=Paper=
{{Paper
|id=None
|storemode=property
|title=Accessing Multilingual Data on the Web for the Semantic Annotation of Cultural Heritage Texts
|pdfUrl=https://ceur-ws.org/Vol-775/paper9.pdf
|volume=Vol-775
|dblpUrl=https://dblp.org/rec/conf/semweb/MoerthDLV11
}}
==Accessing Multilingual Data on the Web for the Semantic Annotation of Cultural Heritage Texts==
https://ceur-ws.org/Vol-775/paper9.pdf
Accessing Multilingual Data on the Web for the
Semantic Annotation of Cultural Heritage Texts
1, 3
Karlheinz Moerth , Thierry Declerck1,2, Piroska Lendvai3, Tamás Váradi
1
ICLTT, Austrian Academy of Sciences, Sonnenfelsgasse 19/8,
1010 Wien, Austria
2
DFKI GmbH, Stuhlsatzenhausweg, 3
66123 Saarbrücken, Germany
3
HASRIL, Hungarian Academy of Sciences, Benczúr u. 33.
H-1068 Budapest, Hungary
Karlheinz.moerth@oeaw.ac.at,
declerck@dfki.de,
piroska@nytud.hu,varadi@nytud.hu
Abstract. Our study targets interoperable semantic annotation of Cultural
Heritage or eHumanities texts in German and Hungarian. A semantic resource
we focus on is the Thompson Motif-index of folk-literature (TMI), the labels of
which are available only in English. We investigate the use lexical data on the
Web in German and Hungarian for supporting semi-automatic translation of
TMI: lexical resources offered by Wiktionary accessed via the Lexvo service,
and discuss shortcomings of those resources. An approach for mapping the
XML dump of Wiktionary onto a TEI and MAF compliant data is presented,
whereby we discuss improvements in the representation of Wiktionary data for
exploiting its multilingual value within the LOD framework.
Keywords: Multilinguality, LOD, Cultural Heritage, Semantic Annotation
1 Introduction
In the context of a cooperation between the Austrian and the Hungarian Academies of
Sciences we investigate the possibility to generate interoperable and multilingual
semantic annotation of Cultural Heritage or eHumanities texts. One of the semantic
resources we consider for this task is the Thompson Motif-index of folk-literature
(TMI)1 [5], which contains around 36,000 terms, cataloguing typical narrative content
of folk tales and myths from around the world. The terms, or ‘labels’, of the
classification system are available only in English.
Our general hypothesis is that converting resources such as TMI into a LOD
compliant combination of multi-layered linguistic annotation and their taxonomic
classes can support the automatic detection and semantic annotation of motifs in
literary work, across genres and languages.
1
An electronic version of TMI is available at: http://www.ruthenia.ru/folklore/thompson/
80
A motif is an element conveying an idea or theme e.g. in film or music, but also in
folklore or scientific texts2. Motifs are cognitively complex notions expressed in
lexically and syntactically highly variable but compact structures. Linguistic features
of motifs have so far not been systematically investigated, but these have been
exposed and aim to be worked out by the authors of this paper, in collaboration with
the international AMICUS network3, with a clear motivation for enhanced indexing
and modelling of cultural heritage data (cf. [1], [3] and [4]).
The TMI catalog focuses on motifs that emphasize ideas or themes. For example,
“K3. Substitute in contest” is one motif in TMI (its parent node being “K0-K99.
Contests won by deception”, subsumed under “K. Deceptions”). Dozens of subtypes
are assigned to this single motif; these catalogue descriptions, or labels, are short
phrases such as “Supernatural substitute in tournament for pious warrior”, “Wise
man disguised as monk beats learned heretic in debate”. The TMI lists 23 main
categories4 and provides a deep hierarchical structure of motifs.
To semantically annotate texts in German and Hungarian with this resource, we
aim to enrich TMI with German and Hungarian labels. Our strategy consists in
providing first for the linguistic annotation of the phrasal heads detected in the
English labels5, and to try to find equivalent lexical entries in German and Hungarian
retrieved from online multilingual lexical resources.
2 Access to Online Lexical Resources in the LOD
The scarcity of freely available professional on-line multilingual lexical data made us
turn to the lexical resources offered by the collaborative dictionary project
Wiktionary, and the access provided to within the Lexvo service6, which has been
deployed within the Linked (Open) Data (LOD) framework 7. Some observations we
could make on this combination of resources are described in this section.
We noted first that in Wiktionary, variants of an entry (e.g. singular or plural
form), often do not feature identical sense or translation information. 8 It is necessary
to link those entries into a consistent unit, and to use an appropriate model for this.
Two candidates can be considered for this modeling: ISO-LMF 9 and lemon
2
Some random examples for motifs in folk tales are e.g. the cruel stepmother, the poor girl
who was chosen as wife in preference to a rich one, or a supernatural who substitutes the
hero in a tournament.
3
http://amicus.uvt.nl
4
E.g. Animal Motifs, Magic, the Dead, Marvels, Tests, the Wise and the Foolish, Deceptions,
Reversals of Fortune
5 The details of this linguistic analysis are described in a submission currently under review.
6
http://www.lexvo.org/
7 http://linkeddata.org/
8
One example is the English Wiktionary entry "creator"
(http://en.wiktionary.org/wiki/creator), which lists the basic morpho-syntactic information,
associated senses and translations whereas the entry "creators”
(http://en.wiktionary.org/wiki/creators#English) only states that it is the plural of "creator".
9
http://en.wikipedia.org/wiki/Lexical_Markup_Framework
81
(developed in the Monnet project and related to the W3C community) 10. An
advantage of the lemon approach would be that one could represent the Wiktionary
data in the RDF format, making Wiktionary data available in the Linked Data
framework. Nevertheless, as a first step we ported the XML dump of Wiktionary into
a TEI11 and MAF12 compliant format (see Section 3).
Lexvo is a service that "brings information about languages, words, characters, and
other human language-related entities to the Linked Data Web and Semantic Web" 13.
Lexvo points to Wiktionary entries, displaying for each word that can be queried (in a
variety of languages) a link to senses that are encoded either in the LOD version of
WordNet14 or/and of OpenCyc15, but in those versions the senses are available only
for English entries. Since the the Wiktionary data is not yet available in a machine-
readable format, Lexvo cannot display the senses available in the resource. This is an
additional argument for porting Wiktionary to RDF. Due to the same reason,
linguistic information associated to each word in WIktionary cannot be made
available in Lexvo. A Lexvo specific shortcoming is the fact that it refers only to the
English version of Wiktionary, regardless of entries that are in fact written in other
languages, ignoring as a consequence several pieces of language-specific information.
3 Porting Wiktionary to a Standardised Representation
Our starting point is the XML dump16 of Wiktionary. Nevertheless, the data do not
really deliver what one might expect from xml data, namely well-formed structured
information. The content is formatted making use of a lightweight markup system
which is used in different Wiki applications, and is neither standardized (various
applications use considerably divergent forms of the wikitext language) nor truly
structure-oriented. It is designed in a format-oriented manner to be transformed into
HTML.
Our initial goal was to transfer these data into an XML format suitable for further
processing. Although, as mentioned above, we consider ISO-LMF and lemon as the
final candidates, for pragmatic reasons, we eventually opted for TEI p5 17 as our
10
http://greententacle.techfak.uni-bielefeld.de/drupal/sites/default/files/lemon-cookbook.pdf
and [8].
11 http://www.tei-c.org/index.xml
12
http://lirics.loria.fr/doc_pub/maf.pdf
13 http://www.lexvo.org
14
http://semanticweb.cs.vu.nl/lod/wn30
15 http://sw.opencyc.org
16
http://dumps.wikimedia.org
17
As the TEI p5 dictionary module was conceptualized as the digital representation of printed
dictionaries, it appears not to be the most natural candidate for the task at hand. However, the
main motive behind adopting the dictionary module of this “de facto” text encoding standard
was that ongoing lexicographic projects of the ICLTT had yielded tools to process this kind of
data. Besides an online dictionary editor geared towards the particular needs of TEI, there are
also a number of thoroughly tested XSLT stylesheets to visualize the particular kind of data. A
second reason, equally important, is the fact that the ICLTT’s dictionary working group has
been working recently on a TEI dictionary schema suitable for use in NLP applications.
82
starting point.. While several attempts at preparing Wiktionary for use in NLP
applications have been made before [2, 5, 7], the tool we present here is – to our
knowledge – the first such application targeting TEI p5, and the first such tool
provided with a graphical user interface.
The actual conversion process is carried out in three main steps. Each of these
steps can be performed separately, which allows the interested user to pursue the
transformation process in detail.
First, the comparatively large database dump (287 MB) was split into manageable
smaller chunks. This process resulted in a collection of roughly 85000 entries.
In the second phase of the conversion, the top-level constituents of these entries
were identified and transformed into XML elements. This task turned out to be pretty
straightforward as the entries (we stick to traditional lexicographic nomenclature
here) display a rather flat hierarchical structure. The resulting chunks each contain a
particular type of data, the main constituents of the dictionary entries. The number of
constituent parts varies with the size of the individual entries (from 3KB up to
338KB). In the result sets, there are chunks containing grammatical data such as for
instance part of speech. There are chunks containing etymological information and/or
usage information. Many entries contain morphological data, in numerous cases
complete inflectional paradigms. The files also hold data concerning hyphenations of
word forms and their pronunciation. However, the central concern of our work here
has been semantic data. This kind of information is stored in sections describing the
various meanings of words. These, in turn, are linked to translations, synonyms,
antonyms, hyperonyms, hyponyms, and often to examples.
The last step in the transformation process has been the conversion of the above
described constituents into TEI p5. Iterating through all the untyped chunks, the
program attempts to identify the right category and subsequently to translate it into
TEI p5. At this point, the main challenge for the programmer was the merging of data
on the same hierarchical level (e.g. meanings and translations) into neatly nested
XML structures. Successful data conversion depends largely on the quality of the
underlying markup. While many errors can be compensated by some trickery in the
program, inconsistencies remain.
The actual tag set applied in our project can also be seen as a contribution aiming
at developing the TEI guidelines towards an encoding system suitable to be used in
NLP applications. We will not go into the gory details of modeling TEI documents
18
here, just one small digression: one particularly useful module of the TEI p5
guidelines was the chapter on feature structures. This mechanism allowed us to model
the representation of the morpho-syntactic data in accordance with the MAF standard
(Morpho-syntactic Annotation Framework, ISO TC 37). Canonical TEI for inflected
word forms such as gingst “(you) went” usually look like this:
18
An initiative towards this end was the workshop Tightening the representation of lexical
data, a TEI perspective at the TEI’s members meeting this year in Würzburg (Germany).
83
We tried to encode such structures in a more MAF-like manner, which is still TEI
conformant:
In this encoding scheme, the morpho-syntactic identifiers used in the ana atribute of
the form element is defined as a set of TEI conformant feature structures. The values
used here refer to a feature value library, which is also linked to the ISO data
categories.
Although the conversion tool already works quite nicely, a number of issues
registered in its requirement specification remain to be solved. It goes without saying
that the first thing that comes to mind, is the issue of other languages, which is on top
of our agenda. First candidates for this are English and French.
The second issue is moving on to LMF which is a project reaching far beyond our
Wiktionary tool. Creating LMF data from TEI is something apparently non-trivial.
One other important task to be achieved in the near future is setting up a service
delivering the data. First steps towards implementing a restful server have been taken.
We hope that by the time this paper is presented, our TEI version of the German-
language Wiktionary will be up and running.
4 Further work on porting Wiktionary to the Semantic Web
Although our work represents a step in making the full Wiktionary information
available for NLP applications, it is not sufficient to represent links between entries
(for example, one entry being the plural of the other, etc), or to make this information
available in the Web or in the LOD and so to establish links between entries and
senses in Wiktionary, WordNet or OpenCyc, on the one, but also between TMI and
LOD data sets on the other hand. Just to name an example: In TMI the concept "A0:
Creator" is the upper class of a large number of (hierarchically ordered) terms. We
collected all the head nouns of those terms, and can build so a kind of domain specific
"WordNet". This list of nouns is for sure very different and more complex than what
we find in WordNet or OpenCyc. We need a way to relate the semantic organization
84
of TMI and WordNet/OpenCyc (or other data sets), also on the base of linguistic
information we can find in the (Semantic) Web. There is therefore a need to port both
Wiktionary and the analyzed labels of TMI onto a LOD compliant RDF. For this we
are getting also advices from the Monnet project19.
Acknowledgments
Part of the described in this paper is supported by the R&D project “Monnet”, which
is co-funded by the European Union under Grant No. 248458, and by the AMICUS
network, which is sponsored by a grant from the Netherlands Organization for
Scientific Research, NWO Humanities, as part of the Internationalization in the
Humanities programme
References
1. Declerck, T., K. Eckart, Lendvai, P., L. Romary, T. Zastrow (2010a). Towards a
Standardised Linguistic Annotation of Fairy Tales. In: Proc. of the LRT standards
workshop at LREC-2010.
2. Krizhanovsky, A. (2010). The comparison of Wiktionary thesauri transformed into the
machine-readable format. (http://arxiv.org/abs/1006.5040)
3. Lendvai, P., Declerck, T., S. Darányi, P. Gervás, R. Hervás, S. Malec, F. Peinado (2010a).
Integration of Linguistic Markup into Semantic Models of Folk Narratives: The Fairy Tale
Use Case. In: Proceedings of the Seventh International conference on Language Resources
and Evaluation, Pages 1996-2001, Valetta, Malta, European Language Resources
Association (ELRA).
4. Lendvai, P. (2010). Granularity Perspectives on Modeling Humanities Concepts. In: S.
Darányi, P. Lendvai, (eds.). First International AMICUS Workshop on Automated Motif
Discovery in Cultural Heritage and Scientific Communication Texts, Vienna, Austria.
University of Szeged, Hungary.
5. Navarro, E., Sajous, F., Gaume, B., Prévot, L., Hsieh, S.-K., Kuo, T.-Y., Magistry, P.,
Huang, C.-R. (2009). Wiktionary and NLP: Improving synonymy networks. In:
Proceedings of the 2009 Workshop on Peoples’s Web Meets NLP, ACL-IJCNLP.
Singapore: pp. 19-27.
6. Thompson, S. (1955-58). Motif-index of folk-literature: A classification of narrative
elements in folktales, ballads, myths, fables, medieval romances, exempla, fabliaux, jest-
books, and local legends. Revised and enlarged edition. Bloomington, Indiana University
Press.
7. Zesch T., Mueller C., Gurevych I. (2008). Extracting lexical semantic knowledge from
Wikipedia and Wiktionary. In: Proceedings of the Conference on Language Resources and
Evaluation. LREC 2008.
8. McCrae, J, Aguado-de-Cea G, Buitelaar P, Cimiano P, Declerck T, Gomez-Perez A,
Gracia J, Hollink L, Montiel-Ponsoda E, Spohr D et al.. In Press. Interchanging lexical
resources on the Semantic Web. Language Resources and Evaluation 2011.
19
http://www.monnet-project.eu
85