Momentum at SemEval-2026 Task 2: LongVA-RoBERTa, a transformer-Based Longitudinal Valence and Arousal Modeling
Year: 2026
Type: conference-paper
Abstract: Emotion modeling is an inherently dynamic phenomenon that changes over time and is influenced by users' affective tendencies.While traditional methods emphasize categorical emotion classification, real-world emotions extend beyond categorical labels, thus can be represented in a continuous valence and arousal space.We propose a method for SemEval 2026 task 2 to address this challenge.The proposed model LongVA-RoBERTa is a transformerbased architecture for Subtask 1 and Subtask 2A.Subtask 1 is a supervised regression problem, in which we predict continuous values for valence and arousal in a two-dimensional space employing partially fine-tuned RoBERTa with attention-pooling mechanism and Multi-Layer Perceptron (MLP).Subtask 2A extends the architecture with a BiLSTM and MLP while keeping the RoBERTa layer unchanged to forecast the short-term longitudinal changes in valence and arousal values.The evaluation metrics used are Pearson correlation coefficient and Mean Absolute Error (MAE).The results for Subtask 1 improved over baseline with competitive performance for Subtask 2A.
Language: English
Subfield: Artificial Intelligence
Field: Computer Science
Domain: Physical Sciences
Open Access status: hybrid
OpenAlex ID: W7166803118