| |
Abstract:
Most current theories of sentence processing assume that
parses are built incrementally under the guidance of
phrase-structure rules and (in some cases) other grammatical
constraints. Consequently, the debate about serial versus parallel
processing has focused on cases of initial substring ambiguity. But
there is another kind of ambiguity that warrants consideration:
local ambiguity, or ambiguity of a noninitial substring that is
disambiguated by preceding material. The prediction that the parser
should be sensitive to local structural ambiguity distinguishes
connectionist models (e.g., Elman, 1991) and some statistical
models involving smoothing (e.g. Charniak, 1993) from the standard
models based on incremental, grammatically-consistent parsing. If
the parser is immediately sensitive to local ambiguity as well as
to initial ambiguity, then parallelism of an especially inclusive
sort may underlie parsing.The simplest case is local lexical
ambiguity. It has long been known that multiple interpretations of
syntactically ambiguous lexical items can support priming even when
prior context selects a single reading (Tanenhaus et al., 1979).
Such effects are predicted by models which assume an
activation-based lexicon, but require no new assumptions about
parsing proper. Recently, Tabor and Richardson (1999) found that
local ambiguity involving two words in sequence ("...fire truck..."
in example 1) slowed readers down in word-by-word self-paced
reading relative to a nonambiguous control ("...hire truck...").
The two-word local ambiguities suggest that the parser is involved
but since the two words form a lexical compound, the result could
also stem from an activation-based lexicon model.
Two new word-by-word self-paced reading experiments support the
hypothesis that local ambiguity involving nonlexical multi-word
sequences can influence the parser. Experiment 1 considered the
local ambiguity of sequences like "the waitress served pea soup" in
the reduced versions of sentences like (2). The reduced versions of
(2) were both read more slowly than the unreduced, but the
reduction effect with an ambiguous embedded verb ("served") was
significantly greater than with an unambiguous ("given") for the
region of words 7-8 (f1(1, 41) = 9.58, p < .005; f2(1, 17) =
6.13, p < .05). These results suggest that readers were
distracted by the clausal interpretation of words 3, 4, 7(ambig),
8, 9 even though this interpretation is ruled out by the preceding
syntax. Experiment 2 considered the local ambiguity of sequences
like "rolls with the punches" in sentences like (3). These words
form an idiom, but the idiomatic reading is ruled out by the
preceding syntax. In this case reading times during the local
ambiguity (words 4-7) were FASTER in the ambiguous versions
("...rolls...") than the nonambiguous ("...tarts..."), though the
result was significant in both subject and item analyses only when
we limited consideration to 14 idioms with identical syntactic
structure (f1(1, 30) = 13.79, p < .001; f2(1, 13) = 5.44, p <
.05). This result is aligned with evidence that idioms are
processed faster than nonidiomatic controls (Gibbs, 1986), but it
is a surprising case because the idiom interpretation is
categorically ruled out by the preceding context.These cases
together suggest that the parser engages in a more inclusive
structural parallelism than is standardly assumed.
(1) We feared they would fire/hire truck drivers
without consulting the union.
(2) Katya watched the waitress (who was)
served/given pea soup by the trainee.
1 2 3 4 5 6 7 8 9 10 11 12
(3) Grandma served the rolls/tarts with the
punches for a snack on Saturday.
1 2 3 4 5 6 7 8 9 10 11 12
|
