composite ontology matching with uncertain mappings recovery

Applied Computing Review 17

Composite Ontology Matchingwith Uncertain Mappings Recovery

Nicola FanizziComputer Science Dept.University of Bari, [email protected]

Claudia d’AmatoComputer Science Dept.University of Bari, Italy

[email protected]

Floriana EspositoComputer Science Dept.University of Bari, Italy

[email protected]

ABSTRACTAn automated ontology matching methodology is presented,supported by various machine learning techniques, as im-plemented in the system MoTo. The methodology is two-tiered. On the first stage it uses a meta-learner to elicitcertain mappings from those predicted by single matchers in-duced by a specific base-learner. Then, uncertain mappingsare recovered passing through a validation process, followedby the aggregation of the individual predictions through lin-guistic quantifiers. Experiments on benchmark ontologiesdemonstrate the effectiveness of the methodology.

Categories and Subject DescriptorsD.2.12 [Interoperability]: Data mapping; I.2.4 [ArtificialIntelligence]: Knowledge Representation Formalisms andMethods; H.3.1 [Content Analysis and Indexing]: Lin-guistic processing; I.2.6 [Artificial Intelligence]: Learning

KeywordsOntology matching, Uncertainty, Validation, Aggregation

1. INTRODUCTIONThe Semantic Web (SW) [3, 30] is the new vision of the Webwhose goal is to make the Web content machine readableand processable besides of human-readable. This view isgrounded on the availability of domain ontologies [18] to beused for semantically annotating resources.

An ontology typically provides a (computer format) vocabu-lary that describes a domain of interest and a specification ofthe meaning of terms used in the vocabulary. The standardlanguage for ontological representation in the SW context isOWL [5] language, which is grounded on Description Logics1

(DLs) [2] as theoretical framework. However, in open andevolving systems, such as the SW, different parties adopt

1Description Logics constitute a fragment of First OrderLogic characterized by a well defined formal semantics anda set of available deductive inference services.

different ontologies. This raises heterogeneity problems at ahigher level [13].

Ontology matching [13] is a promising direction towards thesolution of the semantic heterogeneity problem. It focuseson finding correspondences between semantically related en-tities of different ontologies thus enabling the knowledge anddata expressed in the matched ontologies to interoperate.

Although a variety of automatic ontology matching systemshave been proposed so far, their performance may vary alot depending on the different domains [9]. This problemis generally tackled by selecting the optimal matcher basedon the nature of the matching task and the different fea-tures of the systems. This selection may involve MachineLearning techniques [26] for finding optimal configurationsof the matchers, determining the appropriate heuristics /parameter values to achieve the best results [10].

We propose a comprehensive approach that differs from theprevious ones for exploiting a combination of multiple match-ers which are able to capture diverse aspects of the align-ment. This should allow to overcome the weakness of theindividual matchers. The idea of ensemble learning is induc-ing multiple classifiers (matchers) so that the accuracy oftheir combination (different classifiers can complement andcomplete one another) may lead to a higher performance.

The proposed methodology is made up of two stages. In thefirst stage it uses individual base-learners and then a meta-learner to elicit certain mappings from those predicted bysingle matchers induced by the base-learners. This phaseadopts the stacking [36], an ensemble learning technique,which seems the most appropriate for composing diverselearners. In the second stage, mappings that were previouslydeemed as uncertain can be recovered through a taxonomic /structural validation process, followed by the aggregation ofthe individual predictions made by linguistic quantifiers [17].This stage may be considered as a way of enriching the fea-tures utilized for the choice of the best matchers, in orderto get a more effective combination [9].

The methodology is validated in a realistic setting on bench-mark ontologies. In particular, we use datasets from pastOAEI campaigns2 that provide a gold standard mappingas well as mappings created by different matching systems.Thus, our system can train a classifier on the outcome of

2http://oaei.ontologymatching.org


different matching systems and learn what combination ofresults from different matchers with the best indication ofa correct correspondence. This is competitive w.r.t. previ-ous attempts of combining matchers which have often beenbased on ad hoc methods or had to be customized manually.

The paper proposes the following contributions:

• a hybrid approach for combining various matching sys-tems using machine learning techniques and linguisticaggregation;

• a methodology which is especially meant to recoveruncertain mappings (also called candidate mappings inthe rest of the paper):

– through a number of validators adopting linguis-tic and structural similarity functions [29];

– and an aggregation operator implementing a largechoice of quantifiers [17];

• experiments on OAEI benchmark ontologies prove thatrecovering mappings (through validation and aggrega-tion) can significantly improve the overall performanceof the ontology matching system (especially in termsof recall).

The remainder of the paper is organized as follows. Firstly,the architecture of the MoTo system is presented in Sect. 2.Then in section 3, we illustrate the algorithms used to vali-date and aggregate multiple matchers. Experiments provingthe effectiveness of the method are illustrated in Sect. 4. InSect. 5 related works are examined. Finally, Sect. 6 con-cludes the work outlining possible developments.

2. AN ONTOLOGY MAPPING SYSTEMLet O· = 〈N ·C , N ·R, N ·I〉 denote the input ontology, whereN ·C , N ·R and N ·I stand, respectively, for the sets of thenames for the concepts (classes), the roles (properties) andthe individuals of the ontology. For simplicity, we will con-sider the problem of matching the concepts of some on-tology O1 = 〈N1

C , N1R, N

1I 〉 to those of another ontology

O2 = 〈N2C , N

2R, N

2I 〉. We will focus on the problem of find-

ing equivalence mappings (≡), although the method couldbe extended to discover subsumption mappings (w).

2.1 Application ContextThe reference application context of our hybrid ontology sys-tem is illustrated by Fig. 1. Given two input ontologies un-der comparison O1 and O2, let us suppose that, in order toassess the semantic correspondences among the entities inthe ontologies, the matching system eventually provides asimilarity matrix for the entities belonging to either ontol-ogy. Namely each element of the matrix contains a valuethat indicates the similarity between the couple of entitiesrelated to the row and column. This may have been com-puted through any specific technique which is aimed at de-termining such a value. On the ground of this matrix, a de-cision making module will determine the certain mappings(those whose similarity value exceeds a given threshold ε)from the others. If the system discarded all other mappings,it would likely commit some errors, especially with those

OO11 OO22

mappingmapping

similaritysimilaritymatrixmatrix

certaincertainmappingsmappings

discardeddiscardedmappingsmappings

candidatecandidatemappingsmappings

decision makingdecision making

Figure 1: The reference application context.

mappings that yielded a similarity which was not far fromthe threshold. Hence a number of uncertain mappings (i.e.a sub-matrix of the whole similarity matrix) are retainedas candidate mappings to be evaluated along further tech-niques. To this purpose a taxonomic validator and a struc-tural validator have been devised. They shall be discussedlater in this section. Likewise, aggregation operators canbe applied in the perspective of combining more validators.In such a perspective, the problem of assigning a degree oftrust to each validator arises. In the following, we discussour method for coping with this problem.

2.2 System ArchitectureMoTo (Mapping ontology To ontology) is a multistrategyontology mapping system. It resorts to a variety of matchingtechniques based on machine learning, linguistics and struc-tural criteria. Additionally, MoTo also implements aggre-gation principles for the mappings in order to evaluate thenature of the uncertain mappings, namely if they have to betaken into account or not. Fig. 2 shows its architecture.

The system is among the composite systems3, according tothe classification in [31, 13]. The main functions imple-mented as separate modules are:

• instance classification. All individuals from an ontologyare collected and the membership of each of them w.r.t.each concept of the ontology is determined along withboth a deductive and an inductive approach. Specif-ically, first of all, by the use of a standard deductivereasoner (e.g. pellet reasoner), the instance checking4

for each individual in the ontology is computed withrespect to each concept in the same ontology. Due to

3Ontology matching systems that combine in parallel el-ementary matchers e.g. linguistic matcher, syntacticalmatcher, structural matcher are called composite matchingsystems. They are opposite to the hybrid matching systemswhere elementary matchers are used in sequence.4Instance checking is the inference procedure which assessif it can be proved that an individual is instance of a certainconcept or not.


OO11 OO22

instanceinstanceclassificationclassification


distributiondistributionestimationestimation ML moduleML module

similaritysimilarityestimationestimation similarity functionssimilarity functions

decisiondecisionmakingmaking

uncertainuncertainmappingsmappings

taxonomictaxonomicvalidationvalidation

structuralstructuralvalidationvalidation

contextualcontextualvalidationvalidation

linguisticlinguisticvalidationvalidation

aggregationaggregation

validatedvalidatedmappingsmappings

certaincertainmappingsmappings

discardeddiscardedmappingsmappings

Figure 2: The MoTo system architecture.

the Open World Assumption, typically made in the SWcontext, there could be some unknown cases, namelyindividuals for which it is not possible to prove nei-ther if they are instance of a given concept nor of thenegated concept (used to asses that an individual isnot instance of the concept). In order to manage theseunknown cases, these individuals are inductively clas-sified with respect to each concept in the ontology bythe use of the k-Nearest neighbor algorithm adapted tothe SW representation [14, 4] and the individuals areassigned to the concepts that are majority voted by thetraining examples that are most similar to the individ-uals to classify (the Tversky measure [34] is adoptedas a similarity measure).

• distribution estimation. This module works on the on-tologies to be matched utilizing the services of a sep-arate ML module to estimate the probability distri-bution, related to each couple of concepts (C1, C2) ∈N1C × N2

C , that will be successively used for comput-ing the similarity matrix. The ML module currentlyincludes four base-learners and one meta-learner (im-plementing the technique of stacking of base-classifiersproduced by base-learners) to combine their predic-tions [36]. The available base-learners produce a k-

Nearest Neighbor classifier, an Artificial Neural Net-work and two Bayesian Classifiers, i.e. a content anda name classifier (in this case the information aboutpossible superclasses is also exploited) [35].

This module is used for estimating, for each couple(C1, C2), the joint probability of membership of theindividuals to either concept. Each base-learner L re-ceives a dataset made up of individuals from O1 di-vided in the group of members of C1 and the group ofnon-members of C1. This is exploited to train a base-classifier. Then the individuals inO2 are divided in twogroups according to their membership to C2 ∈ N2

C .The classifier trained on the instances in O1 is thenused to classify the instances in O2. Thus, four in-stance sets are formed, UC1,C2

2 , U¬C1,C22 , UC1,¬C2

2 and

U¬C1,¬C22 , respectively corresponding to the individu-

als that belong to both C1 and C2, the individuals thatdo not belong to C1 but belong to C2 and vice-versaand the individuals that belong neither to C1 nor toC2.

The same procedure is repeated swapping the roles ofthe concepts (and related instances), namely by train-ing each base-learner on the individuals that are in-stances and not instances of C2 and classifying the


individuals that are instances and not instances of C1.Thus, the four instance sets UC1,C2

1 , U¬C1,C21 , UC1,¬C2

1 ,

U¬C1,¬C21 are obtained. The whole procedure can be

iterated for each of the available base-learner. This islikely to lead to different outcomes, i.e. different U -sets.A simple procedure to make a decision in controversialclassification cases may be based on a majority vote,which proves a sub-optimal but often a quite effectivemethod [9]. Alternatively, a more effective techniquesuch as stacking [35] may come into service (for in-stance when the number of the base-learner is even andthere is no majority among the classification resultsof the base-learner). It consists in adopting a meta-learner which produces the final decision on the groundof the determination of the base-learners. The resultsprovided by the ML-module are passed to the distri-bution estimator which computes the joint distribu-tions related to the membership / non-membership to

the two concepts as follows: P (C1, C2) = (|UC1,C21 | +

|UC1,C22 |)/(|N1

I | + |N2I |) where | · | stands for the set

cardinality. The other joint probability distributions(P (¬C1, C2), P (C1,¬C2), P (¬C1,¬C2)) can be de-rived analogously.

• similarity estimation. This module receives the jointprobability distributions computed by the distributionestimation module and exploit them for determiningthe concept similarity. Specifically, the following simi-larity measures can be considered [23, 16]:

– Jaccard(C1, C2) = a/(a+ b+ c)

– Dice(C1, C2) = (b+ c)/(2a+ b+ c)

– Ochiai(C1, C2) = a/√

(a+ b)(a+ c)

– GowerLegendre(C1, C2) = (a− (b+ c) + d)/(a+b+ c+ d)

where a = P (C1, C2), b = P (C1,¬C2), c = P (¬C1, C2)and d = P (¬C1,¬C2).

A similarity matrix S ∈ R|N1C |×|N

2C | is computed where

each element sij ∈ [0, 1] represents the similarity of thecouple of concepts (Ci, Cj) ∈ N1

C ×N2C .

Two thresholds (θmin, θmax) are determined to sep-arate certain {(Ci, Cj) ∈ N1

C × N2C |sij ≥ θmax}, dis-

carded {(Ci, Cj) ∈ N1C×N2

C |sij ≤ θmin} and uncertainmappings {(Ci, Cj) ∈ N1

C × N2C |θmin ≤ sij ≤ θmax}.

The first group can be output, the second one is sim-ply discarded, while the last one is subject to furthercomputation (candidate mappings).

• validation. This module is responsible for the furtherevaluation of uncertain mappings. Specifically, the sim-ilarity of the concepts involved in uncertain mappingis re-computed by the use of different validators eachone grounded of different criteria.

Two types of validators were developed: text-based val-idators that use the vocabulary in conjunction withthe ontological model (it will not be further detailedin the following) and structure-based validators thatuse the taxonomic structure of the input ontologies oralso the entire graph established by the relationships(properties) therein.

• aggregation. The aggregation operator is activated atthe end of the whole process for the composition ofthe results provided by the different validators to givea single decision regarding the uncertain mappings.

3. VALIDATION AND AGGREGATION3.1 Taxonomic and Structural ValidationIn order to compute the similarity between concepts involvedin uncertain mappings, different criteria from the ones usedfor calculating the similarity matrix have to be taken intoaccount. Following [13], five criteria have been consideredfor the comparison of two concepts:#1 most of their direct super-concepts are similar ;#2 most of their direct sub-concepts are similar;#3 most of their super-concepts are similar ;#4 most of their sub-concepts are similar ;#5 all of their related concepts and properties are similar.Criteria #1-#4 merely employ the taxonomic relationshipbetween concepts in the ontology. Validation following thesecriteria can be determined by a taxonomic validator. Thefinal criterion considers also other relationships determinedby the properties in the ontologies. This is implemented ina structural validator.

3.1.1 Taxonomic ValidationThe taxonomic validation module is based on the idea of com-paring two concepts based on their respective position withinthe related subsumption hierarchy. The candidate mappingsfound to be uncertain in the previous phase are input tothis validator to make a decision on their final rejection or apossible recover. By analyzing the taxonomies (parenthoodrelationships between the concepts) the module re-computesa similarity value for the candidate couples of concepts alongthe criteria #1-#4 above.

The taxonomic validation module requires the initial similar-ity matrix St (sub-matrix of the similarity matrix S) con-taining the values that determined certain mappings andthe couples of uncertain mappings that are to be validated.Each concept Ci ∈ O1 is assigned with a unique conceptCj ∈ O2 so that St is transformed in a sparse matrix (since,for each row, only one value is different from zero).

Fig. 3 depicts the algorithm for the n-th criterion. Thisalgorithm computes the similarity of two concepts on theground of the similarity of their (direct) parenthood. More-over, both the observed average similarity and the rate ofmatches found are taken into account. They are carefullycombined since it should be considered that although theaverage similarity may be high the hierarchical structure ofthe ontologies should be also taken into account. This meansthat even in presence of a high similarity value on average,this value could be decreased if the structure of the parent-hood of the two considered concepts is very different (forinstance in the number of concepts that are included).

After computing the similarity of all couples of conceptsfrom either ontology according to simi(C1, C2), i = 1, . . . , 4,the final similarity value given by the validator is a lin-ear combination of these values with an optional additionalterm, which stands for a similarity value computed via an-other function σ covering a different aspect.


simn(C1, C2) : [0, 1]input: C1, C2: concepts; // under comparison

n: integer; // criterion indexoutput: value: [0, 1]; // estimated similaritybeginRC1 ← set of concepts related to C1 according to criterion n;RC2 ← set of concepts related to C2 according to criterion n;nMatches← 0;sum← 0;average← 0;rate← 0;if max(|RC1|, |RC2|) = 0 then return 0;else

for each (Ci, Cj) ∈ RC1 ×RC2 doif St(Ci, Cj) > 0 then

sum← sum + St(Ci, Cj);nMatches← nMatches + 1;

if nMatches = 0 then return 0else

rate← nMatches/max(|RC1|, |RC2|);avg← sum/nMatches;

return α · avg + β · perc;end

Figure 3: Taxonomic similarity algorithm.

3.1.2 Structural ValidationThe structural validation module requires a similarity matrixSs (sub-matrix of the initial similarity matrix S) as input,containing the values that determined the certain mappingsamong the others. This matrix is sparse since each conceptof O1 is mapped to a single concept (having the highestsimilarity value) in O2 (the other elements in the row areset to 0).

The role of the structural validator is to assess a new simi-larity value for the concepts involved in the uncertain map-pings on the bases of the criterion #5 (see [13, 19]). Themeasure employed in this validator estimates the conceptsimilarity through transitive relationships in the respectiveontologies O1 and O2 and it is based on the notion of Infor-mation Content (IC) [29]. Indeed, the adopted measure isa variation (adapted for the context of ontology matching)of the Resnik similarity measure [29]. The main idea of themeasure proposed by Resnik is that the similarity of twoconcepts is given by the variation of their IC with respect tothe IC of their least common superconcept (LCS) that is thefirst parent node in the hierarchy that is common to bothconcepts. The lower is such a variation the higher is thesimilarity value of the concepts.

In our context, the concepts of which we want to assessthe similarity value belong to different ontologies, hence thenotion of LCS cannot be directly used. We generalize thenotion of LCS working on both ontologies. Such a gener-alization is grounded on the notion of related concepts of agiven concept introduced in the following.

Given a couple of concepts (C1, C2) where C1 ∈ O1 andC2 ∈ O2 it is said that (RC1, RC2) is a couple of relatedconcepts w.r.t. (C1, C2) and S iff

1. RC1 ∈ N1C and RC2 ∈ N2

C

2. (RC1, RC2) are corresponding concepts in a mappingdetermined by the input matrix Ss

3. in the concept graph related to the ontologies theremust exist a path from Ci to RCi or viceversa, fori = 1, 2.

• a list representing one such path (whose first ele-ment denotes the direction) is a relation; we willconsider limited relations up to a certain maxlenwith no loops.

4. there is at least a couple of optimal relations connect-ing C1 and C2 through RC1 and RC2 such that:

• the reduced relations, i.e. those obtained by elimi-nating isA links and repeated transitive roles, areequal

• the total length of the relations that satisfy theprevious condition is the minimal total length (de-noted with lensum(·, ·))

Given the set RE(C1, C2) of the couples of related conceptsw.r.t. C1 and C2, the structural similarity measure sims :N1C ×N2

C 7→ R is defined as follows

sims(C1, C2) = w0 · str(C1, C2) + w1 ·map(C1, C2)

where

str(C1, C2) =∑

(RC1,RC2)∈RE(C1,C2)

Ss(RC1, RC2)α · IC(RC1, RC2)

lensum(RC1, RC2)β

with α, β ∈ [0, 1], IC(Ci, Cj) =√

log p(Ci) · log p(Cj) andmap(C1, C2) is a similarity matrix computed by the use ofthe joint probability distributions determined by the relatedmodule described before.

The IC function is generalized considering the estimatedprobability p(C) = (freq(C) +Mµ)/(|N ·C |+M), where theLaplace estimator technique (for some M ∈ N, µ ∈ [0, 1]) isused to avoid values for which the function or its logarithmis undefined and freq(C) returns the number of conceptslinked via a path with C.

3.2 Linguistic and Contextual ValidationThe uncertain mappings can be analyzed adopting a seman-tic viewpoint, based on external resources such as Word-Net [15].

3.2.1 Linguistic ValidationThe linguistic validation module aims at comparing entities(classes and properties, relations) at a linguistic level. En-tities are denoted by names that may differ, e.g., for thewriting conventions or the terms used. Hence, the linguisticvalidation module requires named entities, say E1 and E2,whose similarity has to be assessed. The validation algo-rithm is made of two steps:

• lexical analysis: the name string is examined and di-vided into tokens based on a set of rules which spec-ify how to detect explicit or implicit internal delim-iters (e.g. a change in the capitalization: CompactDisc


yields the tokens [Compact, Disc]); these rules pro-duce the token vectors [t11, . . . , t1n] and [t21, . . . , t2m];

• similarity assessment : the n × m matrix AL = [aij ](linguistic affinity matrix) is obtained from the twovectors using WordNet, with aij = simwn(t1i, t2j) (seebelow). Then the linguistic similarity between the en-tities is computed by:

simL(E1, E2) =

∑mj=1 maxni=1 aij +

∑ni=1 maxmj=1 aij

n+m

The linguistic affinity of two entities exploits the semanticsof the words exploiting their terminological relationships asprovided by WordNet. These serve to create paths that linkthe terms to be compared. The first step finds the set ofpaths Pt1↔t2 between the two terms (otherwise the similar-ity is null). Then, the strength of the paths is measured inorder to select the best one. This is given in terms of theterminological relationships that are involved. Each kind ofrelationship may have its weight so that, say, synonymy isstronger than meronymy. The default weights considered inthe module are:

relationship r weight wr

synonymy 1.0iponymy 0.8

iperonymy 0.8meronymy 0.5

others 0.3

The strength of a path p is computed as the product of theweights of the relationships rk that occur in it. The higheststrength value determines the similarity of the two terms:

simwn(t1, t2) = maxp∈Pt1↔t2

∏rk∈p

wrk

3.2.2 Contextual ValidationThe contextual validation module combines the linguistic in-formation discovered in the previous phase with relationalaspects of the given concepts (C1 and C2) under compari-son. The module uses the two ontologiesO1 andO2 (reducedto a unique model which allows for language-independence)and the linguistic affinity matrix AL previously computedcombined with a relational affinity matrix AR to producea similarity value for the concepts under comparison. Thisnew matrix encodes the kind of relationship holding betweenentities related to the two concepts.

The model the ontologies are reduced to is made up of atriple M = 〈C,P,R〉, where C denotes the set of concepts,P stands for the set of properties (unary relations) of theconcepts, and R contains the (binary) relations betweenconcepts (allowed relations are same-as, kind-of, part-of,generic, with their natural semantics).

In the computation of the similarity, concepts and proper-ties which are directly related to C1 and C2 are considered.Their names are used to compute linguistic affinity matrix,as previously illustrated. The computation of the relationalaffinity matrix AR considers the kind of relation directly

linking the concepts to the other entities in the respectiveontology. The contextual relations are assigned with differ-ent default weights, as follows:

relation weight

property 1.0same-as 1.0kind-of 0.8part-of 0.7generic 0.3

The relational similarity between two relations (properties)r1 and r2, with weights, resp., w1 and w2 is simrel(r1, r2) =1−|w1−w2|. Hence the similarity of all couples of relations(properties) directly related to either concept is computedto fill the relational affinity matrix AR.

Finally, a measure of the contextual similarity is obtained.The contexts of the two concepts are elicited (entities thatare directly related) as represented in the unique model:

c(C1) = 〈e11, . . . , el11 〉, c(C2) = 〈e12, . . . , el22 〉, where each e·· =(s, r), i.e. the name string and the kind of entity involved.The contextual similarity is computed as follows:

simC(C1, C2) =∑

(s1,r1)∈c(C1)(s2,r2)∈c(C2)

AL(s1, s2) ·AR(r1, r2)

l1 · l2

3.3 AggregationSimilarity values computed according to different perspec-tives (matchers) ought to be aggregated to provide a finalvalue. Various methods have been proposed to compute suchaggregate values (see [7, 1, 33]). All require the computa-tion of weights (confidence levels) for the different match-ers implemented. The main problem is then how to de-termine these values automatically (e.g. by recurring to ma-chine learning techniques). One possibility is the usage of anad hoc ensemble learning technique like stacking [36]. Theproblem with these methods is the amount of data requiredfor training the individual and the meta-learner.

We resort to Yager’s OWA (Ordered Weights Aggregation)operators [22].

Definition 1 (OWA operator). Given a n-tuple ofnormalized weights ~w, an OWA operator is a function F :[0, 1]n 7→ [0, 1] such that: F (a1, . . . , an) =

∑ni=1 wibi where

the tuple (b1, . . . , bn) is obtained from (a1, . . . , an) by sortingits elements in descending order.

Note that weights are associated to the positions rather thanto the values.

The aggregation operator is based on a vector of weightscomputed through linguistic quantifiers [17]. A quantifierθ can be represented as Q ⊆ I = [0, 1], where for eachr ∈ I, Q(r) indicates the degree of satisfaction of the crite-rion specified by θ. So if θ = most then Q(.8) = 1 meansthat 80% of the objects are compatible with this criterion.Suppose that n matchers produce n similarity values for the


aggregation(Cube, θ)input: Cube: similarity cube; // m× n× n

θ: quantifier;output: Sc: Rn×n; // final similarity matrixbeginCompute weight vector ~w according to θ;for each candidate mapping (C1, C2) do

beginfor i← 1 to m do

beginSi ← extract(Cube, i);ai ← Si(C1, C2);end

~a← sort(~a); // sort in descending orderS(C1, C2)c ← θ(C1, C2) = Fθ(~a);end

return Sc;end

Figure 5: Aggregation algorithm.

compared concepts to be aggregated (σ1, . . . , σn), the valueof the OWA operator is computed as before: θ(C1, C2) =Fθ(a1, . . . , an) =

∑ni=1 wibi. The weights are determined by

the equations:

wi = Q(i/n)−Q((i− 1)/n) i = 1, . . . , n

Q(r) =

0 r < a(r − a)/(b− a) a ≤ r ≤ b

1 r > ba, b, r ∈ [0, 1]

where a and b are pre-defined thresholds for the single quan-tifiers. Nine quantifiers are implemented in MoTo. Maxsatisfy at least one matcher; Min: satisfy all matchers; Avg :identity – treats all similarity values equally; Most : satisfymost of the matchers; Alh: satisfy at least half of the match-ers; Amap: satisfy as many as possible matchers; Few : sat-isfy few matchers; NH : satisfy nearly half of the matchers;75Perc: satisfy 75% few matchers (similar to Most).

The aggregation process consists of the following steps (seeFigures 4):

1. get the m similarity matrices from the validation mod-ules;

2. reduce them to sparse ones such that each concept C1

in one ontology corresponds to a single C2 in the other;

3. create a similarity cube by composition of such matri-ces;

4. given a selected quantifier θ: compute the final sim-ilarity matrix based on the associated function (seealgorithm illustrated in Fig. 5).

4. EXPERIMENTSThe experiments evaluated the improvement yielded by theadoption of the taxonomic and structural validators withrespect to the performance of the base system. Moreover,we also wanted to find the related best linguistic quantifier.

Table 1: Comparing the taxonomic validator (tv) tothe base system (bs) in terms of precision.

a-n-c k-a-c k-a-n k-n-c k-a-n-cbs tv bs tv bs tv bs tv bs tv

204 1.0 1.0 1.0 1.0 1.0 1.0 1.0 1.0 1.0 1.0

205 .82 .80 .88 .85 .81 .78 .87 .84 .86 .88

206 1.0 1.0 1.0 1.0 1.0 1.0 1.0 1.0 1.0 1.0

221 1.0 1.0 1.0 1.0 1.0 1.0 1.0 1.0 1.0 1.0

222 1.0 1.0 1.0 .91 1.0 .97 1.0 .90 1.0 .96

223 .92 .89 .97 .97 .97 .97 .92 .89 1.0 .94

228 1.0 1.0 1.0 1.0 1.0 1.0 1.0 1.0 1.0 1.0

233 1.0 1.0 1.0 1.0 1.0 1.0 1.0 1.0 1.0 1.0

239 1.0 .94 1.0 .90 1.0 .93 1.0 .90 1.0 1.0

240 .91 .88 .97 .97 .97 .97 .91 .88 1.0 .97

301 .91 .79 .83 .79 .83 .79 .83 .79 1.0 .92

304 .91 .92 .91 .92 .90 .91 .90 .91 .96 .96

4.1 SetupSome of ontologies from the OAEI campaign5 were selected.Specifically the suite was made up of a reference ontology(101) and a number of variants obtained by omitting prop-erties, replacing names with synonyms, changing the graphstructures, etc.. numbered 204, 205, 206, 221, 222, 223, 228,233, 239, 240, 301, 304.

In order to compute the initial mappings, the four learn-ers were set to their default parameters: Bayesian Name (n)and Bayesian Content (c) learners, k-Nearest Neighbor (k);Artificial Neural Network (a). Various combinations of thebase-learners were possible. We will show the best ones, ob-tained by employing a choice of 3 or all four learners. As asimilarity measure for computing the similarity matrix, theJaccard similarity measure has been used.

Concept similarity was computed by the taxonomic validationmodule assuming the values computed by the system front-end (weights .3 and .4). The resulting mappings are selectedwhen above of a threshold of .55. The Structural validationmodule was run with weights .3 and .4, resp., and maxlen= 12. The threshold for the selection of the results was setto .5. As regards the experiments with the linguistic andcontextual validation modules, the similarity assessed usingeither approach was combined with even weights. Finally,the aggregation operator was tested using the presented lin-guistic quantifiers.

As metrics for evaluating the validity of the obtained map-pings, precision, recall and F-measure as defined in [11] havebeen adopted.

4.2 Taxonomic ValidatorIn Tables 1, 2, and 3 the results of the base matching system(bs) are compared to those of the taxonomic validation (tv) interms of precision, recall and F-measure, varying the choiceof basic-learners.

The table 3 shows that tv improves with respect to bs in allbut a couple of ontologies – namely, 221 and 233 – whichhave been obtained by eliminating the hierarchy. The im-provement reaches the 15% for ontologies 223 and 240. Evenmore so, the improvement for 240 has been observed com-bining only 3 base-learners. This was probably due to thefact that the ontologies present a richer taxonomic structurethan the original one. Conversely, for the same reason, slightdecreases were observed for ontologies 222 and 239 because

5http://oaei.ontologymatching.org


final final similaritysimilarity

matrixmatrix

uncertainuncertainmappingsmappings

taxonomictaxonomicvalidationvalidation

structuralstructuralvalidationvalidation

contextualcontextualvalidationvalidation

linguisticlinguisticvalidationvalidation

similaritysimilaritycubecube





Figure 4: Aggregation scheme.

Table 2: Comparing the taxonomic validator (tv) tothe base system (bs) in terms of recall.


204 .81 .83 .83 .83 .86 .86 .81 .83 .83 .83

205 .39 .44 .42 .47 .36 .39 .36 .44 .33 .42

206 .42 .47 .50 .58 .47 .50 .47 .58 .42 .47

221 .39 .39 .44 .44 .42 .42 .42 .42 .67 .67

222 1.0 1.0 .88 .91 .97 .97 .75 .88 .78 .84

223 .64 .86 .97 .97 .97 .97 .64 .86 .81 .89

228 .83 .97 .94 .97 .97 .97 .86 .97 .83 .89

233 .22 .22 .25 .25 .22 .22 .22 .22 .33 .33

239 .88 .91 .66 .88 .72 .88 .78 .84 .63 .69

240 .56 .81 .92 .94 .94 .94 .58 .81 .75 .83

301 .45 .50 .45 .50 .45 .50 .45 .50 .50 .55

304 .68 .74 .68 .74 .61 .65 .58 .68 .71 .71

Table 3: Comparing the taxonomic validator (tv) tothe base system (bs) in terms of F.5-measure.


204 .89 .91 .91 .91 .93 .93 .89 .91 .91 .91

205 .53 .57 .57 .61 .50 .52 .51 .58 .48 .57

206 .59 .64 .67 .74 .64 .67 .64 .74 .59 .64

221 .56 .56 .62 .62 .59 .59 .59 .59 .80 .80

222 1.0 1.0 .93 .91 .98 .97 .86 .89 .88 .90

223 .75 .87 .97 .97 .97 .97 .75 .87 .89 .91

228 .91 .99 .97 .99 .99 .99 .93 .99 .91 .94

233 .36 .36 .40 .40 .36 .36 .36 .36 .50 .50

239 .93 .92 .79 .89 .84 .90 .88 .87 .77 .81

240 .69 .84 .94 .96 .96 .96 .71 .94 .86 .90

301 .61 .61 .59 .61 .59 .61 .59 .61 .67 .69

304 .78 .82 .78 .82 .73 .75 .71 .78 .81 .81

of their poorer taxonomic structures.

In terms of precision and recall the tv generally did not yieldan improvement of the precision, as it may suggest erroneousmappings. Besides, the precision of bs with the ensemble oflearners is already quite high, hence difficult to improve.The improvement is much more evident in terms of recall,as bs is not equally efficient.

4.3 Structural ValidatorTables 4, 5, 6 report the outcomes of the experiments, interms of precision, recall and F-measure, comparing the re-sults of (bs) to those of the structural validation (sv) varyingthe choice of basic-learners.

Table 4: Comparing the structural validator (sv) to thebase system (bs) in terms of precision.

a-n-c k-a-c k-a-n k-n-c k-a-n-cbs sv bs sv bs sv bs sv bs sv

204 1.0 1.0 1.0 1.0 1.0 1.0 1.0 1.0 1.0 1.0

205 .82 .84 .88 .83 .81 .78 .87 .80 .86 .76

206 1.0 1.0 1.0 1.0 1.0 1.0 1.0 1.0 1.0 1.0

221 1.0 .93 1.0 .94 1.0 .94 1.0 .94 1.0 1.0

222 1.0 1.0 1.0 1.0 1.0 1.0 1.0 1.0 1.0 1.0

223 .92 .94 .97 .97 .97 .97 .92 .94 1.0 .97

228 1.0 1.0 1.0 1.0 1.0 1.0 1.0 1.0 1.0 1.0

233 1.0 1.0 1.0 1.0 1.0 1.0 1.0 1.0 1.0 1.0

239 1.0 1.0 1.0 1.0 1.0 1.0 1.0 1.0 1.0 1.0

240 .91 .93 .97 .97 .97 .97 .91 .93 1.0 .96

301 .91 .93 .83 .85 .83 .85 .83 .86 1.0 1.0

304 .91 .93 .91 .93 .90 .92 .90 .92 .96 .96

Table 5: Comparing the structural validator (sv) to thebase system (bs) in terms of recall.


204 .81 .83 .83 .86 .86 .89 .81 .83 .83 .86

205 .39 .44 .42 .42 .36 .39 .36 .44 .33 .36

206 .42 .47 .50 .53 .47 .50 .47 .47 .42 .44

221 .39 .39 .44 .44 .42 .42 .42 .42 .67 .69

222 1.0 1.0 .88 1.0 .97 1.0 .75 1.0 .78 .88

223 .64 .92 .97 .97 .97 .97 .64 .92 .81 .94

228 .83 .94 .94 .97 .97 .97 .86 .94 .83 .86

233 .22 .22 .25 .25 .22 .22 .22 .22 .33 .33

239 .88 .97 .66 .97 .72 .97 .78 .97 .63 .69

240 .56 .75 .92 .94 .94 .94 .58 .75 .75 .75

301 .45 .59 .45 .50 .45 .50 .45 .55 .50 .50

304 .68 .84 .68 .81 .61 .77 .58 .74 .71 .81

Table 6 shows that sv improves the performance of the sys-tem, except for the ontology 233 for the same reasons out-lined before. This was evident especially for ontologies 223,239, and 240 (with a peak of +19% for 239) with the combi-nation k-a-c. In particular, ontology 223 presents an exten-sive hierarchy which helped finding related concepts. A littledecay was observed for ontology 221, one where the taxon-omy was eliminated w.r.t. the original one, with oppositeeffects compared to the mentioned 223.

Again, disaggregating these outcomes in terms of precisionand recall, one observers that precision of bs was alreadyquite high and so difficult to be further improved: a singleerroneously validated candidate mapping may even worsenthe precision. Recall is improved by the validator w.r.t. the


Table 6: Comparing the structural validator (sv) to thebase system (bs) in terms of F.5-measure.


204 .89 .91 .91 .93 .93 .94 .89 .91 .91 .93

205 .53 .58 .57 .56 .50 .52 .51 .57 .48 .49

206 .59 .64 .67 .69 .64 .67 .64 .64 .59 .62

221 .56 .55 .62 .60 .59 .58 .59 .58 .80 .82

222 1.0 1.0 .93 1.0 .98 1.0 .86 1.0 .88 .93

223 .75 .93 .97 .97 .97 .97 .75 .93 .89 .96

228 .91 .97 .97 .99 .99 .99 .93 .97 .91 .93

233 .36 .36 .40 .40 .36 .36 .36 .36 .50 .50

239 .93 .98 .79 .98 .84 .98 .88 .98 .77 .81

240 .69 .83 .94 .96 .96 .96 .71 .83 .86 .84

301 .61 .72 .59 .63 .59 .63 .59 .67 .67 .67

304 .78 .88 .78 .86 .73 .84 .71 .82 .81 .88

Table 7: Comparing the linguistic+contextual validators(lcv) to the base system (bs) in terms of precision.

a-n-c k-a-c k-a-n k-n-c k-a-n-cbs lcv bs lcv bs lcv bs lcv bs lcv

204 1.0 1.0 1.0 1.0 1.0 1.0 1.0 1.0 1.0 1.0

205 .82 .80 .88 .84 .81 .83 .87 .80 .86 .81

206 1.0 1.0 1.0 1.0 1.0 1.0 1.0 1.0 1.0 1.0

221 1.0 1.0 1.0 1.0 1.0 1.0 1.0 1.0 1.0 1.0

222 1.0 1.0 1.0 1.0 1.0 1.0 1.0 1.0 1.0 1.0

223 .92 .94 .97 .97 .97 .97 .92 .94 1.0 1.0

228 1.0 1.0 1.0 1.0 1.0 1.0 1.0 1.0 1.0 1.0

233 1.0 1.0 1.0 1.0 1.0 1.0 1.0 1.0 1.0 1.0

239 1.0 1.0 1.0 1.0 1.0 1.0 1.0 1.0 1.0 1.0

240 .91 .92 .97 .97 .97 .97 .91 .92 1.0 1.0

301 .91 .91 .83 .83 .83 .83 .83 .83 1.0 1.0

304 .91 .88 .91 .89 .90 .93 .90 .87 .96 .89

base system up to a 31% observed on ontology 239 (with thek-a-c combination) although the restriction of the hierarchi-cal structure of this ontology w.r.t. the original one and theelimination of the properties might lead to predict this as adifficult case for sv.

4.4 Linguistic + Contextual ValidatorsAnother experiment evaluated the performance of the othervalidators by combining the decisions made using both thelinguistic validation and contextual validation modules (lcv)and comparing them to those made by base matching system(bs), with the same combinations of basic-learners, in termsof precision, recall and F-measure. Tables 7, 8 and 9 reportthe outcomes of such experiments.

Starting from the evaluation of the F-measure results (seeTable 9), it is evident an improvement for almost all of theontologies (11 out of 12). This is mostly due to an improvedrecall valued obtained using these validators (see Table 8)which are improved in all cases but one (ont. 301) wherefigures remain the same. Indeed the validator could actuallyrecover a good number of uncertain mappings. As regardsthe precision (see Table 7), for almost all ontologies there isno improvement except for some cases in which it is not that

Table 8: Comparing the linguistic+contextual validators(lcv) to the base system (bs) in terms of recall.


204 .81 .83 .83 .83 .86 .86 .81 .83 .83 .83

205 .39 .44 .42 .44 .36 .42 .36 .44 .33 .36

206 .42 .44 .50 .50 .47 .47 .47 .50 .42 .44

221 .39 .78 .44 .69 .42 .58 .42 .69 .67 .69

222 1.0 1.0 .88 1.0 .97 1.0 .75 .94 .78 .85

223 .64 .88 .97 .97 .97 .97 .64 .88 .81 .95

228 .83 .88 .94 1.0 .97 1.0 .86 .88 .83 .86

233 .22 .33 .25 .35 .22 .35 .22 .33 .33 .36

239 .88 .87 .66 .94 .72 .97 .78 .78 .63 .62

240 .56 .61 .92 .97 .94 .97 .58 .71 .75 .78

301 .45 .45 .45 .45 .45 .45 .45 .45 .50 .50

304 .68 .84 .68 .81 .61 .74 .58 .71 .71 .81

Table 9: Comparing the linguistic+contextual validators(lcv) to the base system (bs) in terms of F.5-measure.


204 .89 .91 .91 .91 .93 .93 .89 .91 .91 .91

205 .53 .57 .57 .58 .50 .55 .51 .57 .48 .50

206 .59 .62 .67 .67 .64 .65 .64 .67 .59 .62

221 .56 .87 .62 .83 .59 .58 .59 .74 .80 .83

222 1.0 1.0 .93 1.0 .98 1.0 .86 .96 .88 .93

223 .75 .93 .97 .97 .97 .97 .75 .93 .89 .97

228 .91 .94 .97 1.0 .99 1.0 .93 .94 .91 .94

233 .36 .50 .40 .54 .36 .54 .36 .50 .50 .54

239 .93 .94 .79 .96 .84 .98 .88 .87 .77 .77

240 .69 .83 .94 .96 .96 .96 .71 .83 .86 .87

301 .61 .61 .59 .59 .59 .59 .59 .59 .67 .67

304 .78 .87 .78 .85 .73 .83 .71 .78 .81 .85

Table 10: Performance of the linguistic quantifiers.

Alh Amap Avg Max Min.82 .75 .78 .85 .75

Most 75Perc Few Nh –.77 .78 .82 .81 -

relevant. By comparing the average values of the F-measureover all of the ontologies the neat increase when using thevalidators is evident, especially the cases of test ontologiesobtained by hierarchy variations and synonym replacements.

4.5 Aggregation OperatorFinally, we made experiments for testing bs together withthe various additional components of the system, and es-pecially the aggregation operator. Preliminarily, we testedthe linguistic quantifiers. This produced a choice of the bestquantifier to be utilized for the aggregation operator in thecomparison with the other components (see Tab. 10). Thishas led to selecting the Max quantifier, although also Alh,Few and Nh produced good results.

Then experiments were performed comparing bs, tv, sv, lcvand aggregation (ao). Table 11, reporting the average F-measure outcomes, shows that using ao the system was of-ten able to produce better results. The most problematicontologies were 222, 223, 239 and 301 for which ao lost somecorrect mappings. However, in general, ao produced themaximum average improvement w.r.t. the performance ofbs. As with the previous experiments with the validators,the values for the precision index were difficult to improve.It is important to note that the choice of the Max quantifiersometimes led to erroneous mappings which diminished theoverall performance of ao. The real gain is then in termsof recall with large improvements in some cases and onlytwo cases were a minimal decay was observed (222 and 223)w.r.t. the performance of sv.

5. RELATED WORKThe two main approaches that are used for coping withthe ontology matching problem are: 1) the schema-levelapproach (only the ontology schema is considered); 2) theelement-and-structure-level approach (ontology instances arealso taken into account) [12]. Most of the existing works fo-cus on the schema-level approach and only few of them alsoconsider ontology instances. Indeed, many works have ad-dressed the problem in the context of ontology design andintegration [27, 28, 25]. The initial trend has been to setup elementary matchers copying with a particular aspect


Table 11: Comparing the base system (bs) and thevalidators to the aggregation operator (ao).

bs lcv tv sv ao204 .91 .91 .91 .92 .92

205 .52 .56 .57 .54 .61

206 .63 .64 .68 .65 .69

221 .63 .81 .63 .63 .94

222 .93 .98 .93 .99 .97

223 .87 .95 .92 .95 .94

228 .94 .96 .98 .97 .98

233 .40 .52 .40 .40 .85

239 .84 .91 .88 .95 .89

240 .83 .86 .90 .88 .92

301 .61 .61 .63 .66 .64

304 .76 .83 .80 .86 .88

i.e. linguistic, syntactic etc. Hence, in order to improvethe efficiency of the ontology matching systems, compositematchers have been developed.

In the last years, the idea of combining individual matchershas been pursued by many existing ontology matching ap-proaches (e.g. see [6, 1, 21]). The points of difference regardthe employed features and the way the results of the individ-ual matchers are combined. In contrast to other proposedapproaches that combine a number of specific predefinedclassifiers, our approach is a more general one, as we do notmake any assumptions about the individual matchers to becombined apart from the fact that they provide their resultsin a standardized format.

In the MoTo system there are two levels of combination:one regards the stacking of base-learners which is used to dis-cern certain mappings from the others; the other regards therecovering of uncertain mappings through additional struc-tural and linguistic validation mechanisms whose output canbe finally aggregated by means of specific operators.

In GLUE [8] a meta-learning approach is applied for gen-erating matching hypotheses on the basis of multiple localclassifiers trained on different aspects of the models to bematched. This approach, requires that the input for meta-learning is generated by specific probabilistic learning meth-ods (naıve Bayes classifiers), that are integrated using a lin-ear combination at the meta-level. In our case, we do notmake any assumptions about the matchers used, as it is pos-sible, in principle, to apply any machine learning techniqueat the meta-level. This makes our approach more general.

Besides, the evaluation in [8] is performed on a rather limitedset of ontologies without the existence of commonly agreedreference alignments. As pointed out in [9], the results ofthat meta-learning approach is dominated by the so-calledcontent learner, which always performs almost as good as theintegrated learning approach making the meta-learning stepless important. As reported in the next section, validationand aggregation perform often significantly better than anyof the local matchers.

In [20] kernel machines are used to learn a classifier for map-ping correctness based on a set of simple similarity measures.

The classifier is evaluated on the benchmark datasets of theOAEI and outperforms existing matching systems. The set-ting of the experiments has been criticized in [9], as beingrather unrealistic since all existing reference mappings arethrown together in one large training set and cross-validationis used for computing the accuracy of the classifier. Further,the approach uses only similarity values as input for learn-ing.

In the system described in [32] various Bayesian classifiersare employed. However, the feature set is based on stringdistance measures typically adopted by matching systemsfor determining a syntactic similarity. The experimental re-sults show that there is a strong correlation between differentmeasures and the machine learning approach cannot signif-icantly improve the results of the best individual measure.As shown in [9], using additional and diverse features cansignificantly improve the classification result.

In [24] decision trees and rule learners are used to learnrules for integrating the results of different matching sys-tems. Confidence values respectively measured similaritiesas the basis for learning. The approach has been evaluatedon a subset of the OAEI benchmark dataset but learningfrom confidence values only does not seem to be a goodchoice [9].

As regards the aggregation operator, we have followed theline of COMA [6] and COMA++ [1] implementing morelinguistic quantifiers. The CMC system [33] was meant toovercome the limitations of LSD and COMA by predictingthe credibility of the individual matchers and then combin-ing (couplewise) the resulting similarity matrices by meansof the same method adopted by COMA.

6. CONCLUSIONSThis work focused on ontology matching validation based onstructural aspects on the ontologies. It also concerned theaggregation of the similarities computed by different match-ers, that are able to reconcile the various aspects targeted byeach matcher through many linguistic quantifiers. The ex-perimentation demonstrates that a combination of differenttechniques yields some added value in the quality of map-pings found. Specifically, the taxonomic validator provedits effectiveness despite of the simplicity of the underlyingidea. Another point in favor is that it can be applied toany kind of ontology being focused on the taxonomic aspectonly. The structural validator goes a step even further as itexploits also other relationships between the concepts. Theaggregation operator can select the best mappings from eachsystem component and allows different types of aggregationby changing the most appropriate quantifier.

We are currently planning an enhancement of the validatorsso that other criteria may be implemented. The structuralvalidator may also be enhanced by taking into account an-notations/comments in natural language. A deeper inves-tigation of the application of ensemble machine methods isalso necessary. This may affect also the choice of weights forthe aggregation operator.


7. ACKNOWLEDGMENTSThe authors wish to thank C. Leone, M. Tempesta andC. Terlizzi who designed and implemented the system MoTo.

8. REFERENCES[1] D. Aumueller, H.H. Do, S. Massmann, and E. Rahm.

Schema and ontology matching with COMA++. InProceedings of the ACM SIGMOD InternationalConference on Management of Data, pages 906–908,2005.

[2] F. Baader, D. Calvanese, D. McGuinness, D. Nardi,and P. Patel-Schneider, editors. The Description LogicHandbook. Cambridge University Press, 2003.

[3] T. Berners-Lee, J. Hendler, and O. Lassila. Thesemantic web. Scientific American Magazine, 2001.

[4] C. d’Amato, N. Fanizzi, and F. Esposito. Queryanswering and ontology population: An inductiveapproach. In S. Bechhofer et al., editors, Proceedingsof the 5th European Semantic Web Conference,ESWC2008, volume 5021 of LNCS, pages 288–302.Springer, 2008.

[5] M. Dean and G. Schreiber (eds.). Web OntologyLanguage Reference, 2004.http://www.w3.org/TR/owl-ref.

[6] H.H. Do and E. Rahm. COMA - a system for flexiblecombination of schema matching approaches. InProceedings of 28th International Conference on VeryLarge Data Bases, VLDB2002, pages 610–621. MorganKaufmann, 2002.

[7] A. Doan, P. Domingos, and A. Y. Halevy. Reconcilingschemas of disparate data sources: A machine-learningapproach. In SIGMOD Conf., pages 509–520, 2001.

[8] A. Doan, J. Madhavan, R. Dhamankar, P. Domingos,and A. Halevy. Learning to match ontologies on thesemantic web. VLDB Journal, 2003.

[9] K. Eckert, C. Meilicke, and H. Stuckenschmidt.Improving ontology matching using meta-levellearning. In Proceedings of the Europ. Semantic WebConf., ESWC2008, volume 5444 of LNCS. Springer,2009.

[10] M. Ehrig, S. Staab, and Y. Sure. Bootstrappingontology alignment methods with apfel. In YolandaGil et al., editors, Proceedings of the Int. SemanticWeb Conf., volume 3729 of LNCS, pages 186–200.Springer, 2005.

[11] J. Euzenat. Semantic precision and recall for ontologyalignment evaluation. In M. Veloso, editor,Proceedings of the 20th International Joint Conferenceon Artificial Intelligence, IJCAI2007, pages 348–353,2007.

[12] J. Euzenat et al. State of the art on ontologyalignment. Deliv. D2.2.3, Knowledge Web NoE, 2004.

[13] J. Euzenat and P. Shvaiko. Ontology matching.Springer, 2007.

[14] N. Fanizzi, C. d’Amato, and F. Esposito.Instance-based retrieval by analogy. In Y. Cho et al.,editors, Proceedings of the 22nd Annual ACMSymposium of Applied Computing, SAC2007, pages1398–1402, Seul, South Korea, 2007. ACM.

[15] C. Fellbaum, editor. WordNet An Electronic LexicalDatabase. MIT Press, Cambridge, MA, 1998.

[16] J. C. Gower and P. Legendre. Metric and euclideanproperties of dissimilarity coefficients. Journal ofClassification, 3:5–48, 1986.

[17] M. Grabisch, S.A. Orlovski, and R.R. Yager. Fuzzyaggregation of numerical preferences. Fuzzy sets indecision analysis, operations research and statistics,pages 31–68, 1999.

[18] T.R. Gruber. A translation approach to portableontology specifications. Knowledge Acquisition,5(2):199–220, 1993.

[19] B.B. Hariri, H. Abolhassani, and A. Khodaei. A newstructural similarity measure for ontology alignment.In H. Arabnia, editor, Proceedings of SWWS 2006,pages 36–42. CSREA Press, 2006.

[20] R. Ichise. Machine learning approach for ontologymapping using multiple concept similarity measures.In Proceedings of the 7th IEEE/ACIS Int. Conf. onComputer and Information Science, ICIS 08, 2008.

[21] Y.R. Jean-Mary and M.R. Kabuka. Asmov: results forOAEI 2008. In Proceedings of the ISWC2008Workshop on Ontology Matching, Karlsruhe,Germany, 2008.

[22] Q. Ji, P. Haase, and G. Qi. Combination of similaritymeasures in ontology matching using the OWAoperator. In L. Magdalena et al., editors, Proceedingsof Information Processing and Management ofUncertainty in Knowledge-based Systems, IPMU08,pages 243–250, 2008.

[23] L. Legendre and P. Legendre. Numerical ecology.Development Environment Models, 3, 1983.

[24] P. Maio, N. Bettencourt, N. Silva, and J. Rocha.Evaluating a confidence value for ontology alignment.In Proceedings of the Int. Workshop on OntologyMatching, 2007.

[25] D. McGuinness, R. Fikes, J. Rice, and S. Wilder. Thechimaera ontology environment. In Proceedings of the17th National Conference on Artificial Intelligence,AAAI2000, pages 1123–1124, 2000.

[26] T.M. Mitchell. Machine Learning. McGraw-Hill, 1997.

[27] N. Noy and M. Musen. Prompt: Algorithm and toolfor automated ontology merging and alignment. InProceedings of the 17th National Conference onArtificial Intelligence, AAAI2000, pages 450–455,2000.

[28] N. Noy and M. Musen. The prompt suite: interactivetools for ontology merging and mapping. InternationalJournal of Human-Computer Studies, 59:983–1024,2003.

[29] P. Resnik. Semantic similarity in a taxonomy: Aninformation-based measure and its application toproblems of ambiguity in natural language. Journal ofArtificial Intelligence Research, 11:95–130, 1999.

[30] N. Shadbolt, T. Berners-Lee, and W. Hall. Thesemantic web revisited. IEEE Intelligent Systems,21(3):96–101, 2006.

[31] P. Shvaiko and J. Euzenat. A survey of schema-basedmatching approaches. Journal on Data Semantics,IV:146–171, 2005.

[32] O. Svab and V. Svatek. Combining ontology mappingmethods using bayesian networks. In Proceedings ofthe ISWC2006 Workshop on Ontology Matching, 2006.


[33] K. Tu and Y. Yu. CMC: Combining multipleschema-matching strategies based on credibilityprediction. In L. Zhou et al., editors, Proceedings ofthe Int. Conf. on Database Systems for AdvancedApplications, DASFAA2005, volume 3453 of LNCS,pages 888–893. Springer, 2005.

[34] A. Tversky. Features od similarity. PsycologicalReview, 84(4):327–352, 1977.

[35] I. H. Witten and E. Frank. Data Mining. MorganKaufmann, 2nd edition, 2005.

[36] D.H. Wolpert. Stacked generalization. Neuralnetworks, 5:241–259, 1992.


About the authors:

Floriana Esposito is author of more than 250 papers which have been published in international journals and Proceedings. On August 2006 she received the ECCAI Fellowship in recognition for Pioneering Work in the field of Artificial Intelligence.

Since 1984 associate professor of Computer Science at University of Bari and since 1994 Full Professor of Computer Science, Chair of "Knowledge Engineering”.

Her scientific interests are in the area of Artificial Intelligence and concern the logical and algebraic foundations of Machine Learning: more in particular, the integration of numerical and symbolic methods in learning and classification, the logical foundations of conceptual learning techniques, the automated revision of logical theories, multi-strategy learning for discovery in data bases, inductive logic programming. The major application fields are Document processing and Digital Libraries and, recently, Bioinformatics.

Nicola Fanizzi (PhD in Computer Science), is assistant professor at the Computer Science Dept. of the University of Bari since 2001. The research area of interest, within the broad field of Artificial Intelligence, is Machine Learning and Knowledge Discovery, particularly the methodologies for complex (multi-relational) representations based on fragments of First-Order Logics. Currently he is working on Semantic-Web mining methodologies, focusing on the problems of classification, clustering, ranking, and uncertainty reasoning. This work was/is being carried out within national as well as international research projects and networks of excellence. He authored 100+ scientific papers appeared on scientific journals, proceedings of national and international conferences and workshops. He served in the OC/PC of several workshops and conferences related to Artificial Intelligence, Machine Learning, Data/Web Mining and to the context of the Semantic Web research. See also www.di.uniba.it/~fanizzi

Claudia d'Amato (PhD in Computer Science), is a research fellow at the University of Bari - Computer Science Department. Her main topics of interest are: similarity measures for concepts and individuals in Description Logics, supervised and unsupervised methods for ontology mining, approximate and uncertain reasoning for the Semantic Web. She is author of more that 50 papers among international collections and journals. She has been invited expert for the W3C Uncertainty Reasoning for the World Wide Web Incubator Group. She is organizer of the international Uncertainty Reasoning for the Semantic Web workshop and of the Workshop on Inductive Reasoning and Machine Learning for the Semantic Web. She was Vice-Chair for ISWC 2009, she is editorial board member of the Semantic Web Journal and she is also serving as program committee member in international conferences such as ISWC'10, AAAI'10, ECAI'10, ECML/PKDD'10, SAC'10, EKAW'10, RR'10, RuleML'10, WI'10, STAIRS'10, ICAI'10, KARE'10. She also received the nomination from the Italian Artificial Intelligence community for her PhD thesis "Similarity-based learning methods for the Semantic Web" as one of the best Italian PhD contribution in the area.

composite ontology matching with uncertain mappings recovery

Documents