|
|
|
|
|
|
SEAL: Structural Enhancement of Automatically-acquired Lexicons
|
|
Overview
|
|
This project is concerned with the development of practical, large-scale
lexicons for use in computer applications that use natural language.
While a number of large-scale lexical fragments already exist
(hand-crafted or derived from dictionaries or corpora), few of them are
practically useful. Reasons for this include insufficient content
density (eg syntax but no semantics), inadequate internal structuring,
or inappropriate level of detail for the application in hand (too much
irrelevant detail can greatly increase the processing load).
Rather than attempting to construct better lexicons from scratch, this
project is developing tools which use these existing lexicons as base
data, and support the development of new lexicons with greater content
density, enhanced structure and application-specific level of detail.
This is achieved by merging lexicons together, and by inducing
additional structure, guided by insights from lexical representation
theory.
The work is being evaluated through two pilot applications in areas
related to other research within the Institute, namely text generation
and information extraction.
|
|
Background
|
|
The recent commercially-motivated growth in interest in applied
computational linguistics (or Language Engineering) has
highlighted the need for realistic large-scale linguistic resources,
notably grammars and lexicons. Significant research activity has thus
been directed towards developing such resources, and a number of
large-scale lexical fragments are beginning to emerge (such as ACQUILEX,
CELEX, WordNet, XTAG). However these lexicons remain only of fairly
limited utility in practical applications. Reasons for this include:
- insufficient content density - most existing
lexicons focus on one particular kind of lexical
information (syntax, semantics, orthography
etc.) and do not provide enough coverage of
other apsects.
- inadequate internal structuring - lexicons tend
to be very `flat', yet structure is important
for maintenance and flexibility of
representation and access.
- inappropriate level of detail for the
application in hand - too little detail limits
the utility of the lexicon, too much detail that
is not relevant to the particular application
task often has severe processing penalties.
The early expectation of this line of research was that it would
ultimately lead to large general purpose lexicons suitable for use in a
wide range of applications. To achieve that goal, a significant amount
of manual post-processing to rectify the above problems might be
tolerable as a one-off enterprise. However, a more recent trend suggests
that the single common lexicon approach might not after all be the most
effective way forward. Rather, different application areas might be
better served by more tailored (but still large-scale) lexicons, each
acquired from domain-specific corpora and knowledge sources. This
focusses attention on the need for tools to address these problems many
times over, in different task domains.
|
|
The project
|
|
The present project is concerned with taking existing lexicons as base
data, and producing new lexicons with greater content density, enhanced
structure and optimal feature detail (relative to a given domain or
task), by merging lexicons and by inducing additional structure. Our
approach is motivated by the following assumptions:
- that theoretical work on lexical description
provides valuable insight into the structure of
large-scale practical lexicons
- that the application domain is limited - this
keeps the problems manageable, and offers scope
for investigation of domain-specific phenomena
The key features of the programme of research are
- the analysis of existing lexicons and structural
comparison with current lexical theory
- the development of software techniques and
subsequently tools for the induction of
additional structure in acquired lexicons
- the evaluation of the methods developed in
applications to natural language generation and
information extraction
The approach taken is to map existing lexicons into a common notation
and develop enhancement techniques in that notation. Both
data-driven (inducing additional structure from the data
provided) and theory-driven (seeking instances of theoretical
proposals in the data) approaches are being explored. Examples of the
kind of enhancements we are particularly interested in include the
following:
- construction of optimal inheritance
hierarchies, guided by metrics relating to both
efficiency (such as maximal generality) and
linguistic theory (such as default cases being
unmarked)
- induction of features theory, including feature
typing, and properties such as mutual
exclusivity, covariation and immutability.
- representation of lexical entry covariation as
abstract lexical relationships
|
|
Staff
|
|
The principal investigator is Roger Evans. Adam Kilgarriff
is also working on the project.
|
|
Financial Support
|
|
The project is supported by the EPSRC under grant GR/K/18931.
|
|
Publications
|
|
R.Evans and A. Kilgarriff, ``MRDs, Dictionaries, and How
To Do Lexical Engineering'' in Proceedings of the 2nd Language
Engineering Convention, pp. 125-132, London, UK, 1995.
A. Kilgarriff, ``Which words are particularly characteristic of a
text? A survey of statistical approaches''
in Proceedings, AISB Workshop on Language Engineering for
Document Analysis and Recognition. Brighton, UK, 1996.
Adam Kilgarriff and Raphael Salkie ``Corpus similarity and homogeneity via word frequency.'' Proceedings of Euralex '96 Gothenberg, Sweden. 1996.
Adam Kilgarriff ``Putting frequencies in the dictionary.''
International Journal of Lexicography. Forthcoming.
Maintained by
Roger Evans
(Roger.Evans@itri.brighton.ac.uk).
Last updated 12 January 1997
©Information Technology Research Institute
ITRI home page |
ITRI research overview
|