BabelNet is a very large, wide-coverage multilingual ontology. This resource is created by linking the largest multilingual Web encyclopedia { i.e., Wikipedia { to the most popular computational lexicon { i.e., WordNet. The integration is performed via an automatic mapping and by lling in lexical gaps in resource-poor languages with the aid of Machine Translation. The result is an \encyclopedic dictionary" that provides babel synsets, i.e., concepts and named entities lexicalized in many languages and connected with large amounts of semantic relations. BabelNet is available online at http://www.babelnet.org. In this paper we present a rst attempt at encoding BabelNet for the multilingual Semantic Web.
This paper provides a contribution to the LLOD vision by presenting a rst
encoding of BabelNet in RDF. BabelNet (http://www.babelnet.org) is a very
large multilingual semantic network obtained as a result of a novel integration
and enrichment methodology. This resource is created by linking the largest
multilingual Web encyclopedia { i.e., Wikipedia { to the most popular
computational lexicon { i.e., WordNet [
While the LOD is centered around DBPedia [
In the next Section we introduce BabelNet and brie y illustrate its features. Then, in Section 3 we provide statistics and in Section 4 we describe the RDF encoding of BabelNet. Finally, we give some conclusions in Section 5. 2
BabelNet [
Wikipedia sentences ... rst hydrogen balloon ight. ...interim balloon altitude... ...a British balloon near Be...
+ SemCor sentences ...at the balloon and... ...like a huge balloon, in...
...the balloon would go up...
Machine Translation system
Wikipedia cluster ballooning Montgol er brothers
Fermi gas balloon is-a gas hot-air balloon wind has-part is-a gasbag high wind is-a blow gas
WordNet Babel Synset balloonen, Ballonde, aerostatoes, globusca, pallone aerostaticoit, ballonfr, montgol erefr pairs of concepts (e.g., balloon is-a lighter-than-air craft). Each edge is labeled with a semantic relation from R, e.g., fis-a, part-of , . . . , g, where denotes an unspeci ed semantic relation. Each node v 2 V contains a set of lexicalizations of the concept for di erent languages, e.g., f balloonen, Ballonde, pallone aerostaticoit, . . . , montgol erefr g. We call such multilingually lexicalized concepts Babel synsets. Concepts and relations in BabelNet are harvested from the largest available semantic lexicon of English, WordNet, and a wide-coverage collaboratively-edited encyclopedia, Wikipedia. In order to build the BabelNet graph, we collect at di erent stages: a. from WordNet, all available word senses (as concepts) and all the lexical and semantic pointers between synsets (as relations); b. from Wikipedia, all encyclopedic entries (i.e., Wikipages, as concepts) and semantically unspeci ed relations from hyperlinked text.
An overview of BabelNet is given in Figure 2. The excerpt highlights that WordNet and Wikipedia can overlap both in terms of concepts and relations: accordingly, in order to provide a uni ed resource, we merge the intersection of these two knowledge sources. Next, to enable multilinguality, we collect the lexical realizations of the available concepts in di erent languages. Finally, we connect the multilingual Babel synsets by establishing semantic relations between them. Thus, our methodology consists of three main steps: 1. We combine WordNet and Wikipedia by automatically acquiring a mapping between WordNet senses and Wikipages. This avoids duplicate concepts and allows their inventories of concepts to complement each other.
Babel synsets) by using (a) the human-generated translations provided by Wikipedia (the so-called inter-language links), as well as (b) a machine translation system to translate occurrences of the concepts within sense-tagged corpora.
Language Lemmas Synsets Word senses English 5,938,324 3,032,406 6,550,579 Catalan 3,518,079 2,214,781 3,777,700 French 3,754,079 2,285,458 4,091,456 German 3,602,447 2,270,159 3,910,485 Italian 3,498,948 2,268,188 3,773,384 Spanish 3,623,734 2,252,632 3,941,039
Total 23,935,611 3,032,406 26,044,643 Table 1. Number of lemmas, synsets and word senses in the 6 languages currently covered by BabelNet.
found in WordNet, as well as all wikipedias in the languages of interest: in order to encode the strength of association between synsets, we compute their degree of correlation using a measure of relatedness based on the Dice coe cient. 3
In this section we provide statistics for BabelNet 1.0.1, obtained by applying the
construction methodology brie y described in the previous Section and detailed
in [
The overall mapping contains 89,226 pairs of Wikipages and WordNet senses they map to, covers 52% of the noun senses in WordNet, with an accuracy of about 82% estimated on a random sample of 1,000 items. 3.2
BabelNet currently covers 6 languages, namely: English, Catalan, French, German, Italian and Spanish. Its lexicon includes lemmas which denote both concepts (e.g., balloon) and named entities (e.g., Montgol er brothers). The second column of Table 1 shows the number of lemmas for each language. The lexicons have the same order of magnitude for the 5 non-English languages, whereas English shows larger numbers due to the lack of inter-language links and annotated sentences for many terms, which prevents our construction approach from providing translations.
In Table 2 we report the number of monosemous and polysemous words divided by part of speech. Given that we work with nominal synsets only, the numbers for verbs, adjectives and adverbs are the same as in WordNet 3.0. As for nouns, we observe a very large number of monosemous words (almost 23 million), but also a large number of polysemous words (more than 1 million).
Both numbers are considerably larger than in WordNet, because { as remarked above { words here denote both concepts (mainly from WordNet) and named entities (mainly from Wikipedia). 3.3
BabelNet contains more than 3 million concepts, i.e., Babel synsets, and more than 26 million word senses (regardless of their language). In Table 1 we report the number of synsets covered for each language (third column) and the number of word senses lexicalized in each language (fourth column). 72.3% of the Babel synsets contain lexicalizations in all 6 languages and the overall number of word senses in English is much higher than those in the other languages (owing to the high number of synonyms available in the English WordNet synsets). Each Babel synset contains 8.6 synonyms, i.e., word senses, on average, in any language. The number of synonyms per synset for each language individually ranges from a maximum 2.2 for English to a minimum 1.7 for Italian, with an average of 1.8 synonyms per language.
In Table 3 we show for each language the number of word senses obtained directly from WordNet, Wikipedia pages and redirections, as well as Wikipedia and WordNet translations. 3.4
We now turn to relations in BabelNet. Relations come either from Wikipedia hyperlinks (in any of the covered languages) or WordNet. All our relations are
English Catalan French German Italian Spanish Total WordNet 364,552 - - - - - 364,552 WordNet glosses 617,785 - - - - - 617,785 Wikipedia 50,104,884 978,006 5,613,873 5,940,612 3,602,395 3,411,612 69,651,382 Total 51,087,221 978,006 5,613,873 5,940,612 3,602,395 3,411,612 70,633,719 Table 4. Number of lexico-semantic relations harvested from WordNet, WordNet glosses and the 6 wikipedias.
English WWoikridpNedetia LAabrgaellotooungihs anotnyrpigeidofbaairgcrallfetdthwaitthregmasaionrs haeloafttedduaeirt.o its buoyancy.
German kEeiinneBnaElloigne nisatnterinieebnviecrhftugste.lbsttragende, gasdichte Hulle, die mit Gas gefullt ist und uber Italian pUrninpcaipllioondeiaAerrocshtiamteicdoe.e un tipo di aeromobile, un aerostato che si solleva da terra grazie al Spanish dUenloaseroustidatoos,doe gAlorbqou maeerdoestsaptiacroa, veosluarn,aenaeteronndaievnednooeplraoirpeulcsoamdao quune sueidsoir.ve del principio Table 5. Glosses for the Babel synset referring to the concept of balloon as aircraft'. semantic, in that they connect Babel synsets (rather than senses), however the relations obtained from Wikipedia are unlabeled.2 In Table 4 we show the number of lexico-semantic relations from WordNet, WordNet glosses and the 6 wikipedias used in our work. We can see that the major contribution comes from the English Wikipedia (50 million relations) and Wikipedias in other languages (a few million relations, depending on their size in terms of number of articles and links therein). 3.5
Each Babel synset naturally comes with one or more glosses (possibly available in many languages). In fact, WordNet provides a textual de nition for each English synset, while in Wikipedia a textual de nition can be reliably obtained from the rst sentence of each Wikipage3. Overall, BabelNet includes 4,683,031 glosses (2,985,243 of which are in English). In Table 5 we show the glosses for the Babel synset which refers to the concept of balloon as `aircraft'. 3.6
BabelNet also includes a sense-tagged corpus containing the sentences input to
the Machine Translation system. The corpus, called BabelCor, is built by
collect2 In a future release of the resource we plan to perform an automatic labeling based
on work in the literature. See [
We now introduce a rst RDF encoding of BabelNet. Other encodings, including
one in the Lemon RDF model [
An excerpt of the RDF Babel synset representation follows:
The excerpt above encodes the three Babel synsets for the concepts of balloon (in the sense of aircraft), rst ying machine and Montgol er brothers. The <pos> tag provides the part of speech tag of the synset, the <source> tag describes the source from which the synset was obtained (WN for WordNet, WIKI for Wikipedia, WIKIWN for the intersection between the two resources), <babelSynsetId> provides the numeric id of the synset, and <mainSense> provides the main sense (either from WordNet or Wikipedia) which univocally identi es the Babel synset.
The rst Babel synset listed above, i.e., the concept of balloon (bn:00008187n), is semantically related to the Montgol er brothers (bn:02955250n), among others, as encoded by the semanticallyRelated relation, and is a lighter-than-air craft (bn:00051149n), as encoded by the hypernym relation. 4.2
An excerpt of the RDF Babel sense representation follows: <bn10schema:pos>NOUN</bn10schema:pos> <bn10schema:lemma>Ballongas</bn10schema:lemma> </bn10schema:BabelSense> <bn10schema:BabelSense rdf:about="http://lcl.uniroma1.it/babelnet/bn10/instance/
Ballon-DE@bn:00008187n"> <bn10schema:babelSynsetId>bn:00008187n</bn10schema:babelSynsetId> <bn10schema:lang>DE</bn10schema:lang> <bn10schema:source>WIKI</bn10schema:source> <bn10schema:pos>NOUN</bn10schema:pos> <bn10schema:lemma>Ballon</bn10schema:lemma> </bn10schema:BabelSense> <bn10schema:BabelSense rdf:about="http://lcl.uniroma1.it/babelnet/bn10/instance/
ballon-FR@bn:00008187n"> <bn10schema:babelSynsetId>bn:00008187n</bn10schema:babelSynsetId> <bn10schema:source>WNTR</bn10schema:source> <bn10schema:lang>FR</bn10schema:lang> <bn10schema:source>WIKITR</bn10schema:source> <bn10schema:pos>NOUN</bn10schema:pos> <bn10schema:lemma>ballon</bn10schema:lemma> </bn10schema:BabelSense> <bn10schema:BabelSense rdf:about="http://lcl.uniroma1.it/babelnet/bn10/instance/
Pallone_aerostatico-IT@bn:00008187n"> <bn10schema:babelSynsetId>bn:00008187n</bn10schema:babelSynsetId> <bn10schema:lang>IT</bn10schema:lang> <bn10schema:source>WIKI</bn10schema:source> <bn10schema:pos>NOUN</bn10schema:pos> <bn10schema:lemma>pallone aerostatico</bn10schema:lemma> </bn10schema:BabelSense> ... </rdf:RDF> where <lang> represents the language in which the sense is lexicalized, <source> is the source from which the sense is obtained (WN for WordNet, WNTR or WIKITR for translations of WordNet- or Wikipedia-annotated text, WIKIRED for a Wikipedia redirection, etc.), <pos> is the part of speech tag of the sense, and <lemma> speci es the lexicalization for the sense. 5
The Web of Data is in need for multilingual lexicalizations for Linked Open
Data. This vision of a Linguistic Linked Open Data (LLOD) has recently been
promoted, among others, by the Open Linguistic Working Group as well as other
researchers [
uniroma1.it/multijedi. Computing Laboratory5 { ts this vision by providing multilingual lexicalizations in RDF for millions of concepts, called Babel synsets, as well as a huge network of semantic relations between them. BabelNet currently covers 6 languages, but is continuously expanded with new information and languages.
Future steps include, among others, the integration of a mapping between BabelNet and other linguistic resources which are already part of the LLOD, such as DBPedia.
The author gratefully acknowledges the support of the ERC Starting Grant MultiJEDI No. 259234. The author wishes to thank Giovanni Stilo for his help with the RDF encoding of BabelNet.