Conclusions
There are some useful and widespread ontologies describing generic objects - Web resources (Dublin Core), people (VCard, FOAF), discussion forum comments (SIOC). Several bold attempts to introduce universal ontologies for E-Learning materials have had only modest success. The reasons why the knowledge about educational materials cannot be easily aggregated across organizational domains seem to be caused by decentralization and diverging goals of different educators. It is not caused by any inherent lack of common meta-attributes - on the contrary, there are plenty of them.
The article does not suggest any particular ontology. It investigates a networking-based approach to gradually introduce existing ontologies. They grow incrementally - property by property in a decentralized "bottom up" way. It involves building a mesh of collaborating tagging services, which evolve independently without any global coordination. The only requirement implied by the suggested architecture is ability for the services to call and to implement simple HTTP-based Web services (REST API). This approach is useful whenever the possible values for the properties can be enumerated (e.g. as controlled tag vocabularies), whenever they correlate with the text of the respective document. Tagging based on Support Vector Machine algorithms becomes less useful when classifying predominantly non-verbal content such as multimedia files or photos. SVM classifiers and tag suggestions are not very useful for properties having infinite range (arbitrary numeric values, phrase search, etc.).
The probabilistic model shows that every collaborating tagging service can retain its bias (i.e. it reflects the different interests and taxonomies of its primary users). Meanwhile the collaboration ensures faster training of the SVM algorithm. Some tagging services may get multi fold increase in their training speed and may become hubs, getting large number of requests from neighboring tagging services. Some others may stay largely isolated and grow their training set only from local requests - so the user community does not gain much, if all the other communities around it have very different bias, but it does not lose anything either.
The end users of the tagging systems should retain freedom to classify knowledge as they see fit (or to comply with their organization's guidelines), but they may benefit of being nudged in the commonly right direction, e.g. given prompts about possible annotations and warned about mistakes or misspellings. Users can be motivated to use tags and fill forms for micro-formats, when they see immediate benefit of doing so.
Analysis of architecture and performance shows that ensuring collaboration in the mesh of tagging services does not cause large hardware or software costs or networking overhead.
