While the Semantic Web requires a large amount of structured knowledge (triples) to allow machine reasoning, the acquisition of this knowledge still represents an open issue. Indeed, expressing expert knowledge in a given formalism is a tedious process. Less structured annotations such as tagging have, however, proved immensely popular, whilst existing unstructured or semi-structured collaborative knowledge bases such as Wikipedia have proven to be useful and scalable. Both processes are often regulated through social mechanisms such as wiki-like operations, recommendations, ratings, and collaborative games. To promote collaborative tagging as a means to acquire unstructured as well as structured knowledge we introduce the notion of Extreme Tagging, which describes systems which allow the tagging of resources, as well as of tags themselves and their relations. We provide a formal description of extreme tagging followed by examples and highlight the necessity of regulatory processes which can be applied to it. We also present a prototype implementation.
The process of building “a new brain for humankind” [
In practice however, building and matching ontologies, appears to be an expert
task, and difficulties related to knowledge acquisition, experienced decades ago in the
artificial intelligence community, resurface. Moreover, while ontologies seem well
suited to the description of scientific domains such as medicine and biology which are
already semi-formal and organized by categories and part-of relationships, some
communities such as geospatial scientists only accept with scepticism the exclusive
usage of ontologies to describe their domains [
In the meantime, so called Web2.0 applications, by motivating users to contribute
information, introducing fine tuned social regulation mechanism, as well as providing
friendly user interfaces, have been experiencing both phenomenal growth and
success. With the advent of Web2.0 the usage of unstructured annotations such as
tagging, spread widely. Although the relation of tagging and social interaction has not, to
our knowledge, been investigated in the literature, it seems to be the only way to
allow users to describe their own content, since the system cannot determine in advance
what this content will be. Collaborative tagging systems, by renouncing the use of
predefined vocabularies, provide a simple way for users to give their own meaning to
their own content [
Therefore, while current research is still trying to alleviate problems related to the practical use of ontologies, the semantic web may benefit from techniques used by Web2.0 applications. We believe that for the semantic web to expand faster, new semantic acquisition approaches, distinct from the centralized ontology development by experts, need to be explored. We also believe that any successful solution will use the social lever which raised the Web and Web2.0 to that level of popularity and usage.
Therefore, we introduce the notion of Extreme Tagging Systems (ETS), as an extension of collaborative tagging systems allowing the collaborative construction of knowledge bases. An ETS offers a superset of the possibilities of collaborative tagging systems in that they allow to collaboratively tag the tags themselves, as well as relations between tags. Unlike previous research on emergent semantics of collaborative tagging systems, ETS are not destined to exclusively produce hierarchical ontologies but strive to allow the expression and retrieval of multiple nuances of meaning, or semantic associations. The production of relevant semantic associations can then be automatically controlled through social network regulation mechanisms.
We first describe collaborative tagging systems. Then show the modifications introduced by extreme tagging systems, providing a formal definition. Accordingly, we explain our prototype implementation, and, before concluding, give some examples of regulation mechanisms that should be applied to the system.
The final hypergraph formed by a collaborative tagging system is defined as G with:
A
U R
T .
G
U R
Collaborative tagging systems (CTS) support multiple users in the activity of tagging,
which is marking content for future navigation, filtering or search [
Tagging systems can be represented as hypergraphs [
U {u1,. . ., u k }, R {r1, ... , rm} , and T {t1, . . . , tl} .
Collaborative tagging systems have proved extremely popular. Their strengths
consist in generating serendipity while browsing – the fact of being able to retrieve what
others have tagged in a similar way, e.g. one can retrieve everything that has been
annotated using the tag “ant” –, as well as the elaboration of desire lines – a non
constrained reflection of the user’s vocabulary – through a dataset [
To go toward “less mess”, approaches have been proposed to find groups of related
tags by using tag co-occurrence for given resources [
Furthermore, some semantic web oriented approaches attempt to extract ontologies
from collaborative tagging systems. In [
Ultimately, ontologies and tagging systems are both symbolic frameworks, and as
such they are subject to the criticism of the lack of a retrievable grounding. Indeed, in
both cases by using symbols (in a given language), the expressed concept, or
signified, remains in people's minds, and the resulting symbol networks may appear –
especially to a machine – as “free floating island[s] of reeds [with] no anchor in reality”
[
An Extreme Tagging System (or ETS) offers a superset of the possibilities of collaborative tagging systems in that it allows to collaboratively tag the tags themselves, as well as relations between them. For example, a media resource representing the closeup of a car may be tagged with “car”, “wheels” and “travel”. The tag “wheels” itself may then be tagged (possibly by a different user) with “car” and “wheel”, and the tag “car” itself could further be tagged with “vehicle” (cf. Figure 1)1. 1 Picture from Flickr user Anjuli: http://www.flickr.com/photos/49502989227@N01/56641591/
The tagging of tags is justified by the fact that a tag can have different meanings in different contexts: tagging the tags and the relations between them is used to disambiguate these contexts. For example the tag “tank” on the Flickr photo sharing service2 is used to tag military vehicles3, fish tanks4 as well as a person5. Tag tagging allows a user to explain the meaning of her or his annotations. It also reveals the multiciplicity of meanings: by tagging “tank” with “fish” and “vehicle”, the ambiguity becomes apparent and users can then decide to filter accordingly.
The operation of tagging introduces a relation which is not only functional (something has been tagged by somebody) but also a meaningful (for some reason, excluding spam, somebody tagged something with this particular tag). Indeed tagging a picture with “wheels” may relate to what the picture depicts, and a tag “travel” may relate to the origin of the picture. However, the meaning of the relation is not made explicit by the user at the moment of tagging: we believe that not having to think precisely to the relation and verbalise it as one would do in an ontology results in a smaller cognitive load for the user and is part of the appeal of tagging. In extreme tagging however, this relation itself can be tagged, later on, by any user. The operation of tagging relations between tags can naturally be expressed by triples, for example, if “_” represents the implicit relation introduced by the tagging operation itself, while “…” is used to represent any tag, relations can be: (resource, {_}, “wheels”), (resource, {_}, “travel”), (resource, {_}, ...), (resource, {“shows”}, “wheels”), (resource, {“represents”}, “wheels”), (resource, {...}, “wheels”) or (resource, {“takenduring”}, “travel”), etc. (cf. Figure 2)
Allowing users to tag the tags and the relations between them leads to the
generation of Semantic Associations. Semantic Associations are chains of relations between
one tag to another, or, in graph theoretic terms, a labelled path between two nodes.
According to the definitions of [
if they are semantically connected, i.e. there exist a path of relations between them, or semantically similar, i.e. two entities are similar if a path from the first one to another is similar to the path from the second one to another. We also call semantic annotations knowledge paths, as in this context they represent a crystallisation of the users’ knowledge. We consider that the tagging relation itself, even if implicit, qualifies as a relation in a knowledge path, while we consider that the notion of semantic similarity can be extended from subclass/superclass relations only to any similarity measure. Collaboratively tagging resources, tags and relations leads to serendipitous discovery of associations between resources and/or tags. An example path between “wheel” and “vehicle” for example would be, expressed as a list of triples <“wheel”, “vehicle”> = [(“wheel”, {“singular-of”}, “wheels”), (“wheels”, {_}, “car”), (“car”, {“is-a”}, “vehicle”)].
The ETS model is defined as a collaborative tagging system with semantic associations. Therefore ETS are extensions of the formal model for collaborative tagging systems, defined as follows:
U ,T , A, D , where A
U T
T and D
U T T
T .
We do not distinguish between the set of resources/entities T and the set of tags: all elements of T are entities, which can be “tags” or “resources”. Indeed the mapping description of each entity by a unique identifier – in practice, a URI – makes the distinction superfluous. A is the set of assignments, as in traditional CTS while D represents directional annotations of relations between entities (tags or resources). According to this definition an ETS becomes a hypergraph:
G E
U
The distinction between A and D reflects the distinction between implicit and explicit relations. An implicit relation occurs when an entity has been tagged while an explicit one appears if the relation between two entities has itself been tagged. A knowledge path is a path consisting of explicit or implicit relations between entities.
As relations between tags constitute triples, the link to RDF becomes obvious.
Indeed ETS have the same goals as those sometime advocated by RDF proponents, “to
allow anyone to say anything about anything” [
Tagopedia6 is a prototype ETS built on top of the Facebook platform. Facebook is a social network web application providing a developer framework allowing the creation of applications which interact with core host features such as profile management and login. As any collaborative tagging system Tagopedia allows to tag resources, represented by URIs (cf. Figure 3).
When clicking on a tag however, the user is asked to tag the chosen relation or to enforce an already existing relation by selecting it (cf. Figure 4). The application then moves to the target tag, showing the linked tags and resources and allowing to define new relations. The user may also choose not to tag the relation and directly reach the target, keeping it implicit. 6 Available at http://apps.facebook.com/tagopedia/ (a Facebook account is required). The name Tagopedia, proposed independently by the authors, has already been proposed in 2005 by Russell Beattie in a blog post, for a related application (http://www.russellbeattie.com/notebook/1008277.html). (grail, {topic-of}, “http://www.imdb.com/title/tt0071853/”), (“http://www.imdb.com/title/tt0071853/”, {has-title}, “Monty Python and the Holy Grail”), (“Monty Python and the Holy Grail”, {directed-by}, “Terry Gilliam”)] sa2: [(“John Boorman”, {is-a}, film-director), (film-director, {includes}, “Terry Gilliam”)] sa3: [(“John Boorman”, {is}, British), (British, {nationality-of}, “Terry Gilliam”)] 5
In ETS, semantics are related to the users’ activity and input. The operations involved in a user’s activity can be classified as: annotation, navigation and control. At each level there is a need for incitation, a means to motivate the user to use the system. As a result of these three operations, tags are created and annotated collaboratively and unconstrained semantics emerge (cf. Figure 5). In this section, we describe each activity in turn as well as the corresponding motivation mechanism:
Through annotation users are given the opportunity to create their personal knowledge base. Indeed, instead of tagging resources at their hosting websites, building unrelated islands of tags, they can relate all their resources with their own meaning. Moreover, they can access, for the same resource, tags from other users, and decide to explore their meaning by navigating to them.
Through navigation, users build or enforce semantic associations. Indeed, by exploring a tag which tags a tag, either the user is looking for an explanation of this tag, or she already knows the relation. We assume that she knows the relation if a) she tagged it before, or b) she chooses to tag it when asked to do so. As previously mentioned, navigation does not happen between a tag and another tag without presenting the relation, which the user can choose to tag or not. The motivation of this additional step is to constrain the meanings obtained. Paths which have been explored and validated, are recorded and displayed the next time a request is made to find the paths from one node to another.
Finally control mechanisms are necessary in order for the system to evolve, some of these control mechanisms can be: 1) total control over ones annotations: the annotations added by a user can be modified or deleted by her. 2) appreciation and depreciation of tags: a user can rank a tag (+ or – only). If the total ranking goes below a given threshold, the tag becomes “private” and does not appear in public searches any more. A similar method is already used by commercial websites7. 3) questions to author: Facebook, just as other social networks provides the notion of “friends”, or “contacts”, i.e. users which acknowledged a mutual relationship. If a user does not understand a tagging made by one of her related users a quick means is provided to send him or her a message to ask for an explanation, i.e. the tagging of this particular relation. The requester is notified as soon as the explanation has been given.
It is assumed that each user is interested in sharing his or her vision of the world and in discovering other ways of perceiving it. To the first interest corresponds the annotation activity as well as some control activities number 1) and 2). As a further incitation annotating increases the user’s ranking, in a similar way as internet forums display titles according to the number of posts (often quite imaginative, for example using a graduated scale going from rookie, to half-god or absolute guru). In parallel, an increase in status can be achieved through navigation only, in a similar manner to some multiplayer computer games which increase the avatar’s status by providing titles according to the percentage of the virtual map explored. Indeed, These two ways of using the system, annotating and creating, combine when a navigator – i.e. a user who mostly navigates the system, comparable to a Wikipedia reader rather than to an editor – earns creation points by completing paths and creators earn more points if the paths they created are navigated (i.e. if they make sense). Further incitation may involve visualisation of the number of elements created, as well as graph presentation of the paths explored. 6
The benefits of pushing tagging to an extreme are the ease with which knowledge is acquired, as well as the comprehensiveness of the resulting KB. Possible caveats, which we believe can be solved by collaborative means, include the difficulty to assess the relevance of the resulting knowledge in a given context. Tagopedia is a first prototype of an Extreme Tagging System and we are waiting to obtain a larger knowledge base to attempt a serious evaluation. However, we used the prototype in a limited environment composed of 5 users, and, from the amount of serendipitous meaning collected, were already convinced of the interest of the system. We are planning to release it to the Facebook community in the following months and explore the
aforementioned control mechanisms through it. We are also working on an RDF export mechanism as well as on the integration of a SPARQL query engine. We also plan to import large amounts of tags from Wikipedia and other websites, using links inside the pages or other structured information in order to populate the knowledge base.
Acknowledgments. The authors are grateful to the anonymous reviewers for their precious comments. They would also like to thank Yiwen Wang, Kaixuan Wang and Fadi Badra, who contributed ideas and energy in the early stages of this project. Further thanks are extended to Sean Bechhofer, Enrico Motta, and John Domingue, who contributed, during SSSW07 and later on, to the success of this enterprise, and to Lyndon Nixon for proofreading this paper.