Elamur Research⌕
☰
home / papers / 2026 / PAI at SemEval-2026 Task 3: An LLM and Data Redistribution Adaptation-Based Predictive Strategy for Valence-Arousal Scores

PAI at SemEval-2026 Task 3: An LLM and Data Redistribution Adaptation-Based Predictive Strategy for Valence-Arousal Scores

Paper · 2026 · Zhihao Ruan, Kaifeng Yang, Cheng Chen et al.

Summary

This SemEval-2026 Task 3 system paper addresses predicting continuous valence-arousal (VA) scores for aspects in text. This shared step appears in every DimABSA track: DimASTE, DimASQP and DimStance. The authors identify two difficulties. Emotional intensity is perceived differently across languages, and VA annotation is subjective, with valence and arousal strongly correlated.

Their system works in two stages:

  1. Fine-tuning. Qwen3 models (8B/14B/32B) are fine-tuned with LoRA-based SFT on Subtask 1 data. These outperformed few-shot prompting of Gemini-3 and Grok-4 variants.
  2. Recalibration. The raw predictions are adjusted to match the training-set label distribution. Gaussian mapping, quantile mapping and the Sinkhorn optimal-transport algorithm were compared. Only Sinkhorn aligns the joint two-dimensional VA distribution rather than each dimension separately, and it proved the most accurate and stable.

The final submission used Qwen3-32B plus Sinkhorn. It ranked first on eight leaderboards, including Russian, Tatar and Ukrainian restaurant data. For Ukrainian NLP, the paper shows that cheap post-hoc distribution matching measurably improves LLM-based sentiment regression.

Key points

  • First places (team PAI, account archfool), eight in total:
    • Subtask 1: rus-res, tat-res, ukr-res
    • Subtask 2: rus-res, ukr-res, zho-res
    • Subtask 3: rus-res, ukr-res
  • Subtask 1 test error (system vs. strongest baseline; lower is better):
    • ukr-res: 1.1888 vs. 1.7805
    • rus-res: 1.219 vs. 1.7768
    • tat-res: 1.5294 vs. 1.9380
  • Dev results with Sinkhorn in Stage II (best of the three mappings): rus-res 1.0372, tat-res 1.3535, ukr-res 1.0834.
  • LoRA SFT setup:
    • r=8, α=16, dropout 0
    • 1 epoch, learning rate 1e-4, AdamW with cosine schedule
    • BF16, effective batch size 8
  • Sinkhorn setup: squared Euclidean cost, ε=1, up to 10,000 iterations.
  • Model choice: Qwen3-14B scored best on dev, but Qwen3-32B was submitted for robustness. Gains from scaling 8B→32B were marginal.

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, topics and concepts

📚 Shelves 6The 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 worksPsychology · 1,552 worksExperimental and Cognitive Psychology · 870 works

🧭 Topics 1Research subjects and languages the work is about.Language models & pretraining · 3,320 works

Cite

Cite as Zhihao Ruan, Kaifeng Yang, Cheng Chen, Wenwen Dai and Wenjia Mao. 2026. PAI at SemEval-2026 Task 3: An LLM and Data Redistribution Adaptation-Based Predictive Strategy for Valence-Arousal Scores. 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-1.193

@inproceedings{ruan2026pai,
  title = {{PAI at SemEval-2026 Task 3: An LLM and Data Redistribution Adaptation-Based Predictive Strategy for Valence-Arousal Scores}},
  author = {Zhihao Ruan, Kaifeng Yang, Cheng Chen, Wenwen Dai and Wenjia Mao},
  booktitle = {Proceedings of the 20th International Workshop on Semantic Evaluation (2026)},
  year = {2026},
  doi = {10.18653/v1/2026.semeval-1.193},
  url = {https://doi.org/10.18653/v1/2026.semeval-1.193}
}

References

7 references, 1 of them with a card of ours. ↗ leads to the work, «find» searches for one we have no id for yet.

  1. Jonas Becker, Liang-Chih Yu, Shamsuddeen Hassan Muhammad, Jan Philip Wahle, Terry Ruas, Idris Abdulmumin, Lung-Hao Lee, Nelson Odhiambo, Lilian Wanzare, Wen-Ni Liu, Tzu-Mi Lin, Zhe-Yu Xu, Ying-Lung Lin, Jin Wang, Maryam Ibrahim Mukhtar, Bela Gipp, and Saif M. Mohammad. 2026. Dimstance: Multilingual… find
  2. Marco Cuturi. 2013. Sinkhorn distances: Lightspeed computation of optimal transport. Advances in neural information processing systems, 26. find
  3. Google DeepMind. 2025. Gemini 3 flash model card. find
  4. Lung-Hao Lee, Liang-Chih Yu, Natalia Loukashe-vich, Ilseyar Alimova, Alexander Panchenko, Tzu-Mi Lin, Zhe-Yu Xu, Jian-Yu Zhou, Guangmin Zheng, Jin Wang, Sharanya Awasthi, Jonas Becker, Jan Philip Wahle, Terry Ruas, Shamsuddeen Hassan Muhammad, and Saif M. Mohammad. 2026. Dimabsa: Building multilingu… find
  5. James A Russell. 1980. A circumplex model of affect. Journal of personality and social psychology, 39(6):1161. xAI. 2025. Grok 4 model card: Advancing reasoning and agentic capabilities. Technical report, xAI. find
  6. An Yang, Anfeng Li, Baosong Yang, Beichen Zhang, Binyuan Hui, Bo Zheng, Bowen Yu, Chang Gao, Chengen Huang, Chenxu Lv, and 1 others. 2025. Qwen3 technical report. arXiv preprint arXiv:2505.09388.
  7. Liang-Chih Yu, Jonas Becker, Shamsuddeen Hassan Muhammad, Idris Abdulmumin, Lung-Hao Lee, Ying-Lung Lin, Jin Wang, Jan Philip Wahle, Terry Ruas, Alexander Panchenko, Ilseyar Alimova, Kai-Wei Chang, Lilian Wanzare, Nelson Odhiambo, Bela Gipp, and Saif M. Mohammad. 2026. SemEval-2026 task 3: Dimension… find

Updated

↑ back to top

This page in Markdown → https://md.research.elamur.ai/free/en/papers/ruan-2026-pai-semeval-2026-task-3/