Séminaire de Recherche en Linguistique

Using Computational Models to Test Syntactic Learnability

What are the computational processes necessary for a person to learn the syntax of their language and process it in real-time? I will discuss work that brings new empirical evidence to bear on these questions through computational modeling by asking whether complex structural generalizations can be learned by a flexible, domain-general learning model from string input alone. I focus on the English filler–gap dependency and its related “island constraints” (Ross, 1967) — structural restrictions on the dependency, which have long been hypothesized as unlearnable without a language-specific acquisition bias. I use an assessment paradigm that treats models like subjects in a psycholinguistics experiment, and demonstrate that contemporary neural-network based language models (algorithms trained to assign probabilities to strings) learn the basic covariation that constitutes the dependency, as well as its potential unboundedness and hierarchical constraints. Turning to syntactic islands, I show that models do not expect fillers and gaps to co-vary inside the majority of island configurations tested. This work demonstrates that many subtle syntactic generalizations can be acquired purely from the task of next-word prediction by a domain-general learning algorithm.