Extracting domain knowledge from tables of contents

1
Extracting domain
knowledge from tables of
contents
Harald Lüngen
luengen@uni-giessen.de
Justus-Liebig-Universität Gießen
Henning Lobin
henning.lobin@germanistik.uni-giessen.de
Justus-Liebig-Universität Gießen
1. Introduction
Knowledge in textual form is always presented
as visually and hierarchically structured units
of text, which is particularly true in the case of
academic texts. One research hypothesis of the
ongoing project
Knowledge ordering in texts—
text structure and structure visualisations as
sources of natural ontologies
1
is that the textual
structure of academic texts effectively mirrors
essential parts of the knowledge structure
that is built up in the text. The structuring
of a modern dissertation thesis (e.g. in the
form of an automatically generated table of
contents -
tocs
), for example, represents a
compromise between requirements of the text
type and the methodological and conceptual
structure of its subject-matter. The aim of the
project is to examine how visual-hierarchical
structuring systems are constructed, how
knowledge structures are encoded in them, and
how they can be exploited to automatically
derive ontological knowledge for navigation,
archiving, or search tasks. The idea to extract
domain concepts and semantic relations mainly
from the structural and linguistic information
gathered from tables of contents represents a
novel approach to ontology learning.
2. Data and annotations
In the present phase, we examine German
academic text books, in later phases,
dissertations, research articles and historical
scientific texts will also be taken into account.
A corpus of digital versions of 32 text books
from 12 different academic disciplines has been
compiled,
2
the textual content and an XML
document structure markup was subsequently
extracted (e.g. using the Adobe Pro software).
Using a series of XSLT style sheets, the initial
XML was converted to XML encoding of
the document structure according to the TEI
P5 guidelines. At the same time, the texts
were annotated with morphological analyses
and phrase chunking markup using the Tree
Tagger and Chunker software from Stuttgart
University
3
and converted to a suitable XML
representation, and also with dependency-
syntactic analysis using the Machinese Syntax
Parser by Connexor Oy,
4
resulting in XML
markup as well. Further linguistic annotation
levels (such as domain terms and lexical-
semantic relations) will be added and combined
in XStandoff documents representing multi-
layer annotations that can be queried using XML
standards and tools as described in Stührenberg
& Jettka (2009).
Presently, all available annotation layers are
stored in an eXist native XML database
5
and
are queried using the Oxygen XML editor
6
as a
database client.
The corpus infrastructure is used to explore
the document applying the method of toc
fragment analysis as described in the following
section, and to implement functions for concept
extraction and semantic relation analysis.
2.1. Analysis of toc fragments
Our method of analysing toc fragments consists
of the following steps:
1.
Identification of a toc fragment
2.
Representation of the fragment meaning as a
MultiNet
3.
Identification of the configuration of elements
on different structural levels that induce the
fragment meaning
4.
Hypothesis about the generalisation of a toc
fragment, a structuring schema
5.
Corpus research to verify the generalisation
hypothesis
[booktitle] Einführung Pädagogik
[5.] Ausgewählte Subdisziplinen und
Fachrichtungen
[5.1.] Literatur
[5.2.] Erlebnispädagogik
[5.2.1.] Begrifflichkeit
[5.2.1.1.] Erlebnis als prioritäre Kategorie

2
[5.2.2.] Historie
[5.2.3.] Theoretische Fundierungen und
Menschenbilder
[5.2.4.] Ziele und Funktionen der
Erlebnispädagogik
[5.2.4.1.] Ziele
[5.2.4.2.] Subjektbezogene Funktionen und
mögliche Wirkungsweisen
[5.2.4.3.] Gesellschaftliche Funktionen
[5.2.5.] Merkmale und Modelle der
Erlebnispädagogik
[5.2.6.] Beispiele erlebnispädagogischer
Angebote
[5.2.6.1.] Outward Bound-Konzeption
[5.2.6.2.] Outdoor Management Development
[5.2.7.] Kritikpunkte
[5.2.8.] Einführungsliteratur (zum
Weiterlesen)
[5.3.] Erwachsenenbildung
[5.3.1.] Begriffsklärung
[5.3.2.] Geschichtliche Entwicklung
[5.3.3.] Struktur und Funktionsperspektiven in
der Erwachsenenbildung
[5.3.4.] Theoretische Orientierungen der
Erwachsenenbildung
[5.3.5.] Forschungsfelder
[5.3.6.] Einführungsliteratur (zum
Weiterlesen)
[5.4.] Gesundheitspädagogik
Figure 1: Section from the table of
contents of Raithel et al. (2007)
Consider the section of the generated table
of contents of the text book
Einführung
Pädagogik
by Raithel et al. (2007) shown
in Figure 1. By choosing the heading 5.
Ausgewählte Subdisziplinen und Fachbereiche
and its immediately superordinated heading
(in this case the title of the book) as well
as its immediately subordinated headings, we
arrive at the toc fragment (or “window”) shown
in Figure 2. In the toc fragment, four terms
from the domain are contained,
Pädagogik
,
Erlebnispädagogik
,
Erwachsenenbildung
,
and
Gesundheitspädagogik
.
7
The terms
identification component must distinguish such
expressions denoting domain-specific concepts
from relational nouns commonly found in
academic and scientific discourse (such as
Einführung
,
Subdisziplin
and
Fachrichtung
)
and from terms denoting text-type structural
categories of academic texts such as
Literatur
).
We employ the semantic network approach
Multilayered Extended Semantic Networks
(acronym: MultiNets) by Helbig (2006) to
represent the domain concepts and semantic
relations between them
[booktitle] Einführung in die Pädagogik
[5.] Ausgewählte Subdisziplinen und
Fachrichtungen
[5.1.] Literatur
[5.2.] Erlebnispädagogik
[5.3.] Erwachsenenbildung
[5.4.] Gesundheitspädagogik
Figure 2: Toc fragment
expressed in a toc fragment. The MultiNet
approach is a fully-fledged semantic theory and
provides a rich and consistent inventory of
semantic entity types, features, relations and
functions, and has been previously employed
in the syntactic-semantic analysis components
of QA systems (Hartrumpf 2005). Using the
graphical MWR editor for designing MultiNets,
8
we represent the semantics of the above toc
fragment as shown in Figure 3.
Figure 3: MultiNet analysis of toc fragment
In the semantic network in Figure
3, the concepts
a3_erlebnispädagogik.1.1
,
a3_erwachsenenbildung.1.1
, and
a3_gesundheitspädagogik.1.1
are related to
a3_pädagogik.1.1
by the SUBS relation
denoting subordination of (abstract) situations;
a3_pädagogik.1.1
is in turn related to
fach.1.2
and
a3_disziplin.1.1
by the SUB
relation denoting the subordination of
concepts representing objects.
9
Furthermore,
the semantic decompositions of the three
compounds are analysed using the relations

3
MCONT (mental or informational content),
BENF (beneficiary) and METH (method), and
the relation between the concept
c2
representing
the textbook as such and
a3_pädagogik.1.1
is
specified as MCONT, the relation between
c2
and its authors as ORIG (mental of
informational origin, cf. Helbig 2006).
On account of this analysis, the following
hypothesis is formed:
Given a potential structuring schema, consisting
of an initial expression N, and an expression
N-1 related to N by a
heading_of
relation on
the document structure level, and an expression
N+1 to which N is related by the
heading_of
relation on the document structure level (cf.
Figure 4), if
N-1
N
N+1
Figure 4: Toc Schema
Figure 5: MultiNet Schema
-
N contains the lexeme Subdisziplin or a
synonym on the lexical level
-
and N-1 contains the domain concept A
-
and N+1 contains the domain concept B,
then, by multiple application, construct a
MultiNet-Schema as represented by the graph in
Figure 5.
<result doc="schruender-
lenzen_schriftspracherwerb_2007" docID="i72">
<head level="n-1">9. Schwierigkeiten des
Schriftspracherwerbs rechtzeitig erkennen und
gezielt helfen</head>
<head level="n">9.2 Zentrale
Wahrnehmungsbereiche und ihr Risikopotential</
head>
<head level="n+1">9.2.1 Visuelle
Wahrnehmung</head>
<head level="n+1">9.2.2 Auditive
Wahrnehmung</head>
</result>
<result
doc="brosius_kommunikationsforschung_2008"
docID="i137">
<head level="n-1">8. Kapitel:
Inhaltsanalyse I: Grundlagen</head>
<head level="n">8.4 Anwendungsgebiete und
typische Fragestellungen</head>
<head level="n+1">8.4.1 Inhaltsanalysen auf
dem Feld der politischen Kommunikation</head>
<head level="n+1">8.4.2 Inhaltsanalysen in
der Gewaltforschung</head>
<head level="n+1">8.4.3 Inhaltsanalysen in
der Minderheitenforschung</head>
</result>
<result doc="raithel_paedagogik_2007"
docID="i180">
<head level="n-1">BOOKTITLE: Einführung
Pädagogik</head>
<head level="n">D Ausgewählte
Subdisziplinen und Fachrichtungen</head>
<head level="n+1"> Literatur</head>
<head level="n+1">Erlebnispädagogik</head>
<head level="n+1">Erwachsenenbildung</
head>
...
Figure 6: Query Result Document
The Hypothesis is verified by formulating the
potential structuring schema as a query to the
corpus using the XQuery query language. The
query result document then contains a set of
toc fragments that can now be inspected as
to whether their semantics conform to the
hypothesis or not, leading to a small statistic
about the validity of the hypothesis. Sometimes
the inspection may also lead to a modification of
the original query. In the first result fragment in
Figure 6, for instance, the superordinate concept
Wahrnehmung
is not contained in N-1, but as
the compound modifier of
Bereich
(a synonym
of
Subdisziplin
).
10
In this example it becomes clear that analyses
on the morphological and lexical-semantic level
interact with the analyses of the structuring
information in that both levels provide
conditions or constraints when building the
semantic analysis of a toc fragment. Our corpus
infrastructure is designed such that information
from multiple linguistic and structural levels can
be taken into account.
2.2. Conclusion
We presently inventorise sets of complex
conditions connecting a structuring schema
with a MultiNet Schema as
constructions
in
the sense of Construction Grammar (CxG).
Construction Grammar (Kay 1995, Östman &
Fried 2004) is a theory of grammar which is
not based on phrase structure rules operating
on lexical elements, but as combinations of
constructions in which form schemata are

4
associated with meaning schemata and is
therefore appropriate for the description task at
hand. The inventory of constructions will then
be employed in ontology learning, particularly
for the task of automatically extracting domain
concepts and semantic relations between them.
Constructions describing document structuring
schemata as described above play a role
similar to the lexico-syntactic “Hearst Patterns”
described in Hearst (1992), which have been
employed for extracting semantic relations from
running text.
References
Brosius, Hans-Bernd, Koschel,
Friederike, Haas, Alexander
(2008).
Soziologie. Methoden der empirischen
Kommunikationsforschung. 4. Aufl.
Wiesbaden.
Glöckner, Ingo, Hartrumpf, Sven,
Helbig, Hermann, Leveling, Johannes,
Osswald, Rainer
(2007). 'Automatic semantic
analysis for NLP applications'.
Zeitschrift für
Sprachwissenschaft.
Jg. 26, H. 2
: 241–266.
Hartrumpf, Sven
(2005). 'University of
Hagen at QA@CLEF 2005: Extending
knowledge and deepening linguistic processing
for question answering.'.
Results of the CLEF
2005 Cross-Language System Evaluation
Campaign, Working Notes for the CLEF
2005 Workshop.
Peters, Carol (ed.). Vienna:
Centromedia.
Hearst, Matti A.
(1992). 'Automatic
acquisition of hyponyms from large text
corpora'.
Proceedings of the 14th International
Conference on Computational Linguistics
.
Helbig, Hermann
(2006).
Knowledge
Representation and the Semantics of Natural
Language.
Cognitive Technologies. Heidelberg:
Springer
Kay, Paul
(1995). 'Construction Grammar'.
Handbook of Pragmatics Manual.
Verschueren, Jef, Östman, Jan-Ola, Blommaert,
Jan (eds.). Amsterdam: John Benjamins, pp.
171–177.
Östman, Jan-Ola, Fried, Mirjam (eds.)
(2004).
Construction Grammars: Cognitive
grounding and theoretical extensions.
Amsterdam: John Benjamins.
Raithel, Jürgen, Dollinger, Bernd,
Hörmann, Georg
(2007).
Einführung in die
Pädagogik. Begriff - Strömungen - Klassiker
- Fachrichtungen. 2. Aufl..
VS Verlag für
Sozialwissenschaften. Wiesbaden: Springer
Schmid, Helmut
(1994). 'Probabilistic Part-
of-Speech Tagging using Decision Trees'.
Proceedings of the International Conference
on New Methods in Language Processing.
Manchester, UK.
Schründer-Lenzen, Agi
(2007).
Schriftspracherwerb. Bausteine professio-
nellen Handlungswissens. 2. Aufl.
VS Verlag für
Sozialwissenschaften. Wiesbaden: Springer
Stührenberg, Maik, Jettka, Daniel
(2009).
'A tookit for multi-dimensional markup: the
development of SGF to XStandoff'.
Proceedings
of Balisage: The Markup Conference 2009.
Balisage Series on Markup Technologies. 3 vols.
Tapanainen, Pasi, Järvinen, Timo
(1997). 'A non-projective dependency parser'.
Proceedings of the fifth conference on Applied
natural language processing.
Washington D.C.
Notes
1.
funded within the framework of LOEWE, the excellence
initiative of the state of Hesse, as part of the
LOEWE-
Schwerpunkt Kulturtechniken und ihre Medialisierung
,
cf.
http://www.zmi.uni-giessen.de/projekte/proj
ekt-36.html
.
2.
We would like to thank the publishers
Facultas
,
Haupt
,
Narr/
Francke/Attempto
,
Springer
,
UTB
,
Vandenhoek & Ruprecht
,
and
Wissenschaftliche Buchgesellschaft
for kindly making
available digital versions of textbooks for us.
3.
http://www.ims.uni-stuttgart.de/projekte/corp
lex/TreeTagger/
4.
http://www.connexor.eu/
5.
http://exist-db.org/
6.
http://www.oxygenxml.com/
7.
We consider terms to be linguistic expressions that refer to
domain concepts.
8.
which was kindly made available for us by Professor Helbig’s
group in Hagen.
9.
An
a3_
prefix in a concept name indicates that the concept
was not found in the required reading in the semantic lexicon
HagenLex (Glöckner et al. 2007) which is consulted by the
MWR tool.
10.
Other titles from our corpus cited in Figure 6 are Brosius
(2008), and Schründer-Lenzen (2008).