The current work describes an automatic system especially designed for aligning cross-lingual ontologies. The CroLOM software, unlike existing systems, uses the Yandex translator, NLP techniques and a similarity computation based on the categories of the words and synonyms. CroLOM participated for the first time in OAEI2016 evaluation campaign and the results obtained are so far been quite promising. The paper also discusses some important issues related to multilingualism treatment.
Recently, with the growing number of ontologies defined in different languages,
multilingualism has become an issue of major interest in ontology matching field.
Multilingual ontology alignment, defined as the process of identification of semantic
correspondences between entities of different ontologies described in different natural language,
represents the solution to the problem of semantic interoperability between different
sources of distributed information [
In this study, we consider the approaches of the first category, since we develop an approach which implements a direct strategy. However, there are many exciting questions regarding these approaches to address the multilingualism issue. These questions are as follows: (1) Which machine translation should be used, (2) which translation path should be considered and (3) which ontologies features and dictionaries can be exploited. In the following paragraphs, we describe the points mentioned above.
First, several translators have been developed to translate automatically the text from one natural language to another. We can mention for example: Google, Bing, SDL and Gengo translators. Each translator has its specific characteristics such as: number of source/target languages and execution time. However, selecting one or several translators (by combining them) remains an open problem. This choice is crucial in ”direct approaches”, since they apply a monolingual matching techniques in cross-lingual ontology mapping.
Second, the translation path also plays an important role to resolve the heterogeneity problem. Two translation paths can be considered, (i) either considering the translation path from one to another or (ii) selecting a pivot language which is often the English language. This choice highly depends on available sources (dictionaries, thesaurus, etc.) in different natural languages. Most matching systems consider the translation path using English as a pivot language due to available sources in English language.
Finally, in some cases, the results of a translation machine could be poor, however, to avoid this situation some ontology features can be exploited such Description Logics.
Most matching systems which implement a direct translation approach uses a wellknown translators mentioned above. The current work uses also a direct matching approach. However, unlike existing approaches, it addresses the multilingualism challenge by using (a) the Yandex translator1, (b) a translation into a pivot language after applying NLP techniques and (c) a similarity computation based on the categories of the words and synonyms.
The rest of the paper is organized as follows. First, in Section 2, we discuss the top systems that participated in the last editions of the multifarm track. In section 3 we describe the CroLOM system. Section 4 contains the experiment results. Finally, some concluding remarks and future work are presented in Section 5. 2
In this section, we continue our previous work [
Table 1 summarizes the results of the top systems in the multifarm track.
The AUTOMSv2 system [
The multifarm track of OAEI 2015 contains our dataset in Arabic language (ADOM)
[
Except these systems, the results of XMap, LogMap and AML systems on multifarm track (includes Arabic) are slightly lower than previous editions of OAEI (i.e. in OAEI2014). According to the results obtained from the systems mentioned above, this is explained by the fact that the Arabic dataset brings an additional complexity to the multifarm track.
We have also observed that the best system (in all OAEI editions including this year) achieved an F-measure of 0.51. This is surprising, in spite of many research works that have been established in the field of multilingual ontology matching. 3
We summarize the process of our approach to provide a general idea of the proposed solution. It consists in the following successive phases: 3.1
CroLOM extracts first the entities of the input ontologies. Then, it employs NLP techniques to normalize the entities described in different natural languages. Unlike existing approaches, we have applied lemmatization, stemming and stopword elimination for each natural language separately before translation step. First, for each language considered by multifarm, we have established the stop words of each language in order to eliminate them from entities labels. Second, we have developed morphological algorithms to obtain lemmatization of the entities words.
This step is important 6, since one of matchers used is (1) based on string comparison algorithm to compute similarity and (2) the categories of the words are stoked in lemma form. 3.2
Once the entities are normalized, CroLOM uses the Yandex translator in order to translate the entities described in different natural languages in English as a pivot language. After translation, CroLOM employs for the second time the normalization step in order to eliminate the stop words of the English language from entities labels.
We have mentioned before that the translation path and the used translator play important role to resolve the multilingualism heterogeneity problem. Our choice for the Yandex translator is justified by the fact that it is ranked as the 4th largest search engine in the world and it has not previously used to align multilingual ontologies. However, we have chosen English as a pivot language because there a lot dictionaries that are available in English language. These dictionaries could be exploited in order to improve our system in the future. In addition, to compute the similarity between entities, we have used dictionaries (word categories and WordNet) in English. Due to automatic translation, we have observed that some stop words can be appeared in translated entities. For this purpose, we have employed the normalization for the second time. 3.3
Once the translation and standardization are carried out, CroLOM applies first, a case conversion by converting all entities words in lower case then it passes to the similarity computation step. Unlike existing systems, which use well known matchers, we have developed a matcher which calculates the similarity between entities based on the categories of the Words, string-based algorithm and synonyms using Wordnet7.
The matcher developed establishes a Cartesian product between the two entities words, then it returns the maximum similarity value using Levenshtein distance, similarity based on WordNet and similarity based on the categories of the words. The similarity based on the categories of the words has been adapted with some modification from the project ”Calculate Semantic Similarity” 8. The project has been developed to 6 This step allows to obtain good results such as the results of our previous work [19] (STRIM system) in instance matching. 7 http://wordnet.princeton.edu/ 8 https://sourceforge.net/projects/semantics/
System CroLOM LogMap AML
Track Multifarm Multifarm Multifarm match sentences, however we have modified the code in order to compute similarity between words. Finally, CroLOM applies a filter to select candidate correspondences which possess the maximum similarity value in each line of Cartesian product between entities. Then it applies a second a filter to identify the correspondences that possess similarity value upper than a given threshold. 4
The results obtained by running our CroLOM system on multifarm tracks of OAEI 2016 evaluation campaign are obtained from the following website: http://oaei. ontologymatching.org/2016/results/multifarm/index.html.
The multifarm[
The results obtained by our CroLOM system on multifarm are quite promising with F-measure equal to 36%. Comparing these results against the results of the systems which have participated in OAEI previous editions (Table 1), CroLOM with this first participation, is among the best systems with respect to F-measure. Regarding this year [Table 2], only AML (F-measure equals to 0.40) and LogMap (F-measure equals to 0.37) systems whose results are slightly better than CroLOM system.
The major drawback of CroLOM system is the execution time compared to other systems. We are working forward to identify this problem and improve our system. 5
In this paper, we have presented our CroLOM system, (not) yet another cross-lingual ontology matching system. CroLOM unlike existing approaches, applies first NLP techniques on each natural language before translation. Then, it uses the Yandex translator in order to translate all entities in English as pivot language. Finally, CroLOM computes the similarity between translated entities based on the category of the words and WordNet.
As future challenges, we aim to (1) improving the quality results of our system and especially the execution time, (2) conduct a survey study that addresses all the issues mentioned above, (3) taking into account the indirect approaches. 16. H. Paulheim, ”WeSeE-Match results for OEAI 2012”, In Proceedings of the 7th Workshop on Ontology Matching ISWC 2012, USA, 2012. 17. S. Hertling and H. Paulheim, ”WikiMatch results for OEAI 2012”, In Proceedings of the 7th
Workshop on Ontology Matching ISWC 2012, pp., USA, 2012. 18. M. El-Abdi, H. Souid, M. Kachroudi and S. Ben-Yahia, ”CLONA results for OAEI 2015”,
In Proceedings of the 10th Workshop on Ontology Matching ISWC 2015, USA, 2015. 19. A. Khiat, M. Benaissa and M. A. Belfdhal, ”STRIM results for OAEI 2015 instance matching evaluation”. In Proceedings of the 10th International Workshop on Ontology Matching colocated with the 14th International Semantic Web Conference (ISWC 2015), USA, 2015.