Search papers, labs, and topics across Lattice.
The paper investigates the relationship between regularity and learnability in recursive numeral systems by training agents using Reinforcement Learning to learn different numeral systems. They demonstrate that human-like regular systems are easier to learn than irregular systems, supporting the hypothesis that learnability influences the prevalence of regular systems in human languages. Furthermore, they find that for highly irregular systems, signal length becomes a more dominant factor in learnability than regularity.
Regularity in numeral systems isn't universally beneficial for learnability; its advantage vanishes for highly irregular systems, where signal length becomes the dominant factor.
Human recursive numeral systems (i.e., counting systems such as English base-10 numerals), like many other grammatical systems, are highly regular. Following prior work that relates cross-linguistic tendencies to biases in learning, we ask whether regular systems are common because regularity facilitates learning. Adopting methods from the Reinforcement Learning literature, we confirm that highly regular human(-like) systems are easier to learn than unattested but possible irregular systems. This asymmetry emerges under the natural assumption that recursive numeral systems are designed for generalisation from limited data to represent all integers exactly. We also find that the influence of regularity on learnability is absent for unnatural, highly irregular systems, whose learnability is influenced instead by signal length, suggesting that different pressures may influence learnability differently in different parts of the space of possible numeral systems. Our results contribute to the body of work linking learnability to cross-linguistic prevalence.
