Contribution
A laboratory paradigm for iterated learning with human participants, showing that a language transmitted down a diffusion chain changes cumulatively, becoming easier to learn and more structured, when no participant is trying to change it. Before it, the claim that iterated learning produces linguistic structure rested on computational and mathematical models, and earlier human diffusion chains (Horner et al. 2006) used tasks with a small set of equally good behaviours, where cumulative adaptation could not show.
Key insight
Give each learner only part of a language and test it on all of it. What it cannot memorize it must generalize, and its generalizations become the next learner’s data. Over generations the language reorganizes so that the part predicts the whole. Which reorganization wins depends on what else constrains it: with no pressure against ambiguity, one word per class of meanings; with ambiguity removed from training, one part of the word per feature of the meaning.
Assumptions
- Participants: 80 university students without linguistics training (46 female, 34 male; mean age 22.5, range 18–40), Edinburgh.
- Meaning space: 27 pictures, a coloured object with a motion arrow: shape (square, circle, triangle) × colour (black, blue, red) × motion (horizontal, bouncing, spiralling).
- Initial languages: one random label per picture, 2–4 syllables concatenated from a set of 9 consonant–vowel syllables. Later labels are whatever participants type.
- Bottleneck: each language is split at random into 14 SEEN and 13 UNSEEN pairs. Three rounds of training, each with two randomized exposures to SEEN (string 1 s, then string and picture 5 s) and a test; rounds 1–2 test half of each set, round 3 tests all 27, and that final output is the next generation’s language, re-split at random.
- Chains: four per experiment, ten generations each, each starting from a different random language.
- Participants’ belief: told to reproduce the alien language as well as possible; not told about the chain; a post-test questionnaire suggested many did not notice they were tested on unseen items.
- Experiment 2 filter: before training, if a string labels more than one picture in SEEN, all but one (chosen at random) are moved to UNSEEN, so training data are one-to-one. Described as “an analogue of a pressure to be expressive that would come from communicative need”.
Key results
- Measures. Transmission error: mean normalized Levenshtein distance between a participant’s strings and the previous generation’s, over meanings (0 identical, 1 maximally different). Structure: Pearson correlation between pairwise string distances (normalized Levenshtein) and pairwise meaning distances (Hamming, over the three features), reported as a z-score against 1,000 Monte Carlo reassignments of strings to meanings; undefined when the shuffled sample has no variance (two cases in Experiment 1).
- Experiment 1 (Figure 2, Table 1). Error fell from first to last generation (mean decrease 0.748, SD 0.147, t(3) = 8.656, P < 0.002); structure rose (mean increase 5.578, SD 2.968, t(3) = 3.7575, P < 0.02). Distinct words fell from 27 to 2, 4, 5 and 4 by generation 10. In some chains later participants reproduced the language exactly, unseen items included. “This underspecification was not random but was systematic, in that similar meanings were given the same label.” Example chain: “tuge” for every horizontal mover by generation 4, “poi” for most spiralling ones by generation 6, and by generations 8–10 horizontal “tuge”, spiralling “poi”, bouncing split by shape. Colour distinctions were lost in all four chains.
- Experiment 2 (Figure 4, Table 2). Error fell (mean decrease 0.427, SD 0.106, t(3) = 8.0557, P < 0.002); structure rose (mean increase 6.805, SD 5.390, t(3) = 2.525, P < 0.05). Distinct words stayed between 10 and 27 across chains and generations (generation 10: 16, 12, 12, 23). The circled generation-9 language segments into colour, shape and motion morphemes, with irregular items (one string for a bouncing red circle); the structure appears by about generation 6 and persists while individual morphemes are lost or reanalysed.
Claims
Not filled: the reading is Skimmed. The paper’s claims, as reported, are that transmission alone made the languages more learnable and more structured (both experiments, four chains each, t-tests with df = 3); that the adaptation is “cumulative with respect to learnability and structure but not with respect to expressivity”; and that this is the “first experimental validation for the idea that cultural transmission can lead to the appearance of design without a designer”.
Concepts
- iterated learning: acquiring a behaviour by observing it in someone who acquired it the same way.
- diffusion chain: each participant’s output is the next participant’s input; one participant per generation. “The first reported use of this methodology was by Bartlett in 1932.”
- transmission error: as above; read as the inverse of learnability.
- structure: the correlation of string similarity with meaning similarity.
- systematic underspecification: one string for a class of similar meanings, so that a learner can recover the whole language from a fragment.
Connections
- Bartlett (1932), LIT-768, is named as the first use of the diffusion-chain method. Bartlett’s serial reproduction passed a story down a chain and watched it drift; this paper passes a language down a chain and measures the drift. The paper does not say what Bartlett found, and the record has not read him (LIT-768 stays Deferred).
- Griffiths & Kalish (2007), LIT-769, read in NOTE-599, is cited only inside the block of model references [4–13]. The paper does not use priors or Bayesian learners. Read against it: the loss of colour in every Experiment 1 chain, which the authors tie to the shape bias in word learning, is the kind of outcome Griffiths and Kalish predict, a learner bias showing through transmission. The Experiment 2 filter is an experimenter-imposed selection on the data, outside the learners, which is exactly what their analysis assumes away (equal fitness, no selection). So Experiment 2 is not a test of convergence to the prior. This reading is mine, not the paper’s.
- Kirby (2000): the earlier model that removed ambiguous strings from the learner’s data, the source of the Experiment 2 filter.
- Galantucci; Garrod et al.; Selten & Warglien: experiments in which participants design a shared communication system on purpose, which this design is meant to differ from.
- Horner et al. (2006): diffusion chains in chimpanzees and children with a two-technique puzzle box.
Bearing on the record
- Bears on THEORY-155 (filed from NOTE-599): consistent with it, not a test of it. The chains show bias-shaped outcomes (colour dropped, shape and motion kept), but no participant’s prior was measured, and the filtered condition adds a pressure the theory excludes. Recorded there under “what this does not say”.
- No instruction for machine-learning practice; not an anthology candidate.
Limitations
- Four chains per experiment; the tests have three degrees of freedom.
- Adult university students who already speak a language. The authors concede that “a participant’s first language may influence the learnability”, and argue it cannot explain both the underspecification of Experiment 1 and the concatenative morphology of Experiment 2.
- The expressivity pressure in Experiment 2 is an artificial filter applied by the experimenter, not communication between participants; expressivity and communicative success are not measured.
- Written strings for a 27-item, three-feature meaning space.
- Not checked here: the SI (instructions, full languages in Tables S1–S8), and the formulas, which were seen only as described.
Open questions
- Does the same paradigm with real communication between participants, rather than a filter, produce compositional structure?
- Could the learners’ biases be measured independently, so that the chains’ end states can be compared with them (the test THEORY-155 asks for)?
- Do the results hold with more chains, larger meaning spaces, or child learners?