|
Abstract:
Right-branching languages such as Japanese can present a
challenge to models of sentence processing, since, for example,
relative clause constructions can only be parsed on the basis of
information occuring later in the sentence. Empirical studies have
indicated that Japanese speakers tend to parse strings such as (1)
as simple clauses until proven otherwise (e.g., Yamashita, 1994,
Kamide, 1998, Hirose, 1995).
In the present paper, we approach the process of mis-analysis
from a different perspective, i.e., by examining the cost of
avoiding a mis- analysis. Obviously, mis-analyzing (1) as a simple
clause will cost some processing resources when it needs to be
reanalyzed. However, it allows the parser not to maintain
unprocessed NPs, or to postulate an additional clause node at least
temporarily. If either of these operations requires significant
processing resources, then mis-analyzing (1) as a simple clause may
turn out to be an efficient strategy overall.
(1) John-ga Mary-o mikaketa.
John-Nom Mary-ACC saw
John caught sight of Mary.
(2) John-ga Mary-o mikaketa shoonen-o yonda.
John-Nom Mary-ACC saw boy-ACC called.
John called the boy who saw Mary.
(3) John-ga Mary-ga mikaketa shoonen-o yonda.
John-Nom Mary-Nom saw boy-ACC called.
John called the boy that Mary saw.
(4) John-ga Mary-o mikaketa takusii-ni noseta.
John-Nom Mary-Acc saw taxi-to put-on
John put Mary on a taxi that he saw.
(5) John-ga Mary-ni mikaketa shoonen-no hanasi-o sita.
John talked to Mary about a boy that he had seen.
Using an eye-movement monitoring experiment and a self-paced
reading experiment, four different types of relative clause
constructions plus a simple clause sentence (1 to 5) were
investigated. (1) is a simple clause with nominative and accusative
marked NP's as the verb's argument. (2 to 5) are relative clause
sentences. In (2), when the head noun 'shoonen' (boy) is received,
the first NP 'John-ga' can no longer be the subject of the first
verb 'mikaketa' (saw), and thus has to be reanalyzed as the
argument of a higher clause. In (3), the second NP has a nominative
case marker 'ga.' This signals the presence of a clause boundary
between the first and the second NP since only some verbs can take
two NPs with a nominative marker as its arguments. In (4),
combination of the head noun and the matrix verbs are such that
both the nominative-marked NP and the accusative-marked NP must be
reanalyzed as arguments of the matrix verb. In (5), the second NP
is marked by a particle 'ni.' The verbs in the study were selected
such that they cannot take a 'ni-' marked NP as a compliment. Thus,
when the first verb is received, the presence of the clause
boundary after the second NP becomes obvious.
Overall, Japanese speakers did not rate the difficulty of the
four types of relative clause sentences differently. However,
phrase-by-phrase reading times and regressive eye-movement patterns
from the two experiments showed that there is a significantly
increased processing cost at the second nominative-marked NP in
(3). However, (2) and (4) were more difficult at the point of the
head noun than (3). (4) and (5) were more difficult than the others
at the matrix verb, and at other positions in the second pass
analyses.
The results indicated that the cost of avoiding mis-analysis in
sentences such as (3) was substantial enough to cancel out the cost
of reanalysis in (2). This could be a driving force for Japanese
speakers' preference for a simple clause interpretation. The
results also showed that the Japanese speakers made a pre-head
(pre-verbal) parsing decision to insert a clause boundary.
|
