Multi-Dimensional Markup: N-way relations as a generalisation over possible relations between annotation layers

Text-technological background
Multi-dimensional markup is a topic often discussed. The main
reasonwhy it is researched is the fact that the most important
markup languages today make the implicit assumption that for
a document, only a single hierarchy of markup elements needs
to be represented. Within the fi eld of Digital Humanities,
however, more and more analyses of a text are expressed
by means of annotations, and as a consequence, segments of
a text are marked up by tags relating to different levels of
description. Often, a text is explicitly or implicitly marked up
several times. When using the TEI P5 as an annotation scheme
one might use markup from different TEI modules concurrently
as ‘msdescription’ for manuscript description, ‘textcrit’ for
Text Criticism, and ‘analysis’ for (linguistic) analysis and
interpretation, because the Guidelines state that “TEI schema
may be constructed using any combination of modules” (TEI
P5 Guidelines).
Abstracting away from limitations of specifi c markup
languages, textual regions annotated according to different
levels of descriptions can stand in various relationsships to
each other. Durusau & O’Donnell (2002) list 13 possible
relationsships between two elements A and B used to
concurrently annotate a text span. Their list comprises the
cases ‘No overlap’ (independence), ‘Element A shares end
point with start point of element B or the other way round’,
‘Classic overlap’, ‘Elements share start point’, ‘Elements share
end point’ and ‘Element share both their start points and end
points’. The latter case is known under the label ‘Identity’. The
possible relationships between A and B can also be partitioned
differently, e.g. into Identity, Region A before region B, or the
other way round. Witt (2004) has alternatively grouped the
relations into three ‘meta-relations’ called ‘Identity’, ‘Inclusion’,
and ‘Overlap’. Meta-relations are generalisations over all the
13 basic relations inventorised by Durusau & O’Donnell.
The reason for introducing meta-relations is to reduce the
number of relations to be analysed to those cases that are
most typically needed when querying annotations of multiply
annotated documents. The query tool described in Witt et al. (2005) provides 7 two-way query predicates for the 13
basic relations from Durusau & O’Donnell (where e.g. the
two relations overlap(A,B) and overlap(B,A) are handled by
one query predicate) and specialised predicates for the three
meta-relations.
As argued above, often n-way relationships between elements
from three or more annotation layers need to be queried.
When the detailed accounts of cases of relations between
two elements described above are extended to cases where
three or more layers are analysed, the number of possible
relationships is subject to a combinatorial explosion and
rises into several hundreds and thousands. Only in the case
of identity(A,B), additional 13 cases of three-way relationships
can be distinguished; for all remaining cases of two-way
relationships, considerably more three-way cases need to be
distinguished. It seems impossible to invent names, let alone to
formulate and implement queries for each one of them. Still, for
a user it would be desirable to have a set of query predicates
for n-way relations available, lest (s)he needs to repeatedly
combine queries for two-way relationships, which often can
be done only with the help of a fully-fl edged programming
language.
Application: Analysing n-way relations
in text parsing
One text-technological application where relations between
elements on more than two elements need to be analysed,
is discourse parsing of argumentative texts. In a bottom-up
operating discourse parser such as the one developed for
German research articles in the SemDok project (Lüngen et al.
2006), it is checked successively whether a discourse relation
holds between two known adjacent discourse segments such
that they can be combined to form a larger segment. Often
this depends on the presence of a lexical discourse marker,
such as the adverb ‘lediglich’ (‘only’), in the second segment.
But with ‘lediglich’ as with numerous other markers, there is
the additional condition that it has to occur in the so-called
vorfeld (fi rst topological fi eld of a German sentence according
to the syntax of German, cf. Hinrichs & Kübler 2006), of the fi rst
sentence of the second discourse segment. Thus, a combination
of information from at least three different information levels
(discourse segments, syntax, and discourse markers) needs to
be checked, i.e. whether the following situation holds:
L1: <ds>..........................
...........................</ds>
L2: <s><vorfeld>.................
....</vorfeld>.............</s>
L3: <dm>lediglich</dm>
This situation corresponds to a meta-relation of three-way
inclusion: <ds> from Layer 1 must include a <vorfeld> from
Layer 2, which in turn must include a <dm> from Layer 3.
Querying n-way relations between
elements of multiple annotations
We have identifi ed a set of n-way meta-relations that are typically
needed in text-technological applications for multiply annotated
documents, namely N-way independence, N-way identity,
N-way inclusion, and N-way overlap, (where independence,
identity, and inclusion hold between the elements from all
n layers, and overlap holds between at least one pair among
the n elements). The proposed poster presentation illustrates
further examples from text-technological applications such as
discourse analysis and corpus linguistic studies, where querying
n-way relations between elements is required. It explains
our set of query predicates that have been implemented in
Prolog for n-way meta-relations, and how they are applied to
the examples. Furthermore it presents an evaluation of their
usability and computational performance.
References
Durusau, Patrick and Matthew Brook O’Donnell (2002).
Concurrent Markup for XML Documents. XML Europe 2002.
Hinrichs, Erhard and Sandra Kübler (2006). What Linguists
Always Wanted to Know About German and Did not Know
How to Estimate. In Mickael Suominen, Antti Arppe, Anu
Airola, Orvokki Heinämäki, Matti Miestamo, Urho Määttä,
Jussi Niemi, Kari K. Pitkänen and Kaius Sinnemäki (eds.): A
Man of Measure : Festschrift in Honour of Fred Karlsson on his
60th Birthday. The Linguistic Association of Finland, Special
Supplement to SKY Journal of Linguistics 19. Turku, Finland.
Lüngen, Harald, Henning Lobin, Maja Bärenfänger, Mirco
Hilbert and Csilla Puskas (2006). Text parsing of a complex
genre. In Bob Martens and Milena Dobreva (eds.): Proceedings
of the Conference on Electronic Publishing (ELPUB 2006). Bansko,
Bulgaria.
Witt, Andreas (2004). Multiple hierarchies: New aspects of
an old solution. In Proceedings of Extreme Markup Languages.
Montreal, Canada.
Witt, Andreas, Harald Lüngen, Daniela Goecke and Felix
Sasaki (2005). Unifi cation of XML Documents with
Concurrent Markup. Literary and Linguistic Computing 20(1), S.
103-116. Oxford, UK.