Elamur Research⌕
☰
home / papers / 2026 / LLM-based Atomic Propositions Help Weak Extractors: Evaluation of a Propositioner for Triplet Extraction

LLM-based Atomic Propositions Help Weak Extractors: Evaluation of a Propositioner for Triplet Extraction

Paper · 2026 · Luc Pommeret, Thomas Gerald, Patrick Paroubek et al.

Summary

Building knowledge graphs from text requires pulling (subject, relation, object) triplets out of dense sentences. Coordination, nested clauses and pronouns often defeat extractors. The authors test whether first rewriting text into atomic propositions makes triplet extraction easier. Atomic propositions are minimal, self-contained statements with coreferences resolved. They ground this idea in Bar-Hillel and Carnap's Semantic Information Theory. Under that framework, splitting conjunctions is allowed, but splitting disjunctions is not, because it would invent more specific facts. They release MPropositionneur-V2, a compact multilingual "propositioner" for six European languages. A Qwen3-32B teacher produced atomizations of Wikipedia passages, which were distilled into a Qwen3-0.6B student. Its output feeds two kinds of extractors: entity-centric GLiREL and generative Qwen3 models, with dependency-parser pipelines as baselines. Evaluation covers SMiLER, FewRel, DocRED and CaRB. Weaker extractors benefit most, gaining relation recall and, in multilingual tests, overall accuracy. Strong LLMs need a fallback combination to avoid losing entity recall. The atomic propositions also serve as an auditable intermediate representation, which makes the pipeline interpretable by design.

Key points

  • The model: MPropositionneur-V2 was distilled from Qwen3-32B into Qwen3-0.6B. It splits text recursively until propositions stop changing, and it also resolves coreference by replacing pronouns with full entity names.
  • Formal grounding: a proof in the annex shows that a proposition is atomic exactly when it is a single clause of a Conjunctive Normal Form (CNF). The authors say they have checked the model's compliance with this only manually and partially.
  • Three-stage pipeline: atomization, then triplet extraction, then KG building. Extraction uses either Qwen3-4B prompting or a baseline that parses with SpaCy/Stanza and applies rules. GLiREL and CoreNLP are also tested.
  • Benchmarks: SMiLER, FewRel and DocRED were used for closed relation classification, and CaRB for OpenIE.
  • Main finding: gains appear for weak extractors (GLiREL, CoreNLP, 0.6B models). For stronger LLMs, a fallback strategy keeps the relation-extraction gains while recovering lost entity recall.

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

🔬Centre National de la Recherche ScientifiqueParis · France168 works · 62,369 citations🎓Université Paris-SaclayGif-sur-Yvette · France45 works · 6,127 citations🔬Laboratoire Interdisciplinaire des Sciences du NumériqueGif-sur-Yvette · France14 works · 507 citations

 Fields, topics and concepts

📚 Shelves 5The 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 worksComputer Vision and Pattern Recognition · 2,869 works

Cite

Cite as Luc Pommeret, Thomas Gerald, Patrick Paroubek, Sahar Ghannay, Christophe Servan and Sophie Rosset. 2026. LLM-based Atomic Propositions Help Weak Extractors: Evaluation of a Propositioner for Triplet Extraction. Proceedings of the Knowledge Graphs and Large Language Models Workshop (KG-LLM) @ LREC26.

The paper itself, not this page. BibTeX from the library's record of the work. DOI 10.63317/3kna3utavhgb

@inproceedings{pommeret2026llm,
  title = {{LLM-based Atomic Propositions Help Weak Extractors: Evaluation of a Propositioner for Triplet Extraction}},
  author = {Luc Pommeret, Thomas Gerald, Patrick Paroubek, Sahar Ghannay, Christophe Servan and Sophie Rosset},
  booktitle = {Proceedings of the Knowledge Graphs and Large Language Models Workshop (KG-LLM) @ LREC26},
  year = {2026},
  doi = {10.63317/3kna3utavhgb},
  url = {https://doi.org/10.63317/3kna3utavhgb}
}

References

22 references, 5 of them with a card of ours. ↗ leads to the work, «find» searches for one we have no id for yet.

  1. Oktie Hassanzadeh. 2021. Building a knowledge graph of events and consequences using wikidata. Wikidata@ ISWC, 2982. find
  2. Yehoshua Bar-Hillel and Rudolf Carnap. 1953. Semantic information. The British Journal for the Philosophy of Science, 4(14):147–157. find
  3. Sangnie Bhardwaj, Samarth Aggarwal, and Mausam. 2019. CaRB: A crowdsourced benchmark for open IE. 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 6262–6267, Hong… find
  4. Jack Boylan, Chris Hokamp, and Demian Gholipour Ghalandari. 2025. GLiREL -generalist model for zero-shot relation extraction. In Proceedings of the 2025 Conference of the Nations of the Americas Chapter of the Association for Computational Linguistics: Human Language Technologies (Volume 1: Long Pap… find
  5. Tong Chen, Hongwei Wang, Sihao Chen, Wenhao Yu, Kaixin Ma, Xinran Zhao, Hongming Zhang, and Dong Yu. 2024. Dense X retrieval: What retrieval granularity should we use? In Proceedings of the 2024 Conference on Empirical Methods in Natural Language Processing, pages 15159–15177, Miami, Florida, USA. A…
  6. René Cori and Daniel Lascar. 2003. Logique mathématique : cours et exercices corrigés, Tome 1 - Calcul propositionnel, algèbres de Boole, calcul des prédicats. Sciences Sup. Dunod, Paris. Avec la collaboration de Jean-Louis Krivine. find
  7. Tanguy Herserant and Vincent Guigue. 2025. Seval-ex : Un paradigme basé sur les phrases atomiques pour une évaluation explicable de la qualité des résumés. In Actes de la 20e Conférence en Recherche d’Information et Applications, CORIA 2025, Marseille, France, Juin 30 -Juillet 4, 2025, pages 217–229… find
  8. Geoffrey E. Hinton, Oriol Vinyals, and Jeffrey Dean. 2015. Distilling the knowledge in a neural network. ArXiv, abs/1503.02531.
  9. Aidan Hogan, Eva Blomqvist, Michael Cochez, Claudia D’amato, Gerard De Melo, Claudio Gutierrez, Sabrina Kirrane, José Emilio Labra Gayo, Roberto Navigli, Sebastian Neumaier, Axel-Cyrille Ngonga Ngomo, Axel Polleres, Sabbir M. Rashid, Anisa Rula, Lukas Schmelzeisen, Juan Sequeda, Steffen Staab, and A… — Knowledge Graphs
  10. Pere-Lluís Huguet Cabot, Simone Tedeschi, Axel-Cyrille Ngonga Ngomo, and Roberto Navigli. 2023. Red fm : a filtered and multilingual relation extraction dataset. In Proc. of the 61st Annual Meeting of the Association for Computational Linguistics: ACL 2023, Toronto, Canada. Association for Computati… find
  11. Sewon Min, Kalpesh Krishna, Xinxi Lyu, Mike Lewis, Wen-tau Yih, Pang Koh, Mohit Iyyer, Luke Zettlemoyer, and Hannaneh Hajishirzi. 2023. FActScore: Fine-grained atomic evaluation of factual precision in long form text generation. In Proceedings of the 2023 Conference on Empirical Methods in Natural L…
  12. Makoto Miwa, Rune Sætre, Yusuke Miyao, and Jun’ichi Tsujii. 2010. Entity-focused sentence simplification for relation extraction. In Proceedings of the 23rd International Conference on Computational Linguistics (Coling 2010), pages 788–796, Beijing, China. find
  13. Coling 2010 Organizing Committee. find
  14. Xu Han, Hao Zhu, Pengfei Yu, Ziyun Wang, Yuan Yao, Zhiyuan Liu, and Maosong Sun. 2018. FewRel: A large-scale supervised few-shot relation classification dataset with state-of-the-art evaluation. In Proceedings of the 2018 Conference on Empirical Methods in Natural Language Processing, pages 4803–480… find
  15. Christina Niklaus, Bernhard Bermeitinger, Siegfried Handschuh, and André Freitas. 2016. A sentence simplification system for improving relation extraction. In Proceedings of COLING 2016, the 26th International Conference on Computational Linguistics: System Demonstrations, pages 170–174, Osaka, Japa… find
  16. The COLING 2016 Organizing Committee. find
  17. Oren Etzioni, Michele Banko, Stephen Soderland, and Daniel S. Weld. 2008. Open information extraction from the web. Commun. ACM, 51(12):68–74.
  18. Alessandro Seganti, Klaudia Firl ag, Helena Skowronska, Michał Satława, and Piotr Andruszkiewicz. 2021. Multilingual entity and relation extraction dataset and model. In Proceedings of the 16th Conference of the European Chapter of the Association for Computational Linguistics: Main Volume, pages 1… find
  19. Open IE (CaRB). The following prompt is used for open-domain triplet extraction: Extract all factual (subject, predicate, object) triples from the sentence. One triple per line in the format: subject \| predicate \| object No explanations. If no triple can be extracted, write nothing. find
  20. Joe Stacey, Pasquale Minervini, Haim Dubossarsky, Oana-Maria Camburu, and Marek Rei. 2023. Atomic inference for nli with generated facts as atoms. In Conference on Empirical Methods in Natural Language Processing. find
  21. Sentence: {text} Andra Waagmeester, Gregory Stupp, Sebastian Burgstaller-Muehlbacher, Benjamin M Good, Malachi Griffith, Obi L Griffith, Kristina Hanspers, Henning Hermjakob, Toby S Hudson, Kevin Hy-biske, et al. 2020. Wikidata as a knowledge graph for the life sciences. Elife, 9:e52614. find
  22. Zhishang Xiang, Chuanjie Wu, Qinggang Zhang, Shengyuan Chen, Zijin Hong, Xiao Huang, and Jinsong Su. 2026. When to use graphs in rag: A comprehensive analysis for graph retrieval-augmented generation. find

Updated

↑ back to top

This page in Markdown → https://md.research.elamur.ai/free/en/papers/pommeret-2026-llm-based-atomic-propositions-help/