Semi-automatic ontology engineering and ontology supported document indexing in a multilingual environment

Conversion of a thesaurus or another already existing vocabulary into an ontology is still more of an art than a well-defined process and has to be carefully considered in each specific case. One approach mapping an Art and Architecture Thesaurus to an ontology using Protégé21, which is similar to the here chosen one is described in Wielinga et al. [WSWS01]. Another interesting methodology for integrating ontologies and thesauri to build RDFS schemas is given in [AF99]. The methodology discussed there is platform independent and describes a process of several steps on how to map thesaurus term to an ontology. In this approach, an ontology structure of super-concepts is first created, followed by the mapping of the thesaurus terms to fit into this high level structure of top concepts. Here a more straightforward approach has been chosen, mapping terms directly to concepts and leaving the thesaurus hierarchy structure as is in the first instance. The resolving of this and establishment and restructuring using super-concepts is left for further assessment as explained before in the engineering framework. The first problem in converting a thesaurus arises with the question of what is a concept. As described before in chapter 2, AGROVOC is a collection of terms, each term being either a descriptor or a non-descriptor. The reason for this distinction comes from the purpose of a thesaurus to be used solely for indexing purposes. An ontology, however, is somewhat more generic and can be used in different ways, therefore not reflecting this very specific differentiation. One possible solution would be to simply have a concept for each keyword and link the former non-descriptors with the same relations (use and used for relation) to the remainder of the mapped concepts. The disadvantage of this approach is, however, that in the core ontology these relationships are unlikely to exist and to be used, since the KAON ontology model provides different modelling constructs of reflecting such relationships. Basically, a descriptor and a non-descriptor together refer to the same concept, deriving from the indexing rule to use a descriptor instead of a non-descriptor for indexing. The synonym concept of the KAON Lexical OIModel reflects such a relationship. Experience has shown that a non-descriptor in most cases can be seen as a synonym of its descriptor. The option chosen here is therefore to map all descriptor terms to a concept with a label in each language provided in the AGROVOC. Each non-descriptor has been mapped to synonyms attached to its mapped descriptor concept. For each translation available in the AGROVOC, a lexical [...]

API either from a database or an RDFS file. For more details on the overall architecture refer to the developer guide20 on the KAON web site. When called with the information above, a new browser session is invoked. The user is then able to browse and search the whole ontology structure in all available languages. At the current stage of the project the parameter handling is not implemented, since it is subject to application specific interfacing and therefore not part of the framework. The KAON Portal has been extended, however, with the functionality to mark terms while browsing the ontology. The user can browse the ontology in all the five FAO languages and compile a query string consisting of an arbitrary number of marked entities of the ontology. The lexicalizations of the entities are made visible to the user. The browsing session stores the entity URIs together with the lexicalizations (labels in the currently active language respectively) to be used by and backed up into the calling application (which would be the CDS system in this case). Figure 15 shows a screenshot of the adjusted KAON Portal for the OFsAPH. [...]

format. In this project, the AGROVOC thesaurus has been chosen as input resource. Since the representation of the AGROVOC as a KAON ontology is not only used for this step but also serves as a resource for the automatic text classification algorithm discussed later in chapter 6, the conversion will be discussed in more detail in a separate section later in this chapter. The aim of this acquisition step is to automatically extract possibly the whole subset of this ontology structure, relevant for the target domain. An ontology pruner, developed in [Volz00] and formerly applied in [KVM00] has been extended and adapted to accomplish this task. The approach is a heuristic, for which reason a detailed and extensive evaluation of the algorithm has been carried out. The algorithm, its application and evaluation will be discussed in detail in chapter 5. The input to the pruner is the converted vocabulary to be reused together with two sets of documents, a domain specific one and a generic one, both explained in detail in chapter 5. The output of this step is a subset of the initial ontology, i.e. the input ontology pruned to contain only the concepts relevant for this domain. The result has to be assessed and validated by subject experts and combined with the core ontology. This is done in the next phase of the engineering cycle. In the prototype project the not yet adapted version of the pruner has been applied, again due to time restrictions and the need for some fast results. The set of domain specific documents used for Text-To-Onto in the previous step has been reused here. An ontology with 504 concepts has been extracted from the AGROVOC in this application. Since an extensive evaluation of the algorithm follows in Chapter 5, I will not go into more detail with this prototype result. The evaluation presented in chapter 5 is actually the main part of the second application of the engineering cycle. [...]