Some Issues on Ontology Integration H. Sofia Pinto Instituto Superior Técnico Departamento de Eng. Informática Grupo de Inteligência Artificial Av. Rovisco Pais, 1049-001 Lisboa, Portugal Tel: (351-1) 8417641 Fax: (351-1) 8417472 sofia@gia.ist.utl.pt Asunción Gómez-Pérez João P. Martins Laboratorio de Inteligencia Artificial Instituto Superior Técnico Facultad de Informática Departamento de Eng. Informática Campus de Montegancedo sn. Grupo de Inteligência Artificial Boadilla del Monte, 28660, Madrid, Spain Av. Rovisco Pais, 1049-001 Lisboa, Portugal Tel (34-1) 3367439 Fax: (34-1) 3367412 Tel: (351-1) 8417472, Fax: (351-1) 8417472 asun@delicias.dia.fi.upm.es jpm@gia.ist.utl.pt three words to distinguish among those meanings: integration, merge and use. Abstract 1 Introduction The word integration has been used with differ- ent meanings in the ontology field. This article Within Knowledge Sharing and Reuse, the field of Onto- aims at clarifying the meaning of the word “inte- logical Engineering (OE) is an active area of research. One gration” and presenting some of the relevant work of its open research topics is ontology integration. Unfor- done in integration. We identify three meanings of tunately, there has been an abusive use of the word in- ontology “integration”: when building a new on- tegration within the community. Integration designates, tology reusing (by assembling, extending, special- not only, the special operations to build ontologies from izing or adapting) other ontologies already avail- other ontologies available in some ontology development able; when building an ontology by merging sev- environments (Farquhar, Fikes & Rice 1997), but also eral ontologies into a single one that unifies all of the process of building ontologies from other preexis- them; when building an application using one or tent ontologies (Borst, Akkermans & Top 1997, Dalia- more ontologies. We discuss the different mean- nis & Persson 1997, Gangemi, Pisanelli & Steve 1998, ings of “integration”, identify the main character- Skuce 1997, Swartout, Patil, Knight & Russ 1997), the set istics of the three different processes and propose of activities within some methodologies that specify how to build ontologies using other publicly available ontolo- This work was partially supported by JNICT grant No. PRAXIS gies (Uschold & King 1995, Gruninger 1996, Fernández, XXI/BD/11202/97 (Sub-Programa Ciência e Tecnologia do Segundo Gómez-Pérez & Juristo 1997), the use of ontologies in ap- Quadro Comunitário de Apoio) and project PRAXIS XXI/1568/95. plications (Bernaras, Laresgoiti & Corera 1996, Uschold, The copyright of this paper belongs to the papers authors. Permission to copy without fee all or part of this material is granted provided that the Healy, Williamson, Clark & Woods 1998), just to name a copies are not made or distributed for direct commercial advantage. few. Integration in ONIONS (Gangemi et al. 1998), doesn’t Proceedings of the IJCAI-99 workshop on mean the same as in the Ontolingua Server (Farquhar, Ontologies and Problem-Solving Methods (KRR5) Fikes, Pratt & Rice 1995) or in PhysSys (Borst 1997). Stockholm, Sweden, August 2, 1999 This article aims at clarifying and characterizing the sev- (V.R. Benjamins, B. Chandrasekaran, A. Gomez-Perez, N. Guarino, M. eral meanings of the word integration. This article is orga- Uschold, eds.) nized as follows. Section 2 identifies the three meanings http://sunsite.informatik.rwth-aachen.de/Publications/CEUR-WS/Vol-18/ usually associated to the word and proposes three words to H.S.Pinto, A.Gomez-Perez, J.P.Martins 7-1 refer to those meanings in the OE field. Sections 3, 4 and built. This ontology tries to unify concepts, terminol- 5 discuss the differences between those meanings. In each ogy, definitions, constraints, etc., from all of them and, one of those sections we review some of the most relevant if implemented, using a particular representation on- work, we characterize each one of the different processes tology. So, ontologies are merged, unified into a single and present our views and conclusions. Section 6 empha- one. sizes our main conclusions. 3. Integration of ontologies into applications. In this case, one wants to introduce into an application one 2 What Have We Been Calling Integration? or more ontologies that underly and are shared among We identified three different situations in which the word several software applications or one uses one or more integration has been used: ontologies to specify or implement a knowledge based system (KBS). For instance, if we want to build an 1. Integration of ontologies when building a new ontol- application on airplanes there is a lot of knowledge ogy reusing other available ontologies. In this case about airplanes in general (what parts are they built one wants to build a new ontology, there are some from, how are they designated, how do they interact, ontologies built and made available which are parts what laws determine the way those parts work, etc.) of the new ontology and such ontologies match the that needs to be formalized and implemented. The appropriate requirements. For instance, we want to knowledge that is not specific to any particular air- build an ontology on Control Systems and EngMath plane should be represented in an ontology. When we (Gruber & Olsen 1994) satisfies our requirements for build another application that also needs knowledge it, as for instance, adequate levels of detail and granu- about airplanes we should use the ontology already larity, it is implemented in an adequate language , etc. built (or adapt it if necessary). This airplane ontology Since this ontology is publicly available it should be will, once again, underlie an application. So the ontol- reused. In some cases a whole ontology can be built ogy(ies) is (are) used or reused to build an application. just from assembling other ontologies. Some other times the reused ontologies must be extended, special- From now on, we will refer to each kind of integration ized or adapted. So, ontologies are reused to build new as integration, merge and use, respectively. ones. 2. Integration of ontologies by merging different ontolo- 3 Integration gies about the same subject into a single one that Although some methodologies to build ontologies ac- “unifies” all of them. In this case, one wants to knowledge the need for an integration step or the impor- build an ontology merging ideas, concepts, distinc- tance of integration activities in the process of building tions, axioms, etc., that is knowledge, from other ex- an ontology, the important problems of integration (how isting ontologies on exactly the same subject. For should integration be performed?, etc.) remain more or less instance, among the considerable number of medi- unsolved. cal ontologies there is the Unified Medical Language System (UMLS) (Humphreys & Lindberg 1992) and 3.1 Tools That Allow Integration the Galen COding REference (CORE) model (Rector, Gangemi, Galeazzi, Glowinski & Rossi-Mori 1994). The work on the Ontolingua Server (Farquhar et al. 1995, There are differences between them, not only in the Farquhar, Fikes & Rice 1996, Farquhar et al. 1997), an basic distinctions but also in the way those terms ontology development environment for collaborative ontol- are defined (in the meaning behind those terms), in ogy construction, addressed the problem of ontology inte- the representation ontologies (van Heist et al. 1997) gration. This tool allows collaborative ontology building used to implement them, etc. When all these different and also provides an ontology library, where tested ontolo- ontologies are “integrated”, in the sense of merged, gies are gathered and made publicly available. To allow unified, a new ontology about the medical domain is reuse of the ontologies made available at the Ontolingua Server library, a set of integration operations was identified, Either the required language, or there are good translators avail- specified, defined and made available to ontology builders. able between the language in which it is implemented and the required Users are allowed three operations (Farquhar et al. 1997): language. By subject we mean what the ontology deals about. We avoid the inclusion, polymorphic refinement and restriction (special- term domain since it is only used to describe domain specific ontologies ization). Inclusion is used when the ontology is included and this kind of integration can be used to build general (generic (van (from the library of ontologies kept by the tool) and used Heist, Schreiber & Wielinga 1997), top-level (Guarino 1998) or upper- level) ontologies. as it is. The inclusion relations between ontologies may Classification criteria imposed on the terms gathered and de- be circular, so one concept in one ontology can point to a fined/described in those ontologies, that is concept classifications. concept in another ontology that, again, points to another H.S.Pinto, A.Gomez-Perez, J.P.Martins 7-2 concept in the first ontology. Polymorphic refinement ex- the KACTUS ontology library) and create the electrical dis- tends one operation so that it can be used with several kinds tribution ontology. of arguments. Restriction makes simplifying assumptions Sometimes it may be possible to mistakenly take inte- that restrict the included axioms. The Ontolingua Server gration for maintenance activities if we just want to im- also provides facilities for local symbol renaming. This fa- prove or slightly modify the integrated ontology. For in- cility enables ontology developers (1) to refer to symbols stance, in the case of CHEMICALS (Fernández 1996, from other ontologies using names that are more appro- Fernández et al. 1997) the length centimeter (“cm”) unit, priate to a given ontology and (2) to specify how naming a length unit commonly used in Europe (but not in the conflicts among symbols from multiple ontologies are to USA), was needed. The Standard-Units ontology (Gruber be resolved. All these simple integration operations have & Olsen 1994) available in the Ontolingua Server library been proposed to allow some sort of ontology integration did not include such unit when CHEMICALS was imple- in the Ontolingua Server. mented in the Ontolingua Server. The solution found, with the operations available, was: to develop a new ontology 3.2 Ontologies Built Through Integration which included Standard-Units and add to it the needed unit. However the right solution, adopted latter, was the The Physical Systems ontology (PhySys) (Borst, Ben- inclusion of this unit in the Standard-Units ontology kept jamin, Wielinga & Akkermans 1996, Borst et al. 1997, in the library. This is the appropriate solution since this is Borst 1997) is based on the reuse of the EngMath ontol- not a specific purpose unit, but a world wide generally ac- ogy (Gruber & Olsen 1994), and on general ontologies cepted one and it applies to all domains that may reuse the such as Mereology, Topology, Systems Theory, Component Standard-Units ontology. and Process (Borst et al. 1996), which were implemented in Ontolingua (Gruber 1993). To allow reuse, some gen- eral integration operations were identified, specified and 3.3 Methodologies That Include Integration defined. For instance, the Mereology ontology is reused The methodology to build ontologies presented in (Uschold (more precisely, extended) by the Topology ontology, The & King 1995) includes an integration step. This methodol- operations allowed are (Borst 1997): include and extend, ogy proposes that integration should be done either during include and specialize and include and map . With include capture (knowledge acquisition), coding (implementation) and extend the “imported ontology is extended with new or both. However no solutions for the problem of how in- concepts and relations”. With include and specialize an tegration is done are proposed or discussed. The problem “abstract theory is imported and applied to the contents of is only recognized as a difficult one. the importing ontology”. With include and map “different The methodology to build ontologies proposed in viewpoints on a domain are joined by including the views (Gruninger 1996) also refers integration. This methodol- in the domain ontology and formalization of their interde- ogy mentions two kinds of integration: “combining on- pendencies”. Since this operation contains a lot of domain tologies that have been designed for the same domain” knowledge it is considered an ontology on its own. In (V., and “combining ontologies from different domains”. Once Gomez-Perez, T. & Pinto 1998) we present the Reference again the problem of integration is considered difficult Ontology that was incorporated into the (KA) ontology since two ontologies may use the same terminology with (Benjamins & Fensel 1998), more precisely into the Prod- different semantics. According to this methodology, on- uct subontology. tologies are built based on ontology building blocks and There can be some less formal and clean ways of ontol- foundational theories. According to the building blocks ogy integration. For instance, (Dalianis & Persson 1997) and foundational theories of the ontologies being inte- describes the construction of an ontology for electrical dis- grated, integration is distinguished into: integration (at the tribution networks. This ontology was built by reusing level) of the building blocks, the most simple; integration an ontology for electrical transmission networks (Bernaras (at the level) of the foundational theories, which is more et al. 1996), more precisely its structural subontology. The difficult and may result in only partial integration; and on- ontology developed for electrical distribution networks is tology translation when the ontologies are so different that also a structural ontology. Although the domains are not they share neither the building blocks nor the foundational the same and some adaptations had to be performed, the theories, which makes integration extremely difficult. percentage of reused concepts was quite high. This hints METHONTOLOGY (Fernández et al. 1997, Blázquez, that perhaps a more general ontology could be defined. The Fernández, Garcı́a-Pinar & Gómez-Pérez 1998, Fernández, KACTUS (Schreiber, Weilinga & Jansweijer 1995) toolkit Gómez-Pérez, Sierra & Sierra 1999) is another methodol- was used to edit the pre-existent ontology (available from ogy to build ontologies that also considers integration. It They were named projections. These projections formalize the de- proposes that the development of an ontology should fol- pendencies between concepts and relations in different ontologies. low an evolving prototyping life cycle and not a waterfall In earlier versions it was named include and project. one. Although in earlier versions integration was consid- H.S.Pinto, A.Gomez-Perez, J.P.Martins 7-3 ered as a state during the development of an ontology (after formalization and before implementation), recent versions O consider it as an activity (as well as knowledge acquisi- D tion and evaluation (Gómez-Pérez, Juristo & Pazos 1995)) that should be performed since specification until mainte- nance. This methodology proposes that ontology building, and therefore ontology integration, should be done prefer- O1 O2 ... On ably at the knowledge level (Newell 1982) (in conceptual- D1 D2 Dk ization) and not at the symbol level (in formalization, when selecting the representation ontology) or at the implemen- tational level (when the ontology is codified in a target lan- Figure 1: Integration guage). O 3.4 Our View O1 In integration we have, on one hand, one (or more) on- O2 tologies that are integrated ( , Figure 1), and on the other hand, the ontology resulting from the integra- tion process (O, Figure 1). The integrated ontology(ies) are those that are being reused. They are a part of the resulting O3 O4 ontology. The ontology resulting from the integration pro- cess is what we want to build and although it is referenced as one ontology it can be composed of several “modules”, that are (sub)ontologies. This happens not only in integra- tion but when building an ontology from scratch. For in- stance, the Enterprise ontology (Uschold, King, Moralee & Zorgios 1998) is composed by several modules, called Figure 2: The resulting ontology in an integration process sections, like Meta-Ontology (where concepts like entity, can identify in the structure of the resulting ontology relationship, role, etc. are represented), Activities and Pro- ( , Figure 2) the four ontologies that were integrated cesses (where concepts like activity, resource, plan, etc. are ( , Figure 2) to build it. In this case, more represented), Organization, etc. However we call the re- knowledge had to be added to form the resulting ontology, sulting ontology the Enterprise ontology and usually do not besides knowledge coming from the integrated ontologies. refer to its parts unless we specifically want to talk about them. The ontologies reused are chosen from those available in ontology libraries that meet a series of requirements, for The domain of the integrated ontology is different from example, domain, abstraction, type (van Heist et al. 1997), the domain of the resulting ontology but there may be a generality, modularity, evaluation, just to name a few. The relation between both domains. When the integrated on- resulting ontology should have all the properties of a good tology is reused by the resulting ontology, the integrated ontology. Not only should it be clear, coherent, exten- concepts can be, among other things, (1) used as they are, sible, comply to the principle of the minimal ontological (2) adapted (or modified), (3) specialized (leading to a more commitment and to the minimal encoding bias , as pro- specific ontology on the same domain) or (4) augmented by posed in (Gruber 1995) but it should also be complete, con- new concepts (either by more general concepts or by con- cise, non-ambiguous (related to coherence), have an ade- cepts at the same level). The domains of the different inte- quate level of detail, be built upon the appropriate basic grated ontologies usually are different among themselves, distinctions, have been evaluated, etc. that is, each ontology integrated in the resulting ontology usually is about a different domain either from the result- The problem is how to integrate several existing on- tologies within a new one that is being built. Problems ing ontology (D, Figure 1) or the various ontologies inte- such as consistency of the resulting ontology, level of de- grated ( , where usually k = n, Figure 1). In tail throughout the whole ontology , etc, have to be dealt integration, the resulting ontology should be such that there is no similar ontology already built, otherwise one should Only describe the vocabulary needed to talk about the domain. simply reuse the existing one. In conceptualization one should only consider the knowledge level; In integration one can identify regions in the result- therefore choices for purely implementational convenience should be avoided. ing ontology that were taken from the integrated ontolo- That is, the ontology doesn’t have “islands” of exaggerated level of gies. Knowledge in those regions was left more or less detail and other parts with an adequate one. It should be stressed that none unchanged. In the example presented in Figure 2, one of the parts should have less level of detail than the one required or else the H.S.Pinto, A.Gomez-Perez, J.P.Martins 7-4 with. The solution seems to be the specification of a set of 1995, Swartout et al. 1997) (a natural language ontol- integration operations that tell how knowledge in the inte- ogy) was built by extracting and merging information from grated ontology is going to be included and combined with existing electronic sources. Several existing resources the knowledge in the resulting ontology. Integration oper- were used since each one had different and important fea- ations can be viewed as composing, combining or assem- tures: PENMAN Upper Model (Bateman, Kasper, Moore bling operations. However these operations should only & Whitney 1989), ONTOS, WordNet (Miller 1990) and an be performed if the integrated ontologies have a series of electronic natural language dictionary. The PENMAN Up- features. Not only will the features assure that the inte- per Model and ONTOS are high-level linguistically-based grated ontology is the most appropriate one but also that ontologies but lack a broad coverage of terms. They pro- the integration operations can be successfully applied and vided the upper level-organization. WordNet is a thesaurus- that the resulting ontology will have the desired charac- like hierarchically organized semantic network lacking the teristics. In (V. et al. 1998) we present a series of fea- upper level structure with broad coverage of terms. It pro- tures (and a WWW broker), not specifically for integra- vided the middle structure and terms. Finally, the electronic tion purposes, that can help the search for suitable ontolo- natural language dictionary, with both broad coverage of gies. Comparing both sets of the proposed integration op- words and Semantic Categories, provided both terms and erations we can see that “restriction” (Farquhar et al. 1997) the upper level organization. and “include&specialize” (Borst 1997) work similarly. As There have also been several efforts to find a top-level described in (Borst 1997), both “include&extend” and “in- ontology that could find the agreement of a broad num- clude&map” are abstract operations and can be performed ber of researchers and of systems. Among these “merged” in the Ontolingua server with the available operations. upper-level ontologies, we briefly describe Skuce’s (Skuce However a larger set of integration operations needs to be 1997) and Guarino’s (Guarino 1997) approaches since they identified, specified and defined. follow very different processes to try to achieve the same As the development of an ontology should follow an purpose: an upper-level ontology. Both upper-level on- evolving prototyping life cycle, the ontology may be con- tologies are unfinished proposals. Skuce (Skuce 1997) sidered for integration in specification, conceptualization, tries to find, at least, one of what is called an Agreed- formalization, implementation and maintenance. That is, Upon-Ontology. This ontology, a rather general one, de- we can have different integration procedures for the same fines the most general Fundamental Ontological Distinc- ontology but in different states of the ontology building tions - FOD’s which any top-level (general and domain- process. However the effort of integration varies: it is more independent) ontology should have. In (Skuce 1997), he significant in the earliest states (specification and concep- presents the definitions of concepts like ontology, primi- tualization) than in the final ones (after implementation). tive, distinction, category and entity and then presents the As we have an evolving prototyping ontology building life fundamental distinctions that he finds important. Some cycle the same ontology can be used in the same state in in- of them are concrete/abstract; atomic/composite; ma- tegration activities more than once. These procedures and terial/place; discrete/continuous; state/process; depen- activities form the overall process of integration. The in- dent/independent; and instance/predicate. tegration process needs to be further studied, namely, the Guarino’s approach (Guarino 1997) is based on the solid integration procedures and activities need to be defined. grounds provided by philosophers that have been address- ing these issues for the past 2000 years. His study on the 4 Merge ontological distinctions issues is organized around a theory Merge is the issue where more work has been developed, of parts, a theory of wholes, a theory of identity, a theory of so far. There is a wide variety of projects from a wide dependence and a theory of universals. From these theories range of domains, for example development of natural lan- he defines a preliminary taxonomy of top-level ontological guage ontologies, like SENSUS (Knight & Luk 1994), or concepts that combines clarity with semantic rigor, gener- ontologies on the medical domain like UMLS (Humphreys ality and common sense. His taxonomy of top-level onto- & Lindberg 1992); the search for agreed upper level ontolo- logical concepts is divided into an ontology of Particulars gies, like (Skuce 1997), (Guarino 1997) or (Sowa 1995); and an ontology of Universals. The backbone of distinc- the search for merge methodologies in the medical domain, tions of particulars around sortal categories considers sub- ONIONS (Gangemi et al. 1998). stract, object and quality. The non-sortal categories include Mereological, Physical, Functional, Biological, Intentional 4.1 Ontologies Built Through Merge and Social Stratums. Concrete objects have yet two other sets of distinctions: singular and plural; and body and fea- SENSUS (Knight & Luk 1994, Knight, Chander, Haines, ture. (Unary) Universals have two basic distinctions: tax- Hatzivassiloglou, Hovy, Iida, Luk, Withney & Yamada ontology would be useless, since it would not have sufficient knowledge Which associates certain word senses with particular fields (medicine, represented. biology, etc.). H.S.Pinto, A.Gomez-Perez, J.P.Martins 7-5 ons and properties. A more detailed description can be III (Cote, Rothwell & Brochu 1994) and GMN (Gabrieli found in (Guarino 1997). 1989) partial hierarchical ontologies; ICD10 classification (WHO 1994), which is a hierarchical ontology; and CORE 4.2 Methodologies To Perform Merge model (Rector et al. 1994) developed under the GALEN project (Rector, Solomon & Nowlan 1995) . In the med- Let us first present how SENSUS was built. First PEN- ical domain most available ontology sources are just tax- MAN, ONTOS and the Semantic Categories of the elec- onomies (UMLS is the exception). The ONIONS 6-step tronic natural language dictionary were merged, by hand. methodology can be summarized as: (1) Analyzing and This produced the ontology base (the upper level of the on- selecting the relevant sets of terms from the various ter- tology). Then, WordNet was merged into this base, again minological sources. All sources are considered in this by hand. Finally, a semi-automatic tool helped merging step. (2) Finding local definitions of the terms by ana- WordNet with the English Dictionary by matching similar- lyzing the classification criteria used to define the terms. ities in the textual definitions and by using the hierarchi- (3) Finding the theories related to the distinctions made cal organization of WordNet. This final merge was finally in the local definitions, that is one tries to find the gen- included in the ontology. In each merge step more than eral (global) ontologies (in contrast to the surface ontolo- one ontology or source of knowledge was considered at the gies associated to the local definitions found in the previ- same time. However, the merge process was subdivided for ous step). For that, one has to try to find theory chunks if each level of the ontology (beginning in the upper ones and no theories are available. (4) Finding the theories for the ending on the lower ones). In an effort to ease and search top-level design using the same procedures as in the pre- for a proper methodology to do this kind of merge process, vious step. The top-level categories found depend on the Hovy (Hovy 1996) tried to identify a set of relevant fea- “taste” of the ontological engineer, so the proposed taxon- tures that should be considered when comparing different omy should be taken as a possible choice and easily modi- ontologies for the purpose of merging ontologies, in partic- fiable. (5) Merging local definitions with the top-level cate- ular, natural language ones. gories. One tries to find the direct correspondences among The methodology followed by Skuce to find the ontolog- local items and elements of the theory chunks found for ical distinctions presented in (Skuce 1997) was brainstorm- the top-level or amends/enlarges theory chunks to allow ing, followed by meetings with other researchers interested local items to have room according to the proposed top- in the problem. The work is still incomplete, so he sug- level. (6) The model is formalized, and eventually, imple- gests to incrementally build from the list presented. The mented. In (Steve & Gangemi 1996), the ontological com- proposed methodology begins with the creation of a group mitments of this methodology are favorably analyzed (a- involving a diverse group of researchers working in differ- posteriori) through the ontological commitment rules pro- ent locations. Each member should develop a list of primi- posed in (Guarino, Carrara & Giaretta 1994). It would tives, distinctions and categories (a classification of FOD’s be interesting to see how general the methodology is, by according to the way they are defined) that should be care- trying to apply it to other domains. Its authors claim that fully chosen, defined and carefully documented (choices their methodological choices restrict the currently feasible and definitions). The choices are presented to the group development of formal ontologies to the merge of explicit for discussion and approval. Only when they are agreed task-oriented expert knowledge. upon can they get to the formalization stage. There can be several iterations of the previous steps. The agreed propos- 4.3 A Definition of Merge als are presented to wider audiences for criticism. Some can go back to the initial stages, some may be ready to be Sowa (Sowa 1997) defined merge as: accepted. The idea is to try to find a standardized upper- The process of finding commonalities be- model that would greatly ease some kinds of integration tween two different ontologies A and B and de- efforts. riving a new ontology C that facilitates inter- ONIONS (ONtologic Integration Of Naive Sources) operability between computer systems that are (Gangemi, Steve & Giacomelli 1996, Steve & Gangemi based on the A and B ontologies. The new on- 1996, Gangemi et al. 1998) is a methodology for merging tology C may replace A or B, or it may be used ontologically-heterogeneous taxonomic knowledge which as an intermediary between a system based on has been used to build the formal medical ontology IMO A and a system based on B. Depending on the (Integrated Medical Ontology) and a library of generic on- amount of change necessary to derive C from A tologies ON9. The ONIONS methodology was success- and B, different levels of “integration” can be fully applied to five medical sources: UMLS (Humphreys distinguished: alignment, partial compatibility, & Lindberg 1992, Humphreys & Lindberg 1993), a medical and unification. Alignment is the weakest form ontology implemented in a semantic network; SNOMED- The researchers of ONIONS were also involved in the GALEN The authors use the word integration. project. H.S.Pinto, A.Gomez-Perez, J.P.Martins 7-6 of “integration”: it requires minimal change, but it can only support limited kinds of interoperabil- O ity. It is useful for classification and information S retrieval, but it does not support deep inferences. Partial compatibility requires more changes in order to support more extensive interoperability, even though there may be some concepts or rela- O1 O2 ... On tions in one system or the other that could cre- S S S ate obstacles to full interoperability. Unifica- tion or total compatibility may require extensive changes or major reorganizations of A and B, but Figure 3: Merge it can result in the most complete interoperabil- computation that can be expressed in either one ity: everything that can be done with one can be can be mapped to an equivalent inference or com- done in an exactly equivalent way with the other. putation in the other. In the last quotation the word “integration” should be 4.4 Our View understood as merge. He further defines alignment as: In the merge process we have, on one hand, a set of ontologies (at least two) that are going to be merged Alignment: A mapping of concepts and rela- ( , Figure 3), and on the other hand, the re- tions between two ontologies A and B that pre- sulting ontology (O, Figure 3). The goal is to make a more serves the partial ordering by subtypes in both general ontology about a subject by gathering into a coher- A and B. If an alignment maps a concept or ent bulk, knowledge from several other ontologies in that relation x in ontology A to a concept or rela- same subject. The subject of both the merged and the re- tion y in ontology B, then x and y are said to sulting ontologies are the same (S, Figure 3) although some be equivalent . The mapping may be partial: ontologies are more general than others, that is, the level of there could be many concepts in A or B that have generality of the merged ontologies may not be the same. no equivalents in the other ontology. Before two In merge it may be difficult to identify regions in the ontologies A and B can be aligned, it may be nec- resulting ontology that were taken from the merged ontolo- essary to introduce new subtypes or supertypes gies and that were left more or less unchanged, specially in of concepts or relations in either A or B in or- the cases of unification. In the case of unification, knowl- der to provide suitable targets for alignment. No edge from the merged ontologies is homogenized and al- other changes to the axioms, definitions, proofs, tered through the influence of one source ontology on an- or computations in either A or B are made during other (is spite of the fact that the source ontologies do influ- the process of alignment. Alignment does not de- ence the knowledge represented in the resulting ontology). pend on the choices of names in either ontology. In other cases the knowledge from one particular source ... ontology is scattered and mingled with the knowledge that He further defines partial compatibility as: comes from the other sources. In the cases of alignment where minimal changes are required it may be possible to Partial compatibility: An alignment of two identify some regions in the resulting ontology that were ontologies A and B that supports equivalent in- taken from the merged ontologies. One can certainly find ferences and computation on all equivalent con- the concepts from the source ontologies unchanged since cepts and relations. If A and B are partially com- no changes are made either to axioms, definitions, proofs, patible, then any inference or computation that or computations in any of the source ontologies. In the hy- can be expressed in one ontology using only the pothetical example presented in Figure 4 we show a possi- aligned concepts and relations can be translated ble merge process of ontologies and into the result- to an equivalent inference or computation in the ing ontology . In this example, concepts from the source other ontology. ontologies can be identified. The way the merge process is performed is still very un- He further defines unification as: clear. So far, it is more of an art. Several different ap- proaches to the problem have been put forward by different Unification: A partial compatibility of two groups addressing the issue in different domains: in the nat- ontologies A and B that has been extended to a ural language domain it was done by hand; in the medical total compatibility that includes all concepts and domain a general methodology was proposed; and in search relations in both A and B. If the ontologies of A for generalized upper-models a group iterative approach to and B have been unified, then any inference or reach consensus was proposed and a one lonely researcher H.S.Pinto, A.Gomez-Perez, J.P.Martins 7-7 different needs and viewpoints arising from their particu- O lar contexts”; inter-operability among “different users that need to change data and who are using different software tools”; and systems engineering related to “the role ontolo- gies play in the operation of software systems”. In (Uschold 1998) a set of ten features is proposed to classify and characterize ontology applications. Once the characterization of applications is made, people new in the O1 O2 area wanting to build an ontology application could look up that information and avoid re-inventing the wheel . Also to ease the use of ontologies, (V. et al. 1998) presents a tax- onomy of seventy features and a WWW-broker that help future users to select the most adequate and suitable ontol- ogy for the application they have in mind. Another prob- lem related to use is the integration of Problem Solving Figure 4: The resulting ontology in a merge process Methods with ontologies (Chandrasekaran, Josepheson & effort based on philosophy was followed. So, one of the Benjamins 1998). problems of merge is what methodologies should be used to do it. There are a few different methodologies but there 5.2 Ontologies To Be Used is no consensus on the methodology to follow to merge on- tologies either on the same or in different domains. It even The ontology developed by ARPA/Rome Laboratory Plan- is not clear whether there could be a domain independent ning Initiative (Lehrer 1993), for representing plans and methodology. The common trace in the merging process is planning information is composed of: (1) an abstract on- that, at least, an initial group of ontologies (more than one) tology setting the more general categories (such as space, is analyzed together in its initial steps. While ONIONS time, agents); (2) a set of modular specialized ontologies starts with all ontologies at the beginning of the merge pro- which enlarges the general categories with sets of con- cess, Skuce begins with a selected initial group of ontolo- cepts and alternative theories of more detailed notions com- gies that is incrementally enlarged. While in integration we monly used by planning systems (for instance, specific on- could repeat the process for several integrated ontologies, tologies of temporal relations). The specialized ontolo- one ontology at a time, in the merge process an initial set of gies also provide definitions of concepts when several al- ontologies is, at least, analyzed together in the initial steps. ternative sets of concepts are commonly used to describe In merge there are no operations and what is mainly the same subject in the abstract ontology. These ontolo- used are abstraction, matching and generalization capabili- gies (abstract and concrete) are used by disparate and com- ties and common sense. A deep knowledge in Philosophy municating agents. This ontology can be classified as an also seems to help in this rather undefined process. Issues inter-operability one, according to the framework of uses concerning generalized upper levels have been studied for in (Uschold & Gruninger 1996). more than 2000 years in that science. Some general guide- Among the various ontologies already built for use we lines to guide the abstraction process, like ONIONS, seem can refer: CYC (Lenat & Guha 1990, Lenat & Guha 1994, to help this rather complex process. The problem seems to Lenat, Guha, Pittman, Pratt & Shepherd 1990) , GUM be to find them. (Bateman, Magnini & Fabris 1995), PIF (Lee, Gruninger, Jin, Malone, Tate, Yost & other members of the PIF Work- 5 Use/Application ing Group 1996) the GALEN project, UMLS. In this section we present some ontologies that were built 5.3 Tools For Use either from scratch, through integration, or through merge that were actually used by real applications. We also KACTUS (Knowledge About Complex Technical sys- present work aiming at easing and characterizing use. We tems for multiple USe) project (Schreiber et al. 1995, should stress that there aren’t many reports of application Laresgoiti, Anjewierden, Bernaras, Corera, Schreiber & of ontologies in the literature and the few existing reports Wielinga 1996) aims at “finding and building methods and do not give enough technical details on how the ontology tools to enable reuse and sharing of technical knowledge”. was used by the application. It began by modeling reasoning processes and then went on For instance, look up the techniques used to build similar 5.1 Easing Use applications. http://www.cyc.com In (Uschold & Gruninger 1996) a series of uses of ontolo- http: gies was identified: communication “between people with //swi.psy.uva.nl/projects/NewKACTUS/home.html H.S.Pinto, A.Gomez-Perez, J.P.Martins 7-8 modeling ontologies. KACTUS developed a methodology, a tool and a library to help the construction of KBS for A complex technical domains. According to the framework of uses in (Uschold & Gruninger 1996), the ontologies kept at the library can be classified as systems engineering ones. ... 5.4 Ontologies That Were (Re)Used O1 O2 On In (Laresgoiti et al. 1996), a case of reuse of a previously built ontology in a new application is presented. The con- clusions were that the reuse of ontologies saved a lot of Figure 5: Use effort in implementing applications. However reuse is al- translation of ontologies involves a significant manual ef- most never complete so “reusing any ontology for differ- fort (which is considerable difficult to automatize). ent purposes than those for which the ontology was built Other ontologies that were actually used to build ap- will always require modifications or tuning for the new pur- plications and whose application is discussed in the lit- poses”. They also concluded that a clear organization and a erature are PhySys (Borst 1997), the Enterprise Ontology clear methodology is needed so that people unfamiliar with (Uschold, King, Moralee & Zorgios 1998), the Reference the ontology are able to use it. Ontology (V. et al. 1998), among others. Another important application that uses ontologies is described in (Bernaras et al. 1996). First, an ontology 5.5 Our View for diagnosis in electrical transmission networks was built. In use, there are one or more ontologies involved Then, an ontology for service recovery planning on the ( , Figure 5) and there is no resulting ontol- same domain was built. Finally, they were both unified in ogy. (A, in Figure 5, is the application using the ontology). the sense of merge&integration. From the unified ontology One cannot draw any conclusions as to the architecture of several applications that needed knowledge about electrical the resulting application because that depends on the appli- power transmission systems were developed. The conclu- cation itself. In the case that several ontologies are used sions were that modularization and hierarchical organiza- they should be compatible among themselves. There are tion seem to be good ontology structuring and design prin- several issues involved when analyzing compatibility: lan- ciples and that abstraction and standardization, although guage, ontological commitments, level of detail, context, good principles, should be used with care. Concrete objects etc. However, we think that there is an order of importance were more usable than more abstract ones (although ab- among those different compatibility criteria. If two ontolo- straction is a basic principle to reusability). The fact is that gies are not compatible in their ontological commitments to implement specific concepts from more generic ones de- then all other criteria are irrelevant and it is meaningless to manded a big design effort, more than implementing those analyze them. The ontologies should also satisfy a set of concepts from other specific concepts from related appli- characteristics such as level of generality, modularity, etc., cations. The conclusions reached in (Bernaras et al. 1996) as discussed in (Bernaras et al. 1996). Finally, only verified about the problem of ontology use are that these ontologies (Gómez-Pérez et al. 1995) ontologies should be considered should be built only a small level up in generality than the in use. one used for a specific application, so that the implemen- So far, there are no operations identified in the literature. tation of the ontology in other applications won’t involve a The ontologies used in applications should probably have big design effort. a set of configurable parameters. In each application de- EngMath was reused to build an application for air- veloped based on that ontology those parameters should be craft design (Uschold, Healy, Williamson, Clark & Woods customized. In what concerns use’s methodological aspects 1998). The process of reusing the ontology can be sum- a lot of work needs to be done. One can probably find some marized as: (1) understanding the ontology and finding guidelines that can ease the process of using the ontology, the kernel of reusable knowledge; (2) translate the ontol- as (Laresgoiti et al. 1996) tries to establish. We think that a ogy (that was initially written in Ontolingua) into Slang set of specific methodologies based on the kind of applica- (Waldinger, Srinivas, Goldberg & Jullig 1996); (3) specify tion involved could also help the use process. Among these and refine the task definitions in an iterative process mov- specific methodologies are methodologies for use in KBS, ing closer and closer to the implementation of that spec- Internet brokers, etc. ification; (4) verify each refinement step; (5) integrate the resulting specification in the specification of the application 6 Conclusions and refine its result into executable code. The conclusions reached were that it actually was cost-effective to reuse the In this article we clarify the meaning of the word integra- ontology instead of building it from scratch; and that the tion in the OE field. The three acknowledged meanings H.S.Pinto, A.Gomez-Perez, J.P.Martins 7-9 associated to the word “Integration” should actually be de- References fined using the following words: Bateman, J. A., Kasper, R. T., Moore, J. D. & Whitney, R. A. (1989), A General Organization of Knowledge 1. Integration - In the case of building a new ontology for Natural Language Processing: The PENMAN Up- reusing (by composing) other available ontologies. per Model. research report, USC/Information Sci- ences Institute, Marina del Rey. 2. Merge - In the case of building an ontology unifying knowledge of several ontologies into a single one. Bateman, J., Magnini, B. & Fabris, G. (1995), The Gen- eralized Upper Model Knowledge Base: Organiza- 3. Use/Application - In the case of integrating ontologies tion and Use, in N. Mars, ed., ‘Towards Very Large in applications. Knowledge Bases’, IOS Press, pp. 60–72. Benjamins, R. & Fensel, D. (1998), The Ontological En- Both integration and merge are processes that aim at gineering Initiative (KA) , in N. Guarino, ed., ‘For- building ontologies from other ontologies. However this is mal Ontology in Information Systems’, IOS Press, where similarities end. These processes are quite different pp. 287–301. one from the other, as we have discussed. Use/Application is a completely different process. The objective is not to Bernaras, A., Laresgoiti, I. & Corera, J. (1996), Build- build an ontology. The aim is to build an application using ing and Reusing Ontologies for Electrical Network ontologies. Applications, in W. Wahlster, ed., ‘ECAI96 Proceed- In this article we present a series of partial conclusions ings’, John Wiley&Sons, pp. 298–302. on the three different processes that we have identified. We have identified the main characteristics of the integration Blázquez, M., Fernández, M., Garcı́a-Pinar, J. M. & process, presented some of the operations proposed in the Gómez-Pérez, A. (1998), Building Ontologies at the literature to do integration and presented how current on- Knowledge Level Using the Ontology Design En- tology building methodologies deal with this issue. Further virnoment, in ‘Knowledge Acquisition Workshop’. work is needed in order to identify a larger set of opera- Borst, P. (1997), Construction of Engineering Ontolo- tions that can be used in the integration process, specify gies for Knowledge Sharing and Reuse, PhD thesis, how those operations should work in each phase of the on- Tweente University. tology development life cycle and specify the overall inte- gration process. Further research in integration is within Borst, P., Akkermans, H. & Top, J. (1997), ‘Engineering our plans. Ontologies’, International Journal of Human Com- We have also identified the main characteristics of the puter Studies special issue “Using Explicit Ontolo- merge process. In the literature, no operations to do merge gies in KBS Development” 46(2/3), 365–406. are proposed and there are very few methodologies. Some work is needed in order to specify more and better defined Borst, P., Benjamin, J., Wielinga, B. & Akkermans, H. (1996), An Application of Ontology Construction, in methodologies for this process of building ontologies and ‘ECAI96’s workshop on Ontological Engineering’. see whether these methodologies can be applied to different subjects, domains and contexts. Chandrasekaran, B., Josepheson, J. R. & Benjamins, Finally, we have identified the main characteristics of V. R. (1998), Ontology of Tasks and Methods, in the use process. As in the previous process, no operations ‘ECAI98’s workshop on Application of Ontologies to do use were proposed although there are some guide- and Problem Solving Methods’. lines in the literature about the features that those ontolo- gies should have. Some work is needed in order to develop Cote, R. A., Rothwell, D. J. & Brochu, L., eds (1994), methodologies to do use of applications in spite of the fact SNOMED International, 3rd edn, Northfield III: Col- that these methodologies will be application dependent. lege of American Pathologists. Dalianis, H. & Persson, F. (1997), Reuse of an Ontology 7 Acknowledgments in an Electrical Network Domain, in ‘AAAI97 Spring Symposium Series, workshop on Ontological Engi- We thank all members of GIA at Instituto Superior Técnico neering’. for their support, in particular Ana Cachopo, João Cachopo and António Leitão. We thank all members of the Knowl- Farquhar, A., Fikes, R., Pratt, W. & Rice, J. (1995), Col- edge Sharing and Reuse group at Lab. de Inteligencia laborative Ontology Construction for Information In- Artificial at UPM for their support, in particular Mariano tegration, Technical Report KSL-95-63, Knowledge Fernández. Systems Laboratory, Stanford University. H.S.Pinto, A.Gomez-Perez, J.P.Martins 7-10 Farquhar, A., Fikes, R. & Rice, J. (1996), The Ontolingua Guarino, N. (1997), Semantic Matching: Formal Ontologi- Server: A Tool for Collaborative Ontology Construc- cal Distinctions for Information Organization, Extrac- tion, in ‘Knowledge Acquisition Workshop’. tion, and Integration, in M. T. Pazienza, ed., ‘Informa- tion Extraction: A Mutidisciplinary Approach to an Farquhar, A., Fikes, R. & Rice, J. (1997), Tools for Assem- Emerging Information Technology’, Springer Verlag. bling Modular Ontologies in Ontolingua, in ‘AAAI97 Proceedings’, pp. 436–441. Guarino, N. (1998), Formal Ontology and Information Sys- tems, in N. Guarino, ed., ‘Formal Ontology in Infor- Fernández, M. (1996), CHEMICALS: Ontologia de Ele- mation Systems’, IOS Press, pp. 3–15. mentos Quı́micos. Proyecto Fin de Carrera, Fac. de Informática, UPM. Guarino, N., Carrara, M. & Giaretta, P. (1994), Meta-Level Categories, in E. S. J. Doyle & P. Torasso, eds, ‘KR94 Fernández, M., Gómez-Pérez, A. & Juristo, N. (1997), Proceedings’, Morgan Kaufmann, pp. 270–279. METHONTOLOGY: From Ontological Art Towards Hovy, E. (1996), What Would it Mean to Measure an On- Ontological Engineering, in ‘AAAI97 Spring Sympo- tology? Work in progress. Unpublished ms, USC/ISI. sium Series, workshop on Ontological Engineering’. Humphreys, B. L. & Lindberg, D. A. B. (1992), The Uni- Fernández, M., Gómez-Pérez, A., Sierra, A. P. & Sierra, fied Medical Language Systems Project, in K. C. Lun, J. P. (1999), ‘Building a Chemical Ontology Using ed., ‘MEDINFO92’, Elsevier Science Publishers. METHONTOLOGY and the Ontology Design Envi- ronment’, IEEE Intelligent Systems 14(1), 37–46. Humphreys, B. L. & Lindberg, D. A. B. (1993), ‘The UMLS Project: Making the Conceptual Connection Gabrieli, E. A. (1989), ‘A New Electronic Medical Nomen- Between Users and the Information They Need’, Bul- clature’, Journal of Medical Sciences 3(6). letin of the Medical Library Association 81(2). Gangemi, A., Pisanelli, D. M. & Steve, G. (1998), On- Knight, K., Chander, I., Haines, M., Hatzivassiloglou, V., tology Integration: Experiences with Medical Termi- Hovy, E. H., Iida, M., Luk, S. K., Withney, R. A. nologies, in N. Guarino, ed., ‘Formal Ontology in In- & Yamada, K. (1995), Filling Knowledge Gaps in a formation Systems’, IOS Press, pp. 163–178. Broad-Coverage MT System, in ‘IJCAI95 Proceed- ings’, pp. 1390–1397. Gangemi, A., Steve, G. & Giacomelli, F. (1996), ONIONS: an ontological methodology for taxonomic knowl- Knight, K. & Luk, S. (1994), Building a Large Knowledge edge integration, in ‘ECAI96’s workshop on Ontolog- Base for Machine Translation, in ‘AAAI94 Proceed- ical Engineering’. ings’, pp. 773–778. Gómez-Pérez, A., Juristo, N. & Pazos, J. (1995), Eval- Laresgoiti, I., Anjewierden, A., Bernaras, A., Corera, J., uation and Assessement of the Knowledge Sharing Schreiber, A. T. & Wielinga, B. J. (1996), Ontolo- Technology, in N. Mars, ed., ‘Towards Very Large gies as Vehicles for Reuse: a mini-experiment, in Knowledge Bases’, IOS Press, pp. 289–296. ‘Kowledge Acquisition Workshop’. Lee, J., Gruninger, M., Jin, Y., Malone, T., Tate, A., Gruber, T. (1993), Ontolingua: A Mechanism to Support Yost, G. & other members of the PIF Working Group Portable Ontologies, Technical Report KSL-91-66, (1996), Process Interchange Format (PIF) for Sharing Knowledge Systems Laboratory, Stanford University. Ontologies, in ‘ECAI96’s workshop on Ontological Gruber, T. (1995), ‘Towards Principles for the Design Engineering’. of Ontologies for Knowledge Sharing’, International Lehrer, N. (1993), KRSL Specification Language, Tech- Journal of Human Computer Studies 43(5/6), 907– nical Report 2.0.2, ISX Corporation. Also available 928. as http://isx.com/pub/ARPI/ARPI-pub/ krsl/krsl-info.html. Gruber, T. & Olsen, G. R. (1994), An Ontology for Engineering Mathematics, in E. S. J. Doyle & Lenat, D. & Guha, R. (1990), Building Large Knowledge- P. Torasso, eds, ‘KR94 Proceedings’, Morgan Kauf- Based Systems, Representation and Inference in the mann, pp. 258–269. CYC Project, Addison Wesley. Gruninger, M. (1996), Designing and Evaluating Generic Lenat, D. & Guha, R. (1994), ‘Enabling Agents to Work Ontologies, in ‘ECAI96’s workshop on Ontological Together’, Communications of the ACM 37 (7), 127– Engineering’. 142. H.S.Pinto, A.Gomez-Perez, J.P.Martins 7-11 Lenat, D., Guha, R., Pittman, K., Pratt, D. & Shepherd, Uschold, M. & King, M. (1995), Towards a Methodology M. (1990), ‘CYC: Toward Programs with Common for Building Ontologies, in ‘IJCAI95’s workshop on Sense’, Communications of the ACM 33 (8), 30–49. Basic Ontological Issues in Knowledge Sharing’. Miller, G. (1990), ‘WordNet’, International Journal of Lex- Uschold, M., King, M., Moralee, S. & Zorgios, Y. (1998), icography 3. The Enterprise Ontology, Technical Report AIAI- TR-195, Artificial Intelligence Applications Institute, Newell, A. (1982), ‘The Knowledge Level’, Artificial In- University of Edinburgh. To appear in the Knowl- telligence 18, 87–127. edge Engineering Review, vol 13, Special Issue on Putting Ontologies to Use (eds. Mike Uschold and Rector, A. L., Gangemi, A., Galeazzi, E., Glowinski, A. & Austin Tate). Rossi-Mori, A. (1994), The GALEN CORE Model Schemata for Anatomy: Towards a Re-Usable Ap- V., J. A., Gomez-Perez, A., T., A. L. & Pinto, H. S. (1998), plication Independent Model of Medical Concepts, in (ONTO) Agent: An Ontology-Based WWW Broker ‘Medical Informatics Europe’. to Select Ontologies, in ‘ECAI98’s workshop on Ap- plication of Ontologies and Problem Solving Meth- Rector, A. L., Solomon, W. D. & Nowlan, W. A. (1995), ‘A ods’. Terminology Server for Medical Language and Med- ical Information Systems’, Methods of Information in van Heist, G., Schreiber, A. T. & Wielinga, B. J. (1997), Medicine 34, 147–157. ‘Using Explicit Ontologies in KBS Development’, Internacional Journal of Human-Computer Studies Schreiber, G., Weilinga, B. & Jansweijer, W. (1995), The 45, 183–292. Kactus View on the ‘o’ World, in ‘IJCAI95’s work- shop on Basic Ontological Issues in Knowledge Shar- Waldinger, R., Srinivas, Y. V., Goldberg, A. & Jullig, R. ing’. (1996), Specware Language Manual. -. Skuce, D. (1997), How We Might Reach Agreement on WHO (1994), Classification of Diseases, 10th edn, WHO. Shared Ontologies: A Fundamental Approach, in ‘AAAI97 Spring Symposium Series, workshop on Ontological Engineering’. Sowa, J. F. (1995), ‘Top-Level Ontological Categories’, International Journal of Human Computer Studies 43(5/6), 669–685. Sowa, J. F. (1997), Electronic communication in the onto-std mailing list, 4th of December. Steve, G. & Gangemi, A. (1996), ONIONS Methodology and the Ontological Commitment of Medical Ontol- ogy ON8.5, in ‘Kowledge Acquisition Workshop’. Swartout, B., Patil, R., Knight, K. & Russ, T. (1997), To- ward Distributed Use of Large-Scale Ontologies, in ‘AAAI97 Spring Symposium Series, workshop on Ontological Engineering’. Uschold, M. (1998), Where are the Killer Apps?, in ‘ECAI98’s workshop on Application of Ontologies and Problem Solving Methods’. Uschold, M. & Gruninger, M. (1996), ‘Ontologies: Prin- ciples, Methods and Applications’, Knowledge Engi- neering Review 11(2). Uschold, M., Healy, M., Williamson, K., Clark, P. & Woods, S. (1998), Ontology Reuse and Application, in N. Guarino, ed., ‘Formal Ontology in Information Systems’, IOS Press, pp. 179–192. H.S.Pinto, A.Gomez-Perez, J.P.Martins 7-12