Elamur Research⌕
☰
home / papers / 2026 / PUEB-DimASR at SemEval-2026 Task 3: Escaping the Mean Regression Trap with Graph-Enhanced Transformers for Dimensional Aspect-Based Sentiment Regression

PUEB-DimASR at SemEval-2026 Task 3: Escaping the Mean Regression Trap with Graph-Enhanced Transformers for Dimensional Aspect-Based Sentiment Regression

Paper · 2026 · Oskar Riewe-Perła and Agata Filipowska

Summary

The paper describes the PUEB-DimASR system for SemEval-2026 Task 3 (Track A, Subtask 1, DimASR). The task is to assign continuous valence and arousal scores on a 1–9 scale to each aspect mentioned in a text. The authors focus on the "mean-regression trap": models trained only with MSE tend to predict values near the dataset average and lose the true spread of emotions.

Their baseline is a fine-tuned mdeberta-v3-base with a normalized sigmoid regression head. The proposed GCN-deBERTa adds two GCN layers built on unlabeled spaCy dependency trees, with the [CLS] token linked to all tokens. They compare three training losses: MSE, CCC, and a hybrid MSE + (1 − CCC). All variants are evaluated with 5-fold cross-validation on six languages, including Ukrainian.

GCN-deBERTa beats the baseline in every language. MSE gives the lowest RMSE for English and Chinese but collapses prediction variance. The hybrid loss restores the distribution and gives the best RMSE for Russian, Tatar and Ukrainian. The authors argue that RMSE alone can hide unrealistic output distributions, so a diversity metric should also be reported.

Key points

  • The experiments use the dataset released by the SemEval-2026 Task 3 organizers (Lee et al., 2026), which spans English, Chinese, Japanese, Russian, Ukrainian and Tatar. It has 23,244 sentences and 39,069 aspects in total; Ukrainian, Russian and Tatar each have 1,240 sentences and 2,487 aspects.
  • Dependency parses come from spaCy: uk_core_news_sm for Ukrainian, and the multilingual xx_ent_wiki_sm for Tatar because Tatar has no dedicated model. Training used AdamW for 5 epochs with batch size 64 and seed 42.
  • With the hybrid loss, the best RMSE scores were 1.136 for Russian, 1.178 for Ukrainian and 1.207 for Tatar. With MSE, the best RMSE scores were 0.876 for English and 0.546 for Chinese.
  • Under MSE, arousal diversity (the ratio of predicted to gold standard deviation) fell to 41.6% for Tatar and 43.6% for Russian. The hybrid loss reached about 99.6% distributional alignment, and arousal was consistently harder to predict than valence.
  • The reported results come from refined runs and differ from the official leaderboard submissions. Code is available at github.com/OskarRiewe/SemEval2026_PUEB_DimASR.

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

🎓Poznań University of Economics and BusinessPoznan · Poland3 works · 2 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 1Research subjects and languages the work is about.Sentiment & emotion · 1,035 works

Cite

Cite as Oskar Riewe-Perła and Agata Filipowska. 2026. PUEB-DimASR at SemEval-2026 Task 3: Escaping the Mean Regression Trap with Graph-Enhanced Transformers for Dimensional Aspect-Based Sentiment Regression. 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-1.53

@inproceedings{rieweperla2026pueb,
  title = {{PUEB-DimASR at SemEval-2026 Task 3: Escaping the Mean Regression Trap with Graph-Enhanced Transformers for Dimensional Aspect-Based Sentiment Regression}},
  author = {Oskar Riewe-Perła and Agata Filipowska},
  booktitle = {Proceedings of the 20th International Workshop on Semantic Evaluation (2026)},
  year = {2026},
  doi = {10.18653/v1/2026.semeval-1.53},
  url = {https://doi.org/10.18653/v1/2026.semeval-1.53}
}

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. Pengcheng He, Jianfeng Gao, and Weizhu Chen. 2021. Debertav3: Improving deberta using electra-style pre-training with gradient-disentangled embedding sharing. Preprint, arXiv:2111.09543.
  2. ENG JPN RUS TAT UKR ZHO Domain laptop restaurant finance hotel restaurant restaurant restaurant finance laptop restaurant RMSE 1.7587 1.7011 1.4505 1.2827 2.2749 2.3347 2.2589 0.8179 1.1343 1.2405 *Table 4: Results of the previous version of the model (submitted for evaluation) from the official Sem… find
  3. Minqing Hu and Bing Liu. 2004. Mining and summarizing customer reviews. KDD ’04, page 168–177, New York, NY, USA. Association for Computing Machinery. find
  4. Yan Cathy Hua, Paul Denny, Jörg Wicker, and Katerina Taskova. 2024. A systematic review of aspect-based sentiment analysis: domains, methods, and trends. Artificial Intelligence Review, 57(11):296. find
  5. I Lawrence and Kuei Lin. 1989. A concordance correlation coefficient to evaluate reproducibility. Biometrics, pages 255–268. 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. Lang ENG JPN RUS TAT UKR ZHO Model deBERTa deBERTa deBERTa GCNdeBERTa v2 GCNdeBERTa v2 GCNdeBERTa v2 deBERTa deBERTa deBERTa GCNdeBERTa v2 GCNdeBERTa v2 GCNdeBERTa v2 deBERTa deBERTa deBERTa GCNdeBERTa v2 GCNdeBERTa v2 GCNdeBERTa v2 deBERTa deBERTa deBERTa deBERTa GCNdeBERTa v2 GCNdeBERTa v2 GCNdeBE… find

Updated

↑ back to top

This page in Markdown → https://md.research.elamur.ai/free/en/papers/perla-2026-pueb-dimasr-semeval-2026-task/