|
CNL Lunch Talks Masaya Yoshida Relative Clause Prediction in Japanese Thursday March 11th, 2004, 12:30 PM, 3416 Marie Mount Hall |
The head-final property of Japanese presents many potential hazards for incremental parsing,
among which the head-final relative clause (RC) structures are perhaps the most notorious. There
is normally no indication of the RC structure until the end of the RC, and this gives rise to widely
discussed cases of processing difficulty (e.g. Inoue, 1991; Mazuka & Itoh, 1995). In this paper
we present three experiments that show that Japanese speakers are able to use cues from numeral
classifiers to anticipate an upcoming RC structure and hence avoid the processing difficulty
normally associated with RCs. Furthermore, these same cues are sufficient to block the formation
of long-distance dependencies that would violate island constraints on scrambling.
The presence of an RC in Japanese normally cannot be detected until the parser reaches
the head of the RC, or in unambiguously embedded structures the complementizer-less verb (1).
(1) [Taroo-ga gap yonda]
hon Taroo-Nom read book
The book that Taroo read
Our studies explore the consequences of the fact that genitive numeral classifiers associated with
the head of the RC may precede the RC. There are cases where the numeral classifier is
semantically incompatible with the subject of the RC, as in (2), where the numeral classifier
(satu) and its potential local host NP (sensee) are mismatched. This mismatch may provide a cue
to the presence of an RC structure.
(2) [san-satu-no [sensee-ga yonda] hon]
three-Classifier(book)-Gen teacher-Nom read book
three books that the teacher read
(san-satu means roughly three copies.)
Experiment 1 (sentence fragment completion, n=121) showed that locally matching vs.
mismatching numeral classifiers successfully regulate expectations for RCs. In classifier mismatch
conditions (3a) 86.5% of completions involved RCs, whereas in classifier-match
conditions (3b) there were almost no RC completions (0.02%).
(3) a. Dono NP-ni NP-top san-satu-no Adj sensee-ga
which NP-Dat three-classifier(book)-Gen teacher-Nom
b. Dono NP-ni NP-top san-nin-no Adj sensee-ga
which NP-Dat three-classifier(human)-Gen teacher-Nom
Experiment 2 (self-paced reading) showed that information from mismatching numeral
classifiers can be used on-line to avoid classic garden path effects associated with RCs. Reading
times at the embedded verb, which disambiguates in favor of the RC structure, showed a
significant facilitation in the classifier-mismatch condition (4a) relative to the classifier-match
condition (4b), F1(1, 40)=4.4, p<.05, F2(1, 23)=4.2, p<.05.
(4)a. NP-top / san-satu-no / Adj / sensee-ga / Adj / NP-Dat / Adv / V / hon-o / NP-Dat / V.
three-Cl(book)-Gen teacher-Nom book-Acc
(4)b. NP-wa / san-nin-no / Adj / sensee-ga / Adj / NP-Dat / Adv / V/ hon-o / NP-Dat / V.
three-Cl(human)-Gen teacher-Nom book-Acc
We next investigated whether information from mismatching numeral classifiers not only
predicts an upcoming RC structure, but also triggers syntactic constraints involving RCs.
Previous studies on Japanese parsing show that speakers favor a long-distance scrambling
analysis of fronted dative NPs (Aoshima et al., 2003). This preference is potentially in conflict
with the constraint on scrambling out of RCs (Saito, 1985). Experiment 3 (self-paced reading)
replicated Aoshima et al.fs finding of a pre-verbal Filled Gap Effect in conditions with matching
numeral classifiers (6), but showed that this effect disappeared in conditions where mismatching
classifiers indicate an upcoming RC structure. The Filled Gap Effect (Crain & Fodor, 1985;
Stowe, 1986) was observed in a slowdown immediately after the embedded dative NP in the
scrambled-match condition (6a) relative to the unscrambled-match condition (6b), F1(1, 86)=4.6,
p<.05, F2(1, 23)=5.02, p<.05. This effect is expected if the fronted dative NP undergoes long-distance
scrambling that places it inside the embedded clause. No such contrast is found in the
classifier-mismatch conditions (5a, 5b), Fs<1, indicating that long-distance scrambling is
blocked when classifiers cue an upcoming RC. Thus, island constraints on movement apply
immediately in Japanese, as found in English (e.g. Stowe, 1986; McElree & Griffith, 1998),
despite the extreme rarity of these constructions in Japanese.
(5) Classifier Mismatch Conditions
a. Wh-Dat / NP-Top / three-Cl(book)-Gen / Adj / teacher-Nom / Adj / NP-Dat / Adv / V / book-Acc / V-Q?
b. NP-Top / wh-Dat / three-Cl(book)-Gen / Adj / teacher-Nom / Adj / NP-Dat / Adv / V / book-Acc / V-Q?
(6) Classifier Match Conditions
a. Wh-Dat / NP-Top/ three-Cl(human)-Gen / Adj / teacher-Nom / Adj / NP-Dat / Adv / V / book-Acc / V-Q?
b. NP-top / wh-Dat / three-Cl(human)-Gen / Adj / teacher-Nom / Adj / NP-Dat / Adv / V / book-Acc / V-Q?
|
