Int. J. Appl. Math. Comput. Sci., , Vol. , No. , –DOI:

ENGINEERING INTELLIGENT SYSTEMS ON THE KNOWLEDGEFORMALIZATION CONTINUUM

JOACHIM BAUMEISTER, JOCHEN REUTELSHOEFER, FRANK PUPPE

Institute of Computer ScienceUniversity of Wurzburg, Germany

e-mail: surname@informatik.uni-wuerzburg.de

Albeit their industrial success, the development of intelligent systems is still a complex and risky task. When buildingintelligent systems, we see that the domain knowledge is often present at different levels of formalization—ranging fromtext documents to explicit rules.In this paper, we describe the knowledge formalization continuum as a metaphor to help the domain specialists during theknowledge acquisition phase. To make use of the knowledge formalization continuum, the agile use of knowledge repre-sentations within a knowledge engineering project is proposed, as well as transitions between the different representations,when required. We motivate that a Semantic Wiki is a flexible tool for engineering knowledge on the knowledge formaliza-tion continuum. Case studies are taken from one industrial and one academic project, and they illustrate the applicabilityand benefits of Semantic Wikis in combination with the knowledge formalization continuum.

Keywords: knowledge engineering, knowledge acquisition, semantic wiki

1. Introduction

Intelligent systems, especially knowledge-based solu-tions, are successfully implemented in today’s enterprises.We experience the application of intelligent systems indifferent ”flavors”, ranging from (semantically) enrichedinformation systems to sophisticated model-based expertsystems. Consequently, the knowledge embodied in thesedifferent types of systems differs strongly from naturallanguage text to explicit knowledge in form of rules ormodels.

Although, such systems proved their applicabil-ity and benefits in many domains, the engineering andmaintenance of the underlying knowledge bases is stilla complex and costly task. In this work, we intro-duce the knowledge formalization continuum as a con-ceptual metaphor that gives domain specialists a flexi-ble mental model of the knowledge that is planned tobe formalized. The knowledge formalization contin-uum emphasizes that usable knowledge ranges from veryinformal representations—such as text and images—tovery explicit representations—such as logic formulas orconsistency-based models. The metaphor frees the spe-cialists and engineers to commit to a particular degree ofknowledge formalization at an early stage of the develop-

ment project, but offers a versatile understanding of theformalization process. Furthermore, we motivate that aSemantic Wiki (Schaffert et al., 2008) is an appropriateworkbench for the development of intelligent systems us-ing the knowledge formalization continuum, since it sup-ports the creation and evolution of knowledge in variousfacets and thus supports the idea of the knowledge formal-ization continuum.

The rest of the paper is organized as follows: In Sec-tion 2, we introduce the concept of the knowledge for-malization continuum in more detail. We explain, possi-ble representations on the knowledge formalization con-tinuum and we motivate methods, that facilitate transi-tions between the representations. We further present theSemantic Wiki KnowWE in Section 3 and we motivateits application within the knowledge formalization con-tinuum. A number of examples and case studies using thecontinuum are reported in Section 4 and we conclude thearticle in Section 5.

2. The Knowledge Formalization Contin-uum

In general, a continuum describes a ”continuous sequencein which adjacent elements are not perceptibly differ-

2 Baumeister et al.

Knowledge Formalization Continuum

TextTags

Semantic annotations

Fault models

Functional modelsDecision

trees

Cases

Segmented text

Tabular data

Semantically equivalent transitions

Images

Mindmaps /Flow charts

LogicRules

Fig. 1. Possible knowledge representations of the knowledge formalization continuum.

ent from each other, but the extremes are quite dis-tinct” (Oxford English Dictionary of Current English,2008).

We interpret this definition of a continuum in the con-text of knowledge representations as follows: The samedomain knowledge can be represented in various typesof knowledge representations, where adjacent representa-tions are similar to each other—for example, tabular dataand XML, but more extreme representations are quite dis-tinct, for instance text vs. logic rules. In such a knowl-edge formalization continuum, gradual transitions on for-malization degrees of the same knowledge are possible,but the knowledge to be modelled experiences no abruptchanges or “discontinuities”. Albeit we interpret the con-tinuum as a mental concept of alternative knowledge rep-resentations, it is important to notice that the representeddomain knowledge remains the same. The knowledge for-malization continuum also emphasizes that the knowledgeundergoes an ongoing, continuous development—wherechanges can also imply the switch of the representation.

We note that the knowledge formalization continuumis neither a physical model nor a methodology for de-veloping knowledge bases. It rather should be seen as ametaphor of the knowledge development process. It helpsthe domain specialists to see even plain data, such as textand multimedia, as first-class knowledge that can be trans-formed by gradual transitions to more formal representa-tions when required. On the one hand, data given by tex-tual documents denotes one of the lowest possibilities offormalization. On the other hand, functional models storeknowledge at a very formal level.

Figure 1 shows an example depiction of differentknowledge representations possible within the knowledgeformalization continuum. This enumeration of represen-tations is certainly not intended to be exhaustive; also, thedepicted order of representations between data and knowl-edge is not meant to be explicit. In fact, it appears diffi-cult/impossible to define the total order of the representa-

tions in a general manner. The depicted order was moti-vated by the level of possible expressiveness with respectto the reasoning power of built systems using the particu-lar representation. For example, text can be used for stan-dard keyword-based search and retrieval, whereas seman-tically annotated text allows for semantic queries and nav-igation. At the right end, knowledge based on rules allowsfor even more complex reasoning capabilities.

Every level of formalization has its own advantagesand drawbacks. For example, textual knowledge can beeasily elicited and often is already available in the domain.No prior knowledge with respect to tools or knowledgerepresentation is necessary. However, automated reason-ing using the textual knowledge is not possible with cur-rent state–of–the–art methods, and the knowledge can beretrieved only by using string-based matching methods butnot by applying semantic queries. Logic rules or modelsare well-suited for automated reasoning, and queries canbe processed at the semantic level. In contrast to textualknowledge, the acquisition of rules and models is typi-cally a complex and time-consuming task. Authors needprior training before effectively building knowledge basesat the explicit level with respect to knowledge engineeringprinciples as well as tools that support such knowledgemodeling. For a given knowledge base, that is formalizedusing a particular knowledge representation, there oftenexist semantically equivalent transitions (indicated by thesecond axis in Figure 1). For example, a fault model basedon set-covering models can be often also represented by arule base which in turn may be modelled by a special pur-pose logic dialect. Often, the knowledge is brought to a se-mantically equivalent representation in order to enable ex-tensions by additional domain knowledge. For example, aknowledge base represented by fault models can be tran-sitioned to a rule base in order to allow for a fine-graineddefinition of conditioning findings for a target concept.

Between the two extremes (text vs. logic) there ex-ists a wide range of formats representing knowledge at

Engineering Intelligent Systems on the Knowledge Formalization Continuum 3

different degrees of formalization. Any degree can bethe most useful representation for a specific applicationproject. For a given project, it is a difficult but an im-portant task to select the most appropriate formalizationlevel as the target representation. Since often (fragmentsof) knowledge are already available in textual or tabularform, the development process focuses on bringing theexisting forms to an appropriate level. Although, it typ-ically becomes necessary to fill in missing parts of theknowledge, the original nature of the knowledge remains.Thus, moving to a more formal representation can requirethe more explicit description of the knowledge and canenrich the resulting knowledge with additional semanticsmade explicit. It is worth noticing, that every transitionis a distinct operation that modifies the knowledge repre-sentation. However, the mental model of the knowledgeremains basically the same.

2.1. Transitions on the Knowledge FormalizationContinuum. Transitions between different levels of for-malization, for example from text to cases, can be done ina manual and sometimes in a semi-automated way. Thetransitions from explicit knowledge to less explicit levelsof knowledge are supported by the following methods:

• Natural language generation techniques, for exam-ple (Reiter and Dale, 2000).

• Visualization techniques, for example (Geroimenkoand Chen, 2006; Zacharias, 2007; Card, 2003;Baumeister and Freiberg, 2010).

• Knowledge explanation methods, for example (Roth-Berghofer, 2004).

The direction of transition into a less formal knowledgerepresentation is required, for example to review/evaluatethe developed knowledge bases. Here, less formal but pre-cise visualizations of the knowledge base help to under-stand the semantics of the implemented knowledge.

The typical direction in knowledge engineeringprojects is the transition of little structured/unstructureddata into a more explicit representation; here, the follow-ing disciplines propose useful methods:

• Text Mining, Ontology Learning, and Natural Lan-guage Processing in general for the machine–enabled extraction of concepts from texts, their tax-onomic ordering, and the discovery of basic rela-tions between found concepts, see (Dale et al., 2000;Buitelaar et al., 2005; Feldman and Sanger, 2006) forexample.

• Controlled Languages to automatically interpret a re-stricted subset of natural language text as logic for-mulas, for example an overview is given in (Fuchset al., 2008).

• Templates/forms to enter knowledge on a basic levelby using as much natural language as possible.

• Refactoring methods to support manual changes ofexplicit knowledge without changing the intended se-mantics, for example (Gil and Tallis, 1997; Baumeis-ter et al., 2004; Reutelshoefer, Baumeister andPuppe, 2009). They are often used to accomplishvertical transitions to a semantically equivalent ver-sion within the same knowledge representation, butare also helpful to restructure the knowledge to aless/more formalized level.

• Manual Knowledge Elicitation methods, that are ap-plied when it is not reasonable or tractable to use(semi-)automated methods as sketched above.

In an example application project we can imagine tohave knowledge already available contained in a textualform such as MS-Word documents and semi-structuredMS-Excel sheets. By using ontology learning methods weare able to extract relevant ontological concepts and basicrelations afterwards. In subsequence, strongly formalizedmodels are (manually) defined to formulate enhanced rela-tions between the concepts. The initial textual knowledgeis still available, but now annotated by the added forms offormalized knowledge.

In the future, we envision that the methods men-tioned above work perfectly together to build and en-volve knowledge bases. Most state-of-the-art methods,however, are currently only capable to provide a seedfor knowledge base development, that requires a manualknowledge refinement and formalization.

2.2. Implications. The knowledge formalization con-tinuum embraces the fact that knowledge is usually rep-resented at varying levels of formality. The continuumsupports the entrance of the knowledge engineering pro-cess at an arbitrary level of formality and offers pos-sible transitions of the knowledge to a level where itscost/benefit principle (Lidwell et al., 2003, p. 56) is (in thebest case) optimal. In typical projects, prior knowledge ofthe domain is already at hand, often in form of text doc-uments, spreadsheets, flow-charts, and data bases. Thesedocuments build the foundational reference of the classicknowledge engineering process, where a knowledge engi-neer models the domain knowledge based on these docu-ments. The actual utility and applicability of the knowl-edge usually depends on the particular application.

The knowledge formalization continuum does notpostulate to transform the entire collection into a knowl-edge base at a specific degree, but to perform transitionson parts of the collection when it is possible and appropri-ate. This takes into account that sometimes not all parts ofa domain can be formalized at a specific level or that theformalization of the whole domain knowledge would be

4 Baumeister et al.

too complex, considering costs and risks. In consequence,a system working on the knowledge formalization con-tinuum is required to support the knowledge engineeringprocess at different levels of formalization. It also shouldbe able to support the knowledge sharing process, i.e., itsactual usage, at varying formalization levels.

Following, the cost/benefit principle it needs to bepossible to transform the parts of the knowledge to a levelof formalization, where the (knowledge engineering) costsare minimized and the benefits of using the system aremaximized. Therefore, the knowledge formalization con-tinuum not only needs to support the transitions of partic-ular parts of the knowledge but also should be able to keepreferences between the less and more formalized parts ofthe entire knowledge collection.

2.3. Related Approaches. The presented concept ofthe knowledge formalization continuum is referenced inthe literature in a more general context. (Uschold andGruninger, 2004) discuss different kinds of ontologiesranging from terms to general logic. They also name thisspectrum of possibile representations a continuum, but donot propose transitions between the particular kinds, norproposing to use them jointly when developing intelligentsystems. Similarly, (McGuinness, 2003) introduces theontology spectrum, but concentrates on the expressivenessof ontology languages. (Gruber, 2008) also correlates theformalization degree of data with the depth of inference,and he discusses the ”value/cost” trade-off of the capturedknowledge. He puts the matrix of the formalization de-gree and the development costs in the context of social se-mantic web applications. Further, (Schaffert et al., 2006)distinguish between model scope, model acceptance andthe level of expressiveness, where the latter defines asubspace of the presented knowledge formalization con-tinuum. The level of expressiveness ranges from light-weight ontologies, with term lists as the least expressiveone, to heavy-weight ontologies with very-expressive con-straints as the most expressive representative. Whereasfunctional models and logic programs can be interpretedas ”very-expressive constraints” in some ways, the knowl-edge formalization continuum also considers textual doc-uments as less expressive occurrences of knowledge apartfrom term lists. (Millard et al., 2005) describe an ap-proach that investigates the formality of documents of hy-pertext systems. A vector-based model of the formalityof semantics of these documents is given. Their scale de-fines a subset of the presented continuum ranging fromplain text to RDF data, which is sufficient in the contextof the considered domain. In the context of the Com-monKADS methodology (Schreiber et al., 2001) describea matrix of knowledge formalizations for knowledge elic-itation techniques. Here, the matrix correlates tacit andexplicit knowledge with concept knowledge and processknowledge. This way, the process of knowledge develop-

ment is explicitly considered. However, the matrix doesnot take transitions between the particular elements intoaccount.

3. Engineering the Knowledge Formaliza-tion Continuum with Semantic Wikis

In the previous section, we introduced the ideas of theknowledge formalization continuum and motivated thatmultiple knowledge representations need to be consideredduring the development of a knowledge base. In this sec-tion, we first describe the concept of Semantic Wikis, andwe show how such systems are used as knowledge engi-neering tools. Consequently, we demonstrate the use ofSemantic Wikis for developing knowledge systems withinthe knowledge formalization continuum.

3.1. Knowledge Engineering with Semantic Wikis.In recent years the advent and success of Web 2.0 appli-cations, such as wikis, blogs, and social networks, haschanged the way people are using the Internet. Whencompared to traditional web sites, Web 2.0 applicationsexplicitly involve the users as primary contributors to thesystem. Thus, the value of the particular systems usuallygrows with the increasing contribution of the users. Notonly private life but also daily business is influenced bythe success of Web 2.0 approaches. One prominent ex-ample is the wide-spread use of wikis as flexible knowl-edge management tools, both in personal life and businessenvironments. The content of a wiki can be created andmodified by clicking an (often mandatory) edit button lo-cated on the web page. Due to the simple markup (lessverbose than HTML), users can author web pages easily.Wikipedia is certainly the most popular example of wikisystems, where informal world knowledge is created andupdated by a wiki.

Albeit their success, standard wiki systems show lim-itations, especially when using and sharing the includedknowledge. Usually, only full-text search is possible forknowledge retrieval, and knowledge across different wikipages cannot be aggregated to a unified result. This is-sue motivated the development of Semantic Wikis thatextend standard wikis by an explicit ontological layer de-fined by semantic annotations of the wiki content. Thanksto semantic annotations, knowledge reuse is improvedby semantic search and semantic navigation (Schaffertet al., 2008). At the same time, Semantic Wikis suc-cessfully serve as ontology development tools, that pro-vide a simple, web-based interface to build semantic ap-plications. Current examples of Semantic Wiki imple-mentations are, for instance, IkeWiki (Schaffert, 2006),KnowWE (Baumeister et al., 2010), MoKi (Ghidiniet al., 2009), PlWiki (Nalepa, 2009), Semantic Medi-aWiki (Krotzsch et al., 2006), and SweetWiki (Buffaet al., 2008).

Engineering Intelligent Systems on the Knowledge Formalization Continuum 5

Fig. 2. The Semantic Wiki KnowWE showing an article that describes the car fault clogged air filter by using multiple forms ofknowledge: (1) plain text, images, and (2) explicit rules. Knowledge can be used by (3) simple text search and by (4) aninteractive knowledge-based interview.

The knowledge in a Semantic Wiki is typically orga-nized as follows: Every concept of the ontology is rep-resented by a wiki article, and the content of the articleinformally describes the concept. Properties of the con-cept are defined by explicit semantic annotations withinthe article, where the annotations often link to other arti-cles and concepts, respectively. In general, most SemanticWiki systems are capable of developing and maintainingontologies with the expressiveness of a subset of the ontol-ogy language OWL (Antoniou and van Harmelen, 2003).

Figure 2 shows the Semantic Wiki KnowWE depict-ing an example system for capturing knowledge about di-agnosing car faults. Each car fault (here clogged air filter)is represented by a wiki article containing multiple knowl-edge formalizations: (1) Using plain text and images and(2) explicit rules. The knowledge can be used by (3) asimple text search interface and by (4) an interactive in-terview, that is provided based on the explicit knowledge.

In Figure 3, we see the same wiki article in the editmode: Here, we see that special markup is used to inte-grate different kinds of knowledge into the page: (1) in-cludes images depicting an air filter, (2) shows a semantic

annotation stating that clogged air filter is a subclass ofthe concept TechnicalProblem. In (3) a block of rules isdefined, that are used to derive the solution clogged airfilter in a given case.

That way, a wiki article captures informal as wellas formal sources of knowledge. During the knowledgeformalization, the knowledge is typically distributed overthe entire wiki, i.e., the knowledge base is partitioned intological units and embedded in corresponding wiki articles.For the partitioning of the knowledge, no prior restrictionsare defined and the actual organization of the knowledgebase depends on the characteristics of the domain. For in-stance, in many projects the partitioning with respect tosolutions is more appropriate, i.e., defining an article foreach solution, where the problem-solving knowledge forthis solution is also embedded. In other projects, an organ-isation among the cardinal symptoms of the domain maybe more appropriate.

Besides its flexible organization of knowledge, theSemantic Wiki KnowWE also provides the possibility todefine variants of the distributed knowledge base. Whenspecifying authoritative knowledge bases in the wiki, we

6 Baumeister et al.

Fig. 3. The edit mode of the Semantic Wiki KnowWE: Text, images, and formal knowledge can be edited using a simple markup.

use the metaphor of masters and servants: A large coher-ent knowledge base—usually intended to be exported outof the wiki—is defined on the basis of imports of knowl-edge bases contained in other wiki articles. This masterthen includes all knowledge of the imported articles, i.e.,the servants, and the master can be interpreted as a sin-gle, testable knowledge base. That way, we are also ableto define variants of knowledge bases by defining differ-ent masters importing varying collections of wiki articles.Figure 4 depicts an example of two masters of the wiki.In total, the shown wiki contains four servant wiki articleswith knowledge bases, i.e., Article 1 to Article 4. Addi-tionally, the wiki page Article 5 defines the master knowl-edge base Master 1 by including the knowledge basesfrom the servant articles 1, 2, and 3. The alternative mas-ter Master 2 defined in page Article 6 only includes theknowledge bases from servant articles 3 and 4, and thusrepresents a different view of the entire knowledge base.

3.2. Semantic Wikis and the Knowledge Formal-ization Continuum. As we showed above, a Seman-tic Wiki offers flexible ways to simply combine less for-

mal forms of knowledge (text, images) with more explicitknowledge representations (rules, models, etc.). Less for-mal representations of knowledge serve the developmentprocess in many ways: 1) as startup document at the be-ginning of a project to informally collect knowledge aboutthe domain, 2) as documentation of the knowledge en-gineering decisions taken, 3) as underlying tacit knowl-edge expressing the informal counter-part, and 4) as de-scribing information for concepts represented by the ar-ticle. In practice, the inclusion of less formal knowledgealso improves the overall maintainability of the distributedknowledge base.

In the following we demonstrate the use of the wikiwhen building an example sports recommender system.Here, the wiki collects articles about forms of sport andalso captures explicit knowledge, that recommends an ap-propriate form of sport for an entered user profile. The ex-ample shows the subsequent transitions within the knowl-edge formalization continuum. Here, the knowledge al-ready available in the continuum describes relevant factsabout the forms of sports, such as accomplished traininggoals, costs, and medical restrictions. In subsequent stepswe drive the existing knowledge to more formalized tran-

Engineering Intelligent Systems on the Knowledge Formalization Continuum 7

Article 5

Master 1

Article 6

Master 2

Article 1

Knowledge Base 1

Article 2

Knowledge Base 2

Article 3

Knowledge Base 3

Article 4

Knowledge Base 4

includes

includesincludes

includes

includes

Wiki Articles

Fig. 4. Definition of knowledge variants by the specification ofmasters and servants of the wiki.

sitions.

Initial Filling. We start by filling the wiki with text andmultimedia (pictures and movies) describing the differ-ent forms of sport, for example Running, Swimming, andCycling. It is reasonable that for every form of sports awiki page should be created, i.e., after the initial fillingphase there exist pages about running, swimming, and cy-cling. However, also wiki articles about further domainfacts exist, for instance allergies or muscles. In general,it is reasonable to set up one distinct wiki article for eachdistinct concept of the domain, thus following a commonparadigm of (semantic) wikis. For example, an excerpt ofan article about swimming is as follows

...Swimming is the most common form of watersports. In particular it is recommended for peoplewith back problems because it trains the back mus-cles ... However, people with skin allergies shouldavoid swimming. ...

At this point the wiki can be used as a simple and tra-ditional information system specialized on sports, whereusers can search and browse through the available content.

Annotating Articles. We propose to annotate every wikiarticle with its semantic concept, thus making explicit thata specific article is about a specific concept. For instance,we annotate the article about swimming with the conceptSwimming. At this point, only a very general ontology ofconcepts is required to represent the domain concepts al-ready contained in the wiki. As a benefit of this step itbecomes possible to offer a low-end version of semanticsearch and navigation, that will be more useful when con-cepts are carefully structured in a hierarchy.

Annotation by Properties. The next step tries to iden-tify the typical features of every concept described in the

available text. These findings are then annotated as prop-erties of the article’s concept. In the example above thetext about swimming would then transform as follows(new/changed text is given in bold letters):

...Swimming is the most common form of [has-Finding::water sports]. In particular it is recom-mended for people with back problems because it[hasFinding::trains the back muscles]. ...However, people with [isContradictedBy::skinallergies] should avoid swimming. ...

In the given example, the text phrases water sports,trains the back muscles, and skin allergies are annotatedby the properties hasFinding and isContradictedBy, re-spectively. Each annotation performs the creation of anRDF triple with the article’s concept (here, Swimming) asthe subject, the property’s name as the predicate, and areference to the particular text phrase as the object. Theuse of properties implies the extension of the simple do-main ontology of sport forms defined before. In the givenexample, we introduced the properties hasFindings andisContradictedBy. With the properties defined in the wikian extended version of semantic search and navigation be-comes possible. For example, we are now able to queryfindings (as text phrases) that exclude a specific form ofsport, i.e., “return all text phrases that represent the con-tradiction of a given sports form”.

In a further step, it is reasonable to “semantify” thetext phrases representing the particular properties of aconcept. Thus, we gradually extend the existing annota-tion by explicit concepts describing the ranges of the prop-erties.

...Swimming is the most common form of wa-ter sports [hasFinding:: Medium = in water].In particular it is recommended for people withback problems because it trains the back mus-cles [hasFinding:: Trained muscles = back]. ...However, people with skin allergies [isContra-dictedBy:: Medical restrictions = skin allergy]should avoid swimming. ...

In the shown example, the text phrase trains the backmuscles is moved out of the annotation and replaced by theexplicit concept Trained muscles having a concrete valueback. Furthermore, the last annotation describes that thetext phrase skin allergies is annotated by the value skin al-lergy assigned to the concept Medical restrictions. Thisimplies the extension of the ontology by appropriate con-cepts representing the findings for the different forms ofsport. If these concepts are defined in advance, then nat-ural language processing methods can be used for a semi-automatic annotation of the text. In consequence, a full-fledged semantic search and navigation becomes possible,where the relation of a specific finding value to all avail-

8 Baumeister et al.

able sport concepts can be queried, for example.

Generating Problem-Solving Knowledge. In somecases, the use of semantic annotations is not sufficientlyexpressive for a given application project. Then, it be-comes necessary to transform to a higher level of for-malization by generating and extending strong problem-solving knowledge out of the existing annotations. In thefollowing, we aim to define knowledge to actually deriveparticular forms of sports based on entered user findings.For this reason, we collect all properties, that set a formof sport in relation with a finding that can be entered bythe user. In the given example, we collect the proper-ties hasFinding and isContradictedBy. The semantic wikiKnowWE offers scripts that automatically convert theseproperties either into set-covering models or rules. Forfurther properties with a different semantics the scriptscertainly need to be adapted. In the initial step, such a con-version denotes the transition of the available knowledgeinto an (almost) semantically equivalent version. How-ever, in this case the target representation allows for richerpossibilities to represent further elements of the knowl-edge base.Set-Covering Models. The following shows a transitionof the annotation to a set-covering model (Peng and Reg-gia, 1990), where a set-covering model describes all typ-ical/relevant findings for a solution. The given textualmarkup to be used in wikis was introduced in (Baumeisteret al., 2007). In our example, the solution concepts arecorresponding to the concepts representing the wiki ar-ticles, and findings are defined as the target concepts ofthe included properties. Each of the collected propertiesis compiled by the script into a line of the set-coveringmodel. The value of a hasFinding property is representedas a simple line (denoting the positive expectation of thisfinding), for example Trained muscles = back. For a is-ContradictedBy property the conversion additionally addsa [- -] at the end of the generated line in order to representthe negative expectation of this finding, for example seeMedical restrictions = skin allergy.

Swimming {Medium of sports = waterType of sport = individualTrained muscles = backRunning costs >= mediumMedical restrictions = skin allergy [- -]

}

Bold-faced letters are (hand-crafted) additions to themodel, that have been made after the transition. For in-stance, two further findings are describing the type ofsport and the running costs. The explicit representation inthe model points to an extension of the formalized knowl-edge, although this information is already available in thetext of the wiki article.

Rules. In the following example block, a rule-based ver-sion of the annotations made is shown. In this simple ex-ample, one rule is created by a script collecting all has-Finding properties as well as one rule for every isContra-dictedBy property. Of course, this naive conversion doesnot necessarily conform with the intended semantics ofthe made annotations and therefore is meant as a startingpoint for further (manual) adaptations.

if Medium of sports = waterand Type of sport = individualand Trained muscles = backand Running costs >= medium

then derive Swimming

if Medical restrictions = skin allergythen exclude Swimming

The transition to more expressive knowledge repre-sentations such as set-covering models and rules becomesnecessary when complex relations of the domain cannotbe expressed by semantic annotations anymore. As a ben-efit, the knowledge can then be used for more effectivereasoning ranging from complex semantic queries to thegeneration of problem-solving interviews, where appro-priate solutions for a given problem are derived based onan interactive interview.

In the example above, we showed the stepwise for-malization of knowledge along the knowledge formaliza-tion continuum. In the next section, we present some casestudies and we also motivate that the counter directionalong the knowledge formalization continuum poses in-teresting applications.

4. Case StudiesIn the following, we describe two case studies and demon-strate the application of the knowledge formalization con-tinuum and the Semantic Wiki KnowWE, respectively.The first case study considers the development process ofthe medical decision-support system CareMate, whereasthe second example shows the use of the knowledge for-malization continuum in the context of the e-learning ap-plication HermesWiki.

4.1. Medical Decision-Support with CareMate.The decision-support system Digitalys CareMateis commercially sold by the company Digitalys(http://www.digitalys.de) as part of an equipment kitfor medical rescue trucks. It is a consultation system formedical rescue missions, when the problem definitionof a particular rescue service is complex and a secondopinion becomes important. The knowledge base cur-rently contains about 260 findings distributed over 33questionnaires. The system is able to derive 145 distinct

Engineering Intelligent Systems on the Knowledge Formalization Continuum 9

Fig. 5. Article of the cardinal symptom stomach pain (”Bauchschmerzen”) showing some text explaining the general structure of thecorresponding decision tree and an overview image, that can be enlarged on click (original screenshot is depicted in germanlanguage).

solutions. The compiled knowledge base resulted in a(merged) decision tree having 2051 possible diagnosticpaths.

The knowledge base is developed using the Seman-tic Wiki KnowWE. In the wiki the knowledge base is par-titioned by their cardinal symptoms. So, every cardinalsymptom is represented by a distinct wiki article. Fur-thermore, there exists one wiki article describing the solu-tions of the system. On every page of a cardinal symptom,there is some explanation text given in addition to the for-malized version of the knowledge as a heuristic decisiontree. For larger cardinal symptoms, there are further pageslinked from the original article, where modular parts of thedecision tree are formalized. In Figure 5 the wiki article ofthe cardinal symptom stomach pain (”Bauchschmerzen”)is shown.

At the beginning of the project, the knowledge wasoriginally captured by transferring textbook knowledgeand expert knowledge into flowchart diagrams using thesoftware MS-Visio. The domain specialist was neither ex-perienced with using (specialized) computer software nortrained in knowledge representation. The use of a stan-dard office software, such as MS-Visio, appeared to bethe most natural approach to capture the initial structureof the knowledge base, since no new software distractedthe specialist, but he was able to concentrate on the actualknowledge structuring. After defining the initial structureand logic of the knowledge base over a number of dia-grams, we decided to transition the knowledge into a deci-sion tree representation to allow for the automated execu-tion of the knowledge by the reasoner d3web (Baumeisteret al., 2008).

10 Baumeister et al.

Within the Semantic Wiki both knowledge formal-izations can be represented very easily: The decision treeswere defined directly within the wiki text using a special-ized markup (Baumeister et al., 2010), whereas the origi-nal MS-Visio diagrams were embedded as attachments tothe corresponding wiki articles. Additional text of the ar-ticles explain relations and design decisions. For instance,Figure 5 shows the article of the logical unit stomach pain(”Bauchschmerzen”), where MS-Visio diagrams, organi-zational comments, and decision tree knowledge of thesame knowledge entity are jointly combined. The MS-Visio diagrams can be viewed and downloaded in the wiki.Comments in the article additionally describe, that the de-cision tree logic was divided into two decision trees han-dling the diagnosis of stomach pain for women and formen, separately. The lower part Figure 5 shows an ex-cerpt of the formalized knowledge base, where first thesex (”Geschlecht”) of the patient is asked.

The CareMate system is running on a touch-screentablet, and therefore releases of the knowledge base needto be exported from the wiki into a runtime version. Here,KnowWE provides a deployment feature, that compilesa predefined master article into an executable knowledgesystem.

Before the deployment of a new version of the Care-Mate system, all possible derivation paths of the knowl-edge base needed to be verified. For this task, we uti-lized the visualization technique DDTrees (Baumeisterand Freiberg, 2010) to transfer the available knowledgeinto a less formalized representation. That way, we movethe knowledge to the representation level that was mostappropriate for the particular task.

In summary, the ideas of the knowledge formaliza-tion continuum were applied in many ways during thedevelopment of the CareMate knowledge base. Start-ing with MS-Visio flowcharts formalized from textbookknowledge, the domain specialist could concentrate on theorganisation and structuring of the knowledge. In a secondstep, the flowcharts were semi-automatically transferredto the formal knowledge representation heuristic decisiontrees in order to allow for an automated reasoning process.For the verification process, the knowledge was visualizedby a less knowledge formalization, here DDTrees.

4.2. The e-Learning Platform HermesWiki. TheHermesWiki (Reutelshoefer et al., 2010) is an e-learningplatform developed by the historians of the Departmentof Ancient History from the University of Wurzburg, Ger-many, supported by the Department of Intelligent Systemsof the Institute of Computer Science. HermesWiki is aKnowWE implementation and consists of a concise andreliable overview of Ancient Greek History. The knowl-edge provided here is intended to serve as teaching mate-rial for undergraduate students. At the time of writing, theHermesWiki contains about 800 pages covering essays,

translated excerpts of original sources, and a glossary.HermesWiki is an interesting example of the applica-

tion of knowledge formalization continuum: The domainspecialists—researches and teachers in ancient history—originally started by filling a standard wiki with textualcontent and images. They defined categories and rela-tions by not using explicit markup, but formatting the wikitext following a discussed and agreed convention. Thisconvention simplified the automated extraction of knowl-edge from the entered text. Due to the extensibility ofthe wiki (Reutelshoefer, Lemmerich, Haupt and Baumeis-ter, 2009) it was easy to extend the engine by specializedplugins to automatically parse and formalize the availabletext.

The following block gives an example of the applied”markup by convention”, where time events of history aredefined by a special text formatting:

<<Lamian War323b-322b

After Alexanders death the Greeksrevolted against Macedonian ruleunder the lead of the Athenians.[....]

SOURCE: Paus::1,25,3-6SOURCE: Diod::18,8-18>> ...

The example describes the event Lamian War, that isgiven as the title in the first line. In the second line the timeevent is annotated by a time stamp specifying the durationof the considered event. The text ”323b-322b” denotesthat the event occurred 323 BC until 322 BC. The follow-ing lines consists of the body of the event, describing itsdescription in free text. Optionally, historical sources areappended to the text, explicitly annotated by the keywordSOURCE. In our example, two sources—from Diodor andPausanias— are included.

As a benefit the markup is human-readable as freetext, but is sufficiently formal to be parsed as ontologicaldescriptions. The ontology—generated by the markups—has multi-faceted applications: The knowledge is used forsemantically enriched wiki pages, for a timeline browser,for semantically annotating Google Maps, and a dynami-cally generated quiz for students.

Figure 6 shows the HermesWiki article aboutAlexander the Great (”Alexander der Große”). Here, theontology is used to geographically highlight the importantevents on a Google Map, where Alexander was involvedin; a less formal knowledge representation (map image) isenriched by ontological knowledge. Underneath, the vi-sualization of time events—defined within the page usingthe markup discussed above—is shown as collapse boxes.

Engineering Intelligent Systems on the Knowledge Formalization Continuum 11

Fig. 6. HermesWiki: An article of Alexander the Great showing a map generated by the ontology. The most important events in thelife of Alexander are geographically highlighted on the map.

The HermesWiki project is an interesting example,where the concept of the knowledge formalization contin-uum was intuitively applied: Starting with simple text andimages, the wiki was filled very easily, some existing doc-uments were already available and could be taken overinto the wiki. Using the light-weight approach of ”con-vention over configuration”, simple markups were agreed,and explicit ontological knowledge was extracted from theformatted text. The ontological knowledge in turn wasused to enrich existing maps by visualizing aggregatedviews of the curriculum vitae of particular persons.

5. Conclusions

This article introduced the concept of the knowledge for-malization continuum, which is a mental concept thathelps knowledge engineers and domain specialists dur-ing the development process of intelligent systems. Theknowledge formalization continuum is neither a physicalmodel of the knowledge nor a methodology for the knowl-

edge formalization process. It rather combines alterna-tive knowledge representations of the same subject into acontinuum, ranging from free text to explicit logic rules.We also motivated that a Semantic Wiki is an appropriatetool to support knowledge engineering following the ideasof the knowledge formalization continuum. We demon-strated its application by the Semantic Wiki KnowWE andtwo case studies: The first case study was taken from themedical domain and considered the development and eval-uation of the decision-support system CareMate. The sec-ond case study demonstrated the use of the knowledge for-malization continuum and KnowWE in the context of theHermesWiki platform—an e-learning project in ancienthistory.

The full power of the knowledge formalization con-tinuum cannot be fully exploited with current state-of-the-art methods. In the future, automated methods, as wehave sketched in Section 2.1, need further improvementand practical applicability. For example, adapted and ef-fective NLP methods are necessary, to declaratively and

12 Baumeister et al.

accurately transfer free text knowledge into more explicitforms. Counter-wise it becomes interesting to providedeclarative methods, that allow for a flexible and tailoreddefinition of explanations and visualizations of more ex-plicit representations of knowledge.

ReferencesAntoniou, G. and van Harmelen, F. (2003). Web ontology lan-

guage: OWL, in S. Staab and R. Studer (eds), Handbookon Ontologies in Information Systems, Springer-Verlag.

Baumeister, J. and Freiberg, M. (2010). Knowledge visualiza-tion for evaluation tasks, Knowledge and Information Sys-tems in press.

Baumeister, J., Reutelshoefer, J. and Puppe, F. (2007). Markupsfor knowledge wikis, SAAKM’07: Proceedings of theSemantic Authoring, Annotation and Knowledge MarkupWorkshop, Whistler, Canada, pp. 7–14.

Baumeister, J., Reutelshoefer, J. and Puppe, F. (2010).KnowWE: A semantic wiki for knowledge engineering,Applied Intelligence .

Baumeister, J., Seipel, D. and Puppe, F. (2004). Refactor-ing methods for knowledge bases, EKAW’04: EngineeringKnowledge in the Age of the Semantic Web: 14th Interna-tional Conference, LNAI 3257, Springer, Berlin, pp. 157–171.

Baumeister, J. et al. (2008). The knowledge mod-eling environment d3web.KnowME, open-source at:http://d3web.sourceforge.net.

Buffa, M., Gandon, F., Ereteo, G., Sander, P. and Faron, C.(2008). SweetWiki: A semantic wiki, Web Semantics8(1): 84–97.

Buitelaar, P., Cimiano, P. and Magnini, B. (2005). OntologyLearning from Text: Methods, Evaluation and Applica-tions, Vol. 123 of Frontiers in Artificial Intelligence andApplications, IOS Press.

Card, S. (2003). Information Visualization, Julie A. Jacko, An-drew Sears: The Human-Computer Interaction Handbook,Lawrence Erlbaum Associates, pp. 544–582.

Dale, R., Moisl, H. and Somers, H. (eds) (2000). A Handbookof Natural Language Processing: Techniques and Appli-cations for the Processing of Language as Text, MarcelDekker Inc.

Feldman, R. and Sanger, J. (2006). The Text Mining Hand-book: Advanced Approaches in Analyzing UnstructuredData, Cambridge University Press.

Fuchs, N. E., Kaljurand, K. and Kuhn, T. (2008). At-tempto Controlled English for knowledge representation,in C. Baroglio, P. A. Bonatti, J. Małuszynski, M. Mar-chiori, A. Polleres and S. Schaffert (eds), Reasoning Web,Fourth International Summer School 2008, number 5224in Lecture Notes in Computer Science, Springer, pp. 104–124.

Geroimenko, V. and Chen, C. (eds) (2006). Visualizing the Se-mantic Web, 2 edn, Springer.

Ghidini, C., Kump, B., Lindstaedt, S. N., Mahbub, N., Pammer,V., Rospocher, M. and Serafini, L. (2009). MoKi: Theenterprise modelling wiki, ESWC’09: The Semantic Web:Research and Applications, Vol. 5554 of LNCS, Springer,pp. 831–835.

Gil, Y. and Tallis, M. (1997). A script-based approach to modi-fying knowledge bases, AAAI/IAAI’97: Proceedings of the14th National Conference on Artificial Intelligence and 9thInnovative Applications of Artificial Intelligence Confer-ence, pp. 377–383.

Gruber, T. (2008). Collective knowledge systems: Where the So-cial Web meets the Semantic Web, Web Semantics 6(1): 4–13.

Krotzsch, M., Vrandecic, D. and Volkel, M. (2006). Seman-tic MediaWiki, ISWC’06: Proceedings of the 5th Inter-national Semantic Web Conference, LNAI 4273, Springer,Berlin, pp. 935–942.

Lidwell, W., Holden, K. and Butler, J. (2003). Universal Princi-ples of Design, Rockport Publishers.

McGuinness, D. L. (2003). Ontologies come of age, inD. Fensel, J. A. Hendler, H. Lieberman and W. Wahlster(eds), Spinning the Semantic Web, MIT Press, pp. 171–194.

Millard, D. E., Gibbins, N. M., Michaelides, D. T. and Weal,M. J. (2005). Mind the semantic gap, HYPERTEXT ’05:Proceedings of the 16th ACM conference on Hypertext andHypermedia, ACM, New York, NY, USA, pp. 54–62.

Nalepa, G. J. (2009). PlWiki - a generic semantic wiki archi-tecture, ICCCI’09: Computational Collective Intelligence.Semantic Web, Social Networks and Multiagent Systems,number LNCS 5796, Springer, pp. 345–356.

Oxford English Dictionary of Current English (2008). 3 edn,Oxford University Press.

Peng, Y. and Reggia, J. A. (1990). Abductive Inference Modelsfor Diagnostic Problem-Solving, Springer, Berlin.

Reiter, E. and Dale, R. (2000). Building Natural Language Gen-eration Systems, Cambridge University Press, Cambridge.

Reutelshoefer, J., Baumeister, J. and Puppe, F. (2009). Adata structure for the refactoring of multimodal knowl-edge, KESE’09: 5th Workshop on Knowledge Engineer-ing and Software Engineering (CEUR proceedings 486),Paderborn.

Reutelshoefer, J., Lemmerich, F., Baumeister, J., Wintjes, J.and Haas, L. (2010). Taking OWL to Athens – Seman-tic Web technology takes ancient greek history to students,ESWC’10: Proceedings of the 7th Extended Semantic WebConference, Springer, pp. 333–347.

Reutelshoefer, J., Lemmerich, F., Haupt, F. and Baumeis-ter, J. (2009). An extensible semantic wiki architecture,SemWiki’09: Fourth Workshop on Semantic Wikis – TheSemantic Wiki Web (CEUR proceedings 464).

Roth-Berghofer, T. R. (2004). Explanations and Case-BasedReasoning: Foundational issues, Advances in Case-BasedReasoning, Springer-Verlag, pp. 389–403.

Engineering Intelligent Systems on the Knowledge Formalization Continuum 13

Schaffert, S. (2006). IkeWiki: A semantic wiki for collabora-tive knowledge management, STICA’06: 1st InternationalWorkshop on Semantic Technologies in Collaborative Ap-plications, Manchester, UK.

Schaffert, S., Bry, F., Baumeister, J. and Kiesel, M. (2008). Se-mantic wikis, IEEE Software 25(4): 8–11.

Schaffert, S., Gruber, A. and Westenthaler, R. (2006). A seman-tic wiki for collaborative knowledge formation, Proceed-ings of SEMANTICS 2005 Conference, Trauner Verlag.

Schreiber, G., Akkermans, H., Anjewierden, A., de Hoog, R.,Shadbolt, N., de Velde, W. V. and Wielinga, B. (2001).Knowledge Engineering and Management - The Com-monKADS Methodology, 2 edn, MIT Press.

Uschold, M. and Gruninger, M. (2004). Ontologies and seman-tics for seamless connectivity, SIGMOD Record 33(4): 58–64.

Zacharias, V. (2007). Visualization of rule bases – the overallstructure, I-KNOW ’07: Proceedings of International Con-ference on Knowledge Management.

Joachim Baumeister received his PhD in2004 from the University of Wurzburg, Ger-many, where he gives lectures on SemanticWeb, Algorithms, and Knowledge-based Sys-tems. He authored and co-authored more than70 reviewed publications in the field of Knowl-edge Engineering and the Semantic Web. Hismain research interests are related to pragmaticknowledge engineering techniques, knowledgebase evaluation, and industrial applications of

knowledge systems. He organized several national conferences andworkshops and is serving as program committee member for interna-tional conferences and workshops.

Jochen Reutelshoefer is a Ph.D. student atUniversity of Wurzburg, Germany, since 2007.His research interest is (pattern-based) knowl-edge formalization in Semantic Wikis. With hiscolleagues from the department of IntelligentSystems, Wurzburg, he works on the develop-ment of the Semantic Wiki KnowWE. He au-thored and co-authored about 15 reviewed pub-lications in the field of Knowledge Engineeringand Semantic Wikis. He was co-organizer of

the national FGWM workshop on knowledge management in 2009 andof the international Workshop on Semantic Wikis (SemWiki) in 2010.

Frank Puppe is full professor for Artifi-cial Intelligence and Applied Informatics at theUniversity of Wurzburg, Germany. His re-search interests include all aspects of knowl-edge & information systems and its use in med-ical and technical domains. Towards this goalhe developed with colleagues several tool kits& frameworks for knowledge-based diagno-sis, scheduling, subgroup analysis, multiagentsimulation, case-based training, and human-

computer-interaction. He is author or co-author of more than 150 publi-cations and books.

Received: