SODA: an OWL-DL based Ontology Matching System

This paper describes SODA a novel ontology alignment method for the OWL-DL format. SODA uses a new approach that consists in computing local and semantic similarities among ontological elements. 1 Presentation of the system SODA [1] (Structural Ontology OWL-DL [2] Alignment), is a new approach that aligns two OWL-DL ontologies using similarity measures [3]. Both OWL-DL ontologies are transformed in two corresponding graphs DL-GRAPH which describe all information in the ontologies. SODA uses the DL-GRAPH to align the two ontologies. It operates into successive steps. The first step, computes local similarity by means of linguistic and structural similarities, whereas the second one computes the semantic similarity. Figure 1 depicts the architecture of SODA system. 1.1 Specific techniques used Each OWL-DL ontology to be aligned is transformed into a non oriented graph called DL-GRAPH. All the information belonging to OWL-DL ontology are faithfully mapped into the DL-GRAPH. Nodes of the proposed graph represent classes, properties and instances. The DL-GRAPH nodes represent six types (named also categories) of entities that may exist in an OWL-DL ontology: i.e., concepts, instances of concepts, data types, values of data types and class properties (object nature and data type nature). Connections between the graph nodes map the relationships between the entities in an OWL-DL ontology. It is worthily noted that an OWL-GRAPH describes all the semantic relations between different entities of an ontology. A graph DL-GRAPH allows to represent four kinds of links: specialization, attribution, instantiation and equivalence. DL-GRAPHS are exploited by the alignment model SODA. Similarity measures are used to compare the components of the graphs in order to obtain the correspondence between them. Nodes and links of the two graphs are compared to get out the correspondence between different ontological entities using similarity measures. The output algorithm is an RDF file containing all the correspondences between the entities and the similarity measure values. Fig. 1. Architecture of SODA SODA explores the structure of DL-GRAPH to compute the similarity values between the nodes of both ontologies. The alignment model associates to each category of nodes an aggregation function. This function takes in consideration all the similarity measures and the structure of couple of nodes to be matched. This aggregation function explores all descriptive information of nodes. SODA operates into two successive steps: local and semantic. The first step, implemented via PHASE1 LINGSIM (see Algorithm 1) and PHASE2 STRUCTSIM (see Algorithm 2) functions, computes the local similarity (linguistic and structural one). The second step, c.f. the PHASE3 SEMSIM function (see Algorithm 3), computes the semantic similarity. Table 1 summarizes the notations that are used in the description of our algorithms. Local similarity The computation of the local similarity is carried out in two phases. The first phase allows to compute the linguistic similarity for each couple of node of the same category. The second phase allows to compute the structural similarity using the structure of neighbors of the nodes to be aligned. O1,O2: two OWL-DL ontologies for alignment VLS : linguistic similarity vector VSS : structural similarity vector VV SEMS : semantic similarity vector Each node of the ontology is characterized by: Type: node type Name: node name Each element of the vectors VSL, VSS et VV SEM is characterized by: Node 1: the node of ontology O1 Node 2: the node of ontology O2 Sim: the similarity value Table 1. Algorithm notations Algorithm 1 (c.f., PHASE1 LINGSIM function) computes the linguistic similarity measure. The name of properties and instances are used to compute linguistic similarity. For classes, the computation of linguistic similarity integrates also comments and labels. The computation of linguistic similarity is done only once for each node of the same category. JARO-WINKLER or MONGE-ELKAN [4] functions are used to compute the linguistic similarity. JARO-WINKLER measure is more adapted for short strings, like those representing names and labels [4]. Besides, MONGE-ELKAN measure is better indicated for long strings, e.g. the comments [4]. PHASE1 LINGSIM function computes the linguistic similarity of couple of nodes of both considered ontologies. PHASE1 LINGSIM function takes as input the two ontologies O1 and O2 and the linguistic similarity function FunctLS . COMPUTELINGSIM function (c.f., line 8 of Algorithm 1) takes as an input two nodes, Node1 et Node2, and linguistic similarity function. PHASE1 LINGSIM function returns as an output linguistic similarity value SimL. This function implements the JARO-WINKLER or the MONGE-ELKAN measures. Linguistic similarity of the different couples of nodes are used after that in the computation of the structural similarity. Structural similarity is computed by using linguistic similarity of the couple of nodes to align and the neighborhood structure. Adjacent neighbor nodes of the entities are grouped by category, c.f. PHASE2 STRUCTSIM. This function takes as input two ontologies O1 and O2 to align, linguistic similarity vector VLS and weights associated for each category ΠC . EXTRACTNODES function, (c.f., lines 9 11 of Algorithm 2), allows to extract for each node, its neighbors and to put them in VNodei , where Nodei is a node of O1 or O2. VNode1 and VNode2 vectors and weights associated for each category, ΠC , are used by the COMPUTESTRUCTSIM function (c.f., line 13 of Algorithm 2) to compute the structural similarity, SimS . To work out, the following ”Match-Based similarity” [5, 6] is used to compute similarity between two categories (one in the first ontology and the other in second one): MSim(E,E′) = ∑ (i,i′)∈Paires(E,E′) Sim(i, i ′) Max(|E|, |E′|) , Function : PHASE1 LINGSIM 1 Data: 1. O1 and O2 : two ontologies to align 2. FunctLS : linguistic similarity function Results: VLS : linguistic similarity vector Begin 2 /*Parse all nodes of the ontology O1*/ 3 forall (Node1 ∈ O1) do 4 /*Parse all nodes of the ontology O2*/ 5 forall (Node2 ∈ O2) do 6 If Node1.type=Node2.type then 7 SimL = COMPUTELINGSIM(Node1,Node2) 8 /*Add Node1,Node2 and SimL to VLS*/ 9 Add((Node1,Node2,SimL),VLS) 10 return(VSL) 11 End 12 Algorithm 1: PHASE1 LINGSIM where E et E′ represent two sets of nodes belonging to the same category in O1 and O2. This function uses the local similarities of the couple (i, i′) already computed. Structural similarity is computed by aggregating the ”Match-Based similarity” of each group of adjacent neighborhood nodes by category. A weight is attributed for each group to have a normalized structural similarity. Each category has the same weight which is equal to 1 over the number of groups (categories). Structural similarity, SimS , is computed as follows: