On the Impact of Ontological Commitment Marian Nodine Jerry Fowler Telcordia Technologies setenv, LLC 106 E. 6th Street, Suite 415 Houston, TX Austin, TX 78733 (713)526-8641 (512)478-8923 gfowler@acm.org nodine@research.telcordia.com ABSTRACT In an agent-based system, common ontologies specify the Ontological commitment, or the agreement to have your ontological commitments of a set of participating agents [16]. An applications and users conform to a common domain ontological commitment is an agreement to use a vocabulary in a understanding as encapsulated in one or more shared ontologies, way that is consistent with an ontology. An agent or human is a noble goal and essential for open agent systems. Our committed to an ontology understands (some subset of) the experiences building ontology-based agent systems in multiple ontology and agrees to use it in a manner consistent with the domains have shown us that the intention for a new application semantics of the ontology. Agents and humans committed to the to locate and conform to some existing ontology or ontologies same ontology can share knowledge among themselves with within its domain has many impediments to its success. For some confidence that they share an underlying understanding of instance, the goals of the designer of a domain ontology include what is being said. Commitment to common, shared ontologies developing a complete and comprehensive domain description; facilitates openness in an agent-based system. however, the application developer may only require a small In this paper, we examine the conflicting requirements and goals fragment of that ontology. Multiple applications that conform to of ontology designers, ontology-committed applications, and the ontology may, in fact, use completely orthogonal fragments of ontology-aware users, and their respective impact on the problem the ontology, and not be able to interact at all. Users may insist of ontology commitment and reuse. As ontological sharing and on importing into the ontology sets of terms that are neither reuse increases, the gap between the ontology designers and the logically consistent nor easily modelable. ontology users grows larger. Our goal is to analyze what issues inhibit reuse and to propose strategies for facilitating reuse. In With these issues in mind, in this paper we propose some particular, we consider the problem of reuse of ontologies whose guidelines for ontology development and evolution paradigm that specification is complete, for applications whose requirements should facilitate ontology reuse. These guidelines could were not considered during the design of the ontology. This underpin a usage model for ontologies; one that enables the problem is not addressed in the ontology design methodologies application designer to reuse ontological concepts from multiple summarized in [17]. We develop guidelines and approaches for ontologies in a more flexible manner, while retaining the agents to use existing ontologies in a more flexible manner. essentially good properties of ontology sharing and reuse. These guidelines affect both the design and use of ontology-based Throughout the paper, we relate the issues to real issues we have applications, as well as the way applications advertise encountered within the context of one of our applications, EDEN themselves to other agents with which they may interoperate. [3]. EDEN1 is an agent-based system developed for the purpose of inter-organizational sharing of environmental data collected, stored and monitored by multiple government agencies and non- 1. INTRODUCTION government scientists spread throughout the US and Europe, and The goal of knowledge representation is to make explicit the relating information from these disparate data sources and semantics of a particular domain of interest for the purposes of schemas at a semantic level as needed by the users. EDEN uses sharing the knowledge among humans and computer artifacts. ontologies to represent the semantics of the underlying Sowa [11] subdivides knowledge representation into categories: information, and real and varied databases to populate those “Logic provides the formal structure and rules of inference. ontologies with instances. Ontology defines the kinds of things that exist in the application domain. 2. ONTOLOGICAL COMMITMENT Because ontologies are meant to facilitate sharing and reuse of Computation supports the applications that distinguish knowledge, it is important that the ontology and its collection of knowledge representation from pure philosophy.” users (both human and agent) align themselves to a shared view There is a strong relationship between some specific ontology of the domain during the process of designing and evolving the and the logical rules and computational artifacts that use that ontology for that domain. However, many existing ontologies ontology, in that when they communicate among themselves, they have been developed either by designers attempting to have some level of assurance that the same terms have the same characterize a domain (with no real computational applications meanings to all. However, this use requires that the logical rules that use them) or by application developers to support individual and the computational artifacts have explicit linkages with the applications (with no real sharing of the ontology with other ontology; often in the form of hard-coding the ontological terms into the rules and/or the application code itself. 1 EDEN was funded jointly by the DOD, DOE, EPA and EEA. applications). The plethora of existing ontologies argues that Environmental Data Registry [4], it was difficult to relate many concepts are already represented within some ontology, so measurements of chemicals between different representations reuse of these ontologies, increasing the ontological commitment because the measurement ontology was “well understood,” and level, is now feasible. For example, some ontologies such as the often explicit in the data representations themselves. Even terms Unified Medical Language System (UMLS) Metathesaurus [7] such as “milligrams per kilogram” were sometimes represented have achieved a higher level of commitment. inconsistently in text fields. This posed a problem in extracting There are several issues that impede this sharing and knowledge. values for computation in this “well understood” measurement First are issues related to the conflicting goals and objectives of system. ontology definers, ontology-committed applications, and users of The desire of the application user is to be able to do his job well such applications. Second are issues of the mutual conflict and easily. Application users have expertise in their own jobs between ontological commitment and ontological evolution. and potentially in the domain of the application, but may have Third are issues of conceptual mismatches between ontologies minimal understanding of knowledge representation or and the applications and users that use them. Unfortunately, application design issues. Typical users want an understandable, these issues stem from fundamental issues and characteristics of natural, easy-to-use interface to the application that facilitates the problem of sharing ontologies so broadly. their work. This implies that they want the scope of the ontology restricted to the exact area within the domain that they are 2.1 Conflict: Definer, Application, User specifically concerned with. Another issue is that a user may The commitment to ontologies is hampered by the conflicting have a comfortable vocabulary of domain-related terms that do goals of ontology definers, developers of ontology-committed not map well to the ontology representation model, to the domain applications and ontology-committed application users, and often model that the ontology developer had in mind or to the needs of by the confusion of users and definers over the demands of the application itself. ontology-committed applications. Here, we use the descriptor “ontology-committed” to mean that something purports to use the 2.2 Conflict: Commitment, Evolution terms in the ontology(-ies) in a manner consistent with its (their) A second hampering factor when considering ontological definitions. commitment itself is the fact that ontological commitment The goal of the ontology designer, at least one working towards impedes ontology evolution, and vice versa. Yet, both of these maximizing the usefulness of his ontology to a wide variety of are necessary for an ontology to be successful. As Mark Tuttle applications, is to completely characterize a particular domain at et.al. say, “if you don’t use it, it won’t improve, if you don’t the semantic level. The ontology designer needs expertise in improve it, it won’t get used.” [13]. In other words, use brings knowledge representation and in the domain of the ontology. His out the need to change, change is required for continued use. intent is to develop a comprehensive and up-to-date ontology, As stated before, a measure of ontological commitment is the with a broad set of acceptable terms. His job is hampered by the number of agents and humans that are committed to using the fact that the different domain experts have different viewpoints ontology. Each of these committed entities has made the effort to of the domain, and these viewpoints must be reconciled within learn the ontology and to use the concepts therein in a manner the ontology itself, placing pressure on the designer to either take consistent with their definition. However, the ontology itself may a commonly agreed-upon but consistent subset of the domain, or have problems, or may be incomplete in some aspect, or not to attempt to placate everyone by including everything. The likely reflect changes in the domain itself. In these situations, it result is either the ontology will express concepts at a higher becomes desirable at some point to evolve the ontology, level than can be used effectively by an application that seeks to providing an updated version. Initially, there is no real attach suitable labels to real data, or the terminology within the commitment to the updated version; rather, the entities ontology will not be logically consistent and thus easily committed to the old ontology must explicitly learn the new incorporatable into an application. Fortunately, our experience ontology, and adapt to the changes therein. This leads to the indicates that many terminology disagreements can be addressed observation that, as commitment to an ontology rises, so does the by synonymy. For instance, what the EPA refers to as a “site” the level of effort required to accommodate an evolutionary step. DOD calls a “facility.” What the EPA calls an “operable unit” (a Development of an application represents a strong degree of specific location contained in an EPA “site”) the DOD calls a ontological commitment. The incorporation of an ontology into an “site.” application may require specific coding with respect to the The goal of the application designer is to use the ontology to ontology, so application developers tend not to appreciate it when support an application. The application designer has expertise in the ontology evolves. This tends to discourage users from building applications, especially within given domains. His intent employing the precision they need to express their views of the is to develop an efficient and useful application. It is likely that data in an application. In EDEN, the first version of our ontology the application will not cover the whole domain, or may span easily evolved into the second, because there were only three multiple domains. Also, applications need to focus on some small database fragments and a single user view in the initial issues that are unimportant to the designer. For instance, a failure demonstration. As the complexity of the user interface increased to take sufficient care of value representation issues in the to make multiple views possible, and as we added larger, more ontology may force the application designer to code this complex resources, the task of evolving the ontology from version information directly into the application, even though it is really two to version three became more demanding due simply to the domain-related knowledge. For example, within the numbers of agents impacted. 2.3 Impedance Mismatch attracted to the same ontology. This affects not only the original There are several areas where mismatch and other problems design, but also the process of evolving the ontology. seem to crop up. This problem has been studied extensively within the heterogeneous database community [19,20]. Some 3.1 Design and Representation Ideally, the design and representation of an ontology should be examples of these issues are as follows: done in collaboration with users and applications; the Uneven concept granularity: Some ontologies and other applications then can incorporate the ontology as designed, and dictionaries were developed with a focus on one particular aspect the users can validate the concepts, relationships and axioms that of the domain. This aspect was well-developed and well-tested, they access in the ontology via the application. These but other aspects of the same domain are ignored. For instance, interrelationships are shown in Figure 1. the ontology could have very strong and detailed concepts for contaminants and hazardous waste sites on land and in lakes, but Representability vs. Usability vs. not for ocean contaminants. However, a general application Implementability Representability concerning toxic waste sites would need all of these concepts. Concept modeling mismatch: Some applications and information sources model individual concepts differently from Domain Speci- Speci- Domain others, thus it was hard to develop a single ontological concept knowledge, fication Ontology fication knowledge, validation Designer validation that would relate well to all the ways it could be represented. For instance, the concept of a set of measurements of contaminant levels taken daily could be represented in the ontology and/or the Implementation data itself either with a class of measurement vectors (one instance per day), or a class of measurements, with one instance Requirements, per measurement type per day, specialized on measurement type. validation Value representation mismatch: Many ontologies we have Application Application found do not even consider issues relating to the representation Implementability vs. Designer User of the values of the attributes in the instances, a requirement for Usability many applications. For instance, it would be easy to say that “area” was an integer; therefore setting its range as the range of positive integers. However, the Americans measured area in Figure 1. Ontological Interactions square feet, etc, while the Europeans measured it in square meters, etc. The application needed to know in which unit the Partly because of the desire to maximize ontological commitment, it is normally the goal of the ontology designer to integer in a particular instance of the area is represented, and represent the domain as completely as possible. Ontologies such how the units relate. These concepts should therefore exist in the domain ontology. In a more complex example, one information as UMLS or environmental thesauri such as the General Multilingual Environmental Thesaurus (GEMET) [6] or the source in the EDEN application had “soil” as one of the media Terminology Reference System (TRS) [12] provide a broad base that could be polluted; but a second had “topsoil” and “subsoil” for discourse over their relative domains, medicine and the as separate concepts. A third site had multiple, more fine-grained environment. Their definition has largely been driven, not by the classifications for soil. These did not all map easily onto any implementers of the computer systems that use the ontologies, single ontological representation for soil, as mapping between the but rather by the practitioners within the domain itself. An different concepts required additional knowledge (e.g. the depth of the soil), and sometimes was lossy (e.g. if everything was just example of the practical impact of this high-level design on the EDEN system was the design of value mapping across conceptual mapped to “soil”). domains. The Environmental Data Register’s (EDR) [4] Terminology mismatch: Sometimes different sets of users use implementation of its value domain model was fully normalized. the same term for different concepts, or the term best suited to This produced response times that were suitable for individual the user is already present in the ontology with a different term lookups and mappings, but the performance scaled meaning. This means that necessary concepts are sometimes inadequately when several agents were performing simultaneous omitted because the lexical terms needed to express them in large-scale term conversions for entire tables. To obtain any human language are all present in other usages. For instance, the tolerable performance at all, we were compelled to revise the UMLS set of Semantic Relations was slow to develop the ontology of the conceptual domain to support a flatter, non- concept of genetic relationship because the terms parent and normal data structure. child existed for their use in logical tree structures. All of these issues conspire to make ontology sharing 3.2 Evolution challenging. As application developers develop applications using ontologies, and users use those applications, they identify weaknesses, 3. ONTOLOGY DEVELOPMENT ISSUES missing concepts, and different issues that need to be addressed. For ontological commitment to become real, ontology designers, For instance, when we developed our initial environmental application designers and users within a domain need all to be ontology, some of the users who were used to the terminology in GEMET requested that we evolve our own ontology to be consistent with GEMET, or otherwise incorporate its terms. This request mandated an incorporation of that ontology into the reuse as much already-defined ontological information as he can environmental ontology. Once this incorporation was complete, within his application, tagging the imported concepts back to the all agents had to be adapted to the new ontology (this required a ontologies from which they were taken to facilitate the process of varying amount of effort depending on how closely the agent was evolving the application within the ontology. Unfortunately, up to hardwired to the ontology). Some of the users who were not used this point the application developer has had no input whatsoever to the GEMET terminology also had to learn it. Thus, the into the ontological design, though he has been working closely evolution of the ontology engendered quite a bit of effort on many with the prospective application users. Thus, the necessary people. One could envision that the evolution of an ontology that collaboration between the ontology designer, the application has a higher level of ontological commitment, such as UMLS, developer and the users is missing in this situation. Furthermore, could get very complex. the users’ input into the application design is much more direct Clearly, the evolution of an ontology to a new version must be than that of the ontology designer. The result often is that there is done carefully and collaboratively, involving the same types of as no good existing ontological “match” to the application and/or its were involved in the original design. This indicates a slower, users. more coarse-grained type of evolution, which in turn affects the Our experiences with our own ontology-based agent applications contents of the ontology. An ontology may contain several types has led us to the following observations: of information useful to the domain that it represents. These 1. Reusing ontologies is necessary for acceptance by domain include: specialists, but many mature ontologies specify a lot more • Concepts/classes, their attributes, and the types of than current agent systems can actually use. representations their attributes can take. 2. Most of the applications we have developed require • Relationships between concepts, including subclassing, concepts from multiple ontologies. synonyms, and containment. 3. Most of the applications that we have developed require • Axioms and functions over those concepts. some concepts defined in no ontologies, because of mismatch issues. • Instances of those concepts the relationships between them, and axioms over the instances. GEMET GPS The first three of these are referred to as the schema information EDEN Positioning (also known as the conceptual model or meta-knowledge) and the last is referred to the instance knowledge. Experience has shown us that instances, because they represent changing objects in the real world, tend to evolve much faster Facility Extension EDR than the concepts themselves. Because of this, it is easier to factor the ontology into two pieces, the schema and the instances. Within the EDEN application, the environmental ontology itself only contained schema-level domain information, and we populated the schema using information from different environmental databases. Because the agents themselves Figure 2. A Compositional Ontology naturally were written only against the schema concepts, and the In order to facilitate reuse, we have designed our applications to schema-level ontology information evolved more slowly, ontology use a compositional notion of ontologies, where we select evolution did not create as significant a problem as it could have. concepts from different ontologies as needed, and supplement the This experience has carried over into numerous other result with any new concepts that we cannot locate elsewhere. applications developed using the InfoSleuth agent system [8,3]; This leads us to support three operations over ontologies thus, we recommend this factoring unless the instance themselves, which we term subset, compose and extend, information is guaranteed to be stable (e.g., constants). respectively, as shown in Figure 2. Ontological representation systems maintain this natural factoring between schema and instances. In OIL [9], the schema 4.1 Subset level is encapsulated within Standard OIL. OKBC clearly Agents are frequently coded as specialists, understanding a distinguishes between schema and instances in its own focused subset of the domain itself. For example, an agent that is vocabulary. Therefore, there is no real barrier to this factoring interfacing with information repositories on environmental approach. remediation techniques would not necessarily understand the information related to companies and their responsibilities for 4. COMPOSITIONAL ONTOLOGIES cleaning up specific toxic waste sites; yet the scope of the domain Given that the best approach to ontology design and use is a encompasses both areas. These agents need not incorporate entire collaborative one among ontology designers, application ontologies; in fact, this may not even be possible for lightweight developers and users, let us now turn our attention to the issue of agents using large ontologies. Standards such as FIPA [5] allow reuse. Consider an application developer, developing an agents to advertise the ontologies they understand, However, application to fulfill some of the needs of a specific class of users such an advertisement may be misleading to other agents if the in a given application domain. The developer has decided to agent only understands a subset of the ontology, in that two agents advertising the same ontology may in fact understand or geographic identifiers made it necessary to compose the orthogonal subsets of the ontology, rendering communication domain ontology incorporating terminology from the field of impossible. hazardous waste pollution and remediation with the EDR Furthermore, too much ontology information may also confuse ontology for value representation. users. As a practical matter, users of several of our agent The ability to compose ontologies has an immediate benefit in applications including EDEN were confused by the presence of modular ontology design. For instance, if you look within the ontological terms for which no agent had advertised any EDR tags, there are tags relating to location (street address, knowledge. The presence of the term implied to them that they postal code) as well as geographic (latitude and longitude). These should be able to extract information (other than definition and tags are not specific to the environmental domain, but are shared ontological relationships) relevant to the term. That no agent had among other domains such as address books and GPS positioning advertised the term was an unsatisfying explanation. Because of systems. Furthermore, such structures seem relatively stable. this, the agents themselves that implement the ontology must be This indicates that it may be much more useful to have, say, a careful how they present themselves to their users. single ontology for geographic positions, their representations, We define this subsetting using ontological fragments. An and the relationships between them that can be incorporated into ontological fragment consists of: environmental applications as well as those in other domains. • The name of the ontology 4.3 Extend • The classes or entities in the ontology that are supported Unfortunately, applications frequently need to incorporate within the application, and their superclasses. concepts that are not represented in any ontology, or they may need to “glue” concepts from different ontologies together using • Constraints on those classes including such things as ranges new concepts. Thus, it is not sufficient to compose subsets of of values allowed for specific slots ontologies; frequently it is necessary also to add some new • The axioms that reference the supported classes/ attributes. concepts. For example, within the setting of our EDEN application, the domain ontologies covered the concepts for The fragment can be expressed as a set of constraints over the which users wanted to retrieve information; for instance, they classes and attributes. As an example from the EDEN system, enabled the user to answer questions such as, “What toxic waste each agent that provided a wrapper for a data resource advertised sites are within 10 miles of Houston, TX?” However, the agent- only a fraction of the domain ontology – precisely that portion for based system that supported the application also was able to ask which it could provide data instances. In some cases, a wrapper higher-level, more abstract questions such as “Are the number of agent advertised only a fraction of the slots of a class, and further toxic waste sites within 10 miles of Houston increasing faster placed semantic constraints on the content of the slots based on than they can be remediated?” These higher-level questions used knowledge of the data instances about which it could report. In concepts present neither in the domain ontology, nor in any other other cases, an agent might fill a slot with a static value from the ontology available at the time. canonical data representation for that slot. For example, because there was no record in a particular remediation database for land The important issue with extension is to do it in a manner that use, one of the attributes of a site, certain government database will not create further problems later. There are two ways to wrappers returned the generic value “Federal Facility” which extend an ontology; extensions that are safe and can easily be was one of the values used in other databases. incorporated into other applications, and extensions that are unsafe and should eventually be agreed upon and then folded Another example decision that needs to be made on any system is back by consensus into the ontology. We define safe and unsafe whether a request for everything matching a query should return extensions with respect to the permissible definitional changes an empty set if an agent does not advertise all elements of a with respect to the underlying ontology(-ies). A safe extension class, or should return nulls for the unadvertised elements. This can be characterized as follows: provoked strong disagreement on our team between those who felt that the former was algebraically preferable and those who Safe extension: an extension incorporating existing concepts felt that the latter was more intuitive to users already intimidated from one or more ontologies, where any new concepts are defined by the challenge of understanding the nature of the results axiomatically against the existing concepts in such a way that produced by a dynamic distributed information system. instances of the new concepts can be determined computationally from instances of the existing concepts. The transformation 4.2 Compose axioms must be invertible; the conversion of any instance Another issue that is brought about by ontological commitment is between its view in the existing concept(s) and its view in the the issue that an agent often requires access to terms from new concept must be lossless in both directions. multiple ontologies. There exist a growing number of useful and Safe extensions are “safe” primarily because they can be shared public ontologies, and it often seems more appropriate to pick among the agents easily, simply by sharing the axioms that allow and choose terms from those ontologies than to build a new one for conversion from the existing classes to the extended classes. from scratch, with just the terms that you need for your ontology. We note that new concepts within such a safe extension cannot The result is a virtual ontology that imports fragments of other have new attributes that would need to be present in any instance ontologies, into some sort of unified and useful whole. For of that concept, as this would violate the lossless conversion example, within the EDEN application the diverse ways of property. So, for example, “milliliters per liter” would be a safe representing values relating to various concepts such as chemical extension for a concept if the concepts “water pollutant” and “parts per million” were already present in the ontology being Figure 3 shows a fragment of an example compositional extended. environmental ontology closely related to the one we used in the In contrast, unsafe extensions are “unsafe” because they are EDEN application. This ontology used subsets of the GPS and harder to share with other agents, as the other agents may not Measurement ontologies, composed together with a fragment of maintain the information necessary to populate instances of the the environmental ontology. The necessary concepts for this unsafe classes. For example, adding a new concept, “Water particular application span all three ontologies. The gray boxes pollutant” with a new property, “concentration” would not be a indicate the three ontologies, and the white boxes within indicate safe extension if there were no concept of containing medium in the classes taken from the individual ontology into the the original ontology. Other agents, committed to the ontology, compositional ontology. Note that the attributes from the may not have the information necessary to provide a “Sampling Point” class refer to concepts in the other ontologies. concentration for each pollutant they have, or may have different 5. IMPLEMENTATION ISSUES ways of modeling the concept of how badly the pollutant is Given that we allow the operations of subsetting, composition contaminating the water. and (safe) extension over ontologies within an agent system, the Because unsafe extensions allow agents to instantiate classes that agents themselves need to be able to represent exactly what have been extended unsafely and thus cannot easily be shared; combination they are committed to, in order to ensure that they these extensions limit the openness of the agent. However, do not mislead others with respect to their capabilities. unsafe extensions are sometimes necessary; for instance, they are There are three issues with representation and implementation. often required to deal with some of the impedance mismatch The first is that the agent needs to be able to represent internally issues discussed earlier. Allowing unsafe extensions increases the full set of details on how the ontology fragments it is using fit the possibility that different agents will create divergent together. This can be done with a suitable abstract ontology extensions to the ontology, and become incompatible with one representation. Secondly, the agent must be able to converse in another. Therefore, unsafe extensions should be treated with some detail with other agents about the parts of the ontologies caution. One approach is to limit such extensions to those that that they understand. This must be done using a suitable are monotonic [16] -- not requiring modifications to existing exchange representation. Thirdly, messages should be able to ontological elements. Idealistically, the need for specific unsafe represent the ontology fragments that the message covers. This extensions should be propagated to the ontology designers, who requires a more specific ontology field within the agent can determine whether or not there is a consensus on the need for communication language. the particular extension, and whether they should be incorporated into the next version. 5.1 Internal Representation Internally, an agent needs to understand in detail the specific 4.4 Example schema-level knowledge to which it is committed. This GPS Ontology understanding may be complicated by the fact that ontological Locator knowledge may be represented using many different isa representational models; for instance, there are subject-verb- canonical unit object models such as DAML, frame-oriented models such and Lat/Long GeoLocation OKBC, and object-oriented models such as UML [14, 1]. These unit all differ in key issues such as whether or not attributes are first- class objects, whether or not multiple inheritance is allowed, and EDEN Ontology location how strongly-typed classes must be. Further complicating this, if the operations of subsetting, composition and extension are used, Measured atPoint Sampling a single application may look into a set of ontologies whose Contamination Point underlying ontological models may diverge. depth Recent discussions have included the notion of developing abstract ontology representations for the internal representation Measurement of ontological knowledge. As with all applications that attempt to Distance span multiple viewpoints, key choices in specifying an abstract Ontology isa ontology representation revolve around whether it should canonical unit incorporate every modeling features present in some ontology Meter Quantity representation, or it should only look at the modeling features unit used by every representation, or it should take some path in the isa unit middle. Wilmott et.al. [15] discuss in depth some of the issues Unit of associated with the specification and implementation of an isa abstract ontology representation from a theoretical standpoint. Measure Foot Fortunately, an abstract ontology representation is implemented within an ontology service, and thus only needs to incorporate the Figure 3. A part of the EDEN compositional ontology. requirements of the agents that it serves. In the InfoSleuth system, our ontology service focused on frame-based ontologies, as they were most compatible with the databases with which we defined is a potentially huge superset of what is actually needed were working. by the agent. Furthermore, it may not be possible to relate the Internally, an agent that is using ontologies need only store the terms in the new ontology back to the ontologies that it imported, subsets of the ontology that it needs, and the glue that puts them so it may be computationally difficult to re-relate terms imported together. So, for instance, in the ontology of Figure 2, the agent from the same ontology into two such composed ontologies. itself only needs to understand the concepts in the shaded areas DAML+OIL, on the other hand, uses namespaces to facilitate the of the GEMET, GPS Positioning, Facility and EDR ontologies, as combination of ontologies and the relating of concepts among well as the additional information associated with their multiple ontologies. However, this ability to use namespaces composition and extension. Managing this information in an relies on having the component ontologies’ contents also within a efficient way is the work of the ontology service. namespace. Our experience has shown us that a good abstract ontology Extension: Safe extensions by definition should be able to be representation and a good ontology service are essential to cope specified in terms of logical axioms over existing concepts in the with the incorporation of existing ontologies and their ontology. While axioms are strongly considered among ontology evolutionary steps. The abstract ontology representation designers, many ontology exchange languages hardly consider engenders a unified methodology by which the agent can absorb them at all; for instance, DAML+OIL at the moment has no good existing ontology information and its particular use within the foundation for representing axioms and is thus unsuitable for application. Agent implementers then use the ontologies in a exchanging knowledge concerning safe extensions. more declarative and less hard-wired manner, which in turn In EDEN, we prepared a number of ad hoc lexical translation facilitates the incorporation of new ontological information. processes to translate between an old version and an extended version of the ontology. This did not yield a reusable method. 5.2 External Representation Agents need to be able to share which parts of the different 5.3 Representation in ACL Messages ontologies that they use or require, and how they put them The last area of application impact concerns the communication together, with their other collaborating agents. This happens of ontological reference information in ACL messages. Currently, when the agent advertises its capabilities to another agent, when agent communication languages such as the FIPA ACL [5] carry a an agent is looking to locate another agent that can help it with specification of the ontology from which the message content some task, or when agents are negotiating over who is takes its concepts in an “:ontology” property. Compositional responsible for what subtasks. We have identified three basic ontologies do not map directly to a single ontology, so the ways that ontologies may be adapted dynamically: subsetting, message format needs to change in one of the following ways: composition, and extension: 1. Give the compositional ontology a name and make it Subsetting: The notion of subsetting an ontology is not reflected explicitly available, and use this in the message’s :ontology in the definition of the ontology itself, but rather in artifacts property. related to the use of the ontology. For example, fragments may be 2. Specify all ontologies using multiple :ontology properties. advertised from one agent to another, so that agents can inform This may not cover concepts in safe extensions to the other agents within their community of their exact capabilities. ontology. Alternatively, an application may be supported at the user end by an agent that interfaces with the user, ensuring that the user is 3. Allow the ontology field to contain a description of the presented only with the fragments of the ontology that are concepts in the in the ontologies, including extensions, that supported within the underlying agent community. Any user are used in the message. This could result in unnecessary interfaces can then be tailored to the ontology fragment. performance impact on communicative actions. Composition: As with subsetting, composition should be 6. CONCLUSIONS specified at the time of use, however, there are some Ontological commitment is the decision by a particular group of representational difficulties at this point. While subsetting can be applications or users within an application domain to use the defined as a set of axioms that overlays a single ontology, terms defined in a given ontology. High ontological commitment composition spans multiple ontologies and thus presents different occurs when many users and/or applications within an representational challenges. For instance, a composition may application domain commit to sharing the same vocabulary of wish to state that the values of the “LatitudeMeasure” slot for the concepts, meanings and relationships defined within a specific “FacilityIdentification” in the EDR may be understood by the ontology. Ontological commitment to a common set of ontologies agent as represented in the “LatLongCoordinate” units defined in is a key feature to provide for openness in an agent system, as it the “GPS Coordinates” ontology. In EDEN, we defined a provides a common vocabulary over which agents can converse. geographic location as comprising a (selectable) coordinate This paper focused mainly on two areas relating to ontological scheme and a coordinate value. commitment and use. The first area is the clash of goals between OKBC and DAML present different challenges with respect to ontology definers, application developers and users. Ontology composition. Within OKBC, composition must be done by definers are concerned with the completeness and purity of their explicitly importing the component ontologies. Existing tools do ontological design. Application developers are often concerned not necessarily facilitate importing ontologies from remote only with specific subsets of the ontology that relate directly to locations, or ones that are represented in different languages the application, and with application-specific representational (e.g., DAML). The resulting new ontology, even if it can be and computational issues. Users are concerned mainly with [4] “EPA Environmental Data Registry”. http://www.epa.gov/edr, sticking with known and familiar terminology and vocabularies, 2002. which may not be logically structured and therefore may not be [5] FIPA. “Agent Communication Language Specifications”. amenable to being reformulated into computationally-accessible http://www.fipa.org/repository/aclspecs.html, 2001. concepts. This clash is complicated by the widening gap between these groups – developers, for instance, may be using ontologies [6] “General Multilingual Environmental Thesaurus”. that are already well-established. Thus, while several good http://www.nu.niedersachsen.de/cds/etc- design paradigms involving ontology designers, application cds_neu/library/select.html developers, and users, exist for initial ontology development [7] Humphreys, B.L. et.al. “Assessing and Enhancing the Value [16,17,18], none of these seem to encompass these more long- of the UMLS Knowledge Sources”. In Proceedings of the 15th term concerns. We recommend the development of an extended Annual Symposium on Computer Applications in Medical Care, ontology lifecycle that fosters evolution. Components required to pp. 78-82, Washington, D.C., November, 1991. support this lifecycle should include long-term ontology design [8] Nodine, M. et.al. “Active Information Gathering in support, a widely-available feedback channel from developers InfoSleuth.” International Journal of Cooperative Information and users back to designers, and an open forum for discussing issues and extensions to the ontology. Systems 91/2, 2000, pp.3-28. An application committed to reuse existing ontologies may [9] “Welcome to the OIL Page”. encounter several difficulties, as search for the perfect ontology http://www.ontoknowledge.org/oil, 2002. for the application – one that is acceptable to both the application [10] “OKBC Home Page”. and its users – is not necessarily fruitful. We described a http://www.ksl.stanford.edu/software/OKBC, 2002. compositional approach to ontological use. Compositional [11] Sowa, J. F. Knowledge Representation. Brooks/Cole, ontologies base themselves in existing ontologies as much as California, 2000, pp. xi-xii. possible, using the operations of ontological subsetting, ontological composition and ontological extension to tailor them [12] “EPA Terminology Reference System.” to the specific needs of the applications. Compositional http://www.epa.gov/trs/index.htm , 2002. ontologies fit well with the needs of the application, and the [13] Tuttle, M. S. et al., “Merging terminologies”. Medinfo. approach has the potential to raise the level of ontological 1995;8 Pt 1:162-6. commitment to existing ontologies. However, they require more [14] Object Management Group. OMG Unified Modeling sophisticated ontology-related exchanges among collaborating Language Specification, version 1.3. agents. In order to incorporate these compositional ontologies, http://www.omg.org/technology/documents/formal/unified_model further work needs to be done on the development of a more ing_language.htm. 2000. formal algebra to support the subset, compose and extend operations over ontologies (in contrast to our current ad-hoc [15] Willmott, S., Constantinescu, I. and Callisti, M. methods). Visser and Cui [19] attacked a related problem of “Multilingual Agents: Ontologies, Languages and Abstractions.” heterogeneous ontology structures, and this may provide further In Proceedings of the First International Workshop on insight into this formal algebra. As a second task, we need to Ontologies in Agent Systems, Autonomous Agents 2001, develop a methodology for ensuring the correctness and Montreal, Canada, May, 2001. consistency of these compositional ontologies. [16] Gruber, T.R. “A Translation Approach to Portable Ontology Many approaches to the representation and specification of Specifications.” Knowledge Acquisition 5(2), 1993. ontologies exist; some common ones for the exchange of [17] Gomez-Perez, A., “Ontological Engineering: A State of the ontological information include OKBC, DAML+OIL, OIL and its Art.” Expert Update, 1999. varieties, and UML. Each of these has features amenable to their adoption as a means to exchange information on ontological [18] Gruninger, M and Fox, M. S. “Methodology for the Design components and extensions, though none cover all of the issues and Evaluation of Ontologies.” In Proceedings of the Workshop addressed in this paper. Therefore, the exchange of such on Basic Ontological Issues and Knowledge Sharing, 1995. information using these standards is still problematic. [19] Hammer, J. and McLeod, D. “An Approach to Resolving Semantic Heterogeneity in a Federation of Autonomous, 7. REFERENCES Heterogeneous Database Systems.” In International Journal of [1] Cranefield, S., Haustein, S and Purvis, M. “UML-based Intelligent and Cooperative Information Systems 2(1), 1993. ontology modeling for software agents”. In Proceedings of the [20] Kent, W. “The Many Forms of a Single Fact”. In workshop on Ontologies in Agent Systems, Autonomous Agents Proceedings of IEEE COMPCON, 1998. 2001, Montreal, Canada, 2001. [2] DAML.org. http://www.daml.org, 2002. [3] Deschaine, L. M., Brice, R. S. and Nodine, M. H. “Use of InfoSleuth to Coordinate Information Acquisition, Tracking and Analysis in Complex Applications.” In Proceedings of Advanced Simulation Technologies Conference, April, 2000.