Elamur Research⌕
☰
home / papers / 2026 / Sema System for the DMR 2026 Shared Task: Multistage UMR Parsing with Qwen3-4B

Sema System for the DMR 2026 Shared Task: Multistage UMR Parsing with Qwen3-4B

Summary

This paper presents Sema, a submission to the DMR 2026 shared task, the first competition focused on converting multilingual text into Uniform Meaning Representation (UMR) graphs. Earlier UMR parsers relied on AMR as an intermediate step. Sema instead trains directly on UMR data. The authors adapt Qwen3-4B with LoRA in three consecutive stages:

  1. Training on the noisy ("dirty") split, capped at 2000 examples per language to limit the Czech-dominated imbalance.
  2. Continuing on the clean split while adding document-level content, mostly modal relations.
  3. A separate stage for alignments, recast as word-to-node pairs so that unaligned words can be marked explicitly.

Graphs are linearized in plain PENMAN style. Rule-based fixes (bracket repair, format conversion, empty fallbacks) make the outputs acceptable to the official scorer. The system reaches a macro-averaged F1 of 0.1943. Italian was never seen during fine-tuning, yet it outperforms several trained languages, which points to transfer from multilingual pretraining. Navajo and Arapaho remain weak. The work offers a simple single-GPU baseline for multilingual UMR parsing.

Key points

  • Training data is highly unbalanced. Czech accounts for 159,906 dirty samples, over 80% of that split. The clean split is tiny: English 180, Chinese 557, Navajo 5, Arapaho 53, Czech 103, Latin 0.
  • Training and decoding setup:
    • LoRA with r=16 and alpha=32, applied to attention and feed-forward projections.
    • Learning rate of 2e-4 for the dirty stage and 1e-4 for the clean and alignment stages.
    • Effective batch size of 32 through gradient accumulation, on a single A100 (40GB).
    • Beam search with beam size 4.
  • Per-language F1: Czech 0.2652 (best), Chinese 0.2585, Italian 0.2137, English 0.1919, Latin 0.1918, Navajo 0.1273, Arapaho 0.1115 (lowest).
  • Output quality and failure modes:
    • Outputs are mostly well-formed; the main defect is cycles where nodes dominate themselves.
    • Very long sentences can loop endlessly and are regenerated.
    • Navajo and Arapaho graphs tend to be generic and sparse.
  • Rejected augmentation: adding Levenshtein-based heuristic alignments to the dirty data brought no improvement and was dropped from the final pipeline.

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

Cited by

1 paper of our library cites it (of the 140673 whose bibliographies we read); each has a page here.

Cite

The paper itself, not this page. BibTeX from the library's record of the work. DOI 10.63317/24cughzsw4ae

@inproceedings{vergnette2026sema,
  title = {{Sema System for the DMR 2026 Shared Task: Multistage UMR Parsing with Qwen3-4B}},
  author = {Vergnette, Rémi de and Amblard, Maxime},
  booktitle = {Proceedings of The Seventh International Workshop on Designing Meaning Representations (DMR 2026) @ LREC 2026},
  year = {2026},
  doi = {10.63317/24cughzsw4ae},
  url = {https://doi.org/10.63317/24cughzsw4ae}
}

References

9 references, 3 of them with a card of ours. ↗ leads to the work, «find» searches for one we have no id for yet.

  1. Laura Banarescu, Claire Bonial, Shu Cai, Madalina Georgescu, Kira Griffitt, Ulf Hermjakob, Kevin Knight, Philipp Koehn, Martha Palmer, and Nathan Schneider. 2013. Abstract Meaning Representation for sembanking. In Proceedings of the 7th Linguistic Annotation Workshop and Interoperability with Discourse, pages 178–186, Sofia, Bulgaria. Association for Computational Linguistics. find our card
  2. Michele Bevilacqua, Rexhina Blloshmi, and Roberto Navigli. 2021. One spring to rule them both: Symmetric amr semantic parsing and gen-clean / alignment 1 32 3 3 1e-4 0.001 constant with warmup 16 32 0 alignment 1024 4 0.7 0.9 Language Arapaho Chinese Czech English Italian Latin Navajo Average F1 0.1115 0.2585 0.2652 0.1919 0.2137 0.1918 0.1273 0.1943 Table 4: Score for each language and the macro-average reported by the shared task. eration without a complex pipeline. Proceedings of the AAAI Conference on Artificial Intelligence, 35(14):12564–12573. find
  3. Julia Bonn, Matthew J. Buchholz, Jayeol Chun, Andrew Cowell, William Croft, Lukas Denk, Sijia Ge, Jan Hajič, Kenneth Lai, James H. Martin, Skatje Myers, Alexis Palmer, Martha Palmer, Claire Benet Post, James Pustejovsky, Kristine Stenzel, Haibo Sun, Zdeňka Urešová, Rosa Vallejos, Jens E. L. Van Gysel, Meagan Vigus, Nianwen Xue, and Jin Zhao. 2024. Building a broad infrastructure for uniform meaning representations. In Proceedings of the 2024 Joint International Conference on Computational Linguistics, Language Resources and Evaluation (LREC-COLING 2024), pages 2537–2547, Torino, Italia. ELRA and ICCL. find
  4. Jayeol Chun and Nianwen Xue. 2024. Uniform meaning representation parsing as a pipelined approach. In Proceedings of TextGraphs-17: Graph-based Methods for Natural Language Processing, pages 40–52, Bangkok, Thailand. Association for Computational Linguistics. find
  5. Edward J Hu, Yelong Shen, Phillip Wallis, Zeyuan Allen-Zhu, Yuanzhi Li, Shean Wang, Liang Wang, Weizhu Chen, et al. 2022. Lora: Low-rank adaptation of large language models. Iclr, 1(2):3. find
  6. Emma Markle, Javier Gutierrez Bach, and Shira Wein. 2026. Setup: Sentence-level english-to-uniform meaning representation parser. ↗
  7. Chunlei Xin, Yaojie Lu, Hongyu Lin, Shuheng Zhou, Huijia Zhu, Weiqiang Wang, Zhongyi Liu, Xianpei Han, and Le Sun. 2024. Beyond full fine-tuning: Harnessing the power of LoRA for multi-task instruction tuning. In Proceedings of the 2024 Joint International Conference on Computational Linguistics, Language Resources and Evaluation (LREC-COLING 2024), pages 2307–2317, Torino, Italia. ELRA and ICCL. find
  8. An Yang, Anfeng Li, Baosong Yang, Beichen Zhang, Binyuan Hui, Bo Zheng, Bowen Yu, Chang Gao, Chengen Huang, Chenxu Lv, Chujie Zheng, Dayiheng Liu, Fan Zhou, Fei Huang, Feng Hu, Hao Ge, Haoran Wei, Huan Lin, Jialong Tang, Jian Yang, Jianhong Tu, Jianwei Zhang, Jianxin Yang, Jiaxi Yang, Jing Zhou, Jingren Zhou, Junyang Lin, Kai Dang, Keqin Bao, Kexin Yang, Le Yu, Lianghao Deng, Mei Li, Mingfeng Xue, Mingze Li, Pei Zhang, Peng Wang, Qin Zhu, Rui Men, Ruize Gao, Shixuan Liu, Shuang Luo, Tianhao Li, Tianyi Tang, Wenbiao Yin, Xingzhang Ren, Xinyu Wang, Xinyu Zhang, Xuancheng Ren, Yang Fan, Yang Su, Yichang Zhang, Yinger Zhang, Yu Wan, Yuqiong Liu, Zekun Wang, Zeyu Cui, Zhenru Zhang, Zhipeng Zhou, and Zihan Qiu. 2025. Qwen3 technical report. ↗ our card
  9. Jan Štěpánek, Daniel Zeman, Markéta Lopatková, Federica Gamba, Hana Hledíková, and Nianwen Xue. 2026. First shared task on UMR parsing. In Proceedings of the Seventh International Workshop on Designing Meaning Representations, Palma, Spain. ELRA. find our card

↑ back to top

This page in Markdown → https://md.research.elamur.ai/free/en/papers/vergnette-2026-sema-system-dmr-2026-shared/