Fedzechkina et al.: learners restructure toward short dependencies

[FNJ17] [FJN12]

Artificial-language learning with two mini-languages sharing a lexicon but differing in constituent order: one places the complex NP where the verb's dependencies stay short, the other where the subject dependency stretches across the intervening complex NP. Learners exposed to a 50/50 mixture converge toward the short-dependency order (about two thirds of productions by the end of training) — a learning bias for dependency-length-minimizing orders.

This file holds the design's critical contrast as trees: the two orders of "the big cat chased the dog" with one complex and one simple NP, and the fact that the preferred order has strictly smaller total dependency length. The convergence rates and their statistics live in the papers.

def FedzechkinaEtAl2017.complexFirst : DependencyGrammar.Graph 6

Complex NP first: "the-big-cat the-dog chased" — the verb is close to both argument heads.

Equations

• One or more equations did not get rendered due to their size.

def FedzechkinaEtAl2017.complexLast : DependencyGrammar.Graph 6

Complex NP last: "the-dog the-big-cat chased" — the subject dependency stretches across the complex NP.

Equations

• One or more equations did not get rendered due to their size.

theorem FedzechkinaEtAl2017.preferred_order_shorter_deps : complexFirst.totalLength < complexLast.totalLength

The order learners converge toward has strictly smaller total dependency length.