|
Abstract:
Comparing the processing of globally ambiguous sentences with
disambiguated structures provides important insights into the
architecture of the sentence processor. For example, some prominent
constraint-based theories (McRae et al, 1998; Spivey &
Tanenhaus, 1998; Tabor et al., 1997) predict that the processing of
globally ambiguous sentences should be disrupted because of a
competition between possible alternatives.
In contrast, serial models predict that ambiguous sentences
should be easy, because the initially adopted structure never has
to be revised. Moreover, ¡race based¢ proposals of
serial architectures (cf. Van Gompel et al., 1999; Lewis, 1999)
maintain that the parser determines its initial choice on a
probabilistic basis, such that reanalysis may -- on a proportion of
trials -- even occur in structures that unambiguously support the
initial bias (Traxler et al., 1998; Van Gompel et al., 1999).
So far, support for probabilistic serial parsing comes from
adjunct attachment ambiguities. It is unclear whether similar
patterns of results can be found for structures that involve
ambiguous arguments. We therefore conducted an eye-tracking
experiment in German which compared sentences containing
syntactically ambiguous arguments (1) -- where the critical
wh-phrase "welches Huhn" can either be the subject or the object of
the sentence -- with sentences in which the wh-pronoun is
unambiguously case-marked as the subject (2) or as the object (3).
First-pass reading times at the wh-pronoun were significantly
faster in (1) than (2) or (3), consistent with the predictions of
the race-based account. In following regions, unambiguous
subject-first sentences (2) were read faster than unambiguous
object-first sentences (3), confirming previous findings from the
literature. Importantly, at no point in the sentence, the ambiguous
condition (1) was harder than either of the disambiguated
conditions (2 or 3).
(1) Der Knecht wollte wissen, welches Huhn gestern die Ente
verscheucht hat. (globally ambiguous)
The farm-hand wanted to know, which [amb.] hen yesterday the duck
chased-away has.
(2) Der Knecht wollte wissen, welcher Hahn gestern die Ente
verscheucht hat. (subject first)
The farm-hand wanted to know, which [subj.] rooster yesterday the
duck chased-away has.
(3) Der Knecht wollte wissen, welchen Hahn gestern die Ente
verscheucht hat. (object first)
The farm-hand wanted to know, which [obj.] rooster yesterday the
duck chased-away has.
A second experiment, which employed word-order ambiguities such
as (4-6), showed similar results.
First-pass reading times for the NP "die hungrige Fuechsin" in
(4), which is ambiguous between a subject and object
interpretation, were faster than for "der hungrige Fuchs" (5),
which is an unambiguous subject and "den hungrigen Fuchs" (6),
which is an unambiguous object. Furthermore, first-pass reading
times for the subject-first condition were shorter than for the
object-first condition.
(4) Die hungrige Fuechsin sah ... / The hungry vixen [amb] saw
...
(5) Der hungrige Fuchs sah ... / The hungry fox [subj] saw ...
(6) Den hungrigen Fuchs sah ... / The hungry fox [obj] saw ...
The results from both experiments are most consistent with
probabilistic race-based parsers which claim that reanalysis may
occur in both unambiguous conditions. The results do not support a
competitive constraint- based architecture, nor a
¡conventional¢ (non-probabilistic) serial processing
account.
|
