The application of annotation models for the construction of databases and
tools: Overview and analysis of MPI work since 1994

Hennie Brugman, Peter Wittenburg
Max Planck Institute for Psycholinguistics
Nijmegen, The Netherlands

At the Max-Planck Institute for Psycholinguistics [1] software development
in the area of linguistic annotation has been ongoing since 1994. This has
resulted in a few generations of models, linguistic databases and software
tools. This paper gives an overview and critical evaluation of this
evolutionary process and relates it to recent efforts to attain
standardization and cooperation.  One of the first tools to be developed at
the MPI is MediaTagger [2], a Macintosh program for annotation of Quicktime
movies. MediaTagger's implicit data model already supported multiple,
user-definable and typed tiers, time constraints between annotations on
different tiers, and association of closed vocabularies for annotations
with tiers.  Shortly after MediaTagger's first release, a first formal data
model for linguistic annotations was constructed, applying methods from
database design. This relational model was implemented with
Oracle. MediaTagger was extended to support working with this database
system in client-server mode over ODBC [3]. A graphical tool for the
specification of multilayer, temporal queries was designed and built using
OracleForms and OracleGraphics. This database currently contains
approximately 80,000 annotations on 3,000 tiers associated with 500 digital
video movies. However, there are considerable limitations to this solution.
Consequently, in 1997 the EUDICO (European Distributed Corpora) project was
initiated [4]. Its aim is to build a software framework for the creation
and exploitation of linguistic annotations of multimedia signals. This
framework should enable annotation file format independence, computer
platform independence, client-server operation on distributed archives on
the Internet and use of streaming media.  New software tools and annotation
file formats can also be added.  To establish this we chose to use a
three-layer architecture as commonly used for database systems. Tools from
the application layer use the generic services of objects existing in a
logical layer. These objects can be instantiated from a number of
differently formatted resources at the physical layer. The objects at the
logical layer are specified by an object-oriented model for linguistic
annotations, the Abstract Corpus Model (ACM).  The paper will present two
generations of the ACM and compare it to similar models, particularly the
ATLAS/AIF [5]. It will evaluate how far EUDICO's aims are achieved by each
generation of the model.  EUDICO technology is currently applied in the
Spoken Dutch Corpus project [6]; the DOBES project [7], and a recently
released annotation tool (EAT, Eudico Annotation Tool).  In the near future
we expect to add metadata support to the EUDICO framework by extending our
ACM model and by building in our IMDI [8] compliant metadata browser.

[1] http://www.mpi.nl
[2] http://www.mpi.nl/world/tg/CAVA/mt/MTgeneral.html
[3] http://www.mpi.nl/world/tg/CAVA/CAVA.html
[4] http://www.mpi.nl/world/tg/lapp/eudico/eudico.html
[5] http://www.nist.gov/speech/atlas/
[6] http://lands.let.kun.nl/cgn/home.htm
[7] http://www.mpi.nl/DOBES/
[8] http://www.mpi.nl/ISLE/