Extracting domain knowledge from tables of contents

Analysis of toc fragments

Our method of analysing toc fragments consists of the following steps:

[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 
		[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                        
                    

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
                       

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.

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 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

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> 
	...     
                    

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.

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 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.