Campus Event Calendar
Event Entry
What and Who
Incremental Nonmonotonic Parsing throughSemantic Self-Organization
Marshall Mayberry
University of Texas
Talk
AG 1, AG 2, AG 3, AG 4
Expert Audience
Note: We use this to send email in the morning.
Date, Time and Location
Friday, 10 January 2003
10:00
-- Not specified --
17.3 - Computerlinguistik
2.11
Saarbrücken
Abstract
I present a subsymbolic computational model of natural language
processing that has been trained on a treebank of semantic
representations. Several criteria have constrained my approach: the
model should be
* cognitively plausible, exhibiting human-like behavior in processing,
as well as human-like errors when lesioned.
* incremental so that the model can exhibit cognitively plausible
behavior such as semantic priming in the course of processing
natural language.
* a multiple-access model of disambiguation. That is, the model should
maintain as many interpretations of a sentence fragment as warranted
by their statistical likelihood as ambiguous information is
encountered in the course of processing input. When
disambiguating input allows resolution, the most active
interpretation should be the most prominent.
* nonmonotonic. The model should allow for revision (addition or
elimination) of semantic components as licensed by the input during
sentence processing.
* robust to incomplete, ungrammatical, and dysfluent input. Whereas
a symbolic parser will return failure on such input, unless
explicitly and exhaustively programmed otherwise, a subsymbolic
parser will assign an interpretation to the input that is graded
according to the difficulty of processing that input.
The corpus I have been using in my research is the recently-developed
LinGO Redwoods Head-driven Phrase Structure Grammar (HPSG) Treebank of
5000 VerbMobil utterances. I have chosen to train my system on the
Minimal Recursion Semantics (MRS) representations of these sentences
because MRS, with its emphasis on underspecified, flat semantics,
proves to be perfectly commensurate with the objectives of my research
as described above. Consequently, my work has demonstrated that a
connectionist approach is quite capable of picking up on the bindings
implied by the subcategorization frames of MRS representations while
still retaining the cognitive plausibility of a subsymbolic model.
Contact
--email hidden
passcode not visible
logged in users only