Abstract
Adaptive hypermedia systems rely on domain models to personalize content, yet manual construction of these models is often a bottleneck. This paper presents an automated pipeline for generating hierarchical domain models from document corpora using segmentation, clustering, and subsumption analysis. The resulting models are visualized via Treemaps, enabling real-time tracking of user navigation and knowledge states.
Introduction
Adaptive hypermedia aims to tailor information delivery to individual user needs. Central to this is the Domain Model—a structured map of available knowledge. Traditionally, these models are hand-crafted by experts, a process that is labor-intensive and difficult to scale. We address this challenge by proposing a method to automatically extract a structured hierarchy of concepts directly from raw text, facilitating the creation of scalable, intelligent hypermedia environments.
Methodology
The proposed pipeline transforms a flat document collection into a hierarchical domain model through four primary stages:
- Segmentation: Documents are partitioned into "homogeneous paragraphs" using the TextTiling algorithm to isolate specific sub-topics.
- Clustering: These segments are grouped into clusters representing "concepts." Unlike keyword-centric approaches, a concept is defined here as a set of significant terms ranked by TF-IDF.
- Hierarchy Generation: A subsumption measure is applied to identify generalization/specialization relationships between concepts. If keywords for Concept A consistently appear within the context of Concept B, Concept A is inferred as a sub-concept.
- Visualization: The domain model is rendered as a Treemap, where nested rectangles represent the hierarchy. This 2D space-filling approach allows for intuitive mapping of user interactions.
Experimental Evaluation
The approach was validated using two datasets:
- Looksmart.com Subset: A corpus of ~100 documents on Artificial Intelligence. This provided a "ground truth" human-designed hierarchy for comparison.
- French Museum Website: A real-world application used to evaluate the model's utility in tracking live user navigation.
Results and Discussion
The automated system successfully generated a hierarchy for the Looksmart data that was more granular than the human-designed original. While the manual model used five broad categories, the algorithm identified deeper sub-categories, such as "Ontologies" and "KDD" within "Knowledge Representation."
Furthermore, the Treemap visualization proved effective for user modeling. By overlaying navigation paths onto the domain map, the system can visualize a user's "knowledge state" in real-time.
Conclusion
Automating domain model construction removes a significant barrier to the deployment of adaptive hypermedia. By treating concepts as sets of keywords and utilizing spatial visualizations, we provide a scalable framework for personalization. Future work will focus on quantitative evaluations of user learning outcomes and optimizing the computational efficiency of the clustering process for larger datasets.
Summary updated