Elamur Research⌕
☰
home / papers / 2026 / DualAxis AI at SemEval-2026 Task 3: Dimensional Aspect-Based Sentiment Analysis

DualAxis AI at SemEval-2026 Task 3: Dimensional Aspect-Based Sentiment Analysis

Paper · 2026 · Yahya Missaoui, Solomon Kebede, Mounika Marreddy et al.

Summary

This paper tackles Dimensional Aspect-Based Sentiment Analysis (DimABSA) from SemEval-2026 Task 3. In this setting, sentiment toward an aspect is encoded as real-valued valence and arousal (each on a 1–9 scale) instead of a positive/negative label. The authors, a Goethe University Frankfurt team, build reproducible baselines for all three subtasks on restaurant and laptop reviews.

For DimASR (aspect-level regression), every annotated aspect becomes its own example by joining it with the review text. A BERT or RoBERTa encoder processes this input, and a feed-forward head on mean-pooled token states is trained with Huber loss on standardized targets.

For DimASTE and DimASQP, Flan-T5 receives an instruction plus the review and writes out triplets or quadruplets as text, which are then parsed back into structured records.

RoBERTa-large reaches the lowest development RMSE. The generative system surpasses the organizer baseline in every test configuration, with the largest margin on laptop quadruplets. The study shows that continuous regression is fairly stable, while joint extraction with continuous scoring remains hard. The released code gives future work a reference point.

Key points

  • Subtask 1 (DimASR, dev set): RoBERTa-large performs best in both domains.
    • Restaurant: average RMSE 0.884, CCC 0.800.
    • Laptop: average RMSE 0.789, CCC 0.778.
    • It outperforms RoBERTa-base and both BERT-base variants (cased and uncased).
  • Subtasks 2–3 (test-set cF1), Flan-T5-large vs. organizer baseline:
    • DimASTE: 0.6024 vs. 0.5442 (restaurant) and 0.5038 vs. 0.4664 (laptop).
    • DimASQP: 0.5426 vs. 0.5048 (restaurant) and 0.5910 vs. 0.2483 (laptop).
    • The abstract cites different figures (0.742/0.648 and 0.604/0.385), probably from another evaluation split.
  • Data format: inputs come as JSONL, with VA labels written as "V#A" strings (e.g., 6.75#6.38) and implicit spans marked NULL.
    • Model selection uses a 90/10 train–validation split.
    • DimASR test labels were not released, so Subtask 1 is reported on dev only.
  • Category effect: on laptops, DimASQP scores higher than DimASTE. The authors credit the closed set of aspect categories (e.g., BATTERY#OPERATION_PERFORMANCE) with acting as a grounding signal.
  • Training and code: training converges within 2–3 epochs. Code is available at github.com/SolomonM-Kebede/ProjectNLP-DimABSA2026.

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

🎓Goethe University FrankfurtFrankfurt am Main · Germany18 works · 373 citations

 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 Yahya Missaoui, Solomon Kebede, Mounika Marreddy and Alexander Mehler. 2026. DualAxis AI at SemEval-2026 Task 3: Dimensional Aspect-Based Sentiment Analysis. 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-1.82

@inproceedings{missaoui2026dualaxis,
  title = {{DualAxis AI at SemEval-2026 Task 3: Dimensional Aspect-Based Sentiment Analysis}},
  author = {Yahya Missaoui, Solomon Kebede, Mounika Marreddy and Alexander Mehler},
  booktitle = {Proceedings of the 20th International Workshop on Semantic Evaluation (2026)},
  year = {2026},
  doi = {10.18653/v1/2026.semeval-1.82},
  url = {https://doi.org/10.18653/v1/2026.semeval-1.82}
}

References

8 references, 3 of them with a card of ours. ↗ leads to the work, «find» searches for one we have no id for yet.

  1. Lu Xu and 1 others. 2020. Position-aware tagging for aspect sentiment triplet extraction. In Proceedings of EMNLP.
  2. Sven Buechel and Udo Hahn. 2017. Emobank: Studying the impact of annotation perspective and representation format on dimensional emotion analysis. In Proceedings of EACL. find
  3. Shaoxiong Chen and 1 others. 2021. Bidirectional machine reading comprehension for aspect sentiment triplet extraction. In Proceedings of AAAI. find
  4. Hyung Won Chung and 1 others. 2022. Scaling instruction-finetuned language models. arXiv preprint arXiv:2210.11416. find
  5. 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.
  6. Liang-Chih Yu and Lung-Hao Lee. 2025. Call for participation – semeval-2026 task 3: Dimensional aspect-based sentiment analysis on customer reviews and stance datasets. ACL Member Portal (Association for Computational Linguistics). Posted Nov 7, 2025. Accessed 2026-02-16. find
  7. Wenxuan Zhang and 1 others. 2021. Aspect sentiment quad prediction as paraphrase generation. arXiv preprint arXiv:2110.00796.
  8. Wenxuan Zhang and 1 others. 2022. A survey on aspect-based sentiment analysis. arXiv preprint arXiv:2203.01054. find

Updated

↑ back to top

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