Models of Language Evolution

Cathleen O’Grady, Kenny Smith2018 · DOI

TL;DR

This paper explores how the complex structure of human language, with its combinatorial and compositional properties, can emerge through cultural evolution, rather than solely relying on innate biological mechanisms. By reviewing computational models and experimental studies, it demonstrates that the pressures of language learning and transmission, particularly through "bottlenecks," can drive the development of linguistic regularity and systematicity.

What Problem Does This Paper Address?

The central question this paper tackles is how human language, with its remarkable complexity and expressive power, evolved. Unlike other animal communication systems, language allows for the creation of an infinite array of meanings from a finite set of components. This "duality of patterning" operates at two levels: combining meaningless sounds (phonemes) into meaningful words (morphemes), and then combining these meaningful units into larger structures like phrases and sentences. The paper questions why only humans possess such a system and proposes that cultural evolution, the transmission of language across generations, plays a crucial role in its emergence and structure. It aims to explain how this structure arises without necessarily invoking specialized, language-specific biological adaptations.

Methods

The paper reviews two primary methodological approaches:

  • Computational Modeling: This involves simulating evolutionary processes acting on communication systems. A key technique is "iterated learning," where artificial agents learn a communication system from data produced by previous generations. This process often incorporates "learning bottlenecks," where only a subset of linguistic data is passed on, forcing learners to generalize and leading to the emergence of compositional structure.
  • Experimental Studies with Human Participants: These studies use "transmission chains," where individuals learn and then reproduce material (e.g., drawings, artificial languages). By introducing learning bottlenecks or communicative interaction, researchers can observe how linguistic systems change and adapt over generations, mirroring the processes simulated in computational models. These experiments help to understand the role of individual learning biases and the pressures of communication itself.

Datasets

Not discussed in the paper. The paper focuses on theoretical models and experimental paradigms using artificial languages or simplified communicative tasks, rather than specific, large-scale datasets.

Key Insights and Ideas

  • Cultural Evolution as a Driver of Structure: The paper strongly suggests that the complex structure of language, including its combinatorial and compositional nature, can emerge through the process of cultural transmission across generations, rather than being solely hardwired by biology.
  • The Power of Learning Bottlenecks: The concept of "learning bottlenecks" is central. When learners receive incomplete input, they are forced to generalize, which over time, leads to the development of regular and compositional linguistic systems. This is a key mechanism for explaining the emergence of structure.
  • Learnability and Communicative Utility as Competing Pressures: The paper highlights a tension between making a language easy to learn (learnability) and making it expressive enough to convey a wide range of meanings (communicative utility). The evolution of structure often involves finding a balance between these pressures.
  • Individual Biases Accumulate: While individual learners might not show strong biases towards regularization or specific structures, these biases can emerge and become significant at the population level through repeated transmission across generations.
  • Iconicity Can Lead to Symbolicity: Experiments show that initially iconic representations (those that resemble their referent) can become abstract and symbolic through repeated use and transmission, driven by convention.
  • Regularity and Systematicity Emerge from Transmission: Even with random initial inputs, iterated learning and transmission chains lead to the development of predictable patterns and systematic structures in communication systems.

Key Results

  • Emergence of Compositionality: Computational models of iterated learning have demonstrated that compositional structure can emerge from initially holistic systems after hundreds of generations.
  • Simplification and Regularization: Experimental transmission chains show that languages tend to become simplified and regularized over generations, with the loss of unpredictable variation. For example, in object labeling tasks, one of two variants for a label might be lost.
  • Development of Combinatoriality: Experiments with artificial whistled languages and graphical communication tasks show that combinatorial structure can emerge, allowing for the creation of more complex signals from simpler components.
  • Convergence on Typological Universals: Studies on word order in artificial sign languages and other linguistic features suggest that learning biases, amplified by cultural transmission, can explain why certain linguistic structures are common across many languages (typological universals). For instance, a preference for SOV or SVO word order can emerge.

Limitations and Future Work

  • Biological Evolution's Role: While the paper emphasizes cultural evolution, it acknowledges that biological evolution has likely provided the necessary cognitive biases and learning abilities that underpin human language. A comprehensive account will need to integrate both.
  • Complexity of Real Languages: The models and experiments often use simplified artificial languages. Future work could explore how these principles apply to the full complexity of natural languages.
  • Understanding Specific Biases: While the paper points to learning biases, further research is needed to precisely identify and quantify these biases and their specific impact on different linguistic features.
  • Cross-Species Comparisons: The paper touches upon bird song as a comparable system. Further comparative studies across species could offer deeper insights into the fundamental principles of cultural transmission and the evolution of communication.
  • The Interplay of Interaction and Transmission: While both are shown to be important, understanding the precise dynamics of how interaction and transmission jointly shape language evolution remains an active area of research.

Summary updated

Abstract

Models of Language Evolution reviews the models that provide evidence for the role of cultural evolution in the emergence of linguistic structure. This chapter discusses the levels of linguistic structure, and why the emergence of structure in language is a central question for evolutionary linguistics. It reviews the computational and experimental models which demonstrate that pressures operating during language learning and language use can give rise to the appearance of design in language, through the repeated cycle of learning and use that characterise language transmission. Finally, it discusses how learning biases at the individual level lead to the presence of typological universals: systematic patterns in how the world’s languages tend to be structured.
BibTeX
@misc{ogrady2018models,
  author = {Cathleen O’Grady and Kenny Smith},
  title = {{Models of Language Evolution}},
  year = {2018},
  doi = {10.31219/osf.io/pfg8x},
}
Figure