In this paper we address the problem of resource discovery in the Linked Open Data cloud (LOD) where data described by different schemas is not always linked. We propose an approach that allows discovery of new links between data. These links can help to match schemas that are conceptually relevant with respect to a given application domain. Furthermore, these links can be exploited during the querying process in order to combine data coming from different sources. In this approach we exploit the semantic knowledge declared in different schemas in order to model: (i) the influences between concept similarities, (ii) the influences between data similarities, and (iii) the influences between data and concept similarities. The similarity scores are computed by an iterative resolution of two non linear equation systems that express the concept similarity computation and the data similarity computation. The proposed approach is illustrated on scientific publication data.
The appearance of Web of documents (WWW) [
In this paper we focus our interest on the problem of resource discovery in the Linked Open Data cloud (LOD) where data described by different schemas is not always linked. We propose an approach1 that allows discovery of new links between data. These links can help to match schemas that are conceptually relevant with respect to a given application domain. 1 in the setting of the ANR (the French National Research Agency) project GeOnto.
Ontology alignment plays a key role for semantic interoperability of this data. Many
approaches have been proposed for automatically identifying mappings between
elements (concepts and relations) described in heterogeneous ontologies [
Applying this approach allows one to infer mappings of equivalence between concepts of different schemas as well as to infer owl:same-as relations between instances that refer to the same entity. The obtained schema mappings allow discovery new resources and inferring if they are relevant with respect to a given application domain.
The paper is organized as follows: in section 2 we present the related work in data linking and ontology reconciliation fields. In section 3, we present the ontology and data model and give a short presentation of N2R method on which our work relies. Section 4 presents the proposed approach of link discovery. Finally, we conclude and give some future work in section 5. 2
We denote by “web data” the network formed by the set of structured datasets described in RDF (Resource Description Framework) and linked by explicit links. Large amount of structured data have been published, including in the project Linking Open Data cloud (LOD).
Datasets are expressed in terms of one or several ontologies for establishing the
vocabulary describing data. Web data requires linking together the various sources
of published data. Given the big amount of published data, it is necessary to provide
methods for automatic data linking. Several tools [
Some other works address the problem of link discovery in the context semantic
Web services. In [
Our proposal in this paper can also be compared to approaches studying the
reference reconciliation problem, i.e., detecting whether different data descriptions refer
to the same real world entity (e.g. the same person, the same paper, the same protein).
Different approaches have been proposed. [
In [
In this section we will present the ontology and data model that we consider in this
work. We will then present the Numerical method for Reference Reconciliation (N2R)
[
The considered OWL ontology consists of a set of concepts (unary relations) organized in a taxonomy and a set of typed properties (binary relations). These properties can also be organized in a taxonomy of properties. Two kinds of properties can be distinguished in OWL: the so-called relations (owl:objectProperty), the domain and the range of which are concepts and the so-called attributes (owl:DatatypeProperty), the domain of which is a concept and the range of which is a set of basic values (e.g. Integer, Date, Literal). In Figure 1, we give an extract O1 of the ontology that is used to describe the RDF data of the local data source of publications (see source 1 Figure 2) which we will use to illustrate our proposal.
We allow the declaration of constraints expressed in OWL-DL or in SWRL in order to enrich the domain ontology by additional and useful knowledge. The constraints that we consider are of the following types: Source S1: Article(S1_a1); title(S1_a1,“Implementing the TEA algorithm on sensors”); Person(S1_p1); Person(S1_p2); year(S1_a1, “2004”); name(S1_p1,“Olga V. Gavrylyako” ); name(S1_p2,“Shuang Liu” ); pageFrom(S1_a1,“64” ); pageTo(S1_a1,“69” ); Conference(S1_c1); confName(S1_c1, “Proceedings of the 42nd Annual Southeast Regional Conference, 2004, Huntsville, Alabama, USA, April 2-3, 2004”); confYear(S1_c1, “2004”); city(S1_c1, “Alabama”) authoredBy(S1_a1,S1_p1 ); authoredBy(S1_a1,S1_p2 ); published(S1_a1,S1_c1 ); Article(S1_a2); title(S1_a2,“Weighted Hyper-sphere SVM for Hypertext Classification”); Person(S1_p3); Person(S1_p4); year(S1_a2, “2008”); name(S1_p3,“Shuang Liu” ); name(S1_p4,“Guoyou Shi” ); pageFrom(S1_a2,“733” ); pageTo(S1_a2,“740” ); Conference(S1_c2); confName(S1_c2, “Advances in Neural Networks - ISNN 2008, 5th International Symposium on Neural Networks, ISNN 2008, Beijing, China, September 24-28, 2008, Proceedings, Part I”); confYear(S1_c2, “2008”) city(S1_c2, “Beijing”) authoredBy(S1_a2,S1_p3); authoredBy(S1_a2,S1_p4); published(S1_a2,S1_c2);
Article(S2_a1); title(S2_a1,“Implementing the TEA algorithm on sensors.’); Person(S2_p1); Person(S2_p2); year(S2_a1, “2004”); name(S2_p1,“Olga V. Gavrylyako” ); name(S2_p2,“Shuang Liu” ); pageFrom(S2_a1,“64” ); pageTo(S2_a1,“69” ); Conference(S2_c1); confName(S2_c1, “42nd Annual Southeast Regional Conference, 2004”); confYear(S2_c1, “2004”); city(S2_c1,“Alabama”) authoredBy(S2_a1,S2_p1 ); authoredBy(S2_a1,S2_p2 ); published(S2_a1,S2_c1 ); – Constraints of disjunction between concepts: DISJOINT(C,D) is used to declare that the two concepts C and D are disjoint. In the ontology O1 we declare that all the concepts Article, Conf erence and P erson are pairwise disjoint. – Constraints of functionality of properties: PF(P) is used to declare that the property P (relation or attribute) is a functional property. In O1, we declare that all the properties are functional except the relation authoredBy which means that one article may have several authors. – Constraints of inverse functionality of properties: PFI(P) is used to declare that the property P (relation or attribute) is an inverse functional property. These constraints can be generalized to a set {P1, . . . , Pn} of relations or attributes to state a combined constraint of inverse functionality that we will denote PFI(P1, . . . , Pn). In O1, we declare that the combinations (title, year) and (conf N ame, conf Y ear) are inverse functional. For example, P F I (title, year) expresses that one title and one year cannot be associated to several articles (i.e. both are needed to identify an article). 3.2
A piece of data has a reference, which has the form of a URI (e.g. http://dblp. l3s.de/d2r/resource/authors/A._Joe_Turner), and a description, which is a set of RDF facts involving its reference. An RDF fact can be either: (i) a conceptfact C(i), where C is a concept and i is a reference, (ii) a relation-fact R(i1, i2), where R is a relation and i1 and i2 are references, or (iii) an attribute-fact A(i, v), where A is an attribute, i a reference and v a basic value (e.g. integer, string, date). We consider the Unique Name Assumption (UNA) which can be declared or not on a data source. Declaring UNA on a data source means that two different data descriptions having two different references, then we infer that they refer to distinct entities.
The data description that we consider is composed of RDF facts coming from the data sources which are enriched by applying the OWL entailment rules. Figure 2, provides examples of data coming from two RDF data sources S1 and S2, which conform to the same ontology describing the scientific publication domain previously mentioned.
In the N2R method which we will present in section 3.3, we consider that the descriptions of data coming from different sources conform to the same OWL ontology (possibly after ontology reconciliation). In the link discovery method, that we will present in section 4, the assumption of prior ontology reconciliation is not fulfilled, i.e. the considered data source do not conform to the same ontology. 3.3
N2R is a numerical method which allows inferring reconciliation decisions between reference coming from different sources that conform to the same ontology, i.e. the problem on ontology reconciliation is already solved.
N2R [
N2R can also take as input a set of reference pairs that are reconciled (sim =1)
by another method (e.g. L2R [
references, its similarity score is modeled by a variable xi and the way it depends on other similarity scores, is modeled by an equation: xi = fi(X ), where i ∈ [1..n] and n is the number of reference pairs for which we apply N2R, and X = (x1, x2, . . . , xn) is the set of their corresponding variables. Each equation xi = fi(X ) is of the form: fi(X ) = max(fi−df (X ), fi−ndf (X ))
The function fi−df (X ) is the maximum of the similarity scores of the value pairs
and the reference pairs of attributes and relations with which the i-th reference pair
is functionally dependent. The maximum function allows propagating the similarity
scores of the values and the references having a strong impact. The function fi−ndf (X )
is defined by a weighted average of the similarity scores of the value pairs (and sets)
and the reference pairs (and sets) of attributes and relations with which the i-th reference
pair is not functionally dependent. See [
Iterative algorithm for reference pairs similarity computation. Solving this equation
system is done by an iterative method inspired from the Jacobi method [
i
In order to illustrate the iterative resolution of the equation system, we consider an extract of RDF data given in Figure 2 corresponding to the set of RDF facts where the references S1_a1, S1_c1, S2_a1 and S2_c1 are involved. By considering the disjunctions between concepts of O1 and the UNA in S1 and S2, we obtain an equation system of six variables: x1 = Simr(S1_a1, S2_a1) ; x2 = Simr(S1_c1, S2_c1) ; x3 = Simr(S1_p1, S2_p1) ; x4 = Simr(S1_p1, S2_p2) ; x5 = Simr(S1_p2, S2_p1) ; x6 = Simr(S1_p2, S2_p2).
We give bellow, the similarity scores of basic values obtained by using the Jaccard similarity measure. For clarity reasons, we denote the value of an attribute A associated to a reference i as: A.val(i). For example, the conf Y ear value associate to the reference S2_c2 is denoted conf Y ear.val(S2_c2) which equals to “2008”. The similarity score of the two conference names that are needed in the equation system and that belong to ]0, 1[ is: Simv(conf N ame.val(S1_c1), conf N ame.val(S2_c1)) = 0.43. All the similarity scores of basic values, that are needed in the computation, are either equal to 1 or equal to 0.
The weights that are used in the weighted average of equations are computed in function of the number of common attributes and common relations of the reference pairs. The similarity computation is illustrated by the equation system (see Table 1) obtained from the data descriptions shown in Figure 2 which conforms to the ontology O1. The detailed equations expressing the similarity computation of two articles and two conferences are as follows: x1 = max( 12 (Simv(title.val(S1_a1), title.val(S2_a1)) + Simv(year.val(S1_a1), year.val(S2_a1))), 16 (x2 + SJ({S1_p1, S1_p2}, {S2_p1, S2_p2}), Simv(pageF rom.val(S1_a1), pageT o.val(S2_a1))) with SJ is the Sof tJ accardo similarity measure between sets of objects (see section 4.2) x2 = max(x1, max( 21 (Simv(conf N ame.val(S1_c1), conf N ame.val(S2_c1))+ Simv(conf Y ear.val(S1_c1), conf Y ear.val(S2_c1))), 14 (Simv(city.val(S1_c1), city(S2_c1))) x3 = 21 ∗ x1 + 21 ∗ Simv(name.val(S1_p1), name.val(S2_p1))
The equation system and the different iterations of the resulting similarity computation are provided in Table 1. We assume that fix-point precision ! equals to 0.005.
The solution of the equation system is X = (1, 1, 1, 1, 0.5, 0.5). This corresponds to the similarity scores of the six reference pairs. The fix-point has been reached after three iterations. If we fix the reconciliation threshold Trec at 0.80, then we obtain four reconciliation decisions: two articles, two conferences and two pairs of persons. 4
We present in this section our LDM approach which aims to discover a LOD source that shares concepts and data with a data source described dy a domain ontology. Our approach compares a local dataset on which domain knowledge can be declared and a LOD dataset by using a combined ontology reconciliation and reference reconciliation method. Since data that is provided by the LOD source and by the domain application source is not described using the same ontology, we have adapted N2R method in order to be able to compute data similarities when data do not belongs to non disjoint concepts but to similar concepts. Furthermore, we have defined how similarities between concepts of two ontologies can be computed when some of their references are common (i.e. same URI or owl:same-as links that have been previously asserted) or similar. The main steps of our link discovering approach are as follows: 1. application of an ontology mapping tool to obtain: (i) the set of equivalent/ comparable properties and (i) initial similarity scores for some concept pairs; 2. building of the two equation systems: the conceptual equation system which expresses the similarity computation between pairs of concepts in function of their labels, their structural similarity and their references; and the instance level equation system one which expresses the similarity computation between pairs of references in function of their common description and the similarity of the concepts they are instance of; 3. iterative resolution of the conceptual equation system until a fix point is reached; 4. iterative resolution of the instance level equation system until a fix point is reached.
The two steps (3) and (4) are iterated until a global fix point is reached, i.e., neither the resolution of the conceptual equation system nor the resolution of the instance level equation system does update the similarity scores.
In the following subsections, we will first describe the elementary similarity measures that are used to compute similarities. Then, we present the two equation systems that have been defined to compute concept similarities and data similarities. Finally, we illustrate our LDM approach on data and ontologies of publication domain. 4.1
We first use an alignment tool which exploits lexical and structural information to find similarity scores between ontology elements (concepts and properties). Given a local ontology O1 and a LOD ontology O2, the used alignment tool finds a set of mappings and each mapping is described by the tuple {e1, e2, co, rel} where e1 is aligned with the confidence co to the element e2 using the type of correspondence rel (e.g. equivalence, subsumption, overlap, closeness, etc.). These scores are used to initialize the similarity score simInit of each pair of concepts and to find a set of properties (relations or attributes) that are very similar (rel = equivalence or subsumption, co ≥ th and th is a high threshold). These properties are then considered as equivalent. 4.2
We present in this section the elementary measures used to compute similarity scores between pairs of concepts of two ontologies. These elementary similarity measures take into account the lexical and the structural knowledge declared in the two ontologies. Most of these elementary similarity measures are based on the SoftJaccard similarity measure which computes similarity between sets of basic values or between sets of objects (e.g., references, concepts).
SoftJaccard similarity measure which is an adaptation of the Jaccard similarity measure in the sense that: (i) instead of considering only basic values we consider sets of basic values and (ii) instead of considering the equality between values we consider a similarity score with respect to a threshold θ.
Let S1 and S2 be two sets of elements which can be basic values or objects. To compute the similarity score between S1 and S2 we compute, first, the set CLOSET (S1, S2, θk) which represents the set of element pairs of S1 × S2 having a similarity score simT ≥ θ.
CLOSET (S1, S2, θ) = {ej | ej ∈ S1 and ∃ek ∈ S2 s.t. SimT (ej, ek) > θ},
with T a parameter which indicates if the sets S1 and S2 contain basic values, then
T = v or contain objects, then T = o. When T = v, the function Simv corresponds
to a similarity measure between basic values like J accard, J aro − W inkler, and so
on [
Sof tJ accardT (S1, S2, θ) = | CLOSE(S1, S2, θ) | | S1 | , with | S1 |≥| S2 |
tween concepts we exploit both the conceptual content which means the sets of ancestors and the sets of descendants but also the sets of shared properties with respect to a given equivalence relation. The similarity score between concepts is also function of the similarity scores of their references, i.e. instance level content.
Similarity of concept labels. In OWL ontologies sets of labels are usually associated to the concepts. In case of concepts where the labels are not given, we consider their corresponding URIs. Let L1 be the set of labels of a concept c1 and L2 be the set of labels of the concept c2. The label similarity simlabel is computed by applying the Sof tJ accard similarity measure on the two sets of basic values L1 and L2: simlabel(c1, c2) = Sof tJ accardv(L1, L2, θ1).
Similarity of concept ancestors. For two concepts, we also compute the similarity of their ancestor sets in the two ontologies. Let A1 be the set of ancestors of the concept c1 and A2 be the set of ancestors of c2. The ancestor similarity simanc is computed by applying Sof tJ accard similarity measure on the two sets of concept ancestors (i.e. objects) which is defined as follows: simanc(c1, c2) = Sof tJ accardo(A1, A2, θ2) Similarity of concept descendants. The similarity score of two concepts also depends on the similarity scores of their descendants in the two ontologies. Let D1 be the set of descendants of the concept c1 and D2 be the set of descendants of c2. The descendant similarity simdesc is computed by applying Sof tJ accard similarity measure on the two sets of concept descendants which is defined as follows: simdesc(c1, c2) = Sof tJ accardo(D1, D2, θ3) Similarity of shared properties of concepts. The similarity of two concepts depends on the proportion of equivalent properties compared to the full number of properties defined for both concepts. Let R1d (resp. R2d) be the set of properties such that the concept c1 (resp. c2) is subsumed by the (equivalent) property domain and let R1r (resp. R2r) the set of properties such that the c1 (resp. c2) is subsumed by one of the range of the (equivalent) property. The relation similarity simrel is defined as follows : simrel(c1, c2) = | (R1d ∩ R2d) ∪ (R1r ∩ R2r) | | (R1d ∪ R2d ∪ R1r ∪ R2r) | Similarity of concept references. The similarity score of two concepts also depends on the set of their references. Let I1 (resp. I2) be the set of instances of c1 (resp. c2), the similarity of c1 and c2 depends on the similarity scores obtained for the pairs of references of I1 × I2 and it is computed by applying the Sof tJ accard similarity measure on the sets I1 and I2 of references (i.e. objects). simref (c1, c2) is defined as follows: simref (c1, c2) = Sof tJ accardo(I1, I2, θ4). 4.3
In LDM approach the similarity of each pair of references is expressed by a variable xi in the instance level equation system. Its value depends on the common description of the pair of references w.r.t the equivalent/ comparable properties (cf. N2R). It depends also on the similarity scores of the concepts sci that are instantiated by the pair of references. An equation of the instance level equation systemxi = gi(X ), where i ∈ [1..n] and n is the number of reference pairs and X = (x1, . . . , xn), is of the form: 1 gi(X ) = (sci, fi(X ))
2 with sci is the similarity score computed by the resolution of the conceptual equation system presented in the following. The function fi(X ) is expressed as in N2R method and we consider that knowledge on the (inverse) functionality of the shared properties declared in the local ontology is also fulfilled in the LOD ontology.
The similarity of each pair of concepts (c, c") is expressed by a variable xcj in the conceptual equation system. Its value depends on the initial similarity score provided by the alignment tool, the similarity of their labels, the set of their equivalent / comparable properties and the similarity of their references represented respectively by the constants simj−init, simj−label, simj−rel and simj−ref . It depends also on the similarity of their ancestors and their descendants represented by the variables X SCj−anc and X SCj−desc computed using SotfJaccard function.
An equation xcj = hj (X C), where j ∈ [1..m] and m is the number of concept pairs and X C = (xc1, . . . , xcm), is of the form: (XSCj−anc + XSCj−desc + simj−rel + simj−label + simj−ref )) The values of the constants simj−init, simj−label, simj−rel and simj−ref are computed using the similarity functions described in the above subsection.
The size m of the conceptual equation system is | C1 × C2 |, where C1 (resp. C2) is the set of concepts of the ontology O1 (resp. O2). The size of the instance level equation system depends on the number k of comparable relations and on the size of their corresponding domain instances and range instances. Let ri1 and ri2 be two comparable relations. Let Ei1 (resp. Ei2) be the set of domain instances of ri1 (resp. of ri2) and Ei3 (resp. Ei4) be the set of range instances of ri3 (resp. ri4). It also depends on the number of comparable attributes k" and on the size of their corresponding domain instances. Let aj1 and aj2 be two comparable attributes. Let Ej1 (resp. Ej2) be the set of domain instances of aj1 (resp. of aj2). The number n of variables of the instance level equation system is:
i=k j=k! n =| ! ((Ei1 × Ei2) ∪ (Ei3 × Ei4)) ∪ ( ! (Ej1 × Ej2)) |
i=1 j=1 The computation complexity of the LDM method is O((n2 ∗ itref ) + (m2 ∗ itc)), with itref is the number of iterations of the instance level equation system and itc is the number of iterations of the conceptual equation system.
One of the most distinguishing characteristic of LDM is its ability to propagate similarities at different levels: (i) between pairs of concepts, (ii) between pairs of references and (iii) between sets of references and sets of concepts. By using two separated equation systems we avoid the propagation between references when we compute the concept similarity scores and we avoid also the propagation between concepts when we compute the reference similarity scores. Thus, we decrease the size of the equation system and we allow a user to visualize and validate the intermediate equation system results.
We present in Figure 3 an extract of the DBLP ontology which is used to describe the DBLP data published in the LOD. The considered data set only contains a collection of conference proceedings and the collection of their corresponding research papers in computer science. In order to illustrate our approach of link discovery, we will compare the local RDF data of the source S1 given in Figure 2 with the extract of DBLP dataset of the LOD given in Figure 4.
The initialization step provides the following initial similarity scores for the concept pairs: siminit(Article, InP roceedings) = 0.3; siminit(Article, P roceedings) = 0.1; siminit(Article, Agent) = 0.1; siminit(P erson, InP roceedings) = 0.0; siminit(P erson, P roceedings) = 0.0; siminit(P erson, Agent) = 0.3; siminit(Conf erence, InP roceedings) = 0.2; siminit(Conf erence, P roceedings) = 0.2; siminit(Conf erence, Agent) = 0.1 LOD source S2: InProceedings(S2_a1); label(S2_a1,“Implementing the TEA algorithm on sensors’); Agent(S2_p1); Agent(S2_p2); issued(S2_a1, “2004”); name(S2_p1,“Olga V. Gavrylyako” ); name(S2_p2,“Shuang Liu” ); Proceedings(S2_c1); label(S2_c1, “42nd Annual Southeast Regional Conference, 2004”); creator(S2_a1,S2_p1 ); creator(S2_a1,S2_p2 ); partOf(S2_a1,S2_c1 ); InProceedings(S2_a2); label(S2_a2,“New Chaos Produced from Synchronization of Chaotic Neural Networks”); Agent(S2_p3); issued(S2_a2, “2008”); name(S2_p3,“Zunshui Cheng” ); Proceedings(S2_c2); label(S2_c2, “Advances in Neural Networks - ISNN 2008, 5th International Symposium on Neural Networks”); creator(S2_a2,S2_p3); partOf(S2_a2,S2_c2);
Since, there are no subsumption relations in O1 and in O2 the conceptual equations do not take into account the similarity scores of the ancestors and of the descendants. For example, the equation expressing the similarity of the two concepts Article and InP roceedings is: xc1 = max(0.3, 13 ( 32 + 0 + sim1−ref )). In this example, the conceptual equation system consists of nine variables (xc1, . . . , xc9).
The instance level equation system consists of twenty-five equations representing
all the reference pairs where the common description is not empty. For example, the
equations expressing:
– The similarity of the two references S1_a1
(Article) and S2_a1 (InProceedings) is: x1 =
12 (sc1, max( 12 (Simv(label.val(S2_a1), title.val(S1_a1))+Simv(issued.val(S2_a1),
year.val(S1_a1)), 41 (SJ({S1_p1, S1_p2}, {S2_p1, S2_p2}) + x14)
– The similarity of the two references S1_c1 (Conference) and S2_c1(Proceedings)
is : x14 = 21 (sc14, max(x1, 12 (Simv(label.val(S2_c1), conf N ame.val(S1_c1)))
– The similarity of the two references S1_p1 (Person) and S2_p1 (Agent) is:
x5 = 21 ((sc5, 12 (x1 + Simv(name.val(S1_p1), name.val(S2_p1))
– The similarity of the two references S1_p1 (Conference) and S2_p1
(InProceedings) is: x18 = 12 (sc18, max(Simv(conf N ame.val(S1_c1), label.val(S2_a1))
In Table 2 we show the iterative resolution of the conceptual equation system ES1
modeling the similarity of all the pairs of concepts of the ontologies O1 and O2. The
column siminit represents the initial similarity score computed by an external concept
alignment tool, like TaxoMap [
The Resolution1 step of ES1 corresponds to the first iterative resolution of ES1 where simref of all the concepts equals to 0. The fix-point of ! = 0.05 is reached in two iterations. The Resolution1 step of ES2 corresponds to the first iterative resolution of ES2 where sci of all the references equals to siminit (c.f. Table 2) . The fix-point of ! = 0.05 is reached in three iterations. The Resolution2 step of ES1 corresponds to the second iterative resolution of ES1 where simref of all the concepts equals to the similarity scores computed by ES2 at the last iteration of Resolution1. The fix-point of ! = 0.05 is also reached in two iterations. The Resolution2 step of ES2 corresponds to the second iterative resolution of ES2 where sci of all the references equals to the similarity scores computed by ES1 at the last iteration of Resolution1. The fix point of ! = 0.05 is reached in two iterations.
The global fix-point is reached after three resolutions. At Resolution3 2 of the two systems ES1 and ES2 we obtain the same similarity scores than the last iteration of their corresponding Resolution2 step. The results obtained by ES1 show that the method obtains the best similarity scores for the most possible equivalent concepts: (Article, I nP roceedings), (P erson, Agent) and (Conf erence, P roceedings). In an analogous way, the results obtained by ES2 show that the best similarity scores are obtained for the most possible owl:same-as references. If we fix the reconciliation 2 The scores are not shown here, they are equal to those obtained in the Resolution2 of ES1 and ES2. threshold at 0.45 we infer the reconciliation of the two papers (S1_a1, S2_a1), of the two persons (S1_p1, S2_p1) and of the two conferences (S1_c1, S2_c1).
In this example we have shown the applicability of the approach even when the considered ontologies are not syntactically close and when they have very poor structure (no subsumption relations) which means that the ancestors and the descendants are not considered. 5
In this paper we have presented a Link Discovering Method (LDM) which allows discovery of new data sources that are published in the Open Linked Data cloud (LOD). Our approach is based on the idea of comparing a local dataset on which domain knowledge can be declared and a LOD dataset by using a combined ontology reconciliation and reference reconciliation method. By using our LDM method one may discover more owl:same-as links with datasets available on the LOD.
One of the most distinguishing characteristic of our link discovery approach resides on its ability to propagate similarities at different levels: (i) between pairs of concepts, (ii) between pairs of references and (iii) between sets of references and sets of concepts. By using two separated equation systems we avoid the propagation between references when we compute the concept similarity scores and we avoid also the propagation between concepts when we compute the reference similarity scores.
As a very short term perspective, we plan to test our LDM approach on real data
sets and evaluate the quality of its results and its scalability. It will be worth to
compare LDM method with those of existing link discovery methods like [