Akira Omaki of Johns Hopkins will be giving a colloquium:
Friday, Sept. 27, 3:30 pm
in Machmer E-37.
Title and abstract are below.
Learning to parse syntactic dependencies
Much psycholinguistic research has investigated the relation between distributional information and expectations in language comprehension. One way to study this relation is to examine how expectations arise during sentence processing in adults as a function of relevant distributional information in corpora. An alternate approach is to explore how such predictive behaviors grow in children as a result of relevant distributional information that accumulates over a longer period of time. This presentation will discuss a series of studies that explore the second approach, with a focus on learning and processing of non-local dependencies. The first part of the talk explores the role of transitional probability information (and its breakdown) in learning a dependency between 'is' and 'ing' (e.g., John is kick-ing the ball). It is shown that 15 month old infants, who generally do not demonstrate sensitivity to the co-occurrence of is and ing, can use statistical (ir)regularities and learn to detect the co-occurrence relationship after a few minutes of exposure to input in the lab. The fact that children rely on distributional information to discover non-local dependencies makes it feasible that the acquired representations with statistical information may guide parsing of such non-local dependencies. The second part of the talk presents developmental research on comprehension of various wh-questions. It is shown that children's syntactic expectations in filler-gap dependency processing may be slightly different from the type of expectations that have been observed in adults, despite the fact that the distributional information available for children should fully support adult-like expectations. I will discuss implications of such findings for psycholinguistic models of syntactic predictions, as well as relevant linguistic and cognitive factors that may need to grow before adult-like predictive behaviors emerge.
