Currently proposed tagging ontologies are mostly focused on the de nition of a common schema for representing the agents involved in a tagging process. In this paper we describe preliminary research around the idea of extending tagging ontologies by incorporating some domain speci c class de nitions and relations. We illustrate our idea with a particular use case where a tag recommendation system is driven by such an ontology. Besides our use case, we believe that such extended tagging ontologies can bring more meaningful structure into folksonomies and improve browsing and organisation functionalities of online platforms relying on tagging systems.
Tagging systems are extensively used in online sharing sites as a means of content
browsing and organisation. In general, tagging systems allow users to annotate
resources with free-form textual labels chosen by the users of the system. The
resulting set of associations between tags, users and resources that arise in tagging
systems is known as a folksonomy. Folksonomies su er from a number of
wellknown issues including tag scarcity, ambiguities with synonymy and polysemy,
typographical errors, the use of user-speci c naming conventions, or even the use
of di erent languages [
In order to bring some structure to folksonomies, some studies have focused
on the analysis of folksonomies to automatically derive structured or
semistructured representations of the knowledge of the domain, typically in the form
of lightweight ontologies or hierarchical taxonomies [2{4]. However, these
methods still tend to require signi cant amount of manual e ort to provide meaningful
representations. Some other studies have proposed modelling folksonomies and
the tagging process using ontologies [
In this paper, we present some preliminary research on extending a tagging ontology by including the possibility to represent the semantics of a speci c domain. The generic idea is presented and discussed in Sec. 2. In Sec. 3 we describe a practical application in a real-world tagging system where the tagging ontology is used to drive a tag recommendation system. Finally, in Sec. 4, we discuss about possible future directions. 2
Our starting point for the extension of the tagging ontology is the Modular
Uni ed Tagging Ontology (MUTO) [
We propose to extend the tagging ontology in two ways. First, we add a
number of subclasses to the muto:Tag class which can be used instead of muto:Tag
(right side of Fig. 1). These subclasses represent di erent tag categories (i.e. with
a narrower scope than the generic muto:Tag class), similarly to the idea of TagSet
introduced in the SCOT ontology [
Second, we propose to extend the tagging ontology by adding object
properties to model semantic relations among tag categories and resource categories
(dashed lines in Fig. 1). These object properties are useful to, for example, model
dependencies between categories of tags and resources. The speci c meaning of
these semantic relations would also depend on the particular application domain
of the extended tagging ontology. In addition to semantic relations between tag
and resource categories, and given that the muto:Tag class inherits from the
Simple Knowledge Organization System (SKOS) [
We applied an extended tagging ontology as described above in a tag
recommendation task in the context of Freesound, an online collaborative sound database
with more than 200,000 uploaded sounds and 3,8 million registered users [
To improve that recommendation system, we used the extended tagging ontology to model the folksonomy and include some domain speci c knowledge. On the one side, we extended the tagging ontology by adding 5 resource subclasses corresponding to the 5 sound categories mentioned above (e.g. :EffectsSound). Moreover, we de ned 26 tag subclasses that are intended to group the tags in categories according to the type of information that they describe about sounds (i.e. grouped in audio properties). These include categories like \instrument", \microphone",\chord", \material", or \action" (e.g. :InstrumentTag). On the other side, we extended the ontology by de ning several object properties that relate resource and tag categories. These object properties indicate that a particular tag category is relevant for one or more resource categories. For example, :InstrumentTag is relevant for :MusicSound audio category, and this is indicated with a :hasInstrument object property that relates instrument tag category with music resource category. Furthermore, we populated the extended ontology by manually classifying the 500 most used tags in Freesound into one of the 26 de ned tag categories and added these tags as individuals (instances) of the corresponding tag category. This last step was necessary to bootstrap the tag recommendation system (see below).
Using this ontology we can extend the tag recommendation system in a way that, given the audio category detected by the classi er and the object properties that relate resource and tag categories, we can guide the annotation process by suggesting tag categories that are relevant for a particular sound. For example, for a sound belonging to the resource category :MusicSound, we can suggest tag categories like :InstrumentTag or :TempoTag, which are particularly relevant for musical sounds. Once tag categories are suggested, users can click on them and get a list of tag recommendations for every category. This list is obtained by computing the intersection of the tags provided by the aforementioned recommendation system (based on the tag-tag similarity matrix), with those that have been manually introduced in the ontology as tag instances of the selected tag category. See Fig. 2 for an screenshot of a prototype interface for that system. by the extended tagging ontology. 4
In this paper we have shown some preliminary research on extending current tagging ontologies with structured knowledge speci c to the domain of application of a tagging system. By incorporating domain speci c knowledge in tagging ontologies, we expect to be able to bring some semantically-meaningful structure into folksonomies. We have illustrated the idea with a use case in the context of an audio clip sharing site where a tag recommendation system is driven by an extended tagging ontology. Formal evaluation of the ontology-driven tag recommendation system is planned for future work. Besides the described use case, we think that using extended tagging ontologies can improve other aspects of online platforms relying on tagging systems such as browsing and organisation functionalities. The main limitation for such improvements is the population of the ontology. In our use case, we use a manually populated ontology to bootstrap the recommender, but the tagging system could further populate the ontology by learning new \tag individuals-tag category" relations when users annotate new sounds. Furthermore, other knowledge extraction techniques could be used to automatically populate the ontology with information coming from other user-generated data (e.g. in our case could be sound comments or textual descriptions), and even from external data sources from linked open data.