Language as an evolutionary system

Henry Brighton, Kenny Smith, Simon KirbyPhysics 2005 · DOI

TL;DR

This paper argues that languages themselves evolve, adapting to become more transmissible and stable through cultural evolution. Using computational models, the authors demonstrate how a "transmission bottleneck" and specific learner biases drive the emergence of compositional structure, a key feature of human language, even without explicit pressure for communicative function.

What Problem Does This Paper Address?

The paper tackles the fundamental question of how the complexity and structure of human language arise and persist across generations. While some theories posit innate linguistic structures (nativism), this work explores an alternative: that language itself is an evolutionary system. It investigates how the process of cultural transmission, rather than solely biological endowment, shapes linguistic structure. The authors aim to explain linguistic universals and the evolution of complexity by viewing language as a system that adapts to its own transmission challenges.

Methods

The core methodology involves computational modeling using the Iterated Learning Model (ILM). This framework simulates language transmission across generations, where each agent learns from the previous generation's linguistic output and then produces its own.

Key components of the models include:

  • Associative Matrix Model: A simplified model of language learning and production where agents learn associations between meanings and signals.
  • Minimum Description Length (MDL) Principle: A normative theory of induction used to model hypothesis selection, favoring hypotheses that offer the most compressed representation of data.
  • Finite State Unification Transducers (FSUTs): A formal representation of languages used in the MDL-based model, allowing for compression and generalization.
  • Transmission Bottleneck: A crucial constraint simulating the fact that learners never observe the complete language of the previous generation.
  • Learner Biases: Specific assumptions about how learners process linguistic data, such as biases against many-to-one or one-to-many mappings.

Datasets

Not discussed in the paper. The models generate their own "data" through simulated linguistic interactions.

Key Insights and Ideas

  • Language as an Evolutionary System: The central idea is that language is not just a tool for communication but an evolving entity that adapts to its transmission environment.
  • Transmission Bottleneck as a Driver of Structure: The limitation on what learners can acquire (the bottleneck) is a powerful force that pressures languages to become more structured and generalizable to survive.
  • Compositionality Emerges from Transmission Pressure: The paper demonstrates that compositional structure (where meaning is a function of parts and their combination) arises as an adaptation to the transmission bottleneck, making languages more robust.
  • Learner Biases are Crucial: Specific biases in language learners (e.g., against ambiguity or one-to-many mappings) are shown to be prerequisites for the evolution of both compositional structure and communicative function.
  • Communicative Function as a Side-Effect: Surprisingly, the models suggest that communicative optimality can emerge as a consequence of the biases that drive compositionality, rather than being the primary selective pressure.
  • Invention Plays a Key Role: Beyond learning, the process of "invention" (producing signals for meanings not explicitly learned) is shown to be critical for the evolution of complex linguistic structures.

Key Results

  • Bottleneck Drives Compositionality: Simulations consistently showed that languages evolved high levels of compositionality when a transmission bottleneck was present, while languages without a bottleneck remained largely holistic.
  • Generalization is Essential: Learners' capacity to generalize was a prerequisite for the evolution of compositional structure. Without it, languages remained random.
  • Bias Against Ambiguity is Necessary: A bias against many-to-one mappings (ambiguity) was crucial for both the evolution of compositional structure and communicative function.
  • Optimal Communication Emerges: In many scenarios, systems evolved that achieved high communicative accuracy, often as a byproduct of the pressures leading to compositionality.
  • MDL Induction Leads to Structure: The MDL-based model, incorporating intelligent invention, demonstrated the evolution of compressible and expressive languages, suggesting that simplicity principles can drive linguistic evolution.

Limitations and Future Work

  • Simplified Models: The authors acknowledge that their models are simplifications of natural language, often focusing on specific aspects like compositionality and using simplified meaning/signal spaces.
  • Interaction of Systems: The paper notes that real languages involve multiple interacting systems of regularity, which are not fully captured in the current models.
  • Origin of Biases: While the models show that certain biases are beneficial for linguistic evolution, the paper touches only briefly on the evolutionary origins of these human biases themselves.
  • Future Work: The authors suggest exploring more complex models that capture interactions between different linguistic systems and investigating the evolutionary origins of learner biases. They also highlight the need for further understanding of how innovation and production mechanisms interact with learning in driving linguistic evolution.

Summary updated

BibTeX
@inproceedings{brighton2005language,
  author = {Henry Brighton and Kenny Smith and Simon Kirby},
  title = {{Language as an evolutionary system}},
  booktitle = {Physics},
  year = {2005},
  doi = {10.1016/j.plrev.2005.06.001},
}
Figure