|
Abstract:
Although word-order flexibility has generated a vast
literature in theoretical linguistics, few studies have
investigated the processing of sentences with scrambled
(non-canonical) word order. In Japanese, results have been
contradictory (Mazuka, Itoh and Kondo, 1998; Nakayama, 1995;
Yamashita, 1997) and could be explained by independent factors.We
report a series of self-paced reading experiments using the
following construction in Japanese.
1a. Canonical: Subject NP-dat NP-acc ditransitive-verb
1b. Scrambled: Subject NP-acc NP-dat ditransitive-verb
(Ditransitive verbs such as ``give'' and ``introduce'' were
used. NP-dat and NP-acc stand for object NPs marked with dative or
accusative case-markers respectively.)
(1a) is the canonical word order (e.g., Hoji, 1985). In (1b),
the NP-acc was scrambled to the position prior to the NP-dat.
We argue that the reading time patterns obtained can be most
naturally explained assuming that the positing of a trace (an empty
category at the canonical position of the scrambled NP-acc) in (1b)
leads to longer reading times. Crucial in this conclusion are the
loci of the slow-down (always in the region where a trace should be
created by the parser) and the nature of the difficulty (which is
influenced by the distance between antecedent and trace).
The present results are unlikely to be explained in terms of
difficulty in directly associating a predicate to its arguments as
in categorical grammar (Pickering and Barry, 1991) because all the
differences observed occur before the actual predicate is read.
Frequency accounts cannot easily explain the interaction between
memory-load and scrambling, or the loci of the slow-downs observed.
Moreover, the results are incompatible with proposals under which
(1b) is also canonical (Miyagawa, 1997).
In Experiment I, in order to increase the overall memory load
(Just and Carpenter, 1992; Gibson, 1998), the ditransitive clauses
in (1a,b) were embedded within two clauses and intermixed with
complex fillers requiring careful reading. Participants took longer
to read the NP-dat in the scrambled condition than the NP-acc in
the canonical condition (ps < 0.05). No differences were
observed in the other regions.In Experiment II, the NP-acc remains
in its canonical position in (2a,c); whereas the locative in (2d)
increases the distance between the scrambled NP-acc and its trace,
when compared to (2b).
2a) Canonical: Subject LOCATIVE NP-dat NP-acc ditran-verb.
2b) Short scrambling: Subject LOCATIVE NP-acc NP-dat ditran-verb.
2c) Canonical: Subject NP-dat LOCATIVE NP-acc ditran-verb.
2d) Long scrambling: Subject NP-acc LOCATIVE NP-dat
ditran-verb.
Reading times at the object prior to the verb are crucial
because that is where a trace may be posited. The 2x2 design avoids
confounds such as differences in case-markers compared, and overall
distance from the arguments to the verb.
Longer reading times were detected at the NP-dat in (2d) than in
(2b), but not at the NP-acc in (2c) versus (2a) (interaction bd X
ac, p1 < 0.05; p2 = 0.09). This suggests that the positing of
the trace is harder when the NP-acc is farther away. (The detection
of a slow-down at the locative head in (2c) indicates where
locatives can be base-generated in relation to the
NP-dat.)Experiments using cleft constructions and relative clauses
will also be reported.
Mazuka, R., Itoh, K., and Kondo, T. 1998. Cost of scrambling in
Japanese sentence processing. Poster presented at the 11th Annual
CUNY Conference on Human Sentence Processing. New Brunswick, New
Jersey.
Nakayama, M. 1995. Scrambling and probe recognition. In R. Mazuka,
and N. Nagai (Eds.), Japanese sentence processing. Lawrence
Erlbaum.
Pickering, M., and Barry, Guy. 1991. Sentence processing without
empty categories. Language and Cognitive Processes 6, 229-259.
|
