Elamur Research⌕
☰
home / papers / 2026 / Khaleesiyali at SemEval-2026 Task 2: Lexicon-Augmented RoBERTa for Valence–Arousal Regression on Ecological Essays

Khaleesiyali at SemEval-2026 Task 2: Lexicon-Augmented RoBERTa for Valence–Arousal Regression on Ecological Essays

Paper · 2026 · Eleale Nusi Tee

Summary

The paper describes a system for SemEval-2026 Task 2, Subtask 1, which asks for continuous valence and arousal scores for short English ecological essays and feeling-word texts. The data comprise 5,285 texts from 182 users. Scoring uses user-aware composite metrics that combine between-user and within-user correlation and error. The author fine-tunes roberta-base for joint two-output regression with an MSE loss. Its sentence vector is the first-token representation, extended with six lexicon statistics: the mean, standard deviation and maximum of word valence and of word arousal, drawn from the NRC VAD lexicon. Each text is handled independently, without user identity or temporal order. On the official test set, RoBERTa+Lex clearly beats the organizers' ridge regression over averaged BERT embeddings, with an average composite r of 0.547 versus 0.428. Arousal improves the most. The work shows that cheap, interpretable lexicon priors can complement a fine-tuned transformer. It also offers a static, non-personalized reference point for longitudinal emotion modeling.

Key points

  • Test results: RoBERTa+Lex reaches composite r of 0.656 for valence and 0.438 for arousal, against 0.557 and 0.299 for the linear(BERT) baseline. Average composite MAE falls from 0.601 to 0.532.
  • Lexicon features: NRC VAD covers about 72.05% of training tokens, and no essay ends up with an all-zero feature vector. Dominance scores are left out.
  • Development ablation: On a 2,211/553 train/dev split, adding lexicon features (M2) raises arousal r from 0.500 to 0.545. Valence r drops slightly (0.708 to 0.693) and valence MAE rises.
  • Breakdown by user and text type: Arousal r is higher for unseen users (0.519) than for seen users (0.386). Words-only texts score better than essays. The essay-only figures in the paper's text (0.643/0.369) differ from Table 3 (0.645/0.395).
  • Setup and limitations: Training uses AdamW at a learning rate of 2e-5, batch size 16, up to 10 epochs with early-stopping patience of 2, on one NVIDIA L4 GPU. The author lists the i.i.d. assumption and lack of personalization as limitations, and proposes user-adaptive layers and RNN-style sequence models as future work.

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

🎓Waseda UniversityTokyo · Japan20 works · 584 citations

 Fields, topics and concepts

📚 Shelves 5The field of science this work belongs to (OpenAlex fields).Computer Science · 16,667 worksArtificial Intelligence · 16,194 worksPsychology · 1,552 worksExperimental and Cognitive Psychology · 870 worksSocial Psychology · 306 works

Cite

Cite as Eleale Nusi Tee. 2026. Khaleesiyali at SemEval-2026 Task 2: Lexicon-Augmented RoBERTa for Valence–Arousal Regression on Ecological Essays. 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-1.75

@inproceedings{tee2026khaleesiyali,
  title = {{Khaleesiyali at SemEval-2026 Task 2: Lexicon-Augmented RoBERTa for Valence–Arousal Regression on Ecological Essays}},
  author = {Eleale Nusi Tee},
  booktitle = {Proceedings of the 20th International Workshop on Semantic Evaluation (2026)},
  year = {2026},
  doi = {10.18653/v1/2026.semeval-1.75},
  url = {https://doi.org/10.18653/v1/2026.semeval-1.75}
}

References

16 references, 4 of them with a card of ours. ↗ leads to the work, «find» searches for one we have no id for yet.

  1. Francisca Adoma Acheampong, Henry Nunoo-Mensah, and Wenyu Chen. 2021. Transformer models for text-based emotion detection: a review of bert-based approaches. Artificial Intelligence Review, 54(8):5789– 5829. find
  2. Sven Buechel and Udo Hahn. 2017. Emobank: Studying the impact of annotation perspective and representation format on dimensional emotion analysis. In Conference of the European Chapter of the Association for Computational Linguistics. find
  3. Yunjo Han, Panyu Zhang, Minseo Park, and Uichin Lee. 2024. Systematic evaluation of personalized deep learning models for affect recognition. Proceedings of the ACM on Interactive, Mobile, Wearable and Ubiquitous Technologies, 8(4):1–35. find
  4. Yinhan Liu, Myle Ott, Naman Goyal, Jingfei Du, Mandar Joshi, Danqi Chen, Omer Levy, Mike Lewis, Luke Zettlemoyer, and Veselin Stoyanov. 2019. Roberta: A robustly optimized bert pretraining approach. arXiv preprint arXiv:1907.11692.
  5. Gonçalo Azevedo Mendes and Bruno Martins. 2023. Quantifying valence and arousal in text with multilingual pre-trained transformers. In European Conference on Information Retrieval, pages 84–100. Springer. find
  6. Saif Mohammad. 2018. Obtaining reliable human ratings of valence, arousal, and dominance for 20,000 english words. In Proceedings of the 56th annual meeting of the association for computational linguistics (volume 1: Long papers), pages 174–184.
  7. Saif M. Mohammad. 2025. Nrc vad lexicon v2: Norms for valence, arousal, and dominance for over 55k english terms. arXiv preprint arXiv:2503.23547.
  8. Emily Öhman. 2021. The validity of lexicon-based sentiment analysis in interdisciplinary research. In Proceedings of the Workshop on Natural Language Processing for Digital Humanities, pages 7–12, NIT Silchar, India. NLP Association of India (NLPAI). find
  9. Ji Ho Park, Peng Xu, and Pascale Fung. 2018. Plusemo2vec at semeval-2018 task 1: Exploiting emotion knowledge from emoji and\# hashtags. arXiv preprint arXiv:1804.08280. find
  10. Seungah Son, Andres Saurez, and Dongsoo Har. 2025. Continuous adversarial text representation learning for affective recognition. In 2025 International Conference on Artificial Intelligence in Information and Communication (ICAIIC), pages 0433–0438. IEEE. find
  11. Nikita Soni, H. Andrew Schwartz, Ryan L. Boyd, Phi Long Bui, Syeda Mahwish, August Håkan Nilsson, Adithya V Ganesan, Lyle Ungar, Niranjan Balasubramanian, and Saif M. Mohammad. 2026. SemEval-2026 task 2: Predicting variation in emotional valence and arousal over time from ecological essays. In Proce…
  12. Jessie Sun, H Andrew Schwartz, Youngseo Son, Margaret L Kern, and Simine Vazire. 2020. The language of well-being: Tracking fluctuations in emotion experience through everyday speech. Journal of Personality and Social Psychology, 118(2):364. find
  13. Ali Kargarandehkordi, Matti Kaisti, and Peter Washington. 2024. Personalization of affective models using classical machine learning: a feasibility study. Applied Sciences, 14(4):1337. find
  14. Daniela Teodorescu, Tiffany Cheng, Alona Fyshe, and Saif Mohammad. 2023. Language and mental health: Measures of emotion dynamics from text as linguistic biosocial markers. In Proceedings of the 2023 Conference on Empirical Methods in Natural Language Processing, pages 3117–3133. find
  15. Joe Li, Peter Washington, and 1 others. 2024. A comparison of personalized and generalized approaches to emotion recognition using consumer wearable devices: Machine learning study. JMIR AI, 3(1):e52171. find
  16. Amy Beth Warriner, Victor Kuperman, and Marc Brysbaert. 2013. Norms of valence, arousal, and dominance for 13,915 english lemmas. Behavior research methods, 45(4):1191–1207. find

Updated

↑ back to top

This page in Markdown → https://md.research.elamur.ai/free/en/papers/tee-2026-khaleesiyali-semeval-2026-task-2/