Ann Lawson
--------------
Improving Dictionary Coverage:
Conclusions drawn from a corpus-based validation study

Cyril Belica and Ann Lawson
Institut f�r deutsche Sprache, Mannheim, Germany

This paper describes the recent corpus-based validation of dictionary
material and examines the methodology, results and the conclusions to be
drawn for future work. The paper comprises three parts, describing first
the practical study undertaken, followed by an evaluation of the results
and a discussion of possible alterations or improvements. It then
concludes with an examination of the implications to be drawn from this
study for future work.

The original study was commissioned by the Klett publishing house as a
corpus-based validation of the German part of the database from which
their bilingual dictionaries are generated. The original lemma-list
contained around 60,000 entries. The corpus resources and tools used in
the study are based at the IDS, the Institut f�r deutsche Sprache, and
include part of the COSMAS corpus resources (around 200 million words for
the purposes of this study) and the IDS-Toolbox of cognitive- and
statistically-based programs and algorithms. The study examined
chronological frequencies, "missing" headwords, found in the corpus but
not in the dictionary, and "redundant" headwords, that is, headwords for
which no examples could be found in the corpus. The resulting lemmas were
extensively documented, with typical contexts, concordances, relative
frequencies and collocations. The relative chronological frequency of each
headword was determined using a diachronic reference corpus of 40 million
words compiled for the purpose.=20

Since an exhaustive linguistic evaluation of the results attained, and a
description of the use of the results in future dictionary editions, has
yet to be completed by the lexicographers involved, we undertake an
initial survey and assessment, with special reference to collocational
information extracted for selected lemmas. The significance of word class
and other linguistic factors (such as extreme high or low frequency) and
their influence on the results is investigated. In particular, perceived
shortcomings or omissions in the collocational information are explored in
order to uncover their origin.

Methods of improving the algorithms for future work are outlined. It is
planned to demonstrate how the fine-tuning of the algorithm in specific
contexts has led to improvements in the recall and relevance of the data
extracted. As a case in point, the influence of span variation according
to word class and context is investigated. A brief report will be made on
the ongoing investigation of the implications of this work in the
language-independent sphere.


**********************************************************************
Dr Ann Lawson     =09=09 Multilinguale Forschung
TELRI-II/SIMPLE/DHYDRO=09=09 Abteilung LEXIK
lawson@ids-mannheim.de           Institut f�r deutsche Sprache
Tel: +49 621 1581 427            R5, 6-13
Fax: +49 621 1581 415            D-68161 Mannheim
**********************************************************************