Publications
Refereed Journal Articles
-
2025Dang, M.-H., Pham, T. H. T., Molli, P., Skaf-Molli, H., & Gaignard, A. (2025). LLM4SchemaOrg: Generating Schema.org Markups with Large Language Models. Semantic Web Journal.
@article{dang2025llm4schema, title = {{LLM4SchemaOrg}: Generating {Schema.org} Markups with Large Language Models}, author = {Dang, Minh-Hoang and Pham, Thi Hoang Thi and Molli, Pascal and Skaf-Molli, Hala and Gaignard, Alban}, journal = {Semantic Web Journal}, year = {2025}, month = jun, doi = {10.1177/22104968251382172} }Integrating Schema.org markup into web pages has resulted in the generation of billions of RDF triples. However, around 75% of web pages still lack this critical markup. Large language models (LLMs) present a promising solution by automatically generating the missing Schema.org markup. Despite this potential, there is currently no benchmark to evaluate the markup quality produced by LLMs. This article introduces LLM4Schema.org, an innovative approach for assessing the performance of LLMs in generating Schema.org markup. Unlike traditional methods, LLM4Schema.org does not require a predefined ground truth. Instead, it compares the quality of LLM-generated markup against human-generated markup. Our findings reveal that 40%–50% of the markup produced by GPT-3.5 and GPT-4 is invalid, non-factual, or non-compliant with the Schema.org ontology. These errors underscore the limitations of LLMs …
Refereed Conference Proceedings
-
2024Aimonier-Davat, J., Dang, M.-H., Molli, P., Nédelec, B., & Skaf-Molli, H. (2024). Fediscount: Shopping Online at a Federated Store Using FedUP as SPARQL Federation Engine. Companion Proceedings of the ACM Web Conference 2024, 1059–1062.
@inproceedings{aimonier2024fediscount, title = {Fediscount: Shopping Online at a Federated Store Using {FedUP} as {SPARQL} Federation Engine}, author = {Aimonier-Davat, Julien and Dang, Minh-Hoang and Molli, Pascal and N{\'e}delec, Brice and Skaf-Molli, Hala}, booktitle = {Companion Proceedings of the ACM Web Conference 2024}, pages = {1059--1062}, year = {2024}, month = may, doi = {10.1145/3589335.3651249} }Processing SPARQL queries over large federations of SPARQL endpoints is essential for maintaining the Semantic Web decentralized. However, existing federation engines struggle to query more than a dozen of endpoints. We recently proposed FedUP, a new type of federation engine based on unions-over-joins query plans that outperforms state-of-the-art federation engines by orders of magnitude on large federations. This demonstration paper introduces Fediscount, a federated online shopping application based on the FedShop benchmark, illustrating the capabilities of FedUP. The application is based on standard Semantic Web technologies, enabling end-users to shop online in a virtual federated store comprising 20, 100, or even 200 SPARQL endpoints. This breakthrough opens up promising new avenues for developing and deploying federated applications.
-
2024Aimonier-Davat, J., Nédelec, B., Dang, M.-H., Molli, P., & Skaf-Molli, H. (2024). FedUP: Querying Large-Scale Federations of SPARQL Endpoints. Proceedings of the ACM on Web Conference 2024, 2315–2324.
@inproceedings{aimonier2024fedup, title = {{FedUP}: Querying Large-Scale Federations of {SPARQL} Endpoints}, author = {Aimonier-Davat, Julien and N{\'e}delec, Brice and Dang, Minh-Hoang and Molli, Pascal and Skaf-Molli, Hala}, booktitle = {Proceedings of the ACM on Web Conference 2024}, pages = {2315--2324}, year = {2024}, month = may, doi = {10.1145/3589334.3645704} }Processing SPARQL queries over large federations of SPARQL endpoints is crucial for keeping the Semantic Web decentralized. Despite the existence of hundreds of SPARQL endpoints, current federation engines only scale to dozens. One major issue comes from the current definition of the source selection problem, i.e., finding the minimal set of SPARQL endpoints to contact per triple pattern. Even if such a source selection is minimal, only a few combinations of sources may return results. Consequently, most of the query processing time is wasted evaluating combinations that return no results. In this paper, we introduce the concept of Result-Aware query plans. This concept ensures that every subquery of the query plan effectively contributes to the result of the query. To compute a Result-Aware query plan, we propose FedUP, a new federation engine able to produce Result-Aware query plans by tracking the …
-
2023Aimonier-Davat, J., Dang, M.-H., Molli, P., Nédelec, B., & Skaf-Molli, H. (2023, November). RAW-JENA: Approximate Query Processing for SPARQL Endpoints. ISWC’23: 22nd International Semantic Web Conference.
@inproceedings{aimonier2023rawjena, title = {{RAW-JENA}: Approximate Query Processing for {SPARQL} Endpoints}, author = {Aimonier-Davat, Julien and Dang, Minh-Hoang and Molli, Pascal and N{\'e}delec, Brice and Skaf-Molli, Hala}, booktitle = {ISWC'23: 22nd International Semantic Web Conference}, year = {2023}, month = nov, url = {https://hal.science/hal-04250060/} }Sampling-based Approximate Query Processing (S-AQP) has many important use cases for RDF, including computing large-scale statistics, embeddings, join orderings, approximate aggregations, summaries, and exploratory queries. However, current SPARQL endpoints have no support for S-AQP, and many queries just time out on public SPARQL endpoints. In this demonstration, we present RAW-JENA: an extension of Apache Jena to support S-AQP for conjunctive SPARQL queries relying on random walks. RAW-JENA delivers partial random results and cardinality estimates in a pay-as-you-go fashion.
-
2023Dang, M.-H., Gaignard, A., Skaf-Molli, H., & Molli, P. (2023, November). Schema.org: How is it used? International Semantic Web Conference (ISWC) 2023.
@inproceedings{dang2023schemaorg, title = {{Schema.org}: How is it used?}, author = {Dang, Minh-Hoang and Gaignard, Alban and Skaf-Molli, Hala and Molli, Pascal}, booktitle = {International Semantic Web Conference (ISWC) 2023}, year = {2023}, month = nov, url = {https://hal.science/hal-04250523/} }Schema.org defines a shared vocabulary for semantically annotating web pages. Due to the vast and diverse nature of the contributed annotations, it is not easy to understand the widespread use of Schema.org. In this poster, we rely on the characteristic sets computed from the web data commons datasets to provide insights into property combinations on various websites. Thanks to in-depth experiments, this poster establishes a comprehensive observatory for schema.org annotations, visually presenting the most frequently used classes, commonly used combinations of properties per class, the average number of filled properties per class, and the classes with the greatest property coverage. These findings are valuable for both the communities involved in defining Schema.org vocabularies and the users of these vocabularies.
-
2023Aimonier-Davat, J., Skaf-Molli, H., Molli, P., Dang, M.-H., & Nédelec, B. (2023). Join Ordering of SPARQL Property Path Queries. European Semantic Web Conference, 38–54.
@inproceedings{aimonier2023joinordering, title = {Join Ordering of {SPARQL} Property Path Queries}, author = {Aimonier-Davat, Julien and Skaf-Molli, Hala and Molli, Pascal and Dang, Minh-Hoang and N{\'e}delec, Brice}, booktitle = {European Semantic Web Conference}, pages = {38--54}, publisher = {Springer Nature Switzerland}, year = {2023}, month = may, doi = {10.1007/978-3-031-33455-9_3} }SPARQL property path queries provide a succinct way to write complex navigational queries over RDF knowledge graphs. However, their evaluation remains difficult as they may involve the execution of transitive closures. As a result, many property path queries just time-out when executed on public online RDF knowledge graphs. One solution to speed up their execution is to find optimal join orders. Although the join ordering problem has been extensively studied for traditional SPARQL queries, the presence of property path patterns biases existing approaches. In this paper we focus on C2RPQ_UF queries (conjunctive …
-
2023Dang, M.-H., Aimonier-Davat, J., Molli, P., Hartig, O., Skaf-Molli, H., & Le Crom, Y. (2023). FedShop: A Benchmark for Testing the Scalability of SPARQL Federation Engines. ISWC’23: 22nd International Semantic Web Conference.
@inproceedings{dang2023fedshop, title = {{FedShop}: A Benchmark for Testing the Scalability of {SPARQL} Federation Engines}, author = {Dang, Minh-Hoang and Aimonier-Davat, Julien and Molli, Pascal and Hartig, Olaf and Skaf-Molli, Hala and Le Crom, Yotlan}, booktitle = {ISWC'23: 22nd International Semantic Web Conference}, year = {2023}, doi = {10.1007/978-3-031-47243-5_16} }While several approaches to query a federation of SPARQL endpoints have been proposed in the literature, very little is known about the effectiveness of these approaches and the behavior of the resulting query engines for cases in which the number of federation members increases. The existing benchmarks that are typically used to evaluate SPARQL federation engines do not consider such a form of scalability. In this paper, we set out to close this knowledge gap by investigating the behavior of 4 state-of-the-art SPARQL federation engines using a novel benchmark designed for scalability experiments. Based on the benchmark, we show that scalability is a challenge for each of these engines, especially with respect to the effectiveness of their source selection & query decomposition approaches. FedShop is freely available online at: https://github.com/GDD-Nantes/FedShop.
Thesis
-
2024Dang, M.-H. (2024). Querying the Web as a Knowledge Graph [PhD thesis, Nantes Université].
@phdthesis{dang2024thesis, title = {Querying the Web as a Knowledge Graph}, author = {Dang, Minh-Hoang}, school = {Nantes Université}, year = {2024}, month = nov, url = {https://theses.hal.science/tel-04993631/} }The Semantic Web aims to enhance the Web with structured data, enabling more advanced search capabilities. Despite significant growth, major gaps remain in everyday information coverage, and querying large federations of SPARQL endpoints presents performance challenges. This thesis addresses two key questions: (1) Can Large Language Models (LLMs) be trusted to generate accurate Schema.org annotations? (2) Howcan efficient querying be achieved across an expanding federation of SPARQL endpoints? To tackle these issues, this work introduces two contributions. First, we propose LLM4Schema.org as a tool validate machinegenerated annotations. Second, FedShop benchmarks SPARQL federation scalability, and FedUP improves query performance with a result-aware algorithm. Together, these innovations enhance data coverage and query efficiency across the Semantic Web.
Other
-
2023Brate, R., Dang, M.-H., Hopper, F., He, Y., Meroño-Peñuela, A., & Sadashivaiah, V. (2023). Improving Language Model Predictions via Prompts Enriched with Knowledge Graphs. Zenodo (CERN European Organization for Nuclear Research).
@misc{brate2023prompts, title = {Improving Language Model Predictions via Prompts Enriched with Knowledge Graphs}, author = {Brate, Ryan and Dang, Minh-Hoang and Hopper, Fabian and He, Yuan and Mero{\~n}o-Pe{\~n}uela, Albert and Sadashivaiah, Vijay}, howpublished = {Zenodo (CERN European Organization for Nuclear Research)}, year = {2023}, month = apr, url = {https://www.academia.edu/download/108516442/Klingon_DL4KG_3.pdf} }Despite advances in deep learning and knowledge graphs (KGs), using language models for natural language understanding and question answering remains a challenging task. Pre-trained language models (PLMs) have shown to be able to leverage contextual information, to complete cloze prompts, next sentence completion and question answering tasks in various domains. Unlike structured data querying in eg KGs, mapping an input question to data that may or may not be stored by the language model is not a simple task. Recent studies have highlighted the improvements that can be made to the quality of information retrieved from PLMs by adding auxiliary data to otherwise naive prompts. In this paper, we explore the effects of enriching prompts with additional contextual information leveraged from the Wikidata KG on language model performance. Specifically, we compare the performance of naive vs. KG-engineered cloze prompts for entity genre classification in the movie domain. Selecting a broad range of commonly available Wikidata properties, we show that enrichment of cloze-style prompts with Wikidata information can result in a significantly higher recall for the investigated BERT and RoBERTa large PLMs. However, it is also apparent that the optimum level of data enrichment differs between models.
-
2023Aimonier-Davat, J., Dang, M. H., Skaf-Molli, H., & Molli, P. (2023). DeKaloG: Source Selection for Decentralized Knowledge Graphs. LS2N-Nantes Université.
@techreport{aimonier2023dekalog, title = {{DeKaloG}: Source Selection for Decentralized Knowledge Graphs}, author = {Aimonier-Davat, Julien and Dang, Minh Hoang and Skaf-Molli, Hala and Molli, Pascal}, institution = {LS2N-Nantes Universit{\'e}}, year = {2023}, month = jan, url = {https://hal.science/hal-03936036/} }As a web search engine can find relevant sources for a keyword query, the web of data needs a source selection engine to find relevant endpoints for a SPARQL query. However, source selection requires getting information about the content of endpoints and currently, it remains difficult to automatically explore the content of endpoints as web robots explore the content of web servers. Thanks to the web preemption principle, we propose to automatically build RDF summaries of endpoints through SPARQL queries. We propose an approach to compute the source selection of a query Q by evaluating a rewriting of Q on summaries. As all queries terminate under the web preemption paradigm,DeKaloG can provide a web automated source selection service relying on preemptable SPARQL servers. We empirically demonstrate that various summaries can be extracted with a data transfer proportional to the size of the summary, and highlight the trade-off between the size of the summaries, the accuracy of source selection, and the execution time of source selection.