The DAtaset of Joint Educational Entities (DAJEE) is a repository which hosts more than 20,000 educational resources crawled from the MOOC platform Coursera. The resources are divided per category according to the MOOC categorization on Coursera, which is, however, very shallow. This contribution focuses on a more meaningful categorization of the resources in DAJEE, tailored to their content. To achieve such goal, our approach enriches the resources in DAJEE with semantic entities by applying state-of-the-art semantic techniques. The result is a significant improvement of the categorization of the resources in DAJEE than the previous version.
The DAtaset of Joint Educational Entities (DAJEE) provides a huge variety of
learning resources coming from the popular Massive Open Online Course (MOOC)
platform Coursera1. The novelty of DAJEE is the contextualization of the delivery of
learning resources in lessons and courses [
In Coursera, MOOCs are grouped in 10 top-level categories, each one with a different number of sub-categories, while the resources do not have any categorization. In DAJEE, the category of a resource is inherited from the related MOOC (reproducing the same categorization stated by authors). However, Coursera offers only a shallow categorization with at most two levels of categories, with some top-level categories without any sub-category. Hence, hundreds of resources are grouped into just one category with no additional diversification among them. Also, educational resources
may belong to categories that differ from the ones of their course. This limited categorization of the resources is currently replicated in DAJEE.
We propose a method to enhance DAJEE with a more-in-depth categorization of
the resources based on their content, instead of their course. The proposed method
firstly exploits Semantic Web techniques and data offered by DBpedia for interlinking
the content of a resource with semantic entities in DBpedia [
Categories in DAJEE
The shallow category structure of Coursera is linked to the courses, not to the resources. Instead, DBpedia offers a much more detailed category structure that can directly categorize the resources. As an example, the category Math and Logic in Coursera has no subcategories, while in DBpedia the equivalent category (named Mathematics) has several subcategories with different levels of depth.
In DBpedia, semantic entities have many categories. For achieving our goal, resources in DAJEE should be associated with categories that are tailored to the semantic entities extracted from their transcripts. To keep trace of the categorization of a semantic entity, a sub-graph of the DBpedia category structure describes an entity. Such sub-graph starts with the source category and ends with the categories stated in the DBpedia page of the entity. Since each transcript has a number of entities, each one with a sub-graph, the main problem is how to properly merge these sub-graphs for a correct and meaningful categorization of a resource or a text.
A simple merge of the category-graphs of the entities associated with the resource
is not an efficient solution. DBpedia entities can present a very wide set of categories
and some of them may be poorly or not at all related to the resource. Let us consider
the resource Generic birthday attack from course Cryptography I as a reference. The
DBpedia entity dbr:Cryptographic_hash_function2 has been found in its transcript,
and Figure 1 shows its category graph. The graph includes Science, Mathematics, and
History, while other categories are descendant of Business and Belief which seem to
be unrelated to the resource. For a more effective categorization of the resource, we
suggest a novel method for filtering DBpedia categories based on the Spreading
Activation (SA) [
steps: 1st step) Dijkstra's algorithm finds the shortest paths among the categories stated in the DBpedia page of the entity, if any; 2nd step) the edges receive a weight which is 1 if they are part of those paths; 3rd step) the weight is set to 1 for all the edges that connect the categories on the DBpedia page with the root of the graph.
For deducing the most important top-level categories for an entity, the activation is spread throughout the category graph opposite to the edges direction (from "child" to "parent"). The activation for a category j is regulated as follows: =
!∈!"#$!"#%&#'((!) !" = 10,, ℎ() > 0 where wij is the weight of an edge from category i to j. The algorithm stops when the OutgoingActivation is 0. Figure 2 (left) reports the activations of the top-level categories, finding that Generic birthday attack is mostly about Mathematics (57% of the activations). The same process can also filter the sub-categories of the most frequent top-level categories, removing edges with weight 0 and sub-categories with no activation from the graph of the resource3. Interestingly, the resource Alpha Beta Pruning is identified as Mathematics as well, but it is different from Generic birthday attack. Our method shows that the category graph for Alpha Beta Pruning4 is focused on graph theory and algorithms, while Generic birthday attack presents many connections to mathematical analysis and algebra. So, we can further distinguish resources 3 The final category graph for the resource Generic birthday attack is available at: http://virtuosa.pa2.itd.cnr.it/iswc17/generic_birthday_attack.png 4 The resulting graph is available at:
http://virtuosa.pa2.itd.cnr.it/iswc17/alpha_beta_pruning_graph.png belonging to a same top-level category, like Generic birthday attack and Alpha Beta Pruning. Applying our methodology, a more meaningful categorization tailored on the resource content is now included in DAJEE.