An Algorithm for Merging and Aligning Ontologies: Automation and Tool Support
Natalya Fridman, Mark A. Musen · 1999
As researchers in the ontology-design field develop the content of a growing number of ontologies, the need for sharing and reusing this body of knowledge becomes increasingly critical. Aligning and merging existing ontologies, which is usually handled manually, often constitutes a large and tedious portion of the sharing process. We have developed SMART, an algorithm that provides a semi-automatic approach to ontology merging and alignment. SMART assists the ontology developer by performing certain tasks automatically and by guiding the developer to other tasks for which his intervention is required. SMART also determines possible inconsistencies in the state of the ontology that may result from the user’s actions, and suggests ways to remedy these inconsistencies. We define the set of basic operations that are performed during merging and alignment of ontologies, and determine the effects that invocation of each of these operations has on the process. SMART is based on an extremely general knowledge model and, therefore, can be applied across various platforms. 1 Merging Versus Alignment In recent years, researchers have developed many ontologies. These different groups of researchers are now beginning to work with one another, so they must bring together these disparate source ontologies. Two approaches are possible: (1) merging the ontologies to create a single coherent ontology, or (2) aligning the ontologies by establishing links between them and allowing them to reuse information from one another. As an illustration of the possible processes that establish correspondence between different ontologies, we consider the ontologies that natural languages embody. A researcher trying to find common ground between two such languages may perform one of several tasks. He may create a mapping between the two languages to be used in, say, a machine-translation system. Differences in the ontologies underlying the two languages often do not allow simple one-to-one correspondence, so a mapping must account for these differences. Alternatively, Esperanto language (an international language that was constructed from words in different European languages) was created through merging: All the languages and their underlying ontologies were combined to create a single language. Aligning languages (ontologies) is a third task. Consider how we learn a new domain language that has an extensive vocabulary, such as the language of medicine. The new ontology (the vocabulary of the medical domain) needs to be linked in our minds to the knowledge that we already have (our existing ontology of the world). The creation of these links is alignment. We consider merging and alignment in this paper. For simplicity, throughout the discussion, we assume that only two ontologies are being merged or aligned at any given time. Figure 1 illustrates the difference between ontology merging and alignment. In merging, a single ontology that is a merged version of the original ontologies is created. Often, the original ontologies cover similar or overlapping domains. For example, the Unified Medical Language System (Humphreys and Lindberg 1993; UMLS 1999) is a large merged ontology that reconciles differences in terminology from various machine-readable biomedical information sources. Another example is the project that was merging the top-most levels of two general commonsense-knowledge ontologies—SENSUS (Knight and Luk 1994) and Cyc (Lenat 1995)—to create a single top-level ontology of world knowledge (Hovy 1997). In alignment, the two original ontologies persist, with links established between them. Alignment usually is performed when the ontologies cover domains that are complementary to each other. For example, part of the High Performance Knowledge Base (HPKB) program sponsored by the Defense Advanced Research Projects Agency (DARPA) (Cohen et al. 1999) is structured around one central ontology, the Cyc knowledge base (Lenat 1995). Several teams of researchers develop ontologies in the domain of military tactics to cover the types of military units and weapons, tasks the units can perform, constraints on the units and tasks, and so on. These developers then align these more domain-specific ontologies to Cyc by establishing links into Cyc’c upperand middle-level ontologies. The domain-specific ontologies do not become part of the Cyc knowledge base; rather, they are separate ontologies that include Cyc and use its top-level distinctions. 1 Most knowledge representation systems would require one ontology to be included in the other for the links to be established.