--- title: "GREYC at CRF Filling 2026: Rewrite Before You Extract -" description: "Case Report Forms (CRFs) are usually filled in by hand from clinical notes, which is slow and inconsistent. 3 references." authors: ["Jesus Lovon-Melgarejo", "Jérémie Pantin", "Gaël Dias"] year: 2026 venue: "Proceedings of the Third Workshop on Patient-Oriented Language Processing (CL4Health) @ LREC 2026" source: "https://aclanthology.org/2026.cl4health-1.34/" doi: "10.63317/323qntyvhp3m" licence: "CC BY-NC 4.0 (ELRA, LREC proceedings)" licence_url: "https://creativecommons.org/licenses/by-nc/4.0/" summary_licence: "Elamur Cosmic License" canonical: "https://research.elamur.ai/free/en/papers/melgarejo-2026-greyc-crf-filling-2026-rewrite/" --- [Elamur Research](https://md.research.elamur.ai/free/en/) · [this page in HTML](https://research.elamur.ai/free/en/papers/melgarejo-2026-greyc-crf-filling-2026-rewrite/) [home](https://md.research.elamur.ai/free/en/) / [papers](https://md.research.elamur.ai/free/en/papers/) / [2026](https://md.research.elamur.ai/free/en/papers/2026/page/32/) / GREYC at CRF Filling 2026: Rewrite Before You Extract - Rewriting Clinical Notes for Automated CRF # GREYC at CRF Filling 2026: Rewrite Before You Extract - Rewriting Clinical Notes for Automated CRF **Authors:** Jesus Lovon-Melgarejo, Jérémie Pantin and [Gaël Dias](https://md.research.elamur.ai/free/en/authors/gael-dias/) · **Year:** 2026 · **Venue:** [Proceedings of the Third Workshop on Patient-Oriented Language Processing (CL4Health) @ LREC 2026](https://md.research.elamur.ai/free/en/venues/cl4health/) ([Workshop on Patient-Oriented Language Processing @ LREC 2026](https://md.research.elamur.ai/free/en/venues/cl4health/2026/)) · **Publisher:** [European Language Resources Association (ELRA)](https://md.research.elamur.ai/free/en/publishers/european-language-resources-association-elra/) · **Licence:** [CC BY-NC 4.0 (ELRA, LREC proceedings)](https://creativecommons.org/licenses/by-nc/4.0/) **Links:** [Publisher's page](https://aclanthology.org/2026.cl4health-1.34/) · [PDF](https://aclanthology.org/2026.cl4health-1.34.pdf) · [DOI 10.63317/323qntyvhp3m](https://doi.org/10.63317/323qntyvhp3m) · [OpenAlex](https://openalex.org/W7166527428) **Citations:** 0 citations (OpenAlex, 2026-10-14) ## Summary Case Report Forms (CRFs) are usually filled in by hand from clinical notes, which is slow and inconsistent. For the CRF Filling 2026 shared task, the GREYC team built a modular pipeline of open-weight LLMs and tested it on the English part of the dataset. The dataset contains emergency-department notes on dyspnea from an Italian hospital, each paired with a 134-item CRF. Each CRF item is handled separately in three steps: - **Rewriter:** an LLM restates the note with that specific item in mind. - **Extractor:** a second LLM uses 3-shot prompting to read the original and rewritten notes together and pick the item's value. - **Judge:** a third LLM checks whether the original note actually supports an answer. If it does not, the output is set to unknown. Rewriting helps the system recover more correct values, but it also adds many spurious ones. The judge makes the system more cautious, cutting those errors sharply in exchange for a moderate loss of correct answers. With all three parts, the system reaches a macro-F1 of about 0.64 on the development set. This shows that enriching the note and verifying evidence work well together in a sensitive clinical setting. ## Key points - **Data:** 290 anonymised notes from San Giovanni Bosco Hospital (2021–2023), in English and Italian, split into 10 train, 90 dev and 200 test notes. Values are binary, ordinal, or measured/unknown, and unknown is allowed for every item. - **Models:** Qwen3-8B, Qwen3-32B, Llama-3.1-70B-Instruct and the medical model Llama3-Med42-8B. The general-domain models consistently beat Med42. - **Dev macro-F1 by configuration:** - Extractor alone: best 0.5046. - Extractor with rewriter, no judge: best 0.4496. - Extractor with judge, no rewriter: 0.6293. - Full pipeline: 0.6445, using Qwen3-8B as rewriter, Llama as extractor and Qwen3-32B as judge. - **Effect of the judge:** false positives drop by up to 92%, while true positives fall by no more than 30%. Precision rises in every configuration. The judge's yes/no decision is read from the logits of the first token it generates. - **Test set:** a light version of the system ranked 21st of 32 public submissions, with a macro-F1 of 0.45. *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 Scientifique](https://md.research.elamur.ai/free/en/labs/centre-national-de-la-recherche-scientifique/) — Research institute, Paris, France · 168 works · 62,369 citations - [Université de Caen Normandie](https://md.research.elamur.ai/free/en/universities/universite-de-caen-normandie/) — University, Caen, France · 3 works - [GREYC](https://md.research.elamur.ai/free/en/labs/greyc/) — Research institute, Caen, France · 2 works - [École Nationale Supérieure d'Ingénieurs de Caen](https://md.research.elamur.ai/free/en/universities/ecole-nationale-superieure-d-ingenieurs-de-caen/) — University, Caen, France · 2 works 📚 Shelves (6): [🖥️ Computer Science](https://md.research.elamur.ai/free/en/shelves/computer-science/) · 16,667 works · [🤖 Artificial Intelligence](https://md.research.elamur.ai/free/en/shelves/computer-science/artificial-intelligence/) · 16,194 works · [🎭 Arts and Humanities](https://md.research.elamur.ai/free/en/shelves/arts-and-humanities/) · 16,039 works · [🗣️ Language and Linguistics](https://md.research.elamur.ai/free/en/shelves/arts-and-humanities/language-and-linguistics/) · 15,891 works · [🧬 Biochemistry, Genetics and Molecular Biology](https://md.research.elamur.ai/free/en/shelves/biochemistry-genetics-and-molecular-biology/) · 234 works · [🧬 Molecular Biology](https://md.research.elamur.ai/free/en/shelves/biochemistry-genetics-and-molecular-biology/molecular-biology/) · 231 works ## Cite Cite as: Jesus Lovon-Melgarejo, Jérémie Pantin and Gaël Dias. 2026. GREYC at CRF Filling 2026: Rewrite Before You Extract - Rewriting Clinical Notes for Automated CRF. Proceedings of the Third Workshop on Patient-Oriented Language Processing (CL4Health) @ LREC 2026. The paper itself, not this page. BibTeX from the library's record of the work. ```bibtex @inproceedings{lovonmelgarejo2026greyc, title = {{GREYC at CRF Filling 2026: Rewrite Before You Extract - Rewriting Clinical Notes for Automated CRF}}, author = {Jesus Lovon-Melgarejo, Jérémie Pantin and Gaël Dias}, booktitle = {Proceedings of the Third Workshop on Patient-Oriented Language Processing (CL4Health) @ LREC 2026}, year = {2026}, doi = {10.63317/323qntyvhp3m}, url = {https://doi.org/10.63317/323qntyvhp3m} } ``` - [← previous Greench-v1: distilling SLMs on Greenwashing Detection](https://md.research.elamur.ai/free/en/papers/raspanti-2026-greench-v1-distilling-slms-greenwashing/) - [next → Gricean views of exhaustification: A note on Westera (2022)](https://md.research.elamur.ai/free/en/papers/zumchak-2026-gricean-views-exhaustification-note-westera/) - [↑ back to papers of 2026](https://md.research.elamur.ai/free/en/papers/2026/page/32/) ## References 3 references, 0 of them with a card of ours. ↗ leads to the work, «find» searches for one we have no id for yet. 1. Alba Gutiérrez-Sacristán, Simran Makwana, Audrey Dionne, Simran Mahanta, Karla J Dyer, Faridis Serrano, Carmen Watrin, Pierre Pages, Sajad Mousavi, Anil Degala, et al. 2024. Development and validation of an open-source pipeline for automatic population of case report forms from electronic health rec… [find](https://scholar.google.com/scholar?q=Alba+Guti%C3%A9rrez%2DSacrist%C3%A1n%2C+Simran+Makwana%2C+Audrey+Dionne%2C+Simran+Mahanta%2C+Karla+J+Dyer%2C+Faridis+Serrano%2C+Carmen+Watrin%2C+Pierre+Pages%2C+Sajad+Mousavi%2C+Anil+Degala%2C+et+al%2E+2024%2E+Development+and+validation) 2. Ran Xu, Wenqi Shi, Yue Yu, Yuchen Zhuang, Bowen Jin, May Dongmei Wang, Joyce Ho, and Carl Yang. 2024. RAM-EHR: Retrieval augmentation meets clinical predictions on electronic health records. In Proceedings of the 62nd Annual Meeting of the Association for Computational Linguistics (Volume 2: Short P… [find](https://scholar.google.com/scholar?q=Ran+Xu%2C+Wenqi+Shi%2C+Yue+Yu%2C+Yuchen+Zhuang%2C+Bowen+Jin%2C+May+Dongmei+Wang%2C+Joyce+Ho%2C+and+Carl+Yang%2E+2024%2E+RAM%2DEHR%3A+Retrieval+augmentation+meets+clinical+predictions+on+electronic+health+records%2E+In+Procee) 3. Gabriela Anna Kaczmarek, Pietro Ferrazzi, Lorenzo Porta, Vicky Rubini, and Bernardo Hanning Zhang, Shizhe Diao, Yong Lin, Yi Fung, Qing Lian, Xingyao Wang, Yangyi Chen, Heng Ji, and Tong Zhang. 2024. R-tuning: Instructing large language models to say ‘I don’t know’. In Proceedings of the 2024 Confer… [find](https://scholar.google.com/scholar?q=Gabriela+Anna+Kaczmarek%2C+Pietro+Ferrazzi%2C+Lorenzo+Porta%2C+Vicky+Rubini%2C+and+Bernardo+Hanning+Zhang%2C+Shizhe+Diao%2C+Yong+Lin%2C+Yi+Fung%2C+Qing+Lian%2C+Xingyao+Wang%2C+Yangyi+Chen%2C+Heng+Ji%2C+and+Tong+Zhang%2E+2024%2E+) *Updated 1 November 2026*