Current Projects

A nearly exhaustive test of the “d-linking” amelioration effect on island effects in English.

Islands are sometimes divided into two types: strong islands, which block extraction of all dependency types, and weak islands, which selectively block certain dependencies, such as simple wh-items (what), but not others, such as complex wh-items (which book). Recent experimental studies provide preliminary indications that the strong/weak island distinction for simple and complex (d-linked) wh-items might be gradient: the extraction of complex wh from weak islands partially reduces the island effect, but does not eliminate it entirely. Given the far-reaching consequences that these findings would have for theories of islands, and for the architecture of grammar in general, we conducted a large-N acceptability judgment study to assess the facts for a wide range of islands in English (38 island types spanning 7 island families). We tested ~18000 English native-speakers (~200 per island/wh-type), and we used Bayesian linear regression to estimate posterior distributions of island effect sizes. Results show that amelioration, when present, is partial, and that some islands show patterns not previously reported (check out this presentation presented at GLOW 46).

The nature of island constraints: a behavioral and computational investigation

Formal theories of grammar and traditional parsing models, insofar as they presuppose a categorical notion of grammar, face the challenge of accounting for gradient judgments of acceptability. This challenge is traditionally met by explaining these effects as deriving from extra-grammatical factors.

We are developing a new model based on Self-Organized Sentence Processing (SOSP) that derives gradient effects from the functioning of the grammar itself through a mechanism that coerces the system to generate a structure even when no optimal parse is available. On this view, gradience reflects the ease with which the system can undergo coercion (check out this publication).