Elamur Research⌕
☰
home / papers / 2026 / lamanhnguyen at SemEval-2026 Task 2: Uncovering Lexical Bias and Momentum Lag in Longitudinal Emotion Prediction using Multi-task DeBERTa

lamanhnguyen at SemEval-2026 Task 2: Uncovering Lexical Bias and Momentum Lag in Longitudinal Emotion Prediction using Multi-task DeBERTa

Paper · 2026 · Lam Anh Nguyen

Summary

This paper describes a system for SemEval-2026 Task 2, which asks participants to predict emotional Valence and Arousal from users' chronological diary entries. The task has three subtasks: current emotional state (Subtask 1), short-term change between entries (2a), and the user's long-run baseline (2b). The author treats all of them as multi-task regression with microsoft/deberta-v3-base, trained on the merged Subtask 1 and 2a training data. Two fine-tuned variants are combined in a 0.6/0.4 weighted ensemble. For the longitudinal subtasks, an Exponential Moving Average (EMA) filter with α = 0.1 smooths the raw predictions.

The system does well on point-in-time prediction, with the second-best Valence correlation. However, it yields negative correlations on both longitudinal subtasks. The error analysis traces part of this to a "Venting Effect": the model reads negative vocabulary as low valence, even when writers report feeling relieved after journaling. It also shows that EMA smoothing trades lower MAE for weaker Pearson r. These findings are a useful warning for anyone building emotion trackers over noisy, real-life text.

Key points

  • Ensemble. The two DeBERTa-v3-base variants were V1 (4 epochs, LR 2e-5) and V3 (6 epochs, LR 1e-5). Weights were tuned by grid search, and ensembling cut validation MSE from 0.183 to 0.172.
  • Subtask 1 results. Valence reached r = 0.687 (MAE 0.639) and Arousal reached r = 0.458 (MAE 0.395). This placed 2nd on Valence and 5th on the overall V&A average.
  • Subtask 2 results. Correlations were negative. In 2a, Valence r was −0.273 and Arousal r was −0.275. In 2b, Valence r was −0.398 and Arousal r was −0.577, both significant at p < 0.01.
  • Venting Effect. In the 5% of cases with the largest errors, about 56% contained strongly negative exhaustion-related words but had neutral or positive gold valence.
  • Smoothing ablation (Subtask 2a, Valence). Without smoothing (α = 0.0), the model scored MAE 1.355 with r = 0.142. With α = 0.1, it scored MAE 1.322 with r = −0.273. The author attributes this to variance shrinkage and lag during abrupt mood shifts.

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 Lam Anh Nguyen. 2026. lamanhnguyen at SemEval-2026 Task 2: Uncovering Lexical Bias and Momentum Lag in Longitudinal Emotion Prediction using Multi-task DeBERTa. 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-1.91

@inproceedings{nguyen2026lamanhnguyen,
  title = {{lamanhnguyen at SemEval-2026 Task 2: Uncovering Lexical Bias and Momentum Lag in Longitudinal Emotion Prediction using Multi-task DeBERTa}},
  author = {Lam Anh Nguyen},
  booktitle = {Proceedings of the 20th International Workshop on Semantic Evaluation (2026)},
  year = {2026},
  doi = {10.18653/v1/2026.semeval-1.91},
  url = {https://doi.org/10.18653/v1/2026.semeval-1.91}
}

References

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

  1. Stevie Chancellor and Munmun De Choudhury. 2020. Methods in predictive techniques for mental health status on social media: A critical review. npj Digital Medicine, 3(1):43. find
  2. Sharath Chandra Guntuku, David B Yaden, Margaret L Kern, Lyle H Ungar, and Johannes C Eichstaedt. 2017. Detecting depression and mental illness on social media: An integrative review. Current Opinion in Behavioral Sciences, 18:43–49. essays. In Proceedings of the 20th International Workshop on Seman… find
  3. Adam Tsakalidis, Federico Nanni, Anthony Hills, Jenny Chim, Jiayu Song, and Maria Liakata. 2022. Identifying moments of change from longitudinal user text. In Proceedings of the 60th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 4647–4660. Association… find
  4. Thomas Wolf, Lysandre Debut, Victor Sanh, Julien Chaumond, Clement Delangue, Anthony Moi, Pier-ric Cistac, Tim Rault, Rémi Louf, Morgan Funtowicz, and 1 others. 2020. Transformers: State-of-the-art natural language processing. In Proceedings of the 2020 Conference on Empirical Methods in Natural Lan…
  5. A Our system was built using the HuggingFace Transformers library. The models were trained on a single NVIDIA T4 GPU. Table 3 details the specific configurations for the ensemble components. Hyperparameter Architecture Optimizer Batch Size Max Sequence Length Model V1 Learning Rate Epochs Model V3 L… — DeBERTaV3: Improving DeBERTa using ELECTRA-Style Pre-Training with Gradient-Disentangled Embedding Sharing
  6. Balaji Lakshminarayanan, Alexander Pritzel, and Charles Blundell. 2017. Simple and scalable predictive uncertainty estimation using deep ensembles. In Advances in Neural Information Processing Systems, volume 30. find
  7. Saif M Mohammad. 2018. NRC-VAD lexicon: Real-valued valence, arousal, and dominance ratings for 20,000 English words. arXiv preprint arXiv:1808.01669. — [arXiv:1808.01669](https://arxiv.org/abs/1808.01669) find
  8. Implementation Details and Hyperparameters Value microsoft/deberta-v3-base AdamW 16 128 2 × 10 −5 4 1 × 10 −5 6 0.1 (Subtask 2a) 0.6 (V1) + 0.4 (V3) *Table 3: Hyperparameters for the final ensembled system.* James W Pennebaker, Martha E Francis, and Roger J Booth. 2001. Linguistic Inquiry and Word C… find
  9. Sebastian Ruder. 2017. An overview of multi-task learning in deep neural networks. arXiv preprint arXiv:1706.05098.
  10. 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

Updated

↑ back to top

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