Multilingual Ontology Matching based on Wiktionary Data ♣ Accessible via SPARQL Endpoint

Abstract

Interoperability is a feature required by the Semantic Web. It is provided by the ontology matching methods and algorithms. But now ontologies are presented not only in English, but in other languages as well. It is important to use an automatic translation for obtaining correct matching pairs in multilingual ontology matching. The translation into many languages could be based on the Google Translate API, the Wiktionary database, etc. From the point of view of the balance of presence of many languages, of manually crafted translations, of a huge size of a dictionary, the most promising resource is the Wiktionary. It is a collaborative project working on the same principles as the Wikipedia. The parser of the Wiktionary was developed and the machine-readable dictionary was designed. The data of the machinereadable Wiktionary are stored in a relational database, but with the help of D2R server the database is presented as an RDF store. Thus, it is possible to get lexicographic information (definitions, translations, synonyms) from web service using SPARQL requests. In the case study, the problem entity is a task of multilingual ontology matching based on Wiktionary data accessible via SPARQL endpoint. Ontology matching results obtained using Wiktionary were compared with results based on Google Translate API.

1 Introduction

Ontology matching is the process of finding correspondences between ontologies to allow them to interoperate. There are different methods, algorithms and systems designed for ontology matching [ 9 ].

A relatively new direction is concerned with an alignment of ontologies presented in different languages, i.e. multilingual ontology matching [ 8 ], [ 24 ]. There are different strategies related to multilingual ontology matching [ 24 ]: (1) the indirect alignment strategy based on composition of alignments, (2) the direct matching between two ontologies, i.e., without intermediary ontologies and with the help of external resources (translations). The latter strategy is used in this work.

The Ontology Alignment Evaluation Initiative (OAEI) 1 was launched in 2004 with the goal of estimating and comparing different techniques and systems related to ontology alignment. OAEI provides some multilingual datasets (ontologies and reference alignments), which were used in this work in order to evaluate the ontology matching system.

The multilingual ontology matching platform is presented in this work. COMS (Context-base Ontology Matching System) [ 15 ] implements the multilingual ontology matching based on Google Translate API and the data of the English Wiktionary and SPARQL technology.

The Wiktionary (www.wiktionary.org) is a multilingual and multifunctional dictionary. The Wiktionary contains not only word’s definitions, semantically related words (synonyms, hypernyms, etc.), translations, but also the pronunciations (phonetic transcriptions, audio files), hyphenations, etymologies, quotations, parallel texts (quotations with translations), figures (which illustrate meaning of the words).

Wiktionary is popular since it is freely available and contains huge database of words with translations to many languages. The salient properties of the Wiktionary are the multilinguality, the size, and the speed of evolution. It is difficult to compare dictionaries with the Wiktionary, since data quickly become outdated. E.g. the PanDictionary was compared with the Wiktionary data obtained in the year 2008, when it has 403 413 translations [ 19 ]. Two years later, in 2010, the English Wiktionary contained twice as much translations (964 019). 2 So, the Wiktionary is permanently growing in number of entries and in the scope of languages. Now the English Wiktionary

1 See http://oaei.ontologymatching.org

2 See http://en.wiktionary.org/wiki/User:AKA_MBG/ Statistics:Translations contains entries in about 770 different languages. The Wiktionary data are used: • In machine translation between Dutch and

Afrikaans [ 21 ]; • In the text parsing system NULEX, where some Wiktionary data (verb tense) were integrated with WordNet and VerbNet [ 18 ]; • In a speech recognition and speech synthesis as a basis for the rapid pronunciation dictionary creation [ 10 ].

The Resource Description Framework is a data model for representing information about World Wide Web resources. SPARQL [ 1 ] is a query language for this data model. It is standardized by the World Wide Web Consortium. Now SPARQL is supported by most RDF triple store.

With the help of D2R server [ 4 ] the data extracted from the Wiktionary are presented in the form of RDF store. So, lexicographic information extracted from the Wiktionary is accessible by using SPARQL requests. In the case study, the problem entity is a task of multilingual ontology matching based on Wiktionary data accessible via SPARQL endpoint.

The next section describes system architecture consisting of the ontology matching system, Wiktionary relational database, D2R server and SPARQL client. Section 3 presents multilingual ontology matching experiments based on Wiktionary and Google Translate API. The discussion concludes the paper.

2 System architecture

In this section the developed platform will be described. The key components are a Wiktionary relational database, COMS ontology matching system [ 15 ], and D2R server which provides access to the machinereadable Wiktionary via SPARQL endpoint.

2.1 Machine-readable Wiktionary

There is an approach where the data are extracted from different types of wiki sites for the further processing and semantic search [ 21 ]. In that approach it was developed special services that export structured data into RDF/XML format. These services were designed and tailored to specific wiki engines (MediaWiki, DokuWiki).

Our work had the more modest goal of extracting data from only one type of wiki site (Wiktionary), moreover, only one Wiktionary language edition (English). The important fact is that Wiktionary entries have well-defined structure. However this structure is specified not at the level of MediaWiki, but at the level of texts of Wiktionary entries. Taking into account the structure of Wiktionary entry yield much more interesting information than just “structured data in RDF/XML format”. The following data was extracted from the English and Russian Wiktionaries: definitions, thesaurus and translations. An example of data extracted from the “beautiful” English Wiktionary entry The process of multilingual ontology matching involves two steps. First COMS translates the entities source to target ontology language. Then it applies automatically the following monolingual matching strategies. Fig. 2 shows two ontologies' automatic matching strategy and evaluation. Jena (http://jena.sourceforge.net) is used to parse ontology elements.

2.2.1 String Matching Strategy

Different string matching algorithms can be used here. There is a good survey [ 3 ] on the different string similarity methods to calculate string distance from edit-distance (e.g. Levenstein distance, Monger-Elkan distance, Jaro-Winkler distance) to token-based distance functions (e.g. Jaccard similarity, TF-IDF or cosine similarity, Jense-Shannon distance).

We use the Jaro-Winkler distance [ 25 ] and SmithWaterman algorithm [ 23 ] implemented in SimMetrics5 and SecondString as our string matching methods. The threshold for Jaro-Winkler distance is 0.9. SmithWaterman algorithm can help find the similar region for two strings.

2.2.2 Structure Matching Strategy

Different structure matching strategies are implemented as following: 1. If two elements of two ontologies' triples (subject, predicate and object) are the same, the third element is assumed the same. For example, if the range and domain of two relations are the same, it means that the relations are the same. In future work, this will be extended to compare the common triples in the hierarchy. 2. If the subclasses of two classes are the same, these two classes are assumed the same. In 5 SimMetrics and SecondString are Java-based opensource packages used for string matching.

future work, this will be extended to compare the common classes in the hierarchy.

Expanding tree method [ 17 ]. Ontology is expanded as a tree and set weights in the tree to calculate ontology concept similarity. The different levels are given different weights depending on the depth of the compared classes. The first level concepts, which get the weight as 3 are the class’ subclasses and each relationship where it is domain or range. The second level concepts which get weight 2, are depending on the first level concepts’ subclasses and their relationship’s ranges. Similarity we can get the third level concepts, with weight 1, based on the second level concepts. We treat ontology matching as asymmetric. For example, a small ontology may perfectly match some parts of large ontology, the similarity between the small ontology and large ontology is 1.0 then, but not vice versa. The similarity between two concepts is computed as: sim ( x, y) = ∑ wmatched- concepts

∑ wxi

2.2.3 Lexical Matching Strategy

One of our ontology matching strategies uses the WordNet (version 3.0). WordNet 6 is based on psycholinguistic theories to define word meaning and models not only word meaning associations but also meaning-meaning associations [ 7 ]. WordNet consists of a set of synsets. Synsets have different semantic relationships such as synonymy (similar) and antonymy (opposite), hypernymy (superconcept)/hyponymy (subconcept) (also called Is-A hierarchy / taxonomy), meronymy (part-of) and holonymy (has-a). The paper [ 16 ] provides an overview of how to apply WordNet in the ontology matching. In COMS, we use WordNet as the lexical dictionary.

6 http://wordnet.princeton.edu

WordNet-Similarity 7 has implemented several WordNet-based similarity measures in a Perl package. Java WordNet::Similarity8 is a Java implementation of WordNet::Similarity. Jiang-Conrath [ 11 ] measure is chosen with threshold 1.0 to find corresponding classes in ontology matching. Jiang-Conrath measure is derived from the edge-based notion by adding the information content as a decision factor. jcn = 1 (IC (synset1) + IC (synset 2) - 2 * IC (lcs))) where lcs is the super concept of synset1 and synset2, IC is the information content (of a synset).

For example, there are seven senses for the entry noun school hypernym relation in WordNet (fragment): Sense 1 school -- (an educational institution; "the school was founded in 1900")

=> educational institution -- (an institution dedicated to education)

=> institution, establishment -- (an organization founded and united for a specific purpose) => organization, organisation -- (a group of people who work together)

=> social group -- (people sharing some social relation)

=> group, grouping -- (any number of entities (members) considered as a unit)

=> abstraction, abstract entity -- (a general concept formed by extracting common features from specific examples)

=> entity -- (that which is perceived or known or inferred to have its own distinct existence (living or nonliving))

Sense 2 school, schoolhouse -- (a building where young people receive education; "the school was built in 1932"; "he walked to school every morning") => building, edifice -- (a structure that has a roof and walls and stands more or less permanently in one place; "there was a three-story building on the corner"; "it was an imposing edifice"). 7 http://www.d.umn.edu/~tpederse/similarity.html 8 http://www.cogs.susx.ac.uk/users/drh21/ D2R server uses RDF and SPARQL languages in order to provide access to the relational database [ 4 ]. System takes SPARQL queries from the web and rewrites them to SQL queries via a specially prepared file (D2RQ mapping file).

The ontology matching system takes translation from the machine-readable Wiktionary with the help of D2R server (Fig. 4).

The D2RQ mapping file has to be created only once. After that it is possible to access to the relational database via SPARQL. SPARQL queries will be automatically translated on-the-fly into SQL by D2RQ platform. Therefore there is no need to replicate the database into RDF store.

Fig. 4. Architecture of the platform integrating the ontology matching system with the machine-readable Wiktionary accessible via SPARQL queries

A simple Wiktionary SPARQL client was written in Java (as a part of COMS ontology matching system). It can obtain a list of translations from the source to the target language using Wiktionary data.

3 Experiments

The experiments are based on one benchmark track of OAEI. The reference ontology “test 101” is in English. This reference ontology contains 33 named classes, 24 object properties, 40 data properties, 56 named individuals and 20 anonymous individuals. The “test 206” of benchmark contains one ontology in French. Therefore one reference ontology (in English) is matched to French ontology.

In the “test 206” ontology in French the most part of words are presented in a canonical form (lemma). There are only a few words which are presented in noncanonical form, e.g. French words “articles”, “auteurs”, “éditeurs”, “réalisateurs”, “pages”, “chapitres”, “communications”. Different word forms are recognized by the Google Translate system, but it is not taken into account by translation system based on the machine-readable Wiktionary.

Thus, our system translates labels from English to French first by using multilingual English Wiktionary, before applying monolingual matching procedures.

3.1 Wiktionary Database and SPARQL queries

The dump of the English Wiktionary (as of October 30, 2010) was the source data for our experiments. The created database of the machine-readable Wiktionary contains: - 1 731 784 total entries; - 269 405 English entries - 154 990 French entries; - 964 019 total number of translations; - 50 617 number of translations from English to

French.

This database was used for translation in the ontology matching system. This database was accessed via SPARQL queries. Most SPARQL queries are simple and short [ 2 ]. However, it turns out that it is not so in our case (Table 1).

Table 1 contains the example of the SPARQL request for the machine-readable Wiktionary. Input data for this request are (i) a language code (with value “en”, i.e. English language), (ii) a Wiktionary entry (“rain cats and dogs”). Different colors of the rows in the Table 1 show different parts of the request, where one part corresponds to one table in the database.

The result of this request is translations of the English phrase “rain cats and dogs” into all languages presented in the Wiktionary. The part of the answer is presented in Table 2. Several SPARQL queries to the Wiktionary are presented on the wiki page of the project.9

9 See http://code.google.com/p/wikokit/wiki/

d2rqMappingSPARQL Ontology labels are often concatenated, e.g. "dateDePublication", "IntervalleDePages", "ExtraitCompilation". Google Translate system can recognize the label and translate directly. The machine-readable Wiktionary can’t understand the concatenated label. In order to properly translate, labels are split into sequence of their constituent words. For example, "dateDePublication" is separated as “date De Publication”.

In the reference alignment, one element is coming from “test 101” that is in English, one element is coming from “test 206” that is in French, and their similarity result. The total number of correct translations (the original French word and translated word compared to reference alignment) before applying ontology matching strategy by English Wiktionary is 44, and correct number by Google is 60.

The correct translation of English Wiktionary is lower, it is because that the Google gives the same word as translation if the word is not in the dictionary, e.g., “isbn”, “url”, “lccn”, etc.. However, there is no translation in English Wiktionary to this case (see table 3, “isbn” example). On the other hand, Google translation API only provides one meaning translation of the words while English Wiktionary provides multiple meanings (if the word has) translation, for example, “Université” is only translated to “University” in Google, while is translated to “university; school” in English Wiktionary (see table 3). Google is good at translation the concatenated word, for example, “nomCourt” is translated “Shortname” directly (see table 3). InBook Part Book Conference Collection isbn key Chapters Editor After we get the translation of the French ontology, COMS applies automatically the following monolingual matching strategies as described in Section 2.2. If there is no translation of the word, the original of element of the ontology is used to string matching, for example, “isbn” in Wiktionary case. Even COMS can get separate meaning of the concatenated word, but COMS doesn’t support the different combination of the translation, for example, “nomCourt” is translated to “noun; name; short; court” and the correct translation is “shortName” (see table 3), COMS can’t achieve to “shortName”. “Film” is interpreted to “movie; film; cinema; flick; motion picture”, COMS can recognize it is “MotionPicture”.

The other matching strategies, such as WordNet is applied, e.g. “school” and “institution” similarity is 1.25 (see section 2.3). Structure matching strategy is applied, for example, in “Test 206”, object property “articles” has domain “Revue” that interpreted as “Review” in Google and range “Article”. In “Test 101” object property “articles” has domain “Journal” and range “Article”. Since “articles” is similar “articles” and “Article” is similar “Article”, even “Review” and “Journal” has no string similarity, based on structure similarity rules, “Revue” and “Journal” is similar. The final alignment result is based on the matching strategies presented in section 2.2.

There are different evaluation measures proposed in the OAEI, e.g., compliance and performance measures. The compliance measures consist of Precision, Recall, Fallout, F-measure, Overall, etc. Based on [ 6 ], the definition of precision and recall are:

Definition (Precision).Given a reference alignment R, the precision of some alignment A is given by

R ∩ A P( A, R) = R( A, R) =

It measures a valid possibility for ex post evaluations.

Definition (Recall). Given a reference alignment R, the recall of some alignment A is given by

R ∩ A | A | | R |

The provided reference alignment has 97 elements, which means | R | = 97.

The retrieved alignment based on English Wiktionary has 54 elements, which means | A |= 54, intersection R ∩ A = 53

Precision is (see table 4):

R ∩ A 53 P( A, R) = =

| A | 54 Recall is (see table 4):

R ∩ A 53 R( A, R) = = | R | 97 process and the mapping activity [ 8 ], by using an information about a domain of the ontology [ 5 ]). 3) Now, in the experiment, (1) French words are translated into English, (2) monolingual matching procedures based on English WordNet were applied. There is an idea to use the free French WordNet 10 as an additional resource for the matching of two ontologies in English and French languages.

The Wiktionary parser development will be continued in future work, aiming at an extraction of quotes and Wiktionary context labels.

References

[1] M. Arenas and J. Perez . Querying Semantic Web Data with SPARQL . 30th ACM Symposium on Principles of Database Systems (PODS) , June 2011 , Athens, Greece. http://web.ing.puc.cl/~jperez/papers/pods11b.pdf

[2] M. Arias , J. D. Fernández , M. A. Martínez-Prieto , P. Fuente . An Empirical Study of Real-World SPARQL Queries . In: 20th International World Wide Web Conference, Hyderabad, India, March 28th, 2011 . http://arxiv.org/abs/1103.5043

[3] W. W. Cohen , P. Ravikumar , S. E. Fienberg , A comparison of string distance metrics for namematching tasks , 2003 .

[4] R. Cyganiak and C. Bizer. D2R Server: A Semantic Web Front-end to Existing Relational Databases . Berliner XML Tage , Berlin, Germany, September 2006 . http://richard.cyganiak.de/2008/papers/d2rserver-bxmlt2006.pdf

[5] M. Espinoza , A. Gomez-Perez , and E. Mena . Enriching an ontology with multilingual information . In Proc. of 5th European Semantic Web Conference (ESWC'08) , Tenerife, (Spain), June 2008 . http://sid.cps.unizar.es/PUBLICATIONS/POSTSC RIPTS/eswc08-localization.pdf

[6] J. Euzenat , R. G. Castro , and M. Ehrig . D2 . 2.2: Specification of a benchmarking methodology for alignment techniques . Technical report, NoE Knowledge Web project deliverable , 2004 .

[7] R. Ferrer- i-Cancho. The structure of syntactic dependency networks: insights from recent advances in network theory . In: L. V., A. G . (Eds.), Problems of quantitative linguistics, 2005 , 60 - 75 .

[8] B. Fu , R. Brennan , D. O'Sullivan. Cross-Lingual Ontology Mapping and Its Use on the Multilingual Semantic Web . In: 1st Workshop on the Multilingual Semantic Web, at the 19th International World Wide Web Conference (WWW 2010 ), Raleigh, USA, April 27th, 2010 , CEUR Vol . 571 , 2010 , 13 - 20 . http://www.tara.tcd.ie/jspui/handle/2262/39188 10 See http://alpage.inria.fr/~sagot/wolf-en.html

[9] M. Gollapalli and X. Li . Survey on Data Linkage and Ontology Matching Techniques . Submitted for VLDB , 2011 .

[10] Qingyue He . Automatic Pronunciation Dictionary Generation from Wiktionary and Wikipedia . Thesis. Karlsruhe Institute of Technology. 2009 . http://csl.anthropomatik.kit.edu/index.php?id= 25

[11] J. J. Jiang and D. W. Conrath . Semantic similarity based on corpus statistics and lexical taxonomy . In: International Conference Research on Computational Linguistics , Taiwan, 1997 .

[12] A. Krizhanovsky . The comparison of Wiktionary thesauri transformed into the machine-readable format . 2010 . http://arxiv.org/abs/1006.5040

[13] A. Krizhanovsky . Transformation of Wiktionary entry structure into tables and relations in a relational database schema . 2010 . http://arxiv.org/abs/1011.1368

[14] Z. Kurmas . Zawilinski: a library for studying grammar in Wiktionary . In: Proceedings of the 6th International Symposium on Wikis and Open Collaboration , Gdansk, Poland, July 07 - 09 , 2010 . http://www.cis.gvsu.edu/~kurmasz/Software/#Zawi linski

[15] Feiyu Lin , J. Butters , K. Sandkuhl , F. Ciravegna . Context-based Ontology Matching: Concept and Application Cases . In: 10th IEEE International Conference on Computer and Information Technology (CIT) . Bradford, UK , June 29-July 1 2010 .

[16] Feiyu Lin and Kurt Sandkuhl . A survey of exploiting wordnet in ontology matching . In: IFIP AI , 2008 , 341 - 350 .

[17] Feiyu Lin and Kurt Sandkuhl . A new expanding tree ontology matching method . On the Move to Meaningful Internet Systems 2007: OTM 2007 Workshops. Lecture Notes in Computer Science , 2007 , Volume 4806 / 2007 , 1329 - 1337 .

[18] C. McFate and K. Forbus. NULEX : An OpenLicense Broad Coverage Lexicon. (accepted). In: The 49th Annual Meeting of the Association for Computational Linguistics: Human Language Technologies . Portland, Oregon, USA, June 19-24, 2011 . http://www.aclweb.org/anthology/P/P11/P11- 2063.pdf

[19] Mausam , S. Soderland , O. Etzioni , D. S. Weld , K. Reiter , M. Skinner , M. Sammer , J. Bilmes . Panlingual Lexical Translation via Probabilistic Inference . Artificial Intelligence Journal (AIJ) . Volume 174 Issues 9-10 , 2010 , 619 - 637 . http://www.aaai.org/ocs/index.php/AAAI/AAAI10/ paper/viewFile/1688/2281

[20] Lourens van der Meij , Antoine Isaac, Claus Zinn . A Web-based Repository Service for Vocabularies and Alignments in the Cultural Heritage Domain . Proceedings of the 7th Extended Semantic Web Conference , (ESWC 2010 ). Heraklion, Greece, 30 May - 3 June 2010. http://www.few.vu.nl/~aisaac/papers/STITCHRepository-ESWC10.pdf

[21] F. Orlandi and A. Passant . Semantic Search on Heterogeneous Wiki Systems. In: Proceedings of the 6th International Symposium on Wikis and Open Collaboration : 1 - 10 , ACM, WikiSym10 . 2010 . http://www.deri.ie/about/team/member/fabrizio_orl andi/

[22] P. Otte and F. M. Tyers . Rapid rule-based machine translation between Dutch and Afrikaans . In: 16th Annual Conference of the European Association of Machine Translation, EAMT11 , 2011 . http://xixona.dlsi.ua.es/~fran/publications/eamt201 1a.pdf

[23] T. F. Smith , M. S. Waterman . Identification of common molecular subsequences . Journal of molecular biology Volume 147 ( 1981 ) 195 - 197 .

[24] C. Trojahn , P. Quaresma , R. Vieira . An API for Multi-lingual Ontology Matching . In Proceedings of LREC , 2010 . http://disi.unitn.it/~p2p/RelatedWork/Matching/Tro jahn_ 691 _Paper.pdf

[25] W. E. Winkler . The State of Record Linkage and Current Research . Problems, Technical Report , 1999 .

♣ Part of this work was financed by the Foundation (The Swedish Institute), project CoReLib . Some parts of the research were carried out under projects funded by grants # 09 - 07 -00066, # 09 - 07 -00436 and # 11 -01 - 00251 of the Russian Foundation for Basic Research, and project of the research program “Intelligent information technologies, mathematical modelling, system analysis and automation” of the Russian Academy of Sciences. This work is partly funded by the project DEON (Development and Evolution of Ontologies in Networked Organizations) based on a grant from STINT (The Swedish Foundation for International Cooperation in Research and Higher Education); grant IG 2008-2013.