The emergence of linguistic structure in an online iterated learning task

Abstract

Previous research by Kirby, Cornish & Smith (2008) has found that strikingly compositional language systems can be developed in the laboratory via iterated learning of an artificial language. However, our reanalysis of the data indicates that while iterated learning prompts an increase in language compositionality, the increase is followed by an apparent decrease. This decrease in compositionality is inexplicable, and seems to arise from chance events in a small dataset (4 transmission chains). The current study thus investigates the iterated emergence of language structure on a larger scale using Amazon Mechanical Turk, encompassing 24 independent chains of learners over 10 generations. This richer dataset provides further evidence that iterated learning causes languages to become more compositional, although the trend levels off before the 10th generation. Moreover, analysis of the data (and reanalysis of Kirby, Cornish & Smith, 2008) reveals that systematic units arise along some meaning dimensions before others, giving insight into the biases of learners.