Elamur Research⌕
☰
home / papers / 2026 / BertKittens at SemEval-2026 Task 3: Multi-Domain Aspect Sentiment with BERT/DeBERTa Ensembles for VA Regression and Aspect–Opinion–VA Triplets

BertKittens at SemEval-2026 Task 3: Multi-Domain Aspect Sentiment with BERT/DeBERTa Ensembles for VA Regression and Aspect–Opinion–VA Triplets

Paper · 2026 · Arseny Sukhodolsky, Ruslan Salimgareev and Tatiana Ianshina

Summary

The paper presents the BertKittens system for Track A of SemEval-2026 Task 3 (DimABSA). This task replaces categorical aspect sentiment labels with continuous Valence–Arousal (VA) scores. The team covered Subtask 1 (VA regression for given aspects) and Subtask 2 (extracting aspect–opinion pairs and assigning VA scores) for English, Japanese, Russian and Tatar.

For Subtask 1, a single shared encoder feeds two heads. A token-classification head tags opinion spans, and a regression head outputs VA. The span-tagging objective serves as auxiliary supervision. Subtask 2 uses three separately fine-tuned models in a chain: aspect tagging, then aspect-conditioned opinion tagging, then VA regression. The encoders were microsoft/deberta-v3-large (English), a Japanese DeBERTa-v3-large, DeepPavlov RuBERT (Russian) and multilingual BERT (Tatar).

In Subtask 1 the system placed among the top English entries. Subtask 2 results were much weaker. Mistakes in aspect extraction carried through to every later stage of the chain. The work shows that span supervision helps affective regression, and that cascaded extraction pipelines are fragile under strict end-to-end scoring.

Key points

  • Subtask 1 English rankings: 3rd on laptop (RMSE_VA 1.2769) and 2nd on restaurant (RMSE_VA 1.1812). Japanese RMSE was competitive (hotel 0.7267), but its ranks were lower: 14th on finance and 12th on hotel.
  • Arousal is the weak point: it was much harder to predict than valence. For example, Japanese finance reached PCC_A 0.1157 against PCC_V 0.8074. The authors attribute this to valence being more lexically explicit.
  • Subtask 2 bottleneck: aspect-tagging macro-F1 was about 0.85 for English and 0.80 for Japanese. Final cF1 scores were 0.4469 (English laptop, rank 20), 0.5628 (English restaurant, rank 17), 0.4202 (Japanese hotel), 0.3137 (Russian) and 0.1692 (Tatar).
  • Training choices:
    • Train and dev splits were merged, with 20% held out for validation.
    • English domains were pooled, which beat per-domain models.
    • Russian data received Google Translator back-translation on half of the training set.
    • Dropout was set to 0.2.
  • Scope: Subtask 3 (quadruplets with Entity&Attribute) was dropped because of data loss and time limits. The Ukrainian (ukr-rest) and Chinese datasets released by the organizers were not used.

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.Sentiment & emotion · 972 works

Cite

Cite as Arseny Sukhodolsky, Ruslan Salimgareev and Tatiana Ianshina. 2026. BertKittens at SemEval-2026 Task 3: Multi-Domain Aspect Sentiment with BERT/DeBERTa Ensembles for VA Regression and Aspect–Opinion–VA Triplets. 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-1.380

@inproceedings{sukhodolsky2026bertkittens,
  title = {{BertKittens at SemEval-2026 Task 3: Multi-Domain Aspect Sentiment with BERT/DeBERTa Ensembles for VA Regression and Aspect–Opinion–VA Triplets}},
  author = {Arseny Sukhodolsky, Ruslan Salimgareev and Tatiana Ianshina},
  booktitle = {Proceedings of the 20th International Workshop on Semantic Evaluation (2026)},
  year = {2026},
  doi = {10.18653/v1/2026.semeval-1.380},
  url = {https://doi.org/10.18653/v1/2026.semeval-1.380}
}

References

8 references, 2 of them with a card of ours. ↗ leads to the work, «find» searches for one we have no id for yet.

  1. Jacob Devlin, Ming-Wei Chang, Kenton Lee, and Kristina Toutanova. 2019. Bert: Pre-training of deep bidirectional transformers for language understanding. In Proceedings of NAACL-HLT.
  2. James A. Russell. 2003. Core affect and the psychological construction of emotion. Psychological Review, 110(1):145–172. https://doi.org/10. 1037/0033-295x.110.1.145. find
  3. Wenxuan Zhang, Xin Li, Yang Deng, and Lidong Bing. 2022. A survey on aspect-based sentiment analysis: Tasks, methods, and challenges. arXiv preprint arXiv:2203.01054. https://arxiv.org/ abs/2203.01054. ↗
  4. Pengcheng He, Jianfeng Gao, and Weizhu Chen. 2023. Debertav3: Improving deberta using electra-style pre-training with gradient-disentangled embedding sharing. arXiv preprint arXiv:2111.09543. russian language. In Computational Linguistics and Intellectual Technologies.
  5. 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
  6. Globis University NLP Team. 2023. Deberta-v3 japanese large. Hugging Face model repository. find
  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
  8. Yuri Kuratov and Mikhail Arkhipov. 2019. Adaptation of deep bidirectional multilingual transformers for find

Updated

↑ back to top

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