|
Abstract:
Many theories of reanalysis assume some kind of Minimal Change
Strategy (MCS) according to which the human parser reanalyses by
changing the mental representation as little as possible (Inoue
(1991), Frazier (1994)).
This poster argues that the parser makes a change in the tree
structure if and only if it allows for more grammatical
constraints to be satisfied (MaxG for Maximal Grammaticality).
The MCS is captured by the necessary condition: if a change
doesn't increase grammaticality, then don't do it. However, MaxG
predicts that the MCS can be overridden as the parser tries to
satisfy an optional constraint and therefore changes the
representation more than the strictly necessary.
The following Japanese sentences were used as stimuli in a
self-paced reading experiment.
(1a) Shokuin-wa kakarichou-ni ocha-o onnanohito-ga dashita-to
shiraseta. employee-top manager-dat tea-acc woman-nom served-comp
told "The employee told (the manager) that the woman served tea
(to the manager)."
(1b) Shokuin-wa onnanohito-ga kakarichou-ni ocha-o dashita-to
shiraseta. "The employee told that the woman served tea to the
manager."
Native Japanese speakers read "served" in (1a) more slowly
than in (1b). (p < 0.05 by subjects and items.) This slow-down
confirms the hypothesis that, prior to "served," subjects assume
that the nominative NP "woman" is the beginning of the embedded
clause. Therefore, in (1a), "manager" and "tea" are interpreted
as being part of the main clause; reanalysis occurs at "served"
so that "tea" becomes part of the embedded clause. Reanalysis
doesn't occur in (1b) because "tea" comes after "woman" and
therefore it is interpreted as being part of the embedded clause
from the start.
However, it's not clear in this experiment whether, during
reanalysis in (1a), "manager" remains in the main clause or is
attached to "served." MaxG predicts the latter. Whereas MCS
predicts the former because the dative NP is an optional argument
(Kamide 1997); and the displacement of "tea" alone is sufficient
to correct the structure.
In an off-line experiment, subjects judged whether "manager"
attached to "served" or "told" in sentences like (1a) above and
(1c) below.
(1c) Shokuin-wa kakarichou-ni onnanohito-ga ocha-o dashita-to
shiraseta. "The employee told (the manager) that the woman served
tea (to the manager)."
In (1c), reanalysis doesn't occur because "tea" is already
inside the embedded clause.
As predicted by MaxG, attachment to "served" is stronger in
(1a) than (1c). (p < 0.01 by subjects and items.) Therefore,
reanalysis in (1a) not only attaches "tea" but also "manager" to
"served" as the parser tries to maximally satisfy the
subcategorization frame of the verb rather than sticking to the
minimal change.
Frazier, L. (1994). Sentence (Re-)Analysis. Unpublished
manuscript.
Inoue, Atsu. (1991). A comparative study of parsing in English
and Japanese. PhD dissertation, University of Connecticut.
Kamide, Yuki. (1997). Verb-argument structures in resolution
of thematic attachment ambiguity. Poster presented at the 10th
Annual CUNY Conference on Human Sentence Processing.
|
