Tagging systems are nowadays a common feature in web sites where user-generated content plays an important role. However, the lack of semantics and multilinguality hamper information retrieval process based on folksonomies. In this paper we propose an approach to bring semantics to multilingual folksonomies. This approach includes a sense disambiguation activity and takes advantage from knowledge generated by the masses in the form of articles, redirection and disambiguation links, and translations in Wikipedia. We use DBpedia[2] as semantic resource to de ne the tag meanings.
The term folksonomy is normally used to refer to the classi cation schemes that emerge from the tagging activity of a user community. Hence folksonomies represent consensual knowledge, but they are still a ected by the lack of semantics. Tagging systems are not aware of: 1) possibly related tags due to relations such as synonyms, broaderthan, narrower-than, and spelling variation, or 2) the use of ambiguous tags.
Despite the fact that tagging systems are web applications with a world wide scope and thus reaching users with multiple languages, semantics of multilingual tags has not been researched. We propose a novel solution for the association of semantics to multilingual tags. Our contribution is twofold: 1) a multilingual sense repository, initially for English and Spanish languages, and 2) Sem4Tags a process for the association of semantics to multilingual tags.
The problem of identifying tag semantics in Folksonomies
has been addressed by researchers in two complementary
ways: 1) by identifying groups of related tags using
clustering techniques in the hope of such grouping expose the
meaning of the tags [
MULTILINGUAL SENSE REPOSITORY Inspired by the Tagora sense repository1 we designed MSR, a multilingual sense repository for English and Spanish based on Wikipedia and DBpedia information. MSR uses: 1) article URLs as sense identi ers, and article words along with their frequency as keywords associated with the sense, 2) articles listed in disambiguation pages as possible senses for ambiguous words, 3) the explicit translations among articles to link senses in languages di erent from English to English senses, and 4) DBpedia resources2 to de ne formally each sense. For each tag to be analyzed the population process is carry out:
Create disambiguation list: First, the list of candidate senses is created. We look for a disambiguation page related to the tag. If this page exists then we extract the possible meanings. Otherwise, we look for a content page related to the tag.
Extract sense information: Then, for each candidate sense we extract the keywords and their frequency from the corresponding article.
Get translations: In addition, for tags in languages di erent than English, we look for English translations in Wikipedia and using the LabelTranslator tool3.
Associate semantic entities: Finally, we extract from DBpedia the resources related to the candidate senses. English and Spanish Wikipedia articles are linked to DBpedia resources by means of the page4 and the wikipage-es5 relations. In case the wikipage-es relation does not exists for an Spanish article, we use the translation found in the previous activity and use the page relation. 4.
SEM4TAGS: A PROCESS FOR THE ASSOCIATION OF SEMANTICS TO MULTILINGUAL TAGS
We designed Sem4Tags, a process aiming at associating tags with semantic resources relying on MSR. The input is a tag, its context, and optionally the language of the tag. As context we use the set of user tags co-occurring when 1http://tagora.ecs.soton.ac.uk 2http://dbpedia.org/ 3http://neon-toolkit.org/wiki/LabelTranslator 4http://xmlns.com/foaf/0.1/page 5http://dbpedia.org/property/wikipage-es annotating a resource. The output is a DBpedia resource representing the intended meaning of the input tag. The Sem4Tags process includes the following activities:
Preprocessing: The tag is preprocessed to nd a normalized representation based on Wikipedia article titles. We bene t from Wikipedia redirection pages when the tag has been considered as an alternative to an article title. In addition, we modify morphologically the tag according to the article title notation. Finally, if after those modi cations we have not found a Wikipedia article, we use the Yahoo! spelling service6 to nd an alternative representation.
Active Context Selection: The context is ltered to
get rid of tags that can a ect the disambiguation activity.
The active context contains the set of most highly
semantically related tags to the input tag according to a web-based
relatedness measure[
Sense Retrieval: We select from MSR the set of candidate senses for the tag. We query MSR using the tag normalized version. If the tag is ambiguous the output of this activity is a set of senses. Otherwise, the output is a unique sense.
Disambiguation activity: This activity select the most
probable sense for a tag from a set of senses. The idea is
that the tag and its context can be compared against each
one of the senses measuring the overlapping of the terms in
the context with the terms in the Wikipedia pages related to
the senses. We use the vector space model to represent the
senses and the tag context [
To evaluate Sem4Tags we carried out an experiment using data extracted from Flickr. We gathered 759 photos tagged with tourist cities in Spain (e.g., Barcelona, Ibiza, etc.). On average those photos were annotated with 12.4 tags.
Our baseline attempts to associate directly tags with DBpedia resources. For doing this we create an URI of the form http://en.wikipedia.org/wiki/tag for English tags and of the form http://es.wikipedia.org/wiki/tag for Spanish tags. Then we query DBpedia for the resource directly related to that URI. For each one of the 2318 tagging activities (i.e., triples of the form huser; tag; photoi) we run the baseline, Sem4Tags without selecting the active context, and Sem4Tags selecting the Active context. The semantic associations between tags and DBpedia resources where evaluated by 14 users. For the 15% of tagging activities the evaluators were not able to identify the meaning. For the rest of tagging activites the results are shown in table 1. 6.In tCerOmNsoCf LcoUveSraIgOe NSeSm4Tags is clearly superior to the baseline. This high coverage is due to: 1) the preprocessing activity where tags are normalized, and 2) the amount
Accuracy Approachnlanguage
Base line
Sem4Tags Sem4Tags & Active Context of information in MSR, speci cally the information about the possible meaning of tags. On the contrary, the baseline approach has that low coverage because tags are directly related to Wikipedia content pages, and therefore ambiguous tags, lacking a default article in Wikipedia, are not processed. With respect to accuracy Sem4Tags with Active Context presents the highest value. The use of active context allows us to increase the accuracy in both languages with respect to Sem4Tags. On the other hand, the accuracy of the baseline is very similar to the achieved by Sem4Tags. This fact can suggest that most of the tags are used in the most frequent meaning presented in Wikipedia. 7.
This work is supported by the GeoBuddies (TSI2007-65677C02) and Espan~a Virtual (ALT0317) projects, and the FPI grant (BES-2008-007622). 8.