Elamur Research⌕
☰
home / papers / 2026 / Sagarmatha at SemEval-2026 Task 9: Heterogeneous Ensembling and Hierarchical Task Conditioning for Multilingual Latent Distributional Divergence Modeling

Sagarmatha at SemEval-2026 Task 9: Heterogeneous Ensembling and Hierarchical Task Conditioning for Multilingual Latent Distributional Divergence Modeling

Paper · 2026 · Sujal Maharjan, Astha Shrestha and Pratikshya Shrestha

Summary

The Sagarmatha team built a system for SemEval-2026 Task 9, which asks models to analyse polarized online discourse in 22 languages at three levels: whether a text is polarized (S1), which polarization types apply (S2), and which rhetorical tactics such as vilification or dehumanization appear (S3). Because S3 labels depend logically on the higher levels, ordinary multitask fine-tuning can produce contradictory outputs. The authors fine-tuned five multilingual encoders (mDeBERTa-v3, ReMBERT, LaBSE, mmBERT, XLM-RoBERTa) on the POLAR benchmark. Each model was trained separately to fit a single Kaggle T4 GPU. The system has two main components. Learnable weights combine the last five hidden layers. A hierarchical conditioning head feeds detached (StopGrad) S1 and S2 predictions into the S3 classifier. A rule at inference blocks labels for texts classified as non-polarized. The submitted mean-probability ensemble ranked first in several language and subtask combinations. However, the weighted-pooling variant on its own scored slightly higher on average for S1 and S2. The authors also introduce a metric, Logical Violation Rate, for checking whether predictions respect the label hierarchy.

Key points

  • Three configurations were compared: R1 (Weighted Polyglot), R2 (R1 plus HTC), and R3 (broad late-fusion ensemble, the final submission). Average F1 for R1 vs R3 was 0.790 vs 0.787 on S1 and 0.575 vs 0.559 on S2.
  • The system was 1st on S3 for English (F1 0.511) and Hausa (0.208), and on Telugu it was 1st in detection (0.905) and 2nd in categorization (0.465).
  • Logical Violation Rate (LVR) is the share of samples that receive manifestation labels despite being classified as non-polarized. The R3 ensemble reached about 33% LVR in Italian, and the hierarchical R2 reduced LVR substantially.
  • Training used AdamW (LR 2e-5), 8 epochs, batch size 16, max length 128 and a single seed (42). The S1 threshold was tuned by grid search over 0.3–0.7, and S2/S3 used a fixed threshold of 0.35.
  • In the label-correlation analysis, Racial/Ethnic polarization was most strongly linked to Stereotyping (0.91) and Religious polarization to Vilification (0.85). Results were weakest for Italian S2 (0.230) and Bengali S3 (0.228). Code is available on GitHub (SUJAL390/Sagarmatha_at_Semeval_task_9).

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

🎓IIMS CollegeKathmandu · Nepal5 works · 5 citations🎓University of BedfordshireLuton · United Kingdom2 works · 49 citations

Also on the byline: Taylor's University.

 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 1Research subjects and languages the work is about.Multilingual & cross-lingual NLP · 828 works

Cite

Cite as Sujal Maharjan, Astha Shrestha and Pratikshya Shrestha. 2026. Sagarmatha at SemEval-2026 Task 9: Heterogeneous Ensembling and Hierarchical Task Conditioning for Multilingual Latent Distributional Divergence Modeling. Proceedings of the 20th International Workshop on Semantic Evaluation (2026).

The paper itself, not this page. BibTeX from the library's record of the work. DOI 10.18653/v1/2026.semeval-1.382

@inproceedings{maharjan2026sagarmatha,
  title = {{Sagarmatha at SemEval-2026 Task 9: Heterogeneous Ensembling and Hierarchical Task Conditioning for Multilingual Latent Distributional Divergence Modeling}},
  author = {Sujal Maharjan, Astha Shrestha and Pratikshya Shrestha},
  booktitle = {Proceedings of the 20th International Workshop on Semantic Evaluation (2026)},
  year = {2026},
  doi = {10.18653/v1/2026.semeval-1.382},
  url = {https://doi.org/10.18653/v1/2026.semeval-1.382}
}

References

10 references, 7 of them with a card of ours. ↗ leads to the work, «find» searches for one we have no id for yet.

  1. We present a compact, reproducible pipeline that integrates heterogeneous backbones with learnable layer pooling and hierarchical conditioning. Our contributions are practical and generalizable: HTC offers a light-weight way to enforce structural constraints in multi-label taxonomies, and weighted p… — SemEval-2019 Task 5: Multilingual Detection of Hate Speech Against Immigrants and Women in Twitter
  2. Hyung Won Chung, Thibault Fevry, Henry Tsai, Melvin Johnson, and Sebastian Ruder. 2020. Rethinking embedding coupling in pre-trained language models. arXiv preprint arXiv:2010.12821.
  3. Pengcheng He, Jianfeng Gao, and Weizhu Chen. 2021. Debertav3: Improving deberta using electra-style pre-training with gradient-disentangled embedding sharing. arXiv preprint arXiv:2111.09543.
  4. Shanto Iyengar, Yphtach Lelkes, Matthew Levendusky, Neil Malhotra, and Sean J Westwood. 2019. The origins and consequences of affective polarization in the united states. Annual review of political science, 22(1):129–146. find
  5. Dan Kondratyuk and Milan Straka. 2019. 75 languages, 1 model: Parsing universal dependencies universally. 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 2779–279…
  6. Yphtach Lelkes. 2016. Mass polarization: Manifestations and measurements. Public Opinion Quarterly, 80(S1):392–410. find
  7. Usman Naseem, Robert Geislinger, Juan Ren, Sarah Kohail, Rudy Garrido Veliz, P Sam Sahil, Yiran Zhang, Marco Antonio Stranisci, Idris Abdulmumin, ”Ozge Alacam, Cengiz Acar”urk, Aisha Jabr, Saba Anwar, Abinew Ali Ayele, Elena Tutubalina, Aung Kyaw Htet, Xintong Wang, Surendrabikram Thapa, Tanmoy Chak… — SemEval-2026 Task 9: Detecting Multilingual, Multicultural and Multievent Online Polarization
  8. Usman Naseem, Robert Geislinger, Juan Ren, Sarah Kohail, Rudy Garrido Veliz, P Sam Sahil, Yiran Zhang, Marco Antonio Stranisci, Idris Abdulmumin, Özge Alacam, Cengiz Acartürk, Aisha Jabr, Saba Anwar, Abinew Ali Ayele, Simona Frenda, Alessandra Teresa Cignarella, Elena Tutubalina, Oleg Rogov, Aung Ky… — POLAR: A Benchmark for Multilingual, Multicultural, and Multi-Event Online Polarization
  9. Steve Rathje, Jay J Van Bavel, and Sander Van Der Linden. 2021. Out-group animosity drives engagement on social media. Proceedings of the national academy of sciences, 118(26):e2024292118. find
  10. Sebastian Ruder. 2017. An overview of multi-task learning in deep neural networks. arXiv preprint arXiv:1706.05098.

Updated

↑ back to top

This page in Markdown → https://md.research.elamur.ai/free/en/papers/maharjan-2026-sagarmatha-semeval-2026-task-9/