From: FLAIRS-01 Proceedings. Copyright © 2001, AAAI (www.aaai.org). All rights reserved. Building a Large KnowledgeBase Semi-Automatically Udo Hahn Stefan Schulz [-~ Text KnowledgeEngineering Lab Freiburg University Werthmannplatz 1 D-79085 Freiburg, Germany hahn@coling, uni-freiburg,de Medical Informatics Department Freiburg University Hospital Stefan-Meier-Str. 26 D-79104 Freiburg, Germany s tschulz@uni-freiburg, de Abstract Wedescribe a knowledgeengineeringapproachby which conceptualknowledge is extracted froman informal, semantically weakmedicalthesaurus (UMLS) and automatically convertedinto a formallysounddescriptionlogics system. Our approachconsists of four steps: concept definitions are automatically generated fromthe UMLS source, integrity checkingof taxonomicand partonomic hierarchiesis performed by the terminologicalclassifier, cyclesand inconsistenciesare eliminated,andincremental refinementof the evolvingknowledge base is performed by a domainexpert. Wereport on experimentswith a terminological knowledgebase composedof 164,000 concepts and76,000relations. Introduction Over several decades, an enormousbody of medical knowledge, e.g, disease taxonomies, medical procedures, anatomical terms etc., has been assembled in a wide variety of medical terminologies, thesauri and classification systems. The conceptual structuring of a domainthey allow is typically restricted to the provision of broader/narrowerterms, related terms or (quasi-)synonymous terms. This is most evident in the UMLS,the Unified Medical LanguageSystem (McCray & Nelson 1995), an umbrella system which covers more than 50 medical thesauri and classifications. Its metathesaurus componentcontains more than 600,000 concepts which are structured in hierarchies by 134 semantic types and 54 relations between semantic types. Their semanticsis shallow and intuitive, whichis due to the fact that their usage is primarily intended for humansengaged in various forms of clinical knowledgemanagement. Givenits size, evolutionary diversity and inherent heterogeneity, there is no surprise that the lack of a formal foundationleads to inconsistencies, circular definitions, etc. (Cimino1998). This maynot cause utterly severe problems whenhumansare in the loop and its use is limited to disease encoding, accountancy or documentretrieval tasks. However, anticipating its use for moreknowledge-intensiveapplications such as natural language understanding of medical narratives (Hahn, Romacker,&Schulz 1999) or medical °Copyright (~) 200I, American Association for Artificial Intelligence(www.aaai.org). All rights reserved. 222 FLAIRS-2001 decision support systems (Reggia & Tuhrim1985), those shortcomingsmight lead to an impasse. As a consequence, formal models for dealing with medical knowledgehave been proposed, using representation mechanismsbased on conceptual graphs, semantic networksor description logics (Volot et al. 1994; Mayset al. 1996; Rector et al. 1997). Not surprisingly, however,there is a price to be paid for more expressiveness and formal rigor, viz, increasing modelingefforts and, hence, increasing maintenance costs. Therefore, concrete systems making full use of this rigid approach, especially those which employ high-end knowledgerepresentation languages are usually restricted to rather small subdomains. The knowledgebases developed within the frameworkof the above-mentionedterminological systems have all been designed from scratch - without makingsystematic use of the large bodyof knowledgecontained in those medical terminologies. An intriguing approach would be to join the massive coverage offered by informal medical terminologies with the high level of expressivenesssupported by formal inferencing systems, as developed in the AI knowledge representation community,in order to develop formally solid medical knowledgebases on a larger scale. This idea has already been fostered by Pisanelli, Gangemi,&Steve (1998) who extracted knowledge from the UMLSsemantic networkas well as from parts of the metathesaurusand merged them with logic-based top-level ontologies from various sources. In a similar way, Spackman& Campbell (1998) describe how SNOMED (C6t6 1993) evolves a multi-axial coding systeminto a formally foundedontology. Unfortunately, the efforts madeso far are entirely focused on generalization-based reasoning along is-a hierarchies and lack a reasonable coverage of partonomies. Part-Whole Reasoning As far as medical knowledge is concerned, two main hierarchy-buildingrelationships can be identified, viz. is-a (taxonomic) and part-whole (partonomic) relations. Unlike generalization-based reasoning in concept taxonomies, no fully conclusive mechanismexists up to nowfor reasoning along part-whole hierarchies in description logic systems. For medical domains, however,the exclusion of part-whole reasoning is far from adequate. Anatomical knowledge, a central portion of medical knowledge,is principally organized along part-whole hierarchies. Hence, any proper Figure 1: onomies SEPTriplets: Partitive Relations within Tax- medical knowledgerepresentation has to take account of both hierarchy types (Haimowitz,Patil, &Szolovits 1988). Various approaches to the reconstruction of partwhole reasoning within the object-centered representation paradigm are discussed by Artale et al. (1996). In the description logics communityseveral language extensions have been proposedbased on special constructors for partwhole reasoning (Rector et al. 1997; Horrocks & Sattler 1999), though at the cost of increasing computationalcomplexity. Motivated by informal approaches sketched by Schmolze & Mark (1991) we formalized a model of partwhole reasoning (Hahn, Schulz, & Romacker1999) that does not exceed the expressiveness of the well-understood, parsimonious concept language .AEC (Schmidt-Schaul3 Smolka 1991)) Our proposal is centered arounda particular data structure, so-called SEPtriplets, especially designed for partwholereasoning(cf. Figure 1). Theydefine a characteristic pattern of IS-A hierarchies which support the emulation of inferences typical of transitive PART-OF relations. In this formalism, the relation ANATOMICAL-PART-OF describes the partitive relation betweenphysical parts of an organism. A triplet consists, first of all, of a composite’structure’ concept, the so-called S-node (e.g., HAND-STRUCTURE). Each ’structure’ concept subsumesboth an anatomical entity and each of the anatomicalparts of this entity. Unlike entities and their parts, ’structures’ have no physical correlate in the real world-- they constitute a representational artifact required for the formal reconstruction of systematic patterns of part-whole reasoning. The two direct subsumees of an S-nodeare the correspondingE-node(’entity’) and node (’part’), e.g., HAND and HAND-PART, respectively. Unlike an S-node, these nodes refer to specific ontological objects. The E-node denotes the whole anatomical entity to be modeled, whereas the P-node is the commonsubsumer of any of the parts of the E-node. Hence, for every P-nodethere exists a correspondingE-nodefor the role ANATOMICAL-PART-OF. Some basic anatomical relations in terms of SEPtriplets are illustrated in Figure 2. I.A£Callowsfor the constructionof hierarchies of concepts and relations, whereE denotessubsumption and -- definitional equivalence. Existential(B) anduniversal(V)quantification,negation (-~), disjunction(I-1) andconjunction (t_l) are supported. filler constraints(e.g., typingby(7) are linkedto the relationname R by a dot, BR.(7. Figure 2: Partonomic Hierarchy of the Concept HAND The reconstruction of the relation ANATOMICAL-PARTOF by taxonomic reasoning proceeds as follows. Let us assume that CE and DE denote E-nodes, Cs and Ds denote the S-nodes that subsumeCEand DE, respectively, and Cp and Dp denote the P-nodes related to CEand DE, respectively, via the role ANATOMICAL-PART-OF (cf. Figure 1). Theseconventions can be captured by the following terminological expressions: CE E_ Us E Dp E_ Ds (1) DE E_ DS (2) The P-node is defined as follows (note the disjointness between D~ and Dp): Dp - Ds n -~DE r7 Banatomical-part-of .DE (3) Since CEis subsumedby Dp(1) we infer that the relation ANATOMICAL-PART-OF holds between CE and DE : C E E 3anatomical-part-of.De (4) Knowledge Import and Refinement Our goal is to extract conceptual knowledge from two highly relevant subdomainsof the UMLS, viz. anatomyand pathology, in order to construct a formally sound knowledge base using a terminological knowledgerepresentation language. This task will be divided into four steps: (I) the automatedgeneration of terminological expressions, (2) their submissionto a terminological classifier for consistency checking, (3) the manualrestitution of formal consistency in case of inconsistencies, and, finally, (4) the manual rectification and refinementof the formal representation structures. KB SYSTEMS223 ~7 I CHD " C0014261 ~7 C0005847 CHD CHD C0014261 C0025862 C0005847 CHD C0026844 ~47 C0005847 CHD GriD C0005847CHD C0005847 CHD ~7 CHD C0005847CHD C0005847CHD CSP98 isa C0026844 partof C0034052 C0035330 rsa C0042366 partof C0042367 pad_of C0042367 C0042449 IS8 CSP98 SNMI98 MSI-IgQ MSI-I~ SNM2 MSH99 MSH99 C6P98 MSH99 M81-1~ CSP98 6NMI98 MSH99 MSI-199 MSH99 SNM2 M81-199 Figure 3: SemanticRelations in the UMLS Metathesaurus Step 1: Automated Generationof TerminologicalExpressions. Sourcesfor concepts and relations were the UMLS semantic network and the mrrel, mrconand mrsty tables of the 1999release of the UMLS metathesaurus. The mrrel table whichcontains approximately7,5 million records (cf. Figure3) exhibits the semanticlinks between two UMLS CUIs(concept unique identifier), 2 the mrcon table contains the conceptnamesand mrsty keepsthe semantictype(s) assignedto each CUI.Thesetables, available as ASCIIfiles, wereimportedinto a MicrosoftAccess relational database and manipulatedusing SQLembedded in the VBAprogramminglanguage. For each CUIin the mrrel subset its alphanumeric codewassubstituted by the Englishpreferred term foundin mrcon. After a manualremodelingof the 135 top-level concepts and 247 relations of the UMLS semanticnetwork, weextracted, froma total of 85,899concepts, 38,059anatomy and 50,087 pathology concepts from the metathesaurus. Thecriterion for the inclusion into oneof these sets was the assignmentto predefinedsemantictypes. Also, 2,247 conceptswerefoundto be includedinto both sets, anatomy and pathology. Since wewantedto keep the two subdomainsstrictly disjoint, wemaintainedthese 2,247concepts duplicated, and prefixed all conceptsby ANAor PAT-accordingto their respective subdomain.This can be justified by the observationthat these hybridconceptsexhibit, indeed, multiple meanings.For instance, TUMOR has the meaningof a malignantdisease on the onehand, and of an anatomicalstructure on the other hand. As target structures for the anatomydomainwechose SEPtriplets. Thesewere expressedin the terminological language LOOM which we had previously extended by a special DEFTRIPLET macro(cf. Table 1 for an example). OnlyUMLS part-of, has-part and is-a relation attributes are consideredfor the constructionof taxonomic andpartonomichierarchies. Hence,for each anatomyconcept, one SEPtriplet is created. Theresult is a mixedIS-AandPARTWHOLE hierarchy. For the pathologydomain,wetreated CHD(child) and RN(narrower relation) from the UMLS as indicators 2As a coding convention in UMLS,any two CUIsmust be connected by at least a shallow relation(in Figure3, CHilD relations in the columnRELare assumedbetween CUIs). These shallow relations maybe refined in the columnRELA, ifa thesaurus is available which contains more precise information. Some CUIs are linked either by pan-of or is-a. In any case, the source thesaurus for the relations and the CUIsinvolved is specified in the columns X and Y (e.g., MeSH1999, SNOMED International 1998). 224 FLAIRS-2001 (deftriplet HEART :is-primitive HOLLOW-VISCUS :has-part (:p-and FIBROUS-SKELETON-OF-HEART WALL-OF-HEART CAVITY-OF-HEART CARDIAC-CHAMBER-NOS LEFT-SIDE-OF-HEART RIGHT-SIDE-OF-HEART AORTIC- VALVE PULMONARY-VALVE Table 1: GeneratedTriplets in LOOM Format taxonomiclinks. Nopart-wholerelations wereconsidered, since this categorydoesnot applyto the pathologydomain. Furthermore, for all anatomy conceptscontainedin the definitional statementsof pathologyconceptsthe ’S-node’is the default conceptto whichthey are linked, thus enabling the propagation of roles across the part-wholehierarchy. In both subdomains,shallowrelations, such as the extremely frequent sibling SIB relation, wereincluded as comments into the codeto give someheuristic guidancefor the manualrefinementphase. Step 2: Submission to the LOOM Classifier. The import of UMLS anatomyconcepts resulted in 38,059 DEFTRIPLET expressions for anatomicalconcepts and and 50,087 DEFCONCEPT expressions for pathological concepts. Each DEFTRIPLET was expandedinto three DEFCONCEPT(S-, E-, and P-nodes), and two DEFRELATION (ANATOMICAL-PART-OF-X,INV- ANATOMICAL-PART-OF- x) expressions, summingup to 114,177concepts. This yielded (togetherwith the conceptsfromthe semanticnetwork)a total of 240,764definitory LOOM expressions. From38,059anatomytriplets, 1219DEFTRIPLET statementsexhibited a :HAS-PART clause followedby a list of a variable numberof triplets, containing morethan one argumentin 823 cases (average cardinality: 3.3). 4043 DEFTRIPLET statements containeda :PART-OF clause, only in 332 cases followedby morethan one argument(average cardinality: 1.1). Theresulting knowledgebase was then submittedto the terminologicalclassifier andchecked for terminologicalcycles and coherence.In the anatomy subdomain,one terminologicalcycle and 2328incoherent conceptswerefound, in the pathologysubdomain 355 terminologicalcycles thoughnot a single incoherentconcept weredetermined(cf. Table2). Step 3: ManualRestitution of Consistency.The inconsistencies of the anatomypart of the knowledgebase identified bythe classifier couldall be tracedbackto the simultaneous linkageof twotriplets by both is-a andpartof links, an encodingthat raises a conflict dueto the disjointness required for correspondingP- and E-nodes.In mostof these cases the affectedparentsbelongedto a class of concepts that obviouslycannot be appropriately modeled as SEPtriplets, e.g., SUBDIVISION-OF-ASCENDINGAORTA, ORGAN-PART. The meaningof each of these concepts almostparaphrasesthat of a P-node,so that in these Triplets defconcept statements cycles inconsistencies II Anatomy Pathology 38,059 114,177 1 2,328 50,087 355 0 \ Table 2: Classification Results for the ConceptImport Anltomlr.al-Pad.Of ( cases the violation of the SEP-internal disjointness condiIsA between P - and S - nodes i Has-Anatomical-Part tion was resolved by substituting the involved triplets with simple LOOM concepts, by matching them with already exFigure 4: Part-whole ReasoningPatterns with SEPTriplets isting P-nodes or by disabling IS-A or PART-OF links. cept definitions required the inclusion of anatomic or In the pathologypart of the knowledgebase, we expected partonomiclinks. Often, necessary taxonomicor partoa large numberof terminological cycles, as a consequence nomic parents were already available, but not coded of interpreting the thesaurus-style narrowerterm and child as UMLSparents or broader concepts. 7 from 100 relations through taxonomic subsumption (IS-A). Bearanatomyconcepts had to be considered as misclassified ing in mind the size of the knowledgebase, we consider by the UMLS. 355 cycles a tolerable amount. Those cycles were primarily due to very similar concepts, e.g., ARTERIOSCLERO- ¯ Check of the :has-part arguments assuming ’real SIS vs. ATHEROSCLEROSIS, AMAUROSIS vs. BLINDNESS, anatomy’. In the UMLS sources part-of and has-part and residual categories ("other", "NOS"= not otherwise relations are considered as symmetric. According to specified). These were directly inherited from the source our transformation rules, the attachment of a role HASterminologies and are notoriously difficult to interpret out ANATOMICAL-PART tO an E-node BE, with its range of their definitional context, e.g., OTHER-MALIGNANT- restricted to AEimplies the existence of a concept A NEOPLASM-OF-SKINvs. MALIGNANT-NEOPLASM-OFfor the definition of a conceptB. Onthe other hand, the SKIN-NOS. The cycles were analyzed and a negative list classification of AEas being subsumedby the P-node which consisted of 630 concept pairs was manuallyderived. Be, the latter being defined via the role ANATOMICALIn a subsequentextraction cycle weincorporated this list in PART-OF restricted to BE, implies the existence of BE the automated construction of the LOOM concept definigiven the existence of AE.These constraints do not altions, and given these new constraints, a fully consistent ways conformto ’real’ anatomy, i.e., anatomical conknowledgebase was generated. cepts that mayexhibit pathological modifications. FigStep 4: ManualRectification and Refinement of the ure 4 (left) sketches a concept A that is necessarily KnowledgeBase. This step - when performed for the ANATOMICAL-PART-OF a concept B, but whose exiswhole knowledge base - is time-consuming and requires tence is not required for the definition of B. This is typbroad and in-depth medical expertise. An analysis of ranical of the results of surgical interventions, e.g., a large domsamples from both subdomainsis currently being perintestine without an appendix. formedby a domainexpert. The preliminary data we supply Results: The analysis of 15 triplet definitions that exrefer to the analysis of two randomsamples of each onehibit automatically generated :HAS-PART clauses rehundred anatomy and one-hundred pathology concepts. vealed that 34% of the concepts should be eliminated From the experience we gained so far, the following from the :HAS-PART list in order not to obviate a coworkflowsteps can be derived: herent classification of pathologically modifiedanatom¯ Checking the correctness of both the taxonomic and ical objects) The opposite situation is also common partitive hierarchies. Taxonomic and partitive links are (cf. Figure 4, right): the definition of AEdoes not manually added or removed. Primitive subsumption imply that the role ANATOMICAL-PART-OF be filled by is substituted by non-primitive subsumptionwhenever BE, but BEdoes imply that the inverse role be filled possible. This is a crucial point, because the automatby AE. As an example, a LYMPH-NODE necessarily ically generated hierarchies contain only information contains LYMPH-FOLLICLES, but there exist LYMPHabout the parent concepts and necessary conditions. As FOLLICLES that are not part of a LYMPH-NODE. an example, the automatically generated definition of ¯ Analysis of the sibling relations and defining concepts DERMATITIS includes the information that it is an INas being disjoint. In UMLS,SIB relates concepts that FLAMMATION, and that the role HAS-LOCATION must share the same parent in a taxonomicor partonomichibe filled by the concept SKIN. An INFLAMMATION that erarchy. Pairs of sibling concepts may have common HAS-LOCATION SKIN, however, cannot automatically be classified as DERMATITIS. 3In the example of Table1, the conceptsprintedin italics, viz. Results: Taxonomiclinks had to be removed from 8 AORTIC-VALVEand PULMONARY-VALVE should be eliminated out of 100 sampled pathology concept definitions, but fromthe :HAS-PART list, becausethey maybe missingin certain from none of the anatomyconcept definitions. On the cases as a result of congenitalmalformations,inflammatory procontrary, in 68 cases from this sample, anatomy concessesor surgicalinterventions. KB SYSTEMS226 descendantsor not. If not, they constitute the root of two disjoint subtrees. In a taxonomichierarchy, this means that one conceptimplies the negation of the other (e.g., a benignant tumor cannot be a malignant one, et vice versa). In a partitive hierarchy, this canbe interpreted as spatial disjointness, viz. one conceptdoes not spatially overlap with another one. As an example, ESOPHAGUS and DUODENUM are spatially disjoint, whereas STOMACHand DUODENUMare not (they share a common transition structure, called PYLORUS), such as all neighbor structures that have a surface or region in common. Spatial disjointness can be modeledsuch that the definition of the S-nodeof the concept A implies the negation of the S-nodeof the concept B. Results: The large numberof sibling concepts (on the average 7.3 siblings per concept in the anatomy,8.8 in the pathology subdomain) makes the modeling of disjointness a time-consumingtask, as every pair of concepts must be analyzed. At first glance, our data indicate that conceptualdisjointness holds for at least two-thirds of the sibling conceptsin both domains,and spatial disjointness for over three quarters in the anatomydomain. ¯ Completion and modification of anatomy-pathologyrelations. Surprisingly, only very few pathology concepts contained an explicit reference to a corresponding anatomy concept. These relations must, therefore, be added by a domainexpert. In each case, the decision must be madewhether the E-node or the S-node has to be addressed as the target concept for modification. In the first case, the propagationof roles across part-whole hierarchies is disabled, in the secondcase it is enabled. Results: In an analysis of a random sample of 100 pathology concepts, only 17 were found to be linked with an anatomyconcept. In 15 cases, the default linkage to the S-node was considered to be correct, in one case the linkage to the E-nodewas preferred. In another case, the linkage was consideredfalse. Conclusions Instead of developing sophisticated medical knowledge bases from scratch, we here propose a ’conservative’ approach -- reuse existing large-scale resources, but refine the data from these resources so that advancedrepresentational requirements imposedby more expressive knowledge representation languagesare met. The knowledge engineering approach we propose does exactly this. It provides a formally solid description logics frameworkwith a modeling extension by SEPtriplets so that both taxonomicand partonomic reasoning are supported equally well. While pure automatic conversion from semi-formal to formal environmentscauses problemsof adequacyof the emergingrepresentation structures, the refinement methodologywe propose already inherits its power from the terminological reasoning framework. In our concrete work, we found the implications of using the terminological classifier, the inference engine which computes subsumptionrelations, of utmost importance and of out226 FLAIRS-2001 standing heuristic value. Hence, the knowledgerefinement cycles are truly semi-automatic,fed by medicalexpertise on the side of the humanknowledgeengineer, but also driven by the reasoning system which makes explicit the consequencesof (im)proper concept definitions. Acknowledgements.Wewould like to thank our colleagues in the CLIFgroupfor fruitful discussionsandinstant support.St. Schulzwas partly supportedby a grant fromDFG(Ha2097/5-2). References Artale, A.; Franconi,E.; Guarino,N.; and Pazzi, L. 1996.Partwholerelations in object-centeredsystems:an overview.Data &KnowledgeEngineering20(3):347-383. Cimino,J. J. 1998.Auditingthe UnifiedMedicalLanguage System with semantic methods.Journal of the AmericanMedical lnformaticsAssociation5(1):41--45. C6t6, R. 1993. SNOMED International. Northfield, IL: College of American Pathologists. Hahn, U.; Romacker,M.; and Schuiz, S. 1999. Howknowledge drives understanding:matchingmedical ontologies with the needsof medicallanguageprocessing.Artificial Intelligence in Medicine15(1):25-51. Hahn,U.; Schuiz, S.; and Romacker,M.1999. Part-wholereasoning: a case study in medical ontology engineering. IEEE Intelligent Systems&their Applications14(5):59-67. Haimowitz, I. J.; Patii, R.S.; andSzolovits,P. 1988.Representing medicalknowledge in a terminologicallanguageis difficult. In Proceedingsof the 12th AnnualSymposium on ComputerApplications in MedicalCare- SCAMC’88, 101-105. Horrocks,I., and Sattler, U. 1999. A description logic with transitive andinverseroles androle hierarchies.Journalof Logic and Computation9(3):385-410. Mays,E.; Weida,R.; Dionne,R.; Laker, M.; White,B.; Liang, C.; and Oles, E J. 1996.Scalableandexpressivemedicalterminologies. In Proceedingsof the 1996AMIAAnnualFall Symposium (formerly SCAMC) - AMIA’96, 259-263. McCray,A. T., and Nelson,S.J. 1995. The representation of meaningin the UMLS. Methodsof Information in Medicine 34( 1/2): 193-201. Pisanelli, D. M.; Gangemi, A.; and Steve, G. 1998.Anontological analysis of the UMLS metathesaurus.In Proceedingsof the 1998AMIAAnnualFall Symposium - AMIA’98, 810--814. Rector,A. L.; Bechhofer, S.; Goble,C. A.; Horrocks,I.; Nowlan, W. A.; and Solomon, W.D. 1997. The GRAILconcept modelling languagefor medicalterminology.Artificial Intelligence in Medicine9:139-171. Reggia,J. A., and Tuhrim,S., eds. 1985. Computer-Assisted MedicalDecisionMaking,volume1 &2. NewYork: Springer. Schmidt-SchauB, M., and Smolka,G. 1991. Attributive concept descriptionswithcomplements. Artificial Intelligence48:1-26. Schmolze,J. G., and Mark,W.S. 1991. TheNIKLexperience. Computational Intelligence 6:48-69. Spackman,K. A., and Campbell, K.E. 1998. Compositional concept representation using SNOMED: towards further convergenceof clinical terminologies.In Proceedings of the 1998 AMIAAnnualFall Symposium - AMIA’98, 740-744. Volot, E; Zweigenbaum, P.; Bachimont,B.; Ben Said, M.; Bouaud,J.; Fieschi, M.; and Boisvieux,J.-F. 1994. Structuration and acquisition of medicalknowledge:using UMLS in the ConceptualGraphformalism. In Proceedingsof the 17th Annual Symposiumon ComputerApplications in MedicalCareSCAMC’93, 710-714.