This paper presents and discusses the results produced by KEPLER for the 2017 Ontology Alignment Evaluation Initiative (OAEI 2017). This method is based on the exploitation of three different strategy levels. The proposed alignment method KEPLER is enhanced by the integration of powerful treatments inherited from other related domains, such as Information Retrieval (IR) [1]. For scaling, the method is equipped with a partitioning module. For the management of multilingualism, KEPLER develops a well-defined strategy based on the use of a translator, and this provides very encouraging results.
1.1
The proposed method, KEPLER, exploits besides the classic techniques, an external resource, i.e., a translator to deal with multilingualism. KEPLER implements an alignment strategy which aims at exploiting all the wealth of the used ontologies. 1.2
The main idea of KEPLER is to exploit the expressiveness of the OWL language to detect and compute the similarity between entities of two given ontologies through 6 complementary modules as presented in figure 1.
Entities are described using OWL primitives with their semantics. We can then consider ontology as a semantic graph where entities are nodes connected by links which are OWL primitives. These links have specified semantic primitives. Consequently, if two ontologies in the same domain are similar, their semantic graphs are also the same.
Parsing and pretreatment This module allows to extract the ontological entities initially represented by a primary form of lists. In other words, at the parsing stage, we seek primarily to transform an OWL ontology in a well defined structure that preserves and highlight all the information contained in this ontology. Furthermore, in the resulting informative format, it has a considerable impact on the results of the similarity computation thereafter. Thus, we get couples formed by the entity name and its associated labels.
Partitioning This module aims at splitting ontologies into smaller parts to support
the alignment task [
Translation : An originality of our system, is to solve the heterogeneity problem
mainly due to multilingualism, given the importance of this research area [
Indexation : Indexing is one of the novelties of our approach. It consists in reducing
the search space through the use of effective search strategy on the built indexes which
represent the input ontologies components. To enable faster searching, the driving idea
that was previously used in some works [
Candidate Mappings Identification : The role of this module is to find the entities in common between the indexes. Once the indexes are set up, the querying step of the latter is activated. Thus, the query implementation satisfies the terminology search and semantic aspects at once as we are querying documents in a vector representation that contain a given ontological entity and its synonyms obtained via WordNet. It is worthy to mention that indexes querying is done in both senses.
Filtering and Recovery : The filtering module consists of two complementary submodules, each one is responsible of a specific task in order to refine the set of primarily aligned candidates. At this stage, once the list of candidates is ready, the alignment method uses the first filter. We should note that indexes querying may includes a set of redundant mappings. Doing so, this filter eliminates the redundancy. It goes through the list of candidates and for each candidate, it checks if there are duplicates. If this is the case, it removes the redundant element(s). At the end of filtering phase, we have a candidates list without redundancy, however, there is always the concern of false positives, 3 https://wordnet.princeton.edu/ 4 https://www.bing.com/translator in fact, there was the need to establish a second filter. Once the redundant candidates are deleted, the system uses the second filter that eliminates false positives. This filter is applied to what we call partially redundant entities. An entity is considered as partially redundant if it belongs to two different mappings (i.e., being given three ontological entities e1, e2 and e3. If on the one hand, e1 is aligned to e2, and secondly, e1 is aligned to e3, this last alignment is qualified as doubtful. We note that our method generates (1 : 1) alignments. To overcome this challenge, the alignment method compares the topology of the two suspicious entities (e3 neighbors with e1 neighbors, e2 neighbors with e1 neighbors ) with respect to the redundant entity e1, and retains the couple having the highest topological proximity value. All candidates are subject of this filter, and as output we have the final alignment file.
Alignment Generation : The result of the alignment process provides a set of mappings, which are serialized in the RDF format. 2
Results 2.1
In this section, we present the results obtained by KEPLER in the OAEI 2017.
This track consists of two real world ontologies to be matched, the source ontology
describing the Adult Mouse Anatomy (with 2744 classes) and the target ontology is the
NCI Thesaurus describing the Human Anatomy (with 3304 classes). For this track, and
according to figure 2, KEPLER succeeded to extract 74% of correct mappings with a
precision about 95%. Figure 2 summarizes the evaluation metrics values for Anatomy
track. To this end, it is important to mention that KEPLER has managed to support the
ontologies of the Anatomy database thanks to the Ontopart module [
This dataset is composed of a subset of the Conference track, translated in nine
different languages (i.e., Chinese, Czech, Dutch, French, German, Portuguese, Russian,
Spanish and Arabic). With a special focus on multilingualism, it is possible to evaluate
and compare the performance of alignment approaches through these test cases. Based
on several previous contributions [
KEPLER uses a specific technique to determine the equivalence between ontology entities described in different natural languages. We chose to use the English as a pivot language. The use of a pivot language ensures greater consistency of obtained translations since it starts from the same text. In the different ontologies case, the method is ranked fourth with a recall value of 0:31 as depicted by figure 3.
Whereas in the same ontologies case, the method occupies the first place with a recall value of 0:52 as flagged by figure 4. In the scalability register, this track consists of finding alignments between the Foundational Model of Anatomy (FMA), SNOMED CT, and the National Cancer Institute Thesaurus (NCI). These ontologies are semantically rich and contain tens of thousands of classes. The Large BioMed Track consists of three matching problems, i.e., (1) FMANCI matching problem, (2) FMA-SNOMED matching problem and (3) SNOMED-NCI matching problem. KEPLER handles large ontologies in two phases: the first phase consists on partitioning the ontologies into a set of blocks and the second phase selects two suitable blocks giving the highest value of similarity to be aligned. KEPLER treated (Task 1: FMA-NCI small fragments)[Precision : 0.96 / Recall : 0.83] according to figure 5.
As depicted by figure 6, KEPLER processed also the task 3 of the LargeBio dataset
(FMA-SNOMED small fragments) with a Precision of 0:82 and Recall of 0:55. In the
Phenotype track, our method succeeds in processing only the DOID-ORDO sub-case
by identifying 1824 matches for 1237 expected ones.
In this paper, we briefly presented the alignment system KEPLER with comments of
the results obtained according to the OAEI 2017 tracks, corresponding to the SEALS
platform evaluation modalities. Several observations regarding these results were
highlighted, in particular the impact of the elimination of any ontological resource on the
similarity values. KEPLER is an ongoing work which borrows its idea from two
previous systems, CLONA [