Contribution
Cross-linguistic work had found that languages of larger communities tend to be simpler and more regular, but could not separate size from network density, contact and the share of adult second-language learners. This is the first experiment, by the authors’ account, to vary group size with human participants while holding network structure and total interaction fixed. It shows a causal effect of size on the systematicity of an emerging language, and offers, with correlational support, a mechanism: the greater variety of forms a member of a larger group must cope with.
Key insight
Talking with more people means facing more variants and sharing less history with each. Memorising every partner’s idiosyncratic labels then fails, and what spreads instead is the predictable: labels whose parts map onto parts of meaning. Size buys structure by making memorisation harder.
Assumptions
- Fully connected groups, members exchanging partners every round, so that the only difference between conditions is size (and with it shared history per pair).
- Equal total interaction: 23 interactions per round per pair, the same number of rounds in both conditions.
- Structure measure: the Pearson correlation, per participant and round, between normalised Levenshtein distances of labels and semantic distances (1 for a shape difference plus the angle difference / 180°), used raw rather than as a Mantel z-score because the meaning space grew over rounds.
- Meaning space: 23 scenes, four shapes, continuous angle of motion, and a unique fill per scene that allows a holistic strategy.
- Participants: native Dutch adults; Dutch and other languages forbidden, enforced by an experimenter present.
Key results
Mixed-effects models (lme4, Kenward–Roger p-values); full models in the supplement, not read.
- Structure increased over rounds (Model 10: β = 4.55, t = 9.46, p < 0.0001), nonlinearly; the increase was faster in larger groups (β = 1.92, t = 3.06, p = 0.004); final structure was higher in larger groups (Model 11: β = 0.11, t = 2.93, p = 0.006); larger groups varied less (Model 12: β = −0.015, p = 0.0002).
- Communicative success rose (β = 0.08, p < 0.0001), not significantly modulated by size; final accuracy did not differ (p = 0.083). Larger groups had lower variance.
- Convergence rose (β = 0.007, p = 0.029); no effect of size; larger groups less variable (β = −0.04, t = −23.68).
- Stability rose; larger groups’ languages less stable overall (β = −0.08, p = 0.047), not by the end (p = 0.24).
- Mechanism. Larger groups had more input variability (Model 13: β = 1.45, t = 15.99). Input variability at round n predicted the increase in structure at n + 1 (Model 14: β = 0.015, p < 0.0001); less shared history did too (Model 15: β = −0.017, p = 0.0004); jointly only input variability was significant (Model 16: β = 0.011, p = 0.012; shared history β = −0.008, p = 0.17).
- Function of structure. Structure predicted convergence (Model 17: β = 0.018, p = 0.027) and accuracy (Model 18: β = 0.436, p < 0.0001), more strongly in later rounds.
- Qualitative. Many groups developed shape-part plus motion-part labels; groups differed in how they carved angle (two-axis or clock-like systems), and “no two languages were identical”.
Claims
| id | claim | strength | support |
|---|---|---|---|
| C1 | With network structure and total interaction fixed, groups of eight develop more systematic languages than groups of four, and faster | moderate | Models 10–11; 12 groups per condition |
| C2 | Larger groups are more alike in outcome; small groups vary more on every measure | moderate | Models 3, 6, 9, 12 |
| C3 | Input variability drives the increase in structure | weak to moderate | lagged correlational Models 14–16; variability was not manipulated independently of size |
| C4 | Group size does not affect final communicative success or convergence | moderate (a null) | Models 1–2, 4–5; p = 0.083 on final accuracy |
| C5 | The results scale to real communities and support size as a driver of language evolution | not supported here | argued in the discussion, with one sign-language comparison cited |
Concepts
- community size: the number of people in a fully connected group, four or eight.
- input variability: per scene and round, the summed normalised Levenshtein distance from the most typical label to all other labels produced for it; averaged over scenes.
- shared history: how many times a pair has interacted so far.
- linguistic structure / systematicity: string–meaning distance correlation; compositionality in the paper’s sense, with no distinction between morphology and syntax (endnote).
Connections
The paradigm descends from the group-communication and iterated-learning experiments it cites (Kirby, Cornish and Smith, LIT-771; Kirby, Tamariz, Cornish and Smith 2015; Raviv, Meyer and Lev-Ari’s own “Compositional structure can emerge without generational transmission”, 2019), but there is no generational turnover here: structure arises by horizontal interaction. The cross-linguistic correlation it tests is Lupyan and Dale (2010); the confounds it removes are network density (Trudgill, Wray and Grace) and second-language learners (Bentz and Winter). The sign-language comparison cited is Meir et al. (2012) on two young sign languages, one of a larger community and one of a small one.
Bearing on the record
- Produces THEORY-164: the size effect and the input-variability mechanism, with what the experiment does not show.
- Against THEORY-155. That theory says the endpoint of transmission between Bayesian samplers is the shared prior, independent of how much data passes. This study is not a chain, so it neither supports nor contradicts it; but it shows a social-structural variable changing the outcome among learners who presumably share their biases, which a reader of THEORY-155 should not assume away for interacting populations.
- No instruction for machine-learning practice; nothing for the anthology.
Limitations
- Two sizes only (four and eight), 12 groups each; “larger” is eight.
- Input variability was measured, not manipulated, so the mechanism rests on lagged associations within the same data.
- The meaning space was designed to reward compositional structure, and the incremental introduction of scenes was “implemented in order to introduce a pressure for developing structured and predictable languages”.
- Participants all spoke Dutch and brought its phonotactics and segmentation habits; the written channel and restricted alphabet are far from spoken or signed language.
- One large group played one round fewer, and one participant’s data were lost; both groups’ data were kept.
- The claim that results “should scale” to real populations is argued, not shown.
Open questions
- Does the effect hold, or grow, beyond eight, and with sparse rather than fully connected networks, where the authors expect network structure to modulate size?
- If input variability is manipulated directly in fixed-size groups (for example by injected variant labels), does structure follow?
- In a chain with generational turnover, does population size per generation change the endpoint, or only the speed, as THEORY-155 would lead one to ask?