Abstract
Massive Open Online Courses (MOOCs) often suffer from "shallow" categorization, where individual learning resources inherit broad labels from their parent courses. This paper details a semantic enrichment pipeline applied to the DAtaset of Joint Educational Entities (DAJEE). By mapping resource transcripts to DBpedia entities and applying graph-based filtering algorithms, we expanded the dataset's classification from 46 to over 2,000 categories. Our results demonstrate a significant increase in hierarchy depth and categorization precision, facilitating more granular resource discovery.
Introduction
The proliferation of MOOCs has made tens of thousands of educational resources available online. However, discovering specific topics within these courses remains a challenge. Currently, most platforms use course-level labeling; for instance, a video on "Modular Arithmetic" within a Cryptography course might only be tagged as "Computer Science." This lack of granularity hinders effective search and recommendation. The DAtaset of Joint Educational Entities (DAJEE), containing over 20,000 Coursera resources, initially faced this limitation with only 46 broad categories. This research aims to move toward content-based labeling by extracting specific topics directly from resource transcripts.
Methodology
The proposed pipeline for semantic enrichment consists of four primary stages:
- Semantic Interlinking: Resource transcripts are processed to identify and map key terms to structured entities in DBpedia.
- Graph Construction: For each identified entity, the system extracts its associated category hierarchy from the DBpedia knowledge base to form a complex category graph.
- Noise Reduction and Filtering: DBpedia's interconnected nature often introduces "semantic noise"—irrelevant links that can misclassify technical content into unrelated fields like history or belief. To mitigate this, we employed Dijkstra’s Algorithm to identify the shortest relevant paths and Spreading Activation (SA) to propagate "energy" from specific entities up through the hierarchy.
- Weighting: Edge weights were applied to prioritize descriptive parent categories over tangential ones, ensuring that the most relevant nodes receive the highest activation.
Results and Discussion
The application of the semantic enrichment pipeline resulted in a dramatic expansion of the DAJEE dataset's metadata. The total number of categories grew from 46 to 2,123, providing a much finer level of granularity for resource classification.
Beyond the raw number of categories, the structure of the hierarchy was significantly improved. The number of top-level categories increased from 10 to 29, allowing for better separation of broad fields (e.g., distinguishing "Arts" from "Humanities"). Furthermore, the maximum depth of the category tree increased from 2 levels to 11 levels, enabling the system to distinguish between highly specific sub-fields.
A case study of the "Generic birthday attack" resource highlights the precision of the filtering algorithms. Despite potential links to "Business" or "History" in the raw DBpedia graph, the Spreading Activation algorithm correctly identified "Mathematics" as the primary category, accounting for 57% of the total activation, effectively filtering out unrelated semantic branches.
Conclusion
This research demonstrates that automated semantic enrichment can successfully transform shallow educational metadata into a deep, precise hierarchy. By moving from course-level to content-level labeling, we have significantly enhanced the discovery potential of the DAJEE dataset. Future work will focus on evaluating the computational scalability of these graph algorithms and conducting human-centric evaluations to measure the impact on recommendation system performance.
Summary updated