Elamur Research⌕
☰
home / papers / 2026 / RPI Team at SemEval-2026 Task 3: An LLM-Encoder Ensemble for Coarse-to-Fine Valence-Arousal Sentiment Prediction

RPI Team at SemEval-2026 Task 3: An LLM-Encoder Ensemble for Coarse-to-Fine Valence-Arousal Sentiment Prediction

Paper · 2026 · Mohammed Shahid Modi and Boleslaw Szymanski

Summary

The paper describes the RPI system for subtask 1 of SemEval-2026 Task 3 (DimABSA). The task is to predict continuous valence and arousal scores (1–9) for individual aspects in short reviews. The data spans five domains and six languages, namely Chinese, English, Japanese, Russian, Tatar and Ukrainian.

The system works in two stages, from coarse to fine:

  • Coarse stage: two LoRA-tuned Qwen3-8B models, one for valence and one for arousal, each predict an integer bin.
  • Fine stage: their outputs are turned into textual guidance features and passed to a multilingual-e5-large-instruct encoder. These features include the predicted digit, the expected score, bin probabilities and uncertainty. The encoder regresses both dimensions jointly.

Trained on the roughly 58.5K-annotation DimABSA data, the system reaches an official overall RMSE VA of 1.20. Its macro-averaged RMSE VA is 1.097, well ahead of the zero-shot LLM baselines.

An ablation compares the joint model with separate valence and arousal encoders trained on coarsened gold labels. Joint modelling markedly improves arousal correlation. However, the separate models do better on the smaller Russian, Ukrainian and Tatar sets. The work shows that compact LLMs can serve as soft priors for a calibrated encoder.

Key points

  • Coarse-stage training: valence bins are 1–8 and arousal bins are 3–8, because arousal in the training data never falls below 3. The valence LLM uses hard cross-entropy, while the arousal LLM uses Gaussian-smoothed soft targets (σ = 0.8). Both add an expected-value MSE term. LoRA settings: r = 16, α = 32.
  • Results against baselines (macro RMSE VA): E5-VA scores 1.097, compared with 1.800 for Kimi-K2 Thinking and 2.055 for Qwen3-14B zero-shot.
  • Joint vs. separate encoders:
    • The joint model raises arousal PCC from 0.459 to 0.619, about 35% higher.
    • Valence PCC stays similar (0.877 vs 0.871).
    • The joint model wins 7 of 10 language-domain pairs.
  • Ukrainian results: Ukrainian restaurant data (4,205 items) gives RMSE VA 1.549 for the joint model versus 1.391 for the separate models. Russian and Tatar show the same pattern, with Tatar the hardest overall (1.780).
  • Code and weights: model weights are promised at github.com/modim2rpi/dimabsa_rpi.

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

🎓Rensselaer Polytechnic InstituteTroy · United States23 works · 5,092 citations

 Fields, topics and concepts

📚 Shelves 6The field of science this work belongs to (OpenAlex fields).Computer Science · 18,614 worksArtificial Intelligence · 17,965 worksArts and Humanities · 17,450 worksLanguage and Linguistics · 17,296 worksPsychology · 1,649 worksExperimental and Cognitive Psychology · 930 works

🧭 Topics 2Research subjects and languages the work is about.Language models & pretraining · 3,894 worksSentiment & emotion · 1,035 works

Cite

Cite as Mohammed Shahid Modi and Boleslaw Szymanski. 2026. RPI Team at SemEval-2026 Task 3: An LLM-Encoder Ensemble for Coarse-to-Fine Valence-Arousal Sentiment Prediction. 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-1.403

@inproceedings{modi2026rpi,
  title = {{RPI Team at SemEval-2026 Task 3: An LLM-Encoder Ensemble for Coarse-to-Fine Valence-Arousal Sentiment Prediction}},
  author = {Mohammed Shahid Modi and Boleslaw Szymanski},
  booktitle = {Proceedings of the 20th International Workshop on Semantic Evaluation (2026)},
  year = {2026},
  doi = {10.18653/v1/2026.semeval-1.403},
  url = {https://doi.org/10.18653/v1/2026.semeval-1.403}
}

References

13 references, 6 of them with a card of ours. ↗ leads to the work, «find» searches for one we have no id for yet.

  1. Eugene Charniak and Mark Johnson. 2005. Coarse-to-fine n-best parsing and MaxEnt discriminative reranking. In Proceedings of the 43rd Annual Meeting of the Association for Computational Linguistics (ACL’05), pages 173–180, Ann Arbor, Michigan. Association for Computational Linguistics.
  2. Tri Dao, Daniel Y. Fu, Stefano Ermon, Atri Rudra, and Christopher Ré. 2023. Flashattention-2: Faster attention with better parallelism and work partitioning. arXiv preprint arXiv:2307.08691.
  3. Li Dong and Mirella Lapata. 2018. Coarse-to-fine decoding for neural semantic parsing. In Proceedings of the 56th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 731–742, Melbourne, Australia. Association for Computational Linguistics. find
  4. 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.
  5. Chloé Kiddon and Pedro Domingos. 2011. Coarse-to-fine inference and learning for first-order probabilistic models. In Proceedings of the Twenty-Fifth AAAI Conference on Artificial Intelligence (AAAI-11), pages 1049–1056. find
  6. 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
  7. Tony C. W. Mok and Albert C. S. Chung. 2022. Affine medical image registration with coarse-to-fine vision transformer. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 20835–20844. find
  8. Konstantinos P. Panousis, Dino Ienco, and Diego Marcos. 2024. Coarse-to-fine concept bottleneck models. In Advances in Neural Information Processing Systems (NeurIPS). find
  9. James A Russell. 1980. A circumplex model of affect. Journal of personality and social psychology, 39(6):1161. find
  10. Timo Schick, Jonathan Zhange, Hinrich Schütze, and Colin Raffel. 2023. Toolformer: Language models can teach themselves to use tools. In International Conference on Learning Representations.
  11. Qwen Team. 2025. Qwen3 technical report. Preprint, arXiv:2505.09388.
  12. Liang Wang, Nan Yang, Xiaolong Huang, Linjun Yang, Rangan Majumder, and Furu Wei. 2024. Multilingual e5 text embeddings: A technical report. arXiv preprint arXiv:2402.05672.
  13. 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/modi-2026-rpi-team-semeval-2026-task/