarse context (no attribute shared). C: An example for a generic concept CIRCLE in a coarse context.
Fig 10 Dataset examples. The three objects in the top row are the target objects that form a target concept together. The three objects in the bottom row are the distractor objects that form the context. A: An example for a specific concept LARGE BLUE CIRCLE in a fine context (two attributes shared). B: An example for a specific concept LARGE BLUE CIRCLE in a coarse context (no attribute shared). C: An example for a generic concept CIRCLE in a coarse context.CC BY · PubMed Central
Fig 11 Information-theoretic scores and ambiguity in language. The set of concepts C is the red circle on the left (red and purple), the set of messages M is the full blue circle on the right (blue and purple). The mutual information I(C,M) (only purple) captures one-to-one mappings between messages and concepts. The conditional entropy H(M|C) (only red) captures many-to-one mappings between messages and concepts. The conditional entropy H(C|M) (only blue) captures one-to-many mappings between messages and concepts.
Fig 11 Information-theoretic scores and ambiguity in language. The set of concepts C is the red circle on the left (red and purple), the set of messages M is the full blue circle on the right (blue and purple). The mutual information I(C,M) (only purple) captures one-to-one mappings between messages and concepts. The conditional entropy H(M|C) (only red) captures many-to-one mappings between messages and concepts. The conditional entropy H(C|M) (only blue) captures one-to-many mappings between messages and concepts.CC BY · PubMed Central

Verification

Claims the paper makes

24 Valid · 2 Not valid · 36 Not checked · 195 Own results and interpretations

127 more claims · 2 with their cited sources checked

Create a free account to see all 257 claims and to verify the 13 references still open.

See all claims — free account
  • Probably not a substantial claim. 65
  • It may state two separate claims as one. 76

Create a free account to see these claims.

Self-check

Percentages, sums and reported differences re-computed from the paper’s own numbers.

Each contradiction between two passages listed here was confirmed by an independent check.

4 of 4

Citations

Cites

Show all 107 references

Metadata

Authors
Kobrock, K., Ohmer, X., Bruni, E., & Gotzner, N.
Affiliations
Osnabrück University (DE)
Corresponding author
Kobrock, Kristina
Venue
PLOS Computational Biology
Published
2026-05-26
Volume
22
Issue
5
Pages
e1014326
ISSN
  • 1553-734X (ISSN-L)
  • 1553-7358
OpenAlex
W4411393719
Cited by
0
References
94
FWCI
0.00
Type
journal article
Language
en
Open access
Locations
  • PLoS Computational Biology (publishedVersion, cc-by)
  • PubMed (publishedVersion)
  • PubMed Central (publishedVersion, cc-by)
  • DOAJ (DOAJ: Directory of Open Access Journals) (submittedVersion)
  • PsyArXiv (OSF Preprints) (submittedVersion, cc-by)
Funders
Deutsche Forschungsgemeinschaft (DFG-GRK 2340, 441607011)
Topics
  • Social Sciences › Cultural Studies › Language and cultural evolution
  • Social Sciences › Arts and Humanities › Language and Linguistics › Syntax, Semantics, Linguistic Variation
  • Social Sciences › Psychology › Experimental and Cognitive Psychology › Categorization, perception, and language
Keywords
emergent communication, referential communication, categorization, referring expressions, Rational Speech Act, language emergence, shared context, multi-agent simulation
MeSH terms
Communication, Computer Simulation, Humans, Language, Speech, Speech Perception, Computational Biology

About these checks

Registry, retraction, Cites and statistics results are computed, not judged, with one exception: a printed reference that no index lists is read once by a model before it counts as a citation error. Claims are extracted by a model and kept only when their quote appears verbatim in the text; each cited source is then checked, with a model’s help, for whether it says what the claim relies on. Those verdicts can be wrong — read the claim’s quote and the cited source before relying on one.

The claim verdicts this report counts are experimental: their accuracy is not yet measured.

A language model drew the claims from the paper on 2026-10-01, under version 3 of the claim definition (model deepseek-v4.1-flash, prompt extract-v3).

A second model labelled each claim on 2026-10-01 (model jev-1.13-20260917, definition revision gold-guideline-v0.2).

Cited sources are checked with a language model’s help (gpt-6-sol).

Statistics and arithmetic are recomputed, not judged.

Confidence bands are shown only for fields calibrated on our annotated test set. None is calibrated yet, so no band is shown.

What counts as a claim ›How we measure confidence ›How we check this ›

Are you an author, or do you see a problem with this page? Request a review

How to cite

Cite CoChat for the checks on this page, and the paper for its own findings.

This page
CoChat. (2026). Citation check: The role of pragmatic mechanisms in referential communication and categorization: An emergent communication model. https://app.cochat.ai/papers/10.1371/journal.pcbi.1014326
The paper
Kobrock, K., Ohmer, X., Bruni, E., & Gotzner, N. (2026). The role of pragmatic mechanisms in referential communication and categorization: An emergent communication model. PLOS Computational Biology. https://doi.org/10.1371/journal.pcbi.1014326

Data sources

Bibliographic data: Semantic Scholar, OpenAlex, Crossref.

Contains information from the Semantic Scholar Academic Graph Datasets, made available under the ODC Attribution License (ODC-BY 1.0).

Semantic Scholar’s data is described in: Kinney, R., et al. (2023). The Semantic Scholar Open Data Platform. arXiv:2301.10140.