Richard Lewis
University of Michigan
What kind of memory system supports the incremental construction of novel
compositional structures in sentence processing? I review evidence for a
working memory that has surprising characteristics contrasting sharply
with traditional views of verbal STM. Two of the most significant
constraints are that there is a distinguished capacity of only about two
(not 7 or even 4) items, and serial order information is not available
sufficiently quickly or reliably to be used in real-time comprehension.
How can a parser operate under such constraints, and still account for
the amazing functional capacity of human linguistic processing? I
describe one theoretical approach that attempts to address this puzzle:
cue-based parsing. The approach is embodied in a computational model that
incorporates two independently motivated constraints on memory
(activation decay and similarity-based interference). The theory yields
a novel combination of cross-linguistic empirical predictions, including
locality and anti-locality effects, similarity retrieval interference,
and intrusions of the ungrammatical on the grammatical. It also provides
a surprising answer to the serial order puzzle: serial order information
is rarely needed in sentence parsing, and the special cases where it is
needed correspond to a well-known class of structures that people fail to
comprehend.
Reception to follow in 1413 Marie Mount Hall.