Introduction
A foundational assumption in linguistics is that the most common features of human languages—linguistic universals—are prevalent because they are the easiest for the human brain to learn. This "learnability bias" suggests that "easy" patterns should naturally crowd out "hard" ones over generations of transmission. However, research presented at the 2011 Workshop on Cognitive Modeling and Computational Linguistics challenges the sufficiency of this theory. The authors demonstrate that individual learning biases can be overwhelmed by transmission noise and the sheer number of alternative linguistic structures.
The Iterated Learning Framework
To investigate the gap between individual bias and population-level universals, the researchers modeled language evolution using an iterated learning paradigm. In this model, language is passed from one generation to the next like a "telephone game": a learner acquires a language from data produced by a teacher and then becomes the teacher for the next learner.
This process is formalized as a Markov chain. Each possible language is a "state," and a transition matrix ($Q$) defines the probability of a learner moving from one state to another. The long-term prevalence of language features is determined by the system's stationary distribution—the equilibrium state reached after many generations.
The "Many Alternatives" Trap
The paper identifies a critical hurdle for the learnability hypothesis: the "Many Alternatives" trap. While a specific "easy" language might be more learnable than any single "hard" language, the cumulative weight of hundreds of alternative "hard" structures can prevent the "easy" one from dominating. If transmission errors (mutations) frequently lead learners away from the "universal" property toward a variety of other forms, that property will never become prevalent at a population level.
Simulation and Human Experiments
The authors tested this dynamic through Bayesian simulations and human experiments. In simulations, they varied the strength of the learning bias ($\lambda$) to see how it affected language convergence in a large state space.
As illustrated in the simulation results, when the number of possible languages is large, even a statistically significant bias fails to produce a majority choice unless that bias is exceptionally strong. The system remains in a state of high variance, scattered across many possible languages.
Human experiments mirrored these findings. While participants were significantly better at learning "vowel harmony" (a common universal) than an arbitrary rule in a lab setting, this advantage disappeared in iterated learning chains. Regardless of the starting point, the languages drifted toward a state where the "easy" rule was only followed about 50% of the time, as the "noise" of alternative forms overwhelmed the human preference for harmony.
Conclusion
The study concludes that learnability is not a guaranteed ticket to becoming a linguistic universal. Individual cognitive biases are only one part of the equation; the dynamics of transmission, mutation rates, and the structural landscape of alternative languages play equally vital roles. To truly explain linguistic universals, researchers must look beyond simple "ease of learning" and account for the complex population-level dynamics of language change.
Summary updated