Continuous Valence–Arousal Modeling of Sinhala YouTube Comments with Hybrid Multi-Task Learning

Modeling emotion recognition in social media text is challenging due to informal language and limited annotated resources, especially for low-resource languages. While many approaches treat emotion analysis as a classification task, affective psychology models emotions along continuous valence--arousal (V--A) dimensions. In this paper, it is investigated a continuous V--A regression for social media comments, with a focus on Sinhala YouTube data. Firstly it is shown that direct emotion classification using discretized V--A categories yields weak generalization. Then reformulate the task as multi-task regression of continuous valence and arousal values and validate the approach on an English benchmark before adapting it to Sinhala. The hybrid model integrates contextual transformer representations with complementary static embeddings and is optimized through ablation and hyperparameter tuning. The final model achieves strong correlation and $R^2$ scores for both affective dimensions, demonstrating the effectiveness of continuous affect modeling in low-resource social media settings.