|
Abstract:
In order for interpretation to be incremental in language
processing, structure building must also be immediate and
incremental. In any situation where XP material has been reached
before the head that licenses it (e.g. Spec-initial phrases), it
is not immediately clear how to structure phrases, so some
predictions must be made. Most parsers (including all that use
standard phrase structure rules) predict categories. However, as
will be shown in this paper, by changing the predictions from
categories to features, a parser can make highly specific, fully
warranted predictions without making unwarranted commitments or
fully underspecifying when a category prediction is not
warranted. For example, if an NP is seen at the beginning of an
S, a VP might be predicted, while in this theory a
Nominative-assigner (i.e. Tense) would be predicted. In German
sentences like (1) and (2), a parser that predicts categories
will be necessarily wrong in at least one of the two cases (e.g.
a garden path theory would have no prediction consistent with
both possibilities.)
(1) Ich habe den Hund dem kleinen Kind gegeben. I have the
Dog-ACC the small child-DAT given "I have given the dog to the
small child."
(2) Ich habe den Hund dem kleinen Kind entfliehen gesehen. I
have the dog-ACC the small child-DAT help seen "I saw the dog
help the small child."
By reducing the granularity of the predictions, this parser is
able to build a single fully-connected phrase marker that
requires no revisions for either continuation of the
sentence:
(3) I have [the dog-ACC [the small child-DAT Assign-DAT]
Assign-ACC]...
A dative-assigner and an accusative-assigner are predicted,
but there are no presumptions made about the number or type of
heads that will do the assigning. Whatever verb comes next will
be merged into the lowest head (Assign-DAT), and will unify with
the next predicted head if possible. Other features, such as
Agreement and theta roles, can also be predicted.
This theory predicts that any time one or more XP's (e.g. the
NP's above) can be licensed in multiple ways by material on the
right, they should all be possible as long as there is a common
denominator (a feature or features) for all the possibilities.
This power will be useful even in English adverbials, which can
be attached to either adjectives or verbs. This contrasts with
other theories where one category must be chosen (even before it
can be confirmed.)
Chomsky N. (1995).
The Minimalist Program.
MIT Press, Cambridge, MA.
Fodor, J.D., and Inoue, A. (1994). "The Diagnosis and Cure of
Garden Paths."
Journal of Psycholinguistic Research,
23:5, 407--434.
Frazier, L. (1979) On comprehending sentences, syntactic
parsing strategies. Indiana University Linguistics Club.
Gorrell, P. (1995)
Syntax and Parsing.
Cambridge University Press: : Cambridge, England.
Marcus, M., Hindle, D, and Fleck, M. (1983) "Talking about
talking about trees."
Association for Computational Linguistics,
21, 129--136.
|
