Introduction
The Linked Open Data (LOD) cloud represents a vast, decentralized knowledge graph. However, because it is populated by autonomous publishers, it suffers from significant schema heterogeneity. While alignment of entities and classes has seen substantial progress, the alignment of properties—the predicates defining relationships between entities—remains a critical bottleneck. This paper investigates the nature of property pairs across diverse datasets and evaluates the effectiveness of various alignment strategies.
Taxonomy of Property Pairs
To better understand the challenges of property alignment, the authors propose a taxonomy based on the difficulty of identification:
- Simple Property Pairs: These exhibit high syntactic similarity (e.g.,
birthPlacevs.placeOfBirth). - Opaque Property Pairs:
- Synonymous: These use different terminology for the same semantic meaning (e.g.,
occupationvs.profession) and typically require lexical resources like dictionaries for identification.
- Synonymous: These use different terminology for the same semantic meaning (e.g.,
- Complex: These pairs lack obvious naming or lexical links; their similarity is only apparent through data usage or domain/range analysis (e.g.,
battlevs.participated in conflict).
Methodology
The study evaluates three primary alignment techniques:
- String Similarity: Employs Jaro-Winkler and Dice coefficients to measure literal character overlap.
- Lexical/Dictionary: Utilizes WordNet similarity measures to find synonyms.
- Extension-based Alignment: A data-driven approach that identifies equivalent properties by analyzing shared subject-object pairs (extensions) across datasets, leveraging existing entity co-reference links.
Experimental Setup
The analysis was conducted on object-type properties from several prominent LOD datasets, including DBpedia, Freebase, LinkedMDB, and DBLP. For each dataset pair, a sample of 5,000 instances was used. Ground truth was established through manual identification and external verification.
Results and Discussion
Property Distribution and the "Exact Match" Trap
The analysis reveals that while "Simple" property pairs constitute the majority (approximately 63%), they are rarely identical.
As shown in the data distribution, even within the "Simple" category, "Exact" matches are infrequent. This highlights the inadequacy of basic string equality checks for property alignment. Furthermore, a significant portion of the data (29% Complex, 8% Synonymous) is entirely opaque to simple naming conventions.
Performance of Alignment Methods
The experimental results underscore the limitations of traditional syntactic and lexical methods.
As illustrated in the performance metrics, string-matching algorithms (Jaro-Winkler and Dice) achieve zero recall for synonymous and complex property pairs. They are fundamentally limited to the literal characters of property names. Surprisingly, WordNet-based approaches also underperformed, often failing to find technical or compound terms (e.g., placeOfBirth) in standard lexical databases.
In contrast, the extension-based algorithm proved to be the most robust, achieving an average precision of 0.84 and recall of 0.71. By focusing on how properties are used in practice rather than how they are labeled, this method successfully identified complex alignments and filtered out coincidental matches.
Conclusion
This preliminary analysis demonstrates that property alignment in the LOD cloud requires moving beyond schema-level labels. The prevalence of non-exact simple matches and opaque property pairs necessitates data-driven approaches. Extension-based alignment offers a promising path forward, leveraging instance-level evidence to overcome the limitations of syntactic and lexical methods.
Summary updated