Extended Abstract

Learning Constraints on Wh-Dependencies by Learning How to Efficiently Represent Wh-Dependencies: A Developmental Modeling Investigation With Fragment Grammars

---

Abstract

It’s hotly contested how children learn constraints on wh-dependencies, called syntactic islands. When learning this knowledge, a prerequisite is knowing how to represent wh-dependencies so that constraints can be hypothesized over those representations. Previous work has explained disparate sets of syntactic island constraints by assuming different wh-dependency representations, without a unified dependency representation capturing all these constraints. We implement a modeled learner who learns a Fragment Grammar (FG) representation of wh-dependencies–a representation comprised of potentially different-sized fragments that combine to form full dependencies–that best accounts for the input while being as compact as possible. This efficient wh-dependency representation can then be used to generate any wh-dependency’s probability, and so predict acceptability patterns for stimuli sets that reveal syntactic island knowledge. We find that the identified FG can generate the attested acceptability judgment patterns for all syntactic islands previously investigated, highlighting how implicit knowledge of wh-dependency constraints can emerge from trying to learn to efficiently represent wh-dependencies more generally. We additionally compare the FG representation’s performance against baselines inspired by previous proposals, finding that one baseline also yields equivalent performance. We discuss how this baseline is similar to and different from the FG representation.

Keywords: developmental modeling, wh-dependencies, efficient representation, Fragment Grammar, syntactic islands

How to Cite:

Dickson, N., Pearl, L. & Futrell, R., (2022) “Learning Constraints on Wh-Dependencies by Learning How to Efficiently Represent Wh-Dependencies: A Developmental Modeling Investigation With Fragment Grammars”, Society for Computation in Linguistics 5(1), 220-224. doi: https://doi.org/10.7275/7fd4-fw49