The differences and similarities in thesaurus and ontology structure: with a method for reengineering thesauri into qualitatively good ontologies

© 2013 Dr. Daniel KlessMOTIVATION: In recent years ontologies have become a strongly investigated instrument to create models that describe the meaning of terms and can be processed by machines. Further, because ontologies are developed and described in logic-based languages like OWL, they are also expected to allow the inference of knowledge using reasoning algorithms. Finally, the design of contemporary languages for the description of ontologies allows the modular development of ontologies and the exchange and integration of such independently developed ontology modules using the Internet. Because modeling ontologies is highly resource-consuming, there is a preference to reuse existing models as ontologies instead of developing ontologies from scratch. Complex thesauri or other types of vocabularies are often considered for reuse as ontologies. Such vocabularies have sometimes been developed and extended over decades and thus often contain a broad set of structured concepts used in a given domain making them useful for reuse as ontologies. The adoption of vocabularies as ontologies has turned out problematic in cases where they are used for logic-based reasoning or automatic processing. This raises doubts about whether vocabularies can become ontologies by describing them in logic-based languages like OWL alone as it is suggested in various methods for the ontological reengineering of thesauri. Overall, there appears to be little clarity of what the similarities and differences between vocabularies and ontologies are and what has to be done in order to reuse vocabularies as ontologies. GOAL: In our study we focus on thesauri as a particular type of structured vocabulary and ontologies described in the description logic OWL (OWL-DL). The goal of our study is to determine the similarities and differences between the structure of a qualitatively good thesaurus and the structure of a qualitatively good ontology. A qualitatively good thesaurus we define as a thesaurus that adheres to the international thesaurus standard ISO 25964-1:2011 or its predecessors and thus serves purposes of information retrieval such as indexing and searching. Modeling a domain correctly with respect to the reality it represents, containing a semantically rich description of the kinds of modeled entities, and the possibility to integrate the ontology with other ontologies are the criteria that we demand from a qualitatively good ontology. As a second goal we intend to develop a method that specifies how a qualitatively good ontology can be engineered based on a thesaurus. RESEARCH APPROACH: The study follows the basic structure of design science research and has two main parts. In the first part, the “suggestion” stage of design science, we compare thesauri and ontologies theoretically by mapping their respective relationships and relata. The reference to the international standard ISO 25964-1:2011 on the side of thesauri and the use of the philosophical stance of realism on the side of ontologies resembles a framework by which we can compare thesauri and ontologies in a generalized way. The differences from the theoretical comparison we translate into initial steps for the ontological reengineering of thesauri. The second part of the study, the “development” and “evaluation” of design science, is a case study where the initial steps for the ontological reengineering of thesauri are combined with general methodical steps for developing qualitatively good ontologies. The sum of these steps (a naïve reengineering method) is applied to reengineer a portion of the AGROVOC thesaurus into an ontology. The result of this reengineering is (a) a more mature method to reengineer thesauri into ontologies and (b) a qualitatively good ontology, the structure of which we can compare with the original thesaurus. From the comparison we can infer further differences and similarities between thesauri and ontologies. The reengineering method developed in that course acts as an embodiment of what is known about the differences and similarities between thesauri and ontologies. FINDINGS: Through the theoretical comparison in the first part of the study we found that thesauri have many similarities with ontologies in the way they define structural relationships, particularly with respect to the is-a relation in ontologies and the generic relationship in thesauri. However, thesauri have significant deficiencies from the perspective of ontological realism, most importantly an indifferent agglomeration of ontologically different entities and relationships. Firstly, thesaurus concepts can correspond to universals, but can also represent a single individual, pure collections or mixed collections in ontologies. Secondly, hierarchical relationships in thesauri may correspond to different kinds of part-of relationships or the is-a relationship or the instance of relationship in an ontology. While hierarchical thesaurus relationships may be transitive, the possible mixing of different kinds of relationships may cause “breaks” in the hierarchical transitivity. In the second part of the study, the reengineering case study, it became obvious that the way in which the intensions of thesaurus concepts can be specified (assigning natural language terms, hierarchical relations, associative relations, qualifiers, scope notes, and definitions) is insufficient for the unambiguous interpretation of the concepts. Ontologies can be more precise by specifying membership conditions of concepts. This has consequences with respect to how consistent the generic relationship in a thesaurus can be applied as opposed to its counterpart in ontologies, the is-a relationship. The is-a relationship can be checked for its consistency to the degree to which a) membership conditions are specified for the universals in an ontology, b) the ontology is aligned to a top-level ontology and a corresponding set of formal relations, and c) the universals referenced in the membership conditions are directly or indirectly aligned to the same top-level ontology. The ontological analysis of the generic relationship in a thesaurus can reveal mistakes that cannot be detected using the all-and-some test recommended in the thesaurus standard ISO 25964-1:2011. From the case study we also found that thesauri and ontologies differ in the way relationships are applied that are not the generic relationship/is-a relationship or the in-stance relationship when modeling specific entities in a domain. Specifically, ontological relationships express necessary membership conditions for their relata, whereas thesaurus relationships are established to guide indexers and searchers. As a result, there may be just minimal or even no overlap between admissible or desirable relationships in thesauri and relationships in ontologies. Further, ontological relationships do not imply the inverse relationship between two relata (their “reciprocity”) as it would generally be assumed for thesaurus relationships. Finally, the labeling of the relata in thesauri and ontologies through natural language terms may differ, depending on the conventions applied to using labels in ontologies. The method that evolved out of our case study suggests the following steps for the process of reengineering a thesaurus ontologically: 1) preparatory refinement and checking of the thesaurus, 2) syntactic conversion, 3) identification of membership conditions, 4) alignment of ontologies to a top-level ontology and formal relations, 5) formal specification of membership conditions, 6) adjustment of spelling, punctuation and other aspects of entity labels, 7) dissolving poly-hierarchies, and 8) de-coupling of independent entities. Only steps 1 and 2 are specific to thesauri. Steps 3 to 5 represent the core of general ontological modeling. Steps 6 to 8 may be considered optional. IMPLICATIONS: A consequence of our results is that thesauri cannot be considered a less expressive type of ontology. Instead, thesauri and ontologies need to be treated as two orthogonal kinds of models with superficially similar structures. Neither is a qualitatively good ontology a good thesaurus, nor is a qualitatively good thesaurus a good ontology. A thesaurus requires significant structural and other content changes in order to become a qualitatively good ontology and vice versa. Our results also show that ontologies specified with the description logic OWL should be seen as models of necessary and sufficient membership conditions of entities, which sets them apart from thesauri and other kinds of vocabularies