The rise of the semantic web and the development of different technologies allow different actors to access knowledge found in different ontologies. This is not always obvious because technical constraints such as data volume and execution time are determining factors in the choice of an alignment algorithm. Among the solutions for scaling, the extraction of ontological entities as well as for partitioning methods can be complementary to alignment techniques, given the reduction in the size of the ontologies to be aligned, and therefore the reduction in execution time. In this article, we propose a new alignment method based on the extraction of concepts and labels as well as the creation of correspondences in an automatic way. Indeed, we emphasize that this method does not require any calculation of similarity distance. The obtained results during the evaluation of our method show its effectiveness and can be a decisive turning point for the different existing alignment methods.
Nowadays, ontologies have become one of the most important research orientations,
especially with the advent of the Semantic Web. An ontology is defined as the
conceptualization of objects recognized as existing in a domain, their properties and the
relationships between them [
This paper is organized as follows: Section 2 is a state of the art that presents the different alignment strategies and focuses on related work. In section 3, we describe our extraction based alignment method for large ontologies. Section 4 is an experimental study that illustrates the results and performance of our method. Section 5 presents a discussion of the results obtained. Finally, Section 6 concludes the document and provides an overview of the directions for future work. 2
Alignment consists in determining the set of correspondences between two ontologies
by using or implementing solutions to different heterogeneity problems. Several
alignment techniques, based on different criteria, are currently proposed in the
literature [
The choice of one technique or another or the composition of several of them is not an
easy task. Several studies complement their alignment results by using WordNet [
One of the solutions for scaling involves the possibility of partitioning ontologies into
blocks before performing alignment [
There are several approaches to partitioning. Graph-based approach applies
graphbased algorithms to decompose ontology [
In this section, we propose an ONTology Extraction Method (ONTEM) based on an extraction strategy that is, to our knowledge, almost non-existent in ontology alignment work. The proposed method consists of four main steps: 1) Preprocessing, 2) Common entities identification, 3) Mapping generation, 4) Alignment generating. The general architecture of ONTEM is illustrated in Figure 1.
O1
O2 g n i s s e c o r p e r P s e i t n iten itcao n i
f o it
n m e om Id C s n g io ipn tra p e a n M eG t n en io t a m r ign en l e A G
Intersection
O1 Mappings O2
To facilitate the process of comparing the ontological terms labelling classes and their
properties (i.e. calculating the distances between their character strings), it is very
important to perform a number of pre-processing operations. They significantly
improve alignment results. In addition, when aligning based on synonym extraction,
preprocessing operations facilitate their recognition by lexical databases and/or synonym
dictionaries. The classes of an ontology are extracted after the conceptualization of
the targeted domain according to the objectives to be achieved and the application that
will use the ontology. We will call anchoring a concept of one ontology matched with
a single concept of another ontology, of the same name and meaning as it.We are
dealing with two types of anchor pairs: those obtained from the intersection of
concepts and those obtained from common labels. The approach exploits the richness of
the concept labels. Labels are made up of several words [
Two linguistic techniques used are. Normalisation which consists of : 1) transforming all the characters of the ontological terms into lower-case letters, 2) stripping of special characters, spaces, and numbers, 3) elimination of coordinating conjunctions, articles, prepositions, 4) elimination of words designating sets or elements (composition words). Tokenisation, which is a lexical analysis that consists in transforming a character stream into a token stream by an analyzer (Tokenizer) that recognizes punctuation, white characters, etc.
The pre-processing phase is illustrated by Algorithm1
Inputs:Ontology O Outputs:List_of_Concepts, List_of_Labels, Index_of_Concepts, Index_of_Labels Begin For Each Concept of O do
CN=Normalize(Concept) CT=Tokenize (CN) Add(Concept,CT) Add((Concept,CT),INDEX_CONCEPTS) For Each Label ofConcept
LN=Normalize (Label) LT=Tokenize (LN) Add(Label,LT)
Add((Label,LT),INDEX_LABELS)
EndFor EndFor Return (List_of_Concepts, List_of_Labels, Index_of_Concepts, Index_of_Labels) End
Since the ontologies treated concern the same field, it is obvious that they share common elements. This observation leads us directly to consider the common elements to both ontologies. The elements common to both ontologies can obviously be concepts, labels, properties, as well as relationships. Given two voluminous ontologies, the objective is to match the concepts of the first ontology with the concepts of the second ontology. To do this, we will use the "Intersection" operation to determine the common concepts to both ontologies. 3.2.1 Intersection of common entities Given a domain D and two ontologies O1ϵ D and O2ϵ D.
Let LCO1 and LCO2 the list of concepts of Ontology 1 and Ontology 2 respectively. Let LLO1 and LLO2 the list of labels of Ontology 1 and Ontology 2 respectively. The set LIC (List of Intersection of Concepts) will form the list of common concepts to both ontologies.
LIC=LCO1∩ LCO2.
In the same way, the set LIL (List of Intersection of Labels) will form the list of common labels to both ontologies.
LIL=LLO1∩ LLO2.
The intersection of common entities (concepts and labels) is performed by Algorithm2 for concepts and Algorithm3 for labels.
Inputs: LCO1, LCO2 Outputs: LIC Begin LIC = Intersection (LCO1,LCO2) Return (LIC) End
Inputs: LLO1, LLO2 Outputs: LIL Begin LIL = Intersection (LLO1,LLO2) Return (LIL)
End 3.2.2 Difference of concepts In this phase, we retain only the non-composed concepts. We will only select concepts that are not composed and do not belong to LIC. A compound concept being a name that contains at least one character" _". Let LNCCO1 the list of non-composed concepts of Ontology 1.
Let LCSSO1, list of Concepts for Searching Synonyms of Ontology 1. The difference on concepts is performed by Algorithm 4.
Inputs: LNCCO1, LIC Outputs: LCSSO1 Begin LCSSO1 = Difference (LNCCO1,LCI) Return (LCSSO1)
End 3.3
The mapping discovery process, called ontology alignment, is a function f that applies to two ontologies O1 and O2, with a set of parameters p (weights, thresholds, etc.) and a set of external resources r, and produces a set of mapping A.
A=f(O1,O2,p,r).
Alignment consists of several steps [
Inputs: LIC, ICO1, ICO2 Outputs: LCC Index_Mappings Begin For Each Conceptϵ LIC do
Entity_Name_1=Read(Concept, ICO1) Entity_Name_2=Read(Concept,ICO2) Add (Entity_Name_1,LCC) Add(Entity_Name_2,LCC)
Add (Entity_Name_1,Entity_Name_2, Index_Mappings) EndFor Return(LCC) End The algorithm for generating concept matches from the list of common labels is represented by the Algorithm 6.
Inputs :LIL, ILO1, ILO2, ICO2, Index_Mappings Outputs :LLC, Index_Mappings /* Updated */ Begin For Each (Label ϵ LIL) do /* browse the labels of LIL */
Entity_Name_1=Read(Label, ILO1)
Entity_Name_2=Read(Label, ILO2) Found=0
While (Entity_Name_2 hasValues Is True) && (Found==0) /* Retrieving Entity_name2 not yet used */
IF (Entity_Name_1,Entity_Name_2) not found in Index_Mappings Add(Entity_Name_1,LLC) Add(Entity_Name_2,LLC) /* insert Entity_Name_1 and Entity_Name2 in LLC */ Add(Entity_Name_1,Entity_Name_2, Index_Mappings) /*Update of the Index_Mappings */
Found=1; Else
Entity_Name_2=Read(Label, ILO2)
EndIf
EndWhile EndFor Return(LCL, Index_Mappings)
End 3.4
The alignment will be created automatically from the LCC, LCL and LCW lists. The algorithm for generating the alignment is illustrated by the Algorithm 7.
Inputs: LCC, LCL, LCW Outputs: F-ALIGNE For Each Concept ϵ LCC do
Entity_Name_1=Read(Concept, LCC) Entity_Name_2=Read(Concept, LCC) Add(Entity_Name_1,F-ALIGNE)
Add(Entity_Name_2,F-ALIGNE) EndFor For Each Concept ϵ LCL do
Entity_Name_1=Read(Concept, LCL)
Entity_Name_2=Read(Concept, LCL) Add(Entity_Name_1,F-ALIGNE)
Add(Entity_Name_2,F-ALIGNE) EndFor For Each Concept ϵLCW do
Entity_Name_1=Read(Concept, LCW) Entity_Name_2=Read(Concept, LCW) Add(Entity_Name_1,F-ALIGNE)
Add(Entity_Name_2,F-ALIGNE) EndFor Return (F-ALIGNE) End
ONTEM is a fully automatic method and requires no user intervention during the alignment process. However, the expert who is a knower of the field can confirm, suggest other alignments 4
To highlight the validity of our method, we will compare the alignments that
ONTEM has produced with a reference alignment contained in the
Oaei_LargeBio_Track_2018 section [
The ONTEM prototype was developed on the Eclipse Helios platform, using the
java and APIjena2.4 programming language, as well as the SPARQL semantic graph
reading language.The machine on which the work was performed has an Intel® Core
TM(2) Duo CPU E7500 2.93 Ghz 2.94 Ghz, 4.0 GB RAM, 32bit operating system,
Windows 7 Professional N.The evaluation metrics Precision, Recall and F-measure
were used to compare our ONTEM method with other pioneering methods in the
field, namely: AML, FCAMapX, LogMapBio, LogMap, LogMapLt, XMap,
POMAP++[
In this article, we have focused on the issue of large-scale ontology alignment for the Semantic Web. Indeed, the variety of ontologies of the same domain in the semantic web has led to heterogeneity and therefore to the development of ontology alignment methods. For more than a decade, ontology alignment methods have been attempting to solve heterogeneity and ontology matching problems. Today in many real applications such as in the medical field, the size of ontologies is very large and current alignment methods are faced with many challenges such as lack of memory and long processing times. We have shown that our ONTEM method stands out from the crowd of existing methods by its originality. It makes it possible to build new architectures based on existing methods that will boost ONTEM for much better results, because the purpose of all the work is to be able to make ontology-based information systems interoperable. To this end, this document provides an appropriate solution to this type of problem.A prototype has been set up to support the proposed approach. With this realization, we were able to evaluate our comments and compare them with other recognized methods in this field such as the OAEI_LargeBio_Track section of the 2018 campaign.
In our future work, we plan to consolidate our method to better support the alignment of full-scale large ontologies. We have already started to address this issue, but updating the test database poses other challenges, in terms of the ontological languages used and the evolving semantic description formalisms.