Introduction
Building formal ontologies—structured frameworks of concepts and their logical relationships—is a labor-intensive process requiring both domain expertise and mastery of formal logic. While Wikipedia provides a massive, human-curated category hierarchy, its structure is informally organized and often underspecified. This paper introduces an automated machine-learning framework designed to transform Wikipedia’s messy category hierarchy into structured, domain-specific ontologies. By treating extraction as a series of binary classification tasks, the authors "bootstrap" ontology skeletons by identifying relevant concepts, distinguishing classes from individuals, and defining formal hierarchical relations.
Methodology: The Uniform Approach
The proposed system employs a sequential three-step process, framing each as a binary classification problem solved using Support Vector Machines (SVM):
- Domain Selection: Starting from a root category (e.g., Computing), the system performs a breadth-first traversal. A classifier determines category relevance based on its depth from the root and the lexical similarity of its title to its parent.
- Node Classification: Relevant categories are classified as either a Class (a set of entities, e.g., Programming Languages) or an Individual (a specific entity, e.g., Python). The authors found that simple character-level suffixes (1-3 letters) were highly effective features for this task.
- Relation Classification: Links between categories are classified into formal types:
subclass_of,instance_of,part_of, orrelated_to. This step utilizes WordNet to calculate semantic similarity and hyponymy relationships between titles.
Key Insights: The "9-Hop" Rule
A critical observation in the study is that relevance drops sharply as one moves further from the root category. For example, a user can navigate from Computing to Buddhism in China in just nine hops. The authors demonstrated that while simple depth-based pruning is common, their automated classifier maintains high accuracy across varying depths, significantly outperforming static depth limits.
Experimental Results
The framework was evaluated against "gold standard" datasets like the ACM Computing Classification System and manually annotated data for the Computing and Music domains.
- Domain Selection: Achieved 92% accuracy in the Computing domain, maintaining performance where depth-based methods fail.
- Node Classification: This was the highest-performing module, reaching 95% accuracy in distinguishing classes from individuals using simple linguistic features.
- Relation Accuracy: The system identified
subclass_ofrelations with 80% accuracy. While performance oninstance_ofwas comparable to existing tools like WikiTaxonomy, the system showed superior overall weighted F1 scores on ground-truth data.
Limitations and Future Directions
The primary challenge remains the identification of part_of relations (meronomy), where accuracy hovered between 45% and 59%. Category titles often lack the semantic context needed to distinguish "part of" from "related to." Future research will explore incorporating the full text of Wikipedia articles and implementing "joint classification" to allow the model to leverage hierarchical logic to improve its predictions.
Summary updated