|
Norsk
versjon her
The aim of this project is the further development and testing of a
method for the automatic derivation of wordnets (semantic nets, concept
nets - i.e., semantically classified lexical databases) from translational
corpora. The method
has been developed by Helge Dyvik.
Wordnets are
a language technology resource of increasing importance, with several
applications. Among other things, they allow content-based information
retrieval, automatic logical inference, and improved machine translation.
Parallel corpora are text collections consisting of originals and translations
in two or more languages, where the originals and their translations have
been aligned on the level of sentences, or more rarely also on the level
of words.
The method takes
translational correspondences from a parallel corpus as a starting point.
On the basis of the network of translational correspondences, word senses
are distinguished and semantic relations are calculated automatically,
e.g. hypero- vs. hyponyms (animal vs. dog, good vs. kind), and the result is represented in a complex
lattice structure. The aim of the project is to apply and test the method
on a large scale against a Norwegian/English parallel corpus (ENPC).
This involves among other things word alignment of the corpus, extraction
and processing of data from the corpora, and evaluation of the algorithms
and the derived lattices.
A successful
result would mean that parts of the work towards the development of wordnets
can be automatised.
|
|
Project
description
(in Norwegian)
Project leader:
Helge Dyvik
Project participants:
Knut Hofland
Paul Meurer
Sindre Sørensen
Martha Thunes
Project period:
April 2001-March 2004
Financing:
2001-2002 financed by
L. Meltzers høyskolefond.
2002-2004 financed by
The Research Council of Norway
Papers
Web Demo
|
