Onto_Folkso_StofA_FLimpens

a)    Freely categorizing

To share and index the large number of resources available on the Web raises several issues that systems based on folksonomies Vanderwal [2004], such as del.icio.us for sharing bookmarks, have recently tried to address. On the other hand, the Semantic Web aims at supporting the exchange of information by developing the interoperability between applications available on the Web. To this end, several methods, tools and principles are proposed, among which formal ontologies play a central role. Generally speaking, ontologies are knowledge representations aiming at "specifying explicitly a conceptualization" Gruber [1993]. More specifically, formal ontologies use formal semantics to specify this conceptualization and make it understandable by machines. The obstacles to a generalization of ontologies lie mainly in their cost of design and maintenance.

The problem we address here is the need for the users of social Web platforms to find an agreement about the knowledge representations that support their collaborative use of the system. To this regard, folksonomies are often seen as the bottom-up approach, while formal ontologies of the Semantic Web are considered to be necessarily a top-down approach. In this report we try to show that opposing folksonomies and ontologies in this way is counterproductive, and the works we present here show the potential of combining both approaches in order to collaboratively build up solid knowledge representations that are both representative of the communities of users, and at the same time allows for better retrieval or exchange of information.

The Web 2.0 consists essentially in a successful evolution of the Web supported by some principles and technologies. Social tagging and the resulting folksonomies can be seen as two of those principles which have emerged and met a growing success within Web 2.0 applications. The simplicity of tagging combined with the culture of exchange allows the mass of users to share their annotations on the mass of resources. However, the exploitation of folksonomies raises several issues given by Mathes [2004] and by Passant [2009](section 2.2.3): (1) the ambiguity of tags, for one tag may refer to several concepts ; (2) the variability of the spelling, for several tags may refer to the same concept; (3) the lack of explicit representations of the knowledge contained in folksonomies; (4) the difficulties to deal with tags from different languages. Another challenge is the need to assist the life-cycle of the folksonomies and the ontologies which support the knowledge bases of social Web applications. Our hypothesis is that the synergy of both folksonomies and ontologies may bring great benefits. Research has been undertaken to tackle the problems posed by the annotation and the exchange of the resources on the Web. The systems or methods they propose strive to reconcile ontology-based models and folksonomy-based models.

b)     The need for a shared vocabulary: tidying up online communities

Most of the research works we present in this report take place within the social Web which includes all types of groups of people communicating online. These communities range from groups of people who do not know each other in the real life but contribute to the same sharing platform (as in Wikipedia or delicious.com where users contribute to an encyclopedia or a social bookmarking database), to collaborators who work together and exchange knowledge online.

One of the most commonly cited notion about communities with respect to knowledge sharing issues is probably the notion of Community of Practice (CoP) proposed by Lave & Wenger [1991]. The notion of CoP defines a group of people gathered by a commitment to a common activity and sharing common interests, proficiencies, and knowledge. However, other notions have emerged to describe the specificity of online communities because the criterion of sharing a common commitment is not always fulfilled in communities communicating online.

Tardini & Cantoni [2005] tried to apply the concepts of semiotics [Saussure, 1916,Hjelmslev, 1963] to describe and characterize online communities. They distinguish two main types of communities. (1) Paradigmatic communities are groups of people simply having something in common, such as the fact of using the same website for the "Wikipedia visitors" community. It is possible to belong to several paradigmatic communities at the same time, and these communities can be embedded in each other such as the community of "eyes specialists surgeons" in the surgeons community. (2) Syntagmatic communities consist in groups of persons who are characterized by their differences and complementarities, and who share a common activity. This type of community is also very close to the concept of CoP, but is less constrained concerning the commitment to a common activity.

The next step consists in finding criterias to evaluate whether a group of online users form a syntagmatic or a paradigmatic community, as this distinction has some consequences about the characterization of the type of knowledge structure which will better fit their needs. For instance, the visitors of a web site form a paradigmatic community which can evolve into a syntagmatic community as soon as the visitors start exchanging more and realize they have a lot of things in common. To this respect, Tardini et al. give five conditions which should be fulfilled for a group of users to form a syntagmatic community: (1) a shared environment of communication, (2) a reasonable level of wealth of exchange, which allows for the discovery of common interests, (3) the arousal of a feeling of belonging to a group, (4) the development of a common symbolic space called the "semio-sphere", and (5) the development of a group identity. The development of a semio-sphere is particularly relevant to the scope of this report in that shared ontologies should depict as closely as possible these semio-spheres, and also in that it seems irrelevant to start building collaboratively an ontology if the community is still at the paradigmatic state.

To this respect, the authors have analyzed several online communities (from users of search engines to online video-game players) and came to the conclusion that out of the five conditions mentioned above, the common interests, the feeling of belonging, and the development of a common identity are the most important to constitute a syntagmatic community. The feature of the semio-sphere of a syntagmatic community tells also a lot about the features of the community itself: the more complex the semio-sphere, the more closed the community; on the contrary, the more simple and affordable to newcomers the semio-sphere, the more open the community. This description of the semio-spheres is also close to the distinction between broad and narrow folksonomies (see section 2.1).

c)     Comparison of different types of knowledge representations used to index resources

Before presenting the different attempts to overcoming the gap between folksonomies and ontologies, let us recall briefly the main types of structured knowledge representations traditionally used to classify or index resources or documents. These knowledge representations are also called "termino-ontological resources" in the literature and differ mostly from each other in their level of formal structuration, or in their purpose, or in the way in which they are elaborated.

1. Epistemic classifications (such as Dewey's classification [Dewey, 1876] used for classifying books in libraries) consist in defining a vocabularies which can be universally shared. This type of classification (but more flexible than Dewey's classification scheme) is met for instance in the Dmoz¹ initiative to build a directory of Web pages where specialists debate about categories which should be used to classify all the Web pages.
2. The origin of thesauruses goorigin of thesauruses goes back to the 4th century, but the first modern thesaurus is attributed to the British Peter Mark Roget². Modern thesauruses and other types of controlled vocabularies, such as taxonomies, consist in notions or concepts which are defined and hierarchically structured. They provide descriptors used to index documents and are aimed mostly at navigation purposes. The notions composing thesauruses can be contrasted with the concepts of formal ontologies in that they are oriented towards the descriptions of resources, and are not aimed at describing "what something is", but rather "what something is about" (see SKOS specification and the definition of the skos:Concept class³). Moreover, the types of semantic relations linking the concepts of thesauruses are usually limited to "broader", "narrower", or "related".
3. Along the expansion of the web, semi-formal and shared knowledge representations have been proposed to organize the information on the Web. Such approaches include Topic maps⁴ [Park & Hunting, 2002], or, with a greater stress on dealing with conflictual views within the communities of users, "semiotic ontologies" Cahier et al. [2005]. Primarily, semiotic ontologies and Topic Maps can be used by themselves. In some other cases they can also be considered as an intermediary representation to formal ontologies, in that they are not extended by a "referential formalization"⁵ but are based on "semiotic expressions" or Topics dealing with another type of semantics which rely mostly on human interpretation. These approaches differ from formal ontologies in their purpose, which is not to obtain a formal and operational scheme, but rather "description networks" used by humans to navigate a corpus of documents and resources.
4. Formal ontologies consist in a specification of the conceptualization of a domain of knowledge with the help of formal concepts and properties linking these concepts [Gruber, 1993]. They are at the core of the original vision of the Semantic Web proposed by Berners-Lee et al. [2001]: "The Semantic Web is an extension of the current Web in which information is given well-defined meaning, better enabling computers and people to work in cooperation". Thus, ontologies are at the interface between humans and machines, and can be seen as the formalization of a field of knowledge, given for a specific problem or task Bachimont [2000]⁶. This formalization of a domain allows in turn for making inferences and expandexpands greatly the possibility of querying when looking for resources annotated with formal ontologies.

In comparison with all the above mentioned structures of knowledge representation, folksonomies can be seen as semiotic representations of the knowledge of a community, but they do not include any semantic structureuring. They are not either truly elaborated collaboratively, since they consist merely in a social aggregation of individual knowledge. However, their indisputable advantage over the other types of representations we mentioned above is to have atheir simplicity (they require a minimal cognitive cost of elaboration [Sinha, 2005]), and they have proved to be likely to bethat made them adopted by a mass of users.

d)    Different ways of considering the link between folksonomies and ontologies

The aim of this report is to give the current main approaches to reconcile folksonomy-based and ontology-based approaches to support social interactions. But theThe bridging of ontologies and folksonomies can be done in different ways:

Deriving ontologies from folksonomies   

It is possible to analyze the multipletotake into account the multiple dimensions of folksonomies as they consist in a triadic structure where tags are associated by people to resources (“who tags what”). This is what Mika [2005], for instance, does in order to extract broader and narrower relationships between tags and to build what he calls "lightweight ontologies", that is, ontologies which consist in an ensemble of terms connected with a limited set of semantic relationships (broader, narrower, related for example).

Synchronizing ontologies life-cycle with folksonomies :  

Folksonomies, thanks to their versatility and their ability to integrate fresh new vocabularies, are a good opportunity to populate ontologies or to provide suggest for new concepts to that could be candidates to be added to in an ontology. Passant [2007] exploited this feature of folksonomies to populate a corporate ontology and to support the folksonomy based system he developped to annotate a corporate blog. In his system, ambiguous tags are associated to clearly defined concepts, by the users, while tagging.

Semantically enriching folksonomies   

Even if ontologies and folksonomies may remain different entities, several approaches have proposed to semantically enrich folksonomies by adding a semantic layer to it, or by attempting to semantically structure them with the helps of others, already available ontologies. For instance, Specia & Motta [2007] have developed a system that apply several semantic treatments to a folksonomy, such as finding equivalent tags or grouping similar tags based on similarity measures computed according to the structure of the folksonomy. Then, they query ontologies on the Semantic Web and try to match the tags from these clusters with concepts from ontologies in order to link the tags with semantic relationships. The main limitation of such an approach is the limited coverage of currently available ontologies.

Semantic Web formalism for interoperability  

Another great benefit of combining ontologies and folksonomies lies in the interoperability brought by the formalism of the Semantic Web. The Linking Open Data project ⁷ is consists in to extending the Web with data sources semantically interconnected together and which that publish varied open data sets in RDF format and following a set of ontologies describing the different types of resources. Members Ontologies  fromof the Linking Open Data initiative and ontologies include SIOC⁸ used to describe online communities exchange, or SKOS ⁹ used to describe thesauruses.

e)    Organization of the report

This report is organized as follows. In Section two, we present the different approaches that analyze the nature of folksonomies and tags. Section three is concerned with the analysis of the semantics inherent to the folksonomies and the relationships between the tags which can be extracted in order to build ontologies. Section four will cover methods which semantically enrich folksonomies or which integrate tagging practices in ontology maturing processes. Section five will give an overview of different types of usages of knowledge sharing platforms, and section six will conclude this report with a discussion.

f)      Folksonomies as collaborative classification means

According to Golder & Huberman [2005], social tagging can be seen as a cognitively lighter alternative system of classification to controlled vocabularies and hierarchical systems, which can be seen in a hierarchy of folders for instance. Social tagging is also about sense making since the goal of a tag for its author is to organize its knowledge sources with labels which are a way of making sense of the resources he tags. Tags are then an important sign of what matters for the users and how he describes it.

But social tagging is also about collaborative sense making, and as such, has the potential of revealing the fuzziness of the manifold individual categories merged under the same tag. In the same trend of ideas, Veres [2006] says that tags are the results of ad hoc categorizations, that is, categories which interface between each user's "world model" in order to achieve a goal. But their linguistic properties revealsproperties reveal that tags can also be similar to standard categories in taxonomies.

Golder & Huberman [2005] detailed seven functions that tags may perform for bookmarks in the context of a typical application of social tagging: (1) "Identifying What (or Who) it is About", that is, the topic of the item tagged; (2) "Identifying What it Is", for example an "article", a "blog" or a "book"; (3) "Identifying Who Owns It", or also to whom this bookmark may be forwarded (see also the "network tags" in delicious.com social bookmarking service); (4) "Refining Categories", that is, tags which refine or qualify existing categories, such as numbers; (5) "Identifying Qualities or Characteristics" such as adjectives characterizing the opinion of the author; (6) "Self Reference", such as tags beginning with my; (7) "Task Organizing" which correspond to a particular type of ad hoc categories, oriented towards a specific task such as "toread".

Folksonomies have also been characterized by Vanderwal [2004] who distinguish "narrow folksonomies", in which the personal use of tags is predominant, and "broad folksonomies" in which the use of tags is oriented towards more collective and social purposes (which may correspond in some cases to the first three functions given by Golder & Hubermann). Folksonomies are thus a combination of terms which can serve collaborative categorization, and other terms which are only useful for their authors.

g)    Formal definition

In order to further analyze the structure of folksonomies, we have to model them formally. Hotho et al. [2006] thus proposed a formal definition of a folksonomy which they model as a tuple F:=(U,T,R,Y) where U, T, and R are finite sets, whose elements are called users, tags and resources, respectively. Y is a ternary relation between them such that Y\subseteqq U×T×R, and is called tag assignment. As a collection of data provided by a group of individuals, a folksonomy can be seen as the collection of the "personomies" of all the users. Let us call Pu the personomy of a given user u  Î U, Pu is the restriction of F to u, i. e., Pu:=(Tu,Ru,Yu), with Yu:=(t,r) Î T×R|(u,t,r) Î Y that is, the ensemble of all the tag assignments of user u. As Mika [2005] showed it, such a model induces a graph structure of folksonomies. Thus, a folksonomy can also be seen as tripartite hypergraph H(F)= á V,E ñ where the vertices are given by V=U È T È R and the edges by E=u,t,r|(u,t,r) Î F (see the graphic representation of a folksonomy given by Halpin et al. [2007] in figure 1).

Figure 1: Tripartite graph structure of a tagging system. An edge linking a user, a tag and a resource (website) represents one tagging instance ([Halpin et al., 2007]

h)    Structure and dynamics of social tagging

Golder & Huberman [2005] proposed one of the earliest quantitative analysis of social tagging in which they discuss its nature as well as the dynamics which can be uncovered with statistical analysis lead on the multi-dimensions structure of folksonomies. Golder & Hubermann give some trends in the use of tags in a social bookmarking system (delicious.com). They remark that users have a tendency to use first more general terms when tagging, the first tag having the greatest frequency of occurrence among all the usersusers’ tags, and successive tags having generally a smaller frequency. They also observed stable patterns in the distribution of tags for a given resource (URL in delicious.com). Empirically, once a URL has been bookmarked more than 100 times, each tags frequency remains in a stable ratio of the total frequency of all the other tags used for this URL.

Halpin et al. [2007] pursued this analysis of the dynamics of folksonomies and looked for distribution laws in the frequency of use of the tags. They make the same hypothesis that Golder & Hubermann suggest, that the most used tags to annotate a resource remain the same after a certain amount of time, and they show that this distribution follows a power law. They verify that hypothesis for the seven to ten tags most often associated to popular Web resources posted on delicious.com. These observations may be explained by an imitation process, augmented, in the case of delicious.com, by popular tag suggestions while tagging.

But, as Golder & Hubermann suggest, the stability observed in the distribution of the most popular tags persists even for less common tags, which are not shown as suggestions. The choice of the same tags may also be explained by the fact that users share some of the knowledge they express individually when tagging bookmarks. Golder & Hubermann add that this stability in the characterization of some items is linked with the stability of the ideas and characteristics symbolized by the tags; and that, likewise, this stability may no longer persist when a new concept emerges for describing the same items. This was the case, for example, when the concept "ajax" emerged within the realm of Web designers to describe a set of technologies which were all previously known but not named under a single term.

It is also interesting to look at the distribution of tags for smaller folksonomies, as for instance, Passant [2009] in the context of a corporate folksonomy. In this foklksonomy,Passant [2009] show that tags follow a distribution in which a lot of tags are used a few times. For example, out of the 12257 tags used to annotate 21614 blog posts, 68% are used at most two times, and only 10% are used more then 10 times. As Hayes et al. [2007] showed, it is more difficult to apply classical clustering techniques on this type of distribution in which tags do not neatly partition the annotated data. Indeed, in these cases one should include the content of the annotated data in the analysis of the folksonomy structure.

Concerning the relationships between the tags in a folksonomy, Halpin et al. looked for semantic relationships between the most used tags with the help of inter-tag correlation graphs. Each node of these graphs representrepresents a tag, and can be seen as a circle whose diameter is weighted by the frequency of occurrence of this tag. The length of the edges of these graphs are is weighted by their degree of cooccurrence. The degree of cooccurrence CoocDegree(T_i,T_j) of a pair of tags T_i,T_j is given by :

Where N(T_i) and N(T_j) denotes the number of times each tag T_i and T_j is used individually to tag all pages, and N(T_i,T_j) denotes the number of times two tags are used to tag the same page, summed over all pages. This visualization (shown in 2) can be seen as a tool for assisting the construction of ontologies out of folksonomies by helping identify visually the most related tags to a given tag.

Figure 2: Visualization of a tag correlation network, considering only the correlations corresponding to one central node "complexity" (data source, delicious.com) [Halpin et al., 2007]

i)       Looking for common associations in folksonomies

Other works proposed to apply data mining methods to the tripartite model of folksonomies in order to retrieve information in their structure. Jäschke et al. [2008] proposed to use formal concept analysis techniques in order to discover the subsets of users sharing the same conceptualizations on the same resources. To do so, they build triples of sets ({R}, {U}, {T}) called tri-concepts where each user of the set {U} has tagged each resource of the set {R} with all the tags of the set {T}. According to the authors, extracting tri-concepts from folksonomies is a first step to build more structured ontologies from folksonomies. Ontologies are thus seen as social constructions where each concept is described by a set of tags which belong to a set of users and are used to characterize a certain kind of resources.

Other data mining techniques have been applied by Schmitz et al. [2006] to extract association rules from folksonomies. The first step is to project the tripartite model (Resources, Users, Tags) onto a two-dimension structure called a context in formal concept analysis [Wille, 1982]. For instance, one can consider all the tuples (Users, Resources) associated to a set of tags T_x. Then Schmitz et al. [2006] apply classical rule mining techniques as proposed by Agrawal & Swami [1993]. An example of association rule that may be derived from this projection is :is: all the users associating tags from the set T_A to a set of resources, often associate the tags from the set T_B to the same set of resources. This kind of association rule may be exploited for example in a recommendation system. Other types of association rules may be a powerful tool to identify sub-groups of users sharing the same tagging practices or interested in the same topics.

j)       Comparison and intermediary conclusions

In table 1, we compare the different approaches presented in this section. We divided these contributions in two categories. First, we can mention the qualitative studies conducted on folksonomies. Golder & Huberman [2005] have analyzed the usages of folksonomies and have proposed seven functions that tags may perform for bookmarks bookmarking in the context of a typical application of social tagging. Vanderwal [2004] distinguished broad folksonomies (when tags tend to be understandable by numerous users) from narrow folksonomies (when tags are more user-centered).Veres [2006] tried to define the linguistic nature of tags and showed that some tags correspond to taxonomic categories, while other tags correspond to ad hoc categories serving user's purposes.

Secondly, we distinguished the contributions which focus more on a quantitative analysis of folksonomies. Mika [2005] and Hotho et al. [2006] proposed a formal definition of folksonomies, but Mika [2005] pointed out their graph-like properties and defined them as tripartite hypergraphs. Halpin et al. [2007] pursued this analysis of the dynamics and usages of folksonomies initiated by Golder & Huberman [2005] and showed that the distribution of most frequent tags of popular web pages on delicious.com follow power laws. Schmitz et al. [2006] applied classical rule mining techniques to discover association rules within folksonomies, and Jäschke et al. [2008] used formal concept analysis methods to unveil similar conceptualizations in the tagging of resources shared by groups of users of a social bookmarking site.

Table 1: Comparison table of the approach of section 2 analyzing folksonomies

k)    Measuring the relatedness between tags

Cattuto et al. [2008], and latter Markines et al. [2009], proposed different ways of measuring the similarity between tags and resources in a folksonomy. These approaches can be seen as generalizations of several other approaches (like Mika [2005],Specia & Motta [2007]). The computation of similarity of tags is often the first step to further process the folksonomy data and infer semantic relationships between tags (see 3.2), or to cluster similar tags (see 3.3).

l)       Inferring subsumption relations

Several approaches have been proposed to infer subsumption relationships between tags:

Mika [2005] used the graph grouping similar communities of interest (as described above in section 3.1.2) to derive subsumption properties between the tags thanks to the inclusion of communities of interest. [m1] In this case, a community of interest may be represented by all the actors having who used the tag "fishing". If the communities of interest "fishing" and "nautic activities" have a number of actors in common, the tags "fishing" and "nautic activities” will be considered as semantically related. Furthermore, if the group of actors using the tag "fishing" is a subset of the group of actors using "nautic activities", "nautic activities" will be set as a broader term than "fishing". Mika also shows that subsumption relations are more relevant when derived from inclusions of communities of interest, than when derived from the co-occurrence patterns (given by the Tag-Tag context similarity measure) between tags.

The algorithm proposed by Heymann & Garcia-Molina [2006] takes as input the list of tags in descending order of their centrality in the similarity graph (based here on cosine distance based on the Tag-Resource context). The hierarchy of tags is built starting from the root node, and each tag, taken in order of centrality, is added either as a child of one of the node or the root node (depending on a threshold value of its similarity to these nodes).

Schmitz [2006] used conditional probability to detect subsumption relationships between tags. Given a tag pair (Ti,Tj), let us call the frequency of occurrence of each tag N(Ti) and N(Tj), and the frequency of cooccurrence of both tags N(TiÇTj). The conditional probability P(Ti|Tj) of having Ti given Tj is calcuted as follows:

And conversely

By comparing both values with each other, we can deduce which of the tags of the pair is more dependent on the other tag. In order to induce a hierarchy from flickr.com tags, have adapted the method proposed by Sanderson & Croft [1999] has been adapted, integrating new statistical thresholds to account for the specificity of folksonomies. Thus, tag Ti potentially subsumes tag Tj if :

with

Where t is a given co-occurrence threshold, N(Ti) and N(Tj) must be greater than a minimum value Tmin, and U(Ti) is the number of users that use tag Ti at least once and must be greater than a minimum value Umin.

Schwarzkopf et al. [2007] also proposed building taxonomies out of folksonomies for user profiling purposes. They pointed out the limitations of the algorithm proposed by Heymann & Garcia-Molina [2006]. Indeed, the cosine similarity measure used in this algorithm does not take into account the popularity of tags, while Mika [2005] suggested that relationships between tags established via users (and thus, accounting for the popularity of use, since a tag can subsume another tag only if it is more often used) are more suitable to infer narrower/broader relationships. Schwarzkopf et al. [2007] thus proposed exploiting the latter method to infer subsumption relationships between tags, such that: "If resources tagged with t0 are often also tagged with t1 but a large number of resources tagged with t1 are not tagged with t0, t1 can be considered to subsume t0". Schwarzkopf et al. [2007] also address the transitivity problem of the inferred subsumption relations and noticed that the "similarity context" of tags is not taken into account when adding a child-tag to a parent-tag, that is, the similarity with all the ancestors of a parent-tag. Thus they combine the users-based associations of tags (Mika, 2005) and the cosine similarity measure to prevent a child-tag from being added to a branch for which it is not related to the origin, such as in design > web > howto > productivity > business, where each link makes sense but the whole chain does not.

m)Clustering tags

Here we briefly describe different ways of clustering either equivalent tags, that areis spelling variants of the same term, or similar tags, that iares, tags which are considered to be the most closely related with each other.

n)    Comparison of the approaches and intermediary conclusions

In this section we have presented several approaches which extract semantic relations between tags by analyzing the structures of folksonomies (in contrast with other types of methods which uses external semantic resources to achieve this task, see section 4.5). All the approaches presented in section 3.2 seek try to find subsumption relationships [m2] between tags. The case of Cattuto et al. [2008] is particular in that they characterize different types of similarity measures according to the type of semantic relationships to which they each correspond. Thus, their method can be used to find related tags which share a subsumption relation with a given tag t, however without being sure whether these related tags may subsume or be subsumed by tag t. On the other side, the approaches of Begelman et al. [2006] and Specia & Motta [2007] propose a method to cluster similar tags.

The type of similarity measure allows to distinguishdistinguishing all these methods. Mika [2005] applied social network analysis on different projections of the tripartite structure of folksonomies. Hotho et al. [2006] adapted the PageRank algorithm to the case of folksonomies in order to find not only relationships between tags, but also between users and resources. Schmitz [2006] used conditional probability methods to induce a hierarchy from Flickr tags. Begelman et al. [2006] look closely at the distribution of the cooccurring tags for a given tag, and calculate the threshold above which its cooccurring tags are strongly related to it. Then, several approaches use distributional measures but with different context of aggregation of the folksonomy data, Heymann & Garcia-Molina [2006] using the resource context of association of tags, Specia & Motta [2007] using the tag context of association of tags, Schwarzkopf et al. [2007] using a composite measure mixing tag context and user context of association of tags. Finally Cattuto et al. [2008] proposed an analysis of the different context of distributional aggregation, while Markines et al. [2009] proposed a new type of measure based on mutual information calculus, and a framework to analyze the different types of similarity measures between resources of a folksonomy.

Table 4: Comparison table of the approach extracting semantic relations between tags by analyzing the structure of folksonomies

o)    Collaborative semantic structuring of folksonomies

Weller & Peters [2008] defines the different aspects of folksonomy improvements taken at a collaborative scale. They definedifferent structural levels on which folksonomies may be improved and edited by the contributors to a folksonomy. (a) Whole document collection vs. single document level. Shall we edit the tags as associated to all the documents, or restrain the editing to tags associated to a single document ?. (b) Personal vs. collaborative level: should we share the edition of tags or should it be personal? (c) Intra and cross-platform level: depending on the platform we are considering, the treatment applied may differ.

To tackle the problems of ambiguity or misuse of tagging (like spam), Gruber [2005] proposed to "tag the tags". It would then be possible to state that this tag is the synonym of this other tag, or that this tag does not suit this object, integrating mechanisms of regulation like those observed on Wikipedia. Tanasescu & Streibel [2007] applied the idea of Gruber and extended social tagging systems with the possibility to tag the tags themselves and the relationships between them. Indeed, classical tagging systems allow their users to add a "tagging relationship”, that is a "is_tagged_by” link between a keyword and a document or a Web resource. But richer information may be obtained from the tagging activity, like the relationships between the tags. These tagging can easily be expressed with triples, such as "car" - "is_a" - "vehicle", all these tags being freely added by the users. This feature allows exploiting the technologies of the Semantic Web to assist navigation and to suggest to the user other terms semantically related to her query. To prevent irrelevant contributions, the authors proposed solutions based on votes for some tags, in order to appreciate or depreciate them, or solutions based on points that will be granted either to contributors to the tagging task, or to evaluators of the tags of others. Other incentives to contribution could be provided with the "games with a purpose", that is activities presented as games but exploited for a utilitarian purpose, such as categorizing content from the Web Siorpaes & Hepp [2008].

Huynh-Kim Bang et al.[2008] also proposed to let the users add semantic information while tagging. The goal of is to provide communities of teachers with a tool to organize the educational documents they share, and this tool should merge the flexibility of social tagging and the possibilities of inference brought by semantic formalisms. Thus, they proposed to use structurable tags, that is are tags which can be linked to other tags with a limited set of semantic relationships (in contrast with the openness of the "extreme tagging" of Tanasescu & Streibel [2007]). Two types of semantic relationships are offered to users, each symbolized by a character that users add while tagging:, : the subsumption of a tag by another tag symbolized by the sign ">" (as in "plane > airbus", meaning that tag "plane" subsumes tag "airbus"), and the synonymy between two tags symbolized by the character "=" (as in "test = tests"). Just as all tags are aggregated within a folksonomy, the semantic relationships created by users are also aggregated, meaning that once a user creates a relation between two tags, this relation will be applied to all the users using the same tags.

In the same trend of ideas, Gnizr¹³ and Semanlink¹⁴ [Servant, 2006] allow to define semantic relationships between tags. Gnizr describe tags and semantic relationships between them with ontologies presented above, such as SKOS for the subsumption relation, and the TagOntolgy for the tags. Semanlink proposes its own model, but which inherits from SKOS. We should also mention here the "machine tags" in Flickr¹⁵, where users can define enriched tags in the form of predicate:attribute=value, such as dct:description=New-York or geo:lat=42.33. This type of tags can easily be translated and modeled into RDF triples via the API Flickurl¹⁶.

p)    Ontologies for modeling folksonomies and online-communities

Gruber [2005] states that there is no opposition between ontologies and folksonomies and proposes constructing an "ontology of folksonomy". The "TagOntology" is a project of an ontology dedicated to formalizing the act of tagging. This model brings in four entities to describe tagging: the tagged object or resource; the term used to tag; the user tagging; and the domain in which the tagging takes place (it can be the service used for instance). Gruber suggests reifying the tagging and to consider each tag as an object as such, and below we will see the different implementation of these ideas.

The Semantically Interlinked On-line Communities (SIOC) project of Breslin et al. [2005] provide developers of social Web platforms a formal and technological framework to describe the resources exchanged within and across on-line communities. The formal scheme they propose uses other ontologies like the Simple Knowledge Organization Scheme SKOS¹⁷ which describes systems of organization of knowledge, and Friend Of A Friend FOAF¹⁸ Brickley & Miller [2004] which describes the multiple identities and acquaintances of a user (see figure 7). SIOC describes the most common elements present on Web sites of communities: the concept of "site", the concept of "post" of a Weblog, the concept of "forum", etc. Starting from this vocabulary, the SIOC project proposes tools to automatically annotate the content of some common Web applications (e.g. wordpress.org) according to the SIOC ontology.

The SCOT¹⁹ project proposed by [Kim et al., 2007] aims at representing a folksonomy model with the help of ontologies. This model of tagging is grounded on the Tagging Ontology proposed by Newman et al. [2005] and has four main entities "tagging" itself as an action performed by a user (modeled with sioc:User class), "tag" (scot:Tag, subclass of the tags:Tag class, itself subclass of the skos:Concept class), "cloud of tags" as the containers for the tags of a user, the resource annotated with tags being modeled as sioc:Item (see figure 8). SCOT exporter allows to mapmapping content from a given Content Management System (eg. Wordpress) into SCOT ontologies, allowing interoperability between different tag spaces and the possibility to form groups of similar or related tag clouds. One of the most direct use case of the SCOT model if the use of meta-search which would allow users to find similar folksonomies, for instance based on the use of tags (number of common tags, that is, the number of tagging using the same scot:Tag instance, since all tags spelled the same will be automatically merged).

Other models of tagging have been proposed, such as the one developed by Echarte et al. [2007] or TagOnt²⁰, but none of them seem to have been used contrary to SCOT or SIOC. The Semantic Desktop project NEPOMUK also proposed a class to describe tags through its ontology NEPOMUK Annotation Ontology²¹: the class nao:Tag and a property nao:has_tag, but without considering the action of tagging as a core element of the model of a folksonomy. Kahan et al. [2002]also proposed "Bookmark", a model to describe the infrastructure of the social bookmarking platform Annotea²². Even if this model does not include the notion of tags, it allows to link a resource with the terms used to annotate it with the class bookmark:Topic and the property bookmark:Topic. This model also proposed to organize the topics with the property bookmark:subTopicOf, similar to the SKOS property skos:broader. We should also mention here the microformat²³ rel:tag. Microformats are the product of a community initiative which defines structured metadata which can be embedded within Web pages via simple html tags attributes ²⁴. Thanks to GRDDL (Gleaning Resource Descriptions from Dialects of Languages [con, 2007]), which allows to transform XML dialects into plain RDF, we can transform annotations written with the rel:tag microformat into RDF triples based on the scot:Tag class for instance.

These ontologies tend to realize the "Web of Linked Data" (now named Linking Open Data²⁵) which consists in a vision of the Web where the sources of data and the schema describing them are located with URIs and interconnected in a decentralized way. This project can be realized thanks to ontologies describing the infrastructures where data is stored (such as SIOC, SCOT, and FOAF describe the actors of the social web and the type of data they exchange), or ontologies describing the content or the topics of the data (such as the DBpedia project [Auer et al., 2007] which publishes the Wikipedia content and category structure in a publicly available RDF data store²⁶). This project aims at enabling users to access content not only via HTML hyperlinks, but also thanks to the concepts which can be attached to them.

Figure 7: Modeling online communities: the SIOC model

Figure 8: Modeling tags and folksonomies: the SCOT (scot:) and TagOntology (tags:) models

q)      Infrastructure for linking tags with ontologies

Passant & Laublet [2008]have proposed the MOAT ontology (moat-project.org) which allows users to link the tags they use with a resource (identified with a URI) which represents the meaning of the tag. The MOAT ontology reuses other ontologies such as the FOAF[Brickley & Miller, 2004] ontology to represent the users, or the TagOntology [Newman et al., 2005] to represent the tagging activity, and specifically the "restricted tagging" which corresponds to the link between a tag (defined with MOAT's own class moat:Tag), a user, a tagged ressourceresource and a meaning resource (see a graphic representation of MOAT in figure 9). The meaning resources can be any Web pages (such as Wikipedia pages), but also concepts of online semantic resources such as ontologies or thesauruses, the main idea of the MOAT project being to contribute to the elaboration of the Web of Linked Data.

We should also remark that in both approaches proposed by Passant & Laublet [2008], the semantic connection is made on the tagging and not merely on the tag itself. The tag is simply taken as a string of characters connecting to an act of tagging and which can be connected to several meanings; tags can be ambiguous, but the act of tagging will be perhaps more imprecise than ambiguous, and thus Passant & Laublet [2008] propose to allow users to precise the intention behind their tagging.

Figure 9: Description of the MOAT ontology to link tags with unambiguous meanings [Passant & Laublet, 2008]

r)      Linking tags with professional vocabularies

Passant [2007] proposes strengthening the social tagging interface of a corporate Weblog with a centralized ontology. In his approach Passant considers tags as character strings linked with formal concepts with semantic properties. This association of tagging and ontologies is used here to disambiguate the different meanings of tags. While tagging, users are suggested to connect the terms with which they are tagging to a controlled vocabulary. Thus, if a tag corresponds to two different concepts (for instance the tag "RDF" may correspond to "Resources Description Framework" or to "Rwanda Defense Forces"), the system asks the user to choose the appropriate concept. When no existing concept matches the user’s concept, users are free to propose a new one to the administrators, who in turn will put it in the right place in the ontology. Social tagging is seen here as en empowerment of the construction of an ontology, and ontologies help disambiguating the possible meanings of a tag.

Similar to the approaches described above, the idea of Good et al. [2007] is to link tags to professional annotations, by providing an access, while tagging, to pre-definedpre-define terminologies organized in formal ontologies. Users can then choose an unambiguous term to use as a tag. Entity Describer uses Connotea, a social bookmarking service specialized in the domain of biology, and enrich the interface (via bookmarklet and a GreaseMonkey Script) by allowing users to access professional terminologies (like MeSH e.g.) in order to use them as a source to select from. When doing so, users can also check the definition of the term they are choosing, which helps them validating or not their choice. The advantages of such an approach are: (1) it does not force users to choose only from a set of controlled terms, but proposes to choose from them. (2) It also allows some extended search capabilities within the collection of bookmarks annotated with controlled terms from ontologies, like finding related terms, narrower or broader terms, etc. (3) It is possible to further exploit the relationships between the terms from the controlled vocabularies and the terms freely chosen by the users: since it is possible to tag an item with both, one may infer semantic relationships between the tags more precisely.

s)     Assisting semantic enrichment of tagging

The methods we present below differ with the ones above in that they seek to assist the users in the task of linking tags with ontologies. Thus, they do not necessarily make use of the infrastructure described in section 4.3, but focus more on the automatization of the process of semantifying tags.

t)      Tagging and collaborative ontology maturing processes

Following the distinctions brought by Weller & Peters [2008] between the individual and the collective level at which folksonomies can be modified, we can distinguish the approaches where the users merely propose new concepts [Passant, 2007], with approaches where users can directly edit the whole shared ontology [Braun et al., 2007], or with approaches where users share their personally maintained ontology (Abbattista et al. [2007]). In the latter case, there will be a need to fine-tune sharing strategies or to use ontology mapping techniques [Euzenat & Shvaiko, 2007] in order to efficiently utilize these shared ontologies.

Braun et al. [2007] address the problem of collaborative ontology editing and criticize current ontology engineering tools in that they do not integrate the collaborative processes. Individual user-oriented methodologies let each user develop her ontology and then share it with others. Semantic wikis are wikis including semantic functionalities, such as an indexing of pages with formal vocabularies, and can also be useful tools to collaboratively build ontologies. Indeed the ontologies elaborated in such a context can be extracted from the categories used to organize or index the context of the wiki pages, such as what have been developed by Auer et al. [2007] with DBpedia.

Braun et al. [2007] propose the following description of the ontology maturing process :

1. the first step is the consolidation of the terminology used in the communities (which could be achieved by analyzing the folksonomy for example),
2. the formalization is performed by identifying the concepts and semantic relationships out of the shared terminology,
3. the axiomatizing consists in formalizing more semantic relations between the shared concepts.

This process should also be integrated in current work processes such as information seeking or distribution. The benefit could be a better motivation from the users to participate in ontology-maturing as they wish to retrieve more accurate content in order to be more efficient, or want to make their own publication more visible. Braun et al. [2007] implemented a prototype which consists in a bookmarking service with some extra capabilities such as (1) suggestion of tags from the already existing ontology, (2) possibility for all users to add or edit new "semantic" tags, (3) knowledge representation models based on SKOS which includes narrower, broader, and related semantic relationships.

In the corporate blog supported by a centralized ontology proposed by Passant [2007], users who tag their posts do not actually directly participate in the ontology maturing process. There, users propose new instances that should be included in the ontology, the actual ontology design being let to the systems administrators. In the same manner, Torniai et al. [2008] propose a method to measure the relatedness between the tags used by the users of an e-learning platform and the concepts of the ontologies supporting the system and maintained by the teachers. The tool they propose assists the maintainer of the ontologies to integrate new terms and to extend the ontology with new concepts conveyed by the tags of the users.

In a more collaborative approach, Buffa et al. [2008a] developed a semantic wiki in which any user can tag the pages and organize the tags of the folksonomy like they would for an ontology. The idea is that each action of the user benefits to all the other users. To this respect, Braun et al. [2007] remark that current collaborative tagging systems offer few functionalities to structure the vocabularies, and when they do, the structuring is not shared among users (for instance in delicious.com, the "super tags", which are used to subsume a bundle of tags, are not shared).

In the same trend of sharing the semantic individual actions, Abbattista et al. [2007]proposed an approach to assist the construction and the evolution of ontologies using collaborative tagging principles. Each user is thus seen as a "knowledge organizer" which contributes to the construction of a collective knowledge base by sharing her structured data. The tool they developed seeks to assist the users in this organization process by (1) providing, for a selected resource, relevant metadata from several repositories, (2) assisting the user in disambiguating the chosen terms using lexical resources (Wordnet [Miller et al., 1990]), (3) suggesting the user to place the terms in relevant location within a personal taxonomy. The user then choose to share parts of her knowledge base, called "binders", that is, groups of annotated resources and the corresponding portion of her personal taxonomy, the result being a shared information space.

u)    Comparison and intermediary conclusions

In table 5 we compare the approach presented above. Gruber [2005] suggested constructing collaboratively an ontology of folksonomy to support more advanced use of tagging. This idea has been implemented by Newman et al. [2005], and further improved by Kim et al. [2007] which integrated their SCOT ontology with SIOC Breslin et al. [2005], another ontology modeling users' interaction on social Web platforms. Later, Passant & Laublet [2008] have extended these interconected schemas with MOAT, an ontology allowing to link tags with online resources to define precisely the meaning of tags and to tie them with the "Web of Linked Data"²⁹, a vision of the Web where resources are linked with each other thanks to the concepts which can be attached to them.

Other approaches focus on user intervention in the process of semantically enriching folksonomies. Huynh-Kim Bang et al. [2008] proposes the concept of structurable tags where users can define semantic relations between tags, and suggest letting the users to tag the links existing between tags. The two latter approaches do not make direct use of semantic Web formalisms, as they focus more on the flexibility of the system than on logical consistency of the knowledge structure obtained. Passant [2007] developed a semantically augmented corporate blog where users can attach their tags to the concepts of centrally maintained ontology, while Good et al. [2007] suggest terms from professional vocabularies fetched online at tagging time. Specia & Motta [2007] and Angeletou et al. [2008] proposed automatic methods to link tags to online ontologies, similarly to Tesconi et al. [2008] and Ronzano et al. [2008] who, first, build sets of terms-meaning by mining Wikipedia, and then link each tag of delicious.com users to a unique meaning. Van Damme et al. [2007], in the same trend, suggest integrating as many semantic online resources as possible, and, at the same time, integrating also users intervention to build, at a reasonable cost, genuine "folks-ontologies". Finally, the approaches presented in section 4.6 focus on the ontology maturing processes and exploit Web 2.0 tools to achieve this task like wikis (Buffa et al. [2008a]), blogs (Passant [2007]), e-learning platforms (Torniai et al. [2008]), personal knowledge organizers (Abbattista et al. [2007]), or social bookmarking sites ()

Table 5: Comparison table of the approach enriching folksonomies which (1) exploit users intervention, and/or (2) make use of external semantic resources, and/or (3) seek the automatization of the process, and/or (4) are based on Semantic Web formalisms

v)       Collaborative information and experts seeking

A problem often posed by collaborative work is expert seeking: how to know "who does what"? The study and the system proposed by Delalonde & Soulier [2007] addresses this problem in the context of a big organization. Delalonde & Soulier [2007] developed "DemonD", a system which aims at creating the conditions of social interactions which yields to capitalized knowledge. DemonD is grounded on personal profiles filled in by the users who state their field of expertise and interests with keywords (which can be seen as tags) and by attaching relevant documents. Then the process starts when one of the user asks a question to the system which selects a list of persons and documents relevant to this question. The selection depends on four main criteria (1) matching keywords in the resource, (2) connectivity with other resources, (3) participation of the person in past interactions, and (4) the reputation evaluated by other peers. Then the system automatically creates a forum of discussion to which the selected persons are invited to participate. The system also includes a step of knowledge capitalization as soon as the original question is answered and that this answer is validated. Thus, this approach includes a collaborative elaboration of knowledge which is partly based on folksonomy-like annotation of the resources. To this respect, Delalonde & Soulier [2007] suggest that the system could be enhanced by suggesting tags when the users build their profiles, and that semi-structured vocabularies could also support the annotation process and help more accurate and more relevant selection of resources.

w)  Sharing on the semantic Web

Other works propose integrating several ontologies to assist the sharing of data. Hausenblas & Rehatschek [2007] designed "mle", a system which automatically treats mailing lists in order to map the structure of email to appropriate concepts of an ontology (SIOC). These annotations, generated in RDF, allow this database to be queried with the language of the Semantic Web SPARQL (www.w3.org/TR/rdf-sparql-query/).

Revyu.com Heath & Motta [2007] proposes applying the principles of the "Web of Linked Data" (see section 4.2) to organize the sharing of reviews of cultural items (books, movies, etc.). Revyu.com includes these principles by (1) allowing anyone to access data stored on other databases in order to prevent redundancies; (2) utilizing RDF to annotate the resources; and (3) keeping open the field of knowledge which can be covered since Revyu.com uses multiple ontologies and other types of knowledge bases to categorize items.

Other approaches allow to semantically structuring the tags in order to enrich social bookmarking services, like GroupeMe!³⁰ Abel et al. [2007] or inter.est³¹Kim et al. [2007]. "inter.est" uses the SCOT ontology which describes the structure of the tag clouds of the users. The goal of inter.est is to help users find groups sharing the same interests by allowing users to aggregate tag clouds, to form groups of exchange and to facilitate the search of similar tag clouds.

x)     Semantic Wikis

Semantic wikis were among the first applications to exploit the potential of ontologies to support collaborative practices. Gaved [2006] thus proposed to develop wikis supporting physical rather than virtual communities, and aimed at providing local information guides which could serve as a community memory for a geographical area. The Open Guides project aims at highlighting the different types of usages and future uses, and to provide a theoretical framework about wikis of locality. The Open Guides were developed after an adaptation of generic wiki principles in order to describe items with locative elements : latitude and longitude, address, opening time, name of the area. These wikis can be considered as semantic wikis since each entry can be exported in RDF/XML, and all the info of each entry is structured following concepts from several vocabularies devoted to the sharing of online resources (FOAF, DublinCore, ChefMoz). Gaved also identified common tasks performed by users of wikis, such as locating, exploring, grazing, monitoring, sharing, and asserting about the information described in each entry of the wiki, leading to truly collaborative semantic processes. The analysis of the usages lead to make some other observations concerning the interface which should empower non-technical experts to contribute, the sustainability of the system which can be enhanced by providing more machine-readable metadata, and the spam of diverse kind which tended to pollute the content of the wikis. This return on experiment is of great usefulness for a designer of collaborative tools and addresses the main problems arising from the use of collaborative semantic tools.

SweetWiki [Buffa et al. 2008b] is another example of semantic wikis: users can edit and modify pages, and also tag any document published on the wiki. The tags are tied together in a folksonomy which is expressed with the languages of the Semantic Web. All the new tags are collected as the labels of new classes which are, by default, subsumed by the class "new concept". All the users are then able to organize the tags of the folksonomy, and to edit them, to add new labels in other languages, to create relations of synonyms, to merge classes, etc. The author of pages can also use tags to keep an eye on the activity of other contributors in a targeted manner: each user can specify in her homepage her topic of interest in the form of tags. For instance, a user interested in wikis will put a tag "wiki" in the field "interested by". Then, whenever a page is tagged with "wiki" or a subclass of "wiki", the user will be notified. This function allows watching content that does not yet exist. By keeping track of created or modified pages, and by analyzing over time the behavior of users, it is possible to detect acquaintance networks or communities of interest. This reveals several possibilities: finding the most active person on a given topic, finding the users using similar tags as others, inferring relationships between tags when they are used by the same users, etc.

y)     Comparison and intermediary conclusions

To conclude this brief overview we can see that, except from Delalonde & Soulier [2007] which propose to assist users in finding experts in the social context of corporate organizations, all the other approaches integrate Semantic Web formalisms to describe their data model. In table 6, we can distinguish these approaches with the type of content they organize or with the type of services they offer. While some applications target no specific social context [Hausenblas & Rehatschek, 2007], some others are set in the Web 2.0 by dealing with the sharing of cultural items [Heath & Motta, 2007] or simply by providing semantically enriched social bookmarking services (Kim et al. [2007] and Abel et al. [2007]). Finally, semantic wikis have been developed to assist the communities of the inhabitants of cities [Gaved, 2006], or to assist the activity of organizations in a broad sense.

Table 6: Comparison table of the approach of section 5.

z)     The best of both worlds

We have seen that it is possible to describe a folksonomy and all the activities occurring on social Web sites with an ontology. In this report we have compared different approaches which aim at bridging ontologies and folksonomies to support and leverage the exchange of knowledge over the social Web. In this regard, these research works are relevant to the design of social Web platforms in that their methods or algorithms can greatly benefit to the final user's experience, by proposing more precise tools to navigate within and across the different platforms. Interoperability is a critical factor for the future of on-line social software, and once adapted to fit the usages, technologies and standards of the Semantic Web can significantly improve the current situation.

The approaches we presented above often complement each other and they can be distinguished against different criteria:

Analysis versus formalization:   

First, we can extract out of the folksonomies a "lightweight ontology thanks to statistical analysis [Specia & Motta, 2007, Mika, 2005, Halpin et al.,2007], or we can directly formalize the tags and their usage among communities of users as with the SIOC and the SCOT projects, or as Gruber [2005] suggested it. Both types of approaches aim at improving information retrieval in folksonomy-based systems.

Type of resources annotated:   

Second, we can distinguish the different types of resources annotated. Breslin et al. [2005] seek to assist the exchange of resources on weblogs and forums, while Heath & Motta [2007] treats the case of reviews. In the same trend, Buffa et al. [2008b] enhanced the collaborative edition of wiki pages with social tagging functionalities and formalisms of the Semantic Web.

Social context:   

Third, we can distinguish different kinds of social contexts. A centralized system works well with a clearly defined field of knowledge [Passant, 2007], while, for instance, the collection of reviews of cultural items or bookmarks will require an open field of knowledge Heath & Motta [2007], and in some cases, the degree of formalism in the knowledge structure would have to be adapted to the level of sharing of the knowledge among the members of the community [Zacklad, 2007].

Integration and design:  

Fourth, we can distinguish the systems with respect to the design aspects. Some approaches can seamlessly integrate current social platforms such as the SIOC plug-ins, which generate metadata about the content organized by some popular Content Management Systems (Wordpress, Drupal). Other works can also simply exploit the data already available Mika [2005],Halpin et al. [2007] and infer extra semantic information which can in turn be used to describe more precisely the users data. Finally, other works propose reconsidering the design of social platforms by embedding in them technologies or formalisms of the Semantic Web Abel et al. [2007],Heath & Motta [2007],Buffa et al. [2008b].

aa)                        Adapting the models and tools to the usages

It is also necessary to keep in mind the social aspects of knowledge sharing, and to strive to design models fitting actual usages. For example, Sinha [2006] proposed a social and cognitive analysis of tagging where she shows that annotating a resource with several keywords requires less cognitive effort than choosing a unique category. Tagging is thus simpler since it allows picking up all the concepts first activated in the mind.

Cahier et al. [2007] distinguish between "information seeking", in which the user does not already and exactly know what she is looking for, and "information retrieval", in which the user knows exactly what she wants to find. In the information retrieval case, a high precision is expected in the results of a query. In the "information seeking" case, the user does not expect a great precision in the result, and she is refining her request along the navigation within each consequent results. Here, a high recall may be favored, and serendipity may also be appreciated, or at least not fought against. Cahier shows how ontologies, in the sense of structured vocabularies, allow for better recall by connecting terms semantically.

Zacklad [2007] proposed a comparison of the different types of classification with respect to their adequacy to the needs of the communities. To this regard, ontologies allow for a richer representation of the knowledge of a community, and the comparison of the different levels of formality of the ontologies (see section 1.3) is highly relevant to the information seeking processes. When the queries deal with information related to places or to resources having clearly and neatly defined properties (such as in the field of natural or physical sciences) formal ontologies provide for the most powerful assistance with inference mechanisms. However, in some other fields where more subjective and disputable criteria are at stake, semi-formal knowledge representations such as semiotic ontologies, Topic-Maps, or thesauruses can be interesting alternatives to help navigate within a corpus of knowledge and to represent the different point of views of the members of the community.

In a corporate context, Van Damme et al. [2008] applied coocurrence-based similarity measures and conditional probability to retrieve a hierarchy between tags used to annotate several types of contents produced within a company (messages, internal notes, etc.). They show that companies may be interested in the outputs of such methods as a management tool, which would help them in decision making (by looking at recurrent patterns within their terminology), in revealing them the emergence of new terms (technological watch and monitoring purposes), and in helping them creating new teams (by detecting proficiencies and communities of interest).

bb)                        Perspectives

Gruber [2008] differentiates collective intelligence from collected intelligence. He gives three characteristics of the current systems which collect knowledge: (1) the production of content performed by the users, (2) a synergy between users and the system, (3) increasing benefit with the size of the domain covered. In order to upgrade this type of system towards a collective intelligence, Gruber proposes adding another feature: the emergence of knowledge beyond the mere collection of each contributor's knowledge. He suggests that this fourth feature directly benefits from the integration of the technologies of the Semantic Web.

Thus, the potential of hybrid systems which exploit the benefit of both the ease of use of folksonomies and the support of the formalisms and the methods of the Semantic Web opens new perspectives for assisting knowledge exchange on the social Web. But several challenges remain. Specia & Motta [2007] showed the efficiency of combining statistical techniques with extra knowledge from the ontologies on the Semantic Web, but since the fields of knowledge that could be appropriate is potentially infinite, we need methods to efficiently select the source of information to help structuring the folksonomies. For instance, Review.com [Heath & Motta, 2007] uses that kind of technique to clearly identify whether the provided Web link is about a movie by querying the IMDB.com database, but identifying concepts dealing with the content of the reviews may be more complex and poses the problem of the selection of the sources of additional information. These issues, plus the need to find similarities between groups of tags or to match tags with elements from other ontologies could also benefit from exploiting some of the "ontology matching field's methods [Euzenat & Shvaiko, 2007], such as the identification of similar tags thanks to more or less strict string matching, or the evaluation of the similarity between two tags according to their relative positions within a graph.

The other challenges that social on-line platforms may be faced with, is the workload needed to administrate or contribute to the system. The current approaches to add semantic information to the resources exchanged in the social Web are: (1) organizing tag data a posteriori, that is, analyzing the tags and their usage Mika [2005],Specia & Motta [2007], or proposing the users organizing the tags Buffa et al. [2008b] or tagging the tags themselves Tanasescu & Streibel [2007]; (2) asking users to raise the ambiguity at tagging time Passant [2007], or to provide more detailed information when submitting content Heath & Motta [2007]. The question social Web platforms designers may ask at this moment is how much effort they can expect from their users. And this question is not simple since the social context plays a role: incentives to contribute to an enterprise weblog or to a platform of shared reviews may largely differs in the amount of effort users may put in providing precise information (workmates may be rewarded for good quality contributions), and (even more complex) in the agreement they may find when dealing with non-consensual knowledge (when commenting on a movie, different and contradictory views may emerge). One of the key to these questions may rely on a balance between top-down-style administration of the knowledge base and bottom-up-style auto-regulation. But both of these components of social Web platforms will need appropriate tools and methods (1) that seamlessly integrate within the everyday tasks of the users, without any extra burden and produce useful extra information as a side effect of the natural use of the systems; and (2) achieve an appropriate combination of precision, diversity, and representativeness of the knowledge representations supporting the activities of the "interconnected on-line communities of the social Web.