| |
Abstract:
Knowing how to use a verb requires, among other things,
knowing what sorts of argument structures may be used with that
verb. Thus, verbs may subcategorize for arguments such as subject,
direct object, and sentential complement. Furthermore, verbs differ
with regard to the frames they permit and the probability with
which they occur in various frames (Connine et al.). Garnsey et al.
(1997) and Pickering and Traxler (1998), for example, have
demonstrated empirically that comprehenders are sensitive to the
frequency with which a particular frame occurs for a given verb,
and that they use their knowledge of a verb's subcategorization
tendencies during normal on-line sentence interpretation.What has
not widely been acknowledged, however, is that subcategorization
frame probabilities can differ for various senses of a verb. Thus,
verbs such as "admit" may occur with either a direct object (DO) or
a sentential complement (SC), but its subcategorization tendencies
differ by sense (the "let in" sense tends to occur with a DO,
whereas the "confess" sense tends to occur with a SC). Roland and
Jurafsky (in press) have demonstrated using corpus analyses that
these sense differences can be significant.
The goal of the present study was to investigate whether (a)
during sentence production, the choice of a verb's
subcategorization frame might be influenced by a preceding context
that is biased toward a specific sense of the verb (given that
sense is correlated with subcategorization frame probability), and
(b) while reading sentences containing a temporary structural
ambiguity, sense-biasing contexts might influence a reader's
expectations.
The study consisted of three components. (1) We analyzed three
parsed text corpora (Wall Street Journal, Brown, WSJ87) to identify
verbs that occur with both DO and SC frames. From these, we
selected 45 verbs that showed a meaning difference associated with
the difference in arguments, based on WordNet (Miller et al.,
1990). (2) Subjects completed sentence fragments containing these
45 verbs. Two contexts were created for each verb, one favoring a
sense associated with a DO frame and the other favoring a SC-biased
sense. An example is presented below. Subjects were shown the
context sentence plus a fragment of the target that ended at the
verb ("admitted"). Twenty-four verbs for which the context
significantly influenced the structure of the completion were
chosen for a reading-time study. (3) In a self-paced moving-window
reading time study, each verb occurred in a target sentence
containing either a structurally unambiguous or a
structurally-ambiguous SC ("that" present versus absent). The
target sentence was preceded by either a DO or SC sense-biasing
context. Context interacted with ambiguity at both the NP-region
("the students") and the second disambiguating region ("chance
of"). These off-line and on-line results shows that people are
sensitive to sense-based subcategorization preferences, and that
they use this knowledge during normal language comprehension. We
are currently simulating these data using a constraint-based model
similar to that of McRae, Spivey-Knowlton, and Tanenhaus
(1998).
SC-context: Up until the final exam, the teacher tried to
reassure the two discouraged freshmen about their chances of
passing the course.
DO-context: The two freshmen on the waiting list had begged the
teacher to let them into the class.
Target: Finally, she admitted (that) the students had little
chance of succeeding.
|
