UKP_Psycontrol at SemEval-2026 Task 2: Modeling Valence and Arousal Dynamics from Text

1. Problem Definition

The paper addresses SemEval-2026 Task 2, which focuses on modeling emotional dynamics in chronologically ordered, self-reported texts (essays and feeling-word lists) from U.S. service-industry workers. The goal is to predict two continuous affective dimensions based on the affective circumplex model:

• Valence: The positivity or negativity of the emotion (0–4 scale).
• Arousal: The intensity or activation level of the emotion (0–2 scale).

The task is divided into two subtasks:

• Subtask 1 (Longitudinal Affect Assessment): Predicting the current valence and arousal for each text in a sequence, for both seen (training) and unseen (new) users.
• Subtask 2A (Forecasting Future Variation): Predicting the change ($\Delta$) in valence and arousal for the next time step based on prior texts and affective states.

The dataset is characterized by small sample sizes (137 users, ~2,764 entries), high inter-user variability, and a mix of free-form essays and short feeling-word lists.

2. Methodology

The authors propose three complementary approaches to tackle the subtasks, leveraging Large Language Models (LLMs), probabilistic graphical models, and neural regression.

A. LLM-Based Prompting

The team utilized GPT-OSS 120B for development and GPT-5 for final submissions. They explored several prompting strategies:

• User-Aware vs. User-Agnostic:
  • User-Aware: Prompts include chronologically ordered examples from the specific target user to capture individual expression patterns.
  • User-Agnostic: Prompts use label-balanced random examples from various users.
• Output Formats:
  • Textual Labels: The model generates emotion labels (e.g., "Very happy"), which are mapped to numerical scores via a predefined grid.
  • Direct Numerical Prediction: The model predicts numbers directly (separately or jointly for valence/arousal).
• Dynamic Context: A sliding window strategy (max $N=15$) was tested to incorporate recent predictions, though this was found to accumulate errors compared to fixed-shot prompting.
• Special Handling: Separate prompts were used for "feeling words" (lists) and "essays" (long-form text) to account for differences in contextual richness.

B. MaxEnt Model with Ising Interactions

A Maximum Entropy (MaxEnt) model with Ising-style pairwise interactions was developed to capture structured dependencies between affective states and semantic features.

• Mechanism: The model defines a probability distribution $P(x) = \frac{1}{Z} \exp(-E(x))$ over binary state vectors.
• Components:
  • Bias terms ($h$): Represent individual effects of valence, arousal, and semantic features.
  • Interaction terms ($J$): Capture how variables co-occur or influence each other (e.g., how a specific semantic cluster correlates with a specific affective transition).
• Semantic Compression: To make exact normalization tractable, high-dimensional semantic vectors (derived from LLM clustering of emotion words) were compressed via an autoencoder, binarized, and appended to the state vector.
• Inference: Predictions are derived via conditional expectations, yielding continuous estimates aligned with the evaluation metrics.

C. Neural Regression Model (Subtask 2A)

A lightweight neural regression model was designed specifically for forecasting affective changes.

• Inputs: A sliding window of recent text embeddings (RoBERTa-base), current valence/arousal, previous state change, and a trainable user embedding.
• Feature Analysis: The team compared three input settings:
  1. No-text baseline: Only affective features and user embeddings.
  2. Text-enhanced: Contextualized text embeddings.
  3. Semantic clusters: Embeddings of emotion cluster names.
• Architecture: A feed-forward network with Layer Normalization, ReLU activations, and dropout.

3. Key Findings and Results

Subtask 1: Longitudinal Assessment

• Performance: The system ranked 1st overall among 25 teams.
  • Arousal: Best performance ($r=0.554$ on test).
  • Valence: 6th best ($r=0.667$ on test).
• LLM Dominance: LLM-based prompting significantly outperformed the MaxEnt model and the BERT baseline.
  • User-Aware vs. Agnostic: User-aware prompting performed marginally better, suggesting LLMs can adapt to individual styles, but user-agnostic models still captured most of the signal.
  • Textual vs. Numerical: Generating textual emotion labels and mapping them to numbers outperformed direct numerical prediction, likely because LLMs are better aligned with natural language descriptors than raw scalar generation.
  • Dynamic Updates: Fixed-shot prompting outperformed dynamic sliding windows, as the latter introduced error propagation.

Subtask 2A: Forecasting Variation

• Performance: The system ranked 1st overall among 14 teams.
  • Valence: $r=0.675$ (Test).
  • Arousal: $r=0.683$ (Test).
• Key Insight: For forecasting change, recent numeric state trajectories were more predictive than textual semantics.
  • The best-performing model used only previous valence, arousal, state change, and user embeddings (no text input).
  • Adding text features or semantic clusters did not improve performance and sometimes degraded it.
• Model Comparison: The neural regression model outperformed the MaxEnt model and LLM-based prompting for this specific subtask. The MaxEnt model showed competitive results for arousal but struggled with valence due to limited sample sizes and the complexity of pairwise interactions.

4. Key Contributions

1. State-of-the-Art Performance: Achieved the top rank in both subtasks of SemEval-2026 Task 2.
2. Empirical Discovery on Dynamics: Demonstrated that while LLMs are superior for static affect assessment (capturing the "what" of the emotion from text), short-term affective dynamics (capturing the "change") are better modeled by numeric trajectories and user-specific embeddings rather than semantic text analysis.
3. Hybrid Methodology: Successfully integrated LLM prompting for static tasks with structured probabilistic (MaxEnt) and regression models for dynamic forecasting, providing a comprehensive toolkit for longitudinal affect analysis.
4. User Embedding Efficacy: Highlighted the critical role of trainable user embeddings in capturing individual emotional baselines and dynamics, especially in low-data regimes.

5. Significance and Limitations

• Significance: The work advances the field of computational affect analysis by distinguishing between the capabilities of LLMs for static interpretation versus the need for trajectory-based modeling for temporal forecasting. It suggests that for mental health monitoring (where this data originates), tracking numerical trends may be more robust than re-analyzing text for every new prediction.
• Limitations:
  • Data Sparsity: 92% of users had data from only a single two-week period, limiting the ability to model long-term mental health trajectories.
  • Proprietary Models: Reliance on GPT-5 and GPT-OSS raises reproducibility and privacy concerns regarding sensitive self-reported data.
  • Stability: LLM performance showed sensitivity to temperature settings, and small leaderboard differences in Subtask 2A were noted to be unstable under minor data perturbations.

In conclusion, the UKP_Psycontrol team successfully demonstrated that a hybrid approach—leveraging the semantic understanding of LLMs for static assessment and the statistical efficiency of regression models for dynamic forecasting—yields the most robust results for modeling emotional dynamics in ecological settings.