Elamur Research⌕
☰
home / papers / 2026 / Archaeology at MWE-2026 PARSEME 2.0 Subtasks 1 and 2: Parsing is for Encoders, Paraphrasing is for LLMs

Archaeology at MWE-2026 PARSEME 2.0 Subtasks 1 and 2: Parsing is for Encoders, Paraphrasing is for LLMs

Summary

This paper presents a two-stage system for Romanian in the PARSEME 2.0 Shared Task. It covers detecting multiword expressions (MWEs) in Subtask 1 and rewriting them in Subtask 2. The authors argue that generative models handle meaning well but are imprecise about where an expression begins and ends. Romanian makes this harder because MWE parts are often separated by long intervening material. For identification, they convert the nested .cupt annotations into BIO labels, normalize diacritics, and fine-tune five encoders. RoBERT-base performed best. GPT-4o was tested only on difficult and unseen cases, and its predictions were less precise than the encoder's. For paraphrasing, RoBERT-base finds the span and its category, and GPT-4o then rewrites only that span using few-shot prompts selected by category. The system ranked first for Romanian, ahead of large multilingual parsers. The authors conclude that a small monolingual encoder can guide an LLM effectively. This could later allow smaller open models to handle the rewriting step.

Key points

  • Official Romanian test result (Subtask 1): RoBERT-base reached a Global MWE-based F1 of 85.65% (P 91.03, R 80.88) and ranked 1st. BeeParser scored 83.60 and MTLB-STRUCT scored 82.27.
  • Unseen MWEs: the system also ranked 1st on unseen expressions with F1 16.00, compared with 9.88 for BeeParser and 4.82 for MTLB-STRUCT.
  • Encoder comparison: the authors tested multilingual BERT, xlm-roberta-base, mdeberta-v3-base, RoBERT-base and bert-base-romanian-cased-v1. RoBERT-base had the highest validation F1 (0.903).
  • GPT-4o stress tests: two targeted subsets were built, one of 170 hard, discontinuous or dense examples and one of 85 unseen MWEs. Scoring used strict and soft (overlap-based) F1. GPT-4o's main weakness was inaccurate span boundaries.
  • Subtask 2 trade-off:
    • The "Minimal" prompt achieved the highest BERTScore (89.25).
    • The "Creative" prompt produced far more unique terms (628 vs 235) and higher entropy (6.13 vs 5.36).
    • The cost of the Creative prompt was a BERTScore about 4.5 points lower (84.73).

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

Fields: Linguistics & languages · Machine translation

Cite

The paper itself, not this page. BibTeX from the library's record of the work. DOI 10.18653/v1/2026.mwe-1.31

@inproceedings{roscan2026archaeology,
  title = {{Archaeology at MWE-2026 PARSEME 2.0 Subtasks 1 and 2: Parsing is for Encoders, Paraphrasing is for LLMs}},
  author = {Roscan, Rares-Alexandru and Nisioi, Sergiu},
  booktitle = {Proceedings of the 22nd Workshop on Multiword Expressions (MWE 2026)},
  year = {2026},
  doi = {10.18653/v1/2026.mwe-1.31},
  url = {https://doi.org/10.18653/v1/2026.mwe-1.31}
}

References

12 references, 0 of them with a card of ours. ↗ leads to the work, «find» searches for one we have no id for yet.

  1. Sweta Agrawal and Marine Carpuat. 2024. Do text simplification systems preserve meaning? a human evaluation via reading comprehension. Transactions of the Association for Computational Linguistics, 12:432– 448. find
  2. Fernando Alva-Manchego, Regina Stodden, Joseph Marvin Imperial, Abdullah Barayan, Kai North, and Harish Tayyar Madabushi. 2025. Findings of the TSAR 2025 shared task on readability-controlled text simplification. In Proceedings of the Fourth Workshop on Text Simplification, Accessibility and Readability (TSAR 2025), pages 116–130, Suzhou, China. Association for Computational Linguistics. find
  3. Fabian Anghel, Cristea Petru-Theodor, Claudiu Cre-anga, and Sergiu Nisioi. 2025. RALS: Resources and baselines for Romanian automatic lexical simplification. In Proceedings of the 2025 Conference on Empirical Methods in Natural Language Processing, pages 31481–31492, Suzhou, China. Association for Computational Linguistics. find
  4. Andrei Avram, Verginica Barbu Mititelu, and Dumitru-Clementin Cercel. 2023. Romanian multiword expression detection using multilingual adversarial training and lateral inhibition. In Proceedings of the 19th Workshop on Multiword Expressions (MWE 2023), pages 7–13, Dubrovnik, Croatia. Association for Computational Linguistics. find
  5. Timothy Baldwin and Su Nam Kim. 2010. Multiword expressions. In Handbook of Natural Language Processing. find
  6. Verginica Barbu Mititelu, Mihaela Cristescu, and Mihaela Onofrei. 2019. The Romanian corpus annotated with verbal multiword expressions. In Proceedings of the Joint Workshop on Multiword Expressions and WordNet (MWE-WN 2019), pages 13–21. find
  7. Verginica Barbu Mititelu, Voula Giouli, Gražina Korvel, Chaya Liebeskind, Irina Lobzhanidze, Rusudan Makhachashvili, Stella Markantonatou, Aleksandra Markovic, and Ivelina Stoyanova. 2025. The challenges of syntactic descriptions of multiword expressions in electronic lexicography. In eLex 2025: Electronic Lexicography in the 21st Century, pages 1–20. Lexical Computing CZ s.r.o. Paper ID: 17; 16–18 Nov 2025, Brno, Czech Republic. find
  8. Chutima Bunparit and Pennapa Riabroi. 2025. Effects of narrow reading on the reading comprehension, vocabulary acquisition, and perceptions of l2 students in an esp classroom. LEARN Journal: Language Education and Acquisition Research Network, 18(2):571–593. ↗
  9. Julen Etxaniz, Gorka Azkune, Aitor Soroa, Oier Lopez de Lacalle, and Mikel Artetxe. 2024. Do multilingual language models think better in English? In Proceedings of the 2024 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies (Volume 2: Short Papers), pages 550–564, Mexico City, Mexico. Association for Computational Linguistics. ↗
  10. Theo Guidroz, Diego Ardila, Jimmy Li, Adam Mansour, Paul Jhun, Nina Gonzalez, Xiang Ji, Mike Sanchez, Sujay Kakarmath, Mathias MJ Bellaiche, Miguel Ángel Garrido, Faruk Ahmed, Divyansh Choudhary, Jay Hartford, Chenwei Xu, Henry Javier Serrano Echev-erria, Yifan Wang, Jeff Shaffer, Eric, and 8 others. 2025. Llm-based text simplification and its effect on user comprehension and cognitive load. Preprint, arXiv:2505.01980. ↗
  11. Shiva Taslimipoor, Sara Bahaadini, and Ekaterina Kochmar. 2020. MTLB-STRUCT @parseme 2020: Capturing unseen multiword expressions using multi-task learning and pre-trained masked language models. In Proceedings of the Joint Workshop on Multiword Expressions and Electronic Lexicons, pages 142–148, online. Association for Computational Linguistics. ↗
  12. Tianyang Zhong, Zhenyuan Yang, Zhengliang Liu, Ruidong Zhang, Weihang You, Yiheng Liu, Haiyang Sun, Yi Pan, Yiwei Li, Yifan Zhou, Hanqi Jiang, Junhao Chen, and Tianming Liu. 2026. Opportunities and challenges of large language models for low-resource languages in humanities research. Preprint, arXiv:2412.04497. A ↗

↑ back to top

This page in Markdown → https://md.research.elamur.ai/free/en/papers/roscan-2026-archaeology-mwe-2026-parseme-2/