Cultural evolution involves both convergent transformation and selection. Both forces can lead to convergence on a preferred cultural variant, but they can work in different directions. Computational models have found that if transformation and selection work in opposite directions, cultural transmission will generally result in a compromise between the two. Here, we present a model of the cultural co-evolution of language and perspective-taking, and show that transformation can strengthen the effect of selection, even if this transformation works towards the variant that is dispreferred by selection.
In this model, languages are transmitted through iterated Bayesian learning, but in order for an agent to learn language successfully, they also need to infer other agents' perspectives on the world. Although perspectives are learned in this model, unlike language, they are not culturally transmitted. Based on experimental studies of child learners, we investigate the case when agents have an egocentric bias, that leads them to initially infer, incorrectly, that other speakers have the same perspective as them. This leads to a tendency for ambiguous languages to evolve in the population. However, adding cultural selection for better perspective taking to the model overcomes this tendency and leads to more informative languages, because these help agents infer perspectives. The effect of this selection is strongest when learners have an egocentric bias, because this makes receiving input from an informative language even more important. Thus, in this case where convergent transformation and selection both indirectly influence the informativeness of language, transformation prepares the ground for selection.