ONTOLOGY is a conceptualization of a domain into a

Transcription

1 842 IEEE TRANSACTIONS ON KNOWLEDGE AND DATA ENGINEERING, VOL. 18, NO. 6, JUNE 2006 Automatic Fuzzy Ontology Generation for Semantic Web Quan Thanh Tho, Siu Cheung Hui, Senior Member, IEEE, A.C.M. Fong, Senior Member, IEEE, and Tru Hoang Cao Abstract Ontology is an effective conceptualism commonly used for the Semantic Web. Fuzzy logic can be incorporated to ontology to represent uncertainty information. Typically, fuzzy ontology is generated from a predefined concept hierarchy. However, to construct a concept hierarchy for a certain domain can be a difficult and tedious task. To tackle this problem, this paper proposes the FOGA (Fuzzy Ontology Generation framework) for automatic generation of fuzzy ontology on uncertainty information. The FOGA framework comprises the following components: Fuzzy Formal Concept Analysis, Concept Hierarchy Generation, and Fuzzy Ontology Generation. We also discuss approximating reasoning for incremental enrichment of the ontology with new upcoming data. Finally, a fuzzy-based technique for integrating other attributes of database to the ontology is proposed. Index Terms Intelligent Web services and semantic Web, ontology design, uncertainty, fuzzy, probabilistic, knowledge representation formalisms and methods, concept learning. æ 1 INTRODUCTION ONTOLOGY is a conceptualization of a domain into a human understandable, machine-readable format consisting of entities, attributes, relationships, and axioms [1]. It is used as a standard knowledge representation for the Semantic Web [2]. However, the conceptual formalism supported by typical ontology may not be sufficient to represent uncertainty information commonly found in many application domains due to the lack of clear-cut boundaries between concepts of the domains. For example, a document can be very relevant, relevant, orirrelevant to a research area. In addition, keywords extracted from scientific publications can be used to infer the corresponding research areas. However, it is inappropriate to treat all keywords equally as some keywords may be more significant than others. To tackle this type of problems, one possible solution is to incorporate fuzzy logic [3] into ontology to handle uncertainty data. Traditionally, fuzzy ontology is generated and used in text retrieval [4] and search engines [5], in which membership values are used to evaluate the similarities between the concepts in a concept hierarchy. However, manual generation of fuzzy ontology from a predefined concept hierarchy is a difficult and tedious task that often requires expert interpretation. So, automatic generation of concept hierarchy and fuzzy ontology from uncertainty data of a domain is highly desirable. In this paper, we propose a framework known as FOGA (Fuzzy Ontology Generation framework) that can automatically generate a fuzzy ontology from uncertainty data based on Formal Concept Analysis (FCA) [6] theory. The. Q.T. Tho, S.C. Hui, and A.C.M. Fong are with the School of Computer Engineering, Nanyang Technological University, Blk N4, #2A-32, Singapore {PA B, asschui, acsmfong}@ntu.edu.sg.. T.H. Cao is with the Faculty of Information Technology, Hochiminh City University of Technology, Vietnam. tru@dit.hcmut.edu.vn. Manuscript received 23 Mar. 2005; revised 1 Sept. 2005; accepted 7 Sept. 2005; published online 20 Apr For information on obtaining reprints of this article, please send to: tkde@computer.org, and reference IEEECS Log Number TKDE generated fuzzy ontology is mapped to a semantic representation in OWL (Web Ontology Language) [7]. The rest of this paper is organized as follows: Section 2 discusses related work on ontology generation and FCA. Section 3 gives some basic definitions and operators of the fuzzy theory. The FOGA framework is presented in Section 4. Section 5 discusses the approximating reasoning technique to incrementally furnish the generated ontology with new instance. The problem of integrating extra attributes in database to the ontology is given in Section 6. Performance evaluation of the proposed FOGA framework is given in Section 7. Finally, Section 8 concludes the paper. 2 RELATED WORK 2.1 Ontology Generation Although editing tools [8], [9] have been developed to help users to create and edit ontology, it is a troublesome task to manually derive ontology from data. Typically, ontology can be generated from various data types such as textual data [10], [11], [12], [13], [14], [15], [16], [17], dictionary [18], [19], knowledge-based [20] semistructured schemata [21], [22], [23], and relational schemata [24], [25], [26], [27]. Compared to other types of data, ontology generation from textual data has attracted the most attention. Among techniques used for processing textual data, clustering is one of the most effective techniques for ontology learning. Conceptual clustering techniques such as COBWEB [28] and CLASSIT [29] are powerful clustering techniques that can conceptualize clusters for ontology generation [16], [17]. 2.2 Formal Concept Analysis FCA is a formal technique for data analysis and knowledge presentation. It defines formal contexts to represent relationships between objects and attributes in a domain. From the formal contexts, FCA can then generate formal concepts and interpret the corresponding concept lattice, so that information can be browsed or retrieved effectively /06/$20.00 ß 2006 IEEE Published by the IEEE Computer Society

2 THO ET AL.: AUTOMATIC FUZZY ONTOLOGY GENERATION FOR SEMANTIC WEB 843 FCA is widely used for various applications, such as, text processing [30], [31], [32], ontology merging [33], [34], manager [35], [36], e-learning [37], Web navigation [38], and expert system [39]. However, as most concept lattices are quite complicated in terms of the number of concepts generated, it is necessary to simplify the lattice generated. In the Iceberg concept lattice [40], association rules are typically used for clustering concepts on the lattice. Conceptual scaling [41] or lattice theory [42] is then used to generate the concept hierarchy in the TOSCANA [43] and GALOIS systems [44], respectively. In order to prune the lattices generated for text mining, clustering is first performed on the data set to generate clusters of documents [45], [46]. Then, feature selection is used to extract frequent keywords (or terms) from documents in each cluster as attributes for the cluster. Traditional FCA-based conceptual clustering approaches are hardly able to represent such vague information. To tackle this problem, fuzzy logic can be incorporated into FCA to handle uncertainty information for conceptual clustering and concept hierarchy generation. Pollandt [47], Burusco and Fuentes-González [48], and Huynh and Nakamori [49] have proposed the L-Fuzzy context as an attempt to combine fuzzy logic with FCA. The L-Fuzzy context uses linguistic variables, which are linguistic terms associated with fuzzy sets, to represent uncertainty in the context. However, human interpretation is required to define the linguistic variables. Moreover, the fuzzy concept lattice generated from the L-fuzzy context usually causes a combinatorial explosion of concepts as compared to the traditional concept lattice. We propose a new technique that combines fuzzy logic and FCA as Fuzzy Formal Concept Analysis (FFCA), in which the uncertainty information is directly represented by a real number of membership value in the range of [0,1]. As such, linguistic variables are no longer needed. Compared to the fuzzy concept lattice generated from the L-fuzzy context, the fuzzy concept lattice generated using FFCA will be simpler in terms of the number of formal concepts. It also supports a formal mechanism for calculating concept similarities. 3 FUZZY THEORY In this section, we review some fundamental knowledge of fuzzy theory [3]. Definition 1 (Fuzzy Set). A fuzzy set A on a domain U, is defined by a membership function from U to [0.1], i.e., each item in A has a membership value given by. We denote ðsþ as a fuzzy set generated from a traditional set of items S. Each item in S has a membership value in [0,1]. S can also be called as a crisp set. Definition 2 (Fuzzy Relation). A fuzzy set A on a domain G M, where G and M are two crisp sets is a fuzzy relation on G; M. Definition 3 (Fuzzy Sets Intersection). The intersection of fuzzy sets A and B, denoted as A \ B, is defined by A\B ðxþ ¼minð A ðxþ; B ðxþþ. Definition 4 (Fuzzy Sets Union). The intersection of fuzzy sets A and B, denoted as A [ B, is defined by A[B ðxþ ¼maxð A ðxþ; B ðxþþ: Definition 5 (Fuzzy Set Cardinality). Let S f be a fuzzy set on the domain U. The cardinality of S f is defined as X ¼ ðxþ; S f x2u where ðxþ is the membership of x in S f. Definition 6 (Fuzzy Sets Similarity). The similarity between two fuzzy sets A and B is defined as EðA; BÞ ¼ A \ B A [ B : Definition 7 (Fuzzy Sets Subsethood). The subsethood of a fuzzy set A of a conceptual cluster B is calculated as subsethoodða; BÞ ¼ ja \ Bj : jbj Definition 8 (Fuzzy Set Max-min Composition). Let P ðx; Y Þ be a fuzzy relation on X; Y and P ðy;zþ be a fuzzy relation on Y;Z. The max-min composition of PðX; Y Þ and QðY;ZÞ, denoted as P Q, is defined by: PQ ðx; zþ ¼max y2y minð P ðx; yþ; q ðy; zþþ; 8x 2 X; y 2 Y: The max-min composition indicates the strength of relation between the element of X and Z. 4 THE FOGA FRAMEWORK Fig. 1 shows the proposed FOGA (Fuzzy Ontology Generation framework), which consists of the following components. 4.1 Fuzzy Formal Concept Analysis The Fuzzy Formal Concept Analysis incorporates fuzzy logic into Formal Concept Analysis to represent vague information. Definition 9 (Fuzzy Formal Context). A fuzzy formal context is a triple K ¼ðG; M; I ¼ ðg MÞÞ, where G is a set of objects, M is a set of attributes, and I is a fuzzy set on domain G M. Each relation ðg; mþ 2I has a membership value ðg; mþ in [0,1]. Definition 10 (Fuzzy Representation of Object). Each object O in a fuzzy formal context K can be represented by a fuzzy set ðoþ as ðoþ ¼fA 1 ð 1 Þ;A 2 ð 2 Þ;...;A m ðmþg, where fa 1 ;A 2 ;...;A m g is the set of attributes in K and i is the membership of O with attribute A i in K. ðoþ is called the fuzzy representation of O. Fuzzy formal context can also be represented as a crosstable as shown in Table 1. The context has three objects representing three documents, D 1, D 2, and D 3. It also has

3 844 IEEE TRANSACTIONS ON KNOWLEDGE AND DATA ENGINEERING, VOL. 18, NO. 6, JUNE 2006 Fig. 1. The FOGA framework. TABLE 1 A Cross-Table of a Fuzzy Formal Context TABLE 2 Fuzzy Formal Context in Table 1 with an -cut a ¼ 0:5 three attributes, Data Mining, Clustering, and Fuzzy Logic representing three research topics. The relationship between an object and an attribute is represented by a membership value in [0, 1]. An -cut can be set to eliminate relations that have low membership values. Table 2 shows the cross-table of the fuzzy formal context given in Table 1 with ¼ 0:5. Generally, we can consider the attributes of a formal concept as the description of the concept. Thus, the relationships between the object and the concept should be the intersection of the relationships between the objects and the attributes of the concept. Since each relationship between the object and an attribute is represented as a membership value in fuzzy formal context, the intersection of these membership values should be the minimum of these membership values as in Definition 3. Definition 11 (Fuzzy Formal Concept). Given a fuzzy formal context K ¼ðG; M; IÞ and a confidence threshold T, we define A ¼fm 2 Mj8g 2 A : ðg; mþ Tg for A G and B ¼fg 2 Gj8m 2 B : ðg; mþ Tg for B M. Afuzzy formal concept (or fuzzy concept) of a fuzzy formal context ðg; M; IÞ with a confidence threshold T is a pair ða f ¼ ðaþ;bþ, where A G, B M, A ¼ B, and B ¼ A. Each object g 2 ðaþ has a membership g defined as g ¼ min ðg; mþ; m2b where ðg; mþ = membership value between object g and attribute m defined in I. IfB ¼fg, then g ¼ 1 for every g. A and B are the extent and intent of the formal concept ð ðaþ;bþ, respectively. Note that the fuzzy formal concept in Definition 1 can be considered a special case of a many-valued context [13]. However, our fuzzy-based modification of FCA as presented in Definitions 9 and 11 preserves differently continuous values of objects memberships, which are crucial for calculating concepts similarities. In a formal context, a concept can have many superconcepts and subconcepts. However, the similarities of a concept to its superconcepts and subconcepts are different. Such information cannot be shown in a traditional concept lattice. With fuzzy concept lattice, we can make use of the fuzzy set theory to calculate the similarities between a concept and its subconcepts. Definition 12 (Fuzzy Formal Concept Cardinality). Since the fuzziness of a fuzzy formal concept is represented by membership values of objects of the concept, the cardinality of a fuzzy formal concept K f ¼ð ðaþ;bþ is defined as jk f j¼j ðaþj. Definition 13 (Fuzzy Formal Concept Similarity). The similarity of a fuzzy formal concept K f1 ¼ð ða 1 Þ;B 1 Þ and its subconcept K f2 ¼ð ða 2 Þ;B 2 Þ is defined as EðK f1 ;K f2 Þ¼Eð ða 1 Þ; ða 2 ÞÞ: Fig. 2 shows the traditional concept lattice generated from Table 1 without membership values. Fig. 3 shows the fuzzy concept lattice generated from the fuzzy formal context given in Table 2, in which the similarities between the concepts are given. Fig. 4 shows the L-fuzzy lattice generated by replacing the membership values in Table 2 with the corresponding linguistic values low, medium,

4 THO ET AL.: AUTOMATIC FUZZY ONTOLOGY GENERATION FOR SEMANTIC WEB 845 Fig. 2. A concept lattice generated from traditional FCA. and high given. 1 Clearly, the fuzzy concept lattice is simpler than the L-fuzzy lattice in terms of the number of formal concepts. It can provide additional information, such as membership values of objects in each fuzzy formal concept and similarities of fuzzy formal concepts, which are important for the construction of concept hierarchy. 4.2 Concept Hierarchy Generation Concept Hierarchy Generation clusters the fuzzy concept lattice generated by FFCA to construct a concept hierarchy in the two following steps Fuzzy Conceptual Clustering As in traditional concept lattice, the fuzzy concept lattice generated using FFCA is sometimes quite complicated due to the large number of fuzzy formal concepts generated. Since the formal concepts are generated mathematically, objects that have small differences in terms of attribute values are classified into distinct formal concepts. Such objects should belong to the same concept when they are interpreted by human. Thus, we cluster formal concepts into conceptual clusters using fuzzy conceptual clustering. Compared to traditional clusters, the conceptual clusters generated have the following properties:. Each conceptual cluster is considered as a human interpretable concept in the domain of the fuzzy concept lattice.. Each conceptual cluster is a sublattice extracted from the fuzzy concept lattice.. A formal concept must belong to at least one conceptual cluster. For example, a scientific document can belong to more than one research area. Conceptual clusters are generated based on the premise that if a formal concept A belongs to a conceptual cluster R, then its subconcept B also belongs to R if B is similar to A. We can use a similarity confidence threshold T s to determine whether two concepts are similar or not. Definition 14 (Conceptual Cluster). A conceptual cluster of a concept lattice K with a similarity confidence threshold T s is a sublattice S K of K which has the following properties: Fig. 3. A fuzzy concept lattice generated from FFCA. 2. Any concept C 6¼ C S in S K must have at least one superconcept C 0 2 S K so that EðC;C 0 Þ >T s. Fig. 5a and Fig. 5b show the conceptual clusters generated from the fuzzy concept lattice given in Fig. 3 with similarity confidence thresholds T s ¼ 0:4 and T s ¼ 0:5, respectively. Fig. 6 gives the algorithm that generates conceptual clusters from C S, the starting concept on a fuzzy concept lattice FðKÞ. To generate all conceptual clusters of FðKÞ, we choose C S ¼ supðfðkþþ. A conceptual cluster can be considered as a set of fuzzy formal concept. Each concept is associated with a set of objects and attributes. As such, each conceptual cluster can also be represented as sets of objects and attributes. Moreover, each object in each conceptual cluster should have a membership value implying the uncertainty degree of the fact the object belongs to the conceptual cluster. Therefore, we define some formal definitions as follows: Definition 15 (Object Set of a Conceptual Cluster). Objects belonging to a conceptual cluster C can be represented as a set S O ðcþ ¼[ K2C GðKÞ, where K is a fuzzy formal concept and GðKÞ is the set of objects in K. S O ðcþ is called an object set of C. Definition 16 (Attribute Set of a Conceptual Cluster). Attributes of objects belonging to a conceptual cluster C can be represented as a set S A ðcþ ¼[ K2C MðKÞ, where K is a fuzzy formal concept and MðKÞ is the set of attributes in K. Definition 17 (Membership Value of Objects in Conceptual Cluster). Given a conceptual cluster C and object set S O ðcþ. The membership value of each object O 2 S O ðcþ in C is defined as C O ¼ arg min K2CðKðOÞÞ, where K is a fuzzy formal concept and O ðkþ is the membership value of O in K. 1. S K has a supremum concept C S that is not similar to any of its superconcepts. 1. We use linguistic terms low, medium, and high to replace membership values in the ranges of [0-0.49], [ ], and [0.8-1] in the concept lattice generated in Fig. 4, respectively. Fig. 4. An L-fuzzy concept lattice.

5 846 IEEE TRANSACTIONS ON KNOWLEDGE AND DATA ENGINEERING, VOL. 18, NO. 6, JUNE 2006 Fig. 5. (a) Conceptual clusters generated from Fig. 3 with confidence threshold T s ¼ 0:4. (b) Conceptual clusters generated from Fig. 3 with confidence threshold T s ¼ 0:5. Fig. 6. The fuzzy conceptual clustering algorithm. Thus, each conceptual cluster is associated with a set of objects and attributes as presented in Fig. 7, which correspond to conceptual clusters generated in Fig. 5b. Membership values of objects are also provided. For example, in Fig. 5b, the conceptual cluster CK 1 has two formal concepts C 1 and C 2, in which the object D 2 has membership values as 0.9 and 0.85, respectively. Therefore, the membership value of D 2 in CK 1 is minf0:9; 0:85g ¼0:85. Also, since the conceptual clusters are generated from a fuzzy concept lattice, there should be fuzziness information stored in the lattice. To integrate such fuzziness information Fig. 7. Conceptual clusters with associated object and attribute sets. into the final fuzzy ontology generated, we define fuzzy representation of the conceptual cluster. Definition 18 (Fuzzy Representation of Conceptual Cluster). Given a conceptual cluster C with S O ðcþ ¼fO 1 ;O 2 ;...;O n g; the fuzzy representation of C is a fuzzy set ðcþ as ðcþ ¼ðO 1 Þ\ðO 2 Þ\...\ ðo n Þ Hierarchical Relation Generation As discussed earlier, fuzzy conceptual clustering generates a set of conceptual clusters S C. To construct a concept hierarchy from the conceptual clusters, we need to find the hierarchy relations from the clusters. We first define a concept hierarchy [50] as follows: Definition 19 (Concept Hierarchy). A concept hierarchy is a poset (partially ordered set) ðh;ffþ, where H is a finite set of concepts and ff is a partial order on H. To construct a concept hierarchy satisfying Definition 8, we must construct hierarchical relations among conceptual clusters as a partial order on S C.

6 THO ET AL.: AUTOMATIC FUZZY ONTOLOGY GENERATION FOR SEMANTIC WEB 847 Fig. 8. (a) Conceptual clusters. (b) Concept hierarchy. Definition 20 (Subconcept and Superconcept on a Concept Hierarchy). Let C 1 and C 2 be two conceptual clusters corresponding to two sublattices L 1 and L 2 of a fuzzy concept lattice FðKÞ. Let the fuzzy formal concept I be the supremum of L 1, i.e., I ¼ supðl 1 Þ. C 1 is the subconcept of C 2, denoted as C 1 ffc 2,ifI is the subconcept of any concept C 0 2 L 2 or I C 0, where is the partial order defined on FðKÞ. Equivalently, C 2 is the superconcept of C 1. According to FCA theory, a fuzzy concept lattice FðKÞ is a complete lattice. That is, any fuzzy concept C F on FðKÞ must have at least one superconcept unless C F ¼ supðfðkþþ. Therefore, the definition of superconcept and subconcept relations on conceptual clusters, which is given in Definition 19, assures that each conceptual cluster has at least one superconcept, unless it corresponds to the root node of the concept hierarchy generated. However, we must prove that the ff relation given in Definition 20 is a partial order. Proposition 1. Let C 1 and C 2 be two conceptual clusters corresponding to two sublattices L 1 and L 2 of a fuzzy concept lattice FðKÞ. Let the fuzzy formal concepts I 1 and I 2 be the supremums of L 1 and L 2, respectively. If C 1 ffc 2, then I 1 I 2, where is the partial order defined on FðKÞ. Proof. Since C 1 ffc 2, then 9C 0 2 L 2 such that I 1 C 0. Since I 2 ¼ supðl 2 Þ, therefore C 0 I 2. Since is a partial order, then I 1 I 2. tu From Proposition1, we realize that the ff relation of conceptual clusters is equivalent to the relation of the supremums of the corresponding sublattices. Since the relation is a partial order, the ff relation is a partial order. Fig. 8b illustrates the hierarchical relations constructed from the conceptual clusters given in Fig. 8a. Each concept in the concept hierarchy is represented by a set of its attributes. The supremum and infimum of the lattice are considered as Thing and Nothing concepts, respectively. 4.3 Fuzzy Ontology Generation This step constructs fuzzy ontology from a fuzzy context using the concept hierarchy created by fuzzy conceptual clustering. This is done based on the characteristic that both FCA and ontology support formal definitions of concepts. However, a concept defined in FCA has both extensional and intensional information in a balanced manner, whereas a concept in ontology emphasizes on its intensional aspect. To construct the fuzzy ontology, we need to convert both intensional and extensional information of FCA concepts into the corresponding classes and relations of the ontology. Thus, we define the fuzzy ontology as follows: Definition 21 (Fuzzy Ontology). A fuzzy ontology F O consists of four elements ðc;a C ;R;XÞ, where C represents a set of concepts, A C represents a collection of attributes sets, one for each concept, and R ¼ðR T ;R N Þ represents a set of relationships, which consists of two elements: R N is a set of nontaxonomy relationships and R T is a set of taxonomy relationships. Each concept c i in C represents a set of objects, or instances, of the same kind. Each object o ij of a concept c i can be described by a set of attributes values denoted by A C ðc i Þ. Each relationship r i ðc p ;c q Þ in R represents a binary association between concepts c p and c q, and the instances of such a relationship are pairs of ðc p ;c q Þ concept objects. Each attribute value of an object or relationship instance is associated with a fuzzy membership value between [0,1] implying the uncertainty degree of this attribute value or relationship. X is a set of axioms. Each axiom in X is a constraint on the concept s and relationship s attribute values or a constraint on the relationships between concept objects. The constraints can be described using the SWRL [51] format. For example, the Scholarly Ontology O S ¼ðC;A C ;R;XÞ is a fuzzy ontology where its components are endowed as follows: C = { Document, Research Area } A C ( Document ) = { Name, Author, Title, Keywords, Abstract, Body, Publisher, Publication Date } A C ( Research Area ) = { Name, Keyword } R N = {belong-to( Document, Research Area ), consist-of( Research Area, Document )} R T = {superarea-of( Research Area, Research Area ), subareaof( Research Area, Research Area )} X = {Implies(Antecedent(consist-of(I-variable(x1) I-variable(x2))) Consequent(belong-to(I-variable(x2) I-variable(x1)))) Implies(Antecedent(belong-to(I-variable(x1) I-variable(x2))) Consequent(consist-of(I-variable(x2) I-variable(x1)))) Implies(Antecedent(superarea(I-variable(x1) I-variable(x2))) Consequent(subarea(I-variable(x2) I-variable(x1)))) Implies(Antecedent(subarea(I-variable(x1) I-variable(x2))) Consequent(superarea(I-variable(x2) I-variable(x1))))}

7 848 IEEE TRANSACTIONS ON KNOWLEDGE AND DATA ENGINEERING, VOL. 18, NO. 6, JUNE 2006 Fig. 9. Fuzzy ontology generation process. The Scholarly Ontology has two concepts (classes): Document and Research Area. The attributes of class Document are major properties of a scientific publication such as author, title, etc. A Research Area is indicated by a set of appropriate keywords. Nontaxonomy relation set R N consists of relations between Document and Research Area implying that a document can belong to some research areas and a research area may consist of various scientific documents. Taxonomy relations set R T define hierarchical relations between research area, in which a research area may be a superarea or subarea of the others. The axiom sets X contains of some basic rules that imply the inversed relations of defined relations. As discussed above, the attributes and relations of instances of the Scholarly Ontology are associated with membership values that indicate their degree of uncertainty. For example, a document and a research area can be represented as objects D and R of concepts Document and Research Area, respectively, as follows: R:Name ¼ 00 Concept 1 00 ; R:Keyword ¼f 00 Data Mining 00 ð0:8þ; 00 Clustering 00 ð0:9þg; D:Name ¼ 00 Document 1 00 ; belong-toðd; RÞð0:85Þ: Clearly, every value of the Keyword attribute of the object R is associated with a membership value. Similarly, the relation belong-to between the object R and D also has a membership value. Those membership values are necessary to reflect the attribute value and relation reasonably and accurately since it is difficult to conclude if a keyword or a document fully belongs to a research area. Besides, the membership value of objects attributes and relations that do not contain uncertainty information are implicitly assigned the default value of 1, as the attribute ID of D and R in the above example. Therefore, the defined fuzzy ontology can also be applied for a domain without uncertainty information. The fuzzy ontology generation is illustrated in Fig. 9. Formally, the fuzzy ontology to be generated is defined as a quadruple ðc;a C ;R¼fR T ;R N g;xþ according to Definition 21. It is generated from a fuzzy context K ¼ ðg; M; I ¼ ðg MÞÞ and a concept hierarchy C H ¼ ðh;ffþ is generated from K. Then, the contents of the ontology are furnished as follows Class Mapping Class Mapping furnishes C ¼fE; Ig in which E and I are classes corresponding to extent and intent of the fuzzy context. For example, the extent class mapped from the extent of the fuzzy context given in Table 1 can be labeled manually as Document. We can use appropriate names to represent keyword attributes and use them to label the intent class names as well. For example, the class Research Area can be used to label the initial intent class Taxonomy Relation Generation Taxonomy Relation Generation furnishes R T ¼fsuperclassðI;IÞ; subclassði;iþg: Thus, the hierarchical relations between instances of intent classes are defined. Also, two rules are added to X accordingly:. superclassðx; Y Þ : subclassðy;xþ.. subclassðx; Y Þ : superclassðy;xþ Nontaxonomy Relation Generation Nontaxonomy Relation Generation furnishes R N ¼fR IE ði;eþ;r EI ðe;iþg; in which R EI is the relation between the extent class and intent class. R IE is the reversed relation of R EI. However, we still need to label the nontaxonomy relation. For example, the relation between class Document and class Research Area can be labeled as belong-to, which implies that a document can belong to one or more research areas. Also, two rules are added to X accordingly:. R EI ðx; Y Þ : R IE ðy;xþ.. R IE ðx; Y Þ : R EI ðy;xþ Instances Generation Instances Generation generates instances set I ¼fI I ;I E g, where I I and I E are instances of the intent and extent class. Then, it furnishes membership values for the instances attributes and relationships. It is performed using the following rules:

8 THO ET AL.: AUTOMATIC FUZZY ONTOLOGY GENERATION FOR SEMANTIC WEB 849 Fig. 10. The ontology for the concept hierarchy given in Fig. 7 b represented by OWL. Rule 1 (Intent Instance Generation). 8C 2 H, add a corresponding instance, denoted as I I ðcþ to I I. The attributes values of I I ðcþ are defined as A C ði I ðcþþ ¼ ðcþ. Rule 2 (Extent Instance Generation). 8O 2 G, add an corresponding instance, denoted as I E ðoþ to I I. The attributes values I E ðoþ correspond to attributes values of O defined in K. Rule 3 (Membership Value Generation for Instance Taxonomy Relations). If C 1 C 2 ; 8C 1 ;C 2 2 H, the membership values of the taxonomy relations subclassði I ðc 1 Þ;I I ðc 2 ÞÞ and superclassði I ðc 2 Þ;I I ðc 1 ÞÞ are defined as ðsuperclassði I ðc 2 Þ;I I ðc 1 ÞÞÞ ¼ ðsubclassði I ðc 1 Þ;I I ðc 1 ÞÞÞ ¼ subsethoodððc 1 Þ; ðc 2 ÞÞ: Rule 4 (Membership Value Generation for Instance Nontaxonomy Relations). 8O 2 G; C 2 H;O 2 P O ðcþ, the membership values of the taxonomy relations R IE ði I ðcþ;i E ðoþþ and R EI ði E ðoþ;i I ðhþþ are defined as ðr IE ði I ðcþ;i E ðoþþþ ¼ ðr IE ði I ðcþ;i E ðoþþþ ¼ C O. 4.4 Semantic Representation Conversion The generated fuzzy ontology provides a conceptual model of knowledge in the corresponding domain. However, to make such knowledge accessible and sharable on the Web environment, we must convert it into a semantic representation that can be embedded into the contents of Web pages. In Semantic Web, ontology description language such as OWL can be used to annotate ontology. Therefore, the generated fuzzy ontology can be automatically converted into the corresponding semantic representation in OWL, in which each class and instance is annotated as shown in Fig APPROXIMATING REASONING FOR ONTOLOGY ENRICHMENT A common problem with ontology generation is how to incrementally deal with new data. As discussed earlier, apart from the fuzzy formal context, the fuzzy ontology is also generated from the concept hierarchy acquired using conceptual clustering. To perform conceptual clustering again to incorporate the new data into the ontology would be time-consuming. To avoid this, we propose to use the fuzzy-based approximating reasoning technique to assign new data into appropriate conceptual clusters, which consists of the two following steps. 5.1 Proposition Extraction A proposition can be represented as a statement x is A, where x is a variable and A is a value. In fuzzy logic, a proposition can be represented as a fuzzy set U, which implies x is U. For reasoning, the proposition that has widely been used in fuzzy logic is the IF-THEN proposition, which can be represented as follows: IF <proposition> THEN <proposition>. Assume that we have a conceptual cluster C i that contains a set of objects fo 1 ;O 2 ;...;O n g. Logically, we can construct an IF-THEN proposition as follows: IF x is O 1 or x is O 2 or... or x is O n THEN x belongs to C i. However, since conceptual clusters can share some formal concepts, they can share some objects. Thus, assume that some objects in C i also belong to a conceptual cluster C j, the propositions extracted should

9 850 IEEE TRANSACTIONS ON KNOWLEDGE AND DATA ENGINEERING, VOL. 18, NO. 6, JUNE 2006 be as follows: IF x is O 1 or x is O 2 or... or x is O n THEN x belongs to C i and x may belong to C j. Theoretically, a proposition x is O i, where O i is an object in a fuzzy formal context K, can be represented as P ðo i Þ¼fA 1 ð i1 Þ;A 2 ð i2 Þ;...;A m ðimþg; where fa1;a 2 ;...;A m g is the set of attributes in K and ij is the membership of O i with attribute A j,ink as discussed in Section 3. Thus, the IF proposition IF x is O 1 or x is O 2 or... or x is O n can be represented as a fuzzy set ðo 1 Þ[ðO 2 Þ[...[ ðo n Þ. Since the aim of the IF-THEN proposition is to conclude if an object belongs to conceptual clusters, the THEN proposition should be a fuzzy set F S on the domain D C, where D C is the set of conceptual clusters. The problem is how to calculate the membership for each conceptual cluster C j in D C when we construct the IF-THEN proposition from a certain conceptual cluster C i. Since the membership value implies that If an object belongs to C i, then how much does that object belong to C j, the membership value of C j in the THEN part of the proposition should be the subsethood of ðc i Þ of ðc j Þ. Example 1. In Fig. 5a, two conceptual clusters C K1 and C K2 are generated. Thus, two IF-THEN propositions P 1 and P 2, are obtained, respectively. Since C K1 has two objects D 1 and D 2, the IF part of P 1 can be represented as P ðd 1 Þ[PðD 2 Þ¼f 00 Data Mining 00 ð0:8þ; 00 Fuzzy Logic 00 ð0:61þg [f 00 Data Mining 00 ð0:9þ; 00 Clustering 00 ð0:85þg ¼f 00 Data Mining 00 ð0:9þ; 00 Clustering 00 ð0:85þ; 00 F uzzy Logic 00 ð0:61þg: Similarly, P 2 has the IF part as We have f 00 Data Mining 00 ð0:8þ; 00 F uzzy Logic 00 ð0:87þg: subsethoodððc K1 Þ; ðc K1 ÞÞ ¼ 1; subsethoodððc K1 Þ; ðc K2 ÞÞ ¼ 0:41; subsethoodððc K2 Þ; ðc K1 ÞÞ ¼ 0:42; and subsethoodððc K2 Þ; ðc K2 ÞÞ ¼ 1. Therefore, P 1 represented as a fuzzy rule: IFf 00 Data Mining 00 ð0:9þ; Clusteringð0:85Þ; 00 Fuzzy Logic 00 ð0:61þg THEN fc K1 ð1þ;c K2 ð0:41þg: Similarly, P 2 is represented as IFf 00 Data Mining 00 ð0:8þ; 00 FuzzyLogic 00 ð0:87þg THEN fc K1 ð0:41þ;c K2 ð1þg: 5.2 Approximating Reasoning After the proposition extraction step, we have a set of propositions as fuzzy rules. The next step is to use the generated rules for reasoning new data. For example, assume that we have a fuzzy rule IF x is A THEN y is B, where A and B are fuzzy sets. Then, if we have a new proposition x is A, we need to find what conclusion we can get about y. is Theoretically, a proposition IF <FP1> THEN <FP2>, where FP1 and FP2 are two fuzzy propositions that can be interpreted as a relation connecting FP1 and FP2. In classical propositional logic, the rule IF x THEN y means x implies y and is written as x! y. When x and y are fuzzy propositions, then x! y is a fuzzy relation. There are several definitions for the fuzzy implication relation (! ) listed as follows:. Dienes-Rescher Implication:. Zadeh Implication: x!y ðu; vþ ¼max½1 x ðuþ; y ðvþš: x!y ðu; vþ ¼max½minð x ðuþ; y ðvþþ; 1 x ðuþš:. Godel implication: x!y ðu; vþ ¼ 1 if xðuþ y ðvþ y ðvþ otherwise:. Mamdami Implication: x!y ðu; vþ ¼min½ x ðuþ; y ðvþš:. Larsen Implication: x!y ðu; vþ ¼ x ðuþ: y ðvþ, where x and y are two propositions represented as two fuzzy sets on two domains U and V ; u and v are two objects in U and V ; x ðuþ and y ðvþ are membership values of u and v in x and y, respectively. Using such implications, we can construct the fuzzy relations R between propositions in the IF part and the THEN part in an IF-THEN proposition. Then, we can perform reasoning for new data object O. The conclusion is C such that C ¼ O R. For example, using the Mamdami implication, we can construct fuzzy relations for topic keywords and conceptual clusters from proposition P 1 from 1 as C K1 C K Data Mining 00 0:9 0:41 R 1 ¼ 00 Clustering :85 0:41 5 : 00 F uzzy Logic 00 0:61 0:41 Assume that we have a new document D 4 which can be represented as a fuzzy set of keywords as D 4 ¼f 00 Data Mining 00 ð0:45þ; 00 Clustering 00 ð0:86þ; 00 F uzzy Logic 00 ð0:14þg: The conclusion C D4 1 of D 4 using proposition P 1 with the Mamdami implication is 2 3 0:9 0: C D4 1 ¼ D 4 R 1 ¼½0:45 0:86 0:14Š40:85 0:41 5 0:61 0:41 ¼½0:85 0:41Š ¼fC K1 ð0:85þ;c K2 ð0:41þg: Hence, the conclusion obtained is that the membership grade of D 4 on C K1 is 0.85 and on C K2 is Since we can obtain a distinct proposition from each conceptual cluster, the final

10 THO ET AL.: AUTOMATIC FUZZY ONTOLOGY GENERATION FOR SEMANTIC WEB 851 TABLE 3 A Fuzzy Formal Context Having Cross Relation with the Fuzzy Formal Context in Table 2 TABLE 4 A Concept Context conclusion should be the union of conclusions obtained from the propositions. From the above example, the fuzzy relation R 2 can be constructed from proposition P 2 as R 2 ¼ C K1 C K Data Mining 00 0:42 0:8 00 Clustering :0 0:0 5 : 00 F uzzy Logic 00 0:42 0:87 Therefore, the conclusion C D4 2 given by P 2 is 2 3 0:42 0:8 6 7 C D4 2 ¼ D 4 R 1 ¼½0:45 0:86 0:14Š4 0:0 0:0 5 0:42 0:87 ¼½0:42 0:45Š ¼fC K1 ð0:42þ;c K2 ð0:45þg: The final conclusion C D4 is the union of C D4 1 and C D4 2 such that C D4 ¼fC K1 ð0:85þ;c K2 ð0:41þg [ C K1 ð0:42þ;c K2 ð0:45þg : ¼fC K1 ð0:85þ;c K2 ð0:45þg: Hence, the final decision is that D 4 should belong to C K1 since C K1 is the closest conceptual cluster of D 4. Using other implication relations instead of Mamdami may result in other membership values in the final conclusion. However, the final decision will generally be the same. 6 INTEGRATING OF EXTRA ATTRIBUTES FROM DATABASE TO ONTOLOGY In the previous section, we presented a technique for constructing fuzzy ontology from a fuzzy formal context. Such fuzzy formal context can be generated automatically from database schemata. However, apart from the attributes that are used in the fuzzy formal context, there are probably some other significant attributes available in the database. For example, besides the keywords, a document may have some other important attributes, or extra attributes, such as its authors, publisher, publication dates, etc. To make the generated fuzzy ontology more effective, it is necessary to integrate these extra attributes to the ontology. Thus, we propose a mathematical model to incorporate extra attributes into the fuzzy ontology generated using FOGA. Definition 22 (Cross Relation). The fuzzy formal context K s ¼ðG s ;M s ;I s Þ is said to have a cross relation to a fuzzy formal context K t ¼ðG t ;M t ;I t Þ if M s ¼ G t. The cross relation represents an intercontext relation that probably occurs between the fuzzy formal contexts when the set of objects of a context is regarded as the set of attributes of an other contexts. For example, the context represented by the cross table shown in Table 3 has cross relation with the context in Table 2, while the documents are used as attributes of the authors. The membership value of 1.0 implies that the author is the first author of the document, while 0.5 implies that the author is the second author. To find relations between extra attributes and a concept hierarchy mathematically, we represent the concept hierarchy as a particular fuzzy formal context, or concept context as: Definition 23 (Concept Context). A concept hierarchy C H ¼ ðh;ffþ that is generated from a fuzzy formal context K ¼ ðg; M; IÞ can be represented as a particular fuzzy formal context, or concept context K C ¼ðG C ;M C ;I C Þ, where G C ¼ G, M C ¼ H. Each relation ðg; mþ 2I C has a membership value ðg; mþ ¼ m g. For example, the fuzzy formal context given by Table 4 is the corresponding concept context of the concept hierarchy in Fig. 8b. As discussed above, the concept hierarchy generated from a fuzzy formal context can be represented as a concept context that has the same object set as the original context. Definition 24 (Composed Context). Given a fuzzy formal context K s ¼ðG s ;M s ;I s Þ that has cross relation with a fuzzy formal context K t ¼ðM s ;M t ;I t Þ, the composed context of K s and K t is the fuzzy formal context K st ¼ðG s ;M t ;I st ¼ I s I t Þ. For example, since the fuzzy formal context K C ¼ ðg C ;M C ;I C Þ given in Table 4 has a cross relation with the fuzzy formal context K A ¼ðG A ;M A ;I A Þ given in Table 3, we can generate the composed context K CA ¼ðG CA ;M CA ;I CA Þ, where

11 852 IEEE TRANSACTIONS ON KNOWLEDGE AND DATA ENGINEERING, VOL. 18, NO. 6, JUNE 2006 TABLE 5 The Composed Context between the Fuzzy Formal Contexts Given by Table 4 and Table 3 Fig. 11. An example of concept hierarchy generated :9 0:61 0:61 1:0 0 1:0 6 7 I CA ¼ I C I A : ¼ 4 0:85 0:0 0:0 5 0:5 1:0 0 0:0 0:87 0:87 0:9 0:87 0:87 ¼ : 0:85 0:5 0:5 Table 5 represents the composed context between the contexts represented in Table 4 and Table 3, respectively. In the composed context, relations between authors and research areas have the appropriate membership values that are generated using fuzzy composition as discussed above. As such, relations between an extra attribute and the generated concept hierarchy have been interpreted. Similar approaches can be applied for other extra attributes. Such relations can be employed to discover more advanced knowledge on the ontology. 7 PERFORMANCE EVALUATION 7.1 Generating Ontology from Citation Database To evaluate the proposed FOGA framework for ontology generation, we have collected a set of 1,400 scientific documents on the research area Information Retrieval published in from the Institute for Scientific Information s (ISI) Web site [52]. The downloaded documents are preprocessed to extract related information such as the title, authors, citation keywords, and other citation information. The extracted information is then stored as a citation database. First, we construct a fuzzy formal context K f ¼fG; M; Ig, with G as the set of documents and M as the set of citation keywords. The membership value of a document D on a citation keyword C K in K f is computed as ðd; C K Þ¼ n 1 n 2, where n 1 is the number of documents that cite D and contain C K and n 2 is the number of documents that cite D. This formula is based on the premise that the more frequent a keyword occurs in the citing paper, the more important the keyword is in the cited paper. Then, conceptual clustering is performed from the fuzzy formal context. Each generated conceptual cluster represents a research area. The generated conceptual clusters form a hierarchy of research areas of documents in the Citation Database, or Research Area Hierarchy. Fig. 11 shows a part of the generated Research Area Hierarchy. Each research area is represented by a set of most frequent keywords occurring in the documents that belong to that research area. In FFCA, subareas inherit keywords from their superareas. Note that the inherited keywords are not shown in Fig. 11 when labeling the concepts. Only keywords specific to the concepts are used for labeling. The generated ontology contains scholarly information as a hierarchy of research areas as well as research areas for each document. Taking the advantages of the Semantic Web, such knowledge can be easily shared and reused by other systems for browsing or retrieval. For example, we can use Protégé-2000 [53] for browsing the scholarly ontology. Fig. 12 depicts a part of the generated concept hierarchy of research areas. We use the keyword that has the highest membership value to label the research area. Nevertheless, users can browse more information of each research area. 7.2 Performance Evaluation of Ontology Generation Performance of the ontology generation is evaluated based on the generated Research Area Hierarchy. First, we measure the typical recall, precision, and F-measure [54] to evaluate the clustering results. Second, we use the relaxation error [55] and the corresponding cluster goodness measure to evaluate the goodness of the conceptual clusters generated. We also show whether the use of fuzzy membership instead of crisp value can help improve cluster goodness. Finally, we use the Average Uninterpolated

12 THO ET AL.: AUTOMATIC FUZZY ONTOLOGY GENERATION FOR SEMANTIC WEB 853 Fig. 12. A part of the concept hierarchy of research areas. TABLE 6 Performance Results Using Precision Measurement Precision (AUP) [56], which is a typical measure for evaluating a hierarchical construct, to evaluate the goodness of the generated concept hierarchy Evaluation Using Recall, Precision, and F-Measure We have classified manually the documents downloaded from ISI into classes based on their research themes. These classes are used as a benchmark to evaluate the clustering results in terms of recall, precision, and F-measure. As discussed earlier, we extract citation keywords of documents as their attributes. Since these attributes are descriptors for the generated clusters, if more keywords are extracted and used, the more meaningful the cluster descriptors are constructed. To verify this, we vary the number of keywords N extracted from documents from 2 to 10, and the similarity threshold T s from 0.2 to 0.9 when performing conceptual clustering. The measured precision, recall and F-measure are presented in Table 6, Table 7, and Table 8, respectively. Precision implies accuracy of the clustering results. Table 6 shows that when N is small, the precision is poor. It implies that noisy data in clusters. The precision is improved when the number of extracted keywords is increased. However, this will also cause the recall to decrease as shown in Table 7. When the number of clusters is gradually increased, the efficiency of the clustering results will gradually be decreased. Table 8 shows that when N is low, the F-measure is quite poor. Nevertheless, the F-measure is stable and good when a sufficient number of keywords are extracted. The results also show that the F-measure tends to have the best performance when T s ¼ 0:5. Note that we only measure the clustering results independently without considering clusters hierarchical information. The use of hierarchical information for performance evaluation will be discussed later Evaluation Using Relaxation Error Relaxation error implies dissimilarities of items in a cluster based on attributes values. Since conceptual clustering techniques typically use a set of attributes for concept generation, the relaxation error is quite commonly used for evaluating the goodness of conceptual clusters. The relaxation error RE of a cluster C is defined as REðCÞ ¼ P a2a P n i¼1 P n j¼1 Pðx iþp ðx j Þd a ðx i ;x j Þ, where A is the set of the attributes of items in C, P ðx i Þ is the probability of item x i occurring in C, and d a ðx i ;x j Þ is the distance of x i and x j on attribute a. In this experiment, d a ðx i ;x j Þ¼jmði; aþ mðj; aþj, where mði; aþ and mðj; aþ are the membership values of objects x i and x j on attribute a, respectively. The cluster goodness G of cluster C is defined as GðCÞ ¼1 REðCÞ.

13 854 IEEE TRANSACTIONS ON KNOWLEDGE AND DATA ENGINEERING, VOL. 18, NO. 6, JUNE 2006 TABLE 7 Performance Results Using Recall Measurement TABLE 8 Performance Results Using F-Measure Measurement As discussed earlier, we extract citation keywords of documents as their attributes. Since these attributes are used as descriptors for the generated clusters, we vary the number of keywords extracted to observe the effect of the keyword generated on cluster goodness (see Fig. 13). Besides, since COBWEB is considered as one of the most popular techniques for conceptual clustering, we also apply COBWEB to the citation database to compare the performance. Fig. 14 shows the performance evaluation results on cluster goodness using FFCA and COBWEB, while the number of extracted keywords is varied from 2 to 10. It shows that FFCA achieves better cluster goodness than COBWEB. This demonstrates the advantage of using fuzzy membership values for representing object attributes. In addition, the results also show that improved cluster goodness is obtained when the number of extracted keywords is small. It is expected because a smaller number of keywords used will cause smaller differences in objects in terms of keywords membership values. Therefore, the Fig. 13. Performance evaluation results on cluster goodness. Fig. 14. Performance evaluation on AUP.

14 THO ET AL.: AUTOMATIC FUZZY ONTOLOGY GENERATION FOR SEMANTIC WEB 855 relaxation error will be smaller. However, as we will see later in the following section, a smaller number of extracted keywords will cause poor retrieval performance. 7.3 Evaluation Using Average Uninterpolated Precision The Average Uninterpolated Precision (AUP) is defined as the sum of the precision value at each point (or node) in a hierarchical structure where a relevant item appears, divided by the total number of relevant items. Typically, AUP implies the goodness of a concept hierarchical structure. For evaluating AUP, we have manually classified the downloaded documents into classes based on their research themes. For each class, we extract five most frequent keywords from the documents in the class. Then, we use these keywords as inputs to form retrieval queries and evaluate the retrieval performance using AUP. This is carried out as follows: For each document, we generate a set of document keywords. There are two ways to generate document keywords. The first is to use the set of keywords, known as attribute keywords, from each conceptual cluster as the document keywords. The second is to use the keywords from each document as the document keywords. Then, we vectorize the document keywords and the input query, and calculate the vectors distance for measuring the retrieval performance. We refer to the AUP measured using the first method (i.e., using attribute keywords) as Hierarchical Average Uninterpolated Precision (AUP ðhþ), as each concept inherits attribute keywords from its superconcepts. Whereas the AUP measured using the second method (i.e., using keywords from documents) is referred to as Unconnected Average Uninterpolated Precision (AUP ðuþ). Fig. 14 gives the performance results for PAUPðHÞ and AUP ðuþ using different numbers of extracted keywords N. It shows that when N gets larger, the performance on AUP ðhþ and AUP ðuþ gets better. In addition, the performance on AUP ðhþ is generally better than that of AUP ðuþ. It means that the attribute keywords generated for conceptual clusters are appropriate concepts for representing the concept hierarchical structure. 8 CONCLUSION In this paper, we have proposed the FOGA framework for fuzzy ontology generation on uncertainty information. FOGA consists of the following steps: Fuzzy Formal Concept Analysis, Fuzzy Conceptual Clustering, Fuzzy Ontology Generation, and Semantic Representation Conversion. In addition, we have also proposed an approximating reasoning technique that allows the generated fuzzy ontology to be incrementally furnished with new instances. Finally, we have also proposed a technique to integrate extra attributes in a database to the ontology. The proposed FOGA framework would be useful to construct ontology from uncertainty data as it can represent uncertainty information and construct a concept hierarchy from the uncertainty information in automatically. Apart from constructing scholarly ontology from citation database as previously stated, FOGA has also been used to generate Machine Service Ontology for Semantic Help-desk and Reuters News Topic Themes Ontology in our research. In addition, the scholarly ontology generated in Section 7 has been partially used to construct the Scholarly Semantic Web, a Semantic Web-based information retrieval system to support scholarly activities in the Semantic Web environment. REFERENCES [1] N. Guarino and P. Giaretta, Ontologies and Knowledge Bases: Towards a Terminological Clarification. Toward Very Large Knowledge Bases: Knowledge Building and Knowledge Sharing. Amsterdam: IOS Press, [2] T. Berners-Lee, J. Hendler, and O. Lassila, The Semantic Web, Scientific Am., ners-lee.html, [3] L.A. Zadeh, Fuzzy Logic and Approximate Reasoning, Synthese, vol. 30, pp , [4] M.Z. Islam and L. Brankovic, A Framework for Privacy Preserving Data Mining, Proc. Australasian Workshop Data Mining and Web Intelligence (DMWI 04), pp , [5] D.H. Widyantoro and J. Yen, A Fuzzy Ontology-Based Abstract Search Engine and Its User Studies, Proc. 10th IEEE Int l Conf. Fuzzy Systems, pp , [6] B. Ganter and R. Wille, Formal Concept Analysis: Mathematical Foundations. Springer, [7] W3C, Web Ontology Language Overview, TR/owl-features/, [8] OntoWeb develiverable 1.3, [9] A. Gomez-Perez, O. Corcho, and M. Fernandez-Lopez, Ontological Engineering: With Examples from the Areas of Knowledge Management, e-commerce and the Semantic Web (Advanced Information and Knowledge Processing). Springer, [10] B. Bachimont, A. Isaac, and R. Troncy, Semantic Commitment for Designing Ontologies: A Proposal, Proc. Int l Conf. Knowledge Eng. and Knowledge Management, pp , [11] D.I. Moldovan and R.C. Girju, An Interactive Tool for the Rapid Development of Knowledge Bases, Int l J. Artificial Intelligence Tools (IJAIT), vol. 10, nos. 1-2, [12] A. Maedche and S. Staab, Ontology Learning for the Semantic Web, IEEE Intelligent Systems, vol. 16, no. 2, [13] R. Navigli, P. Velardi, and A. Gangemi, Ontology Learning and Its Application to Automated Terminology Translation, IEEE Intelligent Systems, vol. 18, no. 1, [14] F. Xu, D. Kurz, J. Piskorski, and S. Schmeier, A Domain Adaptive Approach to Automatic Acquisition of Domain Relevant Terms and Their Relations with Bootstrapping, Proc. Third Int l Conf. Language Resources and Evaluation (LREC 2002), [15] L. Khan and F. Luo, Ontology Construction for Information Selection, Proc. 14th IEEE Int l Conf. Tools with Artificial Intelligence, pp , [16] P. Clerkin, P. Cunningham, and C. Hayes, Ontology Discovery for the Semantic Web Using Hierarchical Clustering, Proc. European Conf. Machine Learning (ECML) and European Conf. Principles and Practice of Knowledge Discovery in Databases (ECML/ PKDD-2001), [17] G. Bisson and C. Nedellec, Designing Clustering Methods for Ontology Building: The Mo K Workbench, Proc. Workshop Ontology Learning, 14th European Conf. Artificial Intelligence (ECAI 00), [18] E. Morin, Automatic Acquisition of Semantic Relations between Terms from Technical Corpora, Proc. Fifth Int l Congress on Terminology and Knowledge Eng. (TKE-99), [19] M.A. Hearst, Automatic Acquisition of Hyponyms from Large Text Corpora, Proc. 14th Int l Conf. Computational Linguistic, [20] H. Suryanto and P. Compton, Discovery of Ontologies from Knowledge Bases, Proc. Fifth Int l Conf. Knowledge Capture, pp , [21] A. Deitel, C. Faron, and R. Dieng, Learning Ontologies from RDF Annotations, Proc. Int l Joint Conf. Artifical Intelligence Workshop Ontology Learning, [22] A. Doan, P. Domingos, and A. Levy, Learning Source Descriptions for Data Integration, Proc. Third Int l Workshop the Web and Databases, pp , [23] C. Papatheodorou, A. Vassiliou, and B. Simon, Discovery of Ontologies for Learning Resources Using Word-Based Clustering, Proc. World Conf. Educational Multimedia, Hypermedia, and Telecomm. (ED-MEDIA 02), Aug

15 856 IEEE TRANSACTIONS ON KNOWLEDGE AND DATA ENGINEERING, VOL. 18, NO. 6, JUNE 2006 [24] P. Johannesson, A Method for Transforming Relational Schemas into Conceptual Schemas, Proc. 10th Int l Conf. Data Eng., pp , [25] V. Kashyap, Design and Creation of Ontologies for Environmental Information Retrieval, Proc. 12th Workshop Knowledge Acquisition, [26] D.L. Rubin, M. Hewett, D.E. Oliver, T.E. Klein, and R.B. Altman, Automatic Data Acquisition into Ontologies from Pharmacogenetics Relational Data Sources Using Declarative Object Definitions and XML, Proc. Pacific Symp. Biology, [27] L. Stojanovic, N. Stojanovic, and R. Volz, Migrating Data- Intensive Web Sites into the Semantic Web, Proc. 17th ACM Symp. Applied Computing (SAC), pp , [28] D. Fisher, Knowledge Acquisition via Incremental Conceptual Clustering, Machine Learning, vol. 2, pp , [29] J.H. Gennari, P. Langley, and D. Fisher, Models of Incremental Concept Formation, Machine Learning: Paradigms and Methods, pp , [30] U. Priss, Linguistic Applications of Formal Concept Analysis, Proc. First Int l Conf. Formal Concept Analysis, [31] C. Sporleder, A Galois Lattice Based Approach to Lexical Inheritance Hierarchy Learning, Proc. ECAI 2002 Workshop Machine Learning and Natural Language Processing for Ontology Eng. (OLT 02), [32] W. Petersen, A Set-Theoretical Approach for the Induction of Inheritance Hierarchies, Theoretical Computer Science, vol. 51, [33] B. Ganter and G. Stumme, Creation and Merging of Ontology Top-Levels, Proc. Int l Conf. Computational Science (ICCS 03), pp , [34] G. Stumme and A. Maedche, FCA Merge: Bottom-Up Merging of Ontologies, Proc. 17th Int l Conf. Artificial Intelligence (IJCAI 01), pp , [35] R. Cole and G. Stumme, CEM a Conceptual Manager, Proc. Int l Conf. Computational Science, pp , [36] K. Jones, View Mail Users Manual. com/vm, [37] C. Schmitz, S. Staab, R. Studer, G. Stumme, and J. Tane, Accessing Distributed Learning Repositories through a Courseware Watchdog, Proc. E-Learn 2002 World Conf. E-Learning in Corporate, Government, Healthcare, and Higher Education, pp , [38] R.E. Kent and C. Neuss, Creating a Web Analysis and Visualization Environment, ComputerNetworks and ISDN Systems, vol. 28, nos. 1-2, pp , [39] D. Richards and P. Compton, Combining Formal Concept Analysis and Ripple Down Rules to Support Reuse, Proc. Ninth Int l Conf. Software Eng. and Knowledge Eng. (SEKE 97), [40] G. Stumme, R. Taouil, Y. Bastide, N. Pasquier, and L. Lakhan, Computing Iceberg Concept Lattice with Titanic, J. Knowledge and Data Eng., vol. 42, no. 2, pp , [41] B. Ganter and R. Wille, Conceptual Scaling, Applications of Combinatorics and Graph Theory to the Biological and Social Sciences, pp , [42] G. Birkhoff, Lattice Theory, third ed. Am. Math. Soc., [43] F. Vogt and R. Wille, TOSCANA: A Graphical Tool for Analyzing and Exploring Data, Proc. DIMACS Int l Workshop Graph Drawing (GraphDrawing 94), pp , [44] C. Carpineto and G. Romano, GALOIS: An Order-Theoric Approach to Conceptual Clustering, Proc. Int l Conf. Machine Learning (ICML 93), [45] A. Hotho, S. Staab, and G. Stumme, Explaining Text Clustering Result Using Semantic Structures, Proc. Seventh European Conf. Principles of Data Mining and Knowledge Discovery (PKDD 03), pp , [46] A. Hotho, S. Staab, and G. Stumme, Text Clustering Based on Background Knowledge, technical report, Univ. of Karlsruhe, [47] S. Pollandt, Fuzzy-Begriffe: Formale Begriffsanalyse Unscharfer Daten. Springer Verlag, [48] A. Burusco and R.F. González, The Study of the L-Fuzzy Concept Lattice, Mathware and Soft Computing, vol. 1, no. 3, pp , [49] V.N. Huynh and Y. Nakamori, Fuzzy Concept Formation Based on Context Model, Knowledge-Based Intelligent Information Eng. Systems and Allied Technologies, pp , [50] Y. Lu, Concept Hierarchy in Data Mining: Specification, Generation and Implementation, edu/3024.html, [51] SWRL: A Semantic Web Rule Language Combining OWL and RuleML, [52] ISI, Institute for Scientific Information, [53] The Protégé Ontology Editor and Knowledge Acquisition System, [54] C. VanRijsbergen, Information Retrieval. London: Utterworths, [55] W. Chu and K. Chiang, Abstraction of High Level Concepts from Numerical Values in Databases, Proc. AAAI Workshop Knowledge Discovery in Databases, pp , [56] N. Nanas, V. Uren, and A. deroeck, Building and Applying a Concept Hierarchy Representation of a User Profile, Proc. 26th Ann. Int l ACM SIGIR Conf. Research and Development in Information Retrieval, Quan Thanh Tho received the BE degree in computer science from the Hochiminh City University of Technology, Vietnam, in He is currently a PhD student in the Nanyang Technological University, Singapore. His research interests include semantic Web, ontology generation, data mining, and Web services. Siu Cheung Hui received the BSc degree in mathematics in 1983 and a DPhil degree in computer science in 1987 from the University of Sussex, United Kingdom. He is an associate professor in the School of Computer Engineering at Nanyang Technological University, Singapore. His current research interests include data mining, Internet technology, and multimedia systems. Previously, he worked in IBM China/ Hong Kong as a system engineer from 1987 to He is a senior member of the IEEE and ACM. A.C.M. Fong received his degrees from the University of Auckland and Imperial College, London. He is an associate professor in the School of Computer Engineering at Nanyang Technological University, Singapore. His research interests include various aspects of Internet technology, information processing, multimedia, and communications. He is a senior member of the IEEE, a member of the IEE, and is a Chartered Engineer. Tru Hoang Cao received the BEng degree in computer science and engineering from Ho Chi Minh City University of Technology in 1990, the Meng degree in computer science from the Asian Institute of Technology in 1995, and the PhD degree in computer science from the University of Queensland in He then did postdoctoral research in the Artificial Intelligence Group at the University of Bristol, and the Berkeley Initiative in Soft Computing Group at the University of California at Berkeley. His research interests include semantic Web, soft computing, and conceptual structures. He has served on the program committees of several international conferences, and published more than 45 papers in international journals and conference proceedings in those areas. Dr. Cao is currently a vice dean of the Faculty of Information Technology, Ho Chi Minh City University of Technology, and the chief investigator of the national key project Vietnamese Semantic Web.