There is a recent trend of using controlled natural language (CNL) interfaces to provide
more intuitive ways for entering abstract knowledge constructs [
The Ordnance Survey is developing a modular topographic domain ontology to
facilitate the description and reuse of its topographic data by third parties [
Following this methodology, the domain expert is engaged in the construction of a conceptual ontology which involves the first four steps. The knowledge engineer is then performing the last two steps which focus on the logical level of the ontology.
A crucial component of the Ordnance Survey methodology is the use of a controlled
language for authoring the conceptual ontology - a CNL, called Rabbit, has been
developed for this purpose [
This paper focuses on the support for creating OWL ontologies in Rabbit. A tool called ROO (Rabbit to OWL Ontology authoring) is outlined, pointing at aspects of Rabbit parsing and the CNL-enabled interaction.
ROO is a Protégé plug-in that assists domain experts to build conceptual ontologies. ROO includes the following usability features to support ontology authoring: • provide easy to understand suggestions and task specific messages to help the user enter correct Rabbit constructs; • show feedback about the parsed structure to help the user recognize CNL patterns; • show warnings when the user sentences are parsed but there may be ambiguity when selecting the corresponding knowledge constructs in OWL; • facilitate the knowledge input process by providing a list of Rabbit templates;
ROO provides a customized interface to support the ontology authoring process by
entering Rabbit sentences. See [
JAPE4 transducers perform the parsing of Rabbit constructs based on the annotations gathered during the preprocessing phase. Most constructs contain key phrases which makes the constructs unambiguous. For example: <instance> and 2 http://gate.ac.uk 3 Information extraction library included in GATE http://gate.ac.uk/sale/tao/index.html#annie 4 Java Annotation Pattern Engine which provides a language for finding annotation patterns during the parsing process. It also provides hooks for invoking Java methods during the parsing of a text. See http://gate.ac.uk/sale/tao/index.html#jape <instance> are different, where we have underlined the key phrases. However, constructs like concepts, relationships and instances are not known a priori and can consist of multiple words e.g. Natural Body of Water. JAPE rules restrict the possible set of these constructs based on the initial POS and morphological annotations (e.g. a noun phrase is expected to be a concept and a relationship is expected to be a verb phrase). This allows for some ambiguity at the initial stage of the parsing, as some part of text may be linked to different Rabbit construct (e.g. compare Transport of Water and Body of Water. Both have the same linguistic structure, but on a conceptual level Body of Water is interpreted as a single concept, while Transport of Water is a composed concept describing a subclass of concept Transport restricted by its relationship to concept Water). The Rabbit parser uses the ontology being defined to decide which interpretation should be used.
The end result of the parsing process is a Java representation of Rabbit sentences. At this level, the parser removes any ambiguity by using the ontology being built, as well as heuristic rules. Consider the sentence Every Irrigation Canal contains Water for Irrigation. The parsed result of the sentence depends on whether the ontology defines relation contains water for or relation contains, and concepts Irrigation Canal, Water, Irrigation or Water for Irrigation. The parser will detect when any of these relations and concepts are missing and, if necessary, will prompt the user to define the missing part. The user will also be warned about ambiguity when several possibilities are already defined in the ontology since the heuristic used to disambiguate might not be correct and this can result in an unwanted OWL translation. The advantage of this approach is that the parser is able to correctly detect sentences even when there are missing parts in the ontology or when parts are linguistically ambiguous. This also enables the parser to provide better error messages and suggestions to the user. A possible drawback is that the user has to be careful not to introduce names which may bring undesired ambiguity, but the parser helps to avoid this by giving warning messages.
The conversion to OWL is done based on the Java representation produced by the parser. The Rabbit language specification defines the OWL equivalent of each Rabbit sentence. For instance, declaring concept Water adds the following axioms to the ontology: SubClassOf(Water owl:Thing) and EntityAnnotation( Class(Water) Label(“Water”)). The current implementation of the conversion to OWL is based on OWLAPI5, because that is the API used by Protégé. The architecture allows for an easy reimplementation where a different API could be used (such as Jena).
Continuous user studies are performed at Ordnance Survey to examine the usability
and the use of Rabbit to define sample domain ontologies [
The study showed that with even a minimal amount of training, in both ROO and ACE View, domain experts were able to perform ontology authoring tasks without the need to learn ontology languages, such as OWL. It is always likely that for complex ontologies knowledge engineering skills will be required. However, the study indicated that if methodical, intelligent support for ontology authoring is embedded in the authoring tool, as this is done in ROO, domain experts would be able to actively engage in the ontology authoring process. There was a strong evidence that the support and guidance offered in ROO led to better usability and had positive effect on the quality of the resultant ontologies. The study also identified requirements for further support that could be provided by improving the natural language analysis and by considering common error patterns when using a CNL to define ontological constructs.
The current version of ROO6 provides support for the full set of Rabbit constructs. Further studies with ROO are planned for early 2009.
During the 1 hour presentation at CNL09, we would like to present more details on the topics discussed in this abstract. We will include a demonstration of how the tool is typically used, highlighting the advantages of using a CNL-based tool for ontology construction. We would also like to discuss issues we have come across during the implementation of ROO such as expressing complex logical constructs using Rabbit without introducing ambiguity and evaluating the resultant ontology.