N. Nicolov, Univ.of Sussex

Lecture
Oesterreichisches Forschungsinstitut fuer Artificial Intelligence (OeFAI)
                      Schottengasse 3, A-1010 Wien
                       Tel.+43-1-53532810,5336112
-------------------------------------------------------------------------

                          gemeinsam mit der

         W i e n e r   S p r a c h g e s e l l s c h a f t :



                                VORTRAG
                                *******


Nicolas NICOLOV
University of Sussex


Parsing Techniques for D-Tree Grammars
--------------------------------------

This talk presents ongoing work on the LexSys system for "Analysis of
Naturally-occurring English Text with Stochastic Lexicalised Grammars"
which is being built in the context of a large project at the
University of Sussex. 

D-Tree Grammar (DTG) is a variant of Tree Adjoining Grammar (TAG). DTG
is designed to share some of the advatages of TAG while overcoming
some of its limitations. DTG assumes similar elementary structures as
in TAG but combines them in a more uniform way. The elementary (and
intermediate) structures are viewed as descriptions of trees and are
composed by equating certain nodes. 

I will briefly introduce the current version of D-Tree Grammar
formalism. I will present the architecture of the system and will
discuss the approach for encoding a wide-coverage lexicalised grammar.
In the rest of the talk I will present a bottom-up parser for D-Tree
Grammars. Advantages from using DTGs for parsing include:

1. Elementary structures have an extended domain of locality and
   allow certain kinds of dependencies to be stated directly which in
   other frameworks can only be enforced by a mechanism of percolation
   of feature values; 

2. A DTG parser need only consider the lexicalised d-trees
   corresponding to the input words; 

3. DTG derivation structures directly reflect predicate argument
   structure due to the uniform treatment of complementation and
   modification.

Time permitting I will present how we are addressing the problem of
duplication of parsing steps by sharing computations across elementary
structures and how we might take into account probabilistic
information. I could also discuss the kind of changes we are
considering to introduce to the linguistic analyses.


Zeit: Dienstag, 5.Mai 1998, 18:00 Uhr c.t.
Ort:  *ACHTUNG*: Seminarraum des Instituts fuer Sprachwissenschaft,
      Berggasse 11, 1.Stiege, 3.Stock
      *******************************


OESTERREICHISCHES FORSCHUNGSINSTITUT
FUER ARTIFICIAL INTELLIGENCE

o.Univ.-Prof.Dr.Robert Trappl