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

2026-06-25 · Source: Paper Index on ACL Anthology · Field: Technology & Digital — Artificial Intelligence & Machine Learning, Natural Language Processing, Data Science & Analytics · Depth: Expert, medium

Summary

The "AGI" Team submitted an approach to SemEval-2026 Task 2, focusing on predicting emotional valence and arousal variations over time from ecological essays. Their model combines RoBERTa-Large for text encoding with a unidirectional GRU for temporal modeling and gated user embeddings for personalization. A four-phase staged training curriculum utilizes ordinal regression for absolute affect prediction and a zero-inflated delta model for change detection. The system achieved competitive performance on Subtask 1 (longitudinal affect assessment) with composite correlations of r=0.600 for valence and r=0.452 for arousal. However, it showed systematic degradation in Subtask 2A (state change detection), exhibiting negative correlations (r=-0.167 for valence, r=-0.147 for arousal). This reveals a fundamental trade-off between stability-oriented representations and change sensitivity in emotional dynamics modeling. Code and trained checkpoints are publicly available.

Key takeaway

For NLP engineers developing systems for longitudinal affect assessment, you should carefully consider the inherent trade-off between model stability and its ability to detect subtle emotional state changes. Your design choices, particularly in temporal modeling and change detection mechanisms, will directly impact performance on these distinct tasks. Prioritize either absolute assessment or change detection based on your application's primary goal, and analyze failure modes empirically.

Key insights

Modeling emotional dynamics reveals a trade-off between stable representations and sensitivity to state changes.

Principles

• Stability-oriented representations hinder change sensitivity.
• Personalization improves temporal affect modeling.

Method

The approach uses RoBERTa-Large, a unidirectional GRU, and gated user embeddings, trained in four phases with ordinal regression and a zero-inflated delta model.

In practice

• Utilize RoBERTa-Large for text encoding.
• Incorporate gated user embeddings.
• Access public code and checkpoints.

Topics

• Emotional Valence
• Emotional Arousal
• SemEval-2026 Task 2
• RoBERTa-Large
• GRU Networks
• Affective Computing
• Temporal Modeling

Best for: Research Scientist, AI Scientist, Machine Learning Engineer, NLP Engineer

Related on AIssential

• From Affect Prediction to Affect Forecasting: Evidence for Distinct Information Sources in Longitudinal Text — Current affect prediction relies on text semantics, while future affective change forecasting benefits from numeric trajectory dynamics.
• Emotional Modulation in Swarm Decision Dynamics — Emotional valence and arousal significantly modulate collective decision dynamics by altering recruitment and inhibition rates.
• Emotion-Attended Stateful Memory (EASM):The Architecture for Hyper-Personalization at Scale — Emotion-attended stateful memory significantly enhances language model personalization by integrating user history and emotional signals.
• How to Fine-Tune an SLM for Emotion Recognition — Fine-tuning SLMs for fine-grained emotion recognition on imbalanced datasets requires a multi-technique approach to data balancing and loss weighting.
• Functional Emotions or Situational Contexts? A Discriminating Test from the Mythos Preview System Card — Model behavior during misalignment may stem from situational contexts, not just functional emotions.
• Beyond Sentiment Classification: A Generative Framework for Emotion Intensity Evaluation in Text — Generative LLMs fine-tuned on continuous emotion intensity scores outperform classification and generalize to unseen affective dimensions.

Open in AIssential →

Editorial summary, takeaway, and curation by AIssential. Original article published by Paper Index on ACL Anthology.