Year: 2026
Type: conference-paper
Abstract: This paper describes our system for the DimABSA 2026 Shared Task (SemEval-2026 Task 3), which requires predicting continuous Valence-Arousal (VA) scores in [1, 9] for aspect-level sentiments instead of categorical labels.We develop an approach to address all three Track A subtasks: aspect-level VA regression, triplet extraction.The DESS (De-BERTa Enhanced Syntactic-Semantic) model with VA Regression Head achieves 8.22% micro cF1 on the development set on the triplet extraction subtask.We analyze the challenges of adapting categorical ABSA (Aspect Based Sentiment Analysis) architectures to continuous dimensional sentiment and discuss directions for improvement.We release our code at https://github.com/ VishalRepos/DimABSA2026.
Institution: University of Moratuwa
Language: English
Subfield: Artificial Intelligence
Field: Computer Science
Domain: Physical Sciences
Open Access status: hybrid
OpenAlex ID: W7166824570