I used acceptability rating tasks and self-paced reading studies to
investigate three working memory-based theories' predictions regarding
sentence processing difficulty in Hindi center-embedding constructions:
Hawkins' Early Immediate Constituents (EIC), Gibson's Discourse Locality
Theory (DLT), and Lewis' Retrieval Interference Theory (RIT). Two main
issues were investigated: (a) the effect of definiteness marking on direct
objects; and (b) the effect of increasing head-dependent distance.
First, definite-marked direct objects were found to be harder to process
than bare (indefinite) direct objects, contra EIC, and contra DLT. I argue
that, due to discourse constraints, indefinites are harder to process when
they are in subject position, whereas definites are harder to process in
the direct-object position.
Second, regarding distance between heads and dependents, distance was
manipulated in two distinct ways: (a) by fronting indirect objects, and by
fronting direct objects in center embeddings like 'siitaa-ne hari-ko kitaab
khariid-neko kahaa', 'Sita told Hari to buy a book' ; and (b) by inserting
an adverb between the final NP and the innermost verb in canonical order
center embeddings.
One finding was that if distance is increased between heads and dependents
by reordering the dependents ((a) above), processing becomes more
difficult, as predicted by EIC and DLT, and contra RIT. However,
processing is, suprisingly, easier when distance is increased between the
head and its dependents by inserting an adverb between them ((b) above).
This goes against EIC, DLT, and RIT's predictions. I explain these results
as follows: fronting indirect or direct objects renders them more similar
to subjects (since fronted objects are in a typical subject position),
causing increased similarity-based interference between the actual subject
and the fronted object; by contrast, the easier processing due to adverb
insertion occurs because the adverb strengthens the activation level of the
current hypothesis (in working memory) regarding the sentence completion.
In sum, EIC, DLT, and RIT are only partly able to correctly characterize
important cross-linguistic aspects of human sentence parsing. This
incomplete coverage of the empirical results motivates a new, more general
model of human sentence parsing that correctly accounts for reading-time
and acceptability rating data from four languages.
|
E-mail this message to a friend
