Semi-Automatic Domain Ontology Generation
by Reusing A Large-Scale Ontology

With the emerging of the semantic web, researchers become more interested in the method about ontology generation; especially for creating new domain ontologies through ontology reuse. As we know, it is usually very difficult for people to build ontologies from scratch. Although many researchers have tried to automate the process of ontology generation by several different methodolgoies, such as machine learning and natural language processing, until now there are still no practically satisfactory resolutions available. In the near future, we may expect that the majority work of ontology construction still needs to depend on humans to create manually. At the same time, however, when more and more ontologies are available, it is reasonable to ask whether we could try to reuse the ontology components that have been pre-created so that we can reduce the load for ontology engineers. This research is such a case study. In this work, we have tried to build several small domain ontologies for individual web pages by reusing a large-scale, generic ontology---the Mikrokosmos ontology. Our methodology shows that it is now possible to automate this ontology reuse process, and we may reduce the load of ontology generation from scratch by as much as 50%. This is a finished project. There are also some other research projects in which I have involved or am involving.

The primary knowledge source in the united knowledge base is the Mikrokosmos ontology---a well designed "broad-coverage" ontology [1]. It contains over 5000 concepts, and each concept connects to an average of 14 other concepts through relationships represented in its slots. The ontology is represented in XML syntax. In addition to the primary knowledge source, we have three additionl knowledge sources to compose the knowledge base. The data-frame library provides a collection of data frames, which describes the extensions of many individual concepts. The WordNet [2] lexicons are adopted to be a synonym dictionary that links multiple knowledge sources together. Also, there are some additional lexicons that complement the Mikrokosmos ontology. For examples, lists of abbreviations and acronyms. The integration of knowledge sources are processed through a human-guided semi-automatic way. We only need to process this integration once, and it is separate to the run-time ontology-generation process. The purpose of this integration is to align the concepts defined in the Mikrokosmos ontology to their extensions, if any, and therefore a run-time data recognition process can directly process them to identify the potential instaniations of each concept within training documents.

Our basic assumption on training documents are that they contain rich information and they are narrow on the expected domain. Otherwise, if training documents are not data-rich, we may face the difficulty of not being to find data and thus not being able to generate an ontology that covers the data of interest. If they are not narrow in topic breadth, the generated ontology may face the problem of trying to link highly unrelated concepts. After we have a set of training documents, we need first pre-process these documents by (1) removing unrelated protion of document content, and (2) decomposing a document to multiple records, if applicable. Again, the purpose of these operations is to emphasize the scope of the expected domain and clear as much as possible on any other information that may lead mistakes. Each individual record thus represents a subpart of the domain of interest.

An interesting application of using this tool is that we add a new knowledge source, SALT, that contains both new concepts and new relationship sets into the knowledge base. In addition, we add the Eurodicautom terminology bank as our termbase. This terminology bank consists of over a million concept entries covering a wide range of topic. Each entry is multilingual in character, containing equivalents in any of several languages. With the new augmented knowledge base, we ran the system on various U.S. Department of Energy (DOE) abstracts to test the performance of our system. Although the final results contains some spurious relationship sets because DOE abstracts cover a wide range of topics, there is an encouraging observation. Because of the use of multiple ontologies, the system is able to suggest some brand new relationship sets, which are also reasonably correct, that are not contained in any of the original knowledge sources.

• Knowledge sources with deep conceptual hierachical structure, like the Mikrokosmos ontology, are not good to be the knowledge base for automatic ontology generation. Many times, we find that the long chain of conceptual hierachies increases the difficulty of concept disambiguation and correct detection of relationships between two concepts.
• A solid integrated knowledge base is not good to generate dynamic subdomain out of its structure. A certain level of flexibility must be added to the knowledge base so that it can fit for the variant reality of real world domains.
• Single knowledge source limits the generation of practical relationship sets. Increasing the diversity of knowledge sources can help improve this performance and even suggest brand new relationship sets that are fitful to the domain of interest.
• The level of generality of leaf concepts decides the level of grain (fine grain vs. coarse grain) that a generated ontology may achieve. Both generalization and specialization are not easy to process by machines. Does this suggest that we may want to declare flexible level grains of conceptual specifications in a knowledge base? This must be a debatable questions.