Consensus ontologies. Reconciling the semantics of Web pages and agents

Consensus ontologies. Reconciling the semantics of Web pages and agents
of 5
All materials on our website are shared by users. If you have any questions about copyright issues, please report us to resolve them. We are always happy to assist you.
  University of South Carolina Scholar Commons Faculty PublicationsComputer Science and Engineering, Department of 1-1-2001 Consensus Ontologies: Reconciling the Semanticsof Web Pages and Agents Larry M. Stevens University of South Carolina - Columbia  , Michael N. Huhns University of South Carolina - Columbia  , Follow this and additional works at:hp:// of theComputer Engineering Commons is Article is brought to you for free and open access by the Computer Science and Engineering, Department of at Scholar Commons. It has beenaccepted for inclusion in Faculty Publications by an authorized administrator of Scholar Commons. For more information, please Publication Info Published in  IEEE Internet Computing   , Volume 5, Issue 5, 2001, pages 92-95.hp://© 2001 by the Institute of Electrical and Electronics Engineers (IEEE)   Agents on the Web ConsensusOntologies Reconciling the Semantics of Web Pages and Agents 92 SEPTEMBER • OCTOBER2001 1089-7801/01/$10.00 © 2001 IEEEIEEEINTERNETCOMPUTING Larry M.Stephens • University of South Carolina • Michael N.Huhns • University of South Carolina • I n an old joke, a drunk is on his hands andknees searching for his keys underneath alamppost. “Is this where you dropped them?”he is asked. “No, I dropped them over there, butthe light is better here.” As you build a Web site, it is worthwhile to ask a similar question: “Should you put your informa-tion where it belongs or where people are most like-ly to look for it?” Our recent research to improvesearch through ontologies is providing some inter-esting results for answering this question. Reconciling Web Semantics  Web searches typically yield pointers to a largenumber of Web sites — only some of which are rel-evant. Search engines might rank the sites, but theresults are otherwise unorganized and too numer-ous for users to investigate manually. Many solu-tions have been proposed to this familiar problem,including constructing more intelligent searchengines, requiring users to specify more precisesearch criteria, or requiring Web sites to describetheir contents more precisely.These approaches all use ontologies to describeboth requirements and sources. Unfortunately, thecomprehensive ontology that could solve the prob-lem of information retrieval does not yet exist.Moreover, the Web’s dynamic and eclectic naturemakes it unlikely that everyone would adhere tosuch an ontology if it did.To overcome these limitations, Web developerscould choose among three possible approaches toassociate, organize, and merge information seman-tically from large numbers of independently devel-oped sources: ■  All Web sites could use the same terminology with agreed-upon semantics — a method con-sidered improbable. ■ Each Web site could use its own terminology and provide translations to a global ontology —a method considered difficult, and thus unlikely. ■ Each Web site could use small, local ontologiesthat can be related indirectly with the assis-tance of agents — a method we describe here.Our methodology is consistent with the envis-aged semantic Web, 1 which presumes that Websources will be annotated with ontological infor-mation. 2,3  We also presume that the independent-ly developed sources and ontologies returned froma Web search are for similar domains — therewould be no interesting relationships among themotherwise — but that they will undoubtedly havedissimilar formulations and terminologies.Our hypothesis is that a multiplicity of ontol-ogy fragments, representing the semantics of theindependent sources, can be related to each other automatically without using a global ontology.That is, even when there is no way to determinea direct relationship between a pair of ontologies,they can be related indirectly through a semantic bridge  consisting of other previously unrelatedontologies. Rather than scale causing a problem,additional ontologies can make it easier — or evenpossible — to relate two ontologies. The resultantmerged ontologies provide a semantic character-ization of the set of sources and their domains,and effectively create a single large ontology toserve as a global hub for interactions. Thismethodology establishes a means for agents and  other information system componentsto interoperate. Reconciling SeparatelyDeveloped Ontologies In agent-assisted information retrieval,a user will describe a need to an agent,which will use terms from the user’slocal ontology to translate the descrip-tion into a set of requests. The agentwill contact online brokers and requesthelp in locating sources that can sat-isfy the requests. The agents must rec-oncile their semantics to communicateabout the request, which seems impos-sible if their ontologies share no con-cepts. If they share concepts with athird ontology, however, that onemight provide a semantic bridge torelate all three. Note that the agentsneed to relate only the portions of their ontologies that are necessary for responding to the request. The difficulty in establishing abridge will depend on the semanticdistance between the concepts, and onthe number of ontologies that consti-tute the bridge. The methodology weare investigating is appropriate withlarge numbers of small ontologies —the situation we expect to find in com-plex information environments. A small ontology is like one piece in a jigsaw puzzle: It is difficult to relatetwo random puzzle pieces until they are constrained by others. We expectthe same to be true for ontologies.Two concepts can have seven mutu-ally exclusive relationships betweenthem: subclass , superclass , equivalence  ,  partOf  , hasPart  , sibling , or other  . If arequest contains three concepts, for example, and the request must be relat-ed to an ontology containing 10 con-cepts, then there are 7  × 3  × 10=210possible relationships among them.Only 30 of these will be correct becauseeach of the three concepts in therequest will have exactly one relation-ship with each of the 10 concepts in thesource’s ontology. The correct relation-ships will be determined automatically by applying constraints among theconcepts within each ontology as wellas constraints discoveredamong multiple ontologies.The relationships of major interest are equivalence and sibling . Where those do notexist, we are interested in themost specific superclass or most specific  partOf  . Consider the example inFigure 1a. The ontology frag-ment on the left would berepresented as  partOf(Wheel,Truck), and the one on theright would be  partOf(Tire, APC). There are no obviousrelationships between thesetwo fragments. The concept Truck could be related to  APC  by equivalence  ,  partOf  , has-Part  , subclass , superclass , or  other  , and there is no way to decidewhich is correct. Now consider theaddition of the middle ontology frag-ment  partOf(Wheel, APC) in Figure 1b .  With this added information, there isevidence that we could link the con-cepts Truck and  APC  as equivalent aswell as the concepts Wheel  and Tire  .This example exploits the  partOf  rela-tion, which is common to all threeontologies. Other domain-independentrelations, such as subclassOf  , instance-Of  , and subrelationOf  , will be necessary for the reconciliation process. Moreover,the following six properties of relationscan help in relating occurrences of therelations to each other: reflexivity, sym-metry, asymmetry, transitivity, irreflex-ivity, and antisymmetry. 4 Domain con-cepts and relations can be related toeach other by converse/inverse, compo-sition, (exhaustive) partition, part-whole(with six subtypes), and temporal atti-tude. All local ontologies and informa-tion system components must under-stand and use some minimum set of these fundamental relations.In attempting to relate two ontolo-gies, a system might not be able tofind correspondences between con-cepts because not enough constraintsand similarities exist among terms. 5 Trying to locate correspondences withother ontologies, however, might yieldenough constraints to relate the srci-nal pair. As more ontologies are relat-ed, additional constraints arise amongthe terms of any pair of ontologies. Inthis way, the presence of many smallontologies becomes an advantage. It isalso a disadvantage in that some con-straints might conflict, but we use thepreponderance of evidence to resolvethese statistically. Our Experiments  We asked 55 graduate students in com-puter science and engineering to con-struct small ontologies in DAML(DARPA Agent Markup Language, the givendomain of People. Figure 2 on the nextpage shows a typical example of one of these ontologies.The 55 component ontologiesdescribed 864 classes. Using a string-matching algorithm and other heuris-tics, we constructed one merged ontol-ogy from these (shown in Figure 3)that contained 281 classes in a singlegraph with the root node #Thing. Thisgraph related all of the concepts fromthe ontologies with no orphans — thatis, there was some relationship (path)between every pair of concepts.  We constructed a consensus ontol-ogy  during the merge operation by counting the number of times classes IEEEINTERNETCOMPUTING SEPTEMBER • OCTOBER2001 93 Consensus Ontologies Truck WheelAPCTireTruck WheelAPCPossibly equivalentequivalencepartOf equivalenceTire (a)(b) Figure 1.Ontology relationships.Two ontology fragments with no obvious relationships (a) canbe related by introducing a third ontology toreveal equivalences between components of thesrcinal two fragments (b).  and subclass links appeared in thecomponent ontologies. The class Per-son, and all similar classes such as Per-sons and Personnel whose namesmatched using our simple string-matching algorithm, appeared 14times, for example. The subclass link from Mammals (and its matches) toHumans (and its matches) appeared 9times. We termed these values the rein- forcement  of a concept.Redundant subclass links wereremoved and the corresponding transi-tive closure links were reinforced. Thatis, if C had subclass A with reinforce-ment 2, C had subclass B, and B hadsubclass A, then the link from C direct-ly to A was removed and the remaininglink reinforcements from C to B and Bto A were each increased by 2. We thenremoved any classes or links that werenot reinforced by appearing multipletimes in the merged ontology. The resultrepresents an implicit consensus amongthe ontology writers about which con-cepts should appear in the domain andhow they should be related.Finally, we applied an equivalence heuristic  for collapsing classes withcommon reinforced superclasses andsubclasses. The merged ontology con-tains both Human and Person, for example. The equivalence heuristicfound that all reinforced subclasses of Person are also reinforced subclasses of Human, and all reinforced superclassesof person are also reinforced super-classes of Humans. It thus deemed thatHuman and Person were the same con-cept. This heuristic is similar to aninexact graph matching technique.Figure 4 shows the collapsed consen-sus ontology, now containing 36 class-es related by 62 subclass links. Discussion of Results In analyzing the 55 ontologies, wenoted immediately that each studenthad a different way of describing andorganizing the domain — even for adomain as familiar and simple as peo-ple. It was also apparent that thedescriptions were inaccurate and con-tradictory. Mammals, for example,were described as both a subclass anda superclass of animals. A consensus ontology is perhaps themost useful for information retrievalby humans because it represents theway most people view the world and 94 SEPTEMBER • OCTOBER2001 IEEEINTERNETCOMPUTING  Agents on the Web Figure 3.Merged ontology from 55 independently constructed ontologies for thePeople domain.Of 281 classes,38 with 71 subclass links appear more than once.Figure 2.Typical ontology.Students used DAML to create small ontologies likethis one to characterize a Web site about People.All links denote subclasses.  its information. If most people wrong-ly believe that crocodiles are a kind of mammal, for example, then most userswould find it easier to locate informa-tion about crocodiles located in amammals grouping, rather than in rep-tiles where it factually belongs.The information retrieval measures of   precision and recall  are based on somedegree of match between a request and aresponse. The length of a semantic bridgebetween two concepts can provide analternative measure of conceptual dis-tance and an improved notion of infor-mation relevance. 3 Previous measuresrelied on the number of properties sharedby, or the number of links separating,two concepts within the same ontology.These measures not only require a com-mon ontology, but also fail to account for the density or paucity of informationabout a concept. Our suggested measuredoes not require a common ontology andis sensitive to the amount of informationavailable in the domain. 6 Conclusion Imagine again that in response to arequest for information, a user receivespointers to more than 1,000 docu-ments. The techniques developed by our research would bring organizationto the information received and wouldreconcile the semantics of each docu-ment. Our goal is to help users retrievedynamically generated informationthat is tailored to their individualneeds and preferences. We believe that it is easier for indi- viduals or small groups to develop their own ontologies, regardless of whether global ones are available, and thatthese can be automatically and ex postfacto related. We are working to deter-mine the efficacy of local annotationfor Web sources, as well as performingreconciliation that is qualified by mea-sures of semantic distance. If success-ful, this research will enable softwareagents to resolve the semantic miscon-ceptions that inhibit successful inter-operation with other agents and thatlimit the effectiveness of searching dis-tributed information sources. Acknowledgment This material is based upon work supported by the U.S. National Science Foundation under grant no. IIS-0083362. References 1.T. Berners-Lee, J. Hendler, and O. Lassila,“The Semantic Web,” Scientific American ,May 2001; available at Heflin and J. Hendler, “Dynamic Ontolo-gies on the Web,” Proc. 17th Nat’l Conf. Artificial Intelligence (AAAI 2000) , AAAIPress, Menlo Park, Calif., 2000, pp.443-449.3.J.M. Pierre, “Practical Issues for AutomatedCategorization of Web Sites,” Electronic Proc. ECDL 2000 Workshop on Semantic Web , Sept. 2000; available at Stephens and Y.F. Chen, “Principles for Organizing Semantic Relations in LargeKnowledge Bases,” IEEE Trans. Knowledge and Data Engineering , vol. 8, no. 3, June1996, pp. 492-496.5.G. Wiederhold, “An Algebra for Ontology Composition,” Proc. 1994 Monterey Work-shop on Formal Methods , U.S. Naval Post-graduate School, 1994, pp. 56-62.6.H.S. Delugach, “An Exploration into Seman-tic Distance,” Lecture Notes in Artificial Intelligence  , no. 754, Springer-Verlag, Berlin,1993, pp. 119-124. Michael N. Huhns is a professor of computer science and engineering at the University of South Carolina, where he also directs theCenter for Information Technology. Larry M. Stephens is a professor of computer science and engineering at the University of South Carolina, where his primary researchinterests are agents and ontologies. IEEEINTERNETCOMPUTING SEPTEMBER • OCTOBER2001 95 Consensus Ontologies Figure 4.Consensus ontology.Weakly reinforced concepts were removed and concepts with common subclasses and superclasses were merged to producethis graph containing 36 classes related by 62 subclass links.
Related Search
We Need Your Support
Thank you for visiting our website and your interest in our free products and services. We are nonprofit website to share and download documents. To the running of this website, we need your help to support us.

Thanks to everyone for your continued support.

No, Thanks

We need your sign to support Project to invent "SMART AND CONTROLLABLE REFLECTIVE BALLOONS" to cover the Sun and Save Our Earth.

More details...

Sign Now!

We are very appreciated for your Prompt Action!