Introduction
Wikipedia's categorization system is a rich source of descriptive information, yet it remains largely opaque to automated systems. Categories like “20th-century English male actors” function as fine-grained classes, but for a computer, they are merely unstructured text strings. The challenge lies in interpretation: determining, for instance, whether "English" refers to a person's place of birth, citizenship, or primary language. This paper presents a method to transform these "flat" category strings into structured, disambiguated property-value pairs by leveraging the structured data available in Wikidata.
Methodology: From Strings to Knowledge
The proposed approach moves beyond simple keyword matching by using Wikidata as a semantic bridge. The system "anchors" ambiguous modifiers to precise meanings through a three-stage process:
- Alignment of Modifiers and Values: The system examines the Wikidata entries for articles belonging to a specific category. It looks for matches between n-grams in the category title (e.g., "English") and the values of properties within those entries (e.g.,
Place of Birth: England). - Disambiguation: Textual modifiers are mapped to specific Wikidata topics. For example, "English" is disambiguated to the Wikidata entity for
Englandand its corresponding Wikipedia article. - Property Extraction: To resolve cases where a modifier matches multiple properties (e.g., "English" matching both
Place of BirthandCitizenship), a voting mechanism is employed. The property with the highest "article support set"—the one most frequently associated with members of that category—is selected.
The method also incorporates automatic expansion to handle temporal and geographical hierarchies, such as recognizing that a specific birth date falls within the "20th century" or that a city belongs to a larger administrative region.
Experimental Evaluation
The system was evaluated using English Wikipedia and Wikidata snapshots from June 2018. A gold standard evaluation set of 700 manually filtered Wikipedia categories was created to represent distinct classes. The proposed method ($R_{prp}$) was compared against a baseline ($B_{wcn}$) that utilizes Wikipedia's internal "X by Y" category hierarchy.
Key Results and Insights
The results demonstrate that the proposed method significantly outperforms the string-based baseline across all performance metrics.
The system achieved a remarkably high macro-precision of 0.925, indicating that the extracted property-value pairs are highly accurate. Furthermore, it successfully provided annotations for 76.5% of all Wikipedia categories, a substantial improvement over the 55.3% coverage of the baseline.
The core insight of this work is the "wisdom of the members": the meaning of a category label is best defined by the shared attributes of the entities it contains. By using Wikidata property IDs (e.g., P569 for date of birth), the system generates machine-readable, language-independent knowledge linked to a global graph.
Limitations and Future Directions
While effective, the method's recall is inherently limited by the completeness of Wikidata; categories containing articles with sparse structured data are harder to interpret. Additionally, the system occasionally selects valid but non-optimal properties (e.g., labeling a group by "place of death" if that data is more prevalent than "place of birth"). Future work will explore integrating additional knowledge sources to increase coverage and extending the method to non-English Wikipedia versions.
Summary updated