Source
Limor Raviv, Antje Meyer and Shiri Lev-Ari (2019), Proc. R. Soc. B 286:20191262, LIT-825; read in NOTE-628 (main text; the supplementary models were not read).
The claim
In the authors’ experiment, when people who shared no language built one by communicating, the size of the group they communicated in shaped the result. With everyone able to talk to everyone and the same total amount of interaction, members of larger groups end with languages whose forms map more systematically onto meanings (parts of labels for parts of meanings), reach that structure sooner, and differ less from one group to the next. Members of small groups can sustain idiosyncratic, holistic labels, because with few partners and much shared history they can memorise them. The proposed cause is input variability: more partners means more competing variants to cope with, which favours forms that are predictable from their meaning.
What was actually shown
24 groups (12 of four, 12 of eight) of native Dutch adults played a typed reference game about 23 scenes of moving shapes over 16 rounds. Structure, the correlation between string distance and meaning distance in each participant’s labels, rose faster in larger groups and ended higher (β = 0.11, p = 0.006); larger groups varied less from each other. The result could have gone the other way: the authors also predicted that larger groups would converge less, and they did not (convergence and accuracy ended equal), and the cross-linguistic literature they test attributes simplification to network density or adult learners, which were absent here. Input variability was higher in larger groups and at round n predicted the gain in structure at n + 1, with shared history adding nothing once it was included.
What this does not say
- It does not say the languages of large natural communities are simpler because of size. The experiment removes the confounds; it does not show that, in natural populations, size rather than network density or adult second-language learning carries the effect.
- It does not establish the mechanism. Input variability was measured, not varied. Its lagged association with structure is consistent with the mechanism and does not exclude others that travel with group size (less shared history per pair, more partners to coordinate with).
- It does not hold for any size. Two sizes were tested, four and eight. Nothing is shown about how the effect grows, saturates or reverses.
- It is not about transmission across generations. No one was replaced; this is horizontal interaction. It does not bear directly on whether transmission chains converge to the learners’ prior (THEORY-155).
- “Systematic” here is a string–meaning distance correlation over a meaning space designed to have two separable dimensions. It does not measure syntax, morphology or regularity in a natural language’s sense.