In the rst part of this paper we report about experiences when applying the DBpedia extraction framework to the Korean Wikipedia. We improved the extraction of non-Latin characters and extended the framework with pluggable internationalization components in order to facilitate the extraction of localized information. With these improvements we almost doubled the amount of extracted triples. We also will present the results of the extraction for Korean. In the second part, we present a conceptual study aimed at understanding the impact of international resource synchronization in DBpedia. In the absence of any information synchronization, each country would construct its own datasets and manage it from its users. Moreover the cooperation across the various countries is adversely a ected.
Wikipedia is the largest encyclopedia of mankind and is written collaboratively
by people all around the world. Everybody can access this knowledge as well as
add and edit articles. Right now Wikipedia is available in 260 languages and the
quality of the articles reached a high level [
Therefore, data which could be obtained from other articles is not included and hence cannot be queried. Another problem is the support for non-Latin characters which among other things results in problems during the extraction process. Wikipedia language editions with a relatively small number of articles (compared to the English version) could bene t from an automatic translation and complementation based on DBpedia. The Korean Wikipedia, for example, was founded in October 2002 and reached ten thousand articles in June 20054 . Since February 2010, it has over 130,000 articles and is the 21st largest Wikipedia5. Despite of this growth, compared to the English version with 3.2 million articles it is still small.
The goal of this paper is two-fold: (1) to improve the DBpedia extraction from non-Latin language editions and (2) to automatically translate information from the English DBpedia in order to complement the Korean Wikipedia.
The rst aim is to improve the quality of the extraction in particular for the
Korean language and to make it easier for other users to add support for their
native languages. For this reason the DBpedia framework will be extended with a
plug-in system. To query the Korean DBpedia dataset a Virtuoso Server[
The paper is structured as follows: In Section 2, we give an overview about the work on the Korean DBpedia. Section 3 explains the complementation of Korean Wikipedia using DBpedia. In Section 4, we review related work. Finally, we discuss future work and conclude in Section 5. 2
The Korean DBpedia uses the same framework to extract datasets from Wikipedia as the English version. However, the framework does not have su cient support for non-English languages, especially for the non-Latin alphabet based languages.
For testing and development purposes, a dump of the Korean Wikipedia was loaded into a local MySQL database. The rst step was to use the current DBpedia extraction framework in order to obtain RDF triples from the database.
At the beginning the focus was on infoboxes, because infobox templates o er
already semi-structured information. But instead of just extracting articles that
have a corresponding article in the English Wikipedia, like the datasets
provided by DBpedia, all articles have been processed. More information about the
DBpedia framework and the extraction process can be found in [
After the extraction process and the evaluation of the RDF triples,
encoding problems have been discovered and xed. In fact most of these problems
will occur not only in Korean, but for all languages with non-Latin
characters. Wikipedia and DBpedia use the UTF-8 and URL encoding. URI's in
DBpedia have the form http://ko.dbpedia.org/resource/Name, where Name
is taken from the URL of the source Wikipedia article, which has the form
http://ko.wikipedia.org/wiki/Name. This approach has certain advantages.
Further information can be found in [
This property URI contains the \%" character and thus cannot be serialized as RDF/XML. For this reason another way has to be found to represent properties with \%" encoding. The solution in use by DBpedia is to replace \%" with \ percent ". This resulted in very long and confusing properties which also produced errors during the extraction process. This has not been a big issue for the English DBpedia, since it contains very few of those characters. For other languages this solution is unsuitable. To solve it, di erent solutions have been discussed. The rst possibility is to just drop the triples that contain such characters. Of course this is not an applicable solution for languages that mainly consist of characters which have to be encoded. The second solution was to use a shorter encoding but with this approach the Wikipedia encoding cannot be maintained. Another possibility is to use the \%" character and add an underscore at the end of the string. With this modi cation, the Wikipedia encoding could be maintained and the RDF/XML can be serialized. At the moment we use this solution during the extraction process. The use of IRI6's instead of URI's is another possibility which we will discuss in Section 5. An option has been added to the framework con guration to control which kind of encoding should be used.
Because languages di er in grammar as well, it is obvious that every language uses its own kind of formats and templates. For example, dates in the English and in the Korean Wikipedia look as follows:
For that reason every language has to de ne its own extraction methods. To realize this, a plug-in system has been added to the DBpedia extraction framework (see Fig. 1).
The plug-in system consists of two parts: the default part and an optional part. The default part contains extraction methods for the English Wikipedia and functions for datatype recognition, for example, currencies and measurements. This part will always be applied rst, independent from which language is actually extracted.
The second part is optional. It will be used automatically if the current language is not English. The extractor will load the plug-in le for the corresponding language if it exists. If the extractor did not nd a match in the default part, it will use the optional part to check the current string for corresponding templates. The same approach is used for sub-templates which are contained in the current template.
After these problems have been resolved and the plug-in system has been added to the DBpedia extraction framework, the dataset derived from the Korean Wikipedia infoboxes consists of more than 103,000 resource descriptions with more than 937,000 RDF triples in total. The old framework only extracted 55,105 resource descriptions with around 485,000 RDF triples. The amount of triples and templates was almost doubled. The extended framework also extracted templates which have not been extracted by the old framework at all. A comparison between some example templates extracted by the old framework and the extended version can be found in Fig. 2.
We already started to extend the support for other extractors. Until now the extractors mentioned in Table 1 are supported for Korean. 3
The infobox is manually created by authors that create or edit an article. As a result, many articles have no infoboxes and other articles contain infoboxes which are not complete. Moreover, even the interlanguage linked articles do not use the same infobox template or contain di erent amount of information. This is what we call the imbalance of information. This problem raises an important issue about multi-lingual access on the Web. In the Korean Wikipedia multi-lingual access is prevented due to the lack of interlanguage links.
For example (see Fig. 3), the \Blue House"7 article in English contains the infobox template Korean Name. However, the \Blue House" could be regarded as either a Building or a Structure. The \White House" article, which is very similar to the former, uses the Historic Building infobox template. Furthermore, the interlanguage linked \@ " (\Blue House" in Korean) page does not contain any infobox.
There are several types of imbalances of infobox information between two di erent languages. According to the presence of infobox, we classi ed interlanguages linked pairs of articles into three groups: Short Infobox (S), Distant Infobox (D), and Missing Infobox (M): 7 The \Blue House" is the executive o ce and o cial residence of the South Korean head of state, the President of the Republic of Korea. { The S-group contains pairs of articles which use the same infobox template but have a di erent amount of information. For example, an English-written article and a non-English-written article, which have an interlanguage link and use the same infobox template, but have a di erent amount of template attributes. { The D-group contains pairs of articles which use di erent infobox templates. D-group emerges due to the di erent degrees of each Wikipedia communities' activities. In communities where many editors lively participate, template categories and formats are more well-de ned and more ne-grained. For example, philosopher, politician, o ceholder and military person templates in English are matched just person template in Korean. It appears not only a Korean Wikipedia but also non-English Wikipedias. { The M-group contains pairs of articles where an infobox exists on only one side.
As a rst step, we concentrate on S-group and D-group. We tried to enrich the infobox using dictionary-based term translation. In this work, DBpedia's English triples in infoboxes are translated into Korean triples. We used bilingual dictionary which is originally created for English-to-Korean translations from Wikipedia interlanguage links. Then we added translation patterns for multiterms using bilingual alignment of collocations. Multi-terms are set of single terms such as Computer Science.
DBpedia[
We can get translated Korean triples over 30 times larger than existing Korean triples. However, a large amount of translated triples has no prede ned templates in Korean. There may be a need to form a template schema to organize the ne-grained template structure.
Thus we have built a template ontology, OntoCloud8, from DBpedia and Wikipedia, which was released on September 2009, to e ciently build the template structure. The construction of OntoClolud consists of the following steps: (1) extracting templates of DBpedia as concepts in an ontology, for example, the Template:Infobox Person (2) extracting attributes of these templates, for example, name of Person. These attributes are mapped to properties in ontology. (3) constructing the concept hierarchy by set inclusion of attributes, for example, Book is a subclass of Book series.
{ Book series = fname, title orig, translator, image, image caption, author, illustrator, cover artist, country, language, genre, publisher, media type, pub date, english pub date, preceded by, followed byg.
For the ontology building, similar types of templates are mapped to a concept. For example, the Template:infobox baseball player and Template:infobox asian baseball player describe baseball player. Moreover di erent format of properties with same meaning should be re ned, for example, `birth place', `birthplace and age' and `place birth' are mapped to `birthPlace'. OntoCloud v0.2 includes 1,927 classes, 74 object properties and 101 data properties.
We provided the rst implementation of the DBpedia/Wikipedia multi-lingual enrichment research. 4
DBpedia[
The rst project is Yago[
Another approach is Semantic MediaWiki [
The third project is Freebase, an online database of structured data. Users can edit this database in a similar way as Wikipedia can be edited.
In the area of cross-language data fusion, another project has been launched
[
The DBpedia ontology has been created manually based on the most
commonly used Infoboxes within Wikipedia. Kylin Ontology Generator[
The last project is CoreOnto. It is the research project about IT ontology
infrastructure and service technology development 9. There are several
components and solutions for semi-automated ontology construction. One of them is
the CAT2ISA[
5
Fig. 4. English-Korean Enrichment in DBpedia/Wikipedia
As future work, it is planned to support more extractors for the Korean language and improve the quality of the extracted datasets. The support for Yago and WordNet could be covered by DBpedia-OntoCloud. The OntoCloud has been linked to WordNet where OntoCloud is an ontology transformed from the templates of infobox (English). To make the dataset accessible for everybody a server will be set up at the SWRC10 at KAIST11.
Because encoding for Korean results in unreadable strings for human beings the idea has been raised to use IRI's instead of URI's. It is still uncertain if all tools of the tool chain can handle IRI's. Nevertheless it is already possible to extract the data with IRI's if desired. These triples contain characters that are not valid in XML. The number of triples with such characters is only 48 and can be ignored for now. We also plan to set up a Virtuoso Server to query the Korean DBpedia over SPARQL.
The results from the translated Infoboxes should be evaluated precisely and improved afterwards. After the veri cation of the triples, the Korean DBpedia 10 http://swrc.kaist.ac.kr 11 Korea Advanced Institute of Science and Technology: http://www.kaist.ac.kr can be updated. This will be helpful to guarantee that the same information can be recognized in di erent languages.
The concept shown in Fig. 4 describes the information enrichment process within the Wikipedia and DBpedia. As described earlier, we rst synchronized two di erent language versions of DBpedia using translation. After that we can create infoboxes using translated values. In the nal study of this project, we will generate new sentences for Wikipedia using newly added data from DBpedia. These sentences will be published in the appropriate Wikipedia articles. It can help authors to edit articles and to create infoboxes when a new article is created. The system can support the authors by suggesting the right template.