1 Introduction

One of the problems with such approaches is that background knowledge is not usually explicitly expressed in the knowledge sources [5]. This could seriously limit the ontologies produced with such approaches and may enforce searching for external knowledge sources to fill such knowledge gaps.

An alternative approach, or perhaps a complimentary one to the approaches above, is to search and reuse ontologies that already exist on the web when constructing new ones. It is often believed that one of the major factors for the initial rapid growth of the web was the ability and ease for users to copy HTML code from existing web pages and reuse it to create their own pages. Users were able to create web pages with little effort, and without the need for deep understanding of the underlying languages. Such ease of reuse needs to be replicated with ontologies to speed up their spread and use [21].

Facilitating reuse of other people's ontologies should encourage more individuals and organisations to participate in the semantic web. After all, ontologies are meant to provide an ``easy to reuse library of class objects for modelling problems and domains'' [27]. Wide dissemination of ontology technology would require methods and tools to facilitate ontology reuse [7]. However, no such tools are currently available, which renders ontology reuse a very hardship task [4],[12].

Harnessing online ontologies might be the first step towards achieving true reuse. There is an increasing number of online libraries for searching and downloading ontologies. Examples of such libraries include Ontolingua¹, Protégé², and DAML³. Few search engines have recently appeared that allow keyword-based search for online ontologies, such as Swoogle [6] and OntoSearch [31]. Swoogle⁴ currently contains over 10000 ontologies covering a wide range of domains. However, the use of such tools is limited to giving back a list of potentially relevant ontologies, with no further support for reuse.

Providing support for reuse during ontology development from specific ontology libraries has been studied before (e.g. [9],[12]. However, the focus was mainly to enable users to reuse or import whole ontologies or ontology modules. They provided no support for ranking available ontologies, or for extracting and merging the ontology parts of interest, or for evaluating the resulting ontology.

This paper outlines a proposal for a new approach for automatic ontology construction. The approach is meant to encourage and support reuse of existing ontologies, using latest semantic web research technology. We propose to build a system that searches online ontologies for representations of certain concepts, ranks the retrieved ontologies according to some criteria, then extract the relevant parts of the top ranked ontologies, and merge those parts to acquire the richest domain representation as possible.

We believe that by relying on existing ontologies we can avoid the problems stated in [5], and bootstrap the process of ontology building by reusing existing ontology fragments.

2 Scenario

Imagine there is a knowledge engineer who is in need of an ontology representing the academic domain. The ontology is to be used for creating a knowledge-base to hold information on staff, projects, conferences, publications, etc. There are many ontologies online that covers various portions of this domain. It would certainly be beneficial if the engineer can quickly and efficiently reuse some of these third-party ontologies, to at least bootstrap the ontology construction task.

One of the recommended first steps towards building an ontology is to write down a list of terms to represent in the ontology [23],[27],[17]. This helps scoping the domain, reaching agreement, and building the class hierarchy. Lets assume that one of the terms that our knowledge engineer wrote was ``Conference''. There could be many ontologies out there that covers this concept to some extent, that our engineer is not aware of. It might speed up his task if some existing representations can be easily gathered and represented to him to accept, modify, or at least learn from.

At the time of writing, when searching for owl ontologies on the term ``Conference'' in Swoogle [6], a list of 34 ontologies was returned. Our system could analyse these ontologies to acquire as much representational knowledge as possible about the given term. In some cases, if there are too many ontologies to analyse (such as in our example) then the system could greatly benefit from an ontology ranking service that can order the ontologies according to some criteria. The system could be set to only analyse and reuse, say, the top five ontologies from the ranked list, but of course such threshold can be modified based on what has been found.

Now the system could commence its analyses of the top ranked ontologies, starting with rank number 1. Based on the size and scope of that ontology, the system may decide to take the ontology as a whole, or only take the section that describes ``Conference''. To achieve the latter, the system will need to use an ontology segmentation service to extract the required section from the ontology. In our case, the first ontology on Swoogle's results list is conference.owl<sup>5</sup>. This ontology comprises of only one class ``Conference'' and a number of datatype properties, such as title, paperDueOn, startDate, etc. (figure 1).

Even though the above owl file does represent the concept ``Conference'', but it might not be sufficient, in terms of coverage, detail, etc. However, many other ontologies were found by the search engine when searching for this concept. Our system could start analysing some of those ontologies to enrich our first ontology with more knowledge. For example, the second ontology found by Swoogle was web04photo.owl<sup>6</sup>. A section of this ontology is shown in figure 2. The class ``Conference'' in this ontology has more detail than in conference.owl. For example, we now have a superclass (Event), some siblings (e.g. Workshop, Presentation, etc.), and new properties (e.g. subEventOf, hasSubEvent, hasOrganizer, etc.). However, note that this ontology does not have all the properties found in conference.owl.

Next, the system will need to compare the two ontologies (or segments of ontologies) to find any additional representations that can be merged into the first ontology. This may result in the creation of new classes and properties in the first ontology to cover these additional knowledge representations.

By iterating the above comparison and merging processes for some of the top ranked ontologies, an ontology might be produced that is perhaps richer and more detailed than any of the already existing ontologies. When the process ends, the resulting ontology can be presented to the knowledge engineer to study and modify or even correct as required.

3 Architecture

1 System Processes

1. Search for Ontologies: First step is to find some potentially relevant ontologies to reuse. This may be done by searching for specific keywords (as in Swoogle [6]) or perhaps using a more complex query mechanism (eg metadata search [12], structure-based queries, query expansion techniques, etc). This process should produce a list of ontology URIs that are potentially relevant to the user's needs.
2. Ontology Ranking: As shown in section 2, a search for ontologies may return too many hits. To enable the system to start with the best ontology, the returned list of ontologies must be ranked according to, for example, how well they represent the concepts in the search query [1], how the ontologies have been rated by users [26], or how well they meet the requirements of certain evaluation tests (e.g. OntoClean [10]). The outcome of this process will be a ranked list of ontologies for the system to analyse and reuse.
3. Ontology Segmentation: Depending on the size of the ontology that is being analysed for reuse, and on the desired scope for the final ontology, the system may deploy some tools to extract only certain parts of the ontology in hand. This could be necessary to limit the size and domain of the resulting ontology to only what the user is interested in. Users can be given some control on the level of generosity of the segmentation methods.

   Several approaches for segmenting ontologies have been investigated and covered in the literature. The type of segmentation required by our proposed system could be very simplistic, such as the spreading activation-based extraction discussed in [19], which cuts an ontology view to a fixed graph length around a selected concept. More complex approaches that take the ontology structure into account can also be deployed, such as those discussed in [3],[22]. Other approaches on ontology segmentation include the use of classical clustering algorithms to break an ontology into a fixed number of parts [25], the use of queries to create specific ontology views [29],[14], and the segmentation of an ontology based on analysing its applications' queries [2].

4. Ontology Mapping and Merging: Once ontology segments are obtained, the system will need to compare and merge these segments to gather more domain representation for the concepts in question. This needs to be handled carefully to avoid overwhelming the user with oversized representations.

   Ontology mapping and merging are hot topics in AI research, aiming to bridge the communication gap between applications that rely on different, but overlapping, knowledge representations. A number of tools have been developed to map and/or merge ontologies semi-automatically, where the system compares the given ontologies and reports back to the user some mapping suggestions. Examples of such systems are the PROMPT Suite [18] which is integrated into the Protégé ontology editor [16] and Chimeara [15], which was built on Ontolingua [9].

5. Ontology Evaluation: The system may need to evaluate the ontology that it constructed from fragments taken from other ontologies, to make sure that the resulting ontology meets some minimum quality checks. The evaluation process may help the system to, for example, resolve some inconsistencies, identify semantic gaps which needs to be filled, etc. OntoClean is a well known methodology for evaluating ontologies based on meta-property tags [10]. The tagging process can be too costly if performed manually, but some attempts to automate it are already under way [28].

The user may want to repeat the steps above to search for additional concepts, or to rerun the process using different thresholds (e.g. analyse more ontologies, use larger segments) then merge the resulting ontology with the one that the system produced so far.

4 Challenges

The system described in this paper will definitely be a challenge to build and run successfully. Most of the processes this system builds on (section 3.1) are dependent on semantic web technologies that are still rather immature and far from perfect. However, technologies can only mature once they are applied to practical and real world tests, which is what this system is hoping to provide. Bringing all these technologies together and integrating them to form a single production line, is one of the indirect aims of this proposed system.

Apart from the well known challenges of searching, ranking, segmentation, merging, and evaluation of ontologies, which are well covered in the literature, there are other fundamental issues that the proposed system may need to deal with.

One of the crucial factors when building new ontologies from existing ones is obviously the availability of ontologies to reuse, in terms of numbers and domain variety. Many of the ontologies constructed by semantic web researchers and developers are never put on the web. This will hopefully change once ontology search engines become more popular, and the benefits of making ontologies available for others become more apparent.

There is also the danger of ending up with a very large and messy ontology as a result of using automated systems, like the one we are suggesting, to construct the ontology. This could happen if many large ontologies are used as a resource. Users might find it hard to clean or modify the resulting ontology if its too large, rendering the whole process less useful. However, this problem can be avoided to some extent using system cut-off thresholds, which can be set based on the size of the ontology to be constructed, where the system stops augmenting the ontology once it reaches a certain size limit. Another possible approach is to enable users to interact with the system throughout the whole process. Users could monitor the construction and augmentation of the ontology, and either permanently stop the system if they are satisfied with the result, or pause the system while they clean the ontology before allowing the system to analyse and merge further ontologies.

Because the system is reusing existing ontologies, the quality of those ontologies will certainly affect the quality of the output ontology. Users might want to restrict the system to only those ontologies that pass certain quality tests, or are provided by specific organisations or authors. Applying proper ontology ranking and evaluation processes might help reducing this problem. However, it can not be guaranteed that the segments extracted from the original ontologies will retain the quality and consistency of their source.

5 Conclusions and Future Work

Users of the proposed system will be expected to modify, delete from, and add to the automatically built ontology as they see fit. The purpose here is to avoid reinventing the wheel by provide users with a tool to help them gather and learn from existing domain knowledge representations, thus bootstrapping their ontology construction task.

We aim to start building the system described in this paper soon and experiment with it to evaluate how well such an approach may work in real life scenarios.

6 Acknowledgments

Bibliography

Footnotes

... Ontolingua¹

http:/www-ksl-svc.stanford.edu:5915/

... Protégé²

http:/protege.stanford.edu/pugins/owl/owl-library/

... DAML³

http:/www.daml.org/ontologies

... Swoogle⁴

http:/swoogle.umbc.edu/

...conference.owl⁵

http:/ebiquity.umbc.edu/ontology/conference.owl

...web04photo.owl⁶

http:/www.mindswap.org/golbeck/web/www04photo.owl