Fuzzy Domain Ontology Discovery for Business Knowledge Management

Raymond Y.K. LauDepartment of Information Systems

City University of Hong KongTat Chee Avenue, Kowloon

Hong KongE-mail: raylau@cityu.edu.hk

Abstract

Ontology plays an essential role in the formalization ofbusiness information (e.g., products, services, relationshipsof businesses) for effective human-computer interactions.However, engineering of domain ontologies turns out tobe very labor intensive and time consuming. Recently,some machine learning methods have been proposed forautomatic discovery of domain ontologies. Nevertheless,the accuracy and computational ef ciency of the existingmethods need to be improved to support large scale ontol-ogy construction for real-world business applications. Thispaper illustrates a novel fuzzy domain ontology discoveryalgorithm for supporting real-world business ontologyengineering. By combining lexico-syntactic and statisticallearning methods, the accuracy and the computationalef ciency of the ontology discovery process is improved.Empirical studies have con rme d that the proposed methodcan discover high quality fuzzy domain ontology whichleads to signi cant improvement in information retrievalperformance.

Keywords: Domain Ontology, Fuzzy Sets, Text Mining, In-formation Retrieval, Knowledge Management.

1 Introduction

Knowledge has been recognized as the most importantcorporate asset and it is the key for organizations to achievesustainable competitive advantage. Knowledge manage-ment is a collection of processes that govern the creation,dissemination, and utilization of knowledge [25, 26]. Tobe able to effectively manage the intellectual capital, busi-nesses need an effective approach to identify and captureinformation and knowledge about business processes, prod-ucts, services, markets, customers, suppliers, and competi-tors, and to share this knowledge to improve the organiza-

tions’ goal achievement. Ontologies allow domain knowl-edge such as products, services, markets, etc. to be cap-tured in an explicit and formal way such that it can be sharedamong human and computer systems.

The notion of ontology is becoming very useful invarious elds such as intelligent information extractionand retrieval, cooperative information systems, electroniccommerce, and knowledge management [38]. Since TimBerners-Lee, the inventor of the World Wide Web (Web),coined the vision of a Semantic Web [3], the proliferationof ontologies has been under tremendous growth. The suc-cess of Semantic Web relies heavily on formal ontologiesto structure data for comprehensive and transportable ma-chine understanding [19]. Although there is not a univer-sal consensus on the de niti on of ontology, it is generallyaccepted that ontology is a speci cat ion of conceptualiza-tion [9]. Ontology can take the simple form of a taxonomy(i.e., knowledge encoded in a minimal hierarchical struc-ture) or as a vocabulary with standardized machine inter-pretable terminology supplemented with natural languagede nitions. Ontology provides a number of potential bene- ts in representing and processing knowledge, including theseparation of domain knowledge from application knowl-edge, sharing of common knowledge of subjects among hu-man and computers, and the reuse of domain knowledgefor a variety of applications. Ontology is often speci edin a declarative form by using semantic markup languagessuch as RDF and OWL [6]. igure 1 shows an example ofthe domain ontology extracted from the Reuters RCV1 cor-pus [16] and Figure 2 depicts the corresponding OWL state-ments.

Domain ontologies specify the knowledge for a partic-ular type of domain [7]. This kind of ontologies general-ize over application tasks in such domains such as medi-cal, tourism, banking, nance, etc. A well-known exam-ple is the Uni ed Medical Language System (UMLS) andits component parts such as the Medical Subject Heading(MeSH). Although domain ontologies are useful in many

IEEE Intelligent Informatics Bulletin November 2007 Vol.8 No.1

Feature Article: Raymond Y.K. Lau 29

EconomicsCorporate

Financial News

Government Market

Strategy Legal Share

Monetary Inflation Employment

SubConcept SuperConcept

. . . . . . . . . . . . . . . . . . . .

Figure 1. A Crisp Domain Ontology from theRCV-1 Corpus

Figure 2. OWL for the Financial News Ontol-ogy

areas, engineering of these ontologies turns out to be verylabor intensive and time consuming. Therefore, many au-tomatic or semi-automatic ontology engineering techniqueshave been proposed. Although fully automatic construc-tion of perfect domain ontology is beyond the current state-of-the-art, we believe that the automatic ontology miningmethod illustrated in this paper can assist ontology engi-neers to build domain ontology quicker and more accu-rately.

Although some learning techniques have been applied tothe extraction of domain ontology [4, 7, 31], these methodsare still subject to further enhancement in terms of compu-tational ef cienc y and accuracy. One of the ways to im-prove automated domain ontology discovery is to exploitcontextual information from the knowledge sources. As do-main ontology captures domain (context) dependent infor-mation, an effective discovery method should exploit con-textual information in order to build relevant ontologies.On the other hand, since the taxonomy relations discov-ered from a text mining method often involve uncertainty,an uncertainty management mechanism is required to ad-dress such an issue. The notions of Fuzzy set and FuzzyRelation are effective to represent knowledge with uncer-tainty [42]. Therefore, a fuzzy ontology rather than a crispontology is discovered by the proposed text mining method.

De nition 1 (Fuzzy Set) A fuzzy set F consists of a set ofobjects drawn from a domain X and the membership ofeach object xi in F is de ned by a membership functionµF : X 7→ [0, 1]. If Y is a crisp set, ϕ(Y ) denotes a fuzzyset generated from the traditional set of items Y .

De nition 2 (Fuzzy Relation) A fuzzy relation is de nedas the fuzzy set G on a domain X × Y where X and Yare two crisp sets.

Figure 3 highlights the fuzzy domain ontology corre-sponding to the one depicted in Figure 2. The current OWLsyntax can easily be extended to represent fuzzy domainontology using approach similar to [8]. However, we willonly focus on the mining of fuzzy concepts and fuzzy tax-onomy relations in this paper. The term µC×C(c2, c1) inFigure 3 denotes the membership value of the taxonomyrelation from subclass c2 to superclass c1. From the textmining perspective, a keyword is an object and it belongs todifferent concepts (a linguistic class) with various member-ships. The subsumption relations among linguistic conceptsare often uncertain and are characterized by the appropriatefuzzy relations.

De nition 3 (Fuzzy Ontology) A fuzzy ontology is aquadruple Ont =< X, C, RXC , RCC >, where X is a setof objects and C is a set of concepts. The fuzzy relationRXC : X × C 7→ [0, 1] maps the set of objects to the setof concepts by assigning the respective membership values,and the fuzzy relation RCC : C × C 7→ [0, 1] denotes thefuzzy taxonomy relations among the set of concepts C.

The main contribution of our research work presented inthis paper is the development of a novel fuzzy domain ontol-ogy discovery method which exploits contextual informa-tion embedded in textual databases (e.g., product descrip-tion databases). By combining lexico-syntactic and statis-

November 2007 Vol.8 No.1 IEEE Intelligent Informatics Bulletin

30 Feature Article: Fuzzy Domain Ontology Discovery for Business Knowledge Management

EconomicsCorporate

Financial News

Government Market

Strategy Legal Share

SubConcept SuperConcept

µC×C(c2, c1) µC×C(c5, c1)µC×C(c4, c1)µC×C(c3, c1)

Figure 3. A Fuzzy Domain Ontology from theRCV-1 Corpus

tical learning approaches, the accuracy and the computa-tional ef cienc y of the ontology discovery process is im-proved [20]. The remainder of the paper is organized asfollows. Section 2 highlights previous research in the re-lated area and compare these research work with ours. Sec-tion 3 gives an overview of our text mining methodology.The cognitive and linguistic foundations of the proposedcontext-sensitive ontology discovery method is describedin Section 4. The computational details of the proposedontology mining method are then illustrated in Section 5.Section 6 reports the empirical testing of our fuzzy domainontology mining method. Finally, we offer concluding re-marks and describe future direction of our research work.

2 Related Research

With the increasing importance of product informationmanagement in eCommerce environment, it is vital that pre-cise de nition of product and services readily available insharable, manageable, e xible, and scalable form, that isin the form of an ontology. Although the idea of utiliz-ing ontology for e-Catalogs has been proposed long ago,an operational product ontology system for a speci c do-main is not yet available. Lee et. al. [15] developed anoperational product ontology system called KOCIS for thegovernment procurement service. It consists of the ontol-ogy construction and management sub-system to build theontology database from the product databases and to man-age the the real-time processing for update operations whilemaintaining a consistency of the ontology data. In addi-tion, the ontology search sub-system can retrieves and nav-igates the product ontology information. The search sub-system addresses the problem of ranking keyword searchresults by modeling the product ontology as a Bayesian be-

lief network. Although, the KOCIS system addresses theoperational aspects of an ontology management and searchsystem, it does not support automated or semi-automateddiscovery of product ontology from information sources.This paper focuses on the development of a fuzzy ontol-ogy discovery algorithm for automatic business knowledgemanagement; the proposed method can be readily appliedto discover product ontology for eCommerce.

Cimiano et al. have presented an automatic taxonomylearning algorithm to extract concept hierarchies from a textcorpus [5]. In particular, their taxonomy learning methodis based on formal concept analysis [40]. Formal conceptanalysis is a systematic method for deriving implicit rela-tionships among objects described by a set of attributes.Formal concept analysis can be seen as a conceptual clus-tering techniques at it provides intensional descriptions forthe abstract concepts. Central to formal concept analysisis the notion of a context which is essentially the promi-nent attributes or features common to a set of objects of thesame class. A formal context is a triple K = (G, M, I)where G and M represent a set of objects and attributesrespectively and I is a binary relation between G and M .Thereby, a formal concept (A,B) is de ned by A = {g ∈G|∀m∈M (g,m) ∈ I} and B = {m ∈ M |∀g∈G(g,m) ∈I}. In order to derive attributes from a certain corpus, part-of-speech tagging and linguistic analysis are performed toextract verb/prepositional phrase complement, verb/objectand verb/subject dependencies. For each noun appearing ashead of the extracted syntactic structures, the correspond-ing verbs are taken as the attributes for building the formalcontext. Their approach is evaluated by comparing the au-tomatically generated concept hierarchies with hand-craftedtaxonomies in a tourism and a nance domain. The fuzzyontology discovery method illustrated in this paper employsa novel subsumption based mechanism rather than the for-mal concept analysis approach to generate concept lattice.Semantically richer context vectors are used to representconcepts in our approach as opposed to the simple verb-based features employed by formal concept analysis. In ad-dition, our concept hierarchy represents a fuzzy taxonomyof relations rather than a crisp taxonomy as proposed in [5].

The FOGA framework for fuzzy ontology generationhas been proposed [37]. The FOGA framework consistsof fuzzy formal concept analysis, fuzzy conceptual cluster-ing, fuzzy ontology generation, and semantic representationconversion. Essentially, the FOGA method extends the for-mal concept analysis approach, which has also been appliedto ontology extraction, with the notions of fuzzy sets. Thenotions of formal context and formal concept have beenfuzzi ed by introducing the respective membership func-tions. In addition, an approximate reasoning method is de-veloped so that the automatically generated fuzzy ontologycan be incrementally furnished with the arrival of new in-

IEEE Intelligent Informatics Bulletin November 2007 Vol.8 No.1

Feature Article: Raymond Y.K. Lau 31

stances. The FOGA framework is evaluated in a small cita-tion database. Our method discussed in this paper differsfrom the FOGA framework in that a more compact rep-resentation of fuzzy ontology is developed. The proposedmethod is based on previous work in computational lin-guistic and with the computational mechanism built on theconcept of fuzzy relations. We believe that the proposedmethod is computationally more ef cient and be able toscale up for huge textual databases which typically consistsof millions of records and thousands of terms. Finally, ourproposed method is validated in a standard benchmark tex-tual database which is considerably larger than the citationdatabase used in [37].

A fuzzy ontology which is an extension of the domainontology with crisp concepts is utilized for news summa-rization purpose [14]. In this semi-automatic ontology dis-covery approach, the domain ontology with various eventsof news is pre-de ned by domain experts. A document pre-processing mechanism will generate the meaningful termsbased on the news corpus and a Chinese news dictionarypre-de ned by the domain experts. The meaningful termsare classi ed according to the events of the news by a termclassi er . Basically, every fuzzy concept has a set of mem-bership degrees associated with the various events of thedomain ontology. The main function of the fuzzy inferencemechanism is to generate the membership degrees (classi- cation) for each event with respect to the fuzzy conceptsde ned in the fuzzy ontology. The standard triangular mem-bership function is used for the classi cation purpose. Themethod discussed in this paper is a fully automatic fuzzy do-main ontology discovery approach. There is no pre-de nedfuzzy concepts and taxonomy of concepts, instead our textmining method will automatically discover such conceptsand generate the taxonomy relations. In addition, there isno need to set the arti c ial threshold values for the triangu-lar membership function, instead our membership functioncan automatically derive the membership values based onthe lexico-syntactic and statistical features of the terms ob-served in a textual database.

An ontology mining technique is proposed to extract pat-terns representing users’ information needs [17]. The ontol-ogy mining method consists of two parts: the top backboneand the base backbone. The former represents the relationsbetween compound classes of the ontology. The latter indi-cates the linkage between primitive classes and compoundclasses. The Dempster-Shafer theory of evidence model isadopted to model the relations among classes. The pre-sented method can effectively synthesizing taxonomic rela-tion and non-taxonomic relation in a single ontology model.In addition, a novel method is proposed to capture the evolv-ing patterns in order to re ne the discovered ontology. Fi-nally, a formal model is developed to assess the relevance ofthe discovered ontology with respect to the user’s informa-

tion needs. The ontology mining method is validated basedon the Reuters RCV-1 benchmark collection. The researchwork presented in this paper focuses on fuzzy domain on-tology discovery rather than the discovery of crisp ontologyrepresenting users’ information needs.

Personalized Abstract Search Services (PASS) is a do-main speci c search engine providing abstracts of papersfrom IEEE Transactions sponsored by the IEEE Neural Net-work Council [39]. The system uses a fuzzy ontology ofterm associations to support semantic based information re-trieval. The fuzzy ontology is automatically built using in-formation obtained from the system’s document collection.The system extracts a set of two or three consecutive wordsexhibiting some linguistic patterns such as “noun noun”,“adj noun”, etc. from a corpus. The system then eliminatesthe phrases that contain at least one stop word from a pre-de ned control le. The notions of narrower and broaderterm relations are introduced and a fuzzy conjunction op-erator is applied to compute the membership values of theterm relations. By evaluating the users’ searching activities,it was found that the fuzzy ontology of term relations signif-icantly contributes to the information retrieval process. Ourwork presented in this paper differs from the PASS systemin the fuzzy concepts (instead of terms) are rst identi edand the taxonomy relations of concepts are then developed.In addition, our fuzzy ontology mining approach has beenevaluated based on a bench-mark collection in the eld ofinformation retrieval.

An ontology based text mining system that extracts fuzzyrelations from biological texts is present [1]. This approachpreserves the basic structured knowledge format for storingdomain knowledge, but allows for update of information atthe same time. The document processor parses the text doc-uments and removes the tags pertaining to the biological do-main. The strength of association between a tag pair Ei andEj representing two biological entities is computed accord-ing to a fuzzy conjunction operator. Basically, the member-ship values of the relations are functions of frequency of co-occurrence of concepts. The fuzzy relations between the bi-ological terms are used to guide information retrieval from amedical document collection called GENIA. The ontologydiscovery method presented in this paper deals with generaltextual databases rather than speci cally tagged biologicaldocuments. Concept extraction in our approach is based onthe lexico-syntactic characteristic of tokens appearing in acorpus rather than the pre-de ned semantic of speci c bio-logical tags.

A semiautomatic ontology engineering environmentcalled OntoEdit has been developed [19, 20]. The work-bench supports ontology import, extraction, pruning, re ne-ment, and evaluation. Merging existing semantic structuresor de ning mapping rules between these structures allowsimporting and reusing available ontologies. Ontology ex-

November 2007 Vol.8 No.1 IEEE Intelligent Informatics Bulletin

32 Feature Article: Fuzzy Domain Discovery for Business Knowledge Management

traction is one of the main tasks of ontology engineering,which deals with learning the appropriate ontologies fromthe domain sources. The initial ontology which results fromimport, reuse, and extraction, is then pruned to better t thepurpose of the particular application. Traditional text pro-cessing techniques such as n-gram [30] is used to extend theset of lexical entries L based on source documents. Hierar-chical clustering is applied to learn the taxonomy relationsHC . In addition, morphological analysis and generalizedassociation rule mining are applied to learn the relations Ramong some concepts C. Our work presented in this pa-per focuses on the ontology extraction stage of the ontologyengineering cycle. Moreover, a subsumption-based compu-tational method rather than the traditional clustering methodis used for the extraction of concept lattice.

3 An Overview of the Text Mining Methodol-ogy

Figure 4 depicts the proposed text mining methodologyfor the automatic discovery of fuzzy domain ontology froma textual database (corpus). A text corpus is parsed to ana-lyze the lexico-syntactic elements. For instance, stop wordssuch as “a, an, the” are removed from the source documentssince these words appear in any contexts and they cannotprovide useful information to describe a domain concept.For our implementation, a stop word le is constructedbased on the standard stop word le used in the SMARTretrieval system [29]. Lexical pattern is identi ed by apply-ing Part-of-Speech (POS) tagging to the source documentsand then followed by token stemming based on the Porterstemming algorithm [28]. We refer to the WordNet lexi-con [21] to tag each word during this process. During thelinguistic pattern ltering stage, certain linguistic patternsare extracted based on the speci c requirements speci edby the ontology engineers. For example, the ontology engi-neers may only focus on the “Noun Noun” and “AdjectiveNoun” patterns instead of all the linguistic patterns. Thisis in fact a good way to gain computational ef cienc y byreducing the number of patterns for further statistical anal-ysis. In addition, to extract relevant domain speci c con-cepts, the appearances of concepts across different domainsshould be taken into account. The basic intuition is thata concept frequently appears in a speci c domain (corpus)rather than many different domains is more likely to be a rel-evant domain concept. The statistical Token Analysis stepemploys the information theoretic measure to compute theco-occurrence statistics of the targeting linguistic patterns.Finally, taxonomy of domain concepts is developed accord-ing to the fuzzy conjunction operator. The details of theproposed ontology mining method will be discussed in Sec-tion 5.

Textual database

Stop Word Removal

POS Tagging, Stemming

Linguistic Pattern Filtering

Statistical Token Analysis

Fuzzy Concept Extraction

Fuzzy TaxonomyExtraction

Fuzzy Domain Ontology

Figure 4. Context-Sensitive Fuzzy OntologyDiscovery Process

4 The Linguistic Foundations

The proposed context-sensitive fuzzy ontology discov-ery method is based on the distributional hypothesis whichassumes that terms (concepts) are similar according to theextent that they share similar linguistic contexts [10]. Inparticular, we borrow the notion of collocational expres-sions from computational linguistic to identify the seman-tics of some lexical elements such as concepts from text cor-pora. For computational linguistic, a term refers to one ormore tokens (words) and a term is also a concept if it carriesrecognizable meaning speci c to a domain [23]. Colloca-tional expressions are groups of words related in meaning,and the constituent words of an expression are frequentlyfound in a near loci of a few adjacent words in a textualunit [33, 35]. The collocational expressions are indeed pro-viding the underlying context of a given concept embeddedin natural language text such as Web documents.

Contextual information has long been recognized as oneof the major contributors to concept learning in the eld ofcomputer science [43]. Nevertheless, to automatically de-tect the semantics (meanings) of a concept is not a trivialtask since the meanings of a concept is context (domain)dependent. For example, the concept “bank” can refer to a nancial institute such as a “commercial bank, or refer tothe raised shelf of ground such as the “river bank”. There-fore, to accurately extract domain ontologies from text, con-textual information must be exploited to disambiguate dif-ferent senses. In this regard, static lexicons (i.e., genericlinguistic ontologies) such as WordNet [21] with meanings(senses) computed a priori may not be able to capture thespeci c semantics of concepts pertaining to a particular ap-plication domain. However, WordNet can be used to boot-

IEEE Intelligent Informatics Bulletin November 2007 Vol.8 No.1

Feature Article: Raymond Y.K. Lau 33

strap the performance of information extraction when do-main ontologies are built [22, 24]. Our general approach isthat the collocational expressions are rst extracted from thesource documents; these collocational expressions whichcarry context-sensitive semantics are then used to de ne themeanings of the concepts.

negotiator

0.65

officer

0.86

chief executive

economist

0.52

architect wolitarsky

0.41 0.39

Figure 5. Domain Specific Semantics of theConcept “Chief Executive”

In the eld of information retrieval (IR), the notionof context vectors [11, 32] has been proposed to givecomputer-based representations of concepts. In this ap-proach, a concept is represented by a vector of words andtheir numerical weights. The weight of a word indicatesthe extent to which the particular word is associated withthe underlying concept. For example, the concept “chiefexecutive” is represented by the words such as of cer ,negotiator, economist, etc. as depicted in Figure 5, which isan interesting example by parsing the Reuters-21578 corpus(http://www.daviddlewis.com/resources/testcollections/).The context vector of “chief executive” is shown as follows:

Concept: chief executiveContext Vector:{(officer, 0.72), (economist, 0.65), (negotiator, 0.63),(architect, 0.61), (wolitarsky, 0.44)}

The context vector can be seen as a point in a multi-dimensional geometric information space with each dimen-sion representing a property term. It should be noted thatthe meanings (senses) of “chief executive” is “head of state”or “presidency” as de ned in WordNet [21], which is quitedifferent from that discovered by our context-sensitive textmining method. The last term in the example context vectoris “wolitarsky” which is the name of the chief executive ofa nancial institution often mentioned in the Reuters nan-cial news in that period. So, our method can really discover

domain speci c relation such as “wolitarsky” is a chief ex-ecutive. Static lexicons such as WordNet can only capturethe lexical knowledge of a concept, but fails to represent do-main speci c non-lexical knowledge. A linguistic conceptsuch as “chief executive” can be taken as a class (set) withrespect to the fuzzy set framework. A term such as “woli-tarsky” will then be treated as an object which belongs tothe set with certain degree.

5 Text Mining for Fuzzy Ontology Discovery

It is believed that the main challenge in mining tax-onomy relations from textual databases is to lter outthe noisy relations[18, 20]. Accordingly, our text miningmethod is speci cally designed to deal with such an issue.After standard document pre-processing such as stopword removal, POS tagging, and word stemming [30],a windowing process is conducted over the collection ofdocuments. The windowing process can help reduce thenumber of noisy term relationships. For each document(e.g., Net news, Web page, email, etc.), a virtual windowof δ words is moved from left to right one word at a timeuntil the end of a textual unit (e.g., a sentence) is reached.Within each window, the statistical information amongtokens is collected to develop collocational expressions.Such a windowing process has successfully been applied totext mining before [13]. The windowing process is repeatedfor each document until the entire collection has beenprocessed. According to previous studies, a text windowof 5 to 10 terms is effective [11, 27], and so we adopt thisrange as the basis to perform our windowing process. Toimprove computational ef cienc y and lte r noisy relations,only the speci c linguistic pattern (e.g., Noun Noun, andAdjective Noun) de ned by an ontology engineer will beanalyzed. The following is an example segment of a newsarticle in the Reuters-21578 collection:

<REUTERS OLDID="5545" NEWID="2"><TEXT><TITLE>STANDARD OIL TO FORM FINANCIALUNIT</TITLE><BODY>Standard Oil Co and BP NorthAmerica Inc said they plan to forma venture to manage the money marketborrowing and investment activitiesof both companies.</BODY></TEXT> </REUTERS>

After parsing the main body of the news article, ourontology extraction program will remove the stop words,apply POS tagging and stem the words. So, the result willlook like:

standard (Adj) oil (N) co (N)

November 2007 Vol.8 No.1 IEEE Intelligent Informatics Bulletin

34 Feature Article: Fuzzy Domain Ontology Discovery for Business Knowledge Management

bp (N) north (Adj) america (N)inc (N) said (V) plan (V) form (V)venture (N) manage (V) money (N)market (N) borrow (V) investment (N)activit (N) compan (N) .

Assuming that the window size of 5 is used and the on-tology engineer speci es the “Noun Noun” linguistic pat-tern as the only focus, the potential concepts “Oil Co” and”Co BP” will be extracted from the rst virtual text window.The concept “Oil Co” might be represented by the featuressuch as “standard”, “bp”, and “north”. After parsing thewhole corpus, the statistical data (by statistical token analy-sis) about the potential concepts can be collected. If a wordhas an association weight lower than a pre-de ned thresh-old value, it will be discarded from the context vector of theconcept. This is equivalent to the α-cut operation for fuzzysets.

For statistical token analysis, several information theo-retic methods are employed. Mutual Information has beenapplied to collocational analysis [27, 36] in previous re-search. Mutual Information is an information theoreticmethod to compute the dependency between two entitiesand is de ned by [34]:

MI(ti, tj) = log2

Pr(ti, tj)Pr(ti)Pr(tj)

(1)

where MI(ti, tj) is the mutual information between termti and term tj . Pr(ti, tj) is the joint probability that bothterms appear in a text window, and Pr(ti) is the probabil-ity that a term ti appears in a text window. The probabilityPr(ti) is estimated based on |wt|

|w| where |wt| is the numberof windows containing the term t and |w| is the total numberof windows constructed from a textual database (i.e., a col-lection). Similarly, Pr(ti, tj) is the fraction of the numberof windows containing both terms out of the total numberof windows.

We develop Balanced Mutual Information (BMI) tocompute the degree of association among tokens. Thismethod considers both term presence and term absence asthe evidence of the implicit term relationships.

µci(tj) ≈ BMI(ti, tj)= β(Pr(ti, tj) log2(

Pr(ti,tj)Pr(ti)Pr(tj)

)+

Pr(¬ti,¬tj) log2(Pr(¬ti,¬tj)

Pr(¬ti)Pr(¬tj))) −

(1− β)(Pr(ti,¬tj) log2(Pr(ti,¬tj)

Pr(ti)Pr(¬tj))+

Pr(¬ti, tj) log2(Pr(¬ti,tj)

Pr(¬ti)Pr(tj)))

(2)where µci(tj) is the membership function to estimatethe degree of a term tj ∈ X belonging to a conceptci ∈ C. µci

(tj) is the computational mechanism for

Algorithm FuzzyOntoMine(D, Para, Ont)Input: corpus D and vector of threshold values ParaOutput: a fuzzy domain ontology OntMain Procedure:

1. Ont = {}2. Foreach document d ∈ D Do

(a) Construct text windows w ∈ d

(b) Remove stop words sw from w

(c) Perform POS tagging for each term ti ∈ w

(d) Apply Porter stemming to each term ti

(e) Accumulate the frequency for ti ∈ w and thejoint frequency for any pair ti, tj ∈ w

(f) IF lower ≤ Feq(ti) ≤ upper, X = X ∪ ti

3. End for

4. Foreach term ti ∈ X Do

(a) compute its context vector ci using BMI, MI,JA, CP, KL, or ECH

(b) C = C ∪ ci

5. End for

6. Foreach ci ∈ C Do /* Concept Pruning - α-cut */

(a) IF ∀ti ∈ ci : µci(ti) < α

(b) THEN C = C − ci

7. End for

8. Foreach pair of concepts ci, cj ∈ C Do

(a) Compute the taxonomy relation R(ci, cj) usingSpec(ci, cj)

(b) IF µC×C(ci, cj) > λ, R = R ∪R(ci, cj)

9. End For

10. Foreach R(ci, cj) ∈ R Do /* Taxonomy Pruning */

(a) IF µC×C(ci, cj) < µC×C(cj , ci)

(b) THEN R = R−R(ci, cj)

(c) IF ∃P (ci → cx, . . . , cy → cj)

(d) AND µC×C(ci, cj) ≤ min({µC×C(ci, cx),µC×C(cx, cy), . . . , µC×C(cy, cj)})

(e) THEN R = R−R(ci, cj)

11. End For

12. Output Ont

Figure 6. The Fuzzy Domain Ontology Discov-ery Algorithm

the relation RXC de ned in the fuzzy ontology Ont =<X, C, RXC , RCC >. The membership function µci

(tj) is

IEEE Intelligent Informatics Bulletin November 2007 Vol.8 No.1

Feature Article: Raymond Y.K. Lau 35

indeed approximated by the BMI score. Pr(ti, tj) is thejoint probability that both terms appear in a text window,and Pr(¬ti,¬tj) is the joint probability that both termsare absent in a text window. The weight factor β > 0.5is used to control the relative importance of two kinds ofevidence (positive and negative). In Eq.(2), each MI valueis then normalized by the corresponding joint probabilities.For the special case where Pr(ti, tj) = 1 is true, the jointprobability value is replaced by a large positive integer be-cause terms ti, tj have the strongest association. An α-cutis applied to discard terms from the potential concept if theirmembership values are below the threshold α. After com-puting all the BMI values in a collection, these values aresubject to linear scaling such that each membership valueis within the unit interval ∀ci∈C,tj∈Xµci

(tj) ∈ [0, 1]. Itshould be noted that the constituent terms of a concept arealways belonging to the concept with the maximal mem-bership 1. Other measures that can be used to estimate themembership values of tj ∈ ci include Jaccard (JA), condi-tional probability (CP), Kullback-Leibler divergence (KL),and Expected Cross Entropy (ECH) [12]:

µci(tj) ≈ Jacc(ci, tj)

= Pr(ci∧tj)Pr(ci∨tj)

(3)

µci(tj) ≈ Pr(ci|tj)= Pr(ci,tj)

Pr(tj)

(4)

µci(tj) ≈ KL(ci||tj)

=∑

ci∈C Pr(ci|tj) log2Pr(ci|tj)Pr(ci)

(5)

µci(tj) ≈ ECH(tj , ci)= Pr(tj)

∑ci∈C Pr(ci|tj) log2

Pr(ci|tj)Pr(ci)

(6)

To further lter the noisy concept relations, only the rel-atively prominent concepts for a domain will be further ex-plored. We adopt the TFIDF [30] like heuristic to lter non-relevant domain concepts. Similar approach has also beenused in ontology learning [24]. For example, if a conceptis signi cant for a particular domain, it will appear morefrequently in that domain when compared with its appear-ance in other domains. The following measure is used tocompute the relevance score of a concept:

Rel(ci, Dj) =Dom(ci, Dj)∑nk=1 Dom(c,Dk)

(7)

where Rel(ci, Dj) is the relevance score of a concept ci

in the domain Dj . The term Dom(ci, Dj) is the domainfrequency of the concept ci (i.e., number of documents con-taining the concept divided by the total number of docu-ments in the corpus). The higher the value of Rel(ci, Dj),the more relevant the concept is for domain Dj . Based

on empirical testing, we can estimate a threshold rel for aparticular domain. Only the concepts with relevance scoregreater than the threshold will be selected. For each se-lected concept, its context vector will be expanded based onthe synonymy relation de ned in WordNet [21]. This is infact a smoothing procedure [5]. The intuition is that somewords that belong to a particular concept may not co-occurwith the concept in a corpus. To make our ontology discov-ery method more robust, we need to consider these missingassociations. For instance, our example context vector for“chief executive” will be expanded with the feature “presi-dency” based on the synonymy relation of WordNet, and adefault membership value will be applied to such a term.

The na l stage towards our ontology discovery method isfuzzy taxonomy generation based on subsumption relationsamong extracted concepts. Let Spec(cx, cy) denotes thatconcept cx is a specialization (sub-class) of another conceptcy . The degree of such a specialization is derived by:

µC×C(cx, cy) ≈ Spec(cx, cy)

=∑

tx∈cx,ty∈cy,tx=tyµcx (tx)⊗µcy (ty)∑

tx∈cxµcx (tx)

(8)where ⊗ is a fuzzy conjunction operator which is equiv-alent to the min function. The above formula states thatthe degree of subsumption (speci city) of cx to cy is basedon the ratio of the sum of the minimal membership valuesof the common terms belonging to the two concepts to thesum of the membership values of terms in the concept cx.For instance, if every object of cx is also an object of cy ,a high speci city value will be derived. The Spec(cx, cy)function takes its values from the unit interval [0, 1] andthe subsumption relation is asymmetric. When the tax-onomy is built, we only select the subsumption relationssuch that Spec(cx, cy) > Spec(cy, cx) and Spec(cx, cy) >λ where λ is a threshold to distinguish signi can t sub-sumption relations. The parameter λ is estimated basedon empirical tests. If Spec(cx, cy) = Spec(cy, cx) andSpec(cx, cy) > λ is established, the equivalent relationbetween cx and cy will be extracted. In addition, a prun-ing step is introduced such that the redundant taxonomyrelations are removed. If the membership of a relationµC×C(c1, c2) ≤ min({µC×C(c1, ci), . . . , µC×C(ci, c2)}),where c1, ci, . . . , c2 form a path P from c1 to c2, the relationR(c1, c2) is removed because it can be derived from otherstronger taxonomy relations in the ontology. The fuzzy do-main ontology mining algorithm is summarized and shownin Figure 6.

6 Evaluation

Since one of the most important applications of domainontology is for intelligent information retrieval, our context-

November 2007 Vol.8 No.1 IEEE Intelligent Informatics Bulletin

2

30 Feature Article: Fuzzy Domain Ontology Discovery for Business Knowledge Management

36 Feature Article: Fuzzy Domain Ontology Discovery for Business Knowledge Management

sensitive fuzzy ontology mining method is evaluated withinthe context of information retrieval. Our rst experiment issimilar to the routing tasks used in the Text REtrieval Con-ference (TREC) (http://trec.nist.gov/) which isa well-known international benchmark forum for informa-tion retrieval systems. The Reuters-21578 standard corpuswith the Lewis-Split subset which contains 19,813 docu-ments is used in our experiments. The training set consistsof 13,625 documents and the test set consists of 6,188 doc-uments. Our fuzzy domain ontology is automatically con-structed based on the training set only. It takes 19 minutesonly to complete the ontology mining process on a Pentium-4 2.2GHz PC. In this experiment, a window size of 5, a termsize of 1, a single Noun pattern, and the (BMI) computa-tional method with β = 0.7 are used.

For our ontology extraction method, a concept’s rele-vance score de ned in Eq. 7 is computed with respect toa variety of domains. Therefore, several other corpora areconstructed based on the Web documents retrieved underdifferent Yahoo categories such as “computer”, “entertain-ment”, “education” etc. For the Reuters-21578 corpus, a setof queries are composed based on the pre-de ned Reuterstopics and the top ve (weighted by TFIDF) terms from onerelevant document of the training set. For each Reuters sub-ject code such as “acq”, the corresponding subject descrip-tion such as “acquisitions or mergers” is retrieved from theReuters-21578 category description le. Each query is thenapplied to the testing set and the documents are ranked withrespect to their relevance to the query. The vector-spacemodel [29] is employed in this routing task. For instance,the standard TFIDF term weighting scheme is used to com-pute the term weights of a document and a query respec-tively, and the cosine similarity measure is used to rank eachdocument:

sim(−→q ,−→d ) =

∑ni=1 wq(ki)× wd(ki)√∑n

i=1(wq(ki))2 ×√∑n

i=1(wd(ki))2(9)

where−→q and−→d are the query vector and the document vec-tor respectively. The term wq(ki) represents the weight ofthe ith keyword ki in the query vector −→q , and the termwd(ki) represents the weight of the ith keyword ki in thedocument vector −→d .

The routing tasks are performed with (the experimentalgroup) and without (the control group) the help of our au-tomatically constructed fuzzy domain ontology. Basically,the domain ontology is used for query expansion [41] forthe routing task. For instance, each term in the originalquery is expanded with respect to the domain ontology toobtain a equivalent, a broader, or a more speci c term. Inthis experiment, the type of relations is selected manuallyfrom the fuzzy domain ontology to optimize the retrievaleffectiveness. Standard performance measures [30] such as

precision, recall, and F-measure are then computed basedon the top 100 documents retrieved in both groups:

Precision =a

a + b(10)

Recall = aa+c (11)

Fη =(1 + η2)Precision×Recall

η2Precision + Recall(12)

where a, b, c represent the number of retrieved relevant doc-uments, the number of retrieved non-relevant documents,and the number of not retrieved relevant documents re-spectively. The Fη=1 measure and the recall results of 15randomly selected Reuters topics are depicted in Table 1.The rst column in Table 1 shows the topic names of theReuters-21578 collection; the second column shows thenumber of true relevant documents for each topic. The re-maining two columns are the Fη=1 and the recall resultsachieved when domain ontology is applied to expand initialquery. The last two columns show the Fη=1 and the recall gures when domain ontology is not used for query expan-sion. Except for the topic of “coffee”, the IR performanceis improved with the help of the fuzzy domain ontology forquery expansion. The reason why there is no improvementfor the “coffee” topic is that the automatically generated do-main ontology does not provide additional knowledge to ex-pand the initial query. The difference of IR performance(both F-measure and Recall) between these two groups isstatistically signi can t (p < 0.01) according to a paired onetail t-test. The average improvement of the Fη=1 measure is58.3%. Therefore, we can conclude that the automaticallydiscovered fuzzy domain ontology is with good quality andit is useful for enhancing information retrieval performance.

In our second experiment, various information theoreticmeasures are tested for the purpose of extracting domainconcepts from a corpus. The same routing task is con-ducted except the use of different computational methodssuch as BMI, MI, JA, CP, and KL to estimate the member-ship of a term for a concept. The topic “carcass” is usedto illustrate the typical performance of these methods. Theprecision-recall graph of these runs is plotted in Figure 7.The x axis indicates the various recall levels and the y axisshows the precision values obtained at the corresponding re-call level. For example, the recall level 0.1 indicates the N thposition where 7 relevant documents (there are 68 relevantrecords for this topic) are found from the ranked list, andthe corresponding precision values indicate the retrieval ef-fectiveness of various methods (e.g., the best precision 0.36is achieved by BMI). In general, the higher the precisioncurve, the better performance the information retrieval sys-tem is. As can be seen, the BMI method leads to the bestperformance because it can take into account both positive

IEEE Intelligent Informatics Bulletin November 2007 Vol.8 No.1

Feature Article: Raymond Y.K. Lau 37

indeed approximated by the BMI score. Pr(ti, tj) is thejoint probability that both terms appear in a text window,and Pr(¬ti,¬tj) is the joint probability that both termsare absent in a text window. The weight factor β > 0.5is used to control the relative importance of two kinds ofevidence (positive and negative). In Eq.(2), each MI valueis then normalized by the corresponding joint probabilities.For the special case where Pr(ti, tj) = 1 is true, the jointprobability value is replaced by a large positive integer be-cause terms ti, tj have the strongest association. An α-cutis applied to discard terms from the potential concept if theirmembership values are below the threshold α. After com-puting all the BMI values in a collection, these values aresubject to linear scaling such that each membership valueis within the unit interval ∀ci∈C,tj∈Xµci

(tj) ∈ [0, 1]. Itshould be noted that the constituent terms of a concept arealways belonging to the concept with the maximal mem-bership 1. Other measures that can be used to estimate themembership values of tj ∈ ci include Jaccard (JA), condi-tional probability (CP), Kullback-Leibler divergence (KL),and Expected Cross Entropy (ECH) [12]:

µci(tj) ≈ Jacc(ci, tj)

= Pr(ci∧tj)Pr(ci∨tj)

(3)

µci(tj) ≈ Pr(ci|tj)= Pr(ci,tj)

Pr(tj)

(4)

µci(tj) ≈ KL(ci||tj)

=∑

ci∈C Pr(ci|tj) log2Pr(ci|tj)Pr(ci)

(5)

µci(tj) ≈ ECH(tj , ci)= Pr(tj)

∑ci∈C Pr(ci|tj) log2

Pr(ci|tj)Pr(ci)

(6)

To further lter the noisy concept relations, only the rel-atively prominent concepts for a domain will be further ex-plored. We adopt the TFIDF [30] like heuristic to lter non-relevant domain concepts. Similar approach has also beenused in ontology learning [24]. For example, if a conceptis signi cant for a particular domain, it will appear morefrequently in that domain when compared with its appear-ance in other domains. The following measure is used tocompute the relevance score of a concept:

Rel(ci, Dj) =Dom(ci, Dj)∑nk=1 Dom(c,Dk)

(7)

where Rel(ci, Dj) is the relevance score of a concept ci

in the domain Dj . The term Dom(ci, Dj) is the domainfrequency of the concept ci (i.e., number of documents con-taining the concept divided by the total number of docu-ments in the corpus). The higher the value of Rel(ci, Dj),the more relevant the concept is for domain Dj . Based

on empirical testing, we can estimate a threshold rel for aparticular domain. Only the concepts with relevance scoregreater than the threshold will be selected. For each se-lected concept, its context vector will be expanded based onthe synonymy relation de ned in WordNet [21]. This is infact a smoothing procedure [5]. The intuition is that somewords that belong to a particular concept may not co-occurwith the concept in a corpus. To make our ontology discov-ery method more robust, we need to consider these missingassociations. For instance, our example context vector for“chief executive” will be expanded with the feature “presi-dency” based on the synonymy relation of WordNet, and adefault membership value will be applied to such a term.

The na l stage towards our ontology discovery method isfuzzy taxonomy generation based on subsumption relationsamong extracted concepts. Let Spec(cx, cy) denotes thatconcept cx is a specialization (sub-class) of another conceptcy . The degree of such a specialization is derived by:

µC×C(cx, cy) ≈ Spec(cx, cy)

=∑

tx∈cx,ty∈cy,tx=tyµcx (tx)⊗µcy (ty)∑

tx∈cxµcx (tx)

(8)where ⊗ is a fuzzy conjunction operator which is equiv-alent to the min function. The above formula states thatthe degree of subsumption (speci city) of cx to cy is basedon the ratio of the sum of the minimal membership valuesof the common terms belonging to the two concepts to thesum of the membership values of terms in the concept cx.For instance, if every object of cx is also an object of cy ,a high speci city value will be derived. The Spec(cx, cy)function takes its values from the unit interval [0, 1] andthe subsumption relation is asymmetric. When the tax-onomy is built, we only select the subsumption relationssuch that Spec(cx, cy) > Spec(cy, cx) and Spec(cx, cy) >λ where λ is a threshold to distinguish signi can t sub-sumption relations. The parameter λ is estimated basedon empirical tests. If Spec(cx, cy) = Spec(cy, cx) andSpec(cx, cy) > λ is established, the equivalent relationbetween cx and cy will be extracted. In addition, a prun-ing step is introduced such that the redundant taxonomyrelations are removed. If the membership of a relationµC×C(c1, c2) ≤ min({µC×C(c1, ci), . . . , µC×C(ci, c2)}),where c1, ci, . . . , c2 form a path P from c1 to c2, the relationR(c1, c2) is removed because it can be derived from otherstronger taxonomy relations in the ontology. The fuzzy do-main ontology mining algorithm is summarized and shownin Figure 6.

6 Evaluation

Since one of the most important applications of domainontology is for intelligent information retrieval, our context-

November 2007 Vol.8 No.1 IEEE Intelligent Informatics Bulletin

2 38 Feature Article: Fuzzy Domain Ontology Discovery for Business Knowledge Management

a. Window Size 20 and windowing across sentences b. Window Size 5 and windowing within sentences

Figure 8. The Impact of Windowing and Taxonomy Pruning

too many noisy taxonomy relations exist in the ontologywhich leads to poor query expansion. On the other hand,if the threshold λ is too high, many useful taxonomy rela-tions are ltere d out such that the ontology is not useful forquery re nement. When the threshold λ = 0.001 is used,a noisy ontology will be generated which leads to retrievalperformance worse than the baseline where no ontology isused for query expansion. If an appropriate window size isemployed and the windowing process is carried out withinsentence boundary, a fuzzy domain ontology with higherquality is generated (as depicted in Figure 8.b).

7 Conclusions

The manipulation and exchange of semantically enrichedbusiness intelligence (e.g., products, services, markets, etc.)can enhance the quality of an eCommerce system and offera high level of inter-operability among different enterprisesystems. Ontology certainly plays an important role in theformalization of business knowledge. However, the biggestchallenge for the wide spread applications of ontologies ison the construction of these ontologies because it is a verylabor intensive and time consuming process. As uncertaintyoften presents in real-world applications, it is less likely thatdomain ontologies with crisp concepts and relations can sat-isfy these applications. This paper illustrates a novel fuzzydomain ontology discovery algorithm to facilitate the ontol-ogy engineering process. In particular, contextual informa-tion of a domain is exploited so that higher quality fuzzydomain ontologies can be automatically constructed. Theproposed discovery method combines lexico-syntactic andstatistical learning approaches so as to reduce the chance ofgenerating noisy concepts and relations. Empirical studieshave been performed to evaluate the quality of the fuzzy do-

main ontology discovered by the proposed ontology miningalgorithm. Our preliminary results show that the automat-ically generated fuzzy domain ontology can signi cantlyimprove the effectiveness in information retrieval. Futurework involves comparing the accuracy and the computa-tional ef cienc y of our fuzzy ontology mining method withthat of the other approaches. In addition, larger scale ofquantitative evaluation of our fuzzy ontology mining algo-rithm in the context of business information managementwill be conducted.

References

[1] Muhammad Abulaish and Lipika Dey. Biologicalontology enhancement with fuzzy relations: A text-mining framework. In Andrzej Skowron, RakeshAgrawal, Michael Luck, Takahira Yamaguchi, PierreMorizet-Mahoudeaux, Jiming Liu, and Ning Zhong,editors, Proceedings of the 2005 IEEE / WIC / ACMInternational Conference on Web Intelligence (WI2005), pages 379–385, Compiegne, France, Septem-ber 19–22 2005. IEEE Computer Society.

[2] R. Agrawal and R. Srikant. Fast algorithms for min-ing association rules in large databases. In Jorge B.Bocca, Matthias Jarke, and Carlo Zaniolo, editors,VLDB’94, Proceedings of 20th International Confer-ence on Very Large Data Bases, pages 487–499, San-tiago de Chile, Chile, September 12–15 1994. MorganKaufmann Publishers.

[3] T. Berners-Lee, J. Hendler, and O. Lassila. The se-mantic web. Scienti c American, 284(5):34–43, 2001.

IEEE Intelligent Informatics Bulletin November 2007 Vol.8 No.1

Feature Article: Raymond Y.K. Lau 39

[4] Shan Chen, Damminda Alahakoon, and Maria In-drawan. Background knowledge driven ontology dis-covery. In Proceedings of the 2005 IEEE Interna-tional Conference on e-Technology, e-Commerce ande-Service, pages 202–207, 2005.

[5] P. Cimiano, A. Hotho, and S. Staab. Learning con-cept hierarchies from text corpora using formal con-cept analysis. Journal of Arti cial Intelligence Re-search, 24:305–339, 2005.

[6] The World Wide Web Consortium. WebOntology Language, 2004. Available fromhttp://www.w3.org/2004/OWL/.

[7] Michael Dittenbach, Helmut Berger, and DieterMerkl. Improving domain ontologies by miningsemantics from text. In Proceedings of the FirstAsia-Paci c Conference on Conceptual Modelling(APCCM2004), pages 91–100, 2004.

[8] Mingxia Gao and Chunnian Liu. Extending OWL byfuzzy description logic. In Proceedings of the 17thIEEE International Conference on Tools with Arti- cial Intelligence (ICTAI’05), pages 562–567. IEEEComputer Society, 2005.

[9] T. R. Gruber. A translation approach to portable ontol-ogy speci cations. Knowledge Acquisition, 5(2):199–220, 1993.

[10] Z. Harris. Mathematical Structures of Language. Wi-ley, New York, 1968.

[11] Hongyan Jing and Evelyne Tzoukermann. Informa-tion retrieval based on context distance and morphol-ogy. In Proceedings of the 22nd Annual Interna-tional ACM SIGIR Conference on Research and De-velopment in Information Retrieval, Language Analy-sis, pages 90–96, 1999.

[12] D. Koller and M. Sahami. Hierarchically classify-ing documents using very few words. In Douglas H.Fisher, editor, Proceedings of ICML-97, 14th Interna-tional Conference on Machine Learning, pages 170–178, Nashville, Tennessee, 1997. Morgan KaufmannPublishers, San Francisco, California.

[13] R.Y.K. Lau. Context-Sensitive Text Mining and Be-lief Revision for Intelligent Information Retrieval onthe Web. Web Intelligence and Agent Systems An In-ternational Journal, 1(3-4):1–22, 2003.

[14] Chang-Shing Lee, Zhi-Wei Jian, and Lin-Kai Huang.A fuzzy ontology and its application to news summa-rization. IEEE Transactions on Systems, Man, and Cy-bernetics, Part B, 35(5):859–880, 2005.

[15] Taehee Lee, Ig hoon Lee, Suekyung Lee, Sang gooLee, Dongkyu Kim, Jonghoon Chun, Hyunja Lee, andJunho Shim. Building an operational product ontologysystem. Electronic Commerce Research and Applica-tions, 5(1):16–28, 2006.

[16] D. Lewis, Y. Yang, T. Rose, and F. Li. Rcv1: A newbenchmark collection for text categorization research.Journal of Machine Learning Research, 5:361–397,2004.

[17] Yuefeng Li and Ning Zhong. Mining ontology forautomatically acquiring web user information needs.IEEE Transactions on Knowledge and Data Engineer-ing, 18(4):554–568, 2006.

[18] A. Maedche, V. Pekar, and S. Staab. Ontology learningpart one on discovering taxonomic relations from theweb. In N. Zhong, J. Liu, and Y. Yao, editors, WebIntelligence, pages 3–24. Springer, 2003.

[19] Alexander Maedche and Steffen Staab. Ontologylearning for the semantic web. IEEE Intelligent Sys-tems, 16(2):72–79, 2001.

[20] Alexander Maedche and Steffen Staab. Ontologylearning. In Handbook on Ontologies, pages 173–190.2004.

[21] G. A. Miller, Beckwith R., C. Fellbaum, D. Gross, andK. J. Miller. Introduction to wordnet: An on-line lexi-cal database. Journal of Lexicography, 3(4):234–244,1990.

[22] Michele Missikoff, Roberto Navigli, and Paola Ve-lardi. Integrated approach to web ontology learn-ing and engineering. IEEE Computer, 35(11):60–63,2002.

[23] Christine A. Montgomery. Concept extraction. Amer-ican Journal of Computational Linguistics, 8(2):70–73, 1982.

[24] Roberto Navigli, Paola Velardi, and Aldo Gangemi.Ontology learning and its application to automatedterminology translation. IEEE Intelligent Systems,18(1):22–31, 2003.

[25] I. Nonaka. A dynamic theory of organizational knowl-edge creation. Organization Science, 5(1):14–37,1994.

[26] I. Nonaka and H. Takeuchi. The Knowledge CreatingCompany: How Japanese Companies Create the Dy-namics of Innovation. Oxford University Press, NewYork, 1995.

November 2007 Vol.8 No.1 IEEE Intelligent Informatics Bulletin

28 40 Feature Article: Fuzzy Domain Ontology Discovery for Business Knowledge Management

[27] Patrick Perrin and Frederick Petry. Extraction and rep-resentation of contextual information for knowledgediscovery in texts. Information Sciences, 151:125–152, 2003.

[28] M. Porter. An algorithm for suf x stripping. Program,14(3):130–137, 1980.

[29] G. Salton. Full text information processing usingthe smart system. Database Engineering Bulletin,13(1):2–9, March 1990.

[30] G. Salton and M.J. McGill. Introduction to ModernInformation Retrieval. McGraw-Hill, New York, NewYork, 1983.

[31] M. Sanderson and B. Croft. Deriving concept hierar-chies from text. In Proceedings of the 22nd Annual In-ternational ACM SIGIR Conference on Research andDevelopment in Information Retrieval, pages 206–213. ACM, 1999.

[32] Hinrich Schutze. Automatic word sense discrimina-tion. Computational Linguistics, 24(1):97–124, 1998.

[33] Satoshi Sekine, Jeremy J. Carroll, So a Ananiadou,and Jun’ichi Tsujii. Automatic learning for semanticcollocation. In Proceedings of the third Conferenceon Applied Natural Language Processing, pages 104–110, Trento, Italy, March 31–April 3 1992. Associa-tion for Computational Linguistics.

[34] C. Shannon. A mathematical theory of communica-tion. Bell System Technology Journal, 27:379–423,1948.

[35] G. Smith. Computers and Human Language. OxfordUniversity Press, New York, New York, 1991.

[36] Mark A. Stairmand. Textual context analysis for infor-mation retrieval. In Proceedings of the 20th Annual In-ternational ACM SIGIR Conference on Research andDevelopment in Information Retrieval, pages 140–147, 1997.

[37] Quan Thanh Tho, Siu Cheung Hui, Alvis Cheuk M.Fong, and Tru Hoang Cao. Automatic fuzzy ontol-ogy generation for semantic web. IEEE Transactionson Knowledge and Data Engineering, 18(6):842–856,2006.

[38] Christopher A. Welty. Ontology research. AI Maga-zine, 24(3):11–12, 2003.

[39] Dwi H. Widyantoro and John Yen. A fuzzy ontology-based abstract search engine and its user studies. InThe 10th IEEE International Conference on FuzzySystems, pages 1291–1294. IEEE Press, 2001.

[40] Rudolf Wille. Formal concept analysis as mathemat-ical theory of concepts and concept hierarchies. InBernhard Ganter, Gerd Stumme, and Rudolf Wille, ed-itors, Formal Concept Analysis, Foundations and Ap-plications, volume 3626, pages 1–33. Springer, 2005.

[41] J. Xu and W.B. Croft. Query expansion using local andglobal document analysis. In Hans-Peter Frei, DonnaHarman, Peter Schauble, and Ross Wilkinson, editors,Proceedings of the 19th Annual International ACM SI-GIR Conference on Research and Development in In-formation Retrieval, pages 4–11, Zurich, Switzerland,1996.

[42] L. A. Zadeh. Fuzzy sets. Journal of Information andControl, 8:338–353, 1965.

[43] Wlodek Zadrozny. Context and ontology in under-standing of dialogs. In Proceedings of the IJCAI’95Workshop on Context in NLP, May 15 1995.

IEEE Intelligent Informatics Bulletin November 2007 Vol.8 No.1

Feature Article: Raymond Y.K. Lau 41