The importance of being variable: Learning nonadjacent dependencies in speech
processing
Padraic Monaghan, Luca Onnis,
Morten H. Christiansen and Nick Chater
Abstract
Statistical learning of adjacent dependencies are robust for sequences of syllables, tones,
and visual stimuli in both infants and adults. However, to account for core aspects of
language learning a statistical learning mechanism must also be capable of tracking
relations among nonadjacent items. There are computational limitations to the
dependencies that can be monitored, and we propose that nonadjacent dependencies will
be processed when adjacent dependencies prove to be uninformative. One such condition
under which nonadjacent structure can be learned is when there is high variability of
intervening items (Gomez, 2002). In a series of experiments incorporating nonadjacent
dependencies between syllables in continuous speech, we show that speech segmentation
and generalisation can be achieved simultaneously when there is high variability in the
intervening syllables. Such results challenge the claim that two separate and consecutive
computational processes are necessary for explaining segmentation and generalisation in
language learning.
Please email suggestions/errors to mhc27@cornell.edu |