Elamur Research⌕
☰
home / papers / 2026 / Towards the LinkEn Knowledge Base. A Neuro-Symbolic Approach to Build a Linked Data Hub for English Lemmas with Large Language Models

Towards the LinkEn Knowledge Base. A Neuro-Symbolic Approach to Build a Linked Data Hub for English Lemmas with Large Language Models

Paper · 2026 · Lorenzo Augello and Marco Passarotti

Summary

English has many lexical resources, but none serves as a LiLa-style lemma bank published as Linked Data. Without such a hub, it is hard to link and query across dictionaries, corpora and annotations. The authors build the first part of LinkEn, a planned knowledge base of interoperable English resources. That first part is a lemma bank modelled on the LiLa Knowledge Base for Latin and on OntoLex-Lemon. Each lemma is an instance of lila:Lemma, a subclass of ontolex:Form. External lexical entries attach to it via ontolex:canonicalForm, and corpus tokens via lila:hasLemma. Manual curation would be costly, so they use a neuro-symbolic pipeline instead. Gemini 2.5 Flash turns selected lemma sets into Turtle RDF under explicit ontological constraints. The output is checked for formal validity and for linguistic correctness. Faulty items go through supervised reprompting and correction. The resulting bank has roughly 21k lemmas and 3.5k hypolemmas, and it is queryable through a SPARQL endpoint. The work matters for two reasons. It gives English a lemma hub compatible with other lemma-centric graphs, and it shows a controlled way to let an LLM write knowledge-graph data.

Key points

  • The model reuses LiLa: lemmas are canonical forms (ontolex:Form), not lexical entries. Other resources link to them through ontolex:canonicalForm and lila:hasLemma.
  • Prompting runs in two stages:
    • one generates hypolemmas from words with specific PoS;
    • the other produces RDF triples and PoS tags for the remaining input words.
  • Every output gets two kinds of checks: Turtle/RDF parsing with ontology alignment, and a test of the LLM's linguistic competence. Incorrect items loop back through reprompting with manual supervision.
  • Reported statistics:
    • about 21,275 lemmas and 3,463 hypolemmas;
    • roughly 24.8k nodes and around 25.4k written variants;
    • only 2 classes and 5 properties;
    • 5.17 triples per node on average, about 128k triples in total.
  • The data is published at linken-lod.eu/sparql with a LodView browser, and the code is on GitHub (lorenzoaugello/LinkEn).

Interesting ideas

  • I1. Let the LLM write structured data only inside a fixed schema, then run a parser and ontology check before anything is accepted, so generation stays scalable but cannot drift from the format.
  • I2. Evaluate generated knowledge on two separate axes: whether it is well-formed and whether its linguistic content is right, since a valid triple can still hold a wrong PoS.
  • I3. Treat the lemma as a stable identifier that tokens and lexical entries all point to, rather than duplicating lexical data per resource.
  • I4. Handle derived forms (hypolemmas) in a separate generation pass seeded by specific parts of speech, rather than mixing them into the general lemma prompt.
  • I5. Keep the hub schema tiny (two classes, five properties) so it stays interoperable and easy to validate, and push richer information into linked resources.

For Mova

How these ideas could serve Mova, our small open language model — its files, stages and tests:

  • I1 (learn): LLM teacher tiers could emit silver annotations in a strict serialisation that must parse and validate before entering training data, with rejects reprompted.
  • I2 (learn): score teacher output separately for format validity and for content agreement with witness parsers, so each error type gets its own repair path.
  • I3 (tokens): crate en could assign each lemma a persistent ID that tagger output, VerbNet classes and WordNet senses all reference, giving one join point for explanations.
  • I4 (tokens): build forms such as participles or derived words in a dedicated lemmatisation pass with its own checks, instead of folding them into the main lemma table.
  • I5 (world): the compiled-knowledge layer could keep a minimal core lemma graph and attach ontology, measures and frames as separate linked layers.

Our summary of the paper, in our own words — not its text. The work belongs to its authors; read it at the links above.

 Institutions

🎓Università Cattolica del Sacro CuoreMilan · Italy7 works · 33 citations

 Fields, topics and concepts

📚 Shelves 4The field of science this work belongs to (OpenAlex fields).Computer Science · 16,667 worksArtificial Intelligence · 16,194 worksArts and Humanities · 16,039 worksLanguage and Linguistics · 15,891 works

🧭 Topics 2Research subjects and languages the work is about.English · 2,263 worksKnowledge graphs & ontologies · 192 works

Cite

Cite as Lorenzo Augello and Marco Passarotti. 2026. Towards the LinkEn Knowledge Base. A Neuro-Symbolic Approach to Build a Linked Data Hub for English Lemmas with Large Language Models. Proceedings of 10th Workshop on Linked Data in Linguistics (LDL-2026).

The paper itself, not this page. BibTeX from the library's record of the work. DOI 10.63317/5ndvctbdvkb7

@inproceedings{augello2026linken,
  title = {{Towards the LinkEn Knowledge Base. A Neuro-Symbolic Approach to Build a Linked Data Hub for English Lemmas with Large Language Models}},
  author = {Lorenzo Augello and Marco Passarotti},
  booktitle = {Proceedings of 10th Workshop on Linked Data in Linguistics (LDL-2026)},
  year = {2026},
  doi = {10.63317/5ndvctbdvkb7},
  url = {https://doi.org/10.63317/5ndvctbdvkb7}
}

References

25 references, 4 of them with a card of ours. ↗ leads to the work, «find» searches for one we have no id for yet.

  1. Gabor Angeli, Melvin Jose Johnson Premkumar, and Christopher D. Manning. 2015. Leveraging linguistic structure for open domain information extraction. In Proceedings of the 53rd Annual Meeting of the Association for Computational Linguistics and the 7th International Joint Conference on Natural Language Processing (Volume 1: Long Papers), pages 344–354, Beijing, China. Association for Computational Linguistics. ↗
  2. Tim Berners-Lee, James Hendler, and Ora Lassila. 2001. The semantic web: A new form of web content that is meaningful to computers will unleash a revolution of new possibilities. ScientificAmerican.com. find
  3. Salvatore Carta, Alessandro Giuliani, Leonardo Piano, Alessandro Sebastian Podda, Livio Pompianu, and Sandro Gabriele Tiddia. 2023. Iterative zero-shot llm prompting for knowledge graph construction. ↗
  4. Christian Chiarcos. 2012. Powla: Modeling linguistic corpora in owl/dl. In The Semantic Web: Research and Applications, pages 225–239, Berlin, Heidelberg. Springer Berlin Heidelberg. find
  5. Christian Chiarcos and Maria Sukhareva. 2015. Olia – ontologies of linguistic annotation. Semantic Web, 6:379–386. ↗
  6. Xin Dong, Evgeniy Gabrilovich, Geremy Heitz, Wilko Horn, Ni Lao, Kevin Murphy, Thomas Strohmann, Shaohua Sun, and Wei Zhang. 2014. Knowledge vault: a web-scale approach to probabilistic knowledge fusion. In Proceedings of the 20th ACM SIGKDD International Conference on Knowledge Discovery and Data Mining, KDD ’14, page 601–610, New York, NY, USA. Association for Computing Machinery. find
  7. Fellbaum, Christiane. 1998. WordNet: An Electronic Lexical Database. The MIT Press. find
  8. Theodorus Fransen, Cormac Anderson, Sacha Beniamine, and Marco Passarotti. 2024. The MOLOR lemma bank: a new LLOD resource for Old Irish. In Proceedings of the 9th Workshop on Linked Data in Linguistics @ LREC-COLING 2024, pages 37–43, Torino, Italia. ELRA and ICCL. find
  9. Jan-Christoph Kalo and Leandra Fichtel. 2022. Kamel: Knowledge analysis with multitoken entities in language models. In 4th Conference on Automated Knowledge Base Construction. find
  10. Eleonora Litta, Marco Passarotti, Paolo Brasolin, Giovanni Moretti, Valerio Basile, Andrea Di Fabio, and Cristina Bosco. 2024. The lemma bank of the LiITA knowledge base of interoperable resources for Italian. In Proceedings of the Tenth Italian Conference on Computational Linguistics (CLiC-it 2024), pages 517–522, Pisa, Italy. CEUR Workshop Proceedings. find
  11. John P. McCrae, Julia Bosque Gil, Jordi Gràcia, Paul Bitelaar, and Philipp Cimiano. 2017. The ontolex-lemon model: Development and applications. find
  12. Jeff Z. Pan, Simon Razniewski, Jan-Christoph Kalo, Sneha Singhania, Jiaoyan Chen, Stefan Dietze, Hajira Jabeen, Janna Omeliyanenko, Wen Zhang, Matteo Lissandrini, Russa Biswas, Gerard de Melo, Angela Bonifati, Edlira Vakaj, Mauro Dragoni, and Damien Graux. 2023. Large language models and knowledge graphs: Opportunities and challenges. ↗
  13. Marco Carlo Passarotti and Francesco Mambrini. 2022. Linking latin: Interoperable lexical resources in the lila project. In Building new resources for historical linguistics, pages 103–124. avia University Press. ↗
  14. Marco Carlo Passarotti, Francesco Mambrini, Greta Franzini, Flavio Massimiliano Cecchini, Eleonora Maria Gabriella Litta Modignani Picozzi, Giovanni Moretti, Paolo Ruffolo, and Rachele Sprugnoli. 2020. Interlinking through lemmas. the lexical collection of the lila knowledge base of linguistic resources for latin. Studi e Saggi Linguistici, 58(1):177–212. ↗
  15. Fabio Petroni, Tim Rocktäschel, Sebastian Riedel, Patrick Lewis, Anton Bakhtin, Yuxiang Wu, and Alexander Miller. 2019. Language models as knowledge bases? In Proceedings of the 2019 Conference on Empirical Methods in Natural Language Processing and the 9th International Joint Conference on Natural Language Processing (EMNLP-IJCNLP), pages 2463–2473, Hong Kong, China. Association for Computational Linguistics. ↗
  16. Chenguang Wang, Xiao Liu, and Dawn Song. 2020. Language models are open knowledge graphs. ↗
  17. Rui Yang, Boming Yang, Aosong Feng, Sixun Ouyang, Moritz Blum, Tianwei She, Yuang Jiang, Freddy Lecue, Jinghui Lu, and Irene Li. 2025. Graphusion: A rag framework for knowledge graph construction with a global perspective. ↗
  18. Lingfeng Zhong, Jia Wu, Qian Li, Hao Peng, and Xindong Wu. 2023. A comprehensive survey on automatic knowledge graph construction. ACM Comput. Surv., 56(4). ↗
  19. Zexuan Zhong, Dan Friedman, and Danqi Chen. 2021. Factual probing is [MASK]: Learning vs. learning to recall. In Proceedings of the 2021 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, pages 5017–5033, Online. Association for Computational Linguistics.
  20. Yuqi Zhu, Xiaohan Wang, Jing Chen, Shuofei Qiao, Yixin Ou, Yunzhi Yao, Shumin Deng, Huajun Chen, and Ningyu Zhang. 2024. Llms for knowledge graph construction and reasoning: Recent capabilities and future opportunities. 9. ↗
  21. Language Resource References BNC Consortium. 2007. find
  22. British National Corpus 1994. Literary and Linguistic Data Service. find
  23. Adam Kilgarriff. 1997. Putting frequencies in the dictionary. Int. J. Lexicogr, 10. ↗
  24. McCrae, John P. and Rademaker, Alexandre and Bond, Francis and Rudnicka, Ewa and Fellbaum, Christiane. 2019. find
  25. English WordNet 2019 – An Open-Source WordNet for English. Global Wordnet Association.

Updated

↑ back to top

This page in Markdown → https://md.research.elamur.ai/free/en/papers/augello-2026-linken-knowledge-base-neuro-symbolic/