Adding population structure to models of language evolution by iterated learning

Andrew Whalen, Thomas L. GriffithsJournal of Mathematical Psychology 2016 · DOI

Introduction

The Iterated Learning Model (ILM) has long served as a foundational framework for understanding how languages evolve as they are transmitted across generations. A landmark finding in this field, established by Griffiths and Kalish (2007), demonstrated that when learners are Bayesian agents, the languages they adopt will eventually converge to reflect their innate cognitive biases, or "priors." However, these early models typically assumed simple linear transmission chains—where one individual teaches the next in a single line.

In reality, human populations are organized into complex social networks. This paper addresses a significant gap in the literature: does the "prior = stationary distribution" rule hold when learners are embedded in structured populations? Furthermore, can these models explain the pervasive geographical clustering of languages observed in the real world?

Modeling Language Evolution on Graphs

To investigate these questions, the authors integrated Bayesian inference with graph theory and statistical physics.

Bayesian Iterated Learning

In this model, agents are Bayesian learners. When an agent hears a linguistic utterance, they use Bayes' rule to infer the underlying language. This inference is shaped by two factors: the data they receive (the utterance) and their prior probability distribution, which represents an innate cognitive bias toward certain linguistic structures.

Population Structures

Unlike traditional linear chains, this study places agents on various types of undirected graphs to simulate different social structures:

  • 2D Lattices: Representing simple spatial proximity.
  • Small-World Graphs: Characterized by high local clustering but short path lengths between any two nodes.
  • Scale-Free Graphs: Featuring "hubs" (highly connected individuals) and "authorities," mimicking more complex social hierarchies.

The Evolutionary Process

The authors used a birth-death process to simulate generational turnover. In each step, a random agent is removed from the graph and replaced by a new, "naive" learner. This newcomer selects one neighbor to serve as a teacher, hears an utterance, and adopts a language based on their resulting Bayesian posterior.

Key Findings: Global Robustness and Local Clustering

The study yielded two primary insights that reconcile cognitive models with linguistic geography.

1. Global Convergence to the Prior

The most striking theoretical result is that the global distribution of languages remains robust regardless of the population's social structure. On any undirected graph, the probability of a learner speaking a specific language still converges to the learner's prior bias. For instance, if the population has a 70% prior bias toward Language A, approximately 70% of the population will speak Language A at equilibrium, whether they are arranged in a simple chain or a complex scale-free network. This confirms that social structure does not override innate cognitive constraints on a global scale.

2. The Emergence of Local Homogeneity

While the global frequency is fixed by the prior, the spatial arrangement of speakers is profoundly influenced by the graph structure. In structured networks like 2D lattices, learners are significantly more likely to share a language with their immediate neighbors than would be expected by chance.

By applying tools from statistical physics—specifically voter models and pair approximation—the authors mathematically predicted this degree of clustering. This finding explains a long-standing paradox: how a globally diverse population (reflecting weak or varied priors) can maintain highly homogeneous local "pockets" of speakers.

Implications and Future Directions

This research provides a mathematical bridge between the cognitive mechanisms of individual learning and the macro-level patterns of geographic linguistics. It suggests that while our brains may set the "menu" of available languages, our social networks determine the "neighborhoods" in which they reside.

However, the authors identify several avenues for future research. The current proof relies on a "single teacher" constraint; if learners aggregate data from multiple teachers simultaneously, the population might converge to a single dominant language, potentially overriding the prior. Additionally, exploring directed graphs (to model one-way influence) and incorporating "fitness" or "utility" parameters (where some languages offer communicative advantages) would further refine our understanding of how social and cognitive forces interact to shape human language.

Summary updated

BibTeX
@article{whalen2016adding,
  author = {Andrew Whalen and Thomas L. Griffiths},
  title = {{Adding population structure to models of language evolution by iterated learning}},
  journal = {Journal of Mathematical Psychology},
  year = {2016},
  doi = {10.1016/j.jmp.2016.10.008},
}
Figure