This paper explores how to discover unexpected information in existing folksonomies (serendipity) using extensive multilingual open source repositories as the underlying knowledge base, overcoming linguistic barriers at the same time. A web application called Flickrpedia is given as a practical example, using Flickr as the folksonomy and diverse natural language Wikipedias as the knowledge base.
Adding meaningful metadata to web content, in order to increase the
utility of information by improve the precision of information retrieval to search
engines, is one of the most desired feature by any user. Folksonomies are a
tentative effort toward this goal. The term ‘folksonomy’ is a fusion of ‘folks’
and ‘taxonomy’ and was originally used in cognitive anthropology studies, but
only very recently it became popular with a specialized meaning [
Unlike traditional taxonomies, as for example the Linnaean system used in
life sciences, there is no explicit hierarchy between tags nor tags are exclusive
– e.g. the photo of a cat may be tagged as ‘cat’ and ‘european’ and ‘animal’,
but there is nothing that say that all cats are animals: tags can be seen as
common facets of the item itself [
Consequently, each tag has two different scopes at the same time: the user’s
defined one – personimy, [
Furthermore, consensus permits serendipity, i.e. users dig the web through
tags finding new, unexpected and useful content, not easily accessible via
traditional search engines. In fact, tags act as filters, i.e. a query on more tags
returns the items tagged with any of the given tags – or with all tags,
depending on the application [
Folksonomies share common problems with traditional taxonomies, due to
the fact tags are words, i.e. alphabetical strings meaningful in some natural
language. In particular, there is no synonym (different word strings, analogue
meaning) nor homograph (identical word string, totally different meaning)
control. In fact, there is no restriction to what people can write as a tag, i.e. no
controlled language: people can externalize their free word association through
tags, which respect their own mental models. Consequently, folksonomies lack
in standardization, i.e. different strategies in tag encoding are possibles, as
for instance dates (28-03-2008, ‘2008March3’, ‘3ˆrd March 2008’ and so on)
or in the case of compounds (‘nice-cat’, ‘nice cat’, ‘nicecat’), not to mention
misspellings, so frequent that tag literacy education was advocated [
One of the existing problems behind folksonomies not fully explored until now is multilingualism. As anedoctical evidence suggests, every tag is written in a human language and users are inclined to write in the languages they are comfortable in. It is certainly desiderable for a user not comfortable in English or other big language (in terms of presence in the web) to search and find tags using a search engine interface in his or her tongue, while the engine searches the corresponding tags in English and in other major human languages.
To do so, the user needs to specify both the tag looked for and the natural language in which it is written in a special web application, which extracts the pairs language-tags in every available language before passing the tags to the folksonomy search engine. Our claim is, when searching in 20 natural languages at same time some interesting photo will be found, that would be undiscovered through a single language search (i.e., serendipity improves).
Flickr, a Yahoo! company, is one of the most popular online photo web
applications – e.g., more than 2 million photos are found if ‘flowers’ are searched, at
2007, April the 11th. In Flickr, users can browse or search photos through tags,
a feature that certainly contributed to its popularity. Moreover, some open
source APIs are available1 and people can make queries to the Flickr
repository through an authentication key given on request. For our application, the
language of choice for the API is Ruby, and the development framework is
Ruby on Rails, as it is easy to produce clean code and reliable web application
very quickly [
In our prototypical web application, Flickrpedia (named derived from ‘Flickr’ and ‘Wikipedia’), users can make queries in Flickr writing a tag specifying its natural language. The system crawls the Wikipedia in the corresponding language and look for an appropriate page. For example, if the user is a German-speaker and he is fond of airplanes, he may put the following pair German:Flugzeug and the system, which can manage case-sensivity, will look for the following page in the German Wikipedia: http://de.wikipedia.org/wiki/Flugzeug where de is the natural language ISO code and Flugzeug indicates the corresponding web page. With the help of regular expressions, Flickrpedia parses the web page and extracts the existing language pairs of the same topic (airplanes) in other languages from the appropriate web page box known as “in other languages”, e.g. English:Airplane, French:Avion – but also minority languages, as Basque:Hegazkin (see Fig. 1). The topic names are passed to Flickr as search queries and thumbnails are given to the user.
While Flickr finds less than 10,000 photos (2007, April the 11th) for the tag ‘flugzeug’ Flickrpedia finds more than 20,000 for the same query, giving a lot of unexpected and relevant photos.
1See http://www.flickr.com/services/api. This paper has shown that serendipity in Flickr can be improved through the exploitation of Wikipedias’s URLs as translation sources. The main advantage is that Flickrpedia should only store the wikipedias according to the existing natural languages – actually, 85. This approach wants to suggest that large and extemporaneus shared information repositories, like Flickr, can be managed through other semi-structured information repositories as the wikipedias – as known, wikipedias are the result of a wide and magmatic community of contributors, even anonymous. Moreover, Flickrpedia, if refined out of its actual prototypical phase, may help users with poor knowledge of major languages to retrieve information only through their lesser-used languages.
Flickrpedia is far from perfect: homographies are still unmanaged, even if wikipedias have disambiguating pages, and it is not clear which wikipedias to choose in order to optimize serendipity. By the moment, the parsed wikipedias are the biggest ones in terms of wiki pages, but this doesn’t give any guarantee of serendipity augmentation. Finally, the API given by Flickr is a severe limit: up to 20 tags can be inserted in a single query request, and up to 60 thumbnails may be given.
However, this approach isn’t limited to Flickr as the underlying folksonomy. Our research direction is towards generalization, i.e. users can choose the appropriate folksonomy performing multilingual queries. Finally, specific and precise metrics for serendipity are needed, in order to achieve more formally sound results.