<![CDATA[Aspect-Based Sentiment Analysis (ABSA) often operates under extreme class imbalance, causing Transformer-based models to favour majority classes while failing to detect minority sentiments disproportionately. This study proposes RoABSA, a robust DeBERTa-v3–based framework that integrates Hybrid Semantic Augmentation, combining Easy Data Augmentation and Back-Translation, with cost-sensitive optimization via Weighted Random Sampling and Focal Loss to enhance diversity representation and recalibrate gradient contributions. Evaluated on four SemEval benchmarks (Lap14, Res14, Res15, Res16), RoABSA achieves consistent state-of-the-art performance with Macro-F1 scores of 84.71, 91.65, 88.40, and 84.24, respectively, outperforming strong graph-based and Transformer-based baselines by margins of up to 10.35 points. Ablation results confirm that robustness emerges from the synergy between augmentation and cost-sensitive learning rather than any single component. At the same time, per-class analysis demonstrates substantial gains in detecting minority sentiments, including cases where baselines failed. These findings highlight that addressing long-tailed distributions in ABSA requires coordinated interventions at both data and optimization levels, and that RoABSA provides a practical, generalizable strategy for improving stability and fairness in sentiment classification.]]>
Copyrights © 2026
