Elamur Research⌕
☰
home / papers / 2026 / “AGI” Team at SemEval-2026 Task 2: Predicting Variation in Emotional Valence and Arousal over Time from Ecological Essays

“AGI” Team at SemEval-2026 Task 2: Predicting Variation in Emotional Valence and Arousal over Time from Ecological Essays

Paper · 2026 · Harsh Rathva

Summary

This system paper addresses SemEval-2026 Task 2, which asks models to track emotional valence and arousal over time. The data are ecological essays and feeling-word lists written by U.S. service-industry workers between 2021 and 2024. The author's AffectDynamics system encodes each text with RoBERTa-Large and passes the encodings through a causal, unidirectional GRU, which keeps future texts out of each prediction. A gated user embedding personalises the model and falls back to a global vector for users unseen in training. Training follows a four-phase curriculum:

  • ordinal regression for absolute affect (Subtask 1),
  • a zero-inflated delta model for consecutive-text changes (Subtask 2A),
  • pooled regression for long-term dispositional shift (Subtask 2B).

The system is reasonably strong on Subtask 1 but produces negative correlations on Subtask 2A and near-zero ones on 2B. The paper's main value is its analysis of this failure. Temporal smoothing and a multiplicative delta head favour stable, user-level predictions and push predicted changes toward zero. This suggests that absolute-affect and change-detection tasks may need different architectures.

Key points

  • Data: the full dataset has 5,285 texts from 182 users; the official training split has 2,764 texts from 137 users. Valence ranges from −2 to +2 and arousal from 0 to 2. The author held out 20% of users for internal validation.
  • Subtask 1 (official test): composite r is 0.600 for valence and 0.452 for arousal. For valence, between-user r (0.712) clearly exceeds within-user r (0.489), which supports the stability-bias explanation.
  • Subtasks 2A and 2B: Subtask 2A gives r = −0.167 (valence) and −0.147 (arousal). Subtask 2B correlations are close to zero. A simple linear(prev) baseline exceeds r = 0.6 on 2A.
  • Diagnostics:
    • The variance of predicted deltas is only about 0.15 (valence) and 0.18 (arousal) of the gold-delta variance.
    • Removing the GRU makes 2A correlation slightly positive (about 0.03 / 0.02 on validation).
    • Adding elapsed-time embeddings helps 2A only marginally (from −0.12 to −0.09).
  • Release: code and checkpoints are public on GitHub and Hugging Face (Haxxsh/AffectDynamics-SemEval2026Task2).

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 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 Harsh Rathva. 2026. “AGI” Team at SemEval-2026 Task 2: Predicting Variation in Emotional Valence and Arousal over Time from 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-1.21

@inproceedings{rathva2026agi,
  title = {{“AGI” Team at SemEval-2026 Task 2: Predicting Variation in Emotional Valence and Arousal over Time from Ecological Essays}},
  author = {Harsh Rathva},
  booktitle = {Proceedings of the 20th International Workshop on Semantic Evaluation (2026)},
  year = {2026},
  doi = {10.18653/v1/2026.semeval-1.21},
  url = {https://doi.org/10.18653/v1/2026.semeval-1.21}
}

References

14 references, 5 of them with a card of ours. ↗ leads to the work, «find» searches for one we have no id for yet.

  1. Glenn Affleck, Alex Zautra, Howard Tennen, and Stephen Armeli. 1999. Multilevel daily process designs for consulting and clinical psychology: A preface for the perplexed. Journal of Consulting and Clinical Psychology, 67(5):746–754. find
  2. Rich Caruana. 1997. Multitask learning. Machine Learning, 28(1):41–75. find
  3. Huimin Chen, Maosong Sun, Cunchao Tu, Yankai Lin, and Zhiyuan Liu. 2016. Neural sentiment classification with user and product attention. In Proceedings of the 2016 Conference on Empirical Methods in Natural Language Processing, pages 1650–1659. Association for Computational Linguistics. find
  4. Kyunghyun Cho, Bart Van Merriënboer, Caglar Gulcehre, Dzmitry Bahdanau, Fethi Bougares, Holger Schwenk, and Yoshua Bengio. 2014. Learning phrase representations using RNN encoder-decoder for statistical machine translation. In Proceedings of the 2014 Conference on Empirical Methods in Natural Langu…
  5. David Watson, Lee Anna Clark, and Auke Tellegen. 1988. Development and validation of brief measures of positive and negative affect: The PANAS scales. Journal of Personality and Social Psychology, 54(6):1063–1070. find
  6. A Hardware. Single NVIDIA A100 GPU (16GB+ VRAM), 3-4 hours training. Software. PyTorch 2.0+, Transformers 4.30+, Py-Torch Lightning 2.0+. find
  7. Jacob Devlin, Ming-Wei Chang, Kenton Lee, and Kristina Toutanova. 2019. BERT: Pre-training of deep bidirectional transformers for language understanding. In Proceedings of the 2019 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies,…
  8. Tsung-Yi Lin, Priya Goyal, Ross Girshick, Kaiming He, and Piotr Dollár. 2017. Focal loss for dense object detection. In Proceedings of the IEEE International Conference on Computer Vision, pages 2980–2988.
  9. B Ilya Loshchilov and Frank Hutter. 2019. Decoupled weight decay regularization. In International Conference on Learning Representations (ICLR). find
  10. Value RoBERTa model GRU hidden size GRU layers User embedding dim Dropout Batch size Total epochs Label smoothing Focal loss γ roberta-large 384 2 24 0.1 4 10 0.05 2.0 Data Statistics Statistic Full Train Total texts Users Avg texts/user Median texts/user 5,285 182 72.8 35.0 2,764 137 58.7 31.0 *Tab… 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 Proc…
  12. Parameter *Table 5: Hyperparameter settings.* 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.
  13. Peter McCullagh. 1980. Regression models for ordinal data. Journal of the Royal Statistical Society: Series B (Methodological), 42(2):109–127. find
  14. Implementation Details Variance and Qualitative Analysis Metric Valence Arousal Var(∆ gold ) ˆ Var( ∆) Ratio 1.42 0.21 0.15 0.89 0.16 0.18 *Table 7: Variance comparison confirming systematic under-prediction of change magnitudes.* Yoshihiko Suhara, Yinzhan Xu, and Alex ’Sandy’ Pentland. 2017. DeepMo… find

Updated

↑ back to top

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