Co-evolution of Language and of the Language Acquisition Device

Ted BriscoeACL 1997 · DOI

Ted Briscoe’s 1997 paper investigates a fundamental question at the intersection of evolutionary biology and linguistics: how did the human "Language Acquisition Device" (LAD)—the innate mental machinery that enables children to learn language—co-evolve with the languages themselves? Using computational simulations, Briscoe explores the emergence of Universal Grammar (UG) through the lens of the Baldwin Effect.

The Problem: Poverty of the Stimulus and the Search Space

Linguists often cite the "poverty of the stimulus" to argue that children learn language faster than their exposure to messy, incomplete data would suggest. The standard theoretical solution is Universal Grammar: an innate set of parameters (e.g., verb-object order) that children "flip" based on linguistic input.

However, this theory faces two significant hurdles:

  1. The Search Problem: With numerous parameters, how do learners avoid "local maxima"—incorrect grammatical rules that appear valid for some sentences but fail globally?
  2. The Evolutionary Problem: How could such a complex, pre-configured system evolve biologically?

Briscoe addresses these by proposing that parameters are organized in a default inheritance hierarchy, where general rules are set first and specific overrides follow.

Computational Framework

To test these ideas, Briscoe developed a sophisticated model centered on Generalized Categorial Grammar (GCG). In this system, words are treated as mathematical functions that combine to form sentences.

A GCG derivation of the sentence "Kim loves Sandy," illustrating how lexical categories combine through functional application.

The model incorporates:

  • Working Memory Load (WML): A shift-reduce parser calculates the cognitive cost of processing a sentence.
  • Evolutionary Simulation: Populations of "Language Agents" (LAgts) interact and reproduce based on fitness. Fitness is determined by communicative success and the cognitive ease (low WML) of the language they use.

Modeling Linguistic Diversity

The simulation utilizes a synthetic "Grammar Set" comprising 20 parameters that define nearly 300 distinct grammars. These parameters capture typological distinctions found in real-world languages, allowing the model to simulate families like SVO (English) or SOV (Japanese).

Typological parameter settings for English, German, and Japanese used to define the simulation's language families.

Key Insights: The Baldwin Effect and Memory Constraints

The central finding is the Baldwin Effect in action. Initially learned traits (specific grammars) can become innate (default parameter settings) over generations. Agents born with "correct" defaults learn faster, achieving higher fitness and eventually dominating the population.

Crucially, limited working memory acts as a beneficial filter. It forces learners to focus on simple "trigger" sentences early in development, preventing them from being overwhelmed by complex or ambiguous data. This constraint drives the co-evolutionary dance: languages evolve to be easier to parse, while learners evolve defaults that make those languages easier to acquire.

Results and Implications

The simulations yielded several key results:

  • Efficiency of Defaults: "Default learners" (starting with pre-set SVO-like parameters) were significantly more successful at converging on correct grammars than "unset learners" (blank slates).
  • Evolutionary Dominance: Under memory constraints, default parameters consistently rose to dominance. Without these limits, the evolutionary pressure for defaults was significantly weaker.
  • Language Selection: Populations tended to settle on "subset languages"—simpler versions of languages that minimize working memory load.

Briscoe’s work suggests that right-branching SVO grammars may represent a "near-optimal" starting point for a species developing complex syntax. While the model simplifies some aspects of expressivity and assumes a fixed learning algorithm, it provides a powerful computational proof-of-concept for how biological evolution and linguistic change are inextricably linked.

Summary updated

BibTeX
@inproceedings{briscoe1997co,
  author = {Ted Briscoe},
  title = {{Co-evolution of Language and of the Language Acquisition Device}},
  booktitle = {Annual Meeting of the Association for Computational Linguistics},
  year = {1997},
  doi = {10.3115/976909.979671},
}
Figure