Apart from solving the multi-label problem, the additional incorporation of background knowledge as provided by domain specific ontologies is the second focus of this work. Since ontologies have been defined many times, we will abstain from giving a formal definition of domain ontologies in favour of introducing the main aspects in a short example. The underlining formal definition, used representation and notions in this work are in accordance with [3]. This is also the basis of our implementation in the KAON Framework[6]. Figure 1 shows a very small extract of the AGROVOC thesaurus, represented as an ontology. Refer to [9] for a detailed discussion of converting the AGROVOC thesaurus into an ontology. An ontology is basically a tree-ordered hierarchy structure of concepts as shown in Figure 1.
Figure 1: Small ontology extract
Each concept in the picture (drawn as a rectangle) has lexical entries (labels, synonyms) attached to it. The picture only shows the English labels of the concepts. The important fact for our purposes - explained in more detail in section 4 - is that a concept itself is actually language independent and internally represented by a URI[7] (Uniform resource identifier). Every concept has a super concept, e.g. "supply balance" is the super concept of "stocks". The highest concept is "root". In addition to the tree structure, an ontology can have arbitrary lateral relationships, as shown here with a ‘related term’ relationship. As opposed to simple thesauri, ontologies allow for many types of relationships, making it a more expressive instrument of abstract domain modelling.
|
[6]
[http://kaon.semanticweb.org/]. [7] See also [http://www.w3.org/Addressing/]. |
