- The paper demonstrates an LTQP architecture that discovers SSSOM mapping documents as Web resources and applies schema correspondences during query execution.
- The system enables SPARQL joins across heterogeneous vocabularies and exposes mapping confidence so users can balance result precision against recall.
- The demonstration establishes feasibility but leaves scalability, conflicting or inaccurate mappings, discovery coverage, and Web-scale performance for future evaluation.
Overview
This demonstration paper by Tam, Colpaert, and Taelman (Ghent University – imec, IDLab) presents a system for performing schema alignment at query time within decentralized knowledge graph (DKG) querying. The work targets the setting of Link Traversal Query Processing (LTQP), where SPARQL queries are evaluated over RDF documents discovered by following links across independent data sources, rather than against a single centralized endpoint. In such settings, sources describe semantically overlapping entities using heterogeneous vocabularies and IRIs, so queries written against one schema fail to match data published under another. The paper demonstrates that ontology matching can be executed online — during traversal rather than as an offline preprocessing step — using mappings serialized in the Simple Standard for Sharing Ontological Mappings (SSSOM).
Problem statement
In LTQP over the Web, a client traverses HTTP links between RDF documents to accumulate triples relevant to a query. Because no central authority governs vocabulary choice, equivalent concepts frequently appear under distinct IRIs: different publishers may denote the same class or property with terms from Schema.org, Dublin Core, or bespoke ontologies. Standard LTQP engines evaluate triple patterns syntactically, so a query pattern referencing one IRI will not bind to triples asserting a semantically equivalent but lexically distinct IRI. Offline schema alignment is impractical in this environment because the set of reachable vocabularies is not known in advance and evolves with the traversal. The paper's contribution is therefore an architecture in which SSSOM mapping documents are themselves discoverable resources on the Web, fetched and applied dynamically during query execution.
Method
The approach integrates three components into the LTQP pipeline:
- Mapping discovery: alongside domain data, the traversal follows links to SSSOM mapping files, treating them as first-class resources in the link queue. Mapping documents declare correspondences between source and target IRIs with confidence scores.
- Online rewriting: when the engine encounters triples whose predicates or types have registered mappings, it rewrites incoming triple patterns (or materializes aligned triples) so that joins across differently-schematized sources become possible. Rewriting is driven by the SSSOM correspondence rules rather than a fixed equivalence relation.
- Confidence-aware evaluation: because automatically derived mappings carry uncertainty, the engine exposes mapping confidence so that results can be filtered or ranked; the authors note this trades recall for precision depending on the confidence threshold chosen.
The implementation builds on an existing LTQP engine (the Comunica framework ecosystem), extending its link-traversal and query operators with a mapping-application stage. This design keeps alignment orthogonal to query planning: the same mechanism applies regardless of the specific SPARQL query shape.
Demonstration
The paper is framed as a demonstration: a live scenario shows a user issuing a SPARQL query against a decentralized collection of sources that use divergent vocabularies. During execution, the client discovers SSSOM documents linked from the data sources, applies the correspondences on the fly, and returns results that join entities described under different schemas — results unobtainable without online alignment. The demonstration interface visualizes the traversal, the mappings applied at each step, and their confidences, making the otherwise opaque rewriting process inspectable.
The paper does not report benchmark-scale quantitative comparisons against offline-alignment baselines; its evidential weight rests on the working end-to-end scenario. Readers should accordingly treat the contribution as a feasibility demonstration of the architecture rather than a performance evaluation.
Limitations and open questions
Several constraints are acknowledged or evident from the design:
- Scalability of online rewriting: applying mappings per-triple during traversal adds overhead, and the cost behavior on large traversals with many mapping documents remains unmeasured.
- Mapping quality and provenance: the approach inherits whatever errors exist in published SSSOM files; there is no mechanism for validating or reconciling conflicting mappings discovered from multiple sources, beyond exposing confidence values.
- Discovery coverage: alignment only occurs for vocabularies for which a discoverable mapping document exists along the traversal path; the open question is how clients find relevant mappings when publishers do not link them explicitly.
- Evaluation scope: the demonstration uses curated scenarios; whether the method holds up on realistic Web-scale crawls with noisy, incomplete mappings is left untested.
Conclusion
The paper demonstrates that schema alignment need not be a preprocessing step in decentralized KG querying: by treating SSSOM mapping documents as traversable web resources and applying them during LTQP execution, queries can span heterogeneous vocabularies without prior knowledge of the source schemas. The result establishes the viability of online, confidence-aware alignment in link-traversal settings, while leaving open its scalability, robustness to imperfect mappings, and behavior at realistic scale.