Elamur Research⌕
☰
home / papers / 2026 / UTokyo Tsuruoka Lab at SemEval-2026 Task 9: Efficient Single Forward Pass Inference for Multi-Label Polarization Classification

UTokyo Tsuruoka Lab at SemEval-2026 Task 9: Efficient Single Forward Pass Inference for Multi-Label Polarization Classification

Paper · 2026 · Howard Tangkulung and Yoshimasa Tsuruoka

Summary

The paper describes the UTokyo Tsuruoka Lab system for SemEval-2026 Task 9, which uses the POLAR benchmark (over 110K texts in 22 languages) to detect online polarization. Subtask 1 asks whether a text is polarized. Subtask 2 asks which polarization types apply, a multi-label problem. The authors fine-tune Gemma-3 LLMs with LoRA. For detection they use SALSA, which compares the logits of the "0" and "1" tokens and applies thresholds tuned per language. For type classification they introduce Single-Forward-Pass (SFP) inference. Each type gets its own single token, and all types share a "None" token. Every label is then scored independently from the logits at one output position, so no sequential JSON-style decoding is needed. Compared with that autoregressive baseline, SFP is about 4.5× faster and gives higher macro F1. It also avoids errors cascading from earlier labels to later ones. Fine-tuning on one language transferred significantly better to related languages. This suggests a practical way to adapt to low-resource languages, such as using Russian or Polish models for other Slavic languages.

Key points

  • Models and training: gemma-3-12b-it and gemma-3-27b-it-bnb-4bit, trained with LoRA and unsloth. The model size submitted was chosen separately for each language and subtask. The metric is macro F1 per language.
  • Subtask 2 results (average macro F1):
    • Validation: SFP reached 58.43 (12B) and 58.56 (27B), against 54.46 for the JSON FT baseline.
    • Test: the submitted system scored 61.29.
    • Speed: with batches of 8, SFP took 0.53 s per batch versus 2.36 s for JSON FT.
  • Subtask 1 results: SALSA only slightly beat standard fine-tuning on validation (81.34 vs 80.72), and the test average was 81.80.
  • Rankings: 1st place in 8 languages for Subtask 1 and 6 for Subtask 2. Top 5 in 16 of the 22 languages across both subtasks.
  • Cross-lingual transfer (Subtask 2): gains were larger for related language pairs. Pairs from the same family branch gained 2.97 points versus 0.79 otherwise, and pairs from the same sub-branch gained 4.05 versus 1.35. Both differences were significant at p < 0.05 under the Mann–Whitney U and Welch's t tests. A limitation is that target-language threshold tuning was still required.

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

🎓The University of TokyoTokyo · Japan90 works · 9,156 citations

 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

Cite

Cite as Howard Tangkulung and Yoshimasa Tsuruoka. 2026. UTokyo Tsuruoka Lab at SemEval-2026 Task 9: Efficient Single Forward Pass Inference for Multi-Label Polarization Classification. 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-1.211

@inproceedings{tangkulung2026utokyo,
  title = {{UTokyo Tsuruoka Lab at SemEval-2026 Task 9: Efficient Single Forward Pass Inference for Multi-Label Polarization Classification}},
  author = {Howard Tangkulung and Yoshimasa Tsuruoka},
  booktitle = {Proceedings of the 20th International Workshop on Semantic Evaluation (2026)},
  year = {2026},
  doi = {10.18653/v1/2026.semeval-1.211},
  url = {https://doi.org/10.18653/v1/2026.semeval-1.211}
}

References

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

  1. Teodorescu, David Ifeoluwa Adelani, Alham Fikri Aji, Felermino Dario Mario Ali, Vladimir Araujo, Abinew Ali Ayele, Oana Ignat, Alexander Panchenko, and 2 others. 2025. SemEval-2025 task 11: Bridging the gap in text-based emotion detection. In Proceedings of the 19th International Workshop on Semanti… find
  2. Dimosthenis Antypas and Jose Camacho-Collados. 2023. Robust hate speech detection in social media: A cross-dataset empirical evaluation. In The 7th Workshop on Online Abuse and Harms (WOAH), pages 231–242, Toronto, Canada. Association for Computational Linguistics. find
  3. Ruslan Berdichevsky, Shai Nahum-Gefen, and Elad Ben Zaken. 2025. Salsa: Single-pass autoregressive llm structured classification. Preprint, arXiv:2510.22691. — [arXiv:2510.22691](https://arxiv.org/abs/2510.22691) find
  4. Mauro Bruno, Elena Catanese, and Francesco Ortame. 2024. Towards a hate speech index with attention-based LSTMs and XLM-RoBERTa. In Proceedings of the Tenth Italian Conference on Computational Linguistics (CLiC-it 2024), pages 106–113, Pisa, Italy. CEUR Workshop Proceedings. find
  5. Alexis Conneau, Kartikay Khandelwal, Naman Goyal, Vishrav Chaudhary, Guillaume Wenzek, Francisco Guzmán, Edouard Grave, Myle Ott, Luke Zettlemoyer, and Veselin Stoyanov. 2020. Unsupervised cross-lingual representation learning at scale. In Proceedings of the 58th Annual Meeting of the Association fo… find
  6. Kiran Garimella. 2018. Polarization on Social Media. Ph.D. thesis, Aalto University, Finland. find
  7. Edward J Hu, yelong shen, Phillip Wallis, Zeyuan Allen-Zhu, Yuanzhi Li, Shean Wang, Lu Wang, and Weizhu Chen. 2022. LoRA: Low-rank adaptation of large language models. In International Conference on Learning Representations.
  8. Julia Jaremko, Dagmar Gromann, and Michael Wiegand. 2025. Revisiting implicitly abusive language detection: Evaluating LLMs in zero-shot and few-shot settings. In Proceedings of the 31st International Conference on Computational Linguistics, pages 3879– 3898, Abu Dhabi, UAE. Association for Computat… find
  9. Marc Marone, Orion Weller, William Fleshman, Eugene Yang, Dawn Lawrie, and Benjamin Van Durme. 2025. mmbert: A modern multilingual encoder with annealed language learning. Preprint, arXiv:2509.06888.
  10. Raquel Martínez-España, Julio Fernández-Pedauye, José Giner-Pérez de Lucía, Jose Miguel Rojo-Martínez, Kaoutar Bakdid-Albane, and Juan José García-Escribano. 2024. Methodology for Measuring Individual Affective Polarization Using Sentiment Analysis in Social Networks. IEEE Access. find
  11. Shamsuddeen Hassan Muhammad, Nedjma Ousidhoum, Idris Abdulmumin, Seid Muhie Yimam, Jan Philip Wahle, Terry Lima Ruas, Meriem Beloucif, Christine De Kock, Tadesse Destaw Belay, Ibrahim Said Ahmad, Nirmal Surange, Daniela Usman Naseem, Robert Geislinger, Juan Ren, Sarah Kohail, Rudy Garrido Veliz, P S… — SemEval-2026 Task 9: Detecting Multilingual, Multicultural and Multievent Online Polarization
  12. 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
  13. Eshaan Tanwar, Subhabrata Dutta, Manish Borthakur, and Tanmoy Chakraborty. 2023. Multilingual LLMs are better cross-lingual in-context learners with alignment. In Proceedings of the 61st Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 6292–6307, Toronto…
  14. Wilson Wongso, David Setiawan, Ananto Joyoadikusumo, and Steven Limcorn. 2025. Lazarus NLP at SemEval-2025 task 11: Fine-tuning large language models for multi-label emotion classification via sentence-label pairing. In Proceedings of the 19th International Workshop on Semantic Evaluation (SemEval-2… find
  15. Jieying Xue, Phuong Nguyen, Minh Nguyen, and Xin Liu. 2025. JNLP at SemEval-2025 task 11: Cross-lingual multi-label emotion detection using generative models. In Proceedings of the 19th International Workshop on Semantic Evaluation (SemEval-2025), pages 20–27, Vienna, Austria. Association for Comput… find
  16. Haris Bin Zia, Ignacio Castro, Arkaitz Zubiaga, and Gareth Tyson. 2022. Improving zero-shot cross-lingual hate speech detection with pseudo-label fine-tuning of transformer language models. Proceedings of the International AAAI Conference on Web and Social Media, 16(1):1435–1439. find

Updated

↑ back to top

This page in Markdown → https://md.research.elamur.ai/free/en/papers/tangkulung-2026-utokyo-tsuruoka-lab-semeval-2026/