Elamur Research⌕
☰
home / papers / 2026 / Draken at SemEval-2026 Task 2: Frozen BERT Embeddings with Ridge Regression for Predicting Emotional Valence and Arousal

Draken at SemEval-2026 Task 2: Frozen BERT Embeddings with Ridge Regression for Predicting Emotional Valence and Arousal

Paper · 2026 · Rajalakshmi Sivanaiah, Angel Deborah S, Krishna Varun R et al.

Summary

This paper describes Team Draken's entry for Subtask 1 of SemEval-2026 Task 2. The subtask asks systems to predict continuous emotional valence and arousal from series of English ecological essays written by individual users over time. The authors wanted to know whether a cheap, reproducible pipeline could handle this task without fine-tuning or sequence modeling. Their system passes each essay, truncated to 128 WordPiece tokens, through a frozen bert-base-uncased encoder. It then mean-pools the last-layer token states into a 768-dimensional vector and feeds that vector to a multi-output Ridge regressor in scikit-learn, which predicts both dimensions jointly.

On the official test set, valence correlations are fairly strong and arousal correlations are moderate. Breaking results down by user shows that the model ranks users by their typical emotional level better than it tracks how a single user's emotions change over time. Ablations suggest that word-level input is the most informative. The work matters as a transparent, low-cost baseline that shows where frozen-feature approaches fall short for longitudinal affect modeling.

Key points

  • Pipeline: a frozen bert-base-uncased encoder with mean pooling feeds a MultiOutputRegressor wrapping L2-regularized Ridge. The setup uses no fine-tuning, no augmentation, no lexicons and no user metadata. One run takes about 15 minutes on a Google Colab T4 GPU.
  • Test-set composite Pearson r: 0.594 for valence and 0.296 for arousal, averaging 0.445.
  • Between-user vs. within-user correlation: between-user r is higher than within-user r for both dimensions (valence 0.689 vs. 0.479; arousal 0.377 vs. 0.211). Overall MAE is 0.721 for valence and 0.457 for arousal.
  • Seen vs. unseen users: results were reported separately for each group. On unseen users, between-user arousal r reaches 0.509.
  • Input ablation: the words-only input scores highest (valence r 0.639, arousal r 0.449). The essay-only input scores lowest (0.503 and 0.193), and the full input falls in between (0.594 and 0.296).

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

🎓Sri Sivasubramaniya Nadar College of EngineeringChennai · India10 works · 9 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

🧭 Topics 1Research subjects and languages the work is about.Sentiment & emotion · 972 works

Cite

Cite as Rajalakshmi Sivanaiah, Angel Deborah S, Krishna Varun R and Krishnaraj N. 2026. Draken at SemEval-2026 Task 2: Frozen BERT Embeddings with Ridge Regression for Predicting Emotional Valence and Arousal. 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-1.55

@inproceedings{sivanaiah2026draken,
  title = {{Draken at SemEval-2026 Task 2: Frozen BERT Embeddings with Ridge Regression for Predicting Emotional Valence and Arousal}},
  author = {Rajalakshmi Sivanaiah, Angel Deborah S, Krishna Varun R and Krishnaraj N},
  booktitle = {Proceedings of the 20th International Workshop on Semantic Evaluation (2026)},
  year = {2026},
  doi = {10.18653/v1/2026.semeval-1.55},
  url = {https://doi.org/10.18653/v1/2026.semeval-1.55}
}

References

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

  1. Rich Caruana. 1997. Multitask learning. Machine Learning, 28(1):41–75. find
  2. Jacob Devlin, Ming-Wei Chang, Kenton Lee, and Kristina Toutanova. 2019. BERT: Pre-training of deep bidirectional transformers for language understanding. arXiv preprint arXiv:1810.04805.
  3. Arthur E. Hoerl and Robert W. Kennard. 1970. Ridge regression: Biased estimation for nonorthogonal problems. Technometrics, 12(1):55–67. find
  4. Nils Reimers and Iryna Gurevych. 2019. Sentence-bert: Sentence embeddings using siamese bert-networks. In Proceedings of the 2019 Conference on Empirical Methods in Natural Language Processing (EMNLP-IJCNLP), pages 3982–3992.
  5. James A. Russell. 1980. A circumplex model of affect. Journal of Personality and Social Psychology, 39(6):1161–1178. find
  6. 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 Pro…

Updated

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