=Paper=
{{Paper
|id=Vol-551/paper-20
|storemode=property
|title=Cross-lingual Dutch to English alignment using EuroWordNet and Dutch Wikipedia
|pdfUrl=https://ceur-ws.org/Vol-551/oaei09_paper13.pdf
|volume=Vol-551
|dblpUrl=https://dblp.org/rec/conf/semweb/Bouma08
}}
==Cross-lingual Dutch to English alignment using EuroWordNet and Dutch Wikipedia==
https://ceur-ws.org/Vol-551/oaei09_paper13.pdf
Cross-lingual Dutch to English alignment using
EuroWordNet and Dutch Wikipedia
Gosse Bouma
Information Science, University of Groningen, g.bouma@rug.nl
Abstract. This paper describes a system for linking the thesaurus of the Nether-
lands Institute for Sound and Vision to English WordNet and dbpedia. We used
EuroWordNet, a multilingual wordnet, and Dutch Wikipedia as intermediaries
for the two alignments. EuroWordNet covers most of the subject terms in the
thesaurus, but the organization of the cross-lingual links makes selection of the
most appropriate English target term almost impossible. Using page titles, redi-
rects, disambiguation pages, and anchor text harvested from Dutch Wikipedia
gives reasonable performance on subject terms and geographical terms. Many
person and organization names in the thesaurus could not be located in (Dutch or
English) Wikipedia.
1 Presentation of the system
This paper describes our system for the very large cross-lingual resources (vlcr) task,
which asked for an alignment between the thesaurus of the Netherlands Institute for
Sound and Vision and English WordNet and (English) dbpedia, a database extracted
from Wikipedia.
We used an ad-hoc system to achieve the alignment. For the mapping to English
WordNet, we used EuroWordNet, a multilingual resource which contains a Dutch word-
net, as well as mappings from Dutch to English WordNet. For the mapping to dbpedia,
we used page titles, redirects, and anchor texts harvested from Dutch Wikipedia, and
mapped Dutch pages to English pages using cross-language links. Most XML prepro-
cessing was done using XQuery. The alignment itself was done using (Sicstus) Prolog.
1.1 Background
For our work on open domain question answering, information extraction, and corefer-
ence resolution, we are interested in creating general, informal, taxonomies of entities
encountered in Dutch texts.1 As part of this work, we created a Dutch counterpart of
the Yago system [4], in which Wikipedia categories are aligned with a Dutch wordnet
[1]. We expected that the techniques we used there (especially stemming and parsing of
labels, and using predominant word senses for sense disambiguation) could be applied
to the present task as well.
1
Some results can be found on www.let.rug.nl/gosse/Ontology
1.2 Aligning GTAA to WordNet via EuroWordNet
The mapping from the thesaurus of the Netherlands Institute for Sound and Vision
(GTAA) and English Wordnet was accomplished using EuroWordNet [6]. We concen-
trated on the subset of the thesaurus that contained subject labels, as these are mostly
common nouns or noun phrases headed by a common noun. The Dutch part of Eu-
roWordNet (EWN) contains hardly any proper names, so we expected the overlap be-
tween EWN and the other parts of the thesaurus (on person names, geographical loca-
tions, and organizations) to be minimal.
The alignment procedure is schematically represented in figure 1.
EWN (near)syn WNMAP
subject label −→ EWN id −→ WN1.5 id −→ WN2.0 id
Fig. 1. Mapping GTAA to WordNet
Entries in the thesaurus are often plurals (afgevaardigden (representatives), spoor-
wegen (rail roads), autobussen (buses)), whereas dictionary entries in EWN are typi-
cally singular. To ensure coverage of these cases, all entries in the subject part of the
thesaurus were stemmed using the Alpino parser [5]. Alpino is a wide-coverage depen-
dency parser for Dutch, which includes a morphological analyzer. As the analyzer also
performs compound analysis (ie. autobussen is analyzed as auto bus), we also parsed
all EWN entries with Alpino. Thus, we can find a subject label in EWN by compar-
ing stems. Note that compound analysis would also allow us to link a compound such
as bedrijfspionage (industrial espionage) to a more general concept such as espionage
(assuming a hypernym relation), but such links were not requested in the task definition.
EuroWordNet is a multilingual wordnet, in which each synset is linked to one or
more inter language index ids (ILIs). ILIs in turn are linked to WordNet 1.5 ids. Links
can express among others a synonym, near-synonym, hyponym or hypernym relation.
We used only the synonym and near-synonym relations. Using the ILIs, each Dutch
synset can be linked to an English WordNet id. As we will explain below, this step is in
general one-to-many, as most Dutch synsets are connected to more than one ILI through
the near-synonym relation. In the final step, we mapped WordNet 1.5 ids to WordNet 2.0
ids (the version of WordNet that was used to create the RDF/OWL version of WordNet
that was the target of the mapping), using the WordNet mappings described in [2].2
1.3 Aligning GTAA to dbpedia via Dutch Wikipedia
For linking GTAA entries to dbpedia, we decided to use Dutch Wikipedia as interme-
diary, and to aim for linking GTAA entries to English Wikipedia pages. Translation of
English Wikipedia pages into dbpedia URI’s is done by means of a small script that adds
the correct prefix, and deals with special characters.
2
available from from www.lsi.upc.es/∼nlp/tools/mapping.html
For our work on automatic annotation of web pages with links to Wikipedia [3],
we had harvested a Dutch Wikipedia dump (august 2008) for cross-language links,
redirects, disambiugation pages, and anchor texts (i.e. terms annoted on a Wikipedia
page with a link to another Wikipedia page). We also used a list of English page names
from a dump of English Wikipedia (january 2008).
The first step in the alignment is to generate all variants of a label. To match a term
in the thesaurus with a Wikipedia page directly, for instance, it is necessary that the
first letter is in upper case. Person names in GTAA are given as Lastname, Firstname,
whereas Wikipedia simply uses Firstname Lastname. Subjects in GTAA are often plu-
ral, whereas they tend to be singular in Wikipedia. Singular forms are obtained from the
parsed version of the subject labels that was also used in the aligment with WordNet.
Finally, alternative labels provided by GTAA are considered as variants of the concept
label.
For all variants of a GTAA concept label, we try to find a matching Dutch Wikipedia
page. This can be achieved by an exact match with a Wikipedia page, by an exact match
with a redirect page (in which case the target of the redirect is the desired Wikipedia
page), by finding a matching anchor text (in which case the most frequent target page
for that anchor is returned) or by an exact match with a disambiguation page (in which
case all options are returned). Given a suitable Dutch page, we find the English page
by following the cross-language link from Dutch to English Wikipedia. In some cases
such a link was absent. If a Dutch page (with a corresponding English page) could not
be found by means of the techniques above, we tried to find a matching page in English
Wikipedia directly, using only page titles.
We expect that there will be a difference in accuracy between the various methods
for finding an English page. Preference (and a high confidence score) is given to direct
matches, followed by redirects, anchors, direct matches in English, and disambiguation
pages.
1.4 Scripts and results
The scripts used to produce the alignment can be found at www.let.rug.nl/gosse/
GTAA. Note that EuroWordNet data is missing, as this is a resource which is not in the
public domain.
The results of our alignment can be found at www.let.rug.nl/gosse/GTAA/
bouma-vlcr.tgz.
2 Results
2.1 vlcr: GTAA to WordNet
We only tried to link GTAA subject entries to WordNet. An overview of the results
is given in table 1. Note that coverage is quite reasonable between GTAA and EWN.
Where no link could be found, this is mostly due to multiword subject labels (such as al-
ternatieve energie (alternative energy) or bedreigde diersoorten (endangered species))
and compounds. Multiword phrases are generally absent from EWN, and we made no
attempt to search for these in English WordNet directly. Other subjects that could not
be linked often consist of a compound noun. As compounding is a productive process,
we do not expect all compounds to be present in EWN. Given the fact that we do have a
morphological analysis, we could have linked compound nouns to a more general con-
cept (i.e. the head noun) by means of a hypernym link. Such links were not part of the
task, however. Together, multiword phrases and compounds account for over 80% of
the subject labels that could not be linked. 5% coverage was lost in the mapping from
WordNet 1.5 to WordNet 2.0.
subject labels 3878
linked to EWN 2617 (67%)
unique ILIs 3703
avg. ambiguity 1.4
linked to WN2.0 2392 (62%)
unique synsets 3676
avg. ambiguity 1.5
Table 1. Alignment results for GTAA to EuroWordNet and WordNet 2.0
Ambiguity of the target is a serious problem. This is not only caused by the fact that
a word may belong to more than one synset (word sense ambiguity), but also by the fact
that the mapping between synsets in EWN and WN through ILI links is highly ambigu-
ous. The Dutch nouns part of EWN contains only 631 synonym relation ILIs (which
tend to be unique), and no less than 4641 near-synonym relation ILIs (which tend to
link to several WN targets). One might consider reducing the ambiguity by selecting
the most appropriate word sense for a given subject label. This is by no means trivial
however (see [1] for some results for Dutch). In this particular case, it is also not very
effective, as many synsets are themselves connected to more than one English synset
through the near-synonym relation. The situation is illustrated in figure 2. The concept
brons is linked to two synsets in EWN. As WN has two synsets for the bronze as well,
one might expect each of these synsets to be linked to a specific WN synset. In reality,
however, each EWN synset is linked to each WN synset. Thus, even if one resolved
the concept brons to the correct EWN synset, it still would be practically impossible to
decide which of the two WN synsets ought to be chosen (as the information on how to
disambiguate synsets between wordnets is simply not given). In our results, both tar-
gets are given as possible alignment, but lower confidence is given to links involving a
near-synonym relation.
2.2 vlcr: GTAA to dbpedia
Table 2 gives some results for linking the four different parts of the GTAA thesaurus
(subject/concepts, names/organisations, locations, and persons) to English Wikipedia.
Coverage is best for subjects and locations. GTAA contains many names of persons and
concept EWN synset ILI WN synset
% 10527 −→ 03038788 −→ bronze-noun-1
brons &
%
& 38608 −→ 08841702 −→ bronze-noun-2
Fig. 2. Linking the concept brons to two EWN synsets, and two WN synsets.
organisations that seem to be absent in both Dutch and English Wikipedia. It should also
be noted that coverage of location names is high only because many location names are
found in English Wikipedia directly. This holds partly for names of organisations as
well, but less so for person names. For 6 - 9% of the concepts, a Dutch Wikipedia target
could be found, but no corresponding English page existed.
subject name location person
link type links % links % links % links
nlpage 2027 52.3 3128 11.5 5135 36.7 7311 7.5
redirect 423 10.9 984 3.6 400 2.9 762 0.8
anchor 621 16.0 616 2.3 357 2.6 176 0.2
enpage 260 6.7 4085 15.1 3705 26.5 9246 9.5
linked 3127 80.6 8830 32.6 9602 68.6 17521 17.9
no-english 357 9.2 2197 8.1 878 6.3 5721 5.9
no-link 394 10.2 16077 59.3 3512 25.1 74375 76.2
total 3878 27104 13992 97617
Table 2. Alignment results for GTAA to Dutch and English Wikipedia
3 Discussion
In general, it seems that even with relatively modest technology, a mapping between
two resources in different languages can be achieved. It should be noted, however, that
the mapping to WordNet owes much to the existence of EuroWordNet, which solves
the most difficult (cross-language) part of the task to a large extent. On the other hand,
EuroWordNet does not help much in deciding which synset for a given English term is
the appropriate one.
Our results for Wikipedia linking could still be improved in a number of ways.
We hardly employed categorical constraints. The GTAA thesaurus comes in four parts.
Each part is a different category. This information could be used to block the link from
A4 in the locations file to A4 (paper format) in Wikipedia. Similarly, concept labels
often come with a skope note. Word overlap could be used to select the correct target
page (i.e. to prefer highway A4 in the Netherlands over that in Austria)). Alternatively,
one might use the information that concepts with the same scope note are likely to be
linked to Wikipedia pages with identical or closely related Wikipedia categories to de-
tect outliers. For selecting the most promising target, we experimented with a simple
preference scheme (which always prefers the link given by the most reliable relation),
and a simple weighting scheme (which adds scores when multiple links to the same tar-
get are found). Weighting was used for the final results. No doubt, more subtle schemes
could be developed. For instance, at the moment we only take into account the most
frequent target of an anchor text. Alternatively, one might consider all targets pointed
to by anchor text as potential targets, and use the frequency of these links as a weight.
Somewhat surprisingly, we discovered that cross-language links are not reversible.
Initially, we used cross-language links harvested from English Wikipedia, as this is the
larger resource, and we expected that this might also be more thorough in providing
cross-language links. However, since English Wikipedia has more pages than Dutch
Wikipedia, several English pages may be linked to the same Dutch page (i.e. Bowl-
ing and Ten pin Bowling both point to the Dutch page Bowling). If one works with
cross-language links harvested from Dutch Wikipedia, this situation does occur less
frequently, although similar problems can occur here as well (i.e. in the versions of
Wikipedia we used, the Dutch A4 highway was linked to an English page which redi-
rected to a general page on Dutch highways).
4 Conclusion
We have presented a method for linking the thesaurus of the Netherlands Institute for
Sound and Vision with two English resources, WordNet and Wikipedia. We used an
ad-hoc method which relied on the existence of cross-language links for similar data,
namely EuroWordNet, a multi-lingual wordnet with cross-language links, and Dutch
Wikipedia, which contains cross-language links to English Wikipedia.
References
1. Gosse Bouma. Linking Dutch Wikipedia Categories to EuroWordNet. In Proceedings of
the 19th Computational Linguistics in the Netherlands meeting (CLIN 19). Groningen, the
Netherlands, 2009.
2. J. Daude, L. Padro, and G. Rigau. Mapping wordnets using structural information. In Pro-
ceedings of the 38th Annual Meeting on Association for Computational Linguistics, pages
504–511. Association for Computational Linguistics Morristown, NJ, USA, 2000.
3. Proscovia Olango, Gerwin Kramer, and Gosse Bouma. Termpedia for interactive docu-
ment enrichment. In Computational Linguistics Applications (CLA) workshop at IMCSIT,
Mragowo, Poland, 2009.
4. Fabian M. Suchanek, Gjergji Kasneci, and Gerhard Weikum. Yago: a core of semantic knowl-
edge. In WWW ’07: Proceedings of the 16th international conference on World Wide Web,
pages 697–706, New York, NY, USA, 2007. ACM Press.
5. Gertjan van Noord. At last parsing is now operational. In Piet Mertens, Cedrick Fairon, Anne
Dister, and Patrick Watrin, editors, TALN06. Verbum Ex Machina. Actes de la 13e conference
sur le traitement automatique des langues naturelles, pages 20–42. 2006.
6. P. Vossen. Eurowordnet a multilingual database with lexical semantic networks, 1998.