OACAS: results for OAEI 2011

Ontologies are the kernel of semantic Web. They allow the explicitation of the semantic purpose for structuring different fields of interest. In order to harmonize them and to guarantee the interoperability between these resources, the topic of alignment of ontologies has emerged as an important process to reduce their heterogeneity and improve their exploitation. The paper introduces a new method of alignment of OWL-DL ontologies, using a combination and aggregation of similarity measures. Both ontologies are transformed into a graph which describes their information. The proposed method operates in two steps: local (linguistic similarity composition and neighborhood similarity) step and the aggregation one. 1 Presentation of the system The method, OACAS [1] (Ontologies Alignment using Composition and Aggregation of Similarities), introduces an alignment algorithm of OWL-DL (OntologyWeb Language Description Logic) ontologies. The main thrust of this method is the application of the most suitable similarity measure depending of the category of the node in the ontology. In addition, the OACAS method explores a wider neighborhood than do the pioneering methods of the literature. Carried out experiments showed that OACAS presents very encouraging values of the commonly used evaluation metrics for the assessment of ontologies alignment. 1.1 Specific techniques used The proposed method, OACAS, alignes two ontologies. Both ontologies are described in the OWL-DL language [2]. Both ontologies are transformed in two graphs O-GRAPHS. The obtained graphs are parsed in order to produce the alignment process out. Mapping of an OWL-DL Ontology to an O-GRAPH. The process of building the graphs allows to faithfully map the considered ontologies to be aligned in two graphs, called O-GRAPHS. An O-GRAPH describes all the information categories included in an OWL-DL ontology: classes, relations and instances. Both classes and instances represent the nodes of the graph. The relations between these different entities are induced by the links of an O-GRAPH. Each entity of the ontology is formalized through an associated notion to the RDF formalism [3]. OWL-DL ontology entities are described thanks to OWL language constructors. These constructors are represented through RDF triplets:. In an OWL-DL ontology, a class or a relation description is an RDF triplet. The subject corresponds to the class or to the relation. Predicates are OWL primitives, which are OWL and RDF properties. Each property, used in a triplet, sketches a knowledge of the described entity. The arrangement of those nuggets of knowledge constitues the entity definition. The representation of an OWLDL ontology through an O-GRAPH permits to load the ontology in main memory only once. An O-GRAPH, stored in main memory, statistically reduces the time required to access initial OWL-DL ontology disk resident file. The alignment method. The introduced OACAS method lays on a composition and an aggregation of similarity computation based model. The method starts by exploring the O-GRAPH structure. It determines the nodes of both ontologies to be aligned and gets out the similarity measures. For each node of the same category (or cluster), the alignment model computes similarity mesures between descriptors by using appropriate functions. Thus, this function considers all the descriptive information of this couple (name, comment and label) as well as its neighborhood structure. An aggregation function combines the similarity measures and the node’s structures of the nodes to be aligned. The algorithm implementing the OACAS method takes as input two OWLDL ontologies to be aligned and produces an RDF file containing the aligned nodes as well as their similarity measures. The alignment method operates into two successive steps. The first one computes the local similarity, whereas the second one computes the aggregation similarity. First step: Local similarity The local similarity computation is performed into two successive stages. The first one computes many linguistic similarity measuresand aggregates them for each couple of nodes belonging to the same category (or type). The second one computes neighborhood similarities by exploiting the structures of the nodes to be aligned. The linguistic similarity computation is carried out once for each node of the same cluster (node of the same type) in the beginning of the alignment process. The linguistic similarity measures of couples of entities of the same type (class, property and instance) are computed. The names of properties and instances are used to compute linguistic similarities. For class category, the computation of the linguistic similarity considers both the comments and labels. The computation of linguistic similarities uses different similarity measures. Those measures are adapted to different descriptors (names, comments and labels) of the entities to be aligned. Different similarity values obtained, for the descriptors, are composed. This composition assigns weights to each similarity measure of descriptors. The sum of the assigned weights to different similarity values is equal to 1. This unit sum guarantees that the composition of the similarity produces a normalized value (between 0 and 1). The LEVENSHTEIN similarity measure [4] is used to compute the similarity value between the names of ontological entities. The Q-GRAM similarity measure [5] computes the similarity value between the comments of the ontological entities. The JARO-WINKLER similarity measure [6] computes the similarity value between the labels of ontological entities. The LINGUISTIC function computes composed linguistic similarity of couples of nodes of both ontologies to be aligned, i.e., O1 and O2. It takes as input (i) both ontologies sketched by two corresponding O-GRAPHS; (ii) linguistics similarity functions (i.e., Funct); and (iii) weighted attributed to the descriptors nodes (i.e., ΠD). As a result, it produces a composed linguistic similarity vector, VCLS , for each couple of n nodes. The similarity function Funct considers two nodes, N1 and N2, and returns the linguistic similarity value of the descriptor, SimLD. LEVENSHTEIN or Q-GRAM or JARO-WINKLER implements the similarity function, Funct, depending of the type of the nodes. Composed linguistic similarity, SimCL, is computed depending of the descriptors of nodes to be aligned and associate weights to each descriptor, ΠD. Both nodes (N1 and N2) and the associated composed linguistic similarity (SimCL) are added to the composed linguistic similarity vector (VCLS). The composed linguistic similarity of different couples of entities will be used to compute the neighborhood similarity as sketched in the following. The neighborhood similarity considers both ontologies to be aligned (i.e., O1 and O2), the composed similarity vector (VCLS), the weights assigned to each category (ΠC) and the weights associated to the neighbor level (ΠL). Therefore, it produces the neighborhood similarity vector, VNS . The neighborhood similarity computation needs composed linguistic similarity of the couple of nodes to be aligned and the nodes structures. Neighborhood nodes are organized by category, node having the same type. The neighborhood similarity computation propagates similarity into two successive neighborhood levels. The first level (level 1) includes direct neighbors of the nodes to be aligned whereas second one (level 2) contains indirect neighbors. Direct neighbors of the first level represent nodes having direct relationship with the node under consideration. Neighbors of the second level represent nodes having relationship with the nodes of the first one. The neighbors entities of the first level are clustered into three categories (classes, instances or properties). Each category (or cluster) includes ontological entities having the same type. After the step of clustering, the neighborhood similarity is computed between those categories. The neighborhood nodes of the level 2 are treated in the same manner as the neighbors of the first one. The neighborhood similarity by group MSim takes nodes from vectors V N1 and V N2 regrouped by category (where V N1 and V N2 denote a vector nodes of O1 and O2). The process computation uses the ”Match-Based similarity” [7] as follows: MSim(E,E′) = ∑ (i,i′)∈Pairs(E,E′) SimCLS(i, i ′) Max(|E|, |E′|) . (1) Both sets E and E′ represent nodes of the same cluster belonging respectively to vectors V N1 and V N2. The neighborhood similarity, SimN , is computed using Equation 2: