BEAM: A First Benchmark for Microdata Entity Alignment with Knowledge Graphs
Résumé
Nearly half of all web pages contain semi-structured data (RDFa, microdata, JSON-LD), yet this information remains poorly aligned with public knowledge graphs (KGs). Existing entity alignment (EA) benchmarks, typically derived from structured KGs such as DBpedia, YAGO, and Wikidata, represent idealized settings with high schema overlap and dense link structures, conditions that rarely occur in realistic cross-KG scenarios. To address this gap, we introduce BEAM, a microdata benchmark for entity alignment methods between Web Data Commons microdata and Wikidata.
Unlike prior benchmarks that rely on potentially erroneous owl:sameAs links, BEAM establishes ground-truth alignments through key-based matching (e.g., IATA codes for airports, ISBNs for books), providing reliable identity resolution. We retain much of the noise, heterogeneity, and structural sparsity of web data, rather than artificially cleaning or rebalancing the graphs. As a proof of concept, the current release covers two classes (Airport, Book) and is accompanied by a reusable pipeline to extend BEAM to further classes when suitable keys exist.
Experimental results show a substantial drop in the performance of state-of-the-art EA models on BEAM compared to curated benchmarks, revealing their limited robustness in unstructured web KGs and highlighting the importance of realistic evaluation settings. BEAM is publicly available and adheres to the FAIR principles (Findable, Accessible, Interoperable, Reusable), providing a reproducible foundation for advancing research in entity alignment between semi-structured web data and knowledge graphs.
| Origine | Fichiers éditeurs autorisés sur une archive ouverte |
|---|---|
| Licence |