Emergence of Cohesion in the Evolutionary Dynamics of Grammar Networks

Language provides a powerful instance of cultural evolution where individuals learn from each other and convergence to functionally similar states is advantageous. Recent studies exploring the evolutionary dynamics of populations of language learners have shown promising results regarding linguistic convergence and diversity. These studies often assume a set of equally related languages. In this paper, we deviate from this assumption and assume that some languages are more closely related than others; in particular we explore the population dynamics of grammar acquisition where the similarity between grammars is described by a ring lattice network. We specify the conditions for the emergence of linguistic coherence in terms of learning fidelity and network structure. In agreement with previous studies, our model exhibits a bifurcation from the symmetric state, where all grammars are in equal frequency, to an asymmetric state under dense grammar networks. However, different patterns of convergence, including a weak cohesion phase where the distribution of grammar frequencies is roughly Gaussian and centered around the most frequent grammar g, emerge in more sparse gramar networks. Weak cohesion is of particular interest from an adaptation standpoint since there is both linguistic coherence and standing heritable variation. In technological applications where agents learn from each other and it is desirable for the overall system to converge, our results may provide a guide to designing properties of the language or state representation depending on the desired degree of convergence. Weak cohesion provides a regime where overall functional convergence is relatively high, yet further adaptation is not hindered by low heritable variation.