SemEval-2026 Task 2: EmoVA

This repository contains the model architecture of our solution for the SemEval-2026 Task 2: EmoVA (Predicting Variation in Emotional Valence and Arousal over Time from Ecological Essays) challenge.

System Description

The system implements a regression-based neural network architecture designed to predict continuous emotional Valence and Arousal (VA) from longitudinal text data. The architecture is split into two primary configurations to address the different subtasks.

Subtask 1 (Longitudinal Affect Assessment)

The primary model processes sequences of user essays to capture emotional variation over time. The pipeline consists of:

• Text Encoder: A Transformer-based language model extracts semantic features from the text. It supports Parameter-Efficient Fine-Tuning using BitFit, which unfreezes only bias terms for memory efficiency.
• Set Attention Pooling: Token embeddings are processed through an Induced Set Attention Block (ISAB) with 32 inducing points to reduce self-attention complexity. The output is compressed into 4 fixed-size document representations per text using Pooling by Multihead Attention (PMA).
• Temporal Modeling: A bidirectional LSTM encoder processes the sequence of document embeddings to capture the longitudinal dynamics of the user's emotional state. It uses pack_padded_sequence to dynamically skip padding tokens and improve computational efficiency.
• Loss Functions: The model is optimized using a combination of Masked Mean Squared Error (MSE) and Concordance Correlation Coefficient (CCC) Loss to maximize both absolute accuracy and relative sequence correlation.

Subtask 2a (Forecasting)

The forecasting model extends the base architecture to predict future emotional state changes.

• Multimodal Fusion: The architecture concatenates the sequence of historical Valence and Arousal scores directly with the PMA-pooled text embeddings before passing them into the LSTM.
• Delta Prediction: The final hidden state of the LSTM is combined with the last known VA values and passed through a specialized prediction head to forecast the single future variation.

---

Repository Structure

src/data/

• __init__.py: Exports dataset and dataloader initialization functions.
• dataset.py: Contains EmoVADataset for loading complete user sequences and EmoVADataset2a for loading sliding-window historical text and VA score sequences.
• collate.py: Defines custom collation functions (create_collate_fn and create_collate_fn_2a) to handle batching, flattening, and padding of variable-length longitudinal text sequences and their targets.
• utils.py: Contains functions like setup_dataloader to initialize the tokenizer, dataset, and batch collator in a single pipeline.

src/models/

• affect_model.py: Defines the main end-to-end PyTorch architectures (AffectModel and AffectModel2a).
• encoder.py: Wrapper for HuggingFace models implementing the frozen backbone and PEFT/BitFit configurations.
• set_attention.py: Implements MAB, ISAB, and PMA layers based on the Set Transformer architecture for sequence pooling.
• lstm.py: Contains the LSTMEncoder utilizing packed sequences to efficiently model temporal patterns while ignoring padding.
• heads.py: Defines the final PredictionHead for outputting VA coordinates.
• tokenizer_wrapper.py: Standardizes tokenization, padding, and truncation logic.

src/evaluation/

• metrics.py: Implements official SemEval metrics, including between-user Pearson correlation, within-user Pearson correlation, and the composite score via Fisher's z-transformation. It also computes metrics for Subtask 2a forecasting.

src/training/

• losses.py: Defines custom objective functions, including a masked Mean Squared Error (masked_mse_loss), a Concordance Correlation Coefficient loss (ccc_loss), and a combined function.
• trainer.py: Execution loop for Subtask 1, handling automatic mixed precision, gradient accumulation, model checkpointing, and evaluation against the composite correlation metric.
• trainer_2a.py: Parallel execution loop adapted for Subtask 2a, optimizing via standard MSE and evaluating via average Pearson correlation.
• utils.py: Provides core training utilities like EarlyStopping and gradient clipping.

Root Directory

• utils.py: General utility functions for root execution scripts.
• eda.ipynb: Exploratory Data Analysis notebook.
• main.ipynb: Primary pipeline notebook for training and evaluating baseline models.
• SEMEVAL2026_EMOVA_ABLATION.ipynb: Execution notebook for isolating and evaluating specific architectural components on Subtask 1.
• SEMEVAL2026_EMOVA_ABLATION_2a.ipynb: Execution notebook for ablation studies specific to the Subtask 2a forecasting models.
• SEMEVAL2026_EMOVA_SUBMISSION.ipynb: Final inference notebook used to format predictions for Codabench submission.
• SemEval2026_EmoVA.pdf: System description paper detailing the architecture and experimental methodology.