| --- |
| language: |
| - en |
|
|
| license: apache-2.0 |
|
|
| tags: |
| - semeval |
| - semeval-2026 |
| - emotion |
| - affect-prediction |
| - temporal-nlp |
| - transformers |
| - roberta |
|
|
| datasets: |
| - semeval |
|
|
| pipeline_tag: text-classification |
|
|
| library_name: transformers |
|
|
| metrics: |
| - pearson-correlation |
| --- |
| |
| # AffectDynamics (Team AGI) — Longitudinal Affect Prediction Model |
|
|
| AffectDynamics is a temporal affect modeling system developed for **SemEval-2026 Task 2: Predicting Variation in Emotional Valence and Arousal over Time from Ecological Essays**. |
|
|
| The model predicts emotional **valence** and **arousal** from longitudinal text written by users across time. It combines transformer-based text encoding with temporal modeling and user-level conditioning to capture both **stable emotional baselines** and **dynamic emotional changes**. |
|
|
| --- |
|
|
| # Model Details |
|
|
| **Model name:** AffectDynamics-SemEval2026Task2 |
| **Developer:** Harsh Rathva |
| **Institution:** Sardar Vallabhbhai National Institute of Technology (SVNIT), Surat |
| **Email:** u24ai036@aid.svnit.ac.in |
|
|
| ## Architecture |
|
|
| The system consists of four main components: |
|
|
| ### 1. Text Encoder |
| - **RoBERTa-Large** transformer encoder |
| - Produces contextual embeddings for each text input. |
|
|
| Different pooling strategies are used depending on text type: |
|
|
| - Essays → CLS / pooler representation |
| - Feeling word lists → mean pooled token embeddings |
|
|
| ### 2. Temporal Encoder |
| - **Unidirectional GRU** |
| - Models longitudinal emotional dynamics across user timelines |
| - Ensures **causal temporal modeling** (no future information leakage) |
|
|
| ### 3. User Conditioning |
| - **Gated user embedding** |
| - Uses user statistics such as: |
| - number of samples |
| - timeline length |
| - emotional entropy |
|
|
| This allows interpolation between **user-specific** and **global representations**. |
|
|
| ### 4. Prediction Heads |
|
|
| | Task | Description | |
| |-----|-------------| |
| | **Subtask 1 (S1)** | Absolute valence and arousal prediction | |
| | **Subtask 2A (S2A)** | Short-term emotional state change prediction | |
| | **Subtask 2B (S2B)** | Long-term dispositional change prediction | |
|
|
| --- |
|
|
| # Training Data |
|
|
| The model was trained using the official **SemEval-2026 Task 2 dataset**. |
|
|
| ### Dataset statistics |
|
|
| - Total texts: **5,285** |
| - Training texts: **2,764** |
| - Users: **182 total (137 training users)** |
| - Time span: **2021–2024** |
|
|
| Each entry contains: |
|
|
| | Field | Description | |
| |------|-------------| |
| | user_id | Anonymous user identifier | |
| | text | Ecological essay or feeling word list | |
| | timestamp | Time of writing | |
| | collection_phase | Study phase | |
| | valence | Emotional valence (-2 to 2) | |
| | arousal | Emotional arousal (0 to 2) | |
|
|
| The texts were written by **U.S. service-industry workers** describing their emotional state. |
|
|
| --- |
|
|
| # Training Details |
|
|
| ### Optimization |
|
|
| - Optimizer: **AdamW** |
| - Scheduler: **OneCycleLR** |
| - Batch size: **4** |
| - Training epochs: **10** |
|
|
| ### Learning Rates |
|
|
| | Component | Learning Rate | |
| |----------|---------------| |
| | RoBERTa encoder | 2e-6 | |
| | GRU | 3e-4 | |
| | Task heads | 2e-5 | |
|
|
| ### Loss Functions |
|
|
| | Task | Loss | |
| |----|----| |
| | Subtask 1 | Ordinal regression with label smoothing | |
| | Subtask 2A | Smooth L1 loss | |
| | Subtask 2B | Mean squared error | |
|
|
| --- |
|
|
| # Evaluation Results |
|
|
| Official evaluation results from **SemEval-2026 Task 2**: |
|
|
| | Task | Metric | Valence | Arousal | |
| |----|----|----|----| |
| | **Subtask 1** | Composite correlation | **0.600** | **0.452** | |
| | **Subtask 2A** | Pearson correlation | -0.167 | -0.147 | |
| | **Subtask 2B** | Pearson correlation | 0.086 | -0.081 | |
|
|
| The model demonstrates strong performance on **absolute affect prediction**, but exhibits limitations in **change detection tasks**, highlighting a trade-off between temporal stability and sensitivity to emotional transitions. |
|
|
| --- |
|
|
| # Intended Use |
|
|
| This model is intended for **research purposes**, including: |
|
|
| - longitudinal affect modeling |
| - emotion prediction from text |
| - temporal NLP modeling |
| - ecological momentary assessment analysis |
|
|
| --- |
|
|
| # Limitations |
|
|
| 1. **Stability bias** |
| - Temporal modeling smooths predictions and reduces sensitivity to abrupt changes. |
|
|
| 2. **Dataset domain** |
| - Data originates from a specific population (U.S. service-industry workers). |
|
|
| 3. **Limited users** |
| - Only **137 users** in training data. |
|
|
| 4. **Change prediction difficulty** |
| - Predicting emotional deltas is harder than predicting absolute states. |
|
|
| --- |
|
|
| # Ethical Considerations |
|
|
| Emotion prediction models must be used responsibly. |
|
|
| Potential concerns include: |
|
|
| - privacy risks from modeling personal emotional data |
| - misuse for manipulation or surveillance |
| - dataset demographic bias |
|
|
| This model **should not be used for clinical or psychological diagnosis**. |
|
|
| --- |
|
|
| # Reproducibility |
|
|
| Code and training pipeline: |
|
|
| https://github.com/ezylopx5/AffectDynamics-SemEval2026Task2 |
|
|
| Model weights: |
|
|
| https://huggingface.co/Haxxsh/AffectDynamics-SemEval2026Task2 |
|
|
| --- |
|
|
| # Citation |
|
|
| If you use this model, please cite: |
