Draken at SemEval-2026 Task 2: Frozen BERT Embeddings with Ridge Regression for Predicting Emotional Valence and Arousal
Year: 2026
Type: conference-paper
Abstract: This paper describes Team Draken's submission to Subtask 1 of SemEval-2026 Task 2 on predicting longitudinal variation in emotional valence and arousal from ecological essays written in English.The task involves modeling changes in emotional dimensions over time for individual users, given their free-form essay responses.We adopt a computationally efficient approach that uses frozen contextual embeddings from a pre-trained BERT-base-uncased encoder combined with a multi-output linear regression model with L2 regularization.The system operates in a feature-based fashion without fine-tuning the transformer parameters, relying on mean-pooled sentence representations and a lightweight regression head.Our model obtains strong correlations for valence and moderate correlations for arousal on the official test set, with detailed analyses across seen and unseen users, as well as different input ablations.The results highlight that valence is consistently easier to predict than arousal and that user-level differences are captured more reliably than within-user temporal dynamics.
Institution: Sri Sivasubramaniya Nadar College of Engineering
Language: English
Subfield: Artificial Intelligence
Field: Computer Science
Domain: Physical Sciences
Open Access status: hybrid
OpenAlex ID: W7166833801