Processing nonadjacent dependencies: A graded, associative account
Luca Onnis, Arnaud Destrebecq, Morten H. Christiansen, Nick CHater, & Axel Cleeremans
Abstract
Theories of Artificial Grammar Learning in adults and infants propose different putative learning mechanisms: rule abstraction, similarity
with stored instances, knowledge of whole items, or sensitivity to low-level statistical properties of the input.
These models, however, cannot account for how subjectsÕ sensitivity to nonadjacent dependencies is modulated by the variability
of the intervening material. Gomez (2002) showed that learning about nonadjacent dependencies improves when the embedded material
is highly variable. Onnis, Christiansen, Chater, & Gomez (2003) further demonstrated that learning is also improved when the embedded
material has a variability of zero. Together, these results thus indicate a U-shaped relationship between variability and performance.
In this paper, we report on simulation studies that demonstrate that this U-shaped relationship can be accounted for by a Simple Recurrent
Network trained to predict the successor to each element of input sequences. These results suggest that variability plays a critical role in
defining the conditions under which long-distance dependencies can be mastered. We discuss these results in light of current theories of
language acquisition, and conclude that distributional learning might be more powerful than previously anticipated
Please email suggestions/errors to lo35@cornell.edu |