Hi!
I'm Ruben Taelman, a Professor at IDLab, imec, Ghent University,
with a focus on Web technologies, decentralization, Linked Data publishing, and querying.

My goal is to make data accessible for everyone by providing
intelligent infrastructure and algorithms for data publication and retrieval.

To support my research, I co-edit specifications such as SPARQL, and develop various open source JavaScript libraries such as streaming RDF parsers and the Comunica engine to query Linked Data on the Web.
As this website itself contains Linked Data, you can query it live with Comunica.

Have a look at my publications or projects
and contact me if any of those topics interest you.

Latest blog posts

• Did AI Crawlers Kill SPARQL Federation?
  Public Knowledge Graph infrastructure is degrading due to AI crawlers.
  13 April 2026
  

  RDF provides the basis for distributing Knowledge Graphs (KGs) across different locations, which is useful when KGs cover different data domains with varying purposes, are managed by different teams and organizations, or are exposed by different access policies. One of the most popular ways of publishing a KG is through a SPARQL endpoint, which offers queryable access. When multiple of these KGs need to be integrated, techniques such as SPARQL federation can be used. While many KGs have been available as public SPARQL endpoints, their openness is currently being challenged by the huge load that is placed on them by modern AI crawlers that power LLMs. Recently, public SPARQL endpoints have started putting in place stricter usage restrictions to avoid going down under this increased server load. While these restrictions limit the range of SPARQL queries that can be executed over them, it becomes especially problematic for SPARQL federated queries, which often involves sending multiple smaller queries to endpoints in a short timeframe.

• The cost of modularity in SPARQL
  How much do modularity and decentralization conflict with centralized speed?
  22 April 2025
  

  The JavaScript-based Comunica SPARQL query engine is designed for querying over decentralized environments, e.g. through federation and link traversal. In addition, it makes use of a modular architecture to achieve high flexibility for developers. To determine the impact of these design decisions, and to be able to put the base level performance of Comunica in perspective, I compared its performance to state-of-the-art centralized SPARQL engines in terms of querying over centralized Knowledge Graphs. Results show that Comunica can closely match the performance of state-of-the-art SPARQL query engines, despite having vastly different optimization criteria.

Highlighted publications

1. Conference Link Traversal Query Processing over Decentralized Environments with Structural Assumptions
   

   1. Ruben Taelman0
   2. Ruben Verborgh1

   In Proceedings of the 22nd International Semantic Web Conference (2023) To counter societal and economic problems caused by data silos on the Web, efforts such as Solid strive to reclaim private data by storing it in permissioned documents over a large number of personal vaults across the Web. Building applications on top of such a decentralized Knowledge Graph involves significant technical challenges: centralized aggregation prior to query processing is excluded for legal reasons, and current federated querying techniques cannot handle this large scale of distribution at the expected performance. We propose an extension to Link Traversal Query Processing (LTQP) that incorporates structural properties within decentralized environments to tackle their unprecedented scale. In this article, we analyze the structural properties of the Solid decentralization ecosystem that are relevant for query execution, we introduce novel LTQP algorithms leveraging these structural properties, and evaluate their effectiveness. Our experiments indicate that these new algorithms obtain accurate results in the order of seconds, which existing algorithms cannot achieve. This work reveals that a traversal-based querying method using structural assumptions can be effective for large-scale decentralization, but that advances are needed in the area of query planning for LTQP to handle more complex queries. These insights open the door to query-driven decentralized applications, in which declarative queries shield developers from the inherent complexity of a decentralized landscape. 2023
   More
2. Conference Comunica: a Modular SPARQL Query Engine for the Web
   

   1. Ruben Taelman0
   2. Joachim Van Herwegen1
   3. Miel Vander Sande2
   4. Ruben Verborgh3

   In Proceedings of the 17th International Semantic Web Conference (2018) Query evaluation over Linked Data sources has become a complex story, given the multitude of algorithms and techniques for single- and multi-source querying, as well as the heterogeneity of Web interfaces through which data is published online. Today’s query processors are insufficiently adaptable to test multiple query engine aspects in combination, such as evaluating the performance of a certain join algorithm over a federation of heterogeneous interfaces. The Semantic Web research community is in need of a flexible query engine that allows plugging in new components such as different algorithms, new or experimental SPARQL features, and support for new Web interfaces. We designed and developed a Web-friendly and modular meta query engine called Comunica that meets these specifications. In this article, we introduce this query engine and explain the architectural choices behind its design. We show how its modular nature makes it an ideal research platform for investigating new kinds of Linked Data interfaces and querying algorithms. Comunica facilitates the development, testing, and evaluation of new query processing capabilities, both in isolation and in combination with others. 2018
   More
3. Journal Triple Storage for Random-Access Versioned Querying of RDF Archives
   

   1. Ruben Taelman0
   2. Miel Vander Sande1
   3. Joachim Van Herwegen2
   4. Erik Mannens3
   5. Ruben Verborgh4

   In Journal of Web Semantics (2018) When publishing Linked Open Datasets on the Web, most attention is typically directed to their latest version. Nevertheless, useful information is present in or between previous versions. In order to exploit this historical information in dataset analysis, we can maintain history in RDF archives. Existing approaches either require much storage space, or they expose an insufficiently expressive or efficient interface with respect to querying demands. In this article, we introduce an RDF archive indexing technique that is able to store datasets with a low storage overhead, by compressing consecutive versions and adding metadata for reducing lookup times. We introduce algorithms based on this technique for efficiently evaluating queries at a certain version, between any two versions, and for versions. Using the BEAR RDF archiving benchmark, we evaluate our implementation, called OSTRICH. Results show that OSTRICH introduces a new trade-off regarding storage space, ingestion time, and querying efficiency. By processing and storing more metadata during ingestion time, it significantly lowers the average lookup time for versioning queries. OSTRICH performs better for many smaller dataset versions than for few larger dataset versions. Furthermore, it enables efficient offsets in query result streams, which facilitates random access in results. Our storage technique reduces query evaluation time for versioned queries through a preprocessing step during ingestion, which only in some cases increases storage space when compared to other approaches. This allows data owners to store and query multiple versions of their dataset efficiently, lowering the barrier to historical dataset publication and analysis. 2018
   More

Latest publications

1. Blue Sky Symbiotic Knowledge Graphs: A Vision for Semantic Brain-Computer Interfaces
   

   1. Ruben Taelman0

   In Proceedings of the 25th International Semantic Web Conference (2026) A primary goal of the Semantic Web was to enable autonomous agents to take up tasks on behalf of humans. Now, 25 years after this initial vision, this goal could be considered partially realized. While these agents can be personal, and associated to specific humans, they are not directly integrated within human cognition. In other words, these intelligent agents are external to the human body and brain, and they require indirect communication through human sensory and motor functions. As such, the Semantic Web vision has set out a course to assist the evolution of human knowledge by creating intelligence in external agents, but the intelligence of individual humans remains untouched, or is even degrading. As such, to avoid human intelligence falling behind on artificial intelligence, there is a need to augment our human intelligence. Thanks to the Semantic Web technology stack, we already have a foundation to make machines understand data, which are often captured in the form of Knowledge Graphs. In this paper, we argue that this stack offer a basis to also make knowledge interoperable across humans and can augment human intelligence through Semantic Brain-Computer Interfaces. We present an interdisciplinary research roadmap for achieving this vision, which spans Augmented Cognition, Cognitive Neuroscience, Brain-Computer Interfaces, and Neuro-symbolic AI. This will enable the existence of Symbiotic Knowledge Graphs, which consist of the integration of human knowledge stored in the brain, and external knowledge stored on the Web. Concretely, we propose a research vision where artificial intelligence and human intelligence can co-evolve in a symbiotic manner, without one replacing the other. 2026
   More
2. Conference Does SPARQL federation work in the real world? A case study over large biological SPARQL endpoints
   

   1. Elias Crum0
   2. Bryan-Elliott Tam1
   3. Jonni Hanski2
   4. Ana-Claudia Sima3
   5. Tarcisio Mendes de Farias4
   6. Jerven Bolleman5
   7. Ruben Taelman6

   In Proceedings of the 25th International Semantic Web Conference (2026) Federated SPARQL querying theoretically enables data integration over distributed knowledge sources without requiring centralized data replication. In practice, federated SPARQL querying can be performed manually (i.e., the user provides SERVICE clauses that specify targets for operations) or algorithmically (i.e., an adaptive approach for automatic source assignment for operations). To date, most evaluations of SPARQL federation, both manual and algorithmic, are conducted under controlled benchmark conditions and say relatively little about how federation behaves against public endpoints in everyday use. We present a longitudinal study that executes 67 real-world federated SPARQL queries that target over 20 widely-used, large public SPARQL endpoints, using both manual and algorithmic federation approaches, at four different time points between Spring 2025 and Spring 2026. The federated queries were obtained from users of these endpoints and the majority represent complex, biologically relevant questions. We found that the current state-of-the-art algorithmic federation approaches perform substantially worse than manual approaches across all time points, and in most cases encounter errors when executing the queries tested. We also found that query execution success decreased over the time points tested for both manual and automatic methods. These findings are framed from both user and endpoint-maintainer perspectives to encourage collaborative, community-driven improvements for users and data source maintainers alike. 2026
   More
3. Conference SolidSessionBench: Realistic Query Sequences for User-Oriented Decentralized Environments
   

   1. Ruben Eschauzier0
   2. Ruben Taelman1

   In Proceedings of the 25th International Semantic Web Conference (2026) To build user-facing applications on decentralized data, clientside decentralized query engines offer a natural integration point. However, optimizing queries in these environments is challenging due to a lack of prior knowledge regarding data distribution and availability. By repeatedly interacting with the same engine from the same user, applications allow the engine to overcome this knowledge gap by exploiting patterns in previously issued queries. However, evaluating these techniques accurately is difficult because public query logs are scarce, and existing benchmarks rely on static, template-based structures that fail to capture realistic behavioral variability. To fill this gap, we introduce an evidence-based simulation of client query usage patterns as a reusable benchmark enabling realistic and reproducible evaluation of client-side optimization techniques. We demonstrate its utility by evaluating clientside caching strategies for link traversal query processing. By comparing our sequence-based approach against traditional, non-sequence-based workloads, we highlight how realistic sequence simulation heavily influences performance evaluations. Ultimately, our benchmark reveals important performance dynamics, such as the impact of session switching and query refinement, that traditional workloads do not. We conclude that modeling realistic user interactions provides an essential foundation for accurately evaluating client-side optimization techniques in decentralized environments. 2026
   More
4. Workshop Strengthening LargeRDFBench for Interoperable Federated SPARQL Evaluation
   

   1. Bryan-Elliott Tam0
   2. Saleem Muhammad1
   3. Ruben Taelman2

   In Proceedings of the 2nd International Workshop on Data Management for Knowledge Graphs (2026) LargeRDFBench is one of the most comprehensive benchmarks for evaluating federated SPARQL query engines, combining a large collection of real, interlinked datasets with a rich query suite that has made it a reference point for the community. As the benchmark continues to be widely adopted and the range of conformant engines grows, the interoperability of its artifacts, and thus the reproducibility of comparisons across engines, becomes ever more important. Yet several of its data dumps violate the RDF specifications, and its expected query results are distributed in an ad hoc format and contain discrepancies with respect to the source datasets. We systematically identify and categorize these data-quality issues and repair them, producing standards-conformant serializations of every affected dataset; we re-encode the benchmark’s expected results in the W3C SPARQL 1.1 Query Results JSON Format and correct their discrepancies. We contribute a standards-compliant edition of LargeRDFBench, produced by a reproducible cleaning pipeline, together with its expected query results in a standard, machine-verifiable format. Every dataset now parses under strict, specification-compliant RDF parsers, and the expected results are machine-verifiable through a standard format, extending the benchmark’s reach to the full range of conformant engines while staying faithful to the original data. Running the modernized benchmark end-to-end, we validate and correct its expected results, and find that the hosting engine can change the answers the benchmark returns, so that results obtained on it are not always reproducible across engines. We further perform a preliminary comparison, not previously explored, of ASK- and COUNT-based source selection in the FedX algorithm. This work strengthens an already valuable community resource by aligning its artifacts with the RDF standards. In doing so, we broaden the set of engines that can be fairly and reproducibly compared, and open the question of how the server-side SPARQL engines that host the data influence algorithmic choices and, through them, federated query performance. 2026
   More
5. Workshop Translating RDF 1.2 Graphs to Property Graphs
   

   1. Maxime Jakubowski0
   2. Dominik Tomaszuk1
   3. Daniel Fernandez Alvarez2
   4. Jose Emilio Labra Gayo3
   5. Ruben Taelman4
   6. Ruben Dedecker5
   7. Katja Hose6

   In Proceedings of the 2nd International Workshop on Data Management for Knowledge Graphs (2026) RDF 1.2 introduces statement-level constructs, such as triple terms, reifiers, and descriptions, that pose new challenges for transforming RDF data into labeled property graphs (LPGs). This work presents and compares three algorithms for translating RDF 1.2 data into LPGs, each interpreting a different amount of the RDF 1.2 reification vocabulary. Algorithm Structural interprets no vocabulary and losslessly supports arbitrary RDF 1.2 graphs by representing triple terms explicitly as nodes; Direct interprets rdf:reifies to recover annotated edges while preserving reifiers as graph objects; and Fold additionally folds reifier descriptions into edge properties, resulting in a more concise, albeit lossy, LPG representation. We define the three translations, analyze which class of RDF 1.2 graphs each preserves, and experimentally evaluate their implementations. The experiments confirm the predicted trade-offs: all variants scale linearly on the tested workloads, the generality of Structural has no runtime penalty, and folding descriptions into edge records can significantly reduce the number of edges on description-heavy data. 2026
   More